1 | /* |
---|
2 | * This file implements the perfmon-2 subsystem which is used |
---|
3 | * to program the IA-64 Performance Monitoring Unit (PMU). |
---|
4 | * |
---|
5 | * The initial version of perfmon.c was written by |
---|
6 | * Ganesh Venkitachalam, IBM Corp. |
---|
7 | * |
---|
8 | * Then it was modified for perfmon-1.x by Stephane Eranian and |
---|
9 | * David Mosberger, Hewlett Packard Co. |
---|
10 | * |
---|
11 | * Version Perfmon-2.x is a rewrite of perfmon-1.x |
---|
12 | * by Stephane Eranian, Hewlett Packard Co. |
---|
13 | * |
---|
14 | * Copyright (C) 1999-2005 Hewlett Packard Co |
---|
15 | * Stephane Eranian <eranian@hpl.hp.com> |
---|
16 | * David Mosberger-Tang <davidm@hpl.hp.com> |
---|
17 | * |
---|
18 | * More information about perfmon available at: |
---|
19 | * http://www.hpl.hp.com/research/linux/perfmon |
---|
20 | * |
---|
21 | * |
---|
22 | * For Xen/IA64 xenoprof |
---|
23 | * Copyright (c) 2006 Isaku Yamahata <yamahata at valinux co jp> |
---|
24 | * VA Linux Systems Japan K.K. |
---|
25 | * |
---|
26 | */ |
---|
27 | |
---|
28 | #include <linux/config.h> |
---|
29 | #include <linux/module.h> |
---|
30 | #include <linux/kernel.h> |
---|
31 | #include <linux/sched.h> |
---|
32 | #include <linux/interrupt.h> |
---|
33 | #include <linux/smp_lock.h> |
---|
34 | #include <linux/proc_fs.h> |
---|
35 | #include <linux/seq_file.h> |
---|
36 | #include <linux/init.h> |
---|
37 | #include <linux/vmalloc.h> |
---|
38 | #include <linux/mm.h> |
---|
39 | #include <linux/sysctl.h> |
---|
40 | #include <linux/list.h> |
---|
41 | #include <linux/file.h> |
---|
42 | #include <linux/poll.h> |
---|
43 | #include <linux/vfs.h> |
---|
44 | #include <linux/pagemap.h> |
---|
45 | #include <linux/mount.h> |
---|
46 | #include <linux/bitops.h> |
---|
47 | #include <linux/capability.h> |
---|
48 | #include <linux/rcupdate.h> |
---|
49 | #include <linux/completion.h> |
---|
50 | |
---|
51 | #ifndef XEN |
---|
52 | #include <asm/errno.h> |
---|
53 | #else |
---|
54 | #include <xen/errno.h> |
---|
55 | #endif |
---|
56 | #include <asm/intrinsics.h> |
---|
57 | #include <asm/page.h> |
---|
58 | #include <asm/perfmon.h> |
---|
59 | #include <asm/processor.h> |
---|
60 | #include <asm/signal.h> |
---|
61 | #include <asm/system.h> |
---|
62 | #include <asm/uaccess.h> |
---|
63 | #include <asm/delay.h> |
---|
64 | |
---|
65 | #ifdef XEN |
---|
66 | #include <xen/guest_access.h> |
---|
67 | #include <asm/hw_irq.h> |
---|
68 | #define CONFIG_PERFMON |
---|
69 | #define pid vcpu_id |
---|
70 | #define thread arch._thread |
---|
71 | #define task_pt_regs vcpu_regs |
---|
72 | #endif |
---|
73 | |
---|
74 | #ifdef CONFIG_PERFMON |
---|
75 | /* |
---|
76 | * perfmon context state |
---|
77 | */ |
---|
78 | #define PFM_CTX_UNLOADED 1 /* context is not loaded onto any task */ |
---|
79 | #define PFM_CTX_LOADED 2 /* context is loaded onto a task */ |
---|
80 | #define PFM_CTX_MASKED 3 /* context is loaded but monitoring is masked due to overflow */ |
---|
81 | #define PFM_CTX_ZOMBIE 4 /* owner of the context is closing it */ |
---|
82 | |
---|
83 | #define PFM_INVALID_ACTIVATION (~0UL) |
---|
84 | |
---|
85 | /* |
---|
86 | * depth of message queue |
---|
87 | */ |
---|
88 | #define PFM_MAX_MSGS 32 |
---|
89 | #define PFM_CTXQ_EMPTY(g) ((g)->ctx_msgq_head == (g)->ctx_msgq_tail) |
---|
90 | |
---|
91 | /* |
---|
92 | * type of a PMU register (bitmask). |
---|
93 | * bitmask structure: |
---|
94 | * bit0 : register implemented |
---|
95 | * bit1 : end marker |
---|
96 | * bit2-3 : reserved |
---|
97 | * bit4 : pmc has pmc.pm |
---|
98 | * bit5 : pmc controls a counter (has pmc.oi), pmd is used as counter |
---|
99 | * bit6-7 : register type |
---|
100 | * bit8-31: reserved |
---|
101 | */ |
---|
102 | #define PFM_REG_NOTIMPL 0x0 /* not implemented at all */ |
---|
103 | #define PFM_REG_IMPL 0x1 /* register implemented */ |
---|
104 | #define PFM_REG_END 0x2 /* end marker */ |
---|
105 | #define PFM_REG_MONITOR (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */ |
---|
106 | #define PFM_REG_COUNTING (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */ |
---|
107 | #define PFM_REG_CONTROL (0x4<<4|PFM_REG_IMPL) /* PMU control register */ |
---|
108 | #define PFM_REG_CONFIG (0x8<<4|PFM_REG_IMPL) /* configuration register */ |
---|
109 | #define PFM_REG_BUFFER (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */ |
---|
110 | |
---|
111 | #define PMC_IS_LAST(i) (pmu_conf->pmc_desc[i].type & PFM_REG_END) |
---|
112 | #define PMD_IS_LAST(i) (pmu_conf->pmd_desc[i].type & PFM_REG_END) |
---|
113 | |
---|
114 | #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags & PFM_REGFL_OVFL_NOTIFY) |
---|
115 | |
---|
116 | /* i assumed unsigned */ |
---|
117 | #define PMC_IS_IMPL(i) (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL)) |
---|
118 | #define PMD_IS_IMPL(i) (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL)) |
---|
119 | |
---|
120 | /* XXX: these assume that register i is implemented */ |
---|
121 | #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING) |
---|
122 | #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING) |
---|
123 | #define PMC_IS_MONITOR(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR) == PFM_REG_MONITOR) |
---|
124 | #define PMC_IS_CONTROL(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL) == PFM_REG_CONTROL) |
---|
125 | |
---|
126 | #define PMC_DFL_VAL(i) pmu_conf->pmc_desc[i].default_value |
---|
127 | #define PMC_RSVD_MASK(i) pmu_conf->pmc_desc[i].reserved_mask |
---|
128 | #define PMD_PMD_DEP(i) pmu_conf->pmd_desc[i].dep_pmd[0] |
---|
129 | #define PMC_PMD_DEP(i) pmu_conf->pmc_desc[i].dep_pmd[0] |
---|
130 | |
---|
131 | #define PFM_NUM_IBRS IA64_NUM_DBG_REGS |
---|
132 | #define PFM_NUM_DBRS IA64_NUM_DBG_REGS |
---|
133 | |
---|
134 | #define CTX_OVFL_NOBLOCK(c) ((c)->ctx_fl_block == 0) |
---|
135 | #define CTX_HAS_SMPL(c) ((c)->ctx_fl_is_sampling) |
---|
136 | #define PFM_CTX_TASK(h) (h)->ctx_task |
---|
137 | |
---|
138 | #define PMU_PMC_OI 5 /* position of pmc.oi bit */ |
---|
139 | |
---|
140 | /* XXX: does not support more than 64 PMDs */ |
---|
141 | #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask) |
---|
142 | #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL) |
---|
143 | |
---|
144 | #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask) |
---|
145 | |
---|
146 | #define CTX_USED_IBR(ctx,n) (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64) |
---|
147 | #define CTX_USED_DBR(ctx,n) (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64) |
---|
148 | #define CTX_USES_DBREGS(ctx) (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1) |
---|
149 | #define PFM_CODE_RR 0 /* requesting code range restriction */ |
---|
150 | #define PFM_DATA_RR 1 /* requestion data range restriction */ |
---|
151 | |
---|
152 | #define PFM_CPUINFO_CLEAR(v) pfm_get_cpu_var(pfm_syst_info) &= ~(v) |
---|
153 | #define PFM_CPUINFO_SET(v) pfm_get_cpu_var(pfm_syst_info) |= (v) |
---|
154 | #define PFM_CPUINFO_GET() pfm_get_cpu_var(pfm_syst_info) |
---|
155 | |
---|
156 | #define RDEP(x) (1UL<<(x)) |
---|
157 | |
---|
158 | /* |
---|
159 | * context protection macros |
---|
160 | * in SMP: |
---|
161 | * - we need to protect against CPU concurrency (spin_lock) |
---|
162 | * - we need to protect against PMU overflow interrupts (local_irq_disable) |
---|
163 | * in UP: |
---|
164 | * - we need to protect against PMU overflow interrupts (local_irq_disable) |
---|
165 | * |
---|
166 | * spin_lock_irqsave()/spin_lock_irqrestore(): |
---|
167 | * in SMP: local_irq_disable + spin_lock |
---|
168 | * in UP : local_irq_disable |
---|
169 | * |
---|
170 | * spin_lock()/spin_lock(): |
---|
171 | * in UP : removed automatically |
---|
172 | * in SMP: protect against context accesses from other CPU. interrupts |
---|
173 | * are not masked. This is useful for the PMU interrupt handler |
---|
174 | * because we know we will not get PMU concurrency in that code. |
---|
175 | */ |
---|
176 | #define PROTECT_CTX(c, f) \ |
---|
177 | do { \ |
---|
178 | DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \ |
---|
179 | spin_lock_irqsave(&(c)->ctx_lock, f); \ |
---|
180 | DPRINT(("spinlocked ctx %p by [%d]\n", c, current->pid)); \ |
---|
181 | } while(0) |
---|
182 | |
---|
183 | #define UNPROTECT_CTX(c, f) \ |
---|
184 | do { \ |
---|
185 | DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \ |
---|
186 | spin_unlock_irqrestore(&(c)->ctx_lock, f); \ |
---|
187 | } while(0) |
---|
188 | |
---|
189 | #define PROTECT_CTX_NOPRINT(c, f) \ |
---|
190 | do { \ |
---|
191 | spin_lock_irqsave(&(c)->ctx_lock, f); \ |
---|
192 | } while(0) |
---|
193 | |
---|
194 | |
---|
195 | #define UNPROTECT_CTX_NOPRINT(c, f) \ |
---|
196 | do { \ |
---|
197 | spin_unlock_irqrestore(&(c)->ctx_lock, f); \ |
---|
198 | } while(0) |
---|
199 | |
---|
200 | |
---|
201 | #define PROTECT_CTX_NOIRQ(c) \ |
---|
202 | do { \ |
---|
203 | spin_lock(&(c)->ctx_lock); \ |
---|
204 | } while(0) |
---|
205 | |
---|
206 | #define UNPROTECT_CTX_NOIRQ(c) \ |
---|
207 | do { \ |
---|
208 | spin_unlock(&(c)->ctx_lock); \ |
---|
209 | } while(0) |
---|
210 | |
---|
211 | |
---|
212 | #ifdef CONFIG_SMP |
---|
213 | |
---|
214 | #define GET_ACTIVATION() pfm_get_cpu_var(pmu_activation_number) |
---|
215 | #define INC_ACTIVATION() pfm_get_cpu_var(pmu_activation_number)++ |
---|
216 | #define SET_ACTIVATION(c) (c)->ctx_last_activation = GET_ACTIVATION() |
---|
217 | |
---|
218 | #else /* !CONFIG_SMP */ |
---|
219 | #define SET_ACTIVATION(t) do {} while(0) |
---|
220 | #define GET_ACTIVATION(t) do {} while(0) |
---|
221 | #define INC_ACTIVATION(t) do {} while(0) |
---|
222 | #endif /* CONFIG_SMP */ |
---|
223 | |
---|
224 | #define SET_PMU_OWNER(t, c) do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0) |
---|
225 | #define GET_PMU_OWNER() pfm_get_cpu_var(pmu_owner) |
---|
226 | #define GET_PMU_CTX() pfm_get_cpu_var(pmu_ctx) |
---|
227 | |
---|
228 | #define LOCK_PFS(g) spin_lock_irqsave(&pfm_sessions.pfs_lock, g) |
---|
229 | #define UNLOCK_PFS(g) spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g) |
---|
230 | |
---|
231 | #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0) |
---|
232 | |
---|
233 | /* |
---|
234 | * cmp0 must be the value of pmc0 |
---|
235 | */ |
---|
236 | #define PMC0_HAS_OVFL(cmp0) (cmp0 & ~0x1UL) |
---|
237 | |
---|
238 | #define PFMFS_MAGIC 0xa0b4d889 |
---|
239 | |
---|
240 | /* |
---|
241 | * debugging |
---|
242 | */ |
---|
243 | #define PFM_DEBUGGING 1 |
---|
244 | #ifdef PFM_DEBUGGING |
---|
245 | #define DPRINT(a) \ |
---|
246 | do { \ |
---|
247 | if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \ |
---|
248 | } while (0) |
---|
249 | |
---|
250 | #define DPRINT_ovfl(a) \ |
---|
251 | do { \ |
---|
252 | if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \ |
---|
253 | } while (0) |
---|
254 | #endif |
---|
255 | |
---|
256 | /* |
---|
257 | * 64-bit software counter structure |
---|
258 | * |
---|
259 | * the next_reset_type is applied to the next call to pfm_reset_regs() |
---|
260 | */ |
---|
261 | typedef struct { |
---|
262 | unsigned long val; /* virtual 64bit counter value */ |
---|
263 | unsigned long lval; /* last reset value */ |
---|
264 | unsigned long long_reset; /* reset value on sampling overflow */ |
---|
265 | unsigned long short_reset; /* reset value on overflow */ |
---|
266 | unsigned long reset_pmds[4]; /* which other pmds to reset when this counter overflows */ |
---|
267 | unsigned long smpl_pmds[4]; /* which pmds are accessed when counter overflow */ |
---|
268 | unsigned long seed; /* seed for random-number generator */ |
---|
269 | unsigned long mask; /* mask for random-number generator */ |
---|
270 | unsigned int flags; /* notify/do not notify */ |
---|
271 | unsigned long eventid; /* overflow event identifier */ |
---|
272 | } pfm_counter_t; |
---|
273 | |
---|
274 | /* |
---|
275 | * context flags |
---|
276 | */ |
---|
277 | typedef struct { |
---|
278 | unsigned int block:1; /* when 1, task will blocked on user notifications */ |
---|
279 | unsigned int system:1; /* do system wide monitoring */ |
---|
280 | unsigned int using_dbreg:1; /* using range restrictions (debug registers) */ |
---|
281 | unsigned int is_sampling:1; /* true if using a custom format */ |
---|
282 | unsigned int excl_idle:1; /* exclude idle task in system wide session */ |
---|
283 | unsigned int going_zombie:1; /* context is zombie (MASKED+blocking) */ |
---|
284 | unsigned int trap_reason:2; /* reason for going into pfm_handle_work() */ |
---|
285 | unsigned int no_msg:1; /* no message sent on overflow */ |
---|
286 | unsigned int can_restart:1; /* allowed to issue a PFM_RESTART */ |
---|
287 | unsigned int reserved:22; |
---|
288 | } pfm_context_flags_t; |
---|
289 | |
---|
290 | #define PFM_TRAP_REASON_NONE 0x0 /* default value */ |
---|
291 | #define PFM_TRAP_REASON_BLOCK 0x1 /* we need to block on overflow */ |
---|
292 | #define PFM_TRAP_REASON_RESET 0x2 /* we need to reset PMDs */ |
---|
293 | |
---|
294 | |
---|
295 | /* |
---|
296 | * perfmon context: encapsulates all the state of a monitoring session |
---|
297 | */ |
---|
298 | |
---|
299 | typedef struct pfm_context { |
---|
300 | spinlock_t ctx_lock; /* context protection */ |
---|
301 | |
---|
302 | pfm_context_flags_t ctx_flags; /* bitmask of flags (block reason incl.) */ |
---|
303 | unsigned int ctx_state; /* state: active/inactive (no bitfield) */ |
---|
304 | |
---|
305 | struct task_struct *ctx_task; /* task to which context is attached */ |
---|
306 | |
---|
307 | unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */ |
---|
308 | |
---|
309 | #ifndef XEN |
---|
310 | struct completion ctx_restart_done; /* use for blocking notification mode */ |
---|
311 | #endif |
---|
312 | |
---|
313 | unsigned long ctx_used_pmds[4]; /* bitmask of PMD used */ |
---|
314 | unsigned long ctx_all_pmds[4]; /* bitmask of all accessible PMDs */ |
---|
315 | unsigned long ctx_reload_pmds[4]; /* bitmask of force reload PMD on ctxsw in */ |
---|
316 | |
---|
317 | unsigned long ctx_all_pmcs[4]; /* bitmask of all accessible PMCs */ |
---|
318 | unsigned long ctx_reload_pmcs[4]; /* bitmask of force reload PMC on ctxsw in */ |
---|
319 | unsigned long ctx_used_monitors[4]; /* bitmask of monitor PMC being used */ |
---|
320 | |
---|
321 | unsigned long ctx_pmcs[IA64_NUM_PMC_REGS]; /* saved copies of PMC values */ |
---|
322 | |
---|
323 | unsigned int ctx_used_ibrs[1]; /* bitmask of used IBR (speedup ctxsw in) */ |
---|
324 | unsigned int ctx_used_dbrs[1]; /* bitmask of used DBR (speedup ctxsw in) */ |
---|
325 | unsigned long ctx_dbrs[IA64_NUM_DBG_REGS]; /* DBR values (cache) when not loaded */ |
---|
326 | unsigned long ctx_ibrs[IA64_NUM_DBG_REGS]; /* IBR values (cache) when not loaded */ |
---|
327 | |
---|
328 | pfm_counter_t ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */ |
---|
329 | |
---|
330 | u64 ctx_saved_psr_up; /* only contains psr.up value */ |
---|
331 | |
---|
332 | unsigned long ctx_last_activation; /* context last activation number for last_cpu */ |
---|
333 | unsigned int ctx_last_cpu; /* CPU id of current or last CPU used (SMP only) */ |
---|
334 | unsigned int ctx_cpu; /* cpu to which perfmon is applied (system wide) */ |
---|
335 | |
---|
336 | int ctx_fd; /* file descriptor used my this context */ |
---|
337 | pfm_ovfl_arg_t ctx_ovfl_arg; /* argument to custom buffer format handler */ |
---|
338 | |
---|
339 | pfm_buffer_fmt_t *ctx_buf_fmt; /* buffer format callbacks */ |
---|
340 | void *ctx_smpl_hdr; /* points to sampling buffer header kernel vaddr */ |
---|
341 | unsigned long ctx_smpl_size; /* size of sampling buffer */ |
---|
342 | void *ctx_smpl_vaddr; /* user level virtual address of smpl buffer */ |
---|
343 | |
---|
344 | #ifndef XEN |
---|
345 | wait_queue_head_t ctx_msgq_wait; |
---|
346 | pfm_msg_t ctx_msgq[PFM_MAX_MSGS]; |
---|
347 | int ctx_msgq_head; |
---|
348 | int ctx_msgq_tail; |
---|
349 | struct fasync_struct *ctx_async_queue; |
---|
350 | |
---|
351 | wait_queue_head_t ctx_zombieq; /* termination cleanup wait queue */ |
---|
352 | #endif |
---|
353 | } pfm_context_t; |
---|
354 | |
---|
355 | /* |
---|
356 | * magic number used to verify that structure is really |
---|
357 | * a perfmon context |
---|
358 | */ |
---|
359 | #define PFM_IS_FILE(f) ((f)->f_op == &pfm_file_ops) |
---|
360 | |
---|
361 | #define PFM_GET_CTX(t) ((pfm_context_t *)(t)->thread.pfm_context) |
---|
362 | |
---|
363 | #ifdef CONFIG_SMP |
---|
364 | #define SET_LAST_CPU(ctx, v) (ctx)->ctx_last_cpu = (v) |
---|
365 | #define GET_LAST_CPU(ctx) (ctx)->ctx_last_cpu |
---|
366 | #else |
---|
367 | #define SET_LAST_CPU(ctx, v) do {} while(0) |
---|
368 | #define GET_LAST_CPU(ctx) do {} while(0) |
---|
369 | #endif |
---|
370 | |
---|
371 | |
---|
372 | #define ctx_fl_block ctx_flags.block |
---|
373 | #define ctx_fl_system ctx_flags.system |
---|
374 | #define ctx_fl_using_dbreg ctx_flags.using_dbreg |
---|
375 | #define ctx_fl_is_sampling ctx_flags.is_sampling |
---|
376 | #define ctx_fl_excl_idle ctx_flags.excl_idle |
---|
377 | #define ctx_fl_going_zombie ctx_flags.going_zombie |
---|
378 | #define ctx_fl_trap_reason ctx_flags.trap_reason |
---|
379 | #define ctx_fl_no_msg ctx_flags.no_msg |
---|
380 | #define ctx_fl_can_restart ctx_flags.can_restart |
---|
381 | |
---|
382 | #define PFM_SET_WORK_PENDING(t, v) do { (t)->thread.pfm_needs_checking = v; } while(0); |
---|
383 | #define PFM_GET_WORK_PENDING(t) (t)->thread.pfm_needs_checking |
---|
384 | |
---|
385 | /* |
---|
386 | * global information about all sessions |
---|
387 | * mostly used to synchronize between system wide and per-process |
---|
388 | */ |
---|
389 | typedef struct { |
---|
390 | spinlock_t pfs_lock; /* lock the structure */ |
---|
391 | |
---|
392 | unsigned int pfs_task_sessions; /* number of per task sessions */ |
---|
393 | unsigned int pfs_sys_sessions; /* number of per system wide sessions */ |
---|
394 | unsigned int pfs_sys_use_dbregs; /* incremented when a system wide session uses debug regs */ |
---|
395 | unsigned int pfs_ptrace_use_dbregs; /* incremented when a process uses debug regs */ |
---|
396 | struct task_struct *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */ |
---|
397 | #ifdef XEN |
---|
398 | #define XENOPROF_TASK ((struct task_struct*)1) |
---|
399 | #endif |
---|
400 | } pfm_session_t; |
---|
401 | |
---|
402 | /* |
---|
403 | * information about a PMC or PMD. |
---|
404 | * dep_pmd[]: a bitmask of dependent PMD registers |
---|
405 | * dep_pmc[]: a bitmask of dependent PMC registers |
---|
406 | */ |
---|
407 | typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs); |
---|
408 | typedef struct { |
---|
409 | unsigned int type; |
---|
410 | int pm_pos; |
---|
411 | unsigned long default_value; /* power-on default value */ |
---|
412 | unsigned long reserved_mask; /* bitmask of reserved bits */ |
---|
413 | pfm_reg_check_t read_check; |
---|
414 | pfm_reg_check_t write_check; |
---|
415 | unsigned long dep_pmd[4]; |
---|
416 | unsigned long dep_pmc[4]; |
---|
417 | } pfm_reg_desc_t; |
---|
418 | |
---|
419 | /* assume cnum is a valid monitor */ |
---|
420 | #define PMC_PM(cnum, val) (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1) |
---|
421 | |
---|
422 | /* |
---|
423 | * This structure is initialized at boot time and contains |
---|
424 | * a description of the PMU main characteristics. |
---|
425 | * |
---|
426 | * If the probe function is defined, detection is based |
---|
427 | * on its return value: |
---|
428 | * - 0 means recognized PMU |
---|
429 | * - anything else means not supported |
---|
430 | * When the probe function is not defined, then the pmu_family field |
---|
431 | * is used and it must match the host CPU family such that: |
---|
432 | * - cpu->family & config->pmu_family != 0 |
---|
433 | */ |
---|
434 | typedef struct { |
---|
435 | unsigned long ovfl_val; /* overflow value for counters */ |
---|
436 | |
---|
437 | pfm_reg_desc_t *pmc_desc; /* detailed PMC register dependencies descriptions */ |
---|
438 | pfm_reg_desc_t *pmd_desc; /* detailed PMD register dependencies descriptions */ |
---|
439 | |
---|
440 | unsigned int num_pmcs; /* number of PMCS: computed at init time */ |
---|
441 | unsigned int num_pmds; /* number of PMDS: computed at init time */ |
---|
442 | unsigned long impl_pmcs[4]; /* bitmask of implemented PMCS */ |
---|
443 | unsigned long impl_pmds[4]; /* bitmask of implemented PMDS */ |
---|
444 | |
---|
445 | char *pmu_name; /* PMU family name */ |
---|
446 | unsigned int pmu_family; /* cpuid family pattern used to identify pmu */ |
---|
447 | unsigned int flags; /* pmu specific flags */ |
---|
448 | unsigned int num_ibrs; /* number of IBRS: computed at init time */ |
---|
449 | unsigned int num_dbrs; /* number of DBRS: computed at init time */ |
---|
450 | unsigned int num_counters; /* PMC/PMD counting pairs : computed at init time */ |
---|
451 | int (*probe)(void); /* customized probe routine */ |
---|
452 | unsigned int use_rr_dbregs:1; /* set if debug registers used for range restriction */ |
---|
453 | } pmu_config_t; |
---|
454 | /* |
---|
455 | * PMU specific flags |
---|
456 | */ |
---|
457 | #define PFM_PMU_IRQ_RESEND 1 /* PMU needs explicit IRQ resend */ |
---|
458 | |
---|
459 | /* |
---|
460 | * debug register related type definitions |
---|
461 | */ |
---|
462 | typedef struct { |
---|
463 | unsigned long ibr_mask:56; |
---|
464 | unsigned long ibr_plm:4; |
---|
465 | unsigned long ibr_ig:3; |
---|
466 | unsigned long ibr_x:1; |
---|
467 | } ibr_mask_reg_t; |
---|
468 | |
---|
469 | typedef struct { |
---|
470 | unsigned long dbr_mask:56; |
---|
471 | unsigned long dbr_plm:4; |
---|
472 | unsigned long dbr_ig:2; |
---|
473 | unsigned long dbr_w:1; |
---|
474 | unsigned long dbr_r:1; |
---|
475 | } dbr_mask_reg_t; |
---|
476 | |
---|
477 | typedef union { |
---|
478 | unsigned long val; |
---|
479 | ibr_mask_reg_t ibr; |
---|
480 | dbr_mask_reg_t dbr; |
---|
481 | } dbreg_t; |
---|
482 | |
---|
483 | |
---|
484 | /* |
---|
485 | * perfmon command descriptions |
---|
486 | */ |
---|
487 | typedef struct { |
---|
488 | int (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs); |
---|
489 | char *cmd_name; |
---|
490 | int cmd_flags; |
---|
491 | unsigned int cmd_narg; |
---|
492 | size_t cmd_argsize; |
---|
493 | int (*cmd_getsize)(void *arg, size_t *sz); |
---|
494 | } pfm_cmd_desc_t; |
---|
495 | |
---|
496 | #define PFM_CMD_FD 0x01 /* command requires a file descriptor */ |
---|
497 | #define PFM_CMD_ARG_READ 0x02 /* command must read argument(s) */ |
---|
498 | #define PFM_CMD_ARG_RW 0x04 /* command must read/write argument(s) */ |
---|
499 | #define PFM_CMD_STOP 0x08 /* command does not work on zombie context */ |
---|
500 | |
---|
501 | |
---|
502 | #define PFM_CMD_NAME(cmd) pfm_cmd_tab[(cmd)].cmd_name |
---|
503 | #define PFM_CMD_READ_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ) |
---|
504 | #define PFM_CMD_RW_ARG(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW) |
---|
505 | #define PFM_CMD_USE_FD(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD) |
---|
506 | #define PFM_CMD_STOPPED(cmd) (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP) |
---|
507 | |
---|
508 | #define PFM_CMD_ARG_MANY -1 /* cannot be zero */ |
---|
509 | |
---|
510 | typedef struct { |
---|
511 | unsigned long pfm_spurious_ovfl_intr_count; /* keep track of spurious ovfl interrupts */ |
---|
512 | unsigned long pfm_replay_ovfl_intr_count; /* keep track of replayed ovfl interrupts */ |
---|
513 | unsigned long pfm_ovfl_intr_count; /* keep track of ovfl interrupts */ |
---|
514 | unsigned long pfm_ovfl_intr_cycles; /* cycles spent processing ovfl interrupts */ |
---|
515 | unsigned long pfm_ovfl_intr_cycles_min; /* min cycles spent processing ovfl interrupts */ |
---|
516 | unsigned long pfm_ovfl_intr_cycles_max; /* max cycles spent processing ovfl interrupts */ |
---|
517 | unsigned long pfm_smpl_handler_calls; |
---|
518 | unsigned long pfm_smpl_handler_cycles; |
---|
519 | char pad[SMP_CACHE_BYTES] ____cacheline_aligned; |
---|
520 | } pfm_stats_t; |
---|
521 | |
---|
522 | /* |
---|
523 | * perfmon internal variables |
---|
524 | */ |
---|
525 | static pfm_stats_t pfm_stats[NR_CPUS]; |
---|
526 | static pfm_session_t pfm_sessions; /* global sessions information */ |
---|
527 | |
---|
528 | #ifndef XEN |
---|
529 | static DEFINE_SPINLOCK(pfm_alt_install_check); |
---|
530 | #endif |
---|
531 | static pfm_intr_handler_desc_t *pfm_alt_intr_handler; |
---|
532 | |
---|
533 | #ifndef XEN |
---|
534 | static struct proc_dir_entry *perfmon_dir; |
---|
535 | #endif |
---|
536 | static pfm_uuid_t pfm_null_uuid = {0,}; |
---|
537 | |
---|
538 | static spinlock_t pfm_buffer_fmt_lock; |
---|
539 | static LIST_HEAD(pfm_buffer_fmt_list); |
---|
540 | |
---|
541 | static pmu_config_t *pmu_conf; |
---|
542 | |
---|
543 | /* sysctl() controls */ |
---|
544 | pfm_sysctl_t pfm_sysctl; |
---|
545 | EXPORT_SYMBOL(pfm_sysctl); |
---|
546 | |
---|
547 | #ifndef XEN |
---|
548 | static ctl_table pfm_ctl_table[]={ |
---|
549 | {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,}, |
---|
550 | {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,}, |
---|
551 | {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,}, |
---|
552 | {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,}, |
---|
553 | { 0, }, |
---|
554 | }; |
---|
555 | static ctl_table pfm_sysctl_dir[] = { |
---|
556 | {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, }, |
---|
557 | {0,}, |
---|
558 | }; |
---|
559 | static ctl_table pfm_sysctl_root[] = { |
---|
560 | {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, }, |
---|
561 | {0,}, |
---|
562 | }; |
---|
563 | static struct ctl_table_header *pfm_sysctl_header; |
---|
564 | |
---|
565 | static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs); |
---|
566 | static int pfm_flush(struct file *filp); |
---|
567 | #endif |
---|
568 | |
---|
569 | #define pfm_get_cpu_var(v) __ia64_per_cpu_var(v) |
---|
570 | #define pfm_get_cpu_data(a,b) per_cpu(a, b) |
---|
571 | |
---|
572 | #ifndef XEN |
---|
573 | static inline void |
---|
574 | pfm_put_task(struct task_struct *task) |
---|
575 | { |
---|
576 | if (task != current) put_task_struct(task); |
---|
577 | } |
---|
578 | |
---|
579 | static inline void |
---|
580 | pfm_set_task_notify(struct task_struct *task) |
---|
581 | { |
---|
582 | struct thread_info *info; |
---|
583 | |
---|
584 | info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE); |
---|
585 | set_bit(TIF_NOTIFY_RESUME, &info->flags); |
---|
586 | } |
---|
587 | |
---|
588 | static inline void |
---|
589 | pfm_clear_task_notify(void) |
---|
590 | { |
---|
591 | clear_thread_flag(TIF_NOTIFY_RESUME); |
---|
592 | } |
---|
593 | |
---|
594 | static inline void |
---|
595 | pfm_reserve_page(unsigned long a) |
---|
596 | { |
---|
597 | SetPageReserved(vmalloc_to_page((void *)a)); |
---|
598 | } |
---|
599 | static inline void |
---|
600 | pfm_unreserve_page(unsigned long a) |
---|
601 | { |
---|
602 | ClearPageReserved(vmalloc_to_page((void*)a)); |
---|
603 | } |
---|
604 | #endif |
---|
605 | |
---|
606 | static inline unsigned long |
---|
607 | pfm_protect_ctx_ctxsw(pfm_context_t *x) |
---|
608 | { |
---|
609 | spin_lock(&(x)->ctx_lock); |
---|
610 | return 0UL; |
---|
611 | } |
---|
612 | |
---|
613 | static inline void |
---|
614 | pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f) |
---|
615 | { |
---|
616 | spin_unlock(&(x)->ctx_lock); |
---|
617 | } |
---|
618 | |
---|
619 | #ifndef XEN |
---|
620 | static inline unsigned int |
---|
621 | pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct) |
---|
622 | { |
---|
623 | return do_munmap(mm, addr, len); |
---|
624 | } |
---|
625 | |
---|
626 | static inline unsigned long |
---|
627 | pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec) |
---|
628 | { |
---|
629 | return get_unmapped_area(file, addr, len, pgoff, flags); |
---|
630 | } |
---|
631 | |
---|
632 | |
---|
633 | static struct super_block * |
---|
634 | pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data) |
---|
635 | { |
---|
636 | return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC); |
---|
637 | } |
---|
638 | |
---|
639 | static struct file_system_type pfm_fs_type = { |
---|
640 | .name = "pfmfs", |
---|
641 | .get_sb = pfmfs_get_sb, |
---|
642 | .kill_sb = kill_anon_super, |
---|
643 | }; |
---|
644 | #endif |
---|
645 | |
---|
646 | DEFINE_PER_CPU(unsigned long, pfm_syst_info); |
---|
647 | DEFINE_PER_CPU(struct task_struct *, pmu_owner); |
---|
648 | DEFINE_PER_CPU(pfm_context_t *, pmu_ctx); |
---|
649 | DEFINE_PER_CPU(unsigned long, pmu_activation_number); |
---|
650 | #ifndef XEN |
---|
651 | EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info); |
---|
652 | |
---|
653 | |
---|
654 | /* forward declaration */ |
---|
655 | static struct file_operations pfm_file_ops; |
---|
656 | #endif |
---|
657 | |
---|
658 | /* |
---|
659 | * forward declarations |
---|
660 | */ |
---|
661 | #ifndef CONFIG_SMP |
---|
662 | static void pfm_lazy_save_regs (struct task_struct *ta); |
---|
663 | #endif |
---|
664 | |
---|
665 | void dump_pmu_state(const char *); |
---|
666 | static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs); |
---|
667 | |
---|
668 | #include "perfmon_itanium.h" |
---|
669 | #include "perfmon_mckinley.h" |
---|
670 | #include "perfmon_montecito.h" |
---|
671 | #include "perfmon_generic.h" |
---|
672 | |
---|
673 | static pmu_config_t *pmu_confs[]={ |
---|
674 | &pmu_conf_mont, |
---|
675 | &pmu_conf_mck, |
---|
676 | &pmu_conf_ita, |
---|
677 | &pmu_conf_gen, /* must be last */ |
---|
678 | NULL |
---|
679 | }; |
---|
680 | |
---|
681 | |
---|
682 | #ifndef XEN |
---|
683 | static int pfm_end_notify_user(pfm_context_t *ctx); |
---|
684 | #endif |
---|
685 | |
---|
686 | static inline void |
---|
687 | pfm_clear_psr_pp(void) |
---|
688 | { |
---|
689 | ia64_rsm(IA64_PSR_PP); |
---|
690 | ia64_srlz_i(); |
---|
691 | } |
---|
692 | |
---|
693 | static inline void |
---|
694 | pfm_set_psr_pp(void) |
---|
695 | { |
---|
696 | ia64_ssm(IA64_PSR_PP); |
---|
697 | ia64_srlz_i(); |
---|
698 | } |
---|
699 | |
---|
700 | static inline void |
