]> git.proxmox.com Git - mirror_qemu.git/blobdiff - exec.c
Expose thread_id in info cpus
[mirror_qemu.git] / exec.c
diff --git a/exec.c b/exec.c
index 137175caa0cd79bb43f45c1761e685d432f24123..964ce318fbc713de6d6bf09cb31b4f5846ee1f48 100644 (file)
--- a/exec.c
+++ b/exec.c
 #include <sys/types.h>
 #include <sys/mman.h>
 #endif
-#include <stdlib.h>
-#include <stdio.h>
-#include <stdarg.h>
-#include <string.h>
-#include <errno.h>
-#include <unistd.h>
-#include <inttypes.h>
 
+#include "qemu-common.h"
 #include "cpu.h"
 #include "exec-all.h"
-#include "qemu-common.h"
 #include "tcg.h"
 #include "hw/hw.h"
+#include "hw/qdev.h"
 #include "osdep.h"
 #include "kvm.h"
 #include "qemu-timer.h"
@@ -79,7 +73,7 @@
 #define SMC_BITMAP_USE_THRESHOLD 10
 
 static TranslationBlock *tbs;
-int code_gen_max_blocks;
+static int code_gen_max_blocks;
 TranslationBlock *tb_phys_hash[CODE_GEN_PHYS_HASH_SIZE];
 static int nb_tbs;
 /* any access to the tbs or the page table must use this lock */
@@ -106,7 +100,7 @@ static uint8_t *code_gen_buffer;
 static unsigned long code_gen_buffer_size;
 /* threshold to flush the translated code buffer */
 static unsigned long code_gen_buffer_max_size;
-uint8_t *code_gen_ptr;
+static uint8_t *code_gen_ptr;
 
 #if !defined(CONFIG_USER_ONLY)
 int phys_ram_fd;
@@ -523,7 +517,8 @@ static void code_gen_alloc(unsigned long tb_size)
             exit(1);
         }
     }
-#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__)
+#elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) \
+    || defined(__DragonFly__) || defined(__OpenBSD__)
     {
         int flags;
         void *addr = NULL;
@@ -536,6 +531,13 @@ static void code_gen_alloc(unsigned long tb_size)
         /* Cannot map more than that */
         if (code_gen_buffer_size > (800 * 1024 * 1024))
             code_gen_buffer_size = (800 * 1024 * 1024);
+#elif defined(__sparc_v9__)
+        // Map the buffer below 2G, so we can use direct calls and branches
+        flags |= MAP_FIXED;
+        addr = (void *) 0x60000000UL;
+        if (code_gen_buffer_size > (512 * 1024 * 1024)) {
+            code_gen_buffer_size = (512 * 1024 * 1024);
+        }
 #endif
         code_gen_buffer = mmap(addr, code_gen_buffer_size,
                                PROT_WRITE | PROT_READ | PROT_EXEC, 
@@ -636,17 +638,46 @@ void cpu_exec_init(CPUState *env)
     env->numa_node = 0;
     QTAILQ_INIT(&env->breakpoints);
     QTAILQ_INIT(&env->watchpoints);
+#ifndef CONFIG_USER_ONLY
+    env->thread_id = qemu_get_thread_id();
+#endif
     *penv = env;
 #if defined(CONFIG_USER_ONLY)
     cpu_list_unlock();
 #endif
 #if defined(CPU_SAVE_VERSION) && !defined(CONFIG_USER_ONLY)
-    vmstate_register(cpu_index, &vmstate_cpu_common, env);
-    register_savevm("cpu", cpu_index, CPU_SAVE_VERSION,
+    vmstate_register(NULL, cpu_index, &vmstate_cpu_common, env);
+    register_savevm(NULL, "cpu", cpu_index, CPU_SAVE_VERSION,
                     cpu_save, cpu_load, env);
 #endif
 }
 
+/* Allocate a new translation block. Flush the translation buffer if
+   too many translation blocks or too much generated code. */
+static TranslationBlock *tb_alloc(target_ulong pc)
+{
+    TranslationBlock *tb;
+
+    if (nb_tbs >= code_gen_max_blocks ||
+        (code_gen_ptr - code_gen_buffer) >= code_gen_buffer_max_size)
+        return NULL;
+    tb = &tbs[nb_tbs++];
+    tb->pc = pc;
+    tb->cflags = 0;
+    return tb;
+}
+
+void tb_free(TranslationBlock *tb)
+{
+    /* In practice this is mostly used for single use temporary TB
+       Ignore the hard cases and just back up if this TB happens to
+       be the last one generated.  */
+    if (nb_tbs > 0 && tb == &tbs[nb_tbs - 1]) {
+        code_gen_ptr = tb->tc_ptr;
+        nb_tbs--;
+    }
+}
+
 static inline void invalidate_page_bitmap(PageDesc *p)
 {
     if (p->code_bitmap) {
@@ -1224,32 +1255,6 @@ static inline void tb_alloc_page(TranslationBlock *tb,
 #endif /* TARGET_HAS_SMC */
 }
 
