#include "mono-codeman.h"
#include "mono-mmap.h"
+#include "mono-counters.h"
#include "dlmalloc.h"
-#include <mono/metadata/class-internals.h>
+#include <mono/io-layer/io-layer.h>
#include <mono/metadata/profiler-private.h>
#ifdef HAVE_VALGRIND_MEMCHECK_H
#include <valgrind/memcheck.h>
#endif
-
+
+#if defined(__native_client_codegen__) && defined(__native_client__)
+#include <malloc.h>
+#include <nacl/nacl_dyncode.h>
+#include <mono/mini/mini.h>
+#endif
+
+static uintptr_t code_memory_used = 0;
+static size_t dynamic_code_alloc_count;
+static size_t dynamic_code_bytes_count;
+static size_t dynamic_code_frees_count;
+
/*
* AMD64 processors maintain icache coherency only for pages which are
* marked executable. Also, windows DEP requires us to obtain executable memory from
#define MIN_PAGES 16
-#if defined(__ia64__) || defined(__x86_64__)
+#if defined(__ia64__) || defined(__x86_64__) || defined (_WIN64)
/*
* We require 16 byte alignment on amd64 so the fp literals embedded in the code are
* properly aligned for SSE2.
#define MIN_ALIGN 16
#else
#define MIN_ALIGN 8
+#endif
+#ifdef __native_client_codegen__
+/* For Google Native Client, all targets of indirect control flow need to */
+/* be aligned to bundle boundary. 16 bytes on ARM, 32 bytes on x86.
+ * MIN_ALIGN was updated to force alignment for calls from
+ * tramp-<arch>.c to mono_global_codeman_reserve() */
+/* and mono_domain_code_reserve(). */
+#undef MIN_ALIGN
+#define MIN_ALIGN kNaClBundleSize
+
#endif
/* if a chunk has less than this amount of free space it's considered full */
int read_only;
CodeChunk *current;
CodeChunk *full;
+ CodeChunk *last;
+#if defined(__native_client_codegen__) && defined(__native_client__)
+ GHashTable *hash;
+#endif
};
#define ALIGN_INT(val,alignment) (((val) + (alignment - 1)) & ~(alignment - 1))
+#if defined(__native_client_codegen__) && defined(__native_client__)
+/* End of text segment, set by linker.
+ * Dynamic text starts on the next allocated page.
+ */
+extern char etext[];
+char *next_dynamic_code_addr = NULL;
+
+/*
+ * This routine gets the next available bundle aligned
+ * pointer in the dynamic code section. It does not check
+ * for the section end, this error will be caught in the
+ * service runtime.
+ */
+void*
+allocate_code(intptr_t increment)
+{
+ char *addr;
+ if (increment < 0) return NULL;
+ increment = increment & kNaClBundleMask ? (increment & ~kNaClBundleMask) + kNaClBundleSize : increment;
+ addr = next_dynamic_code_addr;
+ next_dynamic_code_addr += increment;
+ return addr;
+}
+
+int
+nacl_is_code_address (void *target)
+{
+ return (char *)target < next_dynamic_code_addr;
+}
+
+/* Fill code buffer with arch-specific NOPs. */
+void
+mono_nacl_fill_code_buffer (guint8 *data, int size);
+
+#ifndef USE_JUMP_TABLES
+const int kMaxPatchDepth = 32;
+__thread unsigned char **patch_source_base = NULL;
+__thread unsigned char **patch_dest_base = NULL;
+__thread int *patch_alloc_size = NULL;
+__thread int patch_current_depth = -1;
+__thread int allow_target_modification = 1;
+
+static void
+nacl_jit_check_init ()
+{
+ if (patch_source_base == NULL) {
+ patch_source_base = g_malloc (kMaxPatchDepth * sizeof(unsigned char *));
+ patch_dest_base = g_malloc (kMaxPatchDepth * sizeof(unsigned char *));
+ patch_alloc_size = g_malloc (kMaxPatchDepth * sizeof(int));
+ }
+}
+#endif
+
+void
+nacl_allow_target_modification (int val)
+{
+#ifndef USE_JUMP_TABLES
+ allow_target_modification = val;
+#endif /* USE_JUMP_TABLES */
+}
+
+/* Given a patch target, modify the target such that patching will work when
+ * the code is copied to the data section.