---|
701 | pfm_clear_psr_up(void) |
---|
702 | { |
---|
703 | ia64_rsm(IA64_PSR_UP); |
---|
704 | ia64_srlz_i(); |
---|
705 | } |
---|
706 | |
---|
707 | static inline void |
---|
708 | pfm_set_psr_up(void) |
---|
709 | { |
---|
710 | ia64_ssm(IA64_PSR_UP); |
---|
711 | ia64_srlz_i(); |
---|
712 | } |
---|
713 | |
---|
714 | static inline unsigned long |
---|
715 | pfm_get_psr(void) |
---|
716 | { |
---|
717 | unsigned long tmp; |
---|
718 | tmp = ia64_getreg(_IA64_REG_PSR); |
---|
719 | ia64_srlz_i(); |
---|
720 | return tmp; |
---|
721 | } |
---|
722 | |
---|
723 | static inline void |
---|
724 | pfm_set_psr_l(unsigned long val) |
---|
725 | { |
---|
726 | ia64_setreg(_IA64_REG_PSR_L, val); |
---|
727 | ia64_srlz_i(); |
---|
728 | } |
---|
729 | |
---|
730 | static inline void |
---|
731 | pfm_freeze_pmu(void) |
---|
732 | { |
---|
733 | ia64_set_pmc(0,1UL); |
---|
734 | ia64_srlz_d(); |
---|
735 | } |
---|
736 | |
---|
737 | static inline void |
---|
738 | pfm_unfreeze_pmu(void) |
---|
739 | { |
---|
740 | ia64_set_pmc(0,0UL); |
---|
741 | ia64_srlz_d(); |
---|
742 | } |
---|
743 | |
---|
744 | static inline void |
---|
745 | pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs) |
---|
746 | { |
---|
747 | int i; |
---|
748 | |
---|
749 | for (i=0; i < nibrs; i++) { |
---|
750 | ia64_set_ibr(i, ibrs[i]); |
---|
751 | ia64_dv_serialize_instruction(); |
---|
752 | } |
---|
753 | ia64_srlz_i(); |
---|
754 | } |
---|
755 | |
---|
756 | static inline void |
---|
757 | pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs) |
---|
758 | { |
---|
759 | int i; |
---|
760 | |
---|
761 | for (i=0; i < ndbrs; i++) { |
---|
762 | ia64_set_dbr(i, dbrs[i]); |
---|
763 | ia64_dv_serialize_data(); |
---|
764 | } |
---|
765 | ia64_srlz_d(); |
---|
766 | } |
---|
767 | |
---|
768 | /* |
---|
769 | * PMD[i] must be a counter. no check is made |
---|
770 | */ |
---|
771 | static inline unsigned long |
---|
772 | pfm_read_soft_counter(pfm_context_t *ctx, int i) |
---|
773 | { |
---|
774 | return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val); |
---|
775 | } |
---|
776 | |
---|
777 | /* |
---|
778 | * PMD[i] must be a counter. no check is made |
---|
779 | */ |
---|
780 | static inline void |
---|
781 | pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val) |
---|
782 | { |
---|
783 | unsigned long ovfl_val = pmu_conf->ovfl_val; |
---|
784 | |
---|
785 | ctx->ctx_pmds[i].val = val & ~ovfl_val; |
---|
786 | /* |
---|
787 | * writing to unimplemented part is ignore, so we do not need to |
---|
788 | * mask off top part |
---|
789 | */ |
---|
790 | ia64_set_pmd(i, val & ovfl_val); |
---|
791 | } |
---|
792 | |
---|
793 | #ifndef XEN |
---|
794 | static pfm_msg_t * |
---|
795 | pfm_get_new_msg(pfm_context_t *ctx) |
---|
796 | { |
---|
797 | int idx, next; |
---|
798 | |
---|
799 | next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS; |
---|
800 | |
---|
801 | DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail)); |
---|
802 | if (next == ctx->ctx_msgq_head) return NULL; |
---|
803 | |
---|
804 | idx = ctx->ctx_msgq_tail; |
---|
805 | ctx->ctx_msgq_tail = next; |
---|
806 | |
---|
807 | DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx)); |
---|
808 | |
---|
809 | return ctx->ctx_msgq+idx; |
---|
810 | } |
---|
811 | |
---|
812 | static pfm_msg_t * |
---|
813 | pfm_get_next_msg(pfm_context_t *ctx) |
---|
814 | { |
---|
815 | pfm_msg_t *msg; |
---|
816 | |
---|
817 | DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail)); |
---|
818 | |
---|
819 | if (PFM_CTXQ_EMPTY(ctx)) return NULL; |
---|
820 | |
---|
821 | /* |
---|
822 | * get oldest message |
---|
823 | */ |
---|
824 | msg = ctx->ctx_msgq+ctx->ctx_msgq_head; |
---|
825 | |
---|
826 | /* |
---|
827 | * and move forward |
---|
828 | */ |
---|
829 | ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS; |
---|
830 | |
---|
831 | DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, msg->pfm_gen_msg.msg_type)); |
---|
832 | |
---|
833 | return msg; |
---|
834 | } |
---|
835 | |
---|
836 | static void |
---|
837 | pfm_reset_msgq(pfm_context_t *ctx) |
---|
838 | { |
---|
839 | ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0; |
---|
840 | DPRINT(("ctx=%p msgq reset\n", ctx)); |
---|
841 | } |
---|
842 | |
---|
843 | static void * |
---|
844 | pfm_rvmalloc(unsigned long size) |
---|
845 | { |
---|
846 | void *mem; |
---|
847 | unsigned long addr; |
---|
848 | |
---|
849 | size = PAGE_ALIGN(size); |
---|
850 | mem = vmalloc(size); |
---|
851 | if (mem) { |
---|
852 | //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem); |
---|
853 | memset(mem, 0, size); |
---|
854 | addr = (unsigned long)mem; |
---|
855 | while (size > 0) { |
---|
856 | pfm_reserve_page(addr); |
---|
857 | addr+=PAGE_SIZE; |
---|
858 | size-=PAGE_SIZE; |
---|
859 | } |
---|
860 | } |
---|
861 | return mem; |
---|
862 | } |
---|
863 | |
---|
864 | static void |
---|
865 | pfm_rvfree(void *mem, unsigned long size) |
---|
866 | { |
---|
867 | unsigned long addr; |
---|
868 | |
---|
869 | if (mem) { |
---|
870 | DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size)); |
---|
871 | addr = (unsigned long) mem; |
---|
872 | while ((long) size > 0) { |
---|
873 | pfm_unreserve_page(addr); |
---|
874 | addr+=PAGE_SIZE; |
---|
875 | size-=PAGE_SIZE; |
---|
876 | } |
---|
877 | vfree(mem); |
---|
878 | } |
---|
879 | return; |
---|
880 | } |
---|
881 | #endif |
---|
882 | |
---|
883 | static pfm_context_t * |
---|
884 | pfm_context_alloc(void) |
---|
885 | { |
---|
886 | pfm_context_t *ctx; |
---|
887 | |
---|
888 | /* |
---|
889 | * allocate context descriptor |
---|
890 | * must be able to free with interrupts disabled |
---|
891 | */ |
---|
892 | ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL); |
---|
893 | if (ctx) { |
---|
894 | memset(ctx, 0, sizeof(pfm_context_t)); |
---|
895 | DPRINT(("alloc ctx @%p\n", ctx)); |
---|
896 | } |
---|
897 | return ctx; |
---|
898 | } |
---|
899 | |
---|
900 | static void |
---|
901 | pfm_context_free(pfm_context_t *ctx) |
---|
902 | { |
---|
903 | if (ctx) { |
---|
904 | DPRINT(("free ctx @%p\n", ctx)); |
---|
905 | kfree(ctx); |
---|
906 | } |
---|
907 | } |
---|
908 | |
---|
909 | #ifndef XEN |
---|
910 | static void |
---|
911 | pfm_mask_monitoring(struct task_struct *task) |
---|
912 | { |
---|
913 | pfm_context_t *ctx = PFM_GET_CTX(task); |
---|
914 | struct thread_struct *th = &task->thread; |
---|
915 | unsigned long mask, val, ovfl_mask; |
---|
916 | int i; |
---|
917 | |
---|
918 | DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid)); |
---|
919 | |
---|
920 | ovfl_mask = pmu_conf->ovfl_val; |
---|
921 | /* |
---|
922 | * monitoring can only be masked as a result of a valid |
---|
923 | * counter overflow. In UP, it means that the PMU still |
---|
924 | * has an owner. Note that the owner can be different |
---|
925 | * from the current task. However the PMU state belongs |
---|
926 | * to the owner. |
---|
927 | * In SMP, a valid overflow only happens when task is |
---|
928 | * current. Therefore if we come here, we know that |
---|
929 | * the PMU state belongs to the current task, therefore |
---|
930 | * we can access the live registers. |
---|
931 | * |
---|
932 | * So in both cases, the live register contains the owner's |
---|
933 | * state. We can ONLY touch the PMU registers and NOT the PSR. |
---|
934 | * |
---|
935 | * As a consequence to this call, the thread->pmds[] array |
---|
936 | * contains stale information which must be ignored |
---|
937 | * when context is reloaded AND monitoring is active (see |
---|
938 | * pfm_restart). |
---|
939 | */ |
---|
940 | mask = ctx->ctx_used_pmds[0]; |
---|
941 | for (i = 0; mask; i++, mask>>=1) { |
---|
942 | /* skip non used pmds */ |
---|
943 | if ((mask & 0x1) == 0) continue; |
---|
944 | val = ia64_get_pmd(i); |
---|
945 | |
---|
946 | if (PMD_IS_COUNTING(i)) { |
---|
947 | /* |
---|
948 | * we rebuild the full 64 bit value of the counter |
---|
949 | */ |
---|
950 | ctx->ctx_pmds[i].val += (val & ovfl_mask); |
---|
951 | } else { |
---|
952 | ctx->ctx_pmds[i].val = val; |
---|
953 | } |
---|
954 | DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n", |
---|
955 | i, |
---|
956 | ctx->ctx_pmds[i].val, |
---|
957 | val & ovfl_mask)); |
---|
958 | } |
---|
959 | /* |
---|
960 | * mask monitoring by setting the privilege level to 0 |
---|
961 | * we cannot use psr.pp/psr.up for this, it is controlled by |
---|
962 | * the user |
---|
963 | * |
---|
964 | * if task is current, modify actual registers, otherwise modify |
---|
965 | * thread save state, i.e., what will be restored in pfm_load_regs() |
---|
966 | */ |
---|
967 | mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER; |
---|
968 | for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) { |
---|
969 | if ((mask & 0x1) == 0UL) continue; |
---|
970 | ia64_set_pmc(i, th->pmcs[i] & ~0xfUL); |
---|
971 | th->pmcs[i] &= ~0xfUL; |
---|
972 | DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i])); |
---|
973 | } |
---|
974 | /* |
---|
975 | * make all of this visible |
---|
976 | */ |
---|
977 | ia64_srlz_d(); |
---|
978 | } |
---|
979 | |
---|
980 | /* |
---|
981 | * must always be done with task == current |
---|
982 | * |
---|
983 | * context must be in MASKED state when calling |
---|
984 | */ |
---|
985 | static void |
---|
986 | pfm_restore_monitoring(struct task_struct *task) |
---|
987 | { |
---|
988 | pfm_context_t *ctx = PFM_GET_CTX(task); |
---|
989 | struct thread_struct *th = &task->thread; |
---|
990 | unsigned long mask, ovfl_mask; |
---|
991 | unsigned long psr, val; |
---|
992 | int i, is_system; |
---|
993 | |
---|
994 | is_system = ctx->ctx_fl_system; |
---|
995 | ovfl_mask = pmu_conf->ovfl_val; |
---|
996 | |
---|
997 | if (task != current) { |
---|
998 | printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid); |
---|
999 | return; |
---|
1000 | } |
---|
1001 | if (ctx->ctx_state != PFM_CTX_MASKED) { |
---|
1002 | printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__, |
---|
1003 | task->pid, current->pid, ctx->ctx_state); |
---|
1004 | return; |
---|
1005 | } |
---|
1006 | psr = pfm_get_psr(); |
---|
1007 | /* |
---|
1008 | * monitoring is masked via the PMC. |
---|
1009 | * As we restore their value, we do not want each counter to |
---|
1010 | * restart right away. We stop monitoring using the PSR, |
---|
1011 | * restore the PMC (and PMD) and then re-establish the psr |
---|
1012 | * as it was. Note that there can be no pending overflow at |
---|
1013 | * this point, because monitoring was MASKED. |
---|
1014 | * |
---|
1015 | * system-wide session are pinned and self-monitoring |
---|
1016 | */ |
---|
1017 | if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) { |
---|
1018 | /* disable dcr pp */ |
---|
1019 | ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP); |
---|
1020 | pfm_clear_psr_pp(); |
---|
1021 | } else { |
---|
1022 | pfm_clear_psr_up(); |
---|
1023 | } |
---|
1024 | /* |
---|
1025 | * first, we restore the PMD |
---|
1026 | */ |
---|
1027 | mask = ctx->ctx_used_pmds[0]; |
---|
1028 | for (i = 0; mask; i++, mask>>=1) { |
---|
1029 | /* skip non used pmds */ |
---|
1030 | if ((mask & 0x1) == 0) continue; |
---|
1031 | |
---|
1032 | if (PMD_IS_COUNTING(i)) { |
---|
1033 | /* |
---|
1034 | * we split the 64bit value according to |
---|
1035 | * counter width |
---|
1036 | */ |
---|
1037 | val = ctx->ctx_pmds[i].val & ovfl_mask; |
---|
1038 | ctx->ctx_pmds[i].val &= ~ovfl_mask; |
---|
1039 | } else { |
---|
1040 | val = ctx->ctx_pmds[i].val; |
---|
1041 | } |
---|
1042 | ia64_set_pmd(i, val); |
---|
1043 | |
---|
1044 | DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n", |
---|
1045 | i, |
---|
1046 | ctx->ctx_pmds[i].val, |
---|
1047 | val)); |
---|
1048 | } |
---|
1049 | /* |
---|
1050 | * restore the PMCs |
---|
1051 | */ |
---|
1052 | mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER; |
---|
1053 | for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) { |
---|
1054 | if ((mask & 0x1) == 0UL) continue; |
---|
1055 | th->pmcs[i] = ctx->ctx_pmcs[i]; |
---|
1056 | ia64_set_pmc(i, th->pmcs[i]); |
---|
1057 | DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i])); |
---|
1058 | } |
---|
1059 | ia64_srlz_d(); |
---|
1060 | |
---|
1061 | /* |
---|
1062 | * must restore DBR/IBR because could be modified while masked |
---|
1063 | * XXX: need to optimize |
---|
1064 | */ |
---|
1065 | if (ctx->ctx_fl_using_dbreg) { |
---|
1066 | pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs); |
---|
1067 | pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs); |
---|
1068 | } |
---|
1069 | |
---|
1070 | /* |
---|
1071 | * now restore PSR |
---|
1072 | */ |
---|
1073 | if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) { |
---|
1074 | /* enable dcr pp */ |
---|
1075 | ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP); |
---|
1076 | ia64_srlz_i(); |
---|
1077 | } |
---|
1078 | pfm_set_psr_l(psr); |
---|
1079 | } |
---|
1080 | #endif |
---|
1081 | |
---|
1082 | static inline void |
---|
1083 | pfm_save_pmds(unsigned long *pmds, unsigned long mask) |
---|
1084 | { |
---|
1085 | int i; |
---|
1086 | |
---|
1087 | ia64_srlz_d(); |
---|
1088 | |
---|
1089 | for (i=0; mask; i++, mask>>=1) { |
---|
1090 | if (mask & 0x1) pmds[i] = ia64_get_pmd(i); |
---|
1091 | } |
---|
1092 | } |
---|
1093 | |
---|
1094 | #ifndef XEN |
---|
1095 | /* |
---|
1096 | * reload from thread state (used for ctxw only) |
---|
1097 | */ |
---|
1098 | static inline void |
---|
1099 | pfm_restore_pmds(unsigned long *pmds, unsigned long mask) |
---|
1100 | { |
---|
1101 | int i; |
---|
1102 | unsigned long val, ovfl_val = pmu_conf->ovfl_val; |
---|
1103 | |
---|
1104 | for (i=0; mask; i++, mask>>=1) { |
---|
1105 | if ((mask & 0x1) == 0) continue; |
---|
1106 | val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i]; |
---|
1107 | ia64_set_pmd(i, val); |
---|
1108 | } |
---|
1109 | ia64_srlz_d(); |
---|
1110 | } |
---|
1111 | |
---|
1112 | /* |
---|
1113 | * propagate PMD from context to thread-state |
---|
1114 | */ |
---|
1115 | static inline void |
---|
1116 | pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx) |
---|
1117 | { |
---|
1118 | struct thread_struct *thread = &task->thread; |
---|
1119 | unsigned long ovfl_val = pmu_conf->ovfl_val; |
---|
1120 | unsigned long mask = ctx->ctx_all_pmds[0]; |
---|
1121 | unsigned long val; |
---|
1122 | int i; |
---|
1123 | |
---|
1124 | DPRINT(("mask=0x%lx\n", mask)); |
---|
1125 | |
---|
1126 | for (i=0; mask; i++, mask>>=1) { |
---|
1127 | |
---|
1128 | val = ctx->ctx_pmds[i].val; |
---|
1129 | |
---|
1130 | /* |
---|
1131 | * We break up the 64 bit value into 2 pieces |
---|
1132 | * the lower bits go to the machine state in the |
---|
1133 | * thread (will be reloaded on ctxsw in). |
---|
1134 | * The upper part stays in the soft-counter. |
---|
1135 | */ |
---|
1136 | if (PMD_IS_COUNTING(i)) { |
---|
1137 | ctx->ctx_pmds[i].val = val & ~ovfl_val; |
---|
1138 | val &= ovfl_val; |
---|
1139 | } |
---|
1140 | thread->pmds[i] = val; |
---|
1141 | |
---|
1142 | DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n", |
---|
1143 | i, |
---|
1144 | thread->pmds[i], |
---|
1145 | ctx->ctx_pmds[i].val)); |
---|
1146 | } |
---|
1147 | } |
---|
1148 | #else |
---|
1149 | static inline void |
---|
1150 | xenpfm_restore_pmds(pfm_context_t* ctx) |
---|
1151 | { |
---|
1152 | int i; |
---|
1153 | unsigned long ovfl_val = pmu_conf->ovfl_val; |
---|
1154 | unsigned long mask = ctx->ctx_all_pmds[0]; |
---|
1155 | unsigned long val; |
---|
1156 | |
---|
1157 | for (i = 0; mask; i++, mask >>= 1) { |
---|
1158 | if ((mask & 0x1) == 0) |
---|
1159 | continue; |
---|
1160 | |
---|
1161 | val = ctx->ctx_pmds[i].val; |
---|
1162 | /* |
---|
1163 | * We break up the 64 bit value into 2 pieces |
---|
1164 | * the lower bits go to the machine state in the |
---|
1165 | * thread (will be reloaded on ctxsw in). |
---|
1166 | * The upper part stays in the soft-counter. |
---|
1167 | */ |
---|
1168 | if (PMD_IS_COUNTING(i)) { |
---|
1169 | ctx->ctx_pmds[i].val = val & ~ovfl_val; |
---|
1170 | val &= ovfl_val; |
---|
1171 | } |
---|
1172 | ia64_set_pmd(i, val); |
---|
1173 | } |
---|
1174 | ia64_srlz_d(); |
---|
1175 | } |
---|
1176 | #endif |
---|
1177 | |
---|
1178 | #ifndef XEN |
---|
1179 | /* |
---|
1180 | * propagate PMC from context to thread-state |
---|
1181 | */ |
---|
1182 | static inline void |
---|
1183 | pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx) |
---|
1184 | { |
---|
1185 | struct thread_struct *thread = &task->thread; |
---|
1186 | unsigned long mask = ctx->ctx_all_pmcs[0]; |
---|
1187 | int i; |
---|
1188 | |
---|
1189 | DPRINT(("mask=0x%lx\n", mask)); |
---|
1190 | |
---|
1191 | for (i=0; mask; i++, mask>>=1) { |
---|
1192 | /* masking 0 with ovfl_val yields 0 */ |
---|
1193 | thread->pmcs[i] = ctx->ctx_pmcs[i]; |
---|
1194 | DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i])); |
---|
1195 | } |
---|
1196 | } |
---|
1197 | |
---|
1198 | |
---|
1199 | |
---|
1200 | static inline void |
---|
1201 | pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask) |
---|
1202 | { |
---|
1203 | int i; |
---|
1204 | |
---|
1205 | for (i=0; mask; i++, mask>>=1) { |
---|
1206 | if ((mask & 0x1) == 0) continue; |
---|
1207 | ia64_set_pmc(i, pmcs[i]); |
---|
1208 | } |
---|
1209 | ia64_srlz_d(); |
---|
1210 | } |
---|
1211 | #else |
---|
1212 | static inline void |
---|
1213 | xenpfm_restore_pmcs(pfm_context_t* ctx) |
---|
1214 | { |
---|
1215 | int i; |
---|
1216 | unsigned long mask = ctx->ctx_all_pmcs[0]; |
---|
1217 | |
---|
1218 | for (i = 0; mask; i++, mask >>= 1) { |
---|
1219 | if ((mask & 0x1) == 0) |
---|
1220 | continue; |
---|
1221 | ia64_set_pmc(i, ctx->ctx_pmcs[i]); |
---|
1222 | DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i])); |
---|
1223 | } |
---|
1224 | ia64_srlz_d(); |
---|
1225 | |
---|
1226 | } |
---|
1227 | #endif |
---|
1228 | |
---|
1229 | static inline int |
---|
1230 | pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b) |
---|
1231 | { |
---|
1232 | return memcmp(a, b, sizeof(pfm_uuid_t)); |
---|
1233 | } |
---|
1234 | |
---|
1235 | static inline int |
---|
1236 | pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs) |
---|
1237 | { |
---|
1238 | int ret = 0; |
---|
1239 | if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs); |
---|
1240 | return ret; |
---|
1241 | } |
---|
1242 | |
---|
1243 | static inline int |
---|
1244 | pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size) |
---|
1245 | { |
---|
1246 | int ret = 0; |
---|
1247 | if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size); |
---|
1248 | return ret; |
---|
1249 | } |
---|
1250 | |
---|
1251 | |
---|
1252 | static inline int |
---|
1253 | pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, |
---|
1254 | int cpu, void *arg) |
---|
1255 | { |
---|
1256 | int ret = 0; |
---|
1257 | if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg); |
---|
1258 | return ret; |
---|
1259 | } |
---|
1260 | |
---|
1261 | static inline int |
---|
1262 | pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags, |
---|
1263 | int cpu, void *arg) |
---|
1264 | { |
---|
1265 | int ret = 0; |
---|
1266 | if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg); |
---|
1267 | return ret; |
---|
1268 | } |
---|
1269 | |
---|
1270 | static inline int |
---|
1271 | pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs) |
---|
1272 | { |
---|
1273 | int ret = 0; |
---|
1274 | if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs); |
---|
1275 | return ret; |
---|
1276 | } |
---|
1277 | |
---|
1278 | static inline int |
---|
1279 | pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs) |
---|
1280 | { |
---|
1281 | int ret = 0; |
---|
1282 | if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs); |
---|
1283 | return ret; |
---|
1284 | } |
---|
1285 | |
---|
1286 | static pfm_buffer_fmt_t * |
---|
1287 | __pfm_find_buffer_fmt(pfm_uuid_t uuid) |
---|
1288 | { |
---|
1289 | struct list_head * pos; |
---|
1290 | pfm_buffer_fmt_t * entry; |
---|
1291 | |
---|
1292 | list_for_each(pos, &pfm_buffer_fmt_list) { |
---|
1293 | entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list); |
---|
1294 | if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0) |
---|
1295 | return entry; |
---|
1296 | } |
---|
1297 | return NULL; |
---|
1298 | } |
---|
1299 | |
---|
1300 | /* |
---|
1301 | * find a buffer format based on its uuid |
---|
1302 | */ |
---|
1303 | static pfm_buffer_fmt_t * |
---|
1304 | pfm_find_buffer_fmt(pfm_uuid_t uuid) |
---|
1305 | { |
---|
1306 | pfm_buffer_fmt_t * fmt; |
---|
1307 | spin_lock(&pfm_buffer_fmt_lock); |
---|
1308 | fmt = __pfm_find_buffer_fmt(uuid); |
---|
1309 | spin_unlock(&pfm_buffer_fmt_lock); |
---|
1310 | return fmt; |
---|
1311 | } |
---|
1312 | |
---|
1313 | int |
---|
1314 | pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt) |
---|
1315 | { |
---|
1316 | int ret = 0; |
---|
1317 | |
---|
1318 | /* some sanity checks */ |
---|
1319 | if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL; |
---|
1320 | |
---|
1321 | /* we need at least a handler */ |
---|
1322 | if (fmt->fmt_handler == NULL) return -EINVAL; |
---|
1323 | |
---|
1324 | /* |
---|
1325 | * XXX: need check validity of fmt_arg_size |
---|
1326 | */ |
---|
1327 | |
---|
1328 | spin_lock(&pfm_buffer_fmt_lock); |
---|
1329 | |
---|
1330 | if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) { |
---|
1331 | printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name); |
---|
1332 | ret = -EBUSY; |
---|
1333 | goto out; |
---|
1334 | } |
---|
1335 | list_add(&fmt->fmt_list, &pfm_buffer_fmt_list); |
---|
1336 | printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name); |
---|
1337 | |
---|
1338 | out: |
---|
1339 | spin_unlock(&pfm_buffer_fmt_lock); |
---|
1340 | return ret; |
---|
1341 | } |
---|
1342 | EXPORT_SYMBOL(pfm_register_buffer_fmt); |
---|
1343 | |
---|
1344 | int |
---|
1345 | pfm_unregister_buffer_fmt(pfm_uuid_t uuid) |
---|
1346 | { |
---|
1347 | pfm_buffer_fmt_t *fmt; |
---|
1348 | int ret = 0; |
---|
1349 | |
---|
1350 | spin_lock(&pfm_buffer_fmt_lock); |
---|
1351 | |
---|
1352 | fmt = __pfm_find_buffer_fmt(uuid); |
---|
1353 | if (!fmt) { |
---|
1354 | printk(KERN_ERR "perfmon: cannot unregister format, not found\n"); |
---|
1355 | ret = -EINVAL; |
---|
1356 | goto out; |
---|
1357 | } |
---|
1358 | list_del_init(&fmt->fmt_list); |
---|
1359 | printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name); |
---|
1360 | |
---|
1361 | out: |
---|
1362 | spin_unlock(&pfm_buffer_fmt_lock); |
---|
1363 | return ret; |
---|
1364 | |
---|
1365 | } |
---|
1366 | EXPORT_SYMBOL(pfm_unregister_buffer_fmt); |
---|
1367 | |
---|
1368 | extern void update_pal_halt_status(int); |
---|
1369 | |
---|
1370 | static int |
---|
1371 | pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu) |
---|
1372 | { |
---|
1373 | unsigned long flags; |
---|
1374 | /* |
---|
1375 | * validy checks on cpu_mask have been done upstream |
---|
1376 | */ |
---|
1377 | LOCK_PFS(flags); |
---|
1378 | |
---|
1379 | DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n", |
---|
1380 | pfm_sessions.pfs_sys_sessions, |
---|
1381 | pfm_sessions.pfs_task_sessions, |
---|
1382 | pfm_sessions.pfs_sys_use_dbregs, |
---|
1383 | is_syswide, |
---|
1384 | cpu)); |
---|
1385 | |
---|
1386 | if (is_syswide) { |
---|
1387 | /* |
---|
1388 | * cannot mix system wide and per-task sessions |
---|
1389 | */ |
---|
1390 | if (pfm_sessions.pfs_task_sessions > 0UL) { |
---|
1391 | DPRINT(("system wide not possible, %u conflicting task_sessions\n", |
---|
1392 | pfm_sessions.pfs_task_sessions)); |
---|
1393 | goto abort; |
---|
1394 | } |
---|
1395 | |
---|
1396 | if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict; |
---|
1397 | |
---|
1398 | DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id())); |
---|
1399 | |
---|
1400 | #ifndef XEN |
---|
1401 | pfm_sessions.pfs_sys_session[cpu] = task; |
---|
1402 | #else |
---|
1403 | pfm_sessions.pfs_sys_session[cpu] = XENOPROF_TASK; |
---|
1404 | #endif |
---|
1405 | |
---|
1406 | pfm_sessions.pfs_sys_sessions++ ; |
---|
1407 | |
---|
1408 | } else { |
---|
1409 | if (pfm_sessions.pfs_sys_sessions) goto abort; |
---|
1410 | pfm_sessions.pfs_task_sessions++; |
---|
1411 | } |
---|
1412 | |
---|
1413 | DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n", |
---|
1414 | pfm_sessions.pfs_sys_sessions, |
---|
1415 | pfm_sessions.pfs_task_sessions, |
---|
1416 | pfm_sessions.pfs_sys_use_dbregs, |
---|
1417 | is_syswide, |
---|
1418 | cpu)); |
---|
1419 | |
---|
1420 | /* |
---|
1421 | * disable default_idle() to go to PAL_HALT |
---|
1422 | */ |
---|
1423 | update_pal_halt_status(0); |
---|
1424 | |
---|
1425 | UNLOCK_PFS(flags); |
---|
1426 | |
---|
1427 | return 0; |
---|
1428 | |
---|
1429 | error_conflict: |
---|
1430 | DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n", |
---|
1431 | #ifndef XEN |
---|
1432 | pfm_sessions.pfs_sys_session[cpu]->pid, |
---|
1433 | #else |
---|
1434 | -1, |
---|
1435 | #endif |
---|
1436 | cpu)); |
---|
1437 | abort: |
---|
1438 | UNLOCK_PFS(flags); |
---|
1439 | |
---|
1440 | return -EBUSY; |
---|
1441 | |
---|
1442 | } |
---|
1443 | |
---|
1444 | static int |
---|
1445 | pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu) |
---|
1446 | { |
---|
1447 | unsigned long flags; |
---|
1448 | /* |
---|
1449 | * validy checks on cpu_mask have been done upstream |
---|
1450 | */ |
---|
1451 | LOCK_PFS(flags); |
---|
1452 | |
---|
1453 | DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n", |
---|
1454 | pfm_sessions.pfs_sys_sessions, |
---|
1455 | pfm_sessions.pfs_task_sessions, |
---|
1456 | pfm_sessions.pfs_sys_use_dbregs, |
---|
1457 | is_syswide, |
---|
1458 | cpu)); |
---|
1459 | |
---|
1460 | |
---|
1461 | if (is_syswide) { |
---|
1462 | pfm_sessions.pfs_sys_session[cpu] = NULL; |
---|
1463 | /* |
---|
1464 | * would not work with perfmon+more than one bit in cpu_mask |
---|
1465 | */ |
---|
1466 | if (ctx && ctx->ctx_fl_using_dbreg) { |
---|
1467 | if (pfm_sessions.pfs_sys_use_dbregs == 0) { |
---|
1468 | printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx); |
---|
1469 | } else { |
---|
1470 | pfm_sessions.pfs_sys_use_dbregs--; |
---|
1471 | } |
---|
1472 | } |
---|
1473 | pfm_sessions.pfs_sys_sessions--; |
---|
1474 | } else { |
---|
1475 | pfm_sessions.pfs_task_sessions--; |
---|
1476 | } |
---|
1477 | DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n", |
---|
1478 | pfm_sessions.pfs_sys_sessions, |
---|
1479 | pfm_sessions.pfs_task_sessions, |
---|
1480 | pfm_sessions.pfs_sys_use_dbregs, |
---|
1481 | is_syswide, |
---|
1482 | cpu)); |
---|
1483 | |
---|
1484 | /* |
---|
1485 | * if possible, enable default_idle() to go into PAL_HALT |
---|
1486 | */ |
---|
1487 | if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0) |
---|
1488 | update_pal_halt_status(1); |
---|
1489 | |
---|
1490 | UNLOCK_PFS(flags); |
---|
1491 | |
---|
1492 | return 0; |
---|
1493 | } |
---|
1494 | |
---|
1495 | #ifndef XEN |
---|
1496 | /* |
---|
1497 | * removes virtual mapping of the sampling buffer. |
---|
1498 | * IMPORTANT: cannot be called with interrupts disable, e.g. inside |
---|
1499 | * a PROTECT_CTX() section. |
---|
1500 | */ |
---|
1501 | static int |
---|
1502 | pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size) |
---|
1503 | { |
---|
1504 | int r; |
---|
1505 | |
---|
1506 | /* sanity checks */ |
---|
1507 | if (task->mm == NULL || size == 0UL || vaddr == NULL) { |
---|
1508 | printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm); |
---|
1509 | return -EINVAL; |
---|
1510 | } |
---|
1511 | |
---|
1512 | DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size)); |
---|
1513 | |
---|
1514 | /* |
---|
1515 | * does the actual unmapping |
---|
1516 | */ |
---|
1517 | down_write(&task->mm->mmap_sem); |
---|
1518 | |
---|
1519 | DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size)); |
---|
1520 | |
---|
1521 | r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0); |
---|
1522 | |
---|
1523 | up_write(&task->mm->mmap_sem); |
---|
1524 | if (r !=0) { |
---|
1525 | printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size); |
---|
1526 | } |
---|
1527 | |
---|
1528 | DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r)); |
---|
1529 | |
---|
1530 | return 0; |
---|
1531 | } |
---|
1532 | #endif |
---|
1533 | |
---|
1534 | /* |
---|
1535 | * free actual physical storage used by sampling buffer |
---|
1536 | */ |
---|
1537 | #if 0 |
---|
1538 | static int |
---|
1539 | pfm_free_smpl_buffer(pfm_context_t *ctx) |
---|
1540 | { |
---|
1541 | pfm_buffer_fmt_t *fmt; |
---|
1542 | |
---|
1543 | if (ctx->ctx_smpl_hdr == NULL) goto invalid_free; |
---|
1544 | |
---|
1545 | /* |
---|
1546 | * we won't use the buffer format anymore |
---|
1547 | */ |
---|
1548 | fmt = ctx->ctx_buf_fmt; |
---|
1549 | |
---|
1550 | DPRINT(("sampling buffer @%p size %lu vaddr=%p\n", |
---|
1551 | ctx->ctx_smpl_hdr, |
---|
1552 | ctx->ctx_smpl_size, |
---|
1553 | ctx->ctx_smpl_vaddr)); |
---|
1554 | |
---|
1555 | pfm_buf_fmt_exit(fmt, current, NULL, NULL); |
---|
1556 | |
---|
1557 | /* |
---|
1558 | * free the buffer |
---|
1559 | */ |
---|
1560 | pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size); |
---|
1561 | |
---|
1562 | ctx->ctx_smpl_hdr = NULL; |
---|
1563 | ctx->ctx_smpl_size = 0UL; |
---|
1564 | |
---|
1565 | return 0; |
---|
1566 | |
---|
1567 | invalid_free: |
---|
1568 | printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid); |
---|
1569 | return -EINVAL; |
---|
1570 | } |
---|
1571 | #endif |
---|
1572 | |
---|
1573 | static inline void |
---|
1574 | pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt) |
---|
1575 | { |
---|
1576 | if (fmt == NULL) return; |
---|
1577 | |
---|
1578 | pfm_buf_fmt_exit(fmt, current, NULL, NULL); |
---|
1579 | |
---|
1580 | } |
---|
1581 | |
---|
1582 | #ifndef XEN |
---|
1583 | /* |
---|
1584 | * pfmfs should _never_ be mounted by userland - too much of security hassle, |
---|
1585 | * no real gain from having the whole whorehouse mounted. So we don't need |
---|
1586 | * any operations on the root directory. However, we need a non-trivial |
---|