-/* Allocate a new translation block. Flush the translation buffer if
-   too many translation blocks or too much generated code. */
-TranslationBlock *tb_alloc(target_ulong pc)
-{
-    TranslationBlock *tb;
-
-    if (nb_tbs >= code_gen_max_blocks ||
-        (code_gen_ptr - code_gen_buffer) >= code_gen_buffer_max_size)
-        return NULL;
-    tb = &tbs[nb_tbs++];
-    tb->pc = pc;
-    tb->cflags = 0;
-    return tb;
-}
-
-void tb_free(TranslationBlock *tb)
-{
-    /* In practice this is mostly used for single use temporary TB
-       Ignore the hard cases and just back up if this TB happens to
-       be the last one generated.  */
-    if (nb_tbs > 0 && tb == &tbs[nb_tbs - 1]) {
-        code_gen_ptr = tb->tc_ptr;
-        nb_tbs--;
-    }
-}
-
 /* add a new TB and link it to the physical page tables. phys_page2 is
    (-1) to indicate that only one page contains the TB. */
 void tb_link_page(TranslationBlock *tb,
@@ -1638,7 +1643,7 @@ void cpu_interrupt(CPUState *env, int mask)
      * If called from iothread context, wake the target cpu in
      * case its halted.
      */
-    if (!qemu_cpu_self(env)) {
+    if (!qemu_cpu_is_self(env)) {
         qemu_cpu_kick(env);
         return;
     }
@@ -1705,8 +1710,8 @@ static QLIST_HEAD(memory_client_list, CPUPhysMemoryClient) memory_client_list
     = QLIST_HEAD_INITIALIZER(memory_client_list);
 
 static void cpu_notify_set_memory(target_phys_addr_t start_addr,
-                                 ram_addr_t size,
-                                 ram_addr_t phys_offset)
+                                  ram_addr_t size,
+                                  ram_addr_t phys_offset)
 {
     CPUPhysMemoryClient *client;
     QLIST_FOREACH(client, &memory_client_list, list) {
@@ -1715,7 +1720,7 @@ static void cpu_notify_set_memory(target_phys_addr_t start_addr,
 }
 
 static int cpu_notify_sync_dirty_bitmap(target_phys_addr_t start,
-                                       target_phys_addr_t end)
+                                        target_phys_addr_t end)
 {
     CPUPhysMemoryClient *client;
     QLIST_FOREACH(client, &memory_client_list, list) {
@@ -1802,17 +1807,17 @@ int cpu_str_to_log_mask(const char *str)
         p1 = strchr(p, ',');
         if (!p1)
             p1 = p + strlen(p);
-       if(cmp1(p,p1-p,"all")) {
-               for(item = cpu_log_items; item->mask != 0; item++) {
-                       mask |= item->mask;
-               }
-       } else {
-        for(item = cpu_log_items; item->mask != 0; item++) {
-            if (cmp1(p, p1 - p, item->name))
-                goto found;
+        if(cmp1(p,p1-p,"all")) {
+            for(item = cpu_log_items; item->mask != 0; item++) {
+                mask |= item->mask;
+            }
+        } else {
+            for(item = cpu_log_items; item->mask != 0; item++) {
+                if (cmp1(p, p1 - p, item->name))
+                    goto found;
+            }
+            return 0;
         }
-        return 0;
-       }
     found:
         mask |= item->mask;
         if (*p1 != ',')
@@ -1906,11 +1911,11 @@ static inline void tlb_flush_jmp_cache(CPUState *env, target_ulong addr)
        overlap the flushed page.  */
     i = tb_jmp_cache_hash_page(addr - TARGET_PAGE_SIZE);
     memset (&env->tb_jmp_cache[i], 0, 
-           TB_JMP_PAGE_SIZE * sizeof(TranslationBlock *));
+            TB_JMP_PAGE_SIZE * sizeof(TranslationBlock *));
 
     i = tb_jmp_cache_hash_page(addr);
     memset (&env->tb_jmp_cache[i], 0, 
-           TB_JMP_PAGE_SIZE * sizeof(TranslationBlock *));
+            TB_JMP_PAGE_SIZE * sizeof(TranslationBlock *));
 }
 
 static CPUTLBEntry s_cputlb_empty_entry = {
@@ -2036,10 +2041,10 @@ void cpu_physical_memory_reset_dirty(ram_addr_t start, ram_addr_t end,
 