+ */
+void*
+nacl_modify_patch_target (unsigned char *target)
+{
+ /*
+ * There's no need in patch tricks for jumptables,
+ * as we always patch same jumptable.
+ */
+#ifndef USE_JUMP_TABLES
+ /* This seems like a bit of an ugly way to do this but the advantage
+ * is we don't have to worry about all the conditions in
+ * mono_resolve_patch_target, and it can be used by all the bare uses
+ * of <arch>_patch.
+ */
+ unsigned char *sb;
+ unsigned char *db;
+
+ if (!allow_target_modification) return target;
+
+ nacl_jit_check_init ();
+ sb = patch_source_base[patch_current_depth];
+ db = patch_dest_base[patch_current_depth];
+
+ if (target >= sb && (target < sb + patch_alloc_size[patch_current_depth])) {
+ /* Do nothing. target is in the section being generated.
+ * no need to modify, the disp will be the same either way.
+ */
+ } else {
+ int target_offset = target - db;
+ target = sb + target_offset;
+ }
+#endif
+ return target;
+}
+
+void*
+nacl_inverse_modify_patch_target (unsigned char *target)
+{
+ /*
+ * There's no need in patch tricks for jumptables,
+ * as we always patch same jumptable.
+ */
+#ifndef USE_JUMP_TABLES
+ unsigned char *sb;
+ unsigned char *db;
+ int target_offset;
+
+ if (!allow_target_modification) return target;
+
+ nacl_jit_check_init ();
+ sb = patch_source_base[patch_current_depth];
+ db = patch_dest_base[patch_current_depth];
+
+ target_offset = target - sb;
+ target = db + target_offset;
+#endif
+ return target;
+}
+
+
+#endif /* __native_client_codegen && __native_client__ */
+
+#define VALLOC_FREELIST_SIZE 16
+
+static CRITICAL_SECTION valloc_mutex;
+static GHashTable *valloc_freelists;
+
+static void*
+codechunk_valloc (void *preferred, guint32 size)
+{
+ void *ptr;
+ GSList *freelist;
+
+ if (!valloc_freelists) {
+ InitializeCriticalSection (&valloc_mutex);
+ valloc_freelists = g_hash_table_new (NULL, NULL);
+ }
+
+ /*
+ * Keep a small freelist of memory blocks to decrease pressure on the kernel memory subsystem to avoid #3321.