1587 | * d_name - pfm: will go nicely and kill the special-casing in procfs. |
---|
1588 | */ |
---|
1589 | static struct vfsmount *pfmfs_mnt; |
---|
1590 | |
---|
1591 | static int __init |
---|
1592 | init_pfm_fs(void) |
---|
1593 | { |
---|
1594 | int err = register_filesystem(&pfm_fs_type); |
---|
1595 | if (!err) { |
---|
1596 | pfmfs_mnt = kern_mount(&pfm_fs_type); |
---|
1597 | err = PTR_ERR(pfmfs_mnt); |
---|
1598 | if (IS_ERR(pfmfs_mnt)) |
---|
1599 | unregister_filesystem(&pfm_fs_type); |
---|
1600 | else |
---|
1601 | err = 0; |
---|
1602 | } |
---|
1603 | return err; |
---|
1604 | } |
---|
1605 | |
---|
1606 | static void __exit |
---|
1607 | exit_pfm_fs(void) |
---|
1608 | { |
---|
1609 | unregister_filesystem(&pfm_fs_type); |
---|
1610 | mntput(pfmfs_mnt); |
---|
1611 | } |
---|
1612 | |
---|
1613 | static ssize_t |
---|
1614 | pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos) |
---|
1615 | { |
---|
1616 | pfm_context_t *ctx; |
---|
1617 | pfm_msg_t *msg; |
---|
1618 | ssize_t ret; |
---|
1619 | unsigned long flags; |
---|
1620 | DECLARE_WAITQUEUE(wait, current); |
---|
1621 | if (PFM_IS_FILE(filp) == 0) { |
---|
1622 | printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid); |
---|
1623 | return -EINVAL; |
---|
1624 | } |
---|
1625 | |
---|
1626 | ctx = (pfm_context_t *)filp->private_data; |
---|
1627 | if (ctx == NULL) { |
---|
1628 | printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid); |
---|
1629 | return -EINVAL; |
---|
1630 | } |
---|
1631 | |
---|
1632 | /* |
---|
1633 | * check even when there is no message |
---|
1634 | */ |
---|
1635 | if (size < sizeof(pfm_msg_t)) { |
---|
1636 | DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t))); |
---|
1637 | return -EINVAL; |
---|
1638 | } |
---|
1639 | |
---|
1640 | PROTECT_CTX(ctx, flags); |
---|
1641 | |
---|
1642 | /* |
---|
1643 | * put ourselves on the wait queue |
---|
1644 | */ |
---|
1645 | add_wait_queue(&ctx->ctx_msgq_wait, &wait); |
---|
1646 | |
---|
1647 | |
---|
1648 | for(;;) { |
---|
1649 | /* |
---|
1650 | * check wait queue |
---|
1651 | */ |
---|
1652 | |
---|
1653 | set_current_state(TASK_INTERRUPTIBLE); |
---|
1654 | |
---|
1655 | DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail)); |
---|
1656 | |
---|
1657 | ret = 0; |
---|
1658 | if(PFM_CTXQ_EMPTY(ctx) == 0) break; |
---|
1659 | |
---|
1660 | UNPROTECT_CTX(ctx, flags); |
---|
1661 | |
---|
1662 | /* |
---|
1663 | * check non-blocking read |
---|
1664 | */ |
---|
1665 | ret = -EAGAIN; |
---|
1666 | if(filp->f_flags & O_NONBLOCK) break; |
---|
1667 | |
---|
1668 | /* |
---|
1669 | * check pending signals |
---|
1670 | */ |
---|
1671 | if(signal_pending(current)) { |
---|
1672 | ret = -EINTR; |
---|
1673 | break; |
---|
1674 | } |
---|
1675 | /* |
---|
1676 | * no message, so wait |
---|
1677 | */ |
---|
1678 | schedule(); |
---|
1679 | |
---|
1680 | PROTECT_CTX(ctx, flags); |
---|
1681 | } |
---|
1682 | DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret)); |
---|
1683 | set_current_state(TASK_RUNNING); |
---|
1684 | remove_wait_queue(&ctx->ctx_msgq_wait, &wait); |
---|
1685 | |
---|
1686 | if (ret < 0) goto abort; |
---|
1687 | |
---|
1688 | ret = -EINVAL; |
---|
1689 | msg = pfm_get_next_msg(ctx); |
---|
1690 | if (msg == NULL) { |
---|
1691 | printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid); |
---|
1692 | goto abort_locked; |
---|
1693 | } |
---|
1694 | |
---|
1695 | DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type)); |
---|
1696 | |
---|
1697 | ret = -EFAULT; |
---|
1698 | if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t); |
---|
1699 | |
---|
1700 | abort_locked: |
---|
1701 | UNPROTECT_CTX(ctx, flags); |
---|
1702 | abort: |
---|
1703 | return ret; |
---|
1704 | } |
---|
1705 | |
---|
1706 | static ssize_t |
---|
1707 | pfm_write(struct file *file, const char __user *ubuf, |
---|
1708 | size_t size, loff_t *ppos) |
---|
1709 | { |
---|
1710 | DPRINT(("pfm_write called\n")); |
---|
1711 | return -EINVAL; |
---|
1712 | } |
---|
1713 | |
---|
1714 | static unsigned int |
---|
1715 | pfm_poll(struct file *filp, poll_table * wait) |
---|
1716 | { |
---|
1717 | pfm_context_t *ctx; |
---|
1718 | unsigned long flags; |
---|
1719 | unsigned int mask = 0; |
---|
1720 | |
---|
1721 | if (PFM_IS_FILE(filp) == 0) { |
---|
1722 | printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid); |
---|
1723 | return 0; |
---|
1724 | } |
---|
1725 | |
---|
1726 | ctx = (pfm_context_t *)filp->private_data; |
---|
1727 | if (ctx == NULL) { |
---|
1728 | printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid); |
---|
1729 | return 0; |
---|
1730 | } |
---|
1731 | |
---|
1732 | |
---|
1733 | DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd)); |
---|
1734 | |
---|
1735 | poll_wait(filp, &ctx->ctx_msgq_wait, wait); |
---|
1736 | |
---|
1737 | PROTECT_CTX(ctx, flags); |
---|
1738 | |
---|
1739 | if (PFM_CTXQ_EMPTY(ctx) == 0) |
---|
1740 | mask = POLLIN | POLLRDNORM; |
---|
1741 | |
---|
1742 | UNPROTECT_CTX(ctx, flags); |
---|
1743 | |
---|
1744 | DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask)); |
---|
1745 | |
---|
1746 | return mask; |
---|
1747 | } |
---|
1748 | |
---|
1749 | static int |
---|
1750 | pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg) |
---|
1751 | { |
---|
1752 | DPRINT(("pfm_ioctl called\n")); |
---|
1753 | return -EINVAL; |
---|
1754 | } |
---|
1755 | |
---|
1756 | /* |
---|
1757 | * interrupt cannot be masked when coming here |
---|
1758 | */ |
---|
1759 | static inline int |
---|
1760 | pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on) |
---|
1761 | { |
---|
1762 | int ret; |
---|
1763 | |
---|
1764 | ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue); |
---|
1765 | |
---|
1766 | DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n", |
---|
1767 | current->pid, |
---|
1768 | fd, |
---|
1769 | on, |
---|
1770 | ctx->ctx_async_queue, ret)); |
---|
1771 | |
---|
1772 | return ret; |
---|
1773 | } |
---|
1774 | |
---|
1775 | static int |
---|
1776 | pfm_fasync(int fd, struct file *filp, int on) |
---|
1777 | { |
---|
1778 | pfm_context_t *ctx; |
---|
1779 | int ret; |
---|
1780 | |
---|
1781 | if (PFM_IS_FILE(filp) == 0) { |
---|
1782 | printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid); |
---|
1783 | return -EBADF; |
---|
1784 | } |
---|
1785 | |
---|
1786 | ctx = (pfm_context_t *)filp->private_data; |
---|
1787 | if (ctx == NULL) { |
---|
1788 | printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid); |
---|
1789 | return -EBADF; |
---|
1790 | } |
---|
1791 | /* |
---|
1792 | * we cannot mask interrupts during this call because this may |
---|
1793 | * may go to sleep if memory is not readily avalaible. |
---|
1794 | * |
---|
1795 | * We are protected from the conetxt disappearing by the get_fd()/put_fd() |
---|
1796 | * done in caller. Serialization of this function is ensured by caller. |
---|
1797 | */ |
---|
1798 | ret = pfm_do_fasync(fd, filp, ctx, on); |
---|
1799 | |
---|
1800 | |
---|
1801 | DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n", |
---|
1802 | fd, |
---|
1803 | on, |
---|
1804 | ctx->ctx_async_queue, ret)); |
---|
1805 | |
---|
1806 | return ret; |
---|
1807 | } |
---|
1808 | |
---|
1809 | #ifdef CONFIG_SMP |
---|
1810 | /* |
---|
1811 | * this function is exclusively called from pfm_close(). |
---|
1812 | * The context is not protected at that time, nor are interrupts |
---|
1813 | * on the remote CPU. That's necessary to avoid deadlocks. |
---|
1814 | */ |
---|
1815 | static void |
---|
1816 | pfm_syswide_force_stop(void *info) |
---|
1817 | { |
---|
1818 | pfm_context_t *ctx = (pfm_context_t *)info; |
---|
1819 | struct pt_regs *regs = task_pt_regs(current); |
---|
1820 | struct task_struct *owner; |
---|
1821 | unsigned long flags; |
---|
1822 | int ret; |
---|
1823 | |
---|
1824 | if (ctx->ctx_cpu != smp_processor_id()) { |
---|
1825 | printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d but on CPU%d\n", |
---|
1826 | ctx->ctx_cpu, |
---|
1827 | smp_processor_id()); |
---|
1828 | return; |
---|
1829 | } |
---|
1830 | owner = GET_PMU_OWNER(); |
---|
1831 | if (owner != ctx->ctx_task) { |
---|
1832 | printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n", |
---|
1833 | smp_processor_id(), |
---|
1834 | owner->pid, ctx->ctx_task->pid); |
---|
1835 | return; |
---|
1836 | } |
---|
1837 | if (GET_PMU_CTX() != ctx) { |
---|
1838 | printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n", |
---|
1839 | smp_processor_id(), |
---|
1840 | GET_PMU_CTX(), ctx); |
---|
1841 | return; |
---|
1842 | } |
---|
1843 | |
---|
1844 | DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid)); |
---|
1845 | /* |
---|
1846 | * the context is already protected in pfm_close(), we simply |
---|
1847 | * need to mask interrupts to avoid a PMU interrupt race on |
---|
1848 | * this CPU |
---|
1849 | */ |
---|
1850 | local_irq_save(flags); |
---|
1851 | |
---|
1852 | ret = pfm_context_unload(ctx, NULL, 0, regs); |
---|
1853 | if (ret) { |
---|
1854 | DPRINT(("context_unload returned %d\n", ret)); |
---|
1855 | } |
---|
1856 | |
---|
1857 | /* |
---|
1858 | * unmask interrupts, PMU interrupts are now spurious here |
---|
1859 | */ |
---|
1860 | local_irq_restore(flags); |
---|
1861 | } |
---|
1862 | |
---|
1863 | static void |
---|
1864 | pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx) |
---|
1865 | { |
---|
1866 | int ret; |
---|
1867 | |
---|
1868 | DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu)); |
---|
1869 | ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1); |
---|
1870 | DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret)); |
---|
1871 | } |
---|
1872 | #endif /* CONFIG_SMP */ |
---|
1873 | |
---|
1874 | /* |
---|
1875 | * called for each close(). Partially free resources. |
---|
1876 | * When caller is self-monitoring, the context is unloaded. |
---|
1877 | */ |
---|
1878 | static int |
---|
1879 | pfm_flush(struct file *filp) |
---|
1880 | { |
---|
1881 | pfm_context_t *ctx; |
---|
1882 | struct task_struct *task; |
---|
1883 | struct pt_regs *regs; |
---|
1884 | unsigned long flags; |
---|
1885 | unsigned long smpl_buf_size = 0UL; |
---|
1886 | void *smpl_buf_vaddr = NULL; |
---|
1887 | int state, is_system; |
---|
1888 | |
---|
1889 | if (PFM_IS_FILE(filp) == 0) { |
---|
1890 | DPRINT(("bad magic for\n")); |
---|
1891 | return -EBADF; |
---|
1892 | } |
---|
1893 | |
---|
1894 | ctx = (pfm_context_t *)filp->private_data; |
---|
1895 | if (ctx == NULL) { |
---|
1896 | printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid); |
---|
1897 | return -EBADF; |
---|
1898 | } |
---|
1899 | |
---|
1900 | /* |
---|
1901 | * remove our file from the async queue, if we use this mode. |
---|
1902 | * This can be done without the context being protected. We come |
---|
1903 | * here when the context has become unreacheable by other tasks. |
---|
1904 | * |
---|
1905 | * We may still have active monitoring at this point and we may |
---|
1906 | * end up in pfm_overflow_handler(). However, fasync_helper() |
---|
1907 | * operates with interrupts disabled and it cleans up the |
---|
1908 | * queue. If the PMU handler is called prior to entering |
---|
1909 | * fasync_helper() then it will send a signal. If it is |
---|
1910 | * invoked after, it will find an empty queue and no |
---|
1911 | * signal will be sent. In both case, we are safe |
---|
1912 | */ |
---|
1913 | if (filp->f_flags & FASYNC) { |
---|
1914 | DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue)); |
---|
1915 | pfm_do_fasync (-1, filp, ctx, 0); |
---|
1916 | } |
---|
1917 | |
---|
1918 | PROTECT_CTX(ctx, flags); |
---|
1919 | |
---|
1920 | state = ctx->ctx_state; |
---|
1921 | is_system = ctx->ctx_fl_system; |
---|
1922 | |
---|
1923 | task = PFM_CTX_TASK(ctx); |
---|
1924 | regs = task_pt_regs(task); |
---|
1925 | |
---|
1926 | DPRINT(("ctx_state=%d is_current=%d\n", |
---|
1927 | state, |
---|
1928 | task == current ? 1 : 0)); |
---|
1929 | |
---|
1930 | /* |
---|
1931 | * if state == UNLOADED, then task is NULL |
---|
1932 | */ |
---|
1933 | |
---|
1934 | /* |
---|
1935 | * we must stop and unload because we are losing access to the context. |
---|
1936 | */ |
---|
1937 | if (task == current) { |
---|
1938 | #ifdef CONFIG_SMP |
---|
1939 | /* |
---|
1940 | * the task IS the owner but it migrated to another CPU: that's bad |
---|
1941 | * but we must handle this cleanly. Unfortunately, the kernel does |
---|
1942 | * not provide a mechanism to block migration (while the context is loaded). |
---|
1943 | * |
---|
1944 | * We need to release the resource on the ORIGINAL cpu. |
---|
1945 | */ |
---|
1946 | if (is_system && ctx->ctx_cpu != smp_processor_id()) { |
---|
1947 | |
---|
1948 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
1949 | /* |
---|
1950 | * keep context protected but unmask interrupt for IPI |
---|
1951 | */ |
---|
1952 | local_irq_restore(flags); |
---|
1953 | |
---|
1954 | pfm_syswide_cleanup_other_cpu(ctx); |
---|
1955 | |
---|
1956 | /* |
---|
1957 | * restore interrupt masking |
---|
1958 | */ |
---|
1959 | local_irq_save(flags); |
---|
1960 | |
---|
1961 | /* |
---|
1962 | * context is unloaded at this point |
---|
1963 | */ |
---|
1964 | } else |
---|
1965 | #endif /* CONFIG_SMP */ |
---|
1966 | { |
---|
1967 | |
---|
1968 | DPRINT(("forcing unload\n")); |
---|
1969 | /* |
---|
1970 | * stop and unload, returning with state UNLOADED |
---|
1971 | * and session unreserved. |
---|
1972 | */ |
---|
1973 | pfm_context_unload(ctx, NULL, 0, regs); |
---|
1974 | |
---|
1975 | DPRINT(("ctx_state=%d\n", ctx->ctx_state)); |
---|
1976 | } |
---|
1977 | } |
---|
1978 | |
---|
1979 | /* |
---|
1980 | * remove virtual mapping, if any, for the calling task. |
---|
1981 | * cannot reset ctx field until last user is calling close(). |
---|
1982 | * |
---|
1983 | * ctx_smpl_vaddr must never be cleared because it is needed |
---|
1984 | * by every task with access to the context |
---|
1985 | * |
---|
1986 | * When called from do_exit(), the mm context is gone already, therefore |
---|
1987 | * mm is NULL, i.e., the VMA is already gone and we do not have to |
---|
1988 | * do anything here |
---|
1989 | */ |
---|
1990 | if (ctx->ctx_smpl_vaddr && current->mm) { |
---|
1991 | smpl_buf_vaddr = ctx->ctx_smpl_vaddr; |
---|
1992 | smpl_buf_size = ctx->ctx_smpl_size; |
---|
1993 | } |
---|
1994 | |
---|
1995 | UNPROTECT_CTX(ctx, flags); |
---|
1996 | |
---|
1997 | /* |
---|
1998 | * if there was a mapping, then we systematically remove it |
---|
1999 | * at this point. Cannot be done inside critical section |
---|
2000 | * because some VM function reenables interrupts. |
---|
2001 | * |
---|
2002 | */ |
---|
2003 | if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size); |
---|
2004 | |
---|
2005 | return 0; |
---|
2006 | } |
---|
2007 | #endif |
---|
2008 | /* |
---|
2009 | * called either on explicit close() or from exit_files(). |
---|
2010 | * Only the LAST user of the file gets to this point, i.e., it is |
---|
2011 | * called only ONCE. |
---|
2012 | * |
---|
2013 | * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero |
---|
2014 | * (fput()),i.e, last task to access the file. Nobody else can access the |
---|
2015 | * file at this point. |
---|
2016 | * |
---|
2017 | * When called from exit_files(), the VMA has been freed because exit_mm() |
---|
2018 | * is executed before exit_files(). |
---|
2019 | * |
---|
2020 | * When called from exit_files(), the current task is not yet ZOMBIE but we |
---|
2021 | * flush the PMU state to the context. |
---|
2022 | */ |
---|
2023 | #ifndef XEN |
---|
2024 | static int |
---|
2025 | pfm_close(struct inode *inode, struct file *filp) |
---|
2026 | #else |
---|
2027 | static int |
---|
2028 | pfm_close(pfm_context_t *ctx) |
---|
2029 | #endif |
---|
2030 | { |
---|
2031 | #ifndef XEN |
---|
2032 | pfm_context_t *ctx; |
---|
2033 | struct task_struct *task; |
---|
2034 | struct pt_regs *regs; |
---|
2035 | DECLARE_WAITQUEUE(wait, current); |
---|
2036 | unsigned long flags; |
---|
2037 | #endif |
---|
2038 | unsigned long smpl_buf_size = 0UL; |
---|
2039 | void *smpl_buf_addr = NULL; |
---|
2040 | int free_possible = 1; |
---|
2041 | int state, is_system; |
---|
2042 | |
---|
2043 | #ifndef XEN |
---|
2044 | DPRINT(("pfm_close called private=%p\n", filp->private_data)); |
---|
2045 | |
---|
2046 | if (PFM_IS_FILE(filp) == 0) { |
---|
2047 | DPRINT(("bad magic\n")); |
---|
2048 | return -EBADF; |
---|
2049 | } |
---|
2050 | |
---|
2051 | ctx = (pfm_context_t *)filp->private_data; |
---|
2052 | if (ctx == NULL) { |
---|
2053 | printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid); |
---|
2054 | return -EBADF; |
---|
2055 | } |
---|
2056 | |
---|
2057 | PROTECT_CTX(ctx, flags); |
---|
2058 | #else |
---|
2059 | BUG_ON(!spin_is_locked(&ctx->ctx_lock)); |
---|
2060 | #endif |
---|
2061 | |
---|
2062 | state = ctx->ctx_state; |
---|
2063 | is_system = ctx->ctx_fl_system; |
---|
2064 | |
---|
2065 | #ifndef XEN |
---|
2066 | task = PFM_CTX_TASK(ctx); |
---|
2067 | regs = task_pt_regs(task); |
---|
2068 | |
---|
2069 | DPRINT(("ctx_state=%d is_current=%d\n", |
---|
2070 | state, |
---|
2071 | task == current ? 1 : 0)); |
---|
2072 | |
---|
2073 | /* |
---|
2074 | * if task == current, then pfm_flush() unloaded the context |
---|
2075 | */ |
---|
2076 | if (state == PFM_CTX_UNLOADED) goto doit; |
---|
2077 | |
---|
2078 | /* |
---|
2079 | * context is loaded/masked and task != current, we need to |
---|
2080 | * either force an unload or go zombie |
---|
2081 | */ |
---|
2082 | |
---|
2083 | /* |
---|
2084 | * The task is currently blocked or will block after an overflow. |
---|
2085 | * we must force it to wakeup to get out of the |
---|
2086 | * MASKED state and transition to the unloaded state by itself. |
---|
2087 | * |
---|
2088 | * This situation is only possible for per-task mode |
---|
2089 | */ |
---|
2090 | if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) { |
---|
2091 | |
---|
2092 | /* |
---|
2093 | * set a "partial" zombie state to be checked |
---|
2094 | * upon return from down() in pfm_handle_work(). |
---|
2095 | * |
---|
2096 | * We cannot use the ZOMBIE state, because it is checked |
---|
2097 | * by pfm_load_regs() which is called upon wakeup from down(). |
---|
2098 | * In such case, it would free the context and then we would |
---|
2099 | * return to pfm_handle_work() which would access the |
---|
2100 | * stale context. Instead, we set a flag invisible to pfm_load_regs() |
---|
2101 | * but visible to pfm_handle_work(). |
---|
2102 | * |
---|
2103 | * For some window of time, we have a zombie context with |
---|
2104 | * ctx_state = MASKED and not ZOMBIE |
---|
2105 | */ |
---|
2106 | ctx->ctx_fl_going_zombie = 1; |
---|
2107 | |
---|
2108 | /* |
---|
2109 | * force task to wake up from MASKED state |
---|
2110 | */ |
---|
2111 | complete(&ctx->ctx_restart_done); |
---|
2112 | |
---|
2113 | DPRINT(("waking up ctx_state=%d\n", state)); |
---|
2114 | |
---|
2115 | /* |
---|
2116 | * put ourself to sleep waiting for the other |
---|
2117 | * task to report completion |
---|
2118 | * |
---|
2119 | * the context is protected by mutex, therefore there |
---|
2120 | * is no risk of being notified of completion before |
---|
2121 | * begin actually on the waitq. |
---|
2122 | */ |
---|
2123 | set_current_state(TASK_INTERRUPTIBLE); |
---|
2124 | add_wait_queue(&ctx->ctx_zombieq, &wait); |
---|
2125 | |
---|
2126 | UNPROTECT_CTX(ctx, flags); |
---|
2127 | |
---|
2128 | /* |
---|
2129 | * XXX: check for signals : |
---|
2130 | * - ok for explicit close |
---|
2131 | * - not ok when coming from exit_files() |
---|
2132 | */ |
---|
2133 | schedule(); |
---|
2134 | |
---|
2135 | |
---|
2136 | PROTECT_CTX(ctx, flags); |
---|
2137 | |
---|
2138 | |
---|
2139 | remove_wait_queue(&ctx->ctx_zombieq, &wait); |
---|
2140 | set_current_state(TASK_RUNNING); |
---|
2141 | |
---|
2142 | /* |
---|
2143 | * context is unloaded at this point |
---|
2144 | */ |
---|
2145 | DPRINT(("after zombie wakeup ctx_state=%d for\n", state)); |
---|
2146 | } |
---|
2147 | else if (task != current) { |
---|
2148 | #ifdef CONFIG_SMP |
---|
2149 | /* |
---|
2150 | * switch context to zombie state |
---|
2151 | */ |
---|
2152 | ctx->ctx_state = PFM_CTX_ZOMBIE; |
---|
2153 | |
---|
2154 | DPRINT(("zombie ctx for [%d]\n", task->pid)); |
---|
2155 | /* |
---|
2156 | * cannot free the context on the spot. deferred until |
---|
2157 | * the task notices the ZOMBIE state |
---|
2158 | */ |
---|
2159 | free_possible = 0; |
---|
2160 | #else |
---|
2161 | pfm_context_unload(ctx, NULL, 0, regs); |
---|
2162 | #endif |
---|
2163 | } |
---|
2164 | #else |
---|
2165 | /* XXX XEN */ |
---|
2166 | /* unload context */ |
---|
2167 | BUG_ON(state != PFM_CTX_UNLOADED); |
---|
2168 | #endif |
---|
2169 | |
---|
2170 | #ifndef XEN |
---|
2171 | doit: |
---|
2172 | #endif |
---|
2173 | /* reload state, may have changed during opening of critical section */ |
---|
2174 | state = ctx->ctx_state; |
---|
2175 | |
---|
2176 | /* |
---|
2177 | * the context is still attached to a task (possibly current) |
---|
2178 | * we cannot destroy it right now |
---|
2179 | */ |
---|
2180 | |
---|
2181 | /* |
---|
2182 | * we must free the sampling buffer right here because |
---|
2183 | * we cannot rely on it being cleaned up later by the |
---|
2184 | * monitored task. It is not possible to free vmalloc'ed |
---|
2185 | * memory in pfm_load_regs(). Instead, we remove the buffer |
---|
2186 | * now. should there be subsequent PMU overflow originally |
---|
2187 | * meant for sampling, the will be converted to spurious |
---|
2188 | * and that's fine because the monitoring tools is gone anyway. |
---|
2189 | */ |
---|
2190 | if (ctx->ctx_smpl_hdr) { |
---|
2191 | smpl_buf_addr = ctx->ctx_smpl_hdr; |
---|
2192 | smpl_buf_size = ctx->ctx_smpl_size; |
---|
2193 | /* no more sampling */ |
---|
2194 | ctx->ctx_smpl_hdr = NULL; |
---|
2195 | ctx->ctx_fl_is_sampling = 0; |
---|
2196 | } |
---|
2197 | |
---|
2198 | DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n", |
---|
2199 | state, |
---|
2200 | free_possible, |
---|
2201 | smpl_buf_addr, |
---|
2202 | smpl_buf_size)); |
---|
2203 | |
---|
2204 | if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt); |
---|
2205 | |
---|
2206 | /* |
---|
2207 | * UNLOADED that the session has already been unreserved. |
---|
2208 | */ |
---|
2209 | if (state == PFM_CTX_ZOMBIE) { |
---|
2210 | pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu); |
---|
2211 | } |
---|
2212 | |
---|
2213 | #ifndef XEN |
---|
2214 | /* |
---|
2215 | * disconnect file descriptor from context must be done |
---|
2216 | * before we unlock. |
---|
2217 | */ |
---|
2218 | filp->private_data = NULL; |
---|
2219 | |
---|
2220 | /* |
---|
2221 | * if we free on the spot, the context is now completely unreacheable |
---|
2222 | * from the callers side. The monitored task side is also cut, so we |
---|
2223 | * can freely cut. |
---|
2224 | * |
---|
2225 | * If we have a deferred free, only the caller side is disconnected. |
---|
2226 | */ |
---|
2227 | UNPROTECT_CTX(ctx, flags); |
---|
2228 | |
---|
2229 | /* |
---|
2230 | * All memory free operations (especially for vmalloc'ed memory) |
---|
2231 | * MUST be done with interrupts ENABLED. |
---|
2232 | */ |
---|
2233 | if (smpl_buf_addr) pfm_rvfree(smpl_buf_addr, smpl_buf_size); |
---|
2234 | #else |
---|
2235 | UNPROTECT_CTX_NOIRQ(ctx); |
---|
2236 | #endif |
---|
2237 | |
---|
2238 | /* |
---|
2239 | * return the memory used by the context |
---|
2240 | */ |
---|
2241 | if (free_possible) pfm_context_free(ctx); |
---|
2242 | |
---|
2243 | return 0; |
---|
2244 | } |
---|
2245 | |
---|
2246 | #ifndef XEN |
---|
2247 | static int |
---|
2248 | pfm_no_open(struct inode *irrelevant, struct file *dontcare) |
---|
2249 | { |
---|
2250 | DPRINT(("pfm_no_open called\n")); |
---|
2251 | return -ENXIO; |
---|
2252 | } |
---|
2253 | |
---|
2254 | |
---|
2255 | |
---|
2256 | static struct file_operations pfm_file_ops = { |
---|
2257 | .llseek = no_llseek, |
---|
2258 | .read = pfm_read, |
---|
2259 | .write = pfm_write, |
---|
2260 | .poll = pfm_poll, |
---|
2261 | .ioctl = pfm_ioctl, |
---|
2262 | .open = pfm_no_open, /* special open code to disallow open via /proc */ |
---|
2263 | .fasync = pfm_fasync, |
---|
2264 | .release = pfm_close, |
---|
2265 | .flush = pfm_flush |
---|
2266 | }; |
---|
2267 | |
---|
2268 | static int |
---|
2269 | pfmfs_delete_dentry(struct dentry *dentry) |
---|
2270 | { |
---|
2271 | return 1; |
---|
2272 | } |
---|
2273 | |
---|
2274 | static struct dentry_operations pfmfs_dentry_operations = { |
---|
2275 | .d_delete = pfmfs_delete_dentry, |
---|
2276 | }; |
---|
2277 | |
---|
2278 | |
---|
2279 | static int |
---|
2280 | pfm_alloc_fd(struct file **cfile) |
---|
2281 | { |
---|
2282 | int fd, ret = 0; |
---|
2283 | struct file *file = NULL; |
---|
2284 | struct inode * inode; |
---|
2285 | char name[32]; |
---|
2286 | struct qstr this; |
---|
2287 | |
---|
2288 | fd = get_unused_fd(); |
---|
2289 | if (fd < 0) return -ENFILE; |
---|
2290 | |
---|
2291 | ret = -ENFILE; |
---|
2292 | |
---|
2293 | file = get_empty_filp(); |
---|
2294 | if (!file) goto out; |
---|
2295 | |
---|
2296 | /* |
---|
2297 | * allocate a new inode |
---|
2298 | */ |
---|
2299 | inode = new_inode(pfmfs_mnt->mnt_sb); |
---|
2300 | if (!inode) goto out; |
---|
2301 | |
---|
2302 | DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode)); |
---|
2303 | |
---|
2304 | inode->i_mode = S_IFCHR|S_IRUGO; |
---|
2305 | inode->i_uid = current->fsuid; |
---|
2306 | inode->i_gid = current->fsgid; |
---|
2307 | |
---|
2308 | snprintf(name, sizeof(name), "[%lu]", inode->i_ino); |
---|
2309 | this.name = name; |
---|
2310 | this.len = strlen(name); |
---|
2311 | this.hash = inode->i_ino; |
---|
2312 | |
---|
2313 | ret = -ENOMEM; |
---|
2314 | |
---|
2315 | /* |
---|
2316 | * allocate a new dcache entry |
---|
2317 | */ |
---|
2318 | file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this); |
---|
2319 | if (!file->f_dentry) goto out; |
---|
2320 | |
---|
2321 | file->f_dentry->d_op = &pfmfs_dentry_operations; |
---|
2322 | |
---|
2323 | d_add(file->f_dentry, inode); |
---|
2324 | file->f_vfsmnt = mntget(pfmfs_mnt); |
---|
2325 | file->f_mapping = inode->i_mapping; |
---|
2326 | |
---|
2327 | file->f_op = &pfm_file_ops; |
---|
2328 | file->f_mode = FMODE_READ; |
---|
2329 | file->f_flags = O_RDONLY; |
---|
2330 | file->f_pos = 0; |
---|
2331 | |
---|
2332 | /* |
---|
2333 | * may have to delay until context is attached? |
---|
2334 | */ |
---|
2335 | fd_install(fd, file); |
---|
2336 | |
---|
2337 | /* |
---|
2338 | * the file structure we will use |
---|
2339 | */ |
---|
2340 | *cfile = file; |
---|
2341 | |
---|
2342 | return fd; |
---|
2343 | out: |
---|
2344 | if (file) put_filp(file); |
---|
2345 | put_unused_fd(fd); |
---|
2346 | return ret; |
---|
2347 | } |
---|
2348 | |
---|
2349 | static void |
---|
2350 | pfm_free_fd(int fd, struct file *file) |
---|
2351 | { |
---|
2352 | struct files_struct *files = current->files; |
---|
2353 | struct fdtable *fdt; |
---|
2354 | |
---|
2355 | /* |
---|
2356 | * there ie no fd_uninstall(), so we do it here |
---|
2357 | */ |
---|
2358 | spin_lock(&files->file_lock); |
---|
2359 | fdt = files_fdtable(files); |
---|
2360 | rcu_assign_pointer(fdt->fd[fd], NULL); |
---|
2361 | spin_unlock(&files->file_lock); |
---|
2362 | |
---|
2363 | if (file) |
---|
2364 | put_filp(file); |
---|
2365 | put_unused_fd(fd); |
---|
2366 | } |
---|
2367 | |
---|
2368 | static int |
---|
2369 | pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size) |
---|
2370 | { |
---|
2371 | DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size)); |
---|
2372 | |
---|
2373 | while (size > 0) { |
---|
2374 | unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT; |
---|
2375 | |
---|
2376 | |
---|
2377 | if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY)) |
---|
2378 | return -ENOMEM; |
---|
2379 | |
---|
2380 | addr += PAGE_SIZE; |
---|
2381 | buf += PAGE_SIZE; |
---|
2382 | size -= PAGE_SIZE; |
---|
2383 | } |
---|
2384 | return 0; |
---|
2385 | } |
---|
2386 | #endif |
---|
2387 | |
---|
2388 | /* |
---|
2389 | * allocate a sampling buffer and remaps it into the user address space of the task |
---|
2390 | */ |
---|
2391 | static int |
---|
2392 | pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr) |
---|
2393 | { |
---|
2394 | #ifndef XEN |
---|
2395 | struct mm_struct *mm = task->mm; |
---|
2396 | struct vm_area_struct *vma = NULL; |
---|
2397 | unsigned long size; |
---|
2398 | void *smpl_buf; |
---|
2399 | |
---|
2400 | |
---|
2401 | /* |
---|
2402 | * the fixed header + requested size and align to page boundary |
---|
2403 | */ |
---|
2404 | size = PAGE_ALIGN(rsize); |
---|
2405 | |
---|
2406 | DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size)); |
---|
2407 | |
---|
2408 | /* |
---|
2409 | * check requested size to avoid Denial-of-service attacks |
---|
2410 | * XXX: may have to refine this test |
---|
2411 | * Check against address space limit. |
---|
2412 | * |
---|
2413 | * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur) |
---|
2414 | * return -ENOMEM; |
---|
2415 | */ |
---|
2416 | if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur) |
---|
2417 | return -ENOMEM; |
---|
2418 | |
---|
2419 | /* |
---|
2420 | * We do the easy to undo allocations first. |
---|
2421 | * |
---|
2422 | * pfm_rvmalloc(), clears the buffer, so there is no leak |
---|
2423 | */ |
---|
2424 | smpl_buf = pfm_rvmalloc(size); |
---|
2425 | if (smpl_buf == NULL) { |
---|
2426 | DPRINT(("Can't allocate sampling buffer\n")); |
---|
2427 | return -ENOMEM; |
---|
2428 | } |
---|
2429 | |
---|
2430 | DPRINT(("smpl_buf @%p\n", smpl_buf)); |
---|
2431 | |
---|
2432 | /* allocate vma */ |
---|
2433 | vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL); |
---|
2434 | if (!vma) { |
---|
2435 | DPRINT(("Cannot allocate vma\n")); |
---|
2436 | goto error_kmem; |
---|
2437 | } |
---|
2438 | memset(vma, 0, sizeof(*vma)); |
---|
2439 | |
---|
2440 | /* |
---|
2441 | * partially initialize the vma for the sampling buffer |
---|
2442 | */ |
---|
