source: trunk/packages/xen-3.1/xen-3.1/tools/ioemu/patches/xen-mapcache @ 34

# HG changeset patch
# User kfraser@localhost.localdomain
# Node ID 67a06a9b7b1dca707e1cd3b08ae0a341d6e97b3d
# Parent  3f0ca90351e268084fbdb733d70fc596cb46537d
[HVM] qemu: Add guest address-space mapping cache.

On IA32 host or IA32 PAE host, at present, generally, we can't create
an HVM guest with more than 2G memory, because generally it's almost
impossible for Qemu to find a large enough and consecutive virtual
address space to map an HVM guest's whole physical address space.
The attached patch fixes this issue using dynamic mapping based on
little blocks of memory.

Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>
Signed-off-by: Dexuan Cui <dexuan.cui@intel.com>
Signed-off-by: Keir Fraser <keir@xensource.com>

Index: ioemu/vl.c
===================================================================
--- ioemu.orig/vl.c     2007-05-03 15:12:21.000000000 +0100
+++ ioemu/vl.c  2007-05-03 15:12:41.000000000 +0100
@@ -286,7 +286,7 @@
     for(i = start; i < start + length; i += size) {
         ioport_write_table[bsize][i] = func;
         if (ioport_opaque[i] != NULL && ioport_opaque[i] != opaque)
-            hw_error("register_ioport_read: invalid opaque");
+            hw_error("register_ioport_write: invalid opaque");
         ioport_opaque[i] = opaque;
     }
     return 0;
@@ -5894,6 +5894,157 @@
     suspend_requested = 1;
 }
 