+ */
+ EnterCriticalSection (&valloc_mutex);
+ freelist = g_hash_table_lookup (valloc_freelists, GUINT_TO_POINTER (size));
+ if (freelist) {
+ ptr = freelist->data;
+ memset (ptr, 0, size);
+ freelist = g_slist_delete_link (freelist, freelist);
+ g_hash_table_insert (valloc_freelists, GUINT_TO_POINTER (size), freelist);
+ } else {
+ ptr = mono_valloc (preferred, size, MONO_PROT_RWX | ARCH_MAP_FLAGS);
+ if (!ptr && preferred)
+ ptr = mono_valloc (NULL, size, MONO_PROT_RWX | ARCH_MAP_FLAGS);
+ }
+ LeaveCriticalSection (&valloc_mutex);
+ return ptr;
+}
+
+static void
+codechunk_vfree (void *ptr, guint32 size)
+{
+ GSList *freelist;
+
+ EnterCriticalSection (&valloc_mutex);
+ freelist = g_hash_table_lookup (valloc_freelists, GUINT_TO_POINTER (size));
+ if (!freelist || g_slist_length (freelist) < VALLOC_FREELIST_SIZE) {
+ freelist = g_slist_prepend (freelist, ptr);
+ g_hash_table_insert (valloc_freelists, GUINT_TO_POINTER (size), freelist);
+ } else {
+ mono_vfree (ptr, size);
+ }
+ LeaveCriticalSection (&valloc_mutex);
+}
+
+static void
+codechunk_cleanup (void)
+{
+ GHashTableIter iter;
+ gpointer key, value;
+
+ if (!valloc_freelists)
+ return;
+ g_hash_table_iter_init (&iter, valloc_freelists);
+ while (g_hash_table_iter_next (&iter, &key, &value)) {
+ GSList *freelist = value;
+ GSList *l;
+
+ for (l = freelist; l; l = l->next) {
+ mono_vfree (l->data, GPOINTER_TO_UINT (key));
+ }
+ g_slist_free (freelist);
+ }
+ g_hash_table_destroy (valloc_freelists);
+}
+
+void
+mono_code_manager_init (void)
+{
+ mono_counters_register ("Dynamic code allocs", MONO_COUNTER_JIT | MONO_COUNTER_WORD, &dynamic_code_alloc_count);
+ mono_counters_register ("Dynamic code bytes", MONO_COUNTER_JIT | MONO_COUNTER_WORD, &dynamic_code_bytes_count);
+ mono_counters_register ("Dynamic code frees", MONO_COUNTER_JIT | MONO_COUNTER_WORD, &dynamic_code_frees_count);
+}
+
+void
+mono_code_manager_cleanup (void)
+{
+ codechunk_cleanup ();
+}
+
/**
* mono_code_manager_new:
*
MonoCodeManager*
mono_code_manager_new (void)
{
- MonoCodeManager *cman = malloc (sizeof (MonoCodeManager));
+ MonoCodeManager *cman = g_malloc0 (sizeof (MonoCodeManager));
if (!cman)
return NULL;
- cman->current = NULL;
- cman->full = NULL;
- cman->dynamic = 0;
- cman->read_only = 0;
+#if defined(__native_client_codegen__) && defined(__native_client__)
+ if (next_dynamic_code_addr == NULL) {
+ const guint kPageMask = 0xFFFF; /* 64K pages */
+ next_dynamic_code_addr = (uintptr_t)(etext + kPageMask) & ~kPageMask;
+#if defined (__GLIBC__)
+ /* TODO: For now, just jump 64MB ahead to avoid dynamic libraries. */
+ next_dynamic_code_addr += (uintptr_t)0x4000000;
+#else
+ /* Workaround bug in service runtime, unable to allocate */
+ /* from the first page in the dynamic code section. */
+ next_dynamic_code_addr += (uintptr_t)0x10000;
+#endif
+ }
+ cman->hash = g_hash_table_new (NULL, NULL);
+# ifndef USE_JUMP_TABLES
+ if (patch_source_base == NULL) {
+ patch_source_base = g_malloc (kMaxPatchDepth * sizeof(unsigned char *));
+ patch_dest_base = g_malloc (kMaxPatchDepth * sizeof(unsigned char *));
+ patch_alloc_size = g_malloc (kMaxPatchDepth * sizeof(int));
+ }
+# endif
+#endif
return cman;
}
mono_profiler_code_chunk_destroy ((gpointer) dead->data);