2443 | vma->vm_mm = mm; |
---|
2444 | vma->vm_flags = VM_READ| VM_MAYREAD |VM_RESERVED; |
---|
2445 | vma->vm_page_prot = PAGE_READONLY; /* XXX may need to change */ |
---|
2446 | |
---|
2447 | /* |
---|
2448 | * Now we have everything we need and we can initialize |
---|
2449 | * and connect all the data structures |
---|
2450 | */ |
---|
2451 | |
---|
2452 | ctx->ctx_smpl_hdr = smpl_buf; |
---|
2453 | ctx->ctx_smpl_size = size; /* aligned size */ |
---|
2454 | |
---|
2455 | /* |
---|
2456 | * Let's do the difficult operations next. |
---|
2457 | * |
---|
2458 | * now we atomically find some area in the address space and |
---|
2459 | * remap the buffer in it. |
---|
2460 | */ |
---|
2461 | down_write(&task->mm->mmap_sem); |
---|
2462 | |
---|
2463 | /* find some free area in address space, must have mmap sem held */ |
---|
2464 | vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0); |
---|
2465 | if (vma->vm_start == 0UL) { |
---|
2466 | DPRINT(("Cannot find unmapped area for size %ld\n", size)); |
---|
2467 | up_write(&task->mm->mmap_sem); |
---|
2468 | goto error; |
---|
2469 | } |
---|
2470 | vma->vm_end = vma->vm_start + size; |
---|
2471 | vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT; |
---|
2472 | |
---|
2473 | DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start)); |
---|
2474 | |
---|
2475 | /* can only be applied to current task, need to have the mm semaphore held when called */ |
---|
2476 | if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) { |
---|
2477 | DPRINT(("Can't remap buffer\n")); |
---|
2478 | up_write(&task->mm->mmap_sem); |
---|
2479 | goto error; |
---|
2480 | } |
---|
2481 | |
---|
2482 | /* |
---|
2483 | * now insert the vma in the vm list for the process, must be |
---|
2484 | * done with mmap lock held |
---|
2485 | */ |
---|
2486 | insert_vm_struct(mm, vma); |
---|
2487 | |
---|
2488 | mm->total_vm += size >> PAGE_SHIFT; |
---|
2489 | vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file, |
---|
2490 | vma_pages(vma)); |
---|
2491 | up_write(&task->mm->mmap_sem); |
---|
2492 | |
---|
2493 | /* |
---|
2494 | * keep track of user level virtual address |
---|
2495 | */ |
---|
2496 | ctx->ctx_smpl_vaddr = (void *)vma->vm_start; |
---|
2497 | *(unsigned long *)user_vaddr = vma->vm_start; |
---|
2498 | |
---|
2499 | return 0; |
---|
2500 | |
---|
2501 | error: |
---|
2502 | kmem_cache_free(vm_area_cachep, vma); |
---|
2503 | error_kmem: |
---|
2504 | pfm_rvfree(smpl_buf, size); |
---|
2505 | |
---|
2506 | return -ENOMEM; |
---|
2507 | #else |
---|
2508 | /* XXX */ |
---|
2509 | return 0; |
---|
2510 | #endif |
---|
2511 | } |
---|
2512 | |
---|
2513 | #ifndef XEN |
---|
2514 | /* |
---|
2515 | * XXX: do something better here |
---|
2516 | */ |
---|
2517 | static int |
---|
2518 | pfm_bad_permissions(struct task_struct *task) |
---|
2519 | { |
---|
2520 | /* inspired by ptrace_attach() */ |
---|
2521 | DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n", |
---|
2522 | current->uid, |
---|
2523 | current->gid, |
---|
2524 | task->euid, |
---|
2525 | task->suid, |
---|
2526 | task->uid, |
---|
2527 | task->egid, |
---|
2528 | task->sgid)); |
---|
2529 | |
---|
2530 | return ((current->uid != task->euid) |
---|
2531 | || (current->uid != task->suid) |
---|
2532 | || (current->uid != task->uid) |
---|
2533 | || (current->gid != task->egid) |
---|
2534 | || (current->gid != task->sgid) |
---|
2535 | || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE); |
---|
2536 | } |
---|
2537 | #endif |
---|
2538 | |
---|
2539 | static int |
---|
2540 | pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx) |
---|
2541 | { |
---|
2542 | int ctx_flags; |
---|
2543 | |
---|
2544 | /* valid signal */ |
---|
2545 | |
---|
2546 | ctx_flags = pfx->ctx_flags; |
---|
2547 | |
---|
2548 | if (ctx_flags & PFM_FL_SYSTEM_WIDE) { |
---|
2549 | |
---|
2550 | /* |
---|
2551 | * cannot block in this mode |
---|
2552 | */ |
---|
2553 | if (ctx_flags & PFM_FL_NOTIFY_BLOCK) { |
---|
2554 | DPRINT(("cannot use blocking mode when in system wide monitoring\n")); |
---|
2555 | return -EINVAL; |
---|
2556 | } |
---|
2557 | } else { |
---|
2558 | } |
---|
2559 | /* probably more to add here */ |
---|
2560 | |
---|
2561 | return 0; |
---|
2562 | } |
---|
2563 | |
---|
2564 | static int |
---|
2565 | pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags, |
---|
2566 | unsigned int cpu, pfarg_context_t *arg) |
---|
2567 | { |
---|
2568 | pfm_buffer_fmt_t *fmt = NULL; |
---|
2569 | unsigned long size = 0UL; |
---|
2570 | void *uaddr = NULL; |
---|
2571 | void *fmt_arg = NULL; |
---|
2572 | int ret = 0; |
---|
2573 | #define PFM_CTXARG_BUF_ARG(a) (pfm_buffer_fmt_t *)(a+1) |
---|
2574 | |
---|
2575 | /* invoke and lock buffer format, if found */ |
---|
2576 | fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id); |
---|
2577 | if (fmt == NULL) { |
---|
2578 | DPRINT(("[%d] cannot find buffer format\n", task->pid)); |
---|
2579 | return -EINVAL; |
---|
2580 | } |
---|
2581 | |
---|
2582 | /* |
---|
2583 | * buffer argument MUST be contiguous to pfarg_context_t |
---|
2584 | */ |
---|
2585 | if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg); |
---|
2586 | |
---|
2587 | ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg); |
---|
2588 | |
---|
2589 | DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret)); |
---|
2590 | |
---|
2591 | if (ret) goto error; |
---|
2592 | |
---|
2593 | /* link buffer format and context */ |
---|
2594 | ctx->ctx_buf_fmt = fmt; |
---|
2595 | |
---|
2596 | /* |
---|
2597 | * check if buffer format wants to use perfmon buffer allocation/mapping service |
---|
2598 | */ |
---|
2599 | ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size); |
---|
2600 | if (ret) goto error; |
---|
2601 | |
---|
2602 | if (size) { |
---|
2603 | /* |
---|
2604 | * buffer is always remapped into the caller's address space |
---|
2605 | */ |
---|
2606 | ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr); |
---|
2607 | if (ret) goto error; |
---|
2608 | |
---|
2609 | /* keep track of user address of buffer */ |
---|
2610 | arg->ctx_smpl_vaddr = uaddr; |
---|
2611 | } |
---|
2612 | ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg); |
---|
2613 | |
---|
2614 | error: |
---|
2615 | return ret; |
---|
2616 | } |
---|
2617 | |
---|
2618 | static void |
---|
2619 | pfm_reset_pmu_state(pfm_context_t *ctx) |
---|
2620 | { |
---|
2621 | int i; |
---|
2622 | |
---|
2623 | /* |
---|
2624 | * install reset values for PMC. |
---|
2625 | */ |
---|
2626 | for (i=1; PMC_IS_LAST(i) == 0; i++) { |
---|
2627 | if (PMC_IS_IMPL(i) == 0) continue; |
---|
2628 | ctx->ctx_pmcs[i] = PMC_DFL_VAL(i); |
---|
2629 | DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i])); |
---|
2630 | } |
---|
2631 | /* |
---|
2632 | * PMD registers are set to 0UL when the context in memset() |
---|
2633 | */ |
---|
2634 | |
---|
2635 | /* |
---|
2636 | * On context switched restore, we must restore ALL pmc and ALL pmd even |
---|
2637 | * when they are not actively used by the task. In UP, the incoming process |
---|
2638 | * may otherwise pick up left over PMC, PMD state from the previous process. |
---|
2639 | * As opposed to PMD, stale PMC can cause harm to the incoming |
---|
2640 | * process because they may change what is being measured. |
---|
2641 | * Therefore, we must systematically reinstall the entire |
---|
2642 | * PMC state. In SMP, the same thing is possible on the |
---|
2643 | * same CPU but also on between 2 CPUs. |
---|
2644 | * |
---|
2645 | * The problem with PMD is information leaking especially |
---|
2646 | * to user level when psr.sp=0 |
---|
2647 | * |
---|
2648 | * There is unfortunately no easy way to avoid this problem |
---|
2649 | * on either UP or SMP. This definitively slows down the |
---|
2650 | * pfm_load_regs() function. |
---|
2651 | */ |
---|
2652 | |
---|
2653 | /* |
---|
2654 | * bitmask of all PMCs accessible to this context |
---|
2655 | * |
---|
2656 | * PMC0 is treated differently. |
---|
2657 | */ |
---|
2658 | ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1; |
---|
2659 | |
---|
2660 | /* |
---|
2661 | * bitmask of all PMDs that are accesible to this context |
---|
2662 | */ |
---|
2663 | ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0]; |
---|
2664 | |
---|
2665 | DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0])); |
---|
2666 | |
---|
2667 | /* |
---|
2668 | * useful in case of re-enable after disable |
---|
2669 | */ |
---|
2670 | ctx->ctx_used_ibrs[0] = 0UL; |
---|
2671 | ctx->ctx_used_dbrs[0] = 0UL; |
---|
2672 | } |
---|
2673 | |
---|
2674 | #ifndef XEN |
---|
2675 | static int |
---|
2676 | pfm_ctx_getsize(void *arg, size_t *sz) |
---|
2677 | { |
---|
2678 | pfarg_context_t *req = (pfarg_context_t *)arg; |
---|
2679 | pfm_buffer_fmt_t *fmt; |
---|
2680 | |
---|
2681 | *sz = 0; |
---|
2682 | |
---|
2683 | if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0; |
---|
2684 | |
---|
2685 | fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id); |
---|
2686 | if (fmt == NULL) { |
---|
2687 | DPRINT(("cannot find buffer format\n")); |
---|
2688 | return -EINVAL; |
---|
2689 | } |
---|
2690 | /* get just enough to copy in user parameters */ |
---|
2691 | *sz = fmt->fmt_arg_size; |
---|
2692 | DPRINT(("arg_size=%lu\n", *sz)); |
---|
2693 | |
---|
2694 | return 0; |
---|
2695 | } |
---|
2696 | |
---|
2697 | |
---|
2698 | |
---|
2699 | /* |
---|
2700 | * cannot attach if : |
---|
2701 | * - kernel task |
---|
2702 | * - task not owned by caller |
---|
2703 | * - task incompatible with context mode |
---|
2704 | */ |
---|
2705 | static int |
---|
2706 | pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task) |
---|
2707 | { |
---|
2708 | /* |
---|
2709 | * no kernel task or task not owner by caller |
---|
2710 | */ |
---|
2711 | if (task->mm == NULL) { |
---|
2712 | DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid)); |
---|
2713 | return -EPERM; |
---|
2714 | } |
---|
2715 | if (pfm_bad_permissions(task)) { |
---|
2716 | DPRINT(("no permission to attach to [%d]\n", task->pid)); |
---|
2717 | return -EPERM; |
---|
2718 | } |
---|
2719 | /* |
---|
2720 | * cannot block in self-monitoring mode |
---|
2721 | */ |
---|
2722 | if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) { |
---|
2723 | DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid)); |
---|
2724 | return -EINVAL; |
---|
2725 | } |
---|
2726 | |
---|
2727 | if (task->exit_state == EXIT_ZOMBIE) { |
---|
2728 | DPRINT(("cannot attach to zombie task [%d]\n", task->pid)); |
---|
2729 | return -EBUSY; |
---|
2730 | } |
---|
2731 | |
---|
2732 | /* |
---|
2733 | * always ok for self |
---|
2734 | */ |
---|
2735 | if (task == current) return 0; |
---|
2736 | |
---|
2737 | if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) { |
---|
2738 | DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state)); |
---|
2739 | return -EBUSY; |
---|
2740 | } |
---|
2741 | /* |
---|
2742 | * make sure the task is off any CPU |
---|
2743 | */ |
---|
2744 | wait_task_inactive(task); |
---|
2745 | |
---|
2746 | /* more to come... */ |
---|
2747 | |
---|
2748 | return 0; |
---|
2749 | } |
---|
2750 | |
---|
2751 | static int |
---|
2752 | pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task) |
---|
2753 | { |
---|
2754 | struct task_struct *p = current; |
---|
2755 | int ret; |
---|
2756 | |
---|
2757 | /* XXX: need to add more checks here */ |
---|
2758 | if (pid < 2) return -EPERM; |
---|
2759 | |
---|
2760 | if (pid != current->pid) { |
---|
2761 | |
---|
2762 | read_lock(&tasklist_lock); |
---|
2763 | |
---|
2764 | p = find_task_by_pid(pid); |
---|
2765 | |
---|
2766 | /* make sure task cannot go away while we operate on it */ |
---|
2767 | if (p) get_task_struct(p); |
---|
2768 | |
---|
2769 | read_unlock(&tasklist_lock); |
---|
2770 | |
---|
2771 | if (p == NULL) return -ESRCH; |
---|
2772 | } |
---|
2773 | |
---|
2774 | ret = pfm_task_incompatible(ctx, p); |
---|
2775 | if (ret == 0) { |
---|
2776 | *task = p; |
---|
2777 | } else if (p != current) { |
---|
2778 | pfm_put_task(p); |
---|
2779 | } |
---|
2780 | return ret; |
---|
2781 | } |
---|
2782 | #endif |
---|
2783 | |
---|
2784 | |
---|
2785 | #ifndef XEN |
---|
2786 | static int |
---|
2787 | pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
2788 | #else |
---|
2789 | static pfm_context_t* |
---|
2790 | pfm_context_create(pfarg_context_t* req) |
---|
2791 | #endif |
---|
2792 | { |
---|
2793 | #ifndef XEN |
---|
2794 | pfarg_context_t *req = (pfarg_context_t *)arg; |
---|
2795 | struct file *filp; |
---|
2796 | #else |
---|
2797 | pfm_context_t *ctx; |
---|
2798 | #endif |
---|
2799 | int ctx_flags; |
---|
2800 | int ret; |
---|
2801 | |
---|
2802 | #ifndef XEN |
---|
2803 | /* let's check the arguments first */ |
---|
2804 | ret = pfarg_is_sane(current, req); |
---|
2805 | if (ret < 0) return ret; |
---|
2806 | #endif |
---|
2807 | |
---|
2808 | ctx_flags = req->ctx_flags; |
---|
2809 | |
---|
2810 | ret = -ENOMEM; |
---|
2811 | |
---|
2812 | ctx = pfm_context_alloc(); |
---|
2813 | if (!ctx) goto error; |
---|
2814 | |
---|
2815 | #ifndef XEN |
---|
2816 | ret = pfm_alloc_fd(&filp); |
---|
2817 | if (ret < 0) goto error_file; |
---|
2818 | |
---|
2819 | req->ctx_fd = ctx->ctx_fd = ret; |
---|
2820 | |
---|
2821 | /* |
---|
2822 | * attach context to file |
---|
2823 | */ |
---|
2824 | filp->private_data = ctx; |
---|
2825 | #endif |
---|
2826 | |
---|
2827 | /* |
---|
2828 | * does the user want to sample? |
---|
2829 | */ |
---|
2830 | if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) { |
---|
2831 | ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req); |
---|
2832 | if (ret) goto buffer_error; |
---|
2833 | } |
---|
2834 | |
---|
2835 | /* |
---|
2836 | * init context protection lock |
---|
2837 | */ |
---|
2838 | spin_lock_init(&ctx->ctx_lock); |
---|
2839 | |
---|
2840 | /* |
---|
2841 | * context is unloaded |
---|
2842 | */ |
---|
2843 | ctx->ctx_state = PFM_CTX_UNLOADED; |
---|
2844 | |
---|
2845 | /* |
---|
2846 | * initialization of context's flags |
---|
2847 | */ |
---|
2848 | ctx->ctx_fl_block = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0; |
---|
2849 | ctx->ctx_fl_system = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0; |
---|
2850 | ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */ |
---|
2851 | ctx->ctx_fl_no_msg = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0; |
---|
2852 | /* |
---|
2853 | * will move to set properties |
---|
2854 | * ctx->ctx_fl_excl_idle = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0; |
---|
2855 | */ |
---|
2856 | |
---|
2857 | #ifndef XEN |
---|
2858 | /* |
---|
2859 | * init restart semaphore to locked |
---|
2860 | */ |
---|
2861 | init_completion(&ctx->ctx_restart_done); |
---|
2862 | #endif |
---|
2863 | |
---|
2864 | /* |
---|
2865 | * activation is used in SMP only |
---|
2866 | */ |
---|
2867 | ctx->ctx_last_activation = PFM_INVALID_ACTIVATION; |
---|
2868 | SET_LAST_CPU(ctx, -1); |
---|
2869 | |
---|
2870 | #ifndef XEN |
---|
2871 | /* |
---|
2872 | * initialize notification message queue |
---|
2873 | */ |
---|
2874 | ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0; |
---|
2875 | init_waitqueue_head(&ctx->ctx_msgq_wait); |
---|
2876 | init_waitqueue_head(&ctx->ctx_zombieq); |
---|
2877 | #endif |
---|
2878 | |
---|
2879 | DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n", |
---|
2880 | ctx, |
---|
2881 | ctx_flags, |
---|
2882 | ctx->ctx_fl_system, |
---|
2883 | ctx->ctx_fl_block, |
---|
2884 | ctx->ctx_fl_excl_idle, |
---|
2885 | ctx->ctx_fl_no_msg, |
---|
2886 | ctx->ctx_fd)); |
---|
2887 | |
---|
2888 | /* |
---|
2889 | * initialize soft PMU state |
---|
2890 | */ |
---|
2891 | pfm_reset_pmu_state(ctx); |
---|
2892 | |
---|
2893 | #ifndef XEN |
---|
2894 | return 0; |
---|
2895 | #else |
---|
2896 | return ctx; |
---|
2897 | #endif |
---|
2898 | |
---|
2899 | buffer_error: |
---|
2900 | #ifndef XEN |
---|
2901 | pfm_free_fd(ctx->ctx_fd, filp); |
---|
2902 | #endif |
---|
2903 | |
---|
2904 | if (ctx->ctx_buf_fmt) { |
---|
2905 | #ifndef XEN |
---|
2906 | pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs); |
---|
2907 | #else |
---|
2908 | pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, NULL); |
---|
2909 | #endif |
---|
2910 | } |
---|
2911 | #ifndef XEN |
---|
2912 | error_file: |
---|
2913 | #endif |
---|
2914 | pfm_context_free(ctx); |
---|
2915 | |
---|
2916 | error: |
---|
2917 | #ifndef XEN |
---|
2918 | return ret; |
---|
2919 | #else |
---|
2920 | return NULL; |
---|
2921 | #endif |
---|
2922 | } |
---|
2923 | |
---|
2924 | static inline unsigned long |
---|
2925 | pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset) |
---|
2926 | { |
---|
2927 | unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset; |
---|
2928 | unsigned long new_seed, old_seed = reg->seed, mask = reg->mask; |
---|
2929 | extern unsigned long carta_random32 (unsigned long seed); |
---|
2930 | |
---|
2931 | if (reg->flags & PFM_REGFL_RANDOM) { |
---|
2932 | new_seed = carta_random32(old_seed); |
---|
2933 | val -= (old_seed & mask); /* counter values are negative numbers! */ |
---|
2934 | if ((mask >> 32) != 0) |
---|
2935 | /* construct a full 64-bit random value: */ |
---|
2936 | new_seed |= carta_random32(old_seed >> 32) << 32; |
---|
2937 | reg->seed = new_seed; |
---|
2938 | } |
---|
2939 | reg->lval = val; |
---|
2940 | return val; |
---|
2941 | } |
---|
2942 | |
---|
2943 | static void |
---|
2944 | pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset) |
---|
2945 | { |
---|
2946 | unsigned long mask = ovfl_regs[0]; |
---|
2947 | unsigned long reset_others = 0UL; |
---|
2948 | unsigned long val; |
---|
2949 | int i; |
---|
2950 | |
---|
2951 | /* |
---|
2952 | * now restore reset value on sampling overflowed counters |
---|
2953 | */ |
---|
2954 | mask >>= PMU_FIRST_COUNTER; |
---|
2955 | for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) { |
---|
2956 | |
---|
2957 | if ((mask & 0x1UL) == 0UL) continue; |
---|
2958 | |
---|
2959 | ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset); |
---|
2960 | reset_others |= ctx->ctx_pmds[i].reset_pmds[0]; |
---|
2961 | |
---|
2962 | DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val)); |
---|
2963 | } |
---|
2964 | |
---|
2965 | /* |
---|
2966 | * Now take care of resetting the other registers |
---|
2967 | */ |
---|
2968 | for(i = 0; reset_others; i++, reset_others >>= 1) { |
---|
2969 | |
---|
2970 | if ((reset_others & 0x1) == 0) continue; |
---|
2971 | |
---|
2972 | ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset); |
---|
2973 | |
---|
2974 | DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n", |
---|
2975 | is_long_reset ? "long" : "short", i, val)); |
---|
2976 | } |
---|
2977 | } |
---|
2978 | |
---|
2979 | static void |
---|
2980 | pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset) |
---|
2981 | { |
---|
2982 | unsigned long mask = ovfl_regs[0]; |
---|
2983 | unsigned long reset_others = 0UL; |
---|
2984 | unsigned long val; |
---|
2985 | int i; |
---|
2986 | |
---|
2987 | DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset)); |
---|
2988 | |
---|
2989 | if (ctx->ctx_state == PFM_CTX_MASKED) { |
---|
2990 | pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset); |
---|
2991 | return; |
---|
2992 | } |
---|
2993 | |
---|
2994 | /* |
---|
2995 | * now restore reset value on sampling overflowed counters |
---|
2996 | */ |
---|
2997 | mask >>= PMU_FIRST_COUNTER; |
---|
2998 | for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) { |
---|
2999 | |
---|
3000 | if ((mask & 0x1UL) == 0UL) continue; |
---|
3001 | |
---|
3002 | val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset); |
---|
3003 | reset_others |= ctx->ctx_pmds[i].reset_pmds[0]; |
---|
3004 | |
---|
3005 | DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val)); |
---|
3006 | |
---|
3007 | pfm_write_soft_counter(ctx, i, val); |
---|
3008 | } |
---|
3009 | |
---|
3010 | /* |
---|
3011 | * Now take care of resetting the other registers |
---|
3012 | */ |
---|
3013 | for(i = 0; reset_others; i++, reset_others >>= 1) { |
---|
3014 | |
---|
3015 | if ((reset_others & 0x1) == 0) continue; |
---|
3016 | |
---|
3017 | val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset); |
---|
3018 | |
---|
3019 | if (PMD_IS_COUNTING(i)) { |
---|
3020 | pfm_write_soft_counter(ctx, i, val); |
---|
3021 | } else { |
---|
3022 | ia64_set_pmd(i, val); |
---|
3023 | } |
---|
3024 | DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n", |
---|
3025 | is_long_reset ? "long" : "short", i, val)); |
---|
3026 | } |
---|
3027 | ia64_srlz_d(); |
---|
3028 | } |
---|
3029 | |
---|
3030 | static int |
---|
3031 | pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3032 | { |
---|
3033 | #ifndef XEN |
---|
3034 | struct thread_struct *thread = NULL; |
---|
3035 | #endif |
---|
3036 | struct task_struct *task; |
---|
3037 | pfarg_reg_t *req = (pfarg_reg_t *)arg; |
---|
3038 | unsigned long value, pmc_pm; |
---|
3039 | unsigned long smpl_pmds, reset_pmds, impl_pmds; |
---|
3040 | unsigned int cnum, reg_flags, flags, pmc_type; |
---|
3041 | int i, can_access_pmu = 0, is_loaded, is_system, expert_mode; |
---|
3042 | int is_monitor, is_counting, state; |
---|
3043 | int ret = -EINVAL; |
---|
3044 | pfm_reg_check_t wr_func; |
---|
3045 | #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z)) |
---|
3046 | |
---|
3047 | state = ctx->ctx_state; |
---|
3048 | is_loaded = state == PFM_CTX_LOADED ? 1 : 0; |
---|
3049 | is_system = ctx->ctx_fl_system; |
---|
3050 | task = ctx->ctx_task; |
---|
3051 | impl_pmds = pmu_conf->impl_pmds[0]; |
---|
3052 | #ifdef XEN |
---|
3053 | task = NULL; |
---|
3054 | BUG_ON(regs != NULL); |
---|
3055 | #endif |
---|
3056 | |
---|
3057 | if (state == PFM_CTX_ZOMBIE) return -EINVAL; |
---|
3058 | |
---|
3059 | #ifndef XEN |
---|
3060 | if (is_loaded) { |
---|
3061 | thread = &task->thread; |
---|
3062 | /* |
---|
3063 | * In system wide and when the context is loaded, access can only happen |
---|
3064 | * when the caller is running on the CPU being monitored by the session. |
---|
3065 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
3066 | */ |
---|
3067 | if (is_system && ctx->ctx_cpu != smp_processor_id()) { |
---|
3068 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
3069 | return -EBUSY; |
---|
3070 | } |
---|
3071 | can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0; |
---|
3072 | } |
---|
3073 | #else |
---|
3074 | /* XXX FIXME */ |
---|
3075 | if (state != PFM_CTX_UNLOADED) { |
---|
3076 | return -EBUSY; |
---|
3077 | } |
---|
3078 | #endif |
---|
3079 | |
---|
3080 | expert_mode = pfm_sysctl.expert_mode; |
---|
3081 | |
---|
3082 | for (i = 0; i < count; i++, req++) { |
---|
3083 | |
---|
3084 | cnum = req->reg_num; |
---|
3085 | reg_flags = req->reg_flags; |
---|
3086 | value = req->reg_value; |
---|
3087 | smpl_pmds = req->reg_smpl_pmds[0]; |
---|
3088 | reset_pmds = req->reg_reset_pmds[0]; |
---|
3089 | flags = 0; |
---|
3090 | |
---|
3091 | |
---|
3092 | if (cnum >= PMU_MAX_PMCS) { |
---|
3093 | DPRINT(("pmc%u is invalid\n", cnum)); |
---|
3094 | goto error; |
---|
3095 | } |
---|
3096 | |
---|
3097 | pmc_type = pmu_conf->pmc_desc[cnum].type; |
---|
3098 | pmc_pm = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1; |
---|
3099 | is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0; |
---|
3100 | is_monitor = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0; |
---|
3101 | |
---|
3102 | /* |
---|
3103 | * we reject all non implemented PMC as well |
---|
3104 | * as attempts to modify PMC[0-3] which are used |
---|
3105 | * as status registers by the PMU |
---|
3106 | */ |
---|
3107 | if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) { |
---|
3108 | DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type)); |
---|
3109 | goto error; |
---|
3110 | } |
---|
3111 | wr_func = pmu_conf->pmc_desc[cnum].write_check; |
---|
3112 | /* |
---|
3113 | * If the PMC is a monitor, then if the value is not the default: |
---|
3114 | * - system-wide session: PMCx.pm=1 (privileged monitor) |
---|
3115 | * - per-task : PMCx.pm=0 (user monitor) |
---|
3116 | */ |
---|
3117 | if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) { |
---|
3118 | DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n", |
---|
3119 | cnum, |
---|
3120 | pmc_pm, |
---|
3121 | is_system)); |
---|
3122 | goto error; |
---|
3123 | } |
---|
3124 | |
---|
3125 | if (is_counting) { |
---|
3126 | /* |
---|
3127 | * enforce generation of overflow interrupt. Necessary on all |
---|
3128 | * CPUs. |
---|
3129 | */ |
---|
3130 | value |= 1 << PMU_PMC_OI; |
---|
3131 | |
---|
3132 | if (reg_flags & PFM_REGFL_OVFL_NOTIFY) { |
---|
3133 | flags |= PFM_REGFL_OVFL_NOTIFY; |
---|
3134 | } |
---|
3135 | |
---|
3136 | if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM; |
---|
3137 | |
---|
3138 | /* verify validity of smpl_pmds */ |
---|
3139 | if ((smpl_pmds & impl_pmds) != smpl_pmds) { |
---|
3140 | DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum)); |
---|
3141 | goto error; |
---|
3142 | } |
---|
3143 | |
---|
3144 | /* verify validity of reset_pmds */ |
---|
3145 | if ((reset_pmds & impl_pmds) != reset_pmds) { |
---|
3146 | DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum)); |
---|
3147 | goto error; |
---|
3148 | } |
---|
3149 | } else { |
---|
3150 | if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) { |
---|
3151 | DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum)); |
---|
3152 | goto error; |
---|
3153 | } |
---|
3154 | /* eventid on non-counting monitors are ignored */ |
---|
3155 | } |
---|
3156 | |
---|
3157 | /* |
---|
3158 | * execute write checker, if any |
---|
3159 | */ |
---|
3160 | if (likely(expert_mode == 0 && wr_func)) { |
---|
3161 | ret = (*wr_func)(task, ctx, cnum, &value, regs); |
---|
3162 | if (ret) goto error; |
---|
3163 | ret = -EINVAL; |
---|
3164 | } |
---|
3165 | |
---|
3166 | /* |
---|
3167 | * no error on this register |
---|
3168 | */ |
---|
3169 | PFM_REG_RETFLAG_SET(req->reg_flags, 0); |
---|
3170 | |
---|
3171 | /* |
---|
3172 | * Now we commit the changes to the software state |
---|
3173 | */ |
---|
3174 | |
---|
3175 | /* |
---|
3176 | * update overflow information |
---|
3177 | */ |
---|
3178 | if (is_counting) { |
---|
3179 | /* |
---|
3180 | * full flag update each time a register is programmed |
---|
3181 | */ |
---|
3182 | ctx->ctx_pmds[cnum].flags = flags; |
---|
3183 | |
---|
3184 | ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds; |
---|
3185 | ctx->ctx_pmds[cnum].smpl_pmds[0] = smpl_pmds; |
---|
3186 | ctx->ctx_pmds[cnum].eventid = req->reg_smpl_eventid; |
---|
3187 | |
---|
3188 | /* |
---|
3189 | * Mark all PMDS to be accessed as used. |
---|
3190 | * |
---|
3191 | * We do not keep track of PMC because we have to |
---|
3192 | * systematically restore ALL of them. |
---|
3193 | * |
---|
3194 | * We do not update the used_monitors mask, because |
---|
3195 | * if we have not programmed them, then will be in |
---|
3196 | * a quiescent state, therefore we will not need to |
---|
3197 | * mask/restore then when context is MASKED. |
---|
3198 | */ |
---|
3199 | CTX_USED_PMD(ctx, reset_pmds); |
---|
3200 | CTX_USED_PMD(ctx, smpl_pmds); |
---|
3201 | /* |
---|
3202 | * make sure we do not try to reset on |
---|
3203 | * restart because we have established new values |
---|
3204 | */ |
---|
3205 | if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum; |
---|
3206 | } |
---|
3207 | /* |
---|
3208 | * Needed in case the user does not initialize the equivalent |
---|
3209 | * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no |
---|
3210 | * possible leak here. |
---|
3211 | */ |
---|
3212 | CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]); |
---|
3213 | |
---|
3214 | /* |
---|
3215 | * keep track of the monitor PMC that we are using. |
---|
3216 | * we save the value of the pmc in ctx_pmcs[] and if |
---|
3217 | * the monitoring is not stopped for the context we also |
---|
3218 | * place it in the saved state area so that it will be |
---|
3219 | * picked up later by the context switch code. |
---|
3220 | * |
---|
3221 | * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs(). |
---|
3222 | * |
---|
3223 | * The value in thread->pmcs[] may be modified on overflow, i.e., when |
---|
3224 | * monitoring needs to be stopped. |
---|
3225 | */ |
---|
3226 | if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum); |
---|
3227 | |
---|
3228 | /* |
---|
3229 | * update context state |
---|
3230 | */ |
---|
3231 | ctx->ctx_pmcs[cnum] = value; |
---|
3232 | |
---|
3233 | #ifndef XEN |
---|
3234 | if (is_loaded) { |
---|
3235 | /* |
---|
3236 | * write thread state |
---|
3237 | */ |
---|
3238 | if (is_system == 0) thread->pmcs[cnum] = value; |
---|
3239 | |
---|
3240 | /* |
---|
3241 | * write hardware register if we can |
---|
3242 | */ |
---|
3243 | if (can_access_pmu) { |
---|
3244 | ia64_set_pmc(cnum, value); |
---|
3245 | } |
---|
3246 | #ifdef CONFIG_SMP |
---|
3247 | else { |
---|
3248 | /* |
---|
3249 | * per-task SMP only here |
---|
3250 | * |
---|
3251 | * we are guaranteed that the task is not running on the other CPU, |
---|
3252 | * we indicate that this PMD will need to be reloaded if the task |
---|
3253 | * is rescheduled on the CPU it ran last on. |
---|
3254 | */ |
---|
3255 | ctx->ctx_reload_pmcs[0] |= 1UL << cnum; |
---|
3256 | } |
---|
3257 | #endif |
---|
3258 | } |
---|
3259 | #endif |
---|
3260 | |
---|
3261 | DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n", |
---|
3262 | cnum, |
---|
3263 | value, |
---|
3264 | is_loaded, |
---|
3265 | can_access_pmu, |
---|
3266 | flags, |
---|
3267 | ctx->ctx_all_pmcs[0], |
---|
3268 | ctx->ctx_used_pmds[0], |
---|
3269 | ctx->ctx_pmds[cnum].eventid, |
---|
3270 | smpl_pmds, |
---|
3271 | reset_pmds, |
---|
3272 | ctx->ctx_reload_pmcs[0], |
---|
3273 | ctx->ctx_used_monitors[0], |
---|
3274 | ctx->ctx_ovfl_regs[0])); |
---|
3275 | } |
---|
3276 | |
---|
3277 | /* |
---|
3278 | * make sure the changes are visible |
---|
3279 | */ |
---|
3280 | if (can_access_pmu) ia64_srlz_d(); |
---|
3281 | |
---|
3282 | return 0; |
---|
3283 | error: |
---|