     /* we modify the TLB cache so that the dirty bit will be set again
        when accessing the range */
-    start1 = (unsigned long)qemu_get_ram_ptr(start);
+    start1 = (unsigned long)qemu_safe_ram_ptr(start);
     /* Chek that we don't span multiple blocks - this breaks the
        address comparisons below.  */
-    if ((unsigned long)qemu_get_ram_ptr(end - 1) - start1
+    if ((unsigned long)qemu_safe_ram_ptr(end - 1) - start1
             != (end - 1) - start) {
         abort();
     }
@@ -2076,6 +2081,36 @@ int cpu_physical_sync_dirty_bitmap(target_phys_addr_t start_addr,
     return ret;
 }
 
+int cpu_physical_log_start(target_phys_addr_t start_addr,
+                           ram_addr_t size)
+{
+    CPUPhysMemoryClient *client;
+    QLIST_FOREACH(client, &memory_client_list, list) {
+        if (client->log_start) {
+            int r = client->log_start(client, start_addr, size);
+            if (r < 0) {
+                return r;
+            }
+        }
+    }
+    return 0;
+}
+
+int cpu_physical_log_stop(target_phys_addr_t start_addr,
+                          ram_addr_t size)
+{
+    CPUPhysMemoryClient *client;
+    QLIST_FOREACH(client, &memory_client_list, list) {
+        if (client->log_stop) {
+            int r = client->log_stop(client, start_addr, size);
+            if (r < 0) {
+                return r;
+            }
+        }
+    }
+    return 0;
+}
+
 static inline void tlb_update_dirty(CPUTLBEntry *tlb_entry)
 {
     ram_addr_t ram_addr;
@@ -2084,7 +2119,7 @@ static inline void tlb_update_dirty(CPUTLBEntry *tlb_entry)
     if ((tlb_entry->addr_write & ~TARGET_PAGE_MASK) == IO_MEM_RAM) {
         p = (void *)(unsigned long)((tlb_entry->addr_write & TARGET_PAGE_MASK)
             + tlb_entry->addend);
-        ram_addr = qemu_ram_addr_from_host(p);
+        ram_addr = qemu_ram_addr_from_host_nofail(p);
         if (!cpu_physical_memory_is_dirty(ram_addr)) {
             tlb_entry->addr_write |= TLB_NOTDIRTY;
         }
@@ -2172,8 +2207,9 @@ void tlb_set_page(CPUState *env, target_ulong vaddr,
         pd = p->phys_offset;
     }
 #if defined(DEBUG_TLB)
-    printf("tlb_set_page: vaddr=" TARGET_FMT_lx " paddr=0x%08x prot=%x idx=%d smmu=%d pd=0x%08lx\n",
-           vaddr, (int)paddr, prot, mmu_idx, is_softmmu, pd);
+    printf("tlb_set_page: vaddr=" TARGET_FMT_lx " paddr=0x" TARGET_FMT_plx
+           " prot=%x idx=%d pd=0x%08lx\n",
+           vaddr, paddr, prot, mmu_idx, pd);
 #endif
 
     address = vaddr;
@@ -2686,21 +2722,23 @@ static long gethugepagesize(const char *path)
     int ret;
 
     do {
-           ret = statfs(path, &fs);
+        ret = statfs(path, &fs);
     } while (ret != 0 && errno == EINTR);
 
     if (ret != 0) {
-           perror(path);
-           return 0;
+        perror(path);
+        return 0;
     }
 
     if (fs.f_type != HUGETLBFS_MAGIC)
-           fprintf(stderr, "Warning: path not on HugeTLBFS: %s\n", path);
+        fprintf(stderr, "Warning: path not on HugeTLBFS: %s\n", path);
 
     return fs.f_bsize;
 }
 
-static void *file_ram_alloc(ram_addr_t memory, const char *path)
+static void *file_ram_alloc(RAMBlock *block,
+                            ram_addr_t memory,
+                            const char *path)
 {
     char *filename;
     void *area;
@@ -2712,7 +2750,7 @@ static void *file_ram_alloc(ram_addr_t memory, const char *path)
 
     hpagesize = gethugepagesize(path);
     if (!hpagesize) {
-       return NULL;
+        return NULL;
     }
 
     if (memory < hpagesize) {
@@ -2725,14 +2763,14 @@ static void *file_ram_alloc(ram_addr_t memory, const char *path)
     }
 
     if (asprintf(&filename, "%s/qemu_back_mem.XXXXXX", path) == -1) {
-       return NULL;
+        return NULL;
     }
 
     fd = mkstemp(filename);
     if (fd < 0) {
-       perror("unable to create backing store for hugepages");
-       free(filename);
-       return NULL;
+        perror("unable to create backing store for hugepages");
+        free(filename);
+        return NULL;
     }
     unlink(filename);
     free(filename);
@@ -2746,7 +2784,7 @@ static void *file_ram_alloc(ram_addr_t memory, const char *path)
      * mmap will fail.
      */
     if (ftruncate(fd, memory))
-       perror("ftruncate");
+        perror("ftruncate");
 