chunk = chunk->next;
if (dead->flags == CODE_FLAG_MMAP) {
- mono_vfree (dead->data, dead->size);
+ codechunk_vfree (dead->data, dead->size);
/* valgrind_unregister(dead->data); */
} else if (dead->flags == CODE_FLAG_MALLOC) {
dlfree (dead->data);
}
+ code_memory_used -= dead->size;
free (dead);
}
}
#if defined(__arm__)
#define BIND_ROOM 8
#endif
+#if defined(TARGET_ARM64)
+#define BIND_ROOM 8
+#endif
static CodeChunk*
-new_codechunk (int dynamic, int size)
+new_codechunk (CodeChunk *last, int dynamic, int size)
{
int minsize, flags = CODE_FLAG_MMAP;
int chunk_size, bsize = 0;
if (size < minsize)
chunk_size = minsize;
else {
+ /* Allocate MIN_ALIGN-1 more than we need so we can still */
+ /* guarantee MIN_ALIGN alignment for individual allocs */
+ /* from mono_code_manager_reserve_align. */
+ size += MIN_ALIGN - 1;
+ size &= ~(MIN_ALIGN - 1);
chunk_size = size;
chunk_size += pagesize - 1;
chunk_size &= ~ (pagesize - 1);
if (!ptr)
return NULL;
} else {
- ptr = mono_valloc (NULL, chunk_size, MONO_PROT_RWX | ARCH_MAP_FLAGS);
+ /* Try to allocate code chunks next to each other to help the VM */
+ if (last)
+ ptr = codechunk_valloc ((guint8*)last->data + last->size, chunk_size);
+ else
+ ptr = codechunk_valloc (NULL, chunk_size);
if (!ptr)
return NULL;
}
chunk->bsize = bsize;
mono_profiler_code_chunk_new((gpointer) chunk->data, chunk->size);
+ code_memory_used += chunk_size;
+ mono_runtime_resource_check_limit (MONO_RESOURCE_JIT_CODE, code_memory_used);
/*printf ("code chunk at: %p\n", ptr);*/
return chunk;
}
void*
mono_code_manager_reserve_align (MonoCodeManager *cman, int size, int alignment)
{
+#if !defined(__native_client__) || !defined(__native_client_codegen__)
CodeChunk *chunk, *prev;
void *ptr;
+ guint32 align_mask = alignment - 1;
g_assert (!cman->read_only);
g_assert (alignment <= MIN_ALIGN);
if (cman->dynamic) {
- ++mono_stats.dynamic_code_alloc_count;
- mono_stats.dynamic_code_bytes_count += size;
+ ++dynamic_code_alloc_count;
+ dynamic_code_bytes_count += size;
}
if (!cman->current) {
- cman->current = new_codechunk (cman->dynamic, size);
+ cman->current = new_codechunk (cman->last, cman->dynamic, size);
if (!cman->current)
return NULL;
+ cman->last = cman->current;
}
for (chunk = cman->current; chunk; chunk = chunk->next) {
if (ALIGN_INT (chunk->pos, alignment) + size <= chunk->size) {
chunk->pos = ALIGN_INT (chunk->pos, alignment);
- ptr = chunk->data + chunk->pos;
- chunk->pos += size;
+ /* Align the chunk->data we add to chunk->pos */
+ /* or we can't guarantee proper alignment */
+ ptr = (void*)((((uintptr_t)chunk->data + align_mask) & ~(uintptr_t)align_mask) + chunk->pos);
+ chunk->pos = ((char*)ptr - chunk->data) + size;
return ptr;
}
}
cman->full = chunk;
break;
}
- chunk = new_codechunk (cman->dynamic, size);
+ chunk = new_codechunk (cman->last, cman->dynamic, size);
if (!chunk)
return NULL;
chunk->next = cman->current;
cman->current = chunk;
+ cman->last = cman->current;
chunk->pos = ALIGN_INT (chunk->pos, alignment);
- ptr = chunk->data + chunk->pos;
- chunk->pos += size;
+ /* Align the chunk->data we add to chunk->pos */
+ /* or we can't guarantee proper alignment */