3284 | PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL); |
---|
3285 | return ret; |
---|
3286 | } |
---|
3287 | |
---|
3288 | static int |
---|
3289 | pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3290 | { |
---|
3291 | #ifndef XEN |
---|
3292 | struct thread_struct *thread = NULL; |
---|
3293 | #endif |
---|
3294 | struct task_struct *task; |
---|
3295 | pfarg_reg_t *req = (pfarg_reg_t *)arg; |
---|
3296 | unsigned long value, hw_value, ovfl_mask; |
---|
3297 | unsigned int cnum; |
---|
3298 | int i, can_access_pmu = 0, state; |
---|
3299 | int is_counting, is_loaded, is_system, expert_mode; |
---|
3300 | int ret = -EINVAL; |
---|
3301 | pfm_reg_check_t wr_func; |
---|
3302 | |
---|
3303 | |
---|
3304 | state = ctx->ctx_state; |
---|
3305 | is_loaded = state == PFM_CTX_LOADED ? 1 : 0; |
---|
3306 | is_system = ctx->ctx_fl_system; |
---|
3307 | ovfl_mask = pmu_conf->ovfl_val; |
---|
3308 | task = ctx->ctx_task; |
---|
3309 | #ifdef XEN |
---|
3310 | task = NULL; |
---|
3311 | BUG_ON(regs != NULL); |
---|
3312 | #endif |
---|
3313 | |
---|
3314 | if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL; |
---|
3315 | |
---|
3316 | #ifndef XEN |
---|
3317 | /* |
---|
3318 | * on both UP and SMP, we can only write to the PMC when the task is |
---|
3319 | * the owner of the local PMU. |
---|
3320 | */ |
---|
3321 | if (likely(is_loaded)) { |
---|
3322 | thread = &task->thread; |
---|
3323 | /* |
---|
3324 | * In system wide and when the context is loaded, access can only happen |
---|
3325 | * when the caller is running on the CPU being monitored by the session. |
---|
3326 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
3327 | */ |
---|
3328 | if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) { |
---|
3329 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
3330 | return -EBUSY; |
---|
3331 | } |
---|
3332 | can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0; |
---|
3333 | } |
---|
3334 | #else |
---|
3335 | /* XXX FIXME */ |
---|
3336 | if (state != PFM_CTX_UNLOADED) { |
---|
3337 | return -EBUSY; |
---|
3338 | } |
---|
3339 | #endif |
---|
3340 | expert_mode = pfm_sysctl.expert_mode; |
---|
3341 | |
---|
3342 | for (i = 0; i < count; i++, req++) { |
---|
3343 | |
---|
3344 | cnum = req->reg_num; |
---|
3345 | value = req->reg_value; |
---|
3346 | |
---|
3347 | if (!PMD_IS_IMPL(cnum)) { |
---|
3348 | DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum)); |
---|
3349 | goto abort_mission; |
---|
3350 | } |
---|
3351 | is_counting = PMD_IS_COUNTING(cnum); |
---|
3352 | wr_func = pmu_conf->pmd_desc[cnum].write_check; |
---|
3353 | |
---|
3354 | /* |
---|
3355 | * execute write checker, if any |
---|
3356 | */ |
---|
3357 | if (unlikely(expert_mode == 0 && wr_func)) { |
---|
3358 | unsigned long v = value; |
---|
3359 | |
---|
3360 | ret = (*wr_func)(task, ctx, cnum, &v, regs); |
---|
3361 | if (ret) goto abort_mission; |
---|
3362 | |
---|
3363 | value = v; |
---|
3364 | ret = -EINVAL; |
---|
3365 | } |
---|
3366 | |
---|
3367 | /* |
---|
3368 | * no error on this register |
---|
3369 | */ |
---|
3370 | PFM_REG_RETFLAG_SET(req->reg_flags, 0); |
---|
3371 | |
---|
3372 | /* |
---|
3373 | * now commit changes to software state |
---|
3374 | */ |
---|
3375 | hw_value = value; |
---|
3376 | |
---|
3377 | /* |
---|
3378 | * update virtualized (64bits) counter |
---|
3379 | */ |
---|
3380 | if (is_counting) { |
---|
3381 | /* |
---|
3382 | * write context state |
---|
3383 | */ |
---|
3384 | ctx->ctx_pmds[cnum].lval = value; |
---|
3385 | |
---|
3386 | /* |
---|
3387 | * when context is load we use the split value |
---|
3388 | */ |
---|
3389 | if (is_loaded) { |
---|
3390 | hw_value = value & ovfl_mask; |
---|
3391 | value = value & ~ovfl_mask; |
---|
3392 | } |
---|
3393 | } |
---|
3394 | /* |
---|
3395 | * update reset values (not just for counters) |
---|
3396 | */ |
---|
3397 | ctx->ctx_pmds[cnum].long_reset = req->reg_long_reset; |
---|
3398 | ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset; |
---|
3399 | |
---|
3400 | /* |
---|
3401 | * update randomization parameters (not just for counters) |
---|
3402 | */ |
---|
3403 | ctx->ctx_pmds[cnum].seed = req->reg_random_seed; |
---|
3404 | ctx->ctx_pmds[cnum].mask = req->reg_random_mask; |
---|
3405 | |
---|
3406 | /* |
---|
3407 | * update context value |
---|
3408 | */ |
---|
3409 | ctx->ctx_pmds[cnum].val = value; |
---|
3410 | |
---|
3411 | /* |
---|
3412 | * Keep track of what we use |
---|
3413 | * |
---|
3414 | * We do not keep track of PMC because we have to |
---|
3415 | * systematically restore ALL of them. |
---|
3416 | */ |
---|
3417 | CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum)); |
---|
3418 | |
---|
3419 | /* |
---|
3420 | * mark this PMD register used as well |
---|
3421 | */ |
---|
3422 | CTX_USED_PMD(ctx, RDEP(cnum)); |
---|
3423 | |
---|
3424 | /* |
---|
3425 | * make sure we do not try to reset on |
---|
3426 | * restart because we have established new values |
---|
3427 | */ |
---|
3428 | if (is_counting && state == PFM_CTX_MASKED) { |
---|
3429 | ctx->ctx_ovfl_regs[0] &= ~1UL << cnum; |
---|
3430 | } |
---|
3431 | |
---|
3432 | /* XXX FIXME */ |
---|
3433 | #ifndef XEN |
---|
3434 | if (is_loaded) { |
---|
3435 | /* |
---|
3436 | * write thread state |
---|
3437 | */ |
---|
3438 | if (is_system == 0) thread->pmds[cnum] = hw_value; |
---|
3439 | |
---|
3440 | /* |
---|
3441 | * write hardware register if we can |
---|
3442 | */ |
---|
3443 | if (can_access_pmu) { |
---|
3444 | ia64_set_pmd(cnum, hw_value); |
---|
3445 | } else { |
---|
3446 | #ifdef CONFIG_SMP |
---|
3447 | /* |
---|
3448 | * we are guaranteed that the task is not running on the other CPU, |
---|
3449 | * we indicate that this PMD will need to be reloaded if the task |
---|
3450 | * is rescheduled on the CPU it ran last on. |
---|
3451 | */ |
---|
3452 | ctx->ctx_reload_pmds[0] |= 1UL << cnum; |
---|
3453 | #endif |
---|
3454 | } |
---|
3455 | } |
---|
3456 | #endif |
---|
3457 | |
---|
3458 | DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx short_reset=0x%lx " |
---|
3459 | "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n", |
---|
3460 | cnum, |
---|
3461 | value, |
---|
3462 | is_loaded, |
---|
3463 | can_access_pmu, |
---|
3464 | hw_value, |
---|
3465 | ctx->ctx_pmds[cnum].val, |
---|
3466 | ctx->ctx_pmds[cnum].short_reset, |
---|
3467 | ctx->ctx_pmds[cnum].long_reset, |
---|
3468 | PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N', |
---|
3469 | ctx->ctx_pmds[cnum].seed, |
---|
3470 | ctx->ctx_pmds[cnum].mask, |
---|
3471 | ctx->ctx_used_pmds[0], |
---|
3472 | ctx->ctx_pmds[cnum].reset_pmds[0], |
---|
3473 | ctx->ctx_reload_pmds[0], |
---|
3474 | ctx->ctx_all_pmds[0], |
---|
3475 | ctx->ctx_ovfl_regs[0])); |
---|
3476 | } |
---|
3477 | |
---|
3478 | /* |
---|
3479 | * make changes visible |
---|
3480 | */ |
---|
3481 | if (can_access_pmu) ia64_srlz_d(); |
---|
3482 | |
---|
3483 | return 0; |
---|
3484 | |
---|
3485 | abort_mission: |
---|
3486 | /* |
---|
3487 | * for now, we have only one possibility for error |
---|
3488 | */ |
---|
3489 | PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL); |
---|
3490 | return ret; |
---|
3491 | } |
---|
3492 | |
---|
3493 | #ifndef XEN |
---|
3494 | /* |
---|
3495 | * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function. |
---|
3496 | * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an |
---|
3497 | * interrupt is delivered during the call, it will be kept pending until we leave, making |
---|
3498 | * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are |
---|
3499 | * guaranteed to return consistent data to the user, it may simply be old. It is not |
---|
3500 | * trivial to treat the overflow while inside the call because you may end up in |
---|
3501 | * some module sampling buffer code causing deadlocks. |
---|
3502 | */ |
---|
3503 | static int |
---|
3504 | pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3505 | { |
---|
3506 | struct thread_struct *thread = NULL; |
---|
3507 | struct task_struct *task; |
---|
3508 | unsigned long val = 0UL, lval, ovfl_mask, sval; |
---|
3509 | pfarg_reg_t *req = (pfarg_reg_t *)arg; |
---|
3510 | unsigned int cnum, reg_flags = 0; |
---|
3511 | int i, can_access_pmu = 0, state; |
---|
3512 | int is_loaded, is_system, is_counting, expert_mode; |
---|
3513 | int ret = -EINVAL; |
---|
3514 | pfm_reg_check_t rd_func; |
---|
3515 | |
---|
3516 | /* |
---|
3517 | * access is possible when loaded only for |
---|
3518 | * self-monitoring tasks or in UP mode |
---|
3519 | */ |
---|
3520 | |
---|
3521 | state = ctx->ctx_state; |
---|
3522 | is_loaded = state == PFM_CTX_LOADED ? 1 : 0; |
---|
3523 | is_system = ctx->ctx_fl_system; |
---|
3524 | ovfl_mask = pmu_conf->ovfl_val; |
---|
3525 | task = ctx->ctx_task; |
---|
3526 | |
---|
3527 | if (state == PFM_CTX_ZOMBIE) return -EINVAL; |
---|
3528 | |
---|
3529 | if (likely(is_loaded)) { |
---|
3530 | thread = &task->thread; |
---|
3531 | /* |
---|
3532 | * In system wide and when the context is loaded, access can only happen |
---|
3533 | * when the caller is running on the CPU being monitored by the session. |
---|
3534 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
3535 | */ |
---|
3536 | if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) { |
---|
3537 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
3538 | return -EBUSY; |
---|
3539 | } |
---|
3540 | /* |
---|
3541 | * this can be true when not self-monitoring only in UP |
---|
3542 | */ |
---|
3543 | can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0; |
---|
3544 | |
---|
3545 | if (can_access_pmu) ia64_srlz_d(); |
---|
3546 | } |
---|
3547 | expert_mode = pfm_sysctl.expert_mode; |
---|
3548 | |
---|
3549 | DPRINT(("ld=%d apmu=%d ctx_state=%d\n", |
---|
3550 | is_loaded, |
---|
3551 | can_access_pmu, |
---|
3552 | state)); |
---|
3553 | |
---|
3554 | /* |
---|
3555 | * on both UP and SMP, we can only read the PMD from the hardware register when |
---|
3556 | * the task is the owner of the local PMU. |
---|
3557 | */ |
---|
3558 | |
---|
3559 | for (i = 0; i < count; i++, req++) { |
---|
3560 | |
---|
3561 | cnum = req->reg_num; |
---|
3562 | reg_flags = req->reg_flags; |
---|
3563 | |
---|
3564 | if (unlikely(!PMD_IS_IMPL(cnum))) goto error; |
---|
3565 | /* |
---|
3566 | * we can only read the register that we use. That includes |
---|
3567 | * the one we explicitely initialize AND the one we want included |
---|
3568 | * in the sampling buffer (smpl_regs). |
---|
3569 | * |
---|
3570 | * Having this restriction allows optimization in the ctxsw routine |
---|
3571 | * without compromising security (leaks) |
---|
3572 | */ |
---|
3573 | if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error; |
---|
3574 | |
---|
3575 | sval = ctx->ctx_pmds[cnum].val; |
---|
3576 | lval = ctx->ctx_pmds[cnum].lval; |
---|
3577 | is_counting = PMD_IS_COUNTING(cnum); |
---|
3578 | |
---|
3579 | /* |
---|
3580 | * If the task is not the current one, then we check if the |
---|
3581 | * PMU state is still in the local live register due to lazy ctxsw. |
---|
3582 | * If true, then we read directly from the registers. |
---|
3583 | */ |
---|
3584 | if (can_access_pmu){ |
---|
3585 | val = ia64_get_pmd(cnum); |
---|
3586 | } else { |
---|
3587 | /* |
---|
3588 | * context has been saved |
---|
3589 | * if context is zombie, then task does not exist anymore. |
---|
3590 | * In this case, we use the full value saved in the context (pfm_flush_regs()). |
---|
3591 | */ |
---|
3592 | val = is_loaded ? thread->pmds[cnum] : 0UL; |
---|
3593 | } |
---|
3594 | rd_func = pmu_conf->pmd_desc[cnum].read_check; |
---|
3595 | |
---|
3596 | if (is_counting) { |
---|
3597 | /* |
---|
3598 | * XXX: need to check for overflow when loaded |
---|
3599 | */ |
---|
3600 | val &= ovfl_mask; |
---|
3601 | val += sval; |
---|
3602 | } |
---|
3603 | |
---|
3604 | /* |
---|
3605 | * execute read checker, if any |
---|
3606 | */ |
---|
3607 | if (unlikely(expert_mode == 0 && rd_func)) { |
---|
3608 | unsigned long v = val; |
---|
3609 | ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs); |
---|
3610 | if (ret) goto error; |
---|
3611 | val = v; |
---|
3612 | ret = -EINVAL; |
---|
3613 | } |
---|
3614 | |
---|
3615 | PFM_REG_RETFLAG_SET(reg_flags, 0); |
---|
3616 | |
---|
3617 | DPRINT(("pmd[%u]=0x%lx\n", cnum, val)); |
---|
3618 | |
---|
3619 | /* |
---|
3620 | * update register return value, abort all if problem during copy. |
---|
3621 | * we only modify the reg_flags field. no check mode is fine because |
---|
3622 | * access has been verified upfront in sys_perfmonctl(). |
---|
3623 | */ |
---|
3624 | req->reg_value = val; |
---|
3625 | req->reg_flags = reg_flags; |
---|
3626 | req->reg_last_reset_val = lval; |
---|
3627 | } |
---|
3628 | |
---|
3629 | return 0; |
---|
3630 | |
---|
3631 | error: |
---|
3632 | PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL); |
---|
3633 | return ret; |
---|
3634 | } |
---|
3635 | |
---|
3636 | int |
---|
3637 | pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs) |
---|
3638 | { |
---|
3639 | pfm_context_t *ctx; |
---|
3640 | |
---|
3641 | if (req == NULL) return -EINVAL; |
---|
3642 | |
---|
3643 | ctx = GET_PMU_CTX(); |
---|
3644 | |
---|
3645 | if (ctx == NULL) return -EINVAL; |
---|
3646 | |
---|
3647 | /* |
---|
3648 | * for now limit to current task, which is enough when calling |
---|
3649 | * from overflow handler |
---|
3650 | */ |
---|
3651 | if (task != current && ctx->ctx_fl_system == 0) return -EBUSY; |
---|
3652 | |
---|
3653 | return pfm_write_pmcs(ctx, req, nreq, regs); |
---|
3654 | } |
---|
3655 | EXPORT_SYMBOL(pfm_mod_write_pmcs); |
---|
3656 | |
---|
3657 | int |
---|
3658 | pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs) |
---|
3659 | { |
---|
3660 | pfm_context_t *ctx; |
---|
3661 | |
---|
3662 | if (req == NULL) return -EINVAL; |
---|
3663 | |
---|
3664 | ctx = GET_PMU_CTX(); |
---|
3665 | |
---|
3666 | if (ctx == NULL) return -EINVAL; |
---|
3667 | |
---|
3668 | /* |
---|
3669 | * for now limit to current task, which is enough when calling |
---|
3670 | * from overflow handler |
---|
3671 | */ |
---|
3672 | if (task != current && ctx->ctx_fl_system == 0) return -EBUSY; |
---|
3673 | |
---|
3674 | return pfm_read_pmds(ctx, req, nreq, regs); |
---|
3675 | } |
---|
3676 | EXPORT_SYMBOL(pfm_mod_read_pmds); |
---|
3677 | #endif |
---|
3678 | |
---|
3679 | /* |
---|
3680 | * Only call this function when a process it trying to |
---|
3681 | * write the debug registers (reading is always allowed) |
---|
3682 | */ |
---|
3683 | int |
---|
3684 | pfm_use_debug_registers(struct task_struct *task) |
---|
3685 | { |
---|
3686 | pfm_context_t *ctx = task->thread.pfm_context; |
---|
3687 | unsigned long flags; |
---|
3688 | int ret = 0; |
---|
3689 | |
---|
3690 | if (pmu_conf->use_rr_dbregs == 0) return 0; |
---|
3691 | |
---|
3692 | DPRINT(("called for [%d]\n", task->pid)); |
---|
3693 | |
---|
3694 | /* |
---|
3695 | * do it only once |
---|
3696 | */ |
---|
3697 | if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0; |
---|
3698 | |
---|
3699 | /* |
---|
3700 | * Even on SMP, we do not need to use an atomic here because |
---|
3701 | * the only way in is via ptrace() and this is possible only when the |
---|
3702 | * process is stopped. Even in the case where the ctxsw out is not totally |
---|
3703 | * completed by the time we come here, there is no way the 'stopped' process |
---|
3704 | * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine. |
---|
3705 | * So this is always safe. |
---|
3706 | */ |
---|
3707 | if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1; |
---|
3708 | |
---|
3709 | LOCK_PFS(flags); |
---|
3710 | |
---|
3711 | /* |
---|
3712 | * We cannot allow setting breakpoints when system wide monitoring |
---|
3713 | * sessions are using the debug registers. |
---|
3714 | */ |
---|
3715 | if (pfm_sessions.pfs_sys_use_dbregs> 0) |
---|
3716 | ret = -1; |
---|
3717 | else |
---|
3718 | pfm_sessions.pfs_ptrace_use_dbregs++; |
---|
3719 | |
---|
3720 | DPRINT(("ptrace_use_dbregs=%u sys_use_dbregs=%u by [%d] ret = %d\n", |
---|
3721 | pfm_sessions.pfs_ptrace_use_dbregs, |
---|
3722 | pfm_sessions.pfs_sys_use_dbregs, |
---|
3723 | task->pid, ret)); |
---|
3724 | |
---|
3725 | UNLOCK_PFS(flags); |
---|
3726 | |
---|
3727 | return ret; |
---|
3728 | } |
---|
3729 | |
---|
3730 | /* |
---|
3731 | * This function is called for every task that exits with the |
---|
3732 | * IA64_THREAD_DBG_VALID set. This indicates a task which was |
---|
3733 | * able to use the debug registers for debugging purposes via |
---|
3734 | * ptrace(). Therefore we know it was not using them for |
---|
3735 | * perfmormance monitoring, so we only decrement the number |
---|
3736 | * of "ptraced" debug register users to keep the count up to date |
---|
3737 | */ |
---|
3738 | int |
---|
3739 | pfm_release_debug_registers(struct task_struct *task) |
---|
3740 | { |
---|
3741 | unsigned long flags; |
---|
3742 | int ret; |
---|
3743 | |
---|
3744 | if (pmu_conf->use_rr_dbregs == 0) return 0; |
---|
3745 | |
---|
3746 | LOCK_PFS(flags); |
---|
3747 | if (pfm_sessions.pfs_ptrace_use_dbregs == 0) { |
---|
3748 | printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid); |
---|
3749 | ret = -1; |
---|
3750 | } else { |
---|
3751 | pfm_sessions.pfs_ptrace_use_dbregs--; |
---|
3752 | ret = 0; |
---|
3753 | } |
---|
3754 | UNLOCK_PFS(flags); |
---|
3755 | |
---|
3756 | return ret; |
---|
3757 | } |
---|
3758 | |
---|
3759 | #ifndef XEN |
---|
3760 | static int |
---|
3761 | pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3762 | { |
---|
3763 | struct task_struct *task; |
---|
3764 | pfm_buffer_fmt_t *fmt; |
---|
3765 | pfm_ovfl_ctrl_t rst_ctrl; |
---|
3766 | int state, is_system; |
---|
3767 | int ret = 0; |
---|
3768 | |
---|
3769 | state = ctx->ctx_state; |
---|
3770 | fmt = ctx->ctx_buf_fmt; |
---|
3771 | is_system = ctx->ctx_fl_system; |
---|
3772 | task = PFM_CTX_TASK(ctx); |
---|
3773 | |
---|
3774 | switch(state) { |
---|
3775 | case PFM_CTX_MASKED: |
---|
3776 | break; |
---|
3777 | case PFM_CTX_LOADED: |
---|
3778 | if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break; |
---|
3779 | /* fall through */ |
---|
3780 | case PFM_CTX_UNLOADED: |
---|
3781 | case PFM_CTX_ZOMBIE: |
---|
3782 | DPRINT(("invalid state=%d\n", state)); |
---|
3783 | return -EBUSY; |
---|
3784 | default: |
---|
3785 | DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state)); |
---|
3786 | return -EINVAL; |
---|
3787 | } |
---|
3788 | |
---|
3789 | /* |
---|
3790 | * In system wide and when the context is loaded, access can only happen |
---|
3791 | * when the caller is running on the CPU being monitored by the session. |
---|
3792 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
3793 | */ |
---|
3794 | if (is_system && ctx->ctx_cpu != smp_processor_id()) { |
---|
3795 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
3796 | return -EBUSY; |
---|
3797 | } |
---|
3798 | |
---|
3799 | /* sanity check */ |
---|
3800 | if (unlikely(task == NULL)) { |
---|
3801 | printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid); |
---|
3802 | return -EINVAL; |
---|
3803 | } |
---|
3804 | |
---|
3805 | if (task == current || is_system) { |
---|
3806 | |
---|
3807 | fmt = ctx->ctx_buf_fmt; |
---|
3808 | |
---|
3809 | DPRINT(("restarting self %d ovfl=0x%lx\n", |
---|
3810 | task->pid, |
---|
3811 | ctx->ctx_ovfl_regs[0])); |
---|
3812 | |
---|
3813 | if (CTX_HAS_SMPL(ctx)) { |
---|
3814 | |
---|
3815 | prefetch(ctx->ctx_smpl_hdr); |
---|
3816 | |
---|
3817 | rst_ctrl.bits.mask_monitoring = 0; |
---|
3818 | rst_ctrl.bits.reset_ovfl_pmds = 0; |
---|
3819 | |
---|
3820 | if (state == PFM_CTX_LOADED) |
---|
3821 | ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs); |
---|
3822 | else |
---|
3823 | ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs); |
---|
3824 | } else { |
---|
3825 | rst_ctrl.bits.mask_monitoring = 0; |
---|
3826 | rst_ctrl.bits.reset_ovfl_pmds = 1; |
---|
3827 | } |
---|
3828 | |
---|
3829 | if (ret == 0) { |
---|
3830 | if (rst_ctrl.bits.reset_ovfl_pmds) |
---|
3831 | pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET); |
---|
3832 | |
---|
3833 | if (rst_ctrl.bits.mask_monitoring == 0) { |
---|
3834 | DPRINT(("resuming monitoring for [%d]\n", task->pid)); |
---|
3835 | |
---|
3836 | if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task); |
---|
3837 | } else { |
---|
3838 | DPRINT(("keeping monitoring stopped for [%d]\n", task->pid)); |
---|
3839 | |
---|
3840 | // cannot use pfm_stop_monitoring(task, regs); |
---|
3841 | } |
---|
3842 | } |
---|
3843 | /* |
---|
3844 | * clear overflowed PMD mask to remove any stale information |
---|
3845 | */ |
---|
3846 | ctx->ctx_ovfl_regs[0] = 0UL; |
---|
3847 | |
---|
3848 | /* |
---|
3849 | * back to LOADED state |
---|
3850 | */ |
---|
3851 | ctx->ctx_state = PFM_CTX_LOADED; |
---|
3852 | |
---|
3853 | /* |
---|
3854 | * XXX: not really useful for self monitoring |
---|
3855 | */ |
---|
3856 | ctx->ctx_fl_can_restart = 0; |
---|
3857 | |
---|
3858 | return 0; |
---|
3859 | } |
---|
3860 | |
---|
3861 | /* |
---|
3862 | * restart another task |
---|
3863 | */ |
---|
3864 | |
---|
3865 | /* |
---|
3866 | * When PFM_CTX_MASKED, we cannot issue a restart before the previous |
---|
3867 | * one is seen by the task. |
---|
3868 | */ |
---|
3869 | if (state == PFM_CTX_MASKED) { |
---|
3870 | if (ctx->ctx_fl_can_restart == 0) return -EINVAL; |
---|
3871 | /* |
---|
3872 | * will prevent subsequent restart before this one is |
---|
3873 | * seen by other task |
---|
3874 | */ |
---|
3875 | ctx->ctx_fl_can_restart = 0; |
---|
3876 | } |
---|
3877 | |
---|
3878 | /* |
---|
3879 | * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e. |
---|
3880 | * the task is blocked or on its way to block. That's the normal |
---|
3881 | * restart path. If the monitoring is not masked, then the task |
---|
3882 | * can be actively monitoring and we cannot directly intervene. |
---|
3883 | * Therefore we use the trap mechanism to catch the task and |
---|
3884 | * force it to reset the buffer/reset PMDs. |
---|
3885 | * |
---|
3886 | * if non-blocking, then we ensure that the task will go into |
---|
3887 | * pfm_handle_work() before returning to user mode. |
---|
3888 | * |
---|
3889 | * We cannot explicitely reset another task, it MUST always |
---|
3890 | * be done by the task itself. This works for system wide because |
---|
3891 | * the tool that is controlling the session is logically doing |
---|
3892 | * "self-monitoring". |
---|
3893 | */ |
---|
3894 | if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) { |
---|
3895 | DPRINT(("unblocking [%d] \n", task->pid)); |
---|
3896 | complete(&ctx->ctx_restart_done); |
---|
3897 | } else { |
---|
3898 | DPRINT(("[%d] armed exit trap\n", task->pid)); |
---|
3899 | |
---|
3900 | ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET; |
---|
3901 | |
---|
3902 | PFM_SET_WORK_PENDING(task, 1); |
---|
3903 | |
---|
3904 | pfm_set_task_notify(task); |
---|
3905 | |
---|
3906 | /* |
---|
3907 | * XXX: send reschedule if task runs on another CPU |
---|
3908 | */ |
---|
3909 | } |
---|
3910 | return 0; |
---|
3911 | } |
---|
3912 | |
---|
3913 | static int |
---|
3914 | pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3915 | { |
---|
3916 | unsigned int m = *(unsigned int *)arg; |
---|
3917 | |
---|
3918 | pfm_sysctl.debug = m == 0 ? 0 : 1; |
---|
3919 | |
---|
3920 | printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off"); |
---|
3921 | |
---|
3922 | if (m == 0) { |
---|
3923 | memset(pfm_stats, 0, sizeof(pfm_stats)); |
---|
3924 | for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL; |
---|
3925 | } |
---|
3926 | return 0; |
---|
3927 | } |
---|
3928 | #endif |
---|
3929 | |
---|
3930 | /* |
---|
3931 | * arg can be NULL and count can be zero for this function |
---|
3932 | */ |
---|
3933 | static int |
---|
3934 | pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
3935 | { |
---|
3936 | #ifndef XEN |
---|
3937 | struct thread_struct *thread = NULL; |
---|
3938 | #endif |
---|
3939 | struct task_struct *task; |
---|
3940 | pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg; |
---|
3941 | unsigned long flags; |
---|
3942 | dbreg_t dbreg; |
---|
3943 | unsigned int rnum; |
---|
3944 | int first_time; |
---|
3945 | int ret = 0, state; |
---|
3946 | int i, can_access_pmu = 0; |
---|
3947 | int is_system, is_loaded; |
---|
3948 | |
---|
3949 | if (pmu_conf->use_rr_dbregs == 0) return -EINVAL; |
---|
3950 | |
---|
3951 | state = ctx->ctx_state; |
---|
3952 | is_loaded = state == PFM_CTX_LOADED ? 1 : 0; |
---|
3953 | is_system = ctx->ctx_fl_system; |
---|
3954 | task = ctx->ctx_task; |
---|
3955 | #ifdef XEN |
---|
3956 | task = NULL; |
---|
3957 | BUG_ON(regs != NULL); |
---|
3958 | /* currently dbrs, ibrs aren't supported */ |
---|
3959 | BUG(); |
---|
3960 | #endif |
---|
3961 | |
---|
3962 | if (state == PFM_CTX_ZOMBIE) return -EINVAL; |
---|
3963 | |
---|
3964 | /* |
---|
3965 | * on both UP and SMP, we can only write to the PMC when the task is |
---|
3966 | * the owner of the local PMU. |
---|
3967 | */ |
---|
3968 | if (is_loaded) { |
---|
3969 | #ifdef XEN |
---|
3970 | /* XXX */ |
---|
3971 | return -EBUSY; |
---|
3972 | #else |
---|
3973 | thread = &task->thread; |
---|
3974 | /* |
---|
3975 | * In system wide and when the context is loaded, access can only happen |
---|
3976 | * when the caller is running on the CPU being monitored by the session. |
---|
3977 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
3978 | */ |
---|
3979 | if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) { |
---|
3980 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
3981 | return -EBUSY; |
---|
3982 | } |
---|
3983 | can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0; |
---|
3984 | #endif |
---|
3985 | } |
---|
3986 | |
---|
3987 | /* |
---|
3988 | * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w |
---|
3989 | * ensuring that no real breakpoint can be installed via this call. |
---|
3990 | * |
---|
3991 | * IMPORTANT: regs can be NULL in this function |
---|
3992 | */ |
---|
3993 | |
---|
3994 | first_time = ctx->ctx_fl_using_dbreg == 0; |
---|
3995 | |
---|
3996 | /* |
---|
3997 | * don't bother if we are loaded and task is being debugged |
---|
3998 | */ |
---|
3999 | #ifndef XEN |
---|
4000 | if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) { |
---|
4001 | DPRINT(("debug registers already in use for [%d]\n", task->pid)); |
---|
4002 | return -EBUSY; |
---|
4003 | } |
---|
4004 | #else |
---|
4005 | /* Currently no support for is_loaded, see -EBUSY above */ |
---|
4006 | #endif |
---|
4007 | |
---|
4008 | /* |
---|
4009 | * check for debug registers in system wide mode |
---|
4010 | * |
---|
4011 | * If though a check is done in pfm_context_load(), |
---|
4012 | * we must repeat it here, in case the registers are |
---|
4013 | * written after the context is loaded |
---|
4014 | */ |
---|
4015 | if (is_loaded) { |
---|
4016 | LOCK_PFS(flags); |
---|
4017 | |
---|
4018 | if (first_time && is_system) { |
---|
4019 | if (pfm_sessions.pfs_ptrace_use_dbregs) |
---|
4020 | ret = -EBUSY; |
---|
4021 | else |
---|
4022 | pfm_sessions.pfs_sys_use_dbregs++; |
---|
4023 | } |
---|
4024 | UNLOCK_PFS(flags); |
---|
4025 | } |
---|
4026 | |
---|
4027 | if (ret != 0) return ret; |
---|
4028 | |
---|
4029 | /* |
---|
4030 | * mark ourself as user of the debug registers for |
---|
4031 | * perfmon purposes. |
---|
4032 | */ |
---|
4033 | ctx->ctx_fl_using_dbreg = 1; |
---|
4034 | |
---|
4035 | /* |
---|
4036 | * clear hardware registers to make sure we don't |
---|
4037 | * pick up stale state. |
---|
4038 | * |
---|
4039 | * for a system wide session, we do not use |
---|
4040 | * thread.dbr, thread.ibr because this process |
---|
4041 | * never leaves the current CPU and the state |
---|
4042 | * is shared by all processes running on it |
---|
4043 | */ |
---|
4044 | if (first_time && can_access_pmu) { |
---|
4045 | #ifndef XEN |
---|
4046 | DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid)); |
---|
4047 | #endif |
---|
4048 | for (i=0; i < pmu_conf->num_ibrs; i++) { |
---|
4049 | ia64_set_ibr(i, 0UL); |
---|
4050 | ia64_dv_serialize_instruction(); |
---|
4051 | } |
---|
4052 | ia64_srlz_i(); |
---|
4053 | for (i=0; i < pmu_conf->num_dbrs; i++) { |
---|
4054 | ia64_set_dbr(i, 0UL); |
---|
4055 | ia64_dv_serialize_data(); |
---|
4056 | } |
---|
4057 | ia64_srlz_d(); |
---|
4058 | } |
---|
4059 | |
---|
4060 | /* |
---|
4061 | * Now install the values into the registers |
---|
4062 | */ |
---|
4063 | for (i = 0; i < count; i++, req++) { |
---|
4064 | |
---|
4065 | rnum = req->dbreg_num; |
---|
4066 | dbreg.val = req->dbreg_value; |
---|
4067 | |
---|
4068 | ret = -EINVAL; |
---|
4069 | |
---|
4070 | if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) { |
---|
4071 | DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n", |
---|
4072 | rnum, dbreg.val, mode, i, count)); |
---|
4073 | |
---|
4074 | goto abort_mission; |
---|
4075 | } |
---|
4076 | |
---|
4077 | /* |
---|
4078 | * make sure we do not install enabled breakpoint |
---|
4079 | */ |
---|
4080 | if (rnum & 0x1) { |
---|
4081 | if (mode == PFM_CODE_RR) |
---|
4082 | dbreg.ibr.ibr_x = 0; |
---|
4083 | else |
---|
4084 | dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0; |
---|
4085 | } |
---|
4086 | |
---|
4087 | PFM_REG_RETFLAG_SET(req->dbreg_flags, 0); |
---|
4088 | |
---|
4089 | /* |
---|
4090 | * Debug registers, just like PMC, can only be modified |
---|
4091 | * by a kernel call. Moreover, perfmon() access to those |
---|
4092 | * registers are centralized in this routine. The hardware |
---|
4093 | * does not modify the value of these registers, therefore, |
---|
4094 | * if we save them as they are written, we can avoid having |
---|
4095 | * to save them on context switch out. This is made possible |
---|
4096 | * by the fact that when perfmon uses debug registers, ptrace() |
---|
4097 | * won't be able to modify them concurrently. |
---|
4098 | */ |
---|
4099 | if (mode == PFM_CODE_RR) { |
---|
4100 | CTX_USED_IBR(ctx, rnum); |
---|
4101 | |
---|
4102 | if (can_access_pmu) { |