 #ifdef MAP_POPULATE
     /* NB: MAP_POPULATE won't exhaustively alloc all phys pages in the case
@@ -2759,69 +2797,219 @@ static void *file_ram_alloc(ram_addr_t memory, const char *path)
     area = mmap(0, memory, PROT_READ | PROT_WRITE, MAP_PRIVATE, fd, 0);
 #endif
     if (area == MAP_FAILED) {
-       perror("file_ram_alloc: can't mmap RAM pages");
-       close(fd);
-       return (NULL);
+        perror("file_ram_alloc: can't mmap RAM pages");
+        close(fd);
+        return (NULL);
     }
+    block->fd = fd;
     return area;
 }
 #endif
 
-ram_addr_t qemu_ram_alloc(ram_addr_t size)
+static ram_addr_t find_ram_offset(ram_addr_t size)
 {
-    RAMBlock *new_block;
+    RAMBlock *block, *next_block;
+    ram_addr_t offset = 0, mingap = ULONG_MAX;
+
+    if (QLIST_EMPTY(&ram_list.blocks))
+        return 0;
+
+    QLIST_FOREACH(block, &ram_list.blocks, next) {
+        ram_addr_t end, next = ULONG_MAX;
+
+        end = block->offset + block->length;
+
+        QLIST_FOREACH(next_block, &ram_list.blocks, next) {
+            if (next_block->offset >= end) {
+                next = MIN(next, next_block->offset);
+            }
+        }
+        if (next - end >= size && next - end < mingap) {
+            offset =  end;
+            mingap = next - end;
+        }
+    }
+    return offset;
+}
+
+static ram_addr_t last_ram_offset(void)
+{
+    RAMBlock *block;
+    ram_addr_t last = 0;
+
+    QLIST_FOREACH(block, &ram_list.blocks, next)
+        last = MAX(last, block->offset + block->length);
+
+    return last;
+}
+
+ram_addr_t qemu_ram_alloc_from_ptr(DeviceState *dev, const char *name,
+                                   ram_addr_t size, void *host)
+{
+    RAMBlock *new_block, *block;
 
     size = TARGET_PAGE_ALIGN(size);
-    new_block = qemu_malloc(sizeof(*new_block));
+    new_block = qemu_mallocz(sizeof(*new_block));
 
-    if (mem_path) {
-#if defined (__linux__) && !defined(TARGET_S390X)
-        new_block->host = file_ram_alloc(size, mem_path);
-        if (!new_block->host) {
-            new_block->host = qemu_vmalloc(size);
-#ifdef MADV_MERGEABLE
-            madvise(new_block->host, size, MADV_MERGEABLE);
-#endif
+    if (dev && dev->parent_bus && dev->parent_bus->info->get_dev_path) {
+        char *id = dev->parent_bus->info->get_dev_path(dev);
+        if (id) {
+            snprintf(new_block->idstr, sizeof(new_block->idstr), "%s/", id);
+            qemu_free(id);
+        }
+    }
+    pstrcat(new_block->idstr, sizeof(new_block->idstr), name);
+
+    QLIST_FOREACH(block, &ram_list.blocks, next) {
+        if (!strcmp(block->idstr, new_block->idstr)) {
+            fprintf(stderr, "RAMBlock \"%s\" already registered, abort!\n",
+                    new_block->idstr);
+            abort();
         }
+    }
+
+    if (host) {
+        new_block->host = host;
+        new_block->flags |= RAM_PREALLOC_MASK;
+    } else {
+        if (mem_path) {
+#if defined (__linux__) && !defined(TARGET_S390X)
+            new_block->host = file_ram_alloc(new_block, size, mem_path);
+            if (!new_block->host) {
+                new_block->host = qemu_vmalloc(size);
+                qemu_madvise(new_block->host, size, QEMU_MADV_MERGEABLE);
+            }
 #else
-        fprintf(stderr, "-mem-path option unsupported\n");
-        exit(1);
+            fprintf(stderr, "-mem-path option unsupported\n");
+            exit(1);
 #endif
-    } else {
+        } else {
 #if defined(TARGET_S390X) && defined(CONFIG_KVM)
-        /* XXX S390 KVM requires the topmost vma of the RAM to be < 256GB */
-        new_block->host = mmap((void*)0x1000000, size,
-                                PROT_EXEC|PROT_READ|PROT_WRITE,
-                                MAP_SHARED | MAP_ANONYMOUS, -1, 0);
+            /* XXX S390 KVM requires the topmost vma of the RAM to be < 256GB */
+            new_block->host = mmap((void*)0x1000000, size,
+                                   PROT_EXEC|PROT_READ|PROT_WRITE,
+                                   MAP_SHARED | MAP_ANONYMOUS, -1, 0);
 #else
-        new_block->host = qemu_vmalloc(size);
-#endif
-#ifdef MADV_MERGEABLE
-        madvise(new_block->host, size, MADV_MERGEABLE);
+            new_block->host = qemu_vmalloc(size);
 #endif
+            qemu_madvise(new_block->host, size, QEMU_MADV_MERGEABLE);
+        }
     }
-    new_block->offset = ram_list.last_offset;
+
+    new_block->offset = find_ram_offset(size);
     new_block->length = size;
 