+ ptr = (void*)((((uintptr_t)chunk->data + align_mask) & ~(uintptr_t)align_mask) + chunk->pos);
+ chunk->pos = ((char*)ptr - chunk->data) + size;
return ptr;
+#else
+ unsigned char *temp_ptr, *code_ptr;
+ /* Round up size to next bundle */
+ alignment = kNaClBundleSize;
+ size = (size + kNaClBundleSize) & (~kNaClBundleMask);
+ /* Allocate a temp buffer */
+ temp_ptr = memalign (alignment, size);
+ g_assert (((uintptr_t)temp_ptr & kNaClBundleMask) == 0);
+ /* Allocate code space from the service runtime */
+ code_ptr = allocate_code (size);
+ /* Insert pointer to code space in hash, keyed by buffer ptr */
+ g_hash_table_insert (cman->hash, temp_ptr, code_ptr);
+
+#ifndef USE_JUMP_TABLES
+ nacl_jit_check_init ();
+
+ patch_current_depth++;
+ patch_source_base[patch_current_depth] = temp_ptr;
+ patch_dest_base[patch_current_depth] = code_ptr;
+ patch_alloc_size[patch_current_depth] = size;
+ g_assert (patch_current_depth < kMaxPatchDepth);
+#endif
+
+ return temp_ptr;
+#endif
}
/**
void
mono_code_manager_commit (MonoCodeManager *cman, void *data, int size, int newsize)
{
+#if !defined(__native_client__) || !defined(__native_client_codegen__)
g_assert (newsize <= size);
if (cman->current && (size != newsize) && (data == cman->current->data + cman->current->pos - size)) {
cman->current->pos -= size - newsize;
}
+#else
+ unsigned char *code;
+ int status;
+ g_assert (NACL_BUNDLE_ALIGN_UP(newsize) <= size);
+ code = g_hash_table_lookup (cman->hash, data);
+ g_assert (code != NULL);
+ mono_nacl_fill_code_buffer ((uint8_t*)data + newsize, size - newsize);
+ newsize = NACL_BUNDLE_ALIGN_UP(newsize);
+ g_assert ((GPOINTER_TO_UINT (data) & kNaClBundleMask) == 0);
+ g_assert ((newsize & kNaClBundleMask) == 0);
+ status = nacl_dyncode_create (code, data, newsize);
+ if (status != 0) {
+ unsigned char *codep;
+ fprintf(stderr, "Error creating Native Client dynamic code section attempted to be\n"
+ "emitted at %p (hex dissasembly of code follows):\n", code);
+ for (codep = data; codep < data + newsize; codep++)
+ fprintf(stderr, "%02x ", *codep);
+ fprintf(stderr, "\n");
+ g_assert_not_reached ();
+ }
+ g_hash_table_remove (cman->hash, data);
+# ifndef USE_JUMP_TABLES
+ g_assert (data == patch_source_base[patch_current_depth]);
+ g_assert (code == patch_dest_base[patch_current_depth]);
+ patch_current_depth--;
+ g_assert (patch_current_depth >= -1);
+# endif
+ free (data);
+#endif
}
+#if defined(__native_client_codegen__) && defined(__native_client__)
+void *
+nacl_code_manager_get_code_dest (MonoCodeManager *cman, void *data)
+{
+ return g_hash_table_lookup (cman->hash, data);
+}
+#endif
+
/**
* mono_code_manager_size:
* @cman: a code manager
return size;
}
+#ifdef __native_client_codegen__
+# if defined(TARGET_ARM)
+/* Fill empty space with UDF instruction used as halt on ARM. */
+void
+mono_nacl_fill_code_buffer (guint8 *data, int size)
+{
+ guint32* data32 = (guint32*)data;
+ int i;
+ g_assert(size % 4 == 0);
+ for (i = 0; i < size / 4; i++)
+ data32[i] = 0xE7FEDEFF;
+}
+# elif (defined(TARGET_X86) || defined(TARGET_AMD64))
+/* Fill empty space with HLT instruction */
+void
+mono_nacl_fill_code_buffer(guint8 *data, int size)
+{
+ memset (data, 0xf4, size);
+}
+# else
+# error "Not ported"
+# endif
+#endif