Allow using mmap for shared memory if hw bound check is disabled (#3029)

For shared memory, the max memory size must be defined in advanced. Re-allocation
for growing memory can't be used as it might change the base address, therefore when
OS_ENABLE_HW_BOUND_CHECK is enabled the memory is mmaped, and if the flag is
disabled, the memory is allocated. This change introduces a flag that allows users to use
mmap for reserving memory address space even if the OS_ENABLE_HW_BOUND_CHECK
is disabled.

.github/workflows/compilation_on_android_ubuntu.yml

@@ -145,6 +145,7 @@ jobs:
             "-DWAMR_BUILD_SIMD=1",
             "-DWAMR_BUILD_TAIL_CALL=1",
             "-DWAMR_DISABLE_HW_BOUND_CHECK=1",
+            "-DWAMR_ENABLE_SHARED_MEMORY_MMAP=1",
           ]
         os: [ubuntu-22.04]
         platform: [android, linux]

build-scripts/config_common.cmake

@@ -248,6 +248,12 @@ if (WAMR_BUILD_SHARED_MEMORY EQUAL 1)
 else ()
   add_definitions (-DWASM_ENABLE_SHARED_MEMORY=0)
 endif ()
+if (WAMR_ENABLE_SHARED_MEMORY_MMAP EQUAL 1)
+  add_definitions (-DWASM_ENABLE_SHARED_MEMORY_MMAP=1)
+  message ("     Shared memory allocated using mmap enabled")
+else ()
+  add_definitions (-DWASM_ENABLE_SHARED_MEMORY_MMAP=0)
+endif ()
 if (WAMR_BUILD_THREAD_MGR EQUAL 1)
   message ("     Thread manager enabled")
 endif ()

core/iwasm/aot/aot_runtime.c

@@ -340,6 +340,9 @@ tables_instantiate(AOTModuleInstance *module_inst, AOTModule *module,
 static void
 memories_deinstantiate(AOTModuleInstance *module_inst)
 {
+#ifdef WASM_LINEAR_MEMORY_MMAP
+    uint64 map_size;
+#endif
     uint32 i;
     AOTMemoryInstance *memory_inst;
 
@@ -362,14 +365,21 @@ memories_deinstantiate(AOTModuleInstance *module_inst)
 
             if (memory_inst->memory_data) {
 #ifndef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
+                if (shared_memory_is_shared(memory_inst)) {
+                    map_size = (uint64)memory_inst->num_bytes_per_page
+                               * memory_inst->max_page_count;
+                    wasm_munmap_linear_memory(memory_inst->memory_data,
+                                              map_size, map_size);
+                }
+                else
+#endif
                     wasm_runtime_free(memory_inst->memory_data);
 #else
-#ifdef BH_PLATFORM_WINDOWS
+                map_size = (uint64)memory_inst->num_bytes_per_page
-                os_mem_decommit(memory_inst->memory_data,
+                           * memory_inst->cur_page_count;
-                                memory_inst->num_bytes_per_page
+                wasm_munmap_linear_memory(memory_inst->memory_data, map_size,
-                                    * memory_inst->cur_page_count);
+                                          8 * (uint64)BH_GB);
-#endif
-                os_munmap(memory_inst->memory_data, 8 * (uint64)BH_GB);
 #endif
             }
         }
@@ -392,10 +402,9 @@ memory_instantiate(AOTModuleInstance *module_inst, AOTModuleInstance *parent,
     uint32 heap_offset = num_bytes_per_page * init_page_count;
     uint64 memory_data_size, max_memory_data_size;
     uint8 *p = NULL, *global_addr;
-#ifdef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
-    uint8 *mapped_mem;
+    uint8 *mapped_mem = NULL;
-    uint64 map_size = 8 * (uint64)BH_GB;
+    uint64 map_size;
-    uint64 page_size = os_getpagesize();
 #endif
 
 #if WASM_ENABLE_SHARED_MEMORY != 0
@@ -519,15 +528,25 @@ memory_instantiate(AOTModuleInstance *module_inst, AOTModuleInstance *parent,
 #ifndef OS_ENABLE_HW_BOUND_CHECK
 #if WASM_ENABLE_SHARED_MEMORY != 0
     if (is_shared_memory) {
+#if WASM_ENABLE_SHARED_MEMORY_MMAP != 0
+        map_size = max_memory_data_size;
+        if (max_memory_data_size > 0
+            && !(p = mapped_mem =
+                     wasm_mmap_linear_memory(map_size, &max_memory_data_size,
+                                             error_buf, error_buf_size))) {
+            return NULL;
+        }
+#else
         /* Allocate maximum memory size when memory is shared */
         if (max_memory_data_size > 0
             && !(p = runtime_malloc(max_memory_data_size, error_buf,
                                     error_buf_size))) {
             return NULL;
         }
+#endif
     }
     else
-#endif
+#endif /* end of WASM_ENABLE_SHARED_MEMORY != 0 */
     {
         /* Allocate initial memory size when memory is not shared */
         if (memory_data_size > 0
@@ -537,42 +556,17 @@ memory_instantiate(AOTModuleInstance *module_inst, AOTModuleInstance *parent,
         }
     }
 #else  /* else of OS_ENABLE_HW_BOUND_CHECK */
-    memory_data_size = (memory_data_size + page_size - 1) & ~(page_size - 1);
-
     /* Totally 8G is mapped, the opcode load/store address range is 0 to 8G:
      *   ea = i + memarg.offset
      * both i and memarg.offset are u32 in range 0 to 4G
      * so the range of ea is 0 to 8G
      */
-    if (!(p = mapped_mem = os_mmap(NULL, map_size, MMAP_PROT_NONE,
+    map_size = 8 * (uint64)BH_GB;
-                                   MMAP_MAP_NONE, os_get_invalid_handle()))) {
+    if (!(p = mapped_mem = wasm_mmap_linear_memory(
+              map_size, &memory_data_size, error_buf, error_buf_size))) {
         set_error_buf(error_buf, error_buf_size, "mmap memory failed");
         return NULL;
     }
-
-#ifdef BH_PLATFORM_WINDOWS
-    if (!os_mem_commit(p, memory_data_size, MMAP_PROT_READ | MMAP_PROT_WRITE)) {
-        set_error_buf(error_buf, error_buf_size, "commit memory failed");
-        os_munmap(mapped_mem, map_size);
-        return NULL;
-    }
-#endif
-
-    if (os_mprotect(p, memory_data_size, MMAP_PROT_READ | MMAP_PROT_WRITE)
-        != 0) {
-        set_error_buf(error_buf, error_buf_size, "mprotect memory failed");
-#ifdef BH_PLATFORM_WINDOWS
-        os_mem_decommit(p, memory_data_size);
-#endif
-        os_munmap(mapped_mem, map_size);
-        return NULL;
-    }
-
-    /* Newly allocated pages are filled with zero by the OS, we don't fill it
-     * again here */
-
-    if (memory_data_size > UINT32_MAX)
-        memory_data_size = UINT32_MAX;
 #endif /* end of OS_ENABLE_HW_BOUND_CHECK */
 
     memory_inst->module_type = Wasm