     QLIST_INSERT_HEAD(&ram_list.blocks, new_block, next);
 
     ram_list.phys_dirty = qemu_realloc(ram_list.phys_dirty,
-        (ram_list.last_offset + size) >> TARGET_PAGE_BITS);
-    memset(ram_list.phys_dirty + (ram_list.last_offset >> TARGET_PAGE_BITS),
+                                       last_ram_offset() >> TARGET_PAGE_BITS);
+    memset(ram_list.phys_dirty + (new_block->offset >> TARGET_PAGE_BITS),
            0xff, size >> TARGET_PAGE_BITS);
 
-    ram_list.last_offset += size;
-
     if (kvm_enabled())
         kvm_setup_guest_memory(new_block->host, size);
 
     return new_block->offset;
 }
 
+ram_addr_t qemu_ram_alloc(DeviceState *dev, const char *name, ram_addr_t size)
+{
+    return qemu_ram_alloc_from_ptr(dev, name, size, NULL);
+}
+
 void qemu_ram_free(ram_addr_t addr)
 {
-    /* TODO: implement this.  */
+    RAMBlock *block;
+
+    QLIST_FOREACH(block, &ram_list.blocks, next) {
+        if (addr == block->offset) {
+            QLIST_REMOVE(block, next);
+            if (block->flags & RAM_PREALLOC_MASK) {
+                ;
+            } else if (mem_path) {
+#if defined (__linux__) && !defined(TARGET_S390X)
+                if (block->fd) {
+                    munmap(block->host, block->length);
+                    close(block->fd);
+                } else {
+                    qemu_vfree(block->host);
+                }
+#else
+                abort();
+#endif
+            } else {
+#if defined(TARGET_S390X) && defined(CONFIG_KVM)
+                munmap(block->host, block->length);
+#else
+                qemu_vfree(block->host);
+#endif
+            }
+            qemu_free(block);
+            return;
+        }
+    }
+
+}
+
+#ifndef _WIN32
+void qemu_ram_remap(ram_addr_t addr, ram_addr_t length)
+{
+    RAMBlock *block;
+    ram_addr_t offset;
+    int flags;
+    void *area, *vaddr;
+
+    QLIST_FOREACH(block, &ram_list.blocks, next) {
+        offset = addr - block->offset;
+        if (offset < block->length) {
+            vaddr = block->host + offset;
+            if (block->flags & RAM_PREALLOC_MASK) {
+                ;
+            } else {
+                flags = MAP_FIXED;
+                munmap(vaddr, length);
+                if (mem_path) {
+#if defined(__linux__) && !defined(TARGET_S390X)
+                    if (block->fd) {
+#ifdef MAP_POPULATE
+                        flags |= mem_prealloc ? MAP_POPULATE | MAP_SHARED :
+                            MAP_PRIVATE;
+#else
+                        flags |= MAP_PRIVATE;
+#endif
+                        area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
+                                    flags, block->fd, offset);
+                    } else {
+                        flags |= MAP_PRIVATE | MAP_ANONYMOUS;
+                        area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
+                                    flags, -1, 0);
+                    }
+#else
+                    abort();
+#endif
+                } else {
+#if defined(TARGET_S390X) && defined(CONFIG_KVM)
+                    flags |= MAP_SHARED | MAP_ANONYMOUS;
+                    area = mmap(vaddr, length, PROT_EXEC|PROT_READ|PROT_WRITE,
+                                flags, -1, 0);
+#else
+                    flags |= MAP_PRIVATE | MAP_ANONYMOUS;
+                    area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
+                                flags, -1, 0);
+#endif
+                }
+                if (area != vaddr) {
+                    fprintf(stderr, "Could not remap addr: %lx@%lx\n",
+                            length, addr);
+                    exit(1);
+                }
+                qemu_madvise(vaddr, length, QEMU_MADV_MERGEABLE);
+            }
+            return;
+        }
+    }
 }
+#endif /* !_WIN32 */
 