+#if defined(MAPCACHE)
+
+#if defined(__i386__) 
+#define MAX_MCACHE_SIZE    0x40000000 /* 1GB max for x86 */
+#define MCACHE_BUCKET_SHIFT 16
+#elif defined(__x86_64__)
+#define MAX_MCACHE_SIZE    0x1000000000 /* 64GB max for x86_64 */
+#define MCACHE_BUCKET_SHIFT 20
+#endif
+
+#define MCACHE_BUCKET_SIZE (1UL << MCACHE_BUCKET_SHIFT)
+
+#define BITS_PER_LONG (sizeof(long)*8)
+#define BITS_TO_LONGS(bits) \
+    (((bits)+BITS_PER_LONG-1)/BITS_PER_LONG)
+#define DECLARE_BITMAP(name,bits) \
+    unsigned long name[BITS_TO_LONGS(bits)]
+#define test_bit(bit,map) \
+    (!!((map)[(bit)/BITS_PER_LONG] & (1UL << ((bit)%BITS_PER_LONG))))
+
+struct map_cache {
+    unsigned long paddr_index;
+    uint8_t      *vaddr_base;
+    DECLARE_BITMAP(valid_mapping, MCACHE_BUCKET_SIZE>>PAGE_SHIFT);
+};
+
+static struct map_cache *mapcache_entry;
+static unsigned long nr_buckets;
+
+/* For most cases (>99.9%), the page address is the same. */
+static unsigned long last_address_index = ~0UL;
+static uint8_t      *last_address_vaddr;
+
+static int qemu_map_cache_init(void)
+{
+    unsigned long size;
+
+    nr_buckets = (((MAX_MCACHE_SIZE >> PAGE_SHIFT) +
+                   (1UL << (MCACHE_BUCKET_SHIFT - PAGE_SHIFT)) - 1) >>
+                  (MCACHE_BUCKET_SHIFT - PAGE_SHIFT));
+    fprintf(logfile, "qemu_map_cache_init nr_buckets = %lx\n", nr_buckets);
+
+    /*
+     * Use mmap() directly: lets us allocate a big hash table with no up-front
+     * cost in storage space. The OS will allocate memory only for the buckets
+     * that we actually use. All others will contain all zeroes.
+     */
+    size = nr_buckets * sizeof(struct map_cache);
+    size = (size + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);
+    mapcache_entry = mmap(NULL, size, PROT_READ|PROT_WRITE,
+                          MAP_SHARED|MAP_ANONYMOUS, 0, 0);
+    if (mapcache_entry == MAP_FAILED) {
+        errno = ENOMEM;
+        return -1;
+    }
+
+    return 0;
+}
+
+static void qemu_remap_bucket(struct map_cache *entry,
+                              unsigned long address_index)
+{
+    uint8_t *vaddr_base;
+    unsigned long pfns[MCACHE_BUCKET_SIZE >> PAGE_SHIFT];
+    unsigned int i, j;
+
+    if (entry->vaddr_base != NULL) {
+        errno = munmap(entry->vaddr_base, MCACHE_BUCKET_SIZE);
+        if (errno) {
+            fprintf(logfile, "unmap fails %d\n", errno);
+            exit(-1);
+        }
+    }
+
+    for (i = 0; i < MCACHE_BUCKET_SIZE >> PAGE_SHIFT; i++)
+        pfns[i] = (address_index << (MCACHE_BUCKET_SHIFT-PAGE_SHIFT)) + i;
+
+    vaddr_base = xc_map_foreign_batch(xc_handle, domid, PROT_READ|PROT_WRITE,
+                                      pfns, MCACHE_BUCKET_SIZE >> PAGE_SHIFT);
+    if (vaddr_base == NULL) {
+        fprintf(logfile, "xc_map_foreign_batch error %d\n", errno);
+        exit(-1);
+    }
+
+    entry->vaddr_base  = vaddr_base;
+    entry->paddr_index = address_index;
+
+    for (i = 0; i < MCACHE_BUCKET_SIZE >> PAGE_SHIFT; i += BITS_PER_LONG) {
+        unsigned long word = 0;
+        j = ((i + BITS_PER_LONG) > (MCACHE_BUCKET_SIZE >> PAGE_SHIFT)) ?
+            (MCACHE_BUCKET_SIZE >> PAGE_SHIFT) % BITS_PER_LONG : BITS_PER_LONG;
+        while (j > 0)
+            word = (word << 1) | !(pfns[i + --j] & 0xF0000000UL);
+        entry->valid_mapping[i / BITS_PER_LONG] = word;
+    }
+}
+
+uint8_t *qemu_map_cache(target_phys_addr_t phys_addr)
+{
+    struct map_cache *entry;
+    unsigned long address_index  = phys_addr >> MCACHE_BUCKET_SHIFT;
+    unsigned long address_offset = phys_addr & (MCACHE_BUCKET_SIZE-1);
+
+    if (address_index == last_address_index)
+        return last_address_vaddr + address_offset;
+
+    entry = &mapcache_entry[address_index % nr_buckets];
+
+    if (entry->vaddr_base == NULL || entry->paddr_index != address_index ||
+        !test_bit(address_offset>>PAGE_SHIFT, entry->valid_mapping))
+        qemu_remap_bucket(entry, address_index);
+
+    if (!test_bit(address_offset>>PAGE_SHIFT, entry->valid_mapping))
+        return NULL;
+
+    last_address_index = address_index;
+    last_address_vaddr = entry->vaddr_base;
+
+    return last_address_vaddr + address_offset;
+}
+
+void qemu_invalidate_map_cache(void)
+{
+    unsigned long i;
+
+    mapcache_lock();
+
+    for (i = 0; i < nr_buckets; i++) {
+        struct map_cache *entry = &mapcache_entry[i];
+
+        if (entry->vaddr_base == NULL)
+            continue;
+
+        errno = munmap(entry->vaddr_base, MCACHE_BUCKET_SIZE);
+        if (errno) {
+            fprintf(logfile, "unmap fails %d\n", errno);
+            exit(-1);
+        }
+
+        entry->paddr_index = 0;
+        entry->vaddr_base  = NULL;
+    }
+
+    last_address_index =  ~0UL;
+    last_address_vaddr = NULL;
+
+    mapcache_unlock();
+}
+
+#endif /* defined(MAPCACHE) */
+
 int main(int argc, char **argv)
 {
 #ifdef CONFIG_GDBSTUB
@@ -5930,8 +6081,11 @@
     unsigned long ioreq_pfn;
     extern void *shared_page;
     extern void *buffered_io_page;
-    extern void *buffered_pio_page;
+#ifdef __ia64__
     unsigned long nr_pages;
+    xen_pfn_t *page_array;
+    extern void *buffered_pio_page;
+#endif
 
     char qemu_dm_logfilename[64];
 
@@ -6221,6 +6375,7 @@
                 break;
             case QEMU_OPTION_m:
                 ram_size = atol(optarg) * 1024 * 1024;
+                ram_size = (uint64_t)atol(optarg) * 1024 * 1024;
                 if (ram_size <= 0)
                     help();
 #ifndef CONFIG_DM
@@ -6482,30 +6637,15 @@
 
 #if defined(__i386__) || defined(__x86_64__)
 