---|
4103 | ia64_set_ibr(rnum, dbreg.val); |
---|
4104 | ia64_dv_serialize_instruction(); |
---|
4105 | } |
---|
4106 | |
---|
4107 | ctx->ctx_ibrs[rnum] = dbreg.val; |
---|
4108 | |
---|
4109 | DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n", |
---|
4110 | rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu)); |
---|
4111 | } else { |
---|
4112 | CTX_USED_DBR(ctx, rnum); |
---|
4113 | |
---|
4114 | if (can_access_pmu) { |
---|
4115 | ia64_set_dbr(rnum, dbreg.val); |
---|
4116 | ia64_dv_serialize_data(); |
---|
4117 | } |
---|
4118 | ctx->ctx_dbrs[rnum] = dbreg.val; |
---|
4119 | |
---|
4120 | DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n", |
---|
4121 | rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu)); |
---|
4122 | } |
---|
4123 | } |
---|
4124 | |
---|
4125 | return 0; |
---|
4126 | |
---|
4127 | abort_mission: |
---|
4128 | /* |
---|
4129 | * in case it was our first attempt, we undo the global modifications |
---|
4130 | */ |
---|
4131 | if (first_time) { |
---|
4132 | LOCK_PFS(flags); |
---|
4133 | if (ctx->ctx_fl_system) { |
---|
4134 | pfm_sessions.pfs_sys_use_dbregs--; |
---|
4135 | } |
---|
4136 | UNLOCK_PFS(flags); |
---|
4137 | ctx->ctx_fl_using_dbreg = 0; |
---|
4138 | } |
---|
4139 | /* |
---|
4140 | * install error return flag |
---|
4141 | */ |
---|
4142 | PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL); |
---|
4143 | |
---|
4144 | return ret; |
---|
4145 | } |
---|
4146 | |
---|
4147 | static int |
---|
4148 | pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4149 | { |
---|
4150 | return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs); |
---|
4151 | } |
---|
4152 | |
---|
4153 | static int |
---|
4154 | pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4155 | { |
---|
4156 | return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs); |
---|
4157 | } |
---|
4158 | |
---|
4159 | int |
---|
4160 | pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs) |
---|
4161 | { |
---|
4162 | pfm_context_t *ctx; |
---|
4163 | |
---|
4164 | if (req == NULL) return -EINVAL; |
---|
4165 | |
---|
4166 | ctx = GET_PMU_CTX(); |
---|
4167 | |
---|
4168 | if (ctx == NULL) return -EINVAL; |
---|
4169 | |
---|
4170 | /* |
---|
4171 | * for now limit to current task, which is enough when calling |
---|
4172 | * from overflow handler |
---|
4173 | */ |
---|
4174 | if (task != current && ctx->ctx_fl_system == 0) return -EBUSY; |
---|
4175 | |
---|
4176 | return pfm_write_ibrs(ctx, req, nreq, regs); |
---|
4177 | } |
---|
4178 | EXPORT_SYMBOL(pfm_mod_write_ibrs); |
---|
4179 | |
---|
4180 | int |
---|
4181 | pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs) |
---|
4182 | { |
---|
4183 | pfm_context_t *ctx; |
---|
4184 | |
---|
4185 | if (req == NULL) return -EINVAL; |
---|
4186 | |
---|
4187 | ctx = GET_PMU_CTX(); |
---|
4188 | |
---|
4189 | if (ctx == NULL) return -EINVAL; |
---|
4190 | |
---|
4191 | /* |
---|
4192 | * for now limit to current task, which is enough when calling |
---|
4193 | * from overflow handler |
---|
4194 | */ |
---|
4195 | if (task != current && ctx->ctx_fl_system == 0) return -EBUSY; |
---|
4196 | |
---|
4197 | return pfm_write_dbrs(ctx, req, nreq, regs); |
---|
4198 | } |
---|
4199 | EXPORT_SYMBOL(pfm_mod_write_dbrs); |
---|
4200 | |
---|
4201 | |
---|
4202 | static int |
---|
4203 | pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4204 | { |
---|
4205 | pfarg_features_t *req = (pfarg_features_t *)arg; |
---|
4206 | |
---|
4207 | req->ft_version = PFM_VERSION; |
---|
4208 | return 0; |
---|
4209 | } |
---|
4210 | |
---|
4211 | #ifndef XEN |
---|
4212 | static int |
---|
4213 | pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4214 | { |
---|
4215 | struct pt_regs *tregs; |
---|
4216 | struct task_struct *task = PFM_CTX_TASK(ctx); |
---|
4217 | int state, is_system; |
---|
4218 | |
---|
4219 | state = ctx->ctx_state; |
---|
4220 | is_system = ctx->ctx_fl_system; |
---|
4221 | |
---|
4222 | /* |
---|
4223 | * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE) |
---|
4224 | */ |
---|
4225 | if (state == PFM_CTX_UNLOADED) return -EINVAL; |
---|
4226 | |
---|
4227 | /* |
---|
4228 | * In system wide and when the context is loaded, access can only happen |
---|
4229 | * when the caller is running on the CPU being monitored by the session. |
---|
4230 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
4231 | */ |
---|
4232 | if (is_system && ctx->ctx_cpu != smp_processor_id()) { |
---|
4233 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
4234 | return -EBUSY; |
---|
4235 | } |
---|
4236 | DPRINT(("task [%d] ctx_state=%d is_system=%d\n", |
---|
4237 | PFM_CTX_TASK(ctx)->pid, |
---|
4238 | state, |
---|
4239 | is_system)); |
---|
4240 | /* |
---|
4241 | * in system mode, we need to update the PMU directly |
---|
4242 | * and the user level state of the caller, which may not |
---|
4243 | * necessarily be the creator of the context. |
---|
4244 | */ |
---|
4245 | if (is_system) { |
---|
4246 | /* |
---|
4247 | * Update local PMU first |
---|
4248 | * |
---|
4249 | * disable dcr pp |
---|
4250 | */ |
---|
4251 | ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP); |
---|
4252 | ia64_srlz_i(); |
---|
4253 | |
---|
4254 | /* |
---|
4255 | * update local cpuinfo |
---|
4256 | */ |
---|
4257 | PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP); |
---|
4258 | |
---|
4259 | /* |
---|
4260 | * stop monitoring, does srlz.i |
---|
4261 | */ |
---|
4262 | pfm_clear_psr_pp(); |
---|
4263 | |
---|
4264 | /* |
---|
4265 | * stop monitoring in the caller |
---|
4266 | */ |
---|
4267 | ia64_psr(regs)->pp = 0; |
---|
4268 | |
---|
4269 | return 0; |
---|
4270 | } |
---|
4271 | /* |
---|
4272 | * per-task mode |
---|
4273 | */ |
---|
4274 | |
---|
4275 | if (task == current) { |
---|
4276 | /* stop monitoring at kernel level */ |
---|
4277 | pfm_clear_psr_up(); |
---|
4278 | |
---|
4279 | /* |
---|
4280 | * stop monitoring at the user level |
---|
4281 | */ |
---|
4282 | ia64_psr(regs)->up = 0; |
---|
4283 | } else { |
---|
4284 | tregs = task_pt_regs(task); |
---|
4285 | |
---|
4286 | /* |
---|
4287 | * stop monitoring at the user level |
---|
4288 | */ |
---|
4289 | ia64_psr(tregs)->up = 0; |
---|
4290 | |
---|
4291 | /* |
---|
4292 | * monitoring disabled in kernel at next reschedule |
---|
4293 | */ |
---|
4294 | ctx->ctx_saved_psr_up = 0; |
---|
4295 | DPRINT(("task=[%d]\n", task->pid)); |
---|
4296 | } |
---|
4297 | return 0; |
---|
4298 | } |
---|
4299 | |
---|
4300 | |
---|
4301 | static int |
---|
4302 | pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4303 | { |
---|
4304 | struct pt_regs *tregs; |
---|
4305 | int state, is_system; |
---|
4306 | |
---|
4307 | state = ctx->ctx_state; |
---|
4308 | is_system = ctx->ctx_fl_system; |
---|
4309 | |
---|
4310 | if (state != PFM_CTX_LOADED) return -EINVAL; |
---|
4311 | |
---|
4312 | /* |
---|
4313 | * In system wide and when the context is loaded, access can only happen |
---|
4314 | * when the caller is running on the CPU being monitored by the session. |
---|
4315 | * It does not have to be the owner (ctx_task) of the context per se. |
---|
4316 | */ |
---|
4317 | if (is_system && ctx->ctx_cpu != smp_processor_id()) { |
---|
4318 | DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu)); |
---|
4319 | return -EBUSY; |
---|
4320 | } |
---|
4321 | |
---|
4322 | /* |
---|
4323 | * in system mode, we need to update the PMU directly |
---|
4324 | * and the user level state of the caller, which may not |
---|
4325 | * necessarily be the creator of the context. |
---|
4326 | */ |
---|
4327 | if (is_system) { |
---|
4328 | |
---|
4329 | /* |
---|
4330 | * set user level psr.pp for the caller |
---|
4331 | */ |
---|
4332 | ia64_psr(regs)->pp = 1; |
---|
4333 | |
---|
4334 | /* |
---|
4335 | * now update the local PMU and cpuinfo |
---|
4336 | */ |
---|
4337 | PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP); |
---|
4338 | |
---|
4339 | /* |
---|
4340 | * start monitoring at kernel level |
---|
4341 | */ |
---|
4342 | pfm_set_psr_pp(); |
---|
4343 | |
---|
4344 | /* enable dcr pp */ |
---|
4345 | ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP); |
---|
4346 | ia64_srlz_i(); |
---|
4347 | |
---|
4348 | return 0; |
---|
4349 | } |
---|
4350 | |
---|
4351 | /* |
---|
4352 | * per-process mode |
---|
4353 | */ |
---|
4354 | |
---|
4355 | if (ctx->ctx_task == current) { |
---|
4356 | |
---|
4357 | /* start monitoring at kernel level */ |
---|
4358 | pfm_set_psr_up(); |
---|
4359 | |
---|
4360 | /* |
---|
4361 | * activate monitoring at user level |
---|
4362 | */ |
---|
4363 | ia64_psr(regs)->up = 1; |
---|
4364 | |
---|
4365 | } else { |
---|
4366 | tregs = task_pt_regs(ctx->ctx_task); |
---|
4367 | |
---|
4368 | /* |
---|
4369 | * start monitoring at the kernel level the next |
---|
4370 | * time the task is scheduled |
---|
4371 | */ |
---|
4372 | ctx->ctx_saved_psr_up = IA64_PSR_UP; |
---|
4373 | |
---|
4374 | /* |
---|
4375 | * activate monitoring at user level |
---|
4376 | */ |
---|
4377 | ia64_psr(tregs)->up = 1; |
---|
4378 | } |
---|
4379 | return 0; |
---|
4380 | } |
---|
4381 | |
---|
4382 | static int |
---|
4383 | pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4384 | { |
---|
4385 | pfarg_reg_t *req = (pfarg_reg_t *)arg; |
---|
4386 | unsigned int cnum; |
---|
4387 | int i; |
---|
4388 | int ret = -EINVAL; |
---|
4389 | |
---|
4390 | for (i = 0; i < count; i++, req++) { |
---|
4391 | |
---|
4392 | cnum = req->reg_num; |
---|
4393 | |
---|
4394 | if (!PMC_IS_IMPL(cnum)) goto abort_mission; |
---|
4395 | |
---|
4396 | req->reg_value = PMC_DFL_VAL(cnum); |
---|
4397 | |
---|
4398 | PFM_REG_RETFLAG_SET(req->reg_flags, 0); |
---|
4399 | |
---|
4400 | DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value)); |
---|
4401 | } |
---|
4402 | return 0; |
---|
4403 | |
---|
4404 | abort_mission: |
---|
4405 | PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL); |
---|
4406 | return ret; |
---|
4407 | } |
---|
4408 | |
---|
4409 | static int |
---|
4410 | pfm_check_task_exist(pfm_context_t *ctx) |
---|
4411 | { |
---|
4412 | struct task_struct *g, *t; |
---|
4413 | int ret = -ESRCH; |
---|
4414 | |
---|
4415 | read_lock(&tasklist_lock); |
---|
4416 | |
---|
4417 | do_each_thread (g, t) { |
---|
4418 | if (t->thread.pfm_context == ctx) { |
---|
4419 | ret = 0; |
---|
4420 | break; |
---|
4421 | } |
---|
4422 | } while_each_thread (g, t); |
---|
4423 | |
---|
4424 | read_unlock(&tasklist_lock); |
---|
4425 | |
---|
4426 | DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx)); |
---|
4427 | |
---|
4428 | return ret; |
---|
4429 | } |
---|
4430 | #endif |
---|
4431 | |
---|
4432 | static int |
---|
4433 | pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4434 | { |
---|
4435 | struct task_struct *task; |
---|
4436 | #ifndef XEN |
---|
4437 | struct thread_struct *thread; |
---|
4438 | #endif |
---|
4439 | struct pfm_context_t *old; |
---|
4440 | unsigned long flags; |
---|
4441 | #ifndef CONFIG_SMP |
---|
4442 | struct task_struct *owner_task = NULL; |
---|
4443 | #endif |
---|
4444 | pfarg_load_t *req = (pfarg_load_t *)arg; |
---|
4445 | unsigned long *pmcs_source, *pmds_source; |
---|
4446 | int the_cpu; |
---|
4447 | int ret = 0; |
---|
4448 | int state, is_system, set_dbregs = 0; |
---|
4449 | |
---|
4450 | state = ctx->ctx_state; |
---|
4451 | is_system = ctx->ctx_fl_system; |
---|
4452 | #ifdef XEN |
---|
4453 | task = NULL; |
---|
4454 | old = NULL; |
---|
4455 | pmcs_source = pmds_source = NULL; |
---|
4456 | #ifndef CONFIG_SMP |
---|
4457 | owner_task = NULL; |
---|
4458 | #endif |
---|
4459 | flags = 0; |
---|
4460 | BUG_ON(count != 0); |
---|
4461 | BUG_ON(regs != NULL); |
---|
4462 | #endif |
---|
4463 | /* |
---|
4464 | * can only load from unloaded or terminated state |
---|
4465 | */ |
---|
4466 | if (state != PFM_CTX_UNLOADED) { |
---|
4467 | DPRINT(("cannot load to [%d], invalid ctx_state=%d\n", |
---|
4468 | req->load_pid, |
---|
4469 | ctx->ctx_state)); |
---|
4470 | return -EBUSY; |
---|
4471 | } |
---|
4472 | |
---|
4473 | #ifndef XEN |
---|
4474 | DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg)); |
---|
4475 | |
---|
4476 | if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) { |
---|
4477 | DPRINT(("cannot use blocking mode on self\n")); |
---|
4478 | return -EINVAL; |
---|
4479 | } |
---|
4480 | |
---|
4481 | ret = pfm_get_task(ctx, req->load_pid, &task); |
---|
4482 | if (ret) { |
---|
4483 | DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret)); |
---|
4484 | return ret; |
---|
4485 | } |
---|
4486 | |
---|
4487 | ret = -EINVAL; |
---|
4488 | |
---|
4489 | /* |
---|
4490 | * system wide is self monitoring only |
---|
4491 | */ |
---|
4492 | if (is_system && task != current) { |
---|
4493 | DPRINT(("system wide is self monitoring only load_pid=%d\n", |
---|
4494 | req->load_pid)); |
---|
4495 | goto error; |
---|
4496 | } |
---|
4497 | |
---|
4498 | thread = &task->thread; |
---|
4499 | #else |
---|
4500 | BUG_ON(!spin_is_locked(&ctx->ctx_lock)); |
---|
4501 | if (!is_system) { |
---|
4502 | ret = -EINVAL; |
---|
4503 | goto error; |
---|
4504 | } |
---|
4505 | #endif |
---|
4506 | |
---|
4507 | ret = 0; |
---|
4508 | #ifndef XEN |
---|
4509 | /* |
---|
4510 | * cannot load a context which is using range restrictions, |
---|
4511 | * into a task that is being debugged. |
---|
4512 | */ |
---|
4513 | if (ctx->ctx_fl_using_dbreg) { |
---|
4514 | if (thread->flags & IA64_THREAD_DBG_VALID) { |
---|
4515 | ret = -EBUSY; |
---|
4516 | DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid)); |
---|
4517 | goto error; |
---|
4518 | } |
---|
4519 | LOCK_PFS(flags); |
---|
4520 | |
---|
4521 | if (is_system) { |
---|
4522 | if (pfm_sessions.pfs_ptrace_use_dbregs) { |
---|
4523 | DPRINT(("cannot load [%d] dbregs in use\n", task->pid)); |
---|
4524 | ret = -EBUSY; |
---|
4525 | } else { |
---|
4526 | pfm_sessions.pfs_sys_use_dbregs++; |
---|
4527 | DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task->pid, pfm_sessions.pfs_sys_use_dbregs)); |
---|
4528 | set_dbregs = 1; |
---|
4529 | } |
---|
4530 | } |
---|
4531 | |
---|
4532 | UNLOCK_PFS(flags); |
---|
4533 | |
---|
4534 | if (ret) goto error; |
---|
4535 | } |
---|
4536 | #else |
---|
4537 | BUG_ON(ctx->ctx_fl_using_dbreg); |
---|
4538 | #endif |
---|
4539 | |
---|
4540 | /* |
---|
4541 | * SMP system-wide monitoring implies self-monitoring. |
---|
4542 | * |
---|
4543 | * The programming model expects the task to |
---|
4544 | * be pinned on a CPU throughout the session. |
---|
4545 | * Here we take note of the current CPU at the |
---|
4546 | * time the context is loaded. No call from |
---|
4547 | * another CPU will be allowed. |
---|
4548 | * |
---|
4549 | * The pinning via shed_setaffinity() |
---|
4550 | * must be done by the calling task prior |
---|
4551 | * to this call. |
---|
4552 | * |
---|
4553 | * systemwide: keep track of CPU this session is supposed to run on |
---|
4554 | */ |
---|
4555 | the_cpu = ctx->ctx_cpu = smp_processor_id(); |
---|
4556 | |
---|
4557 | ret = -EBUSY; |
---|
4558 | /* |
---|
4559 | * now reserve the session |
---|
4560 | */ |
---|
4561 | ret = pfm_reserve_session(current, is_system, the_cpu); |
---|
4562 | if (ret) goto error; |
---|
4563 | |
---|
4564 | /* |
---|
4565 | * task is necessarily stopped at this point. |
---|
4566 | * |
---|
4567 | * If the previous context was zombie, then it got removed in |
---|
4568 | * pfm_save_regs(). Therefore we should not see it here. |
---|
4569 | * If we see a context, then this is an active context |
---|
4570 | * |
---|
4571 | * XXX: needs to be atomic |
---|
4572 | */ |
---|
4573 | #ifndef XEN |
---|
4574 | DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n", |
---|
4575 | thread->pfm_context, ctx)); |
---|
4576 | |
---|
4577 | ret = -EBUSY; |
---|
4578 | old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *)); |
---|
4579 | if (old != NULL) { |
---|
4580 | DPRINT(("load_pid [%d] already has a context\n", req->load_pid)); |
---|
4581 | goto error_unres; |
---|
4582 | } |
---|
4583 | |
---|
4584 | pfm_reset_msgq(ctx); |
---|
4585 | #endif |
---|
4586 | |
---|
4587 | ctx->ctx_state = PFM_CTX_LOADED; |
---|
4588 | |
---|
4589 | /* |
---|
4590 | * link context to task |
---|
4591 | */ |
---|
4592 | ctx->ctx_task = task; |
---|
4593 | |
---|
4594 | if (is_system) { |
---|
4595 | /* |
---|
4596 | * we load as stopped |
---|
4597 | */ |
---|
4598 | PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE); |
---|
4599 | PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP); |
---|
4600 | |
---|
4601 | if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE); |
---|
4602 | } else { |
---|
4603 | #ifndef XEN |
---|
4604 | thread->flags |= IA64_THREAD_PM_VALID; |
---|
4605 | #else |
---|
4606 | BUG(); |
---|
4607 | #endif |
---|
4608 | } |
---|
4609 | |
---|
4610 | #ifndef XEN |
---|
4611 | /* |
---|
4612 | * propagate into thread-state |
---|
4613 | */ |
---|
4614 | pfm_copy_pmds(task, ctx); |
---|
4615 | pfm_copy_pmcs(task, ctx); |
---|
4616 | |
---|
4617 | pmcs_source = thread->pmcs; |
---|
4618 | pmds_source = thread->pmds; |
---|
4619 | |
---|
4620 | /* |
---|
4621 | * always the case for system-wide |
---|
4622 | */ |
---|
4623 | if (task == current) { |
---|
4624 | |
---|
4625 | if (is_system == 0) { |
---|
4626 | |
---|
4627 | /* allow user level control */ |
---|
4628 | ia64_psr(regs)->sp = 0; |
---|
4629 | DPRINT(("clearing psr.sp for [%d]\n", task->pid)); |
---|
4630 | |
---|
4631 | SET_LAST_CPU(ctx, smp_processor_id()); |
---|
4632 | INC_ACTIVATION(); |
---|
4633 | SET_ACTIVATION(ctx); |
---|
4634 | #ifndef CONFIG_SMP |
---|
4635 | /* |
---|
4636 | * push the other task out, if any |
---|
4637 | */ |
---|
4638 | owner_task = GET_PMU_OWNER(); |
---|
4639 | if (owner_task) pfm_lazy_save_regs(owner_task); |
---|
4640 | #endif |
---|
4641 | } |
---|
4642 | /* |
---|
4643 | * load all PMD from ctx to PMU (as opposed to thread state) |
---|
4644 | * restore all PMC from ctx to PMU |
---|
4645 | */ |
---|
4646 | pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]); |
---|
4647 | pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]); |
---|
4648 | |
---|
4649 | ctx->ctx_reload_pmcs[0] = 0UL; |
---|
4650 | ctx->ctx_reload_pmds[0] = 0UL; |
---|
4651 | |
---|
4652 | /* |
---|
4653 | * guaranteed safe by earlier check against DBG_VALID |
---|
4654 | */ |
---|
4655 | if (ctx->ctx_fl_using_dbreg) { |
---|
4656 | pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs); |
---|
4657 | pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs); |
---|
4658 | } |
---|
4659 | /* |
---|
4660 | * set new ownership |
---|
4661 | */ |
---|
4662 | SET_PMU_OWNER(task, ctx); |
---|
4663 | |
---|
4664 | DPRINT(("context loaded on PMU for [%d]\n", task->pid)); |
---|
4665 | } else { |
---|
4666 | /* |
---|
4667 | * when not current, task MUST be stopped, so this is safe |
---|
4668 | */ |
---|
4669 | regs = task_pt_regs(task); |
---|
4670 | |
---|
4671 | /* force a full reload */ |
---|
4672 | ctx->ctx_last_activation = PFM_INVALID_ACTIVATION; |
---|
4673 | SET_LAST_CPU(ctx, -1); |
---|
4674 | |
---|
4675 | /* initial saved psr (stopped) */ |
---|
4676 | ctx->ctx_saved_psr_up = 0UL; |
---|
4677 | ia64_psr(regs)->up = ia64_psr(regs)->pp = 0; |
---|
4678 | } |
---|
4679 | #else |
---|
4680 | BUG_ON(!is_system); |
---|
4681 | |
---|
4682 | /* load pmds, pmcs */ |
---|
4683 | xenpfm_restore_pmds(ctx); |
---|
4684 | xenpfm_restore_pmcs(ctx); |
---|
4685 | |
---|
4686 | ctx->ctx_reload_pmcs[0] = 0UL; |
---|
4687 | ctx->ctx_reload_pmds[0] = 0UL; |
---|
4688 | |
---|
4689 | BUG_ON(ctx->ctx_fl_using_dbreg); |
---|
4690 | |
---|
4691 | SET_PMU_OWNER(NULL, ctx); |
---|
4692 | #endif |
---|
4693 | |
---|
4694 | ret = 0; |
---|
4695 | |
---|
4696 | #ifndef XEN |
---|
4697 | error_unres: |
---|
4698 | if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu); |
---|
4699 | #endif |
---|
4700 | error: |
---|
4701 | #ifndef XEN |
---|
4702 | /* |
---|
4703 | * we must undo the dbregs setting (for system-wide) |
---|
4704 | */ |
---|
4705 | if (ret && set_dbregs) { |
---|
4706 | LOCK_PFS(flags); |
---|
4707 | pfm_sessions.pfs_sys_use_dbregs--; |
---|
4708 | UNLOCK_PFS(flags); |
---|
4709 | } |
---|
4710 | /* |
---|
4711 | * release task, there is now a link with the context |
---|
4712 | */ |
---|
4713 | if (is_system == 0 && task != current) { |
---|
4714 | pfm_put_task(task); |
---|
4715 | |
---|
4716 | if (ret == 0) { |
---|
4717 | ret = pfm_check_task_exist(ctx); |
---|
4718 | if (ret) { |
---|
4719 | ctx->ctx_state = PFM_CTX_UNLOADED; |
---|
4720 | ctx->ctx_task = NULL; |
---|
4721 | } |
---|
4722 | } |
---|
4723 | } |
---|
4724 | #else |
---|
4725 | BUG_ON(set_dbregs); |
---|
4726 | #endif |
---|
4727 | return ret; |
---|
4728 | } |
---|
4729 | |
---|
4730 | /* |
---|
4731 | * in this function, we do not need to increase the use count |
---|
4732 | * for the task via get_task_struct(), because we hold the |
---|
4733 | * context lock. If the task were to disappear while having |
---|
4734 | * a context attached, it would go through pfm_exit_thread() |
---|
4735 | * which also grabs the context lock and would therefore be blocked |
---|
4736 | * until we are here. |
---|
4737 | */ |
---|
4738 | static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx); |
---|
4739 | |
---|
4740 | static int |
---|
4741 | pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs) |
---|
4742 | { |
---|
4743 | struct task_struct *task = PFM_CTX_TASK(ctx); |
---|
4744 | struct pt_regs *tregs; |
---|
4745 | int prev_state, is_system; |
---|
4746 | int ret; |
---|
4747 | |
---|
4748 | #ifndef XEN |
---|
4749 | DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task->pid : -1)); |
---|
4750 | #else |
---|
4751 | task = NULL; |
---|
4752 | tregs = NULL; |
---|
4753 | BUG_ON(arg != NULL); |
---|
4754 | BUG_ON(count != 0); |
---|
4755 | BUG_ON(regs != NULL); |
---|
4756 | #endif |
---|
4757 | |
---|
4758 | prev_state = ctx->ctx_state; |
---|
4759 | is_system = ctx->ctx_fl_system; |
---|
4760 | |
---|
4761 | /* |
---|
4762 | * unload only when necessary |
---|
4763 | */ |
---|
4764 | if (prev_state == PFM_CTX_UNLOADED) { |
---|
4765 | DPRINT(("ctx_state=%d, nothing to do\n", prev_state)); |
---|
4766 | return 0; |
---|
4767 | } |
---|
4768 | |
---|
4769 | /* |
---|
4770 | * clear psr and dcr bits |
---|
4771 | */ |
---|
4772 | #ifndef XEN |
---|
4773 | ret = pfm_stop(ctx, NULL, 0, regs); |
---|
4774 | if (ret) return ret; |
---|
4775 | #else |
---|
4776 | /* caller does it by hand */ |
---|
4777 | ret = 0; |
---|
4778 | #endif |
---|
4779 | |
---|
4780 | ctx->ctx_state = PFM_CTX_UNLOADED; |
---|
4781 | |
---|
4782 | /* |
---|
4783 | * in system mode, we need to update the PMU directly |
---|
4784 | * and the user level state of the caller, which may not |
---|
4785 | * necessarily be the creator of the context. |
---|
4786 | */ |
---|
4787 | if (is_system) { |
---|
4788 | |
---|
4789 | /* |
---|
4790 | * Update cpuinfo |
---|
4791 | * |
---|
4792 | * local PMU is taken care of in pfm_stop() |
---|
4793 | */ |
---|
4794 | PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE); |
---|
4795 | PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE); |
---|
4796 | |
---|
4797 | /* |
---|
4798 | * save PMDs in context |
---|
4799 | * release ownership |
---|
4800 | */ |
---|
4801 | pfm_flush_pmds(current, ctx); |
---|
4802 | |
---|
4803 | /* |
---|
4804 | * at this point we are done with the PMU |
---|
4805 | * so we can unreserve the resource. |
---|
4806 | */ |
---|
4807 | if (prev_state != PFM_CTX_ZOMBIE) |
---|
4808 | pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu); |
---|
4809 | |
---|
4810 | #ifndef XEN |
---|
4811 | /* |
---|
4812 | * disconnect context from task |
---|
4813 | */ |
---|
4814 | task->thread.pfm_context = NULL; |
---|
4815 | #endif |
---|
4816 | /* |
---|
4817 | * disconnect task from context |
---|
4818 | */ |
---|
4819 | ctx->ctx_task = NULL; |
---|
4820 | |
---|
4821 | /* |
---|
4822 | * There is nothing more to cleanup here. |
---|
4823 | */ |
---|
4824 | return 0; |
---|
4825 | } |
---|
4826 | |
---|
4827 | #ifndef XEN |
---|
4828 | /* |
---|
4829 | * per-task mode |
---|
4830 | */ |
---|
4831 | tregs = task == current ? regs : task_pt_regs(task); |
---|
4832 | |
---|
4833 | if (task == current) { |
---|
4834 | /* |
---|
4835 | * cancel user level control |
---|
4836 | */ |
---|
4837 | ia64_psr(regs)->sp = 1; |
---|
4838 | |
---|
4839 | DPRINT(("setting psr.sp for [%d]\n", task->pid)); |
---|
4840 | } |
---|
4841 | /* |
---|
4842 | * save PMDs to context |
---|
4843 | * release ownership |
---|
4844 | */ |
---|
4845 | pfm_flush_pmds(task, ctx); |
---|
4846 | |
---|
4847 | /* |
---|
4848 | * at this point we are done with the PMU |
---|
4849 | * so we can unreserve the resource. |
---|
4850 | * |
---|
4851 | * when state was ZOMBIE, we have already unreserved. |
---|
4852 | */ |
---|
4853 | if (prev_state != PFM_CTX_ZOMBIE) |
---|
4854 | pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu); |
---|
4855 | |
---|
4856 | /* |
---|
4857 | * reset activation counter and psr |
---|
4858 | */ |
---|
4859 | ctx->ctx_last_activation = PFM_INVALID_ACTIVATION; |
---|
4860 | SET_LAST_CPU(ctx, -1); |
---|
4861 | |
---|
4862 | /* |
---|
4863 | * PMU state will not be restored |
---|
4864 | */ |
---|
4865 | task->thread.flags &= ~IA64_THREAD_PM_VALID; |
---|
4866 | |
---|
4867 | /* |
---|
4868 | * break links between context and task |
---|
4869 | */ |
---|
4870 | task->thread.pfm_context = NULL; |
---|
4871 | ctx->ctx_task = NULL; |
---|
4872 | |
---|
4873 | PFM_SET_WORK_PENDING(task, 0); |
---|
4874 | |
---|
4875 | ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE; |
---|
4876 | ctx->ctx_fl_can_restart = 0; |
---|
4877 | ctx->ctx_fl_going_zombie = 0; |
---|
4878 | |
---|
4879 | DPRINT(("disconnected [%d] from context\n", task->pid)); |
---|
4880 | |
---|
4881 | return 0; |
---|
4882 | #else |
---|
4883 | BUG(); |
---|
4884 | return -EINVAL; |
---|
4885 | #endif |
---|
4886 | } |
---|
4887 | |
---|
4888 | |
---|
4889 | #ifndef XEN |
---|
4890 | /* |
---|
4891 | * called only from exit_thread(): task == current |
---|
4892 | * we come here only if current has a context attached (loaded or masked) |
---|
4893 | */ |
---|
4894 | void |
---|
4895 | pfm_exit_thread(struct task_struct *task) |
---|
4896 | { |
---|
4897 | pfm_context_t *ctx; |
---|
4898 | unsigned long flags; |
---|
4899 | struct pt_regs *regs = task_pt_regs(task); |
---|
4900 | int ret, state; |
---|
4901 | int free_ok = 0; |
---|
4902 | |
---|
4903 | ctx = PFM_GET_CTX(task); |
---|
4904 | |
---|
4905 | PROTECT_CTX(ctx, flags); |
---|
4906 | |
---|
4907 | DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task->pid)); |
---|
4908 | |
---|
4909 | state = ctx->ctx_state; |
---|
4910 | switch(state) { |
---|
4911 | case PFM_CTX_UNLOADED: |
---|
4912 | /* |
---|
4913 | * only comes to thios function if pfm_context is not NULL, i.e., cannot |
---|
4914 | * be in unloaded state |
---|
4915 | */ |
---|
4916 | printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task->pid); |
---|
4917 | break; |
---|
4918 | case PFM_CTX_LOADED: |
---|
4919 | case PFM_CTX_MASKED: |
---|
4920 | ret = pfm_context_unload(ctx, NULL, 0, regs); |
---|
4921 | if (ret) { |
---|
4922 | printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret); |
---|
4923 | } |
---|
4924 | DPRINT(("ctx unloaded for current state was %d\n", state)); |
---|
4925 | |
---|
4926 | pfm_end_notify_user(ctx); |
---|
4927 | break; |
---|
4928 | case PFM_CTX_ZOMBIE: |
---|
4929 | ret = pfm_context_unload(ctx, NULL, 0, regs); |
---|
4930 | if (ret) { |
---|
4931 | printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task->pid, state, ret); |
---|
4932 | } |
---|
4933 | free_ok = 1; |
---|
4934 | break; |
---|
4935 | default: |
---|
4936 | printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task->pid, state); |
---|
4937 | break; |
---|
4938 | } |
---|
4939 | UNPROTECT_CTX(ctx, flags); |
---|
4940 | |
---|
4941 | { u64 psr = pfm_get_psr(); |
---|
4942 | BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP)); |
---|
4943 | BUG_ON(GET_PMU_OWNER()); |
---|
4944 | BUG_ON(ia64_psr(regs)->up); |
---|
4945 | BUG_ON(ia64_psr(regs)->pp); |
---|
4946 | } |
---|
4947 | |
---|
4948 | /* |
---|
4949 | * All memory free operations (especially for vmalloc'ed memory) |
---|
4950 | * MUST be done with interrupts ENABLED. |
---|
4951 | */ |
---|
4952 | if (free_ok) pfm_context_free(ctx); |
---|
4953 | } |
---|
4954 | |
---|
4955 | /* |
---|
4956 | * functions MUST be listed in the increasing order of their index (see permfon.h) |
---|
4957 | */ |
---|
4958 | #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz } |
---|
4959 | #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL } |
---|
4960 | #define PFM_CMD_PCLRWS (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP) |
---|
4961 | #define PFM_CMD_PCLRW (PFM_CMD_FD|PFM_CMD_ARG_RW) |
---|
4962 | #define PFM_CMD_NONE { NULL, "no-cmd", 0, 0, 0, NULL} |
---|
4963 | |
---|
4964 | static pfm_cmd_desc_t pfm_cmd_tab[]={ |
---|
4965 | /* 0 */PFM_CMD_NONE, |
---|
4966 | /* 1 */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL), |
---|
4967 | /* 2 */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL), |
---|
4968 | /* 3 */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL), |
---|
4969 | /* 4 */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS), |
---|
4970 | /* 5 */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS), |
---|
4971 | /* 6 */PFM_CMD_NONE, |
---|
4972 | /* 7 */PFM_CMD_NONE, |
---|
4973 | /* 8 */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize), |
---|
4974 | /* 9 */PFM_CMD_NONE, |
---|
4975 | /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW), |
---|
4976 | /* 11 */PFM_CMD_NONE, |
---|
4977 | /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL), |
---|
4978 | /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL), |
---|
4979 | /* 14 */PFM_CMD_NONE, |
---|
4980 | /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL), |
---|
4981 | /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL), |
---|
4982 | /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS), |
---|