 /* Return a host pointer to ram allocated with qemu_ram_alloc.
    With the exception of the softmmu code in this file, this should
@@ -2837,8 +3025,11 @@ void *qemu_get_ram_ptr(ram_addr_t addr)
 
     QLIST_FOREACH(block, &ram_list.blocks, next) {
         if (addr - block->offset < block->length) {
-            QLIST_REMOVE(block, next);
-            QLIST_INSERT_HEAD(&ram_list.blocks, block, next);
+            /* Move this entry to to start of the list.  */
+            if (block != QLIST_FIRST(&ram_list.blocks)) {
+                QLIST_REMOVE(block, next);
+                QLIST_INSERT_HEAD(&ram_list.blocks, block, next);
+            }
             return block->host + (addr - block->offset);
         }
     }
@@ -2849,23 +3040,50 @@ void *qemu_get_ram_ptr(ram_addr_t addr)
     return NULL;
 }
 
-/* Some of the softmmu routines need to translate from a host pointer
-   (typically a TLB entry) back to a ram offset.  */
-ram_addr_t qemu_ram_addr_from_host(void *ptr)
+/* Return a host pointer to ram allocated with qemu_ram_alloc.
+ * Same as qemu_get_ram_ptr but avoid reordering ramblocks.
+ */
+void *qemu_safe_ram_ptr(ram_addr_t addr)
+{
+    RAMBlock *block;
+
+    QLIST_FOREACH(block, &ram_list.blocks, next) {
+        if (addr - block->offset < block->length) {
+            return block->host + (addr - block->offset);
+        }
+    }
+
+    fprintf(stderr, "Bad ram offset %" PRIx64 "\n", (uint64_t)addr);
+    abort();
+
+    return NULL;
+}
+
+int qemu_ram_addr_from_host(void *ptr, ram_addr_t *ram_addr)
 {
     RAMBlock *block;
     uint8_t *host = ptr;
 
     QLIST_FOREACH(block, &ram_list.blocks, next) {
         if (host - block->host < block->length) {
-            return block->offset + (host - block->host);
+            *ram_addr = block->offset + (host - block->host);
+            return 0;
         }
     }
+    return -1;
+}
 
-    fprintf(stderr, "Bad ram pointer %p\n", ptr);
-    abort();
+/* Some of the softmmu routines need to translate from a host pointer
+   (typically a TLB entry) back to a ram offset.  */
+ram_addr_t qemu_ram_addr_from_host_nofail(void *ptr)
+{
+    ram_addr_t ram_addr;
 
-    return 0;
+    if (qemu_ram_addr_from_host(ptr, &ram_addr)) {
+        fprintf(stderr, "Bad ram pointer %p\n", ptr);
+        abort();
+    }
+    return ram_addr;
 }
 
 static uint32_t unassigned_mem_readb(void *opaque, target_phys_addr_t addr)
@@ -3201,6 +3419,8 @@ static int subpage_register (subpage_t *mmio, uint32_t start, uint32_t end,
     printf("%s: %p start %08x end %08x idx %08x eidx %08x mem %ld\n", __func__,
            mmio, start, end, idx, eidx, memory);
 #endif
+    if ((memory & ~TARGET_PAGE_MASK) == IO_MEM_RAM)
+        memory = IO_MEM_UNASSIGNED;
     memory = (memory >> IO_MEM_SHIFT) & (IO_MEM_NB_ENTRIES - 1);
     for (; idx <= eidx; idx++) {
         mmio->sub_io_index[idx] = memory;
@@ -3220,7 +3440,8 @@ static subpage_t *subpage_init (target_phys_addr_t base, ram_addr_t *phys,
     mmio = qemu_mallocz(sizeof(subpage_t));
 
     mmio->base = base;
-    subpage_memory = cpu_register_io_memory(subpage_read, subpage_write, mmio);
+    subpage_memory = cpu_register_io_memory(subpage_read, subpage_write, mmio,
+                                            DEVICE_NATIVE_ENDIAN);
 #if defined(DEBUG_SUBPAGE)
     printf("%s: %p base " TARGET_FMT_plx " len %08x %d\n", __func__,
            mmio, base, TARGET_PAGE_SIZE, subpage_memory);
@@ -3244,6 +3465,106 @@ static int get_free_io_mem_idx(void)
     return -1;
 }
 