-    nr_pages = ram_size/PAGE_SIZE;
-
-    page_array = (xen_pfn_t *)malloc(nr_pages * sizeof(xen_pfn_t));
-    if (page_array == NULL) {
-        fprintf(logfile, "malloc returned error %d\n", errno);
-        exit(-1);
-    }
-
-    for ( i = 0; i < nr_pages; i++)
-        page_array[i] = i;
-
-    phys_ram_base = xc_map_foreign_batch(xc_handle, domid,
-                                         PROT_READ|PROT_WRITE, page_array,
-                                         nr_pages);
-    if (phys_ram_base == NULL) {
-        fprintf(logfile, "batch map guest memory returned error %d\n", errno);
+    if (qemu_map_cache_init()) {
+        fprintf(logfile, "qemu_map_cache_init returned: error %d\n", errno);
         exit(-1);
     }
 
     xc_get_hvm_param(xc_handle, domid, HVM_PARAM_IOREQ_PFN, &ioreq_pfn);
     fprintf(logfile, "shared page at pfn %lx\n", ioreq_pfn);
     shared_page = xc_map_foreign_range(xc_handle, domid, PAGE_SIZE,
-                                       PROT_READ|PROT_WRITE,
-                                       page_array[ioreq_pfn]);
+                                       PROT_READ|PROT_WRITE, ioreq_pfn);
     if (shared_page == NULL) {
         fprintf(logfile, "map shared IO page returned error %d\n", errno);
         exit(-1);
@@ -6514,15 +6654,12 @@
     xc_get_hvm_param(xc_handle, domid, HVM_PARAM_BUFIOREQ_PFN, &ioreq_pfn);
     fprintf(logfile, "buffered io page at pfn %lx\n", ioreq_pfn);
     buffered_io_page = xc_map_foreign_range(xc_handle, domid, PAGE_SIZE,
-                                            PROT_READ|PROT_WRITE,
-                                            page_array[ioreq_pfn]);
+                                            PROT_READ|PROT_WRITE, ioreq_pfn);
     if (buffered_io_page == NULL) {
         fprintf(logfile, "map buffered IO page returned error %d\n", errno);
         exit(-1);
     }
 
-    free(page_array);
-
 #elif defined(__ia64__)
 
     nr_pages = ram_size/PAGE_SIZE;
Index: ioemu/target-i386-dm/exec-dm.c
===================================================================
--- ioemu.orig/target-i386-dm/exec-dm.c 2007-05-03 15:10:22.000000000 +0100
+++ ioemu/target-i386-dm/exec-dm.c      2007-05-03 15:12:34.000000000 +0100
@@ -36,6 +36,7 @@
 
 #include "cpu.h"
 #include "exec-all.h"
+#include "vl.h"
 
 //#define DEBUG_TB_INVALIDATE
 //#define DEBUG_FLUSH
@@ -127,10 +128,17 @@
 FILE *logfile;
 int loglevel;
 
+#ifdef MAPCACHE
+pthread_mutex_t mapcache_mutex;
+#endif
+
 void cpu_exec_init(CPUState *env)
 {
     CPUState **penv;
     int cpu_index;
+#ifdef MAPCACHE
+    pthread_mutexattr_t mxattr; 
+#endif
 
     env->next_cpu = NULL;
     penv = &first_cpu;
@@ -144,6 +152,14 @@
 
     /* alloc dirty bits array */
     phys_ram_dirty = qemu_malloc(phys_ram_size >> TARGET_PAGE_BITS);
+
+#ifdef MAPCACHE
+    /* setup memory access mutex to protect mapcache */
+    pthread_mutexattr_init(&mxattr); 
+    pthread_mutexattr_settype(&mxattr, PTHREAD_MUTEX_RECURSIVE);
+    pthread_mutex_init(&mapcache_mutex, &mxattr); 
+    pthread_mutexattr_destroy(&mxattr); 
+#endif
 }
 
 /* enable or disable low levels log */
@@ -409,16 +425,11 @@
         return 0;
 }
 