4983 | /* 18 */PFM_CMD_NONE, |
---|
4984 | /* 19 */PFM_CMD_NONE, |
---|
4985 | /* 20 */PFM_CMD_NONE, |
---|
4986 | /* 21 */PFM_CMD_NONE, |
---|
4987 | /* 22 */PFM_CMD_NONE, |
---|
4988 | /* 23 */PFM_CMD_NONE, |
---|
4989 | /* 24 */PFM_CMD_NONE, |
---|
4990 | /* 25 */PFM_CMD_NONE, |
---|
4991 | /* 26 */PFM_CMD_NONE, |
---|
4992 | /* 27 */PFM_CMD_NONE, |
---|
4993 | /* 28 */PFM_CMD_NONE, |
---|
4994 | /* 29 */PFM_CMD_NONE, |
---|
4995 | /* 30 */PFM_CMD_NONE, |
---|
4996 | /* 31 */PFM_CMD_NONE, |
---|
4997 | /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL), |
---|
4998 | /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL) |
---|
4999 | }; |
---|
5000 | #define PFM_CMD_COUNT (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t)) |
---|
5001 | |
---|
5002 | static int |
---|
5003 | pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags) |
---|
5004 | { |
---|
5005 | struct task_struct *task; |
---|
5006 | int state, old_state; |
---|
5007 | |
---|
5008 | recheck: |
---|
5009 | state = ctx->ctx_state; |
---|
5010 | task = ctx->ctx_task; |
---|
5011 | |
---|
5012 | if (task == NULL) { |
---|
5013 | DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state)); |
---|
5014 | return 0; |
---|
5015 | } |
---|
5016 | |
---|
5017 | DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n", |
---|
5018 | ctx->ctx_fd, |
---|
5019 | state, |
---|
5020 | task->pid, |
---|
5021 | task->state, PFM_CMD_STOPPED(cmd))); |
---|
5022 | |
---|
5023 | /* |
---|
5024 | * self-monitoring always ok. |
---|
5025 | * |
---|
5026 | * for system-wide the caller can either be the creator of the |
---|
5027 | * context (to one to which the context is attached to) OR |
---|
5028 | * a task running on the same CPU as the session. |
---|
5029 | */ |
---|
5030 | if (task == current || ctx->ctx_fl_system) return 0; |
---|
5031 | |
---|
5032 | /* |
---|
5033 | * we are monitoring another thread |
---|
5034 | */ |
---|
5035 | switch(state) { |
---|
5036 | case PFM_CTX_UNLOADED: |
---|
5037 | /* |
---|
5038 | * if context is UNLOADED we are safe to go |
---|
5039 | */ |
---|
5040 | return 0; |
---|
5041 | case PFM_CTX_ZOMBIE: |
---|
5042 | /* |
---|
5043 | * no command can operate on a zombie context |
---|
5044 | */ |
---|
5045 | DPRINT(("cmd %d state zombie cannot operate on context\n", cmd)); |
---|
5046 | return -EINVAL; |
---|
5047 | case PFM_CTX_MASKED: |
---|
5048 | /* |
---|
5049 | * PMU state has been saved to software even though |
---|
5050 | * the thread may still be running. |
---|
5051 | */ |
---|
5052 | if (cmd != PFM_UNLOAD_CONTEXT) return 0; |
---|
5053 | } |
---|
5054 | |
---|
5055 | /* |
---|
5056 | * context is LOADED or MASKED. Some commands may need to have |
---|
5057 | * the task stopped. |
---|
5058 | * |
---|
5059 | * We could lift this restriction for UP but it would mean that |
---|
5060 | * the user has no guarantee the task would not run between |
---|
5061 | * two successive calls to perfmonctl(). That's probably OK. |
---|
5062 | * If this user wants to ensure the task does not run, then |
---|
5063 | * the task must be stopped. |
---|
5064 | */ |
---|
5065 | if (PFM_CMD_STOPPED(cmd)) { |
---|
5066 | if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) { |
---|
5067 | DPRINT(("[%d] task not in stopped state\n", task->pid)); |
---|
5068 | return -EBUSY; |
---|
5069 | } |
---|
5070 | /* |
---|
5071 | * task is now stopped, wait for ctxsw out |
---|
5072 | * |
---|
5073 | * This is an interesting point in the code. |
---|
5074 | * We need to unprotect the context because |
---|
5075 | * the pfm_save_regs() routines needs to grab |
---|
5076 | * the same lock. There are danger in doing |
---|
5077 | * this because it leaves a window open for |
---|
5078 | * another task to get access to the context |
---|
5079 | * and possibly change its state. The one thing |
---|
5080 | * that is not possible is for the context to disappear |
---|
5081 | * because we are protected by the VFS layer, i.e., |
---|
5082 | * get_fd()/put_fd(). |
---|
5083 | */ |
---|
5084 | old_state = state; |
---|
5085 | |
---|
5086 | UNPROTECT_CTX(ctx, flags); |
---|
5087 | |
---|
5088 | wait_task_inactive(task); |
---|
5089 | |
---|
5090 | PROTECT_CTX(ctx, flags); |
---|
5091 | |
---|
5092 | /* |
---|
5093 | * we must recheck to verify if state has changed |
---|
5094 | */ |
---|
5095 | if (ctx->ctx_state != old_state) { |
---|
5096 | DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state)); |
---|
5097 | goto recheck; |
---|
5098 | } |
---|
5099 | } |
---|
5100 | return 0; |
---|
5101 | } |
---|
5102 | |
---|
5103 | /* |
---|
5104 | * system-call entry point (must return long) |
---|
5105 | */ |
---|
5106 | asmlinkage long |
---|
5107 | sys_perfmonctl (int fd, int cmd, void __user *arg, int count) |
---|
5108 | { |
---|
5109 | struct file *file = NULL; |
---|
5110 | pfm_context_t *ctx = NULL; |
---|
5111 | unsigned long flags = 0UL; |
---|
5112 | void *args_k = NULL; |
---|
5113 | long ret; /* will expand int return types */ |
---|
5114 | size_t base_sz, sz, xtra_sz = 0; |
---|
5115 | int narg, completed_args = 0, call_made = 0, cmd_flags; |
---|
5116 | int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs); |
---|
5117 | int (*getsize)(void *arg, size_t *sz); |
---|
5118 | #define PFM_MAX_ARGSIZE 4096 |
---|
5119 | |
---|
5120 | /* |
---|
5121 | * reject any call if perfmon was disabled at initialization |
---|
5122 | */ |
---|
5123 | if (unlikely(pmu_conf == NULL)) return -ENOSYS; |
---|
5124 | |
---|
5125 | if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) { |
---|
5126 | DPRINT(("invalid cmd=%d\n", cmd)); |
---|
5127 | return -EINVAL; |
---|
5128 | } |
---|
5129 | |
---|
5130 | func = pfm_cmd_tab[cmd].cmd_func; |
---|
5131 | narg = pfm_cmd_tab[cmd].cmd_narg; |
---|
5132 | base_sz = pfm_cmd_tab[cmd].cmd_argsize; |
---|
5133 | getsize = pfm_cmd_tab[cmd].cmd_getsize; |
---|
5134 | cmd_flags = pfm_cmd_tab[cmd].cmd_flags; |
---|
5135 | |
---|
5136 | if (unlikely(func == NULL)) { |
---|
5137 | DPRINT(("invalid cmd=%d\n", cmd)); |
---|
5138 | return -EINVAL; |
---|
5139 | } |
---|
5140 | |
---|
5141 | DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n", |
---|
5142 | PFM_CMD_NAME(cmd), |
---|
5143 | cmd, |
---|
5144 | narg, |
---|
5145 | base_sz, |
---|
5146 | count)); |
---|
5147 | |
---|
5148 | /* |
---|
5149 | * check if number of arguments matches what the command expects |
---|
5150 | */ |
---|
5151 | if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count))) |
---|
5152 | return -EINVAL; |
---|
5153 | |
---|
5154 | restart_args: |
---|
5155 | sz = xtra_sz + base_sz*count; |
---|
5156 | /* |
---|
5157 | * limit abuse to min page size |
---|
5158 | */ |
---|
5159 | if (unlikely(sz > PFM_MAX_ARGSIZE)) { |
---|
5160 | printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", current->pid, sz); |
---|
5161 | return -E2BIG; |
---|
5162 | } |
---|
5163 | |
---|
5164 | /* |
---|
5165 | * allocate default-sized argument buffer |
---|
5166 | */ |
---|
5167 | if (likely(count && args_k == NULL)) { |
---|
5168 | args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL); |
---|
5169 | if (args_k == NULL) return -ENOMEM; |
---|
5170 | } |
---|
5171 | |
---|
5172 | ret = -EFAULT; |
---|
5173 | |
---|
5174 | /* |
---|
5175 | * copy arguments |
---|
5176 | * |
---|
5177 | * assume sz = 0 for command without parameters |
---|
5178 | */ |
---|
5179 | if (sz && copy_from_user(args_k, arg, sz)) { |
---|
5180 | DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg)); |
---|
5181 | goto error_args; |
---|
5182 | } |
---|
5183 | |
---|
5184 | /* |
---|
5185 | * check if command supports extra parameters |
---|
5186 | */ |
---|
5187 | if (completed_args == 0 && getsize) { |
---|
5188 | /* |
---|
5189 | * get extra parameters size (based on main argument) |
---|
5190 | */ |
---|
5191 | ret = (*getsize)(args_k, &xtra_sz); |
---|
5192 | if (ret) goto error_args; |
---|
5193 | |
---|
5194 | completed_args = 1; |
---|
5195 | |
---|
5196 | DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz)); |
---|
5197 | |
---|
5198 | /* retry if necessary */ |
---|
5199 | if (likely(xtra_sz)) goto restart_args; |
---|
5200 | } |
---|
5201 | |
---|
5202 | if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd; |
---|
5203 | |
---|
5204 | ret = -EBADF; |
---|
5205 | |
---|
5206 | file = fget(fd); |
---|
5207 | if (unlikely(file == NULL)) { |
---|
5208 | DPRINT(("invalid fd %d\n", fd)); |
---|
5209 | goto error_args; |
---|
5210 | } |
---|
5211 | if (unlikely(PFM_IS_FILE(file) == 0)) { |
---|
5212 | DPRINT(("fd %d not related to perfmon\n", fd)); |
---|
5213 | goto error_args; |
---|
5214 | } |
---|
5215 | |
---|
5216 | ctx = (pfm_context_t *)file->private_data; |
---|
5217 | if (unlikely(ctx == NULL)) { |
---|
5218 | DPRINT(("no context for fd %d\n", fd)); |
---|
5219 | goto error_args; |
---|
5220 | } |
---|
5221 | prefetch(&ctx->ctx_state); |
---|
5222 | |
---|
5223 | PROTECT_CTX(ctx, flags); |
---|
5224 | |
---|
5225 | /* |
---|
5226 | * check task is stopped |
---|
5227 | */ |
---|
5228 | ret = pfm_check_task_state(ctx, cmd, flags); |
---|
5229 | if (unlikely(ret)) goto abort_locked; |
---|
5230 | |
---|
5231 | skip_fd: |
---|
5232 | ret = (*func)(ctx, args_k, count, task_pt_regs(current)); |
---|
5233 | |
---|
5234 | call_made = 1; |
---|
5235 | |
---|
5236 | abort_locked: |
---|
5237 | if (likely(ctx)) { |
---|
5238 | DPRINT(("context unlocked\n")); |
---|
5239 | UNPROTECT_CTX(ctx, flags); |
---|
5240 | } |
---|
5241 | |
---|
5242 | /* copy argument back to user, if needed */ |
---|
5243 | if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT; |
---|
5244 | |
---|
5245 | error_args: |
---|
5246 | if (file) |
---|
5247 | fput(file); |
---|
5248 | |
---|
5249 | kfree(args_k); |
---|
5250 | |
---|
5251 | DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret)); |
---|
5252 | |
---|
5253 | return ret; |
---|
5254 | } |
---|
5255 | |
---|
5256 | static void |
---|
5257 | pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs) |
---|
5258 | { |
---|
5259 | pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt; |
---|
5260 | pfm_ovfl_ctrl_t rst_ctrl; |
---|
5261 | int state; |
---|
5262 | int ret = 0; |
---|
5263 | |
---|
5264 | state = ctx->ctx_state; |
---|
5265 | /* |
---|
5266 | * Unlock sampling buffer and reset index atomically |
---|
5267 | * XXX: not really needed when blocking |
---|
5268 | */ |
---|
5269 | if (CTX_HAS_SMPL(ctx)) { |
---|
5270 | |
---|
5271 | rst_ctrl.bits.mask_monitoring = 0; |
---|
5272 | rst_ctrl.bits.reset_ovfl_pmds = 0; |
---|
5273 | |
---|
5274 | if (state == PFM_CTX_LOADED) |
---|
5275 | ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs); |
---|
5276 | else |
---|
5277 | ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs); |
---|
5278 | } else { |
---|
5279 | rst_ctrl.bits.mask_monitoring = 0; |
---|
5280 | rst_ctrl.bits.reset_ovfl_pmds = 1; |
---|
5281 | } |
---|
5282 | |
---|
5283 | if (ret == 0) { |
---|
5284 | if (rst_ctrl.bits.reset_ovfl_pmds) { |
---|
5285 | pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET); |
---|
5286 | } |
---|
5287 | if (rst_ctrl.bits.mask_monitoring == 0) { |
---|
5288 | DPRINT(("resuming monitoring\n")); |
---|
5289 | if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current); |
---|
5290 | } else { |
---|
5291 | DPRINT(("stopping monitoring\n")); |
---|
5292 | //pfm_stop_monitoring(current, regs); |
---|
5293 | } |
---|
5294 | ctx->ctx_state = PFM_CTX_LOADED; |
---|
5295 | } |
---|
5296 | } |
---|
5297 | |
---|
5298 | /* |
---|
5299 | * context MUST BE LOCKED when calling |
---|
5300 | * can only be called for current |
---|
5301 | */ |
---|
5302 | static void |
---|
5303 | pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs) |
---|
5304 | { |
---|
5305 | int ret; |
---|
5306 | |
---|
5307 | DPRINT(("entering for [%d]\n", current->pid)); |
---|
5308 | |
---|
5309 | ret = pfm_context_unload(ctx, NULL, 0, regs); |
---|
5310 | if (ret) { |
---|
5311 | printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", current->pid, ret); |
---|
5312 | } |
---|
5313 | |
---|
5314 | /* |
---|
5315 | * and wakeup controlling task, indicating we are now disconnected |
---|
5316 | */ |
---|
5317 | wake_up_interruptible(&ctx->ctx_zombieq); |
---|
5318 | |
---|
5319 | /* |
---|
5320 | * given that context is still locked, the controlling |
---|
5321 | * task will only get access when we return from |
---|
5322 | * pfm_handle_work(). |
---|
5323 | */ |
---|
5324 | } |
---|
5325 | |
---|
5326 | static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds); |
---|
5327 | /* |
---|
5328 | * pfm_handle_work() can be called with interrupts enabled |
---|
5329 | * (TIF_NEED_RESCHED) or disabled. The down_interruptible |
---|
5330 | * call may sleep, therefore we must re-enable interrupts |
---|
5331 | * to avoid deadlocks. It is safe to do so because this function |
---|
5332 | * is called ONLY when returning to user level (PUStk=1), in which case |
---|
5333 | * there is no risk of kernel stack overflow due to deep |
---|
5334 | * interrupt nesting. |
---|
5335 | */ |
---|
5336 | void |
---|
5337 | pfm_handle_work(void) |
---|
5338 | { |
---|
5339 | pfm_context_t *ctx; |
---|
5340 | struct pt_regs *regs; |
---|
5341 | unsigned long flags, dummy_flags; |
---|
5342 | unsigned long ovfl_regs; |
---|
5343 | unsigned int reason; |
---|
5344 | int ret; |
---|
5345 | |
---|
5346 | ctx = PFM_GET_CTX(current); |
---|
5347 | if (ctx == NULL) { |
---|
5348 | printk(KERN_ERR "perfmon: [%d] has no PFM context\n", current->pid); |
---|
5349 | return; |
---|
5350 | } |
---|
5351 | |
---|
5352 | PROTECT_CTX(ctx, flags); |
---|
5353 | |
---|
5354 | PFM_SET_WORK_PENDING(current, 0); |
---|
5355 | |
---|
5356 | pfm_clear_task_notify(); |
---|
5357 | |
---|
5358 | regs = task_pt_regs(current); |
---|
5359 | |
---|
5360 | /* |
---|
5361 | * extract reason for being here and clear |
---|
5362 | */ |
---|
5363 | reason = ctx->ctx_fl_trap_reason; |
---|
5364 | ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE; |
---|
5365 | ovfl_regs = ctx->ctx_ovfl_regs[0]; |
---|
5366 | |
---|
5367 | DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state)); |
---|
5368 | |
---|
5369 | /* |
---|
5370 | * must be done before we check for simple-reset mode |
---|
5371 | */ |
---|
5372 | if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE) goto do_zombie; |
---|
5373 | |
---|
5374 | |
---|
5375 | //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking; |
---|
5376 | if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking; |
---|
5377 | |
---|
5378 | /* |
---|
5379 | * restore interrupt mask to what it was on entry. |
---|
5380 | * Could be enabled/diasbled. |
---|
5381 | */ |
---|
5382 | UNPROTECT_CTX(ctx, flags); |
---|
5383 | |
---|
5384 | /* |
---|
5385 | * force interrupt enable because of down_interruptible() |
---|
5386 | */ |
---|
5387 | local_irq_enable(); |
---|
5388 | |
---|
5389 | DPRINT(("before block sleeping\n")); |
---|
5390 | |
---|
5391 | /* |
---|
5392 | * may go through without blocking on SMP systems |
---|
5393 | * if restart has been received already by the time we call down() |
---|
5394 | */ |
---|
5395 | ret = wait_for_completion_interruptible(&ctx->ctx_restart_done); |
---|
5396 | |
---|
5397 | DPRINT(("after block sleeping ret=%d\n", ret)); |
---|
5398 | |
---|
5399 | /* |
---|
5400 | * lock context and mask interrupts again |
---|
5401 | * We save flags into a dummy because we may have |
---|
5402 | * altered interrupts mask compared to entry in this |
---|
5403 | * function. |
---|
5404 | */ |
---|
5405 | PROTECT_CTX(ctx, dummy_flags); |
---|
5406 | |
---|
5407 | /* |
---|
5408 | * we need to read the ovfl_regs only after wake-up |
---|
5409 | * because we may have had pfm_write_pmds() in between |
---|
5410 | * and that can changed PMD values and therefore |
---|
5411 | * ovfl_regs is reset for these new PMD values. |
---|
5412 | */ |
---|
5413 | ovfl_regs = ctx->ctx_ovfl_regs[0]; |
---|
5414 | |
---|
5415 | if (ctx->ctx_fl_going_zombie) { |
---|
5416 | do_zombie: |
---|
5417 | DPRINT(("context is zombie, bailing out\n")); |
---|
5418 | pfm_context_force_terminate(ctx, regs); |
---|
5419 | goto nothing_to_do; |
---|
5420 | } |
---|
5421 | /* |
---|
5422 | * in case of interruption of down() we don't restart anything |
---|
5423 | */ |
---|
5424 | if (ret < 0) goto nothing_to_do; |
---|
5425 | |
---|
5426 | skip_blocking: |
---|
5427 | pfm_resume_after_ovfl(ctx, ovfl_regs, regs); |
---|
5428 | ctx->ctx_ovfl_regs[0] = 0UL; |
---|
5429 | |
---|
5430 | nothing_to_do: |
---|
5431 | /* |
---|
5432 | * restore flags as they were upon entry |
---|
5433 | */ |
---|
5434 | UNPROTECT_CTX(ctx, flags); |
---|
5435 | } |
---|
5436 | |
---|
5437 | static int |
---|
5438 | pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg) |
---|
5439 | { |
---|
5440 | if (ctx->ctx_state == PFM_CTX_ZOMBIE) { |
---|
5441 | DPRINT(("ignoring overflow notification, owner is zombie\n")); |
---|
5442 | return 0; |
---|
5443 | } |
---|
5444 | |
---|
5445 | DPRINT(("waking up somebody\n")); |
---|
5446 | |
---|
5447 | if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait); |
---|
5448 | |
---|
5449 | /* |
---|
5450 | * safe, we are not in intr handler, nor in ctxsw when |
---|
5451 | * we come here |
---|
5452 | */ |
---|
5453 | kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN); |
---|
5454 | |
---|
5455 | return 0; |
---|
5456 | } |
---|
5457 | |
---|
5458 | static int |
---|
5459 | pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds) |
---|
5460 | { |
---|
5461 | pfm_msg_t *msg = NULL; |
---|
5462 | |
---|
5463 | if (ctx->ctx_fl_no_msg == 0) { |
---|
5464 | msg = pfm_get_new_msg(ctx); |
---|
5465 | if (msg == NULL) { |
---|
5466 | printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n"); |
---|
5467 | return -1; |
---|
5468 | } |
---|
5469 | |
---|
5470 | msg->pfm_ovfl_msg.msg_type = PFM_MSG_OVFL; |
---|
5471 | msg->pfm_ovfl_msg.msg_ctx_fd = ctx->ctx_fd; |
---|
5472 | msg->pfm_ovfl_msg.msg_active_set = 0; |
---|
5473 | msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds; |
---|
5474 | msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL; |
---|
5475 | msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL; |
---|
5476 | msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL; |
---|
5477 | msg->pfm_ovfl_msg.msg_tstamp = 0UL; |
---|
5478 | } |
---|
5479 | |
---|
5480 | DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n", |
---|
5481 | msg, |
---|
5482 | ctx->ctx_fl_no_msg, |
---|
5483 | ctx->ctx_fd, |
---|
5484 | ovfl_pmds)); |
---|
5485 | |
---|
5486 | return pfm_notify_user(ctx, msg); |
---|
5487 | } |
---|
5488 | |
---|
5489 | static int |
---|
5490 | pfm_end_notify_user(pfm_context_t *ctx) |
---|
5491 | { |
---|
5492 | pfm_msg_t *msg; |
---|
5493 | |
---|
5494 | msg = pfm_get_new_msg(ctx); |
---|
5495 | if (msg == NULL) { |
---|
5496 | printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n"); |
---|
5497 | return -1; |
---|
5498 | } |
---|
5499 | /* no leak */ |
---|
5500 | memset(msg, 0, sizeof(*msg)); |
---|
5501 | |
---|
5502 | msg->pfm_end_msg.msg_type = PFM_MSG_END; |
---|
5503 | msg->pfm_end_msg.msg_ctx_fd = ctx->ctx_fd; |
---|
5504 | msg->pfm_ovfl_msg.msg_tstamp = 0UL; |
---|
5505 | |
---|
5506 | DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n", |
---|
5507 | msg, |
---|
5508 | ctx->ctx_fl_no_msg, |
---|
5509 | ctx->ctx_fd)); |
---|
5510 | |
---|
5511 | return pfm_notify_user(ctx, msg); |
---|
5512 | } |
---|
5513 | #else |
---|
5514 | #define pfm_ovfl_notify_user(ctx, ovfl_pmds) do {} while(0) |
---|
5515 | #endif |
---|
5516 | |
---|
5517 | /* |
---|
5518 | * main overflow processing routine. |
---|
5519 | * it can be called from the interrupt path or explicitely during the context switch code |
---|
5520 | */ |
---|
5521 | static void |
---|
5522 | pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx, u64 pmc0, struct pt_regs *regs) |
---|
5523 | { |
---|
5524 | pfm_ovfl_arg_t *ovfl_arg; |
---|
5525 | unsigned long mask; |
---|
5526 | unsigned long old_val, ovfl_val, new_val; |
---|
5527 | unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds; |
---|
5528 | unsigned long tstamp; |
---|
5529 | pfm_ovfl_ctrl_t ovfl_ctrl; |
---|
5530 | unsigned int i, has_smpl; |
---|
5531 | int must_notify = 0; |
---|
5532 | #ifdef XEN |
---|
5533 | BUG_ON(task != NULL); |
---|
5534 | #endif |
---|
5535 | |
---|
5536 | if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring; |
---|
5537 | |
---|
5538 | /* |
---|
5539 | * sanity test. Should never happen |
---|
5540 | */ |
---|
5541 | if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check; |
---|
5542 | |
---|
5543 | tstamp = ia64_get_itc(); |
---|
5544 | mask = pmc0 >> PMU_FIRST_COUNTER; |
---|
5545 | ovfl_val = pmu_conf->ovfl_val; |
---|
5546 | has_smpl = CTX_HAS_SMPL(ctx); |
---|
5547 | |
---|
5548 | DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s " |
---|
5549 | "used_pmds=0x%lx\n", |
---|
5550 | pmc0, |
---|
5551 | task ? task->pid: -1, |
---|
5552 | (regs ? regs->cr_iip : 0), |
---|
5553 | CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking", |
---|
5554 | ctx->ctx_used_pmds[0])); |
---|
5555 | |
---|
5556 | |
---|
5557 | /* |
---|
5558 | * first we update the virtual counters |
---|
5559 | * assume there was a prior ia64_srlz_d() issued |
---|
5560 | */ |
---|
5561 | for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) { |
---|
5562 | |
---|
5563 | /* skip pmd which did not overflow */ |
---|
5564 | if ((mask & 0x1) == 0) continue; |
---|
5565 | |
---|
5566 | /* |
---|
5567 | * Note that the pmd is not necessarily 0 at this point as qualified events |
---|
5568 | * may have happened before the PMU was frozen. The residual count is not |
---|
5569 | * taken into consideration here but will be with any read of the pmd via |
---|
5570 | * pfm_read_pmds(). |
---|
5571 | */ |
---|
5572 | old_val = new_val = ctx->ctx_pmds[i].val; |
---|
5573 | new_val += 1 + ovfl_val; |
---|
5574 | ctx->ctx_pmds[i].val = new_val; |
---|
5575 | |
---|
5576 | /* |
---|
5577 | * check for overflow condition |
---|
5578 | */ |
---|
5579 | if (likely(old_val > new_val)) { |
---|
5580 | ovfl_pmds |= 1UL << i; |
---|
5581 | if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i; |
---|
5582 | } |
---|
5583 | |
---|
5584 | DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n", |
---|
5585 | i, |
---|
5586 | new_val, |
---|
5587 | old_val, |
---|
5588 | ia64_get_pmd(i) & ovfl_val, |
---|
5589 | ovfl_pmds, |
---|
5590 | ovfl_notify)); |
---|
5591 | } |
---|
5592 | |
---|
5593 | /* |
---|
5594 | * there was no 64-bit overflow, nothing else to do |
---|
5595 | */ |
---|
5596 | if (ovfl_pmds == 0UL) return; |
---|
5597 | |
---|
5598 | /* |
---|
5599 | * reset all control bits |
---|
5600 | */ |
---|
5601 | ovfl_ctrl.val = 0; |
---|
5602 | reset_pmds = 0UL; |
---|
5603 | |
---|
5604 | /* |
---|
5605 | * if a sampling format module exists, then we "cache" the overflow by |
---|
5606 | * calling the module's handler() routine. |
---|
5607 | */ |
---|
5608 | if (has_smpl) { |
---|
5609 | unsigned long start_cycles, end_cycles; |
---|
5610 | unsigned long pmd_mask; |
---|
5611 | int j, k, ret = 0; |
---|
5612 | int this_cpu = smp_processor_id(); |
---|
5613 | |
---|
5614 | pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER; |
---|
5615 | ovfl_arg = &ctx->ctx_ovfl_arg; |
---|
5616 | |
---|
5617 | prefetch(ctx->ctx_smpl_hdr); |
---|
5618 | |
---|
5619 | for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) { |
---|
5620 | |
---|
5621 | mask = 1UL << i; |
---|
5622 | |
---|
5623 | if ((pmd_mask & 0x1) == 0) continue; |
---|
5624 | |
---|
5625 | ovfl_arg->ovfl_pmd = (unsigned char )i; |
---|
5626 | ovfl_arg->ovfl_notify = ovfl_notify & mask ? 1 : 0; |
---|
5627 | ovfl_arg->active_set = 0; |
---|
5628 | ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */ |
---|
5629 | ovfl_arg->smpl_pmds[0] = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0]; |
---|
5630 | |
---|
5631 | ovfl_arg->pmd_value = ctx->ctx_pmds[i].val; |
---|
5632 | ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval; |
---|
5633 | ovfl_arg->pmd_eventid = ctx->ctx_pmds[i].eventid; |
---|
5634 | |
---|
5635 | /* |
---|
5636 | * copy values of pmds of interest. Sampling format may copy them |
---|
5637 | * into sampling buffer. |
---|
5638 | */ |
---|
5639 | if (smpl_pmds) { |
---|
5640 | for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) { |
---|
5641 | if ((smpl_pmds & 0x1) == 0) continue; |
---|
5642 | ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ? pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j); |
---|
5643 | DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1])); |
---|
5644 | } |
---|
5645 | } |
---|
5646 | |
---|
5647 | pfm_stats[this_cpu].pfm_smpl_handler_calls++; |
---|
5648 | |
---|
5649 | start_cycles = ia64_get_itc(); |
---|
5650 | |
---|
5651 | /* |
---|
5652 | * call custom buffer format record (handler) routine |
---|
5653 | */ |
---|
5654 | ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp); |
---|
5655 | |
---|
5656 | end_cycles = ia64_get_itc(); |
---|
5657 | |
---|
5658 | /* |
---|
5659 | * For those controls, we take the union because they have |
---|
5660 | * an all or nothing behavior. |
---|
5661 | */ |
---|
5662 | ovfl_ctrl.bits.notify_user |= ovfl_arg->ovfl_ctrl.bits.notify_user; |
---|
5663 | ovfl_ctrl.bits.block_task |= ovfl_arg->ovfl_ctrl.bits.block_task; |
---|
5664 | ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring; |
---|
5665 | /* |
---|
5666 | * build the bitmask of pmds to reset now |
---|
5667 | */ |
---|
5668 | if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask; |
---|
5669 | |
---|
5670 | pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles; |
---|
5671 | } |
---|
5672 | /* |
---|
5673 | * when the module cannot handle the rest of the overflows, we abort right here |
---|
5674 | */ |
---|
5675 | if (ret && pmd_mask) { |
---|
5676 | DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n", |
---|
5677 | pmd_mask<<PMU_FIRST_COUNTER)); |
---|
5678 | } |
---|
5679 | /* |
---|
5680 | * remove the pmds we reset now from the set of pmds to reset in pfm_restart() |
---|
5681 | */ |
---|
5682 | ovfl_pmds &= ~reset_pmds; |
---|
5683 | } else { |
---|
5684 | /* |
---|
5685 | * when no sampling module is used, then the default |
---|
5686 | * is to notify on overflow if requested by user |
---|
5687 | */ |
---|
5688 | ovfl_ctrl.bits.notify_user = ovfl_notify ? 1 : 0; |
---|
5689 | ovfl_ctrl.bits.block_task = ovfl_notify ? 1 : 0; |
---|
5690 | ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */ |
---|
5691 | ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1; |
---|
5692 | /* |
---|
5693 | * if needed, we reset all overflowed pmds |
---|
5694 | */ |
---|
5695 | if (ovfl_notify == 0) reset_pmds = ovfl_pmds; |
---|
5696 | } |
---|
5697 | |
---|
5698 | DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds)); |
---|
5699 | |
---|
5700 | /* |
---|
5701 | * reset the requested PMD registers using the short reset values |
---|
5702 | */ |
---|
5703 | if (reset_pmds) { |
---|
5704 | unsigned long bm = reset_pmds; |
---|
5705 | pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET); |
---|
5706 | } |
---|
5707 | |
---|
5708 | if (ovfl_notify && ovfl_ctrl.bits.notify_user) { |
---|
5709 | #ifndef XEN |
---|
5710 | /* |
---|
5711 | * keep track of what to reset when unblocking |
---|
5712 | */ |
---|
5713 | ctx->ctx_ovfl_regs[0] = ovfl_pmds; |
---|
5714 | |
---|
5715 | /* |
---|
5716 | * check for blocking context |
---|
5717 | */ |
---|
5718 | if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) { |
---|
5719 | |
---|
5720 | ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK; |
---|
5721 | |
---|
5722 | /* |
---|
5723 | * set the perfmon specific checking pending work for the task |
---|
5724 | */ |
---|
5725 | PFM_SET_WORK_PENDING(task, 1); |
---|
5726 | |
---|
5727 | /* |
---|
5728 | * when coming from ctxsw, current still points to the |
---|
5729 | * previous task, therefore we must work with task and not current. |
---|
5730 | */ |
---|
5731 | pfm_set_task_notify(task); |
---|
5732 | } |
---|
5733 | /* |
---|
5734 | * defer until state is changed (shorten spin window). the context is locked |
---|
5735 | * anyway, so the signal receiver would come spin for nothing. |
---|
5736 | */ |
---|
5737 | must_notify = 1; |
---|
5738 | #else |
---|
5739 | gdprintk(XENLOG_INFO, "%s check!\n", __func__); |
---|
5740 | #endif |
---|
5741 | } |
---|
5742 | |
---|
5743 | DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n", |
---|
5744 | #ifndef XEN |
---|
5745 | GET_PMU_OWNER() ? GET_PMU_OWNER()->pid : -1, |
---|
5746 | PFM_GET_WORK_PENDING(task), |
---|
5747 | #else |
---|
5748 | -1, 0UL, |
---|
5749 | #endif |
---|
5750 | ctx->ctx_fl_trap_reason, |
---|
5751 | ovfl_pmds, |
---|
5752 | ovfl_notify, |
---|
5753 | ovfl_ctrl.bits.mask_monitoring ? 1 : 0)); |
---|
5754 | /* |
---|
5755 | * in case monitoring must be stopped, we toggle the psr bits |
---|
5756 | */ |
---|
5757 | if (ovfl_ctrl.bits.mask_monitoring) { |
---|
5758 | #ifndef XEN |
---|
5759 | pfm_mask_monitoring(task); |
---|
5760 | ctx->ctx_state = PFM_CTX_MASKED; |
---|
5761 | ctx->ctx_fl_can_restart = 1; |
---|
5762 | #else |
---|
5763 | gdprintk(XENLOG_INFO, "%s check!\n", __func__); |
---|
5764 | #endif |
---|
5765 | } |
---|
5766 | |
---|
5767 | /* |
---|
5768 | * send notification now |
---|
5769 | */ |
---|
5770 | if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify); |
---|
5771 | |
---|
5772 | return; |
---|
5773 | |
---|
5774 | sanity_check: |
---|
5775 | printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n", |
---|
5776 | smp_processor_id(), |
---|
5777 | task ? task->pid : -1, |
---|
5778 | pmc0); |
---|
5779 | return; |
---|
5780 | |
---|
5781 | stop_monitoring: |
---|
5782 | /* |
---|
5783 | * in SMP, zombie context is never restored but reclaimed in pfm_load_regs(). |
---|
5784 | * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can |
---|
5785 | * come here as zombie only if the task is the current task. In which case, we |
---|
5786 | * can access the PMU hardware directly. |
---|
5787 | * |
---|
5788 | * Note that zombies do have PM_VALID set. So here we do the minimal. |