+/*
+ * Usually, devices operate in little endian mode. There are devices out
+ * there that operate in big endian too. Each device gets byte swapped
+ * mmio if plugged onto a CPU that does the other endianness.
+ *
+ * CPU          Device           swap?
+ *
+ * little       little           no
+ * little       big              yes
+ * big          little           yes
+ * big          big              no
+ */
+
+typedef struct SwapEndianContainer {
+    CPUReadMemoryFunc *read[3];
+    CPUWriteMemoryFunc *write[3];
+    void *opaque;
+} SwapEndianContainer;
+
+static uint32_t swapendian_mem_readb (void *opaque, target_phys_addr_t addr)
+{
+    uint32_t val;
+    SwapEndianContainer *c = opaque;
+    val = c->read[0](c->opaque, addr);
+    return val;
+}
+
+static uint32_t swapendian_mem_readw(void *opaque, target_phys_addr_t addr)
+{
+    uint32_t val;
+    SwapEndianContainer *c = opaque;
+    val = bswap16(c->read[1](c->opaque, addr));
+    return val;
+}
+
+static uint32_t swapendian_mem_readl(void *opaque, target_phys_addr_t addr)
+{
+    uint32_t val;
+    SwapEndianContainer *c = opaque;
+    val = bswap32(c->read[2](c->opaque, addr));
+    return val;
+}
+
+static CPUReadMemoryFunc * const swapendian_readfn[3]={
+    swapendian_mem_readb,
+    swapendian_mem_readw,
+    swapendian_mem_readl
+};
+
+static void swapendian_mem_writeb(void *opaque, target_phys_addr_t addr,
+                                  uint32_t val)
+{
+    SwapEndianContainer *c = opaque;
+    c->write[0](c->opaque, addr, val);
+}
+
+static void swapendian_mem_writew(void *opaque, target_phys_addr_t addr,
+                                  uint32_t val)
+{
+    SwapEndianContainer *c = opaque;
+    c->write[1](c->opaque, addr, bswap16(val));
+}
+
+static void swapendian_mem_writel(void *opaque, target_phys_addr_t addr,
+                                  uint32_t val)
+{
+    SwapEndianContainer *c = opaque;
+    c->write[2](c->opaque, addr, bswap32(val));
+}
+
+static CPUWriteMemoryFunc * const swapendian_writefn[3]={
+    swapendian_mem_writeb,
+    swapendian_mem_writew,
+    swapendian_mem_writel
+};
+
+static void swapendian_init(int io_index)
+{
+    SwapEndianContainer *c = qemu_malloc(sizeof(SwapEndianContainer));
+    int i;
+
+    /* Swap mmio for big endian targets */
+    c->opaque = io_mem_opaque[io_index];
+    for (i = 0; i < 3; i++) {
+        c->read[i] = io_mem_read[io_index][i];
+        c->write[i] = io_mem_write[io_index][i];
+
+        io_mem_read[io_index][i] = swapendian_readfn[i];
+        io_mem_write[io_index][i] = swapendian_writefn[i];
+    }
+    io_mem_opaque[io_index] = c;
+}
+
+static void swapendian_del(int io_index)
+{
+    if (io_mem_read[io_index][0] == swapendian_readfn[0]) {
+        qemu_free(io_mem_opaque[io_index]);
+    }
+}
+
 /* mem_read and mem_write are arrays of functions containing the
    function to access byte (index 0), word (index 1) and dword (index
    2). Functions can be omitted with a NULL function pointer.
@@ -3254,7 +3575,7 @@ static int get_free_io_mem_idx(void)
 static int cpu_register_io_memory_fixed(int io_index,
                                         CPUReadMemoryFunc * const *mem_read,
                                         CPUWriteMemoryFunc * const *mem_write,
-                                        void *opaque)
+                                        void *opaque, enum device_endian endian)
 {
     int i;
 
@@ -3278,14 +3599,30 @@ static int cpu_register_io_memory_fixed(int io_index,
     }
     io_mem_opaque[io_index] = opaque;
 