-static inline int paddr_is_ram(target_phys_addr_t addr)
-{
-    /* Is this guest physical address RAM-backed? */
-#if defined(CONFIG_DM) && (defined(__i386__) || defined(__x86_64__))
-    return ((addr < HVM_BELOW_4G_MMIO_START) ||
-            (addr >= HVM_BELOW_4G_MMIO_START + HVM_BELOW_4G_MMIO_LENGTH));
-#else
-    return (addr < ram_size);
+#if defined(__i386__) || defined(__x86_64__)
+#define phys_ram_addr(x) (qemu_map_cache(x))
+#elif defined(__ia64__)
+#define phys_ram_addr(x) ((addr < ram_size) ? (phys_ram_base + (x)) : NULL)
 #endif
-}
 
 void cpu_physical_memory_rw(target_phys_addr_t addr, uint8_t *buf, 
                             int len, int is_write)
@@ -426,13 +437,15 @@
     int l, io_index;
     uint8_t *ptr;
     uint32_t val;
-    
+
+    mapcache_lock();
+
     while (len > 0) {
         /* How much can we copy before the next page boundary? */
         l = TARGET_PAGE_SIZE - (addr & ~TARGET_PAGE_MASK); 
         if (l > len)
             l = len;
-       
+
         io_index = iomem_index(addr);
         if (is_write) {
             if (io_index) {
@@ -452,11 +465,11 @@
                     io_mem_write[io_index][0](io_mem_opaque[io_index], addr, val);
                     l = 1;
                 }
-            } else if (paddr_is_ram(addr)) {
+            } else if ((ptr = phys_ram_addr(addr)) != NULL) {
                 /* Reading from RAM */
-                memcpy(phys_ram_base + addr, buf, l);
+                memcpy(ptr, buf, l);
 #ifdef __ia64__
-                sync_icache((unsigned long)(phys_ram_base + addr), l);
+                sync_icache(ptr, l);
 #endif 
             }
         } else {
@@ -477,9 +490,9 @@
                     stb_raw(buf, val);
                     l = 1;
                 }
-            } else if (paddr_is_ram(addr)) {
+            } else if ((ptr = phys_ram_addr(addr)) != NULL) {
                 /* Reading from RAM */
-                memcpy(buf, phys_ram_base + addr, l);
+                memcpy(buf, ptr, l);
             } else {
                 /* Neither RAM nor known MMIO space */
                 memset(buf, 0xff, len); 
@@ -489,6 +502,8 @@
         buf += l;
         addr += l;
     }
+
+    mapcache_unlock();
 }
 #endif
 
Index: ioemu/vl.h
===================================================================
--- ioemu.orig/vl.h     2007-05-03 15:12:20.000000000 +0100
+++ ioemu/vl.h  2007-05-03 15:12:34.000000000 +0100
@@ -156,6 +156,28 @@
 
 extern FILE *logfile;
 
+
+#if defined(__i386__) || defined(__x86_64__)
+
+#define MAPCACHE
+
+uint8_t *qemu_map_cache(target_phys_addr_t phys_addr);
+void     qemu_invalidate_map_cache(void);
+
+#include <pthread.h>
+extern  pthread_mutex_t mapcache_mutex;
+#define mapcache_lock() pthread_mutex_lock(&mapcache_mutex)
+#define mapcache_unlock() pthread_mutex_unlock(&mapcache_mutex)
+
+#else 
+
+#define qemu_invalidate_map_cache() ((void)0)
+
+#define mapcache_lock()   ((void)0)
+#define mapcache_unlock() ((void)0)
+
+#endif
+
 extern int xc_handle;
 extern int domid;
 
Index: ioemu/target-i386-dm/cpu.h
===================================================================
--- ioemu.orig/target-i386-dm/cpu.h     2007-05-03 15:10:22.000000000 +0100
+++ ioemu/target-i386-dm/cpu.h  2007-05-03 15:12:21.000000000 +0100
@@ -25,7 +25,8 @@
 #ifdef TARGET_X86_64
 #define TARGET_LONG_BITS 64
 #else
-#define TARGET_LONG_BITS 32
+/* #define TARGET_LONG_BITS 32 */
+#define TARGET_LONG_BITS 64 /* for Qemu map cache */
 #endif
 
 /* target supports implicit self modifying code */
Index: ioemu/target-i386-dm/helper2.c
===================================================================
--- ioemu.orig/target-i386-dm/helper2.c 2007-05-03 15:12:19.000000000 +0100
+++ ioemu/target-i386-dm/helper2.c      2007-05-03 15:12:21.000000000 +0100
@@ -526,6 +526,9 @@
     case IOREQ_TYPE_TIMEOFFSET:
         cpu_ioreq_timeoffset(env, req);
         break;
+    case IOREQ_TYPE_INVALIDATE:
+        qemu_invalidate_map_cache();
+        break;
     default:
         hw_error("Invalid ioreq type 0x%x\n", req->type);
     }