---|
5789 | * |
---|
5790 | * In case the context was zombified it could not be reclaimed at the time |
---|
5791 | * the monitoring program exited. At this point, the PMU reservation has been |
---|
5792 | * returned, the sampiing buffer has been freed. We must convert this call |
---|
5793 | * into a spurious interrupt. However, we must also avoid infinite overflows |
---|
5794 | * by stopping monitoring for this task. We can only come here for a per-task |
---|
5795 | * context. All we need to do is to stop monitoring using the psr bits which |
---|
5796 | * are always task private. By re-enabling secure montioring, we ensure that |
---|
5797 | * the monitored task will not be able to re-activate monitoring. |
---|
5798 | * The task will eventually be context switched out, at which point the context |
---|
5799 | * will be reclaimed (that includes releasing ownership of the PMU). |
---|
5800 | * |
---|
5801 | * So there might be a window of time where the number of per-task session is zero |
---|
5802 | * yet one PMU might have a owner and get at most one overflow interrupt for a zombie |
---|
5803 | * context. This is safe because if a per-task session comes in, it will push this one |
---|
5804 | * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide |
---|
5805 | * session is force on that CPU, given that we use task pinning, pfm_save_regs() will |
---|
5806 | * also push our zombie context out. |
---|
5807 | * |
---|
5808 | * Overall pretty hairy stuff.... |
---|
5809 | */ |
---|
5810 | DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task->pid: -1)); |
---|
5811 | pfm_clear_psr_up(); |
---|
5812 | ia64_psr(regs)->up = 0; |
---|
5813 | ia64_psr(regs)->sp = 1; |
---|
5814 | return; |
---|
5815 | } |
---|
5816 | |
---|
5817 | static int |
---|
5818 | pfm_do_interrupt_handler(int irq, void *arg, struct pt_regs *regs) |
---|
5819 | { |
---|
5820 | struct task_struct *task; |
---|
5821 | pfm_context_t *ctx; |
---|
5822 | unsigned long flags; |
---|
5823 | u64 pmc0; |
---|
5824 | int this_cpu = smp_processor_id(); |
---|
5825 | int retval = 0; |
---|
5826 | |
---|
5827 | pfm_stats[this_cpu].pfm_ovfl_intr_count++; |
---|
5828 | |
---|
5829 | /* |
---|
5830 | * srlz.d done before arriving here |
---|
5831 | */ |
---|
5832 | pmc0 = ia64_get_pmc(0); |
---|
5833 | |
---|
5834 | #ifndef XEN |
---|
5835 | task = GET_PMU_OWNER(); |
---|
5836 | #else |
---|
5837 | task = NULL; |
---|
5838 | #endif |
---|
5839 | ctx = GET_PMU_CTX(); |
---|
5840 | |
---|
5841 | /* |
---|
5842 | * if we have some pending bits set |
---|
5843 | * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1 |
---|
5844 | */ |
---|
5845 | #ifndef XEN |
---|
5846 | if (PMC0_HAS_OVFL(pmc0) && task) { |
---|
5847 | #else |
---|
5848 | if (PMC0_HAS_OVFL(pmc0)) { |
---|
5849 | #endif |
---|
5850 | /* |
---|
5851 | * we assume that pmc0.fr is always set here |
---|
5852 | */ |
---|
5853 | |
---|
5854 | /* sanity check */ |
---|
5855 | if (!ctx) goto report_spurious1; |
---|
5856 | |
---|
5857 | #ifndef XEN |
---|
5858 | if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) |
---|
5859 | goto report_spurious2; |
---|
5860 | #endif |
---|
5861 | |
---|
5862 | PROTECT_CTX_NOPRINT(ctx, flags); |
---|
5863 | |
---|
5864 | pfm_overflow_handler(task, ctx, pmc0, regs); |
---|
5865 | |
---|
5866 | UNPROTECT_CTX_NOPRINT(ctx, flags); |
---|
5867 | |
---|
5868 | } else { |
---|
5869 | pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++; |
---|
5870 | retval = -1; |
---|
5871 | } |
---|
5872 | /* |
---|
5873 | * keep it unfrozen at all times |
---|
5874 | */ |
---|
5875 | pfm_unfreeze_pmu(); |
---|
5876 | |
---|
5877 | return retval; |
---|
5878 | |
---|
5879 | report_spurious1: |
---|
5880 | #ifndef XEN |
---|
5881 | printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n", |
---|
5882 | this_cpu, task->pid); |
---|
5883 | #endif |
---|
5884 | pfm_unfreeze_pmu(); |
---|
5885 | return -1; |
---|
5886 | #ifndef XEN /* XEN path doesn't take this goto */ |
---|
5887 | report_spurious2: |
---|
5888 | printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", |
---|
5889 | this_cpu, |
---|
5890 | task->pid); |
---|
5891 | pfm_unfreeze_pmu(); |
---|
5892 | return -1; |
---|
5893 | #endif |
---|
5894 | } |
---|
5895 | |
---|
5896 | static irqreturn_t |
---|
5897 | pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs) |
---|
5898 | { |
---|
5899 | unsigned long start_cycles, total_cycles; |
---|
5900 | unsigned long min, max; |
---|
5901 | int this_cpu; |
---|
5902 | int ret; |
---|
5903 | |
---|
5904 | this_cpu = get_cpu(); |
---|
5905 | if (likely(!pfm_alt_intr_handler)) { |
---|
5906 | min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min; |
---|
5907 | max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max; |
---|
5908 | |
---|
5909 | start_cycles = ia64_get_itc(); |
---|
5910 | |
---|
5911 | ret = pfm_do_interrupt_handler(irq, arg, regs); |
---|
5912 | |
---|
5913 | total_cycles = ia64_get_itc(); |
---|
5914 | |
---|
5915 | /* |
---|
5916 | * don't measure spurious interrupts |
---|
5917 | */ |
---|
5918 | if (likely(ret == 0)) { |
---|
5919 | total_cycles -= start_cycles; |
---|
5920 | |
---|
5921 | if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles; |
---|
5922 | if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles; |
---|
5923 | |
---|
5924 | pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles; |
---|
5925 | } |
---|
5926 | } |
---|
5927 | else { |
---|
5928 | (*pfm_alt_intr_handler->handler)(irq, arg, regs); |
---|
5929 | } |
---|
5930 | |
---|
5931 | put_cpu_no_resched(); |
---|
5932 | return IRQ_HANDLED; |
---|
5933 | } |
---|
5934 | |
---|
5935 | #ifndef XEN |
---|
5936 | /* |
---|
5937 | * /proc/perfmon interface, for debug only |
---|
5938 | */ |
---|
5939 | |
---|
5940 | #define PFM_PROC_SHOW_HEADER ((void *)NR_CPUS+1) |
---|
5941 | |
---|
5942 | static void * |
---|
5943 | pfm_proc_start(struct seq_file *m, loff_t *pos) |
---|
5944 | { |
---|
5945 | if (*pos == 0) { |
---|
5946 | return PFM_PROC_SHOW_HEADER; |
---|
5947 | } |
---|
5948 | |
---|
5949 | while (*pos <= NR_CPUS) { |
---|
5950 | if (cpu_online(*pos - 1)) { |
---|
5951 | return (void *)*pos; |
---|
5952 | } |
---|
5953 | ++*pos; |
---|
5954 | } |
---|
5955 | return NULL; |
---|
5956 | } |
---|
5957 | |
---|
5958 | static void * |
---|
5959 | pfm_proc_next(struct seq_file *m, void *v, loff_t *pos) |
---|
5960 | { |
---|
5961 | ++*pos; |
---|
5962 | return pfm_proc_start(m, pos); |
---|
5963 | } |
---|
5964 | |
---|
5965 | static void |
---|
5966 | pfm_proc_stop(struct seq_file *m, void *v) |
---|
5967 | { |
---|
5968 | } |
---|
5969 | |
---|
5970 | static void |
---|
5971 | pfm_proc_show_header(struct seq_file *m) |
---|
5972 | { |
---|
5973 | struct list_head * pos; |
---|
5974 | pfm_buffer_fmt_t * entry; |
---|
5975 | unsigned long flags; |
---|
5976 | |
---|
5977 | seq_printf(m, |
---|
5978 | "perfmon version : %u.%u\n" |
---|
5979 | "model : %s\n" |
---|
5980 | "fastctxsw : %s\n" |
---|
5981 | "expert mode : %s\n" |
---|
5982 | "ovfl_mask : 0x%lx\n" |
---|
5983 | "PMU flags : 0x%x\n", |
---|
5984 | PFM_VERSION_MAJ, PFM_VERSION_MIN, |
---|
5985 | pmu_conf->pmu_name, |
---|
5986 | pfm_sysctl.fastctxsw > 0 ? "Yes": "No", |
---|
5987 | pfm_sysctl.expert_mode > 0 ? "Yes": "No", |
---|
5988 | pmu_conf->ovfl_val, |
---|
5989 | pmu_conf->flags); |
---|
5990 | |
---|
5991 | LOCK_PFS(flags); |
---|
5992 | |
---|
5993 | seq_printf(m, |
---|
5994 | "proc_sessions : %u\n" |
---|
5995 | "sys_sessions : %u\n" |
---|
5996 | "sys_use_dbregs : %u\n" |
---|
5997 | "ptrace_use_dbregs : %u\n", |
---|
5998 | pfm_sessions.pfs_task_sessions, |
---|
5999 | pfm_sessions.pfs_sys_sessions, |
---|
6000 | pfm_sessions.pfs_sys_use_dbregs, |
---|
6001 | pfm_sessions.pfs_ptrace_use_dbregs); |
---|
6002 | |
---|
6003 | UNLOCK_PFS(flags); |
---|
6004 | |
---|
6005 | spin_lock(&pfm_buffer_fmt_lock); |
---|
6006 | |
---|
6007 | list_for_each(pos, &pfm_buffer_fmt_list) { |
---|
6008 | entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list); |
---|
6009 | seq_printf(m, "format : %02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x-%02x %s\n", |
---|
6010 | entry->fmt_uuid[0], |
---|
6011 | entry->fmt_uuid[1], |
---|
6012 | entry->fmt_uuid[2], |
---|
6013 | entry->fmt_uuid[3], |
---|
6014 | entry->fmt_uuid[4], |
---|
6015 | entry->fmt_uuid[5], |
---|
6016 | entry->fmt_uuid[6], |
---|
6017 | entry->fmt_uuid[7], |
---|
6018 | entry->fmt_uuid[8], |
---|
6019 | entry->fmt_uuid[9], |
---|
6020 | entry->fmt_uuid[10], |
---|
6021 | entry->fmt_uuid[11], |
---|
6022 | entry->fmt_uuid[12], |
---|
6023 | entry->fmt_uuid[13], |
---|
6024 | entry->fmt_uuid[14], |
---|
6025 | entry->fmt_uuid[15], |
---|
6026 | entry->fmt_name); |
---|
6027 | } |
---|
6028 | spin_unlock(&pfm_buffer_fmt_lock); |
---|
6029 | |
---|
6030 | } |
---|
6031 | |
---|
6032 | static int |
---|
6033 | pfm_proc_show(struct seq_file *m, void *v) |
---|
6034 | { |
---|
6035 | unsigned long psr; |
---|
6036 | unsigned int i; |
---|
6037 | int cpu; |
---|
6038 | |
---|
6039 | if (v == PFM_PROC_SHOW_HEADER) { |
---|
6040 | pfm_proc_show_header(m); |
---|
6041 | return 0; |
---|
6042 | } |
---|
6043 | |
---|
6044 | /* show info for CPU (v - 1) */ |
---|
6045 | |
---|
6046 | cpu = (long)v - 1; |
---|
6047 | seq_printf(m, |
---|
6048 | "CPU%-2d overflow intrs : %lu\n" |
---|
6049 | "CPU%-2d overflow cycles : %lu\n" |
---|
6050 | "CPU%-2d overflow min : %lu\n" |
---|
6051 | "CPU%-2d overflow max : %lu\n" |
---|
6052 | "CPU%-2d smpl handler calls : %lu\n" |
---|
6053 | "CPU%-2d smpl handler cycles : %lu\n" |
---|
6054 | "CPU%-2d spurious intrs : %lu\n" |
---|
6055 | "CPU%-2d replay intrs : %lu\n" |
---|
6056 | "CPU%-2d syst_wide : %d\n" |
---|
6057 | "CPU%-2d dcr_pp : %d\n" |
---|
6058 | "CPU%-2d exclude idle : %d\n" |
---|
6059 | "CPU%-2d owner : %d\n" |
---|
6060 | "CPU%-2d context : %p\n" |
---|
6061 | "CPU%-2d activations : %lu\n", |
---|
6062 | cpu, pfm_stats[cpu].pfm_ovfl_intr_count, |
---|
6063 | cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles, |
---|
6064 | cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min, |
---|
6065 | cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max, |
---|
6066 | cpu, pfm_stats[cpu].pfm_smpl_handler_calls, |
---|
6067 | cpu, pfm_stats[cpu].pfm_smpl_handler_cycles, |
---|
6068 | cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count, |
---|
6069 | cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count, |
---|
6070 | cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0, |
---|
6071 | cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0, |
---|
6072 | cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0, |
---|
6073 | cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1, |
---|
6074 | cpu, pfm_get_cpu_data(pmu_ctx, cpu), |
---|
6075 | cpu, pfm_get_cpu_data(pmu_activation_number, cpu)); |
---|
6076 | |
---|
6077 | if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) { |
---|
6078 | |
---|
6079 | psr = pfm_get_psr(); |
---|
6080 | |
---|
6081 | ia64_srlz_d(); |
---|
6082 | |
---|
6083 | seq_printf(m, |
---|
6084 | "CPU%-2d psr : 0x%lx\n" |
---|
6085 | "CPU%-2d pmc0 : 0x%lx\n", |
---|
6086 | cpu, psr, |
---|
6087 | cpu, ia64_get_pmc(0)); |
---|
6088 | |
---|
6089 | for (i=0; PMC_IS_LAST(i) == 0; i++) { |
---|
6090 | if (PMC_IS_COUNTING(i) == 0) continue; |
---|
6091 | seq_printf(m, |
---|
6092 | "CPU%-2d pmc%u : 0x%lx\n" |
---|
6093 | "CPU%-2d pmd%u : 0x%lx\n", |
---|
6094 | cpu, i, ia64_get_pmc(i), |
---|
6095 | cpu, i, ia64_get_pmd(i)); |
---|
6096 | } |
---|
6097 | } |
---|
6098 | return 0; |
---|
6099 | } |
---|
6100 | |
---|
6101 | struct seq_operations pfm_seq_ops = { |
---|
6102 | .start = pfm_proc_start, |
---|
6103 | .next = pfm_proc_next, |
---|
6104 | .stop = pfm_proc_stop, |
---|
6105 | .show = pfm_proc_show |
---|
6106 | }; |
---|
6107 | |
---|
6108 | static int |
---|
6109 | pfm_proc_open(struct inode *inode, struct file *file) |
---|
6110 | { |
---|
6111 | return seq_open(file, &pfm_seq_ops); |
---|
6112 | } |
---|
6113 | #endif |
---|
6114 | |
---|
6115 | |
---|
6116 | /* |
---|
6117 | * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens |
---|
6118 | * during pfm_enable() hence before pfm_start(). We cannot assume monitoring |
---|
6119 | * is active or inactive based on mode. We must rely on the value in |
---|
6120 | * local_cpu_data->pfm_syst_info |
---|
6121 | */ |
---|
6122 | void |
---|
6123 | pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin) |
---|
6124 | { |
---|
6125 | struct pt_regs *regs; |
---|
6126 | unsigned long dcr; |
---|
6127 | unsigned long dcr_pp; |
---|
6128 | |
---|
6129 | dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0; |
---|
6130 | |
---|
6131 | /* |
---|
6132 | * pid 0 is guaranteed to be the idle task. There is one such task with pid 0 |
---|
6133 | * on every CPU, so we can rely on the pid to identify the idle task. |
---|
6134 | */ |
---|
6135 | if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) { |
---|
6136 | regs = task_pt_regs(task); |
---|
6137 | ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0; |
---|
6138 | return; |
---|
6139 | } |
---|
6140 | /* |
---|
6141 | * if monitoring has started |
---|
6142 | */ |
---|
6143 | if (dcr_pp) { |
---|
6144 | dcr = ia64_getreg(_IA64_REG_CR_DCR); |
---|
6145 | /* |
---|
6146 | * context switching in? |
---|
6147 | */ |
---|
6148 | if (is_ctxswin) { |
---|
6149 | /* mask monitoring for the idle task */ |
---|
6150 | ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP); |
---|
6151 | pfm_clear_psr_pp(); |
---|
6152 | ia64_srlz_i(); |
---|
6153 | return; |
---|
6154 | } |
---|
6155 | /* |
---|
6156 | * context switching out |
---|
6157 | * restore monitoring for next task |
---|
6158 | * |
---|
6159 | * Due to inlining this odd if-then-else construction generates |
---|
6160 | * better code. |
---|
6161 | */ |
---|
6162 | ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP); |
---|
6163 | pfm_set_psr_pp(); |
---|
6164 | ia64_srlz_i(); |
---|
6165 | } |
---|
6166 | } |
---|
6167 | |
---|
6168 | #ifndef XEN |
---|
6169 | #ifdef CONFIG_SMP |
---|
6170 | |
---|
6171 | static void |
---|
6172 | pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs) |
---|
6173 | { |
---|
6174 | struct task_struct *task = ctx->ctx_task; |
---|
6175 | |
---|
6176 | ia64_psr(regs)->up = 0; |
---|
6177 | ia64_psr(regs)->sp = 1; |
---|
6178 | |
---|
6179 | if (GET_PMU_OWNER() == task) { |
---|
6180 | DPRINT(("cleared ownership for [%d]\n", ctx->ctx_task->pid)); |
---|
6181 | SET_PMU_OWNER(NULL, NULL); |
---|
6182 | } |
---|
6183 | |
---|
6184 | /* |
---|
6185 | * disconnect the task from the context and vice-versa |
---|
6186 | */ |
---|
6187 | PFM_SET_WORK_PENDING(task, 0); |
---|
6188 | |
---|
6189 | task->thread.pfm_context = NULL; |
---|
6190 | task->thread.flags &= ~IA64_THREAD_PM_VALID; |
---|
6191 | |
---|
6192 | DPRINT(("force cleanup for [%d]\n", task->pid)); |
---|
6193 | } |
---|
6194 | |
---|
6195 | |
---|
6196 | /* |
---|
6197 | * in 2.6, interrupts are masked when we come here and the runqueue lock is held |
---|
6198 | */ |
---|
6199 | void |
---|
6200 | pfm_save_regs(struct task_struct *task) |
---|
6201 | { |
---|
6202 | pfm_context_t *ctx; |
---|
6203 | struct thread_struct *t; |
---|
6204 | unsigned long flags; |
---|
6205 | u64 psr; |
---|
6206 | |
---|
6207 | |
---|
6208 | ctx = PFM_GET_CTX(task); |
---|
6209 | if (ctx == NULL) return; |
---|
6210 | t = &task->thread; |
---|
6211 | |
---|
6212 | /* |
---|
6213 | * we always come here with interrupts ALREADY disabled by |
---|
6214 | * the scheduler. So we simply need to protect against concurrent |
---|
6215 | * access, not CPU concurrency. |
---|
6216 | */ |
---|
6217 | flags = pfm_protect_ctx_ctxsw(ctx); |
---|
6218 | |
---|
6219 | if (ctx->ctx_state == PFM_CTX_ZOMBIE) { |
---|
6220 | struct pt_regs *regs = task_pt_regs(task); |
---|
6221 | |
---|
6222 | pfm_clear_psr_up(); |
---|
6223 | |
---|
6224 | pfm_force_cleanup(ctx, regs); |
---|
6225 | |
---|
6226 | BUG_ON(ctx->ctx_smpl_hdr); |
---|
6227 | |
---|
6228 | pfm_unprotect_ctx_ctxsw(ctx, flags); |
---|
6229 | |
---|
6230 | pfm_context_free(ctx); |
---|
6231 | return; |
---|
6232 | } |
---|
6233 | |
---|
6234 | /* |
---|
6235 | * save current PSR: needed because we modify it |
---|
6236 | */ |
---|
6237 | ia64_srlz_d(); |
---|
6238 | psr = pfm_get_psr(); |
---|
6239 | |
---|
6240 | BUG_ON(psr & (IA64_PSR_I)); |
---|
6241 | |
---|
6242 | /* |
---|
6243 | * stop monitoring: |
---|
6244 | * This is the last instruction which may generate an overflow |
---|
6245 | * |
---|
6246 | * We do not need to set psr.sp because, it is irrelevant in kernel. |
---|
6247 | * It will be restored from ipsr when going back to user level |
---|
6248 | */ |
---|
6249 | pfm_clear_psr_up(); |
---|
6250 | |
---|
6251 | /* |
---|
6252 | * keep a copy of psr.up (for reload) |
---|
6253 | */ |
---|
6254 | ctx->ctx_saved_psr_up = psr & IA64_PSR_UP; |
---|
6255 | |
---|
6256 | /* |
---|
6257 | * release ownership of this PMU. |
---|
6258 | * PM interrupts are masked, so nothing |
---|
6259 | * can happen. |
---|
6260 | */ |
---|
6261 | SET_PMU_OWNER(NULL, NULL); |
---|
6262 | |
---|
6263 | /* |
---|
6264 | * we systematically save the PMD as we have no |
---|
6265 | * guarantee we will be schedule at that same |
---|
6266 | * CPU again. |
---|
6267 | */ |
---|
6268 | pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]); |
---|
6269 | |
---|
6270 | /* |
---|
6271 | * save pmc0 ia64_srlz_d() done in pfm_save_pmds() |
---|
6272 | * we will need it on the restore path to check |
---|
6273 | * for pending overflow. |
---|
6274 | */ |
---|
6275 | t->pmcs[0] = ia64_get_pmc(0); |
---|
6276 | |
---|
6277 | /* |
---|
6278 | * unfreeze PMU if had pending overflows |
---|
6279 | */ |
---|
6280 | if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu(); |
---|
6281 | |
---|
6282 | /* |
---|
6283 | * finally, allow context access. |
---|
6284 | * interrupts will still be masked after this call. |
---|
6285 | */ |
---|
6286 | pfm_unprotect_ctx_ctxsw(ctx, flags); |
---|
6287 | } |
---|
6288 | |
---|
6289 | #else /* !CONFIG_SMP */ |
---|
6290 | void |
---|
6291 | pfm_save_regs(struct task_struct *task) |
---|
6292 | { |
---|
6293 | pfm_context_t *ctx; |
---|
6294 | u64 psr; |
---|
6295 | |
---|
6296 | ctx = PFM_GET_CTX(task); |
---|
6297 | if (ctx == NULL) return; |
---|
6298 | |
---|
6299 | /* |
---|
6300 | * save current PSR: needed because we modify it |
---|
6301 | */ |
---|
6302 | psr = pfm_get_psr(); |
---|
6303 | |
---|
6304 | BUG_ON(psr & (IA64_PSR_I)); |
---|
6305 | |
---|
6306 | /* |
---|
6307 | * stop monitoring: |
---|
6308 | * This is the last instruction which may generate an overflow |
---|
6309 | * |
---|
6310 | * We do not need to set psr.sp because, it is irrelevant in kernel. |
---|
6311 | * It will be restored from ipsr when going back to user level |
---|
6312 | */ |
---|
6313 | pfm_clear_psr_up(); |
---|
6314 | |
---|
6315 | /* |
---|
6316 | * keep a copy of psr.up (for reload) |
---|
6317 | */ |
---|
6318 | ctx->ctx_saved_psr_up = psr & IA64_PSR_UP; |
---|
6319 | } |
---|
6320 | |
---|
6321 | static void |
---|
6322 | pfm_lazy_save_regs (struct task_struct *task) |
---|
6323 | { |
---|
6324 | pfm_context_t *ctx; |
---|
6325 | struct thread_struct *t; |
---|
6326 | unsigned long flags; |
---|
6327 | |
---|
6328 | { u64 psr = pfm_get_psr(); |
---|
6329 | BUG_ON(psr & IA64_PSR_UP); |
---|
6330 | } |
---|
6331 | |
---|
6332 | ctx = PFM_GET_CTX(task); |
---|
6333 | t = &task->thread; |
---|
6334 | |
---|
6335 | /* |
---|
6336 | * we need to mask PMU overflow here to |
---|
6337 | * make sure that we maintain pmc0 until |
---|
6338 | * we save it. overflow interrupts are |
---|
6339 | * treated as spurious if there is no |
---|
6340 | * owner. |
---|
6341 | * |
---|
6342 | * XXX: I don't think this is necessary |
---|
6343 | */ |
---|
6344 | PROTECT_CTX(ctx,flags); |
---|
6345 | |
---|
6346 | /* |
---|
6347 | * release ownership of this PMU. |
---|
6348 | * must be done before we save the registers. |
---|
6349 | * |
---|
6350 | * after this call any PMU interrupt is treated |
---|
6351 | * as spurious. |
---|
6352 | */ |
---|
6353 | SET_PMU_OWNER(NULL, NULL); |
---|
6354 | |
---|
6355 | /* |
---|
6356 | * save all the pmds we use |
---|
6357 | */ |
---|
6358 | pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]); |
---|
6359 | |
---|
6360 | /* |
---|
6361 | * save pmc0 ia64_srlz_d() done in pfm_save_pmds() |
---|
6362 | * it is needed to check for pended overflow |
---|
6363 | * on the restore path |
---|
6364 | */ |
---|
6365 | t->pmcs[0] = ia64_get_pmc(0); |
---|
6366 | |
---|
6367 | /* |
---|
6368 | * unfreeze PMU if had pending overflows |
---|
6369 | */ |
---|
6370 | if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu(); |
---|
6371 | |
---|
6372 | /* |
---|
6373 | * now get can unmask PMU interrupts, they will |
---|
6374 | * be treated as purely spurious and we will not |
---|
6375 | * lose any information |
---|
6376 | */ |
---|
6377 | UNPROTECT_CTX(ctx,flags); |
---|
6378 | } |
---|
6379 | #endif /* CONFIG_SMP */ |
---|
6380 | |
---|
6381 | #ifdef CONFIG_SMP |
---|
6382 | /* |
---|
6383 | * in 2.6, interrupts are masked when we come here and the runqueue lock is held |
---|
6384 | */ |
---|
6385 | void |
---|
6386 | pfm_load_regs (struct task_struct *task) |
---|
6387 | { |
---|
6388 | pfm_context_t *ctx; |
---|
6389 | struct thread_struct *t; |
---|
6390 | unsigned long pmc_mask = 0UL, pmd_mask = 0UL; |
---|
6391 | unsigned long flags; |
---|
6392 | u64 psr, psr_up; |
---|
6393 | int need_irq_resend; |
---|
6394 | |
---|
6395 | ctx = PFM_GET_CTX(task); |
---|
6396 | if (unlikely(ctx == NULL)) return; |
---|
6397 | |
---|
6398 | BUG_ON(GET_PMU_OWNER()); |
---|
6399 | |
---|
6400 | t = &task->thread; |
---|
6401 | /* |
---|
6402 | * possible on unload |
---|
6403 | */ |
---|
6404 | if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return; |
---|
6405 | |
---|
6406 | /* |
---|
6407 | * we always come here with interrupts ALREADY disabled by |
---|
6408 | * the scheduler. So we simply need to protect against concurrent |
---|
6409 | * access, not CPU concurrency. |
---|
6410 | */ |
---|
6411 | flags = pfm_protect_ctx_ctxsw(ctx); |
---|
6412 | psr = pfm_get_psr(); |
---|
6413 | |
---|
6414 | need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND; |
---|
6415 | |
---|
6416 | BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP)); |
---|
6417 | BUG_ON(psr & IA64_PSR_I); |
---|
6418 | |
---|
6419 | if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) { |
---|
6420 | struct pt_regs *regs = task_pt_regs(task); |
---|
6421 | |
---|
6422 | BUG_ON(ctx->ctx_smpl_hdr); |
---|
6423 | |
---|
6424 | pfm_force_cleanup(ctx, regs); |
---|
6425 | |
---|
6426 | pfm_unprotect_ctx_ctxsw(ctx, flags); |
---|
6427 | |
---|
6428 | /* |
---|
6429 | * this one (kmalloc'ed) is fine with interrupts disabled |
---|
6430 | */ |
---|
6431 | pfm_context_free(ctx); |
---|
6432 | |
---|
6433 | return; |
---|
6434 | } |
---|
6435 | |
---|
6436 | /* |
---|
6437 | * we restore ALL the debug registers to avoid picking up |
---|
6438 | * stale state. |
---|
6439 | */ |
---|
6440 | if (ctx->ctx_fl_using_dbreg) { |
---|
6441 | pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs); |
---|
6442 | pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs); |
---|
6443 | } |
---|
6444 | /* |
---|
6445 | * retrieve saved psr.up |
---|
6446 | */ |
---|
6447 | psr_up = ctx->ctx_saved_psr_up; |
---|
6448 | |
---|
6449 | /* |
---|
6450 | * if we were the last user of the PMU on that CPU, |
---|
6451 | * then nothing to do except restore psr |
---|
6452 | */ |
---|
6453 | if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) { |
---|
6454 | |
---|
6455 | /* |
---|
6456 | * retrieve partial reload masks (due to user modifications) |
---|
6457 | */ |
---|
6458 | pmc_mask = ctx->ctx_reload_pmcs[0]; |
---|
6459 | pmd_mask = ctx->ctx_reload_pmds[0]; |
---|
6460 | |
---|
6461 | } else { |
---|
6462 | /* |
---|
6463 | * To avoid leaking information to the user level when psr.sp=0, |
---|
6464 | * we must reload ALL implemented pmds (even the ones we don't use). |
---|
6465 | * In the kernel we only allow PFM_READ_PMDS on registers which |
---|
6466 | * we initialized or requested (sampling) so there is no risk there. |
---|
6467 | */ |
---|
6468 | pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0]; |
---|
6469 | |
---|
6470 | /* |
---|
6471 | * ALL accessible PMCs are systematically reloaded, unused registers |
---|
6472 | * get their default (from pfm_reset_pmu_state()) values to avoid picking |
---|
6473 | * up stale configuration. |
---|
6474 | * |
---|
6475 | * PMC0 is never in the mask. It is always restored separately. |
---|
6476 | */ |
---|
6477 | pmc_mask = ctx->ctx_all_pmcs[0]; |
---|
6478 | } |
---|
6479 | /* |
---|
6480 | * when context is MASKED, we will restore PMC with plm=0 |
---|
6481 | * and PMD with stale information, but that's ok, nothing |
---|
6482 | * will be captured. |
---|
6483 | * |
---|
6484 | * XXX: optimize here |
---|
6485 | */ |
---|
6486 | if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask); |
---|
6487 | if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask); |
---|
6488 | |
---|
6489 | /* |
---|
6490 | * check for pending overflow at the time the state |
---|
6491 | * was saved. |
---|
6492 | */ |
---|
6493 | if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) { |
---|
6494 | /* |
---|
6495 | * reload pmc0 with the overflow information |
---|
6496 | * On McKinley PMU, this will trigger a PMU interrupt |
---|
6497 | */ |
---|
6498 | ia64_set_pmc(0, t->pmcs[0]); |
---|
6499 | ia64_srlz_d(); |
---|
6500 | t->pmcs[0] = 0UL; |
---|
6501 | |
---|
6502 | /* |
---|
6503 | * will replay the PMU interrupt |
---|
6504 | */ |
---|
6505 | if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR); |
---|
6506 | |
---|
6507 | pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++; |
---|
6508 | } |
---|
6509 | |
---|
6510 | /* |
---|
6511 | * we just did a reload, so we reset the partial reload fields |
---|
6512 | */ |
---|
6513 | ctx->ctx_reload_pmcs[0] = 0UL; |
---|
6514 | ctx->ctx_reload_pmds[0] = 0UL; |
---|
6515 | |
---|
6516 | SET_LAST_CPU(ctx, smp_processor_id()); |
---|
6517 | |
---|
6518 | /* |
---|
6519 | * dump activation value for this PMU |
---|
6520 | */ |
---|
6521 | INC_ACTIVATION(); |
---|
6522 | /* |
---|
6523 | * record current activation for this context |
---|
6524 | */ |
---|
6525 | SET_ACTIVATION(ctx); |
---|
6526 | |
---|
6527 | /* |
---|
6528 | * establish new ownership. |
---|
6529 | */ |
---|
6530 | SET_PMU_OWNER(task, ctx); |
---|
6531 | |
---|
6532 | /* |
---|
6533 | * restore the psr.up bit. measurement |
---|
6534 | * is active again. |
---|
6535 | * no PMU interrupt can happen at this point |
---|
6536 | * because we still have interrupts disabled. |
---|
6537 | */ |
---|
6538 | if (likely(psr_up)) pfm_set_psr_up(); |
---|
6539 | |
---|
6540 | /* |
---|
6541 | * allow concurrent access to context |
---|
6542 | */ |
---|
6543 | pfm_unprotect_ctx_ctxsw(ctx, flags); |
---|
6544 | } |
---|
6545 | #else /* !CONFIG_SMP */ |
---|
6546 | /* |
---|
6547 | * reload PMU state for UP kernels |
---|
6548 | * in 2.5 we come here with interrupts disabled |
---|
6549 | */ |
---|
6550 | void |
---|
6551 | pfm_load_regs (struct task_struct *task) |
---|
6552 | { |
---|
6553 | struct thread_struct *t; |
---|
6554 | pfm_context_t *ctx; |
---|
6555 | struct task_struct *owner; |
---|
6556 | unsigned long pmd_mask, pmc_mask; |
---|
6557 | u64 psr, psr_up; |
---|
6558 | int need_irq_resend; |
---|
6559 | |
---|
6560 | owner = GET_PMU_OWNER(); |
---|
6561 | ctx = PFM_GET_CTX(task); |
---|
6562 | t = &task->thread; |
---|
6563 | psr = pfm_get_psr(); |
---|
6564 | |
---|
6565 | BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP)); |
---|
6566 | BUG_ON(psr & IA64_PSR_I); |
---|
6567 | |
---|
6568 | /* |
---|
6569 | * we restore ALL the debug registers to avoid picking up |
---|
6570 | * stale state. |
---|
6571 | * |
---|
6572 | * This must be done even when the task is still the owner |
---|
6573 | * as the registers may have been modified via ptrace() |
---|
6574 | * (not perfmon) by the previous task. |
---|
6575 | */ |
---|
6576 | if (ctx->ctx_fl_using_dbreg) { |
---|
6577 | pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs); |
---|
6578 | pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs); |
---|
6579 | } |
---|
6580 | |
---|
6581 | /* |
---|
6582 | * retrieved saved psr.up |
---|
6583 | */ |
---|
6584 | psr_up = ctx->ctx_saved_psr_up; |
---|
6585 | need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND; |
---|
6586 | |
---|
6587 | /* |
---|
6588 | * short path, our state is still there, just |
---|
6589 | * need to restore psr and we go |
---|
6590 | * |
---|
6591 | * we do not touch either PMC nor PMD. the psr is not touched |
---|
6592 | * by the overflow_handler. So we are safe w.r.t. to interrupt |
---|
6593 | * concurrency even without interrupt masking. |
---|
6594 | */ |
---|
6595 | if (likely(owner == task)) { |
---|
6596 | if (likely(psr_up)) pfm_set_psr_up(); |
---|
6597 | return; |
---|
6598 | } |
---|
6599 | |
---|
6600 | /* |
---|
6601 | * someone else is still using the PMU, first push it out and |
---|
6602 | * then we'll be able to install our stuff ! |
---|
6603 | * |
---|
6604 | * Upon return, there will be no owner for the current PMU |
---|
6605 | */ |
---|
6606 | if (owner) pfm_lazy_save_regs(owner); |
---|
6607 | |
---|
6608 | /* |
---|
6609 | * To avoid leaking information to the user level when psr.sp=0, |
---|
6610 | * we must reload ALL implemented pmds (even the ones we don't use). |
---|
6611 | * In the kernel we only allow PFM_READ_PMDS on registers which |
---|
6612 | * we initialized or requested (sampling) so there is no risk there. |
---|
6613 | */ |
---|
6614 | pmd_mask = pfm_sysctl.fastctxsw ? ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0]; |
---|
6615 | |
---|
6616 | /* |
---|
6617 | * ALL accessible PMCs are systematically reloaded, unused registers |
---|
6618 | * get their default (from pfm_reset_pmu_state()) values to avoid picking |
---|