+    switch (endian) {
+    case DEVICE_BIG_ENDIAN:
+#ifndef TARGET_WORDS_BIGENDIAN
+        swapendian_init(io_index);
+#endif
+        break;
+    case DEVICE_LITTLE_ENDIAN:
+#ifdef TARGET_WORDS_BIGENDIAN
+        swapendian_init(io_index);
+#endif
+        break;
+    case DEVICE_NATIVE_ENDIAN:
+    default:
+        break;
+    }
+
     return (io_index << IO_MEM_SHIFT);
 }
 
 int cpu_register_io_memory(CPUReadMemoryFunc * const *mem_read,
                            CPUWriteMemoryFunc * const *mem_write,
-                           void *opaque)
+                           void *opaque, enum device_endian endian)
 {
-    return cpu_register_io_memory_fixed(0, mem_read, mem_write, opaque);
+    return cpu_register_io_memory_fixed(0, mem_read, mem_write, opaque, endian);
 }
 
 void cpu_unregister_io_memory(int io_table_address)
@@ -3293,6 +3630,8 @@ void cpu_unregister_io_memory(int io_table_address)
     int i;
     int io_index = io_table_address >> IO_MEM_SHIFT;
 
+    swapendian_del(io_index);
+
     for (i=0;i < 3; i++) {
         io_mem_read[io_index][i] = unassigned_mem_read[i];
         io_mem_write[io_index][i] = unassigned_mem_write[i];
@@ -3305,14 +3644,21 @@ static void io_mem_init(void)
 {
     int i;
 
-    cpu_register_io_memory_fixed(IO_MEM_ROM, error_mem_read, unassigned_mem_write, NULL);
-    cpu_register_io_memory_fixed(IO_MEM_UNASSIGNED, unassigned_mem_read, unassigned_mem_write, NULL);
-    cpu_register_io_memory_fixed(IO_MEM_NOTDIRTY, error_mem_read, notdirty_mem_write, NULL);
+    cpu_register_io_memory_fixed(IO_MEM_ROM, error_mem_read,
+                                 unassigned_mem_write, NULL,
+                                 DEVICE_NATIVE_ENDIAN);
+    cpu_register_io_memory_fixed(IO_MEM_UNASSIGNED, unassigned_mem_read,
+                                 unassigned_mem_write, NULL,
+                                 DEVICE_NATIVE_ENDIAN);
+    cpu_register_io_memory_fixed(IO_MEM_NOTDIRTY, error_mem_read,
+                                 notdirty_mem_write, NULL,
+                                 DEVICE_NATIVE_ENDIAN);
     for (i=0; i<5; i++)
         io_mem_used[i] = 1;
 
     io_mem_watch = cpu_register_io_memory(watch_mem_read,
-                                          watch_mem_write, NULL);
+                                          watch_mem_write, NULL,
+                                          DEVICE_NATIVE_ENDIAN);
 }
 
 #endif /* !defined(CONFIG_USER_ONLY) */
@@ -3612,7 +3958,7 @@ void cpu_physical_memory_unmap(void *buffer, target_phys_addr_t len,
 {
     if (buffer != bounce.buffer) {
         if (is_write) {
-            ram_addr_t addr1 = qemu_ram_addr_from_host(buffer);
+            ram_addr_t addr1 = qemu_ram_addr_from_host_nofail(buffer);
             while (access_len) {
                 unsigned l;
                 l = TARGET_PAGE_SIZE;
@@ -3997,8 +4343,7 @@ void cpu_io_recompile(CPUState *env, void *retaddr)
 
 #if !defined(CONFIG_USER_ONLY)
 
-void dump_exec_info(FILE *f,
-                    int (*cpu_fprintf)(FILE *f, const char *fmt, ...))
+void dump_exec_info(FILE *f, fprintf_function cpu_fprintf)
 {
     int i, target_code_size, max_target_code_size;
     int direct_jmp_count, direct_jmp2_count, cross_page;
@@ -4025,14 +4370,14 @@ void dump_exec_info(FILE *f,
     }
     /* XXX: avoid using doubles ? */
     cpu_fprintf(f, "Translation buffer state:\n");
-    cpu_fprintf(f, "gen code size       %ld/%ld\n",
+    cpu_fprintf(f, "gen code size       %td/%ld\n",
                 code_gen_ptr - code_gen_buffer, code_gen_buffer_max_size);
     cpu_fprintf(f, "TB count            %d/%d\n", 
                 nb_tbs, code_gen_max_blocks);
     cpu_fprintf(f, "TB avg target size  %d max=%d bytes\n",
                 nb_tbs ? target_code_size / nb_tbs : 0,
                 max_target_code_size);
-    cpu_fprintf(f, "TB avg host size    %d bytes (expansion ratio: %0.1f)\n",
+    cpu_fprintf(f, "TB avg host size    %td bytes (expansion ratio: %0.1f)\n",
                 nb_tbs ? (code_gen_ptr - code_gen_buffer) / nb_tbs : 0,
                 target_code_size ? (double) (code_gen_ptr - code_gen_buffer) / target_code_size : 0);
     cpu_fprintf(f, "cross page TB count %d (%d%%)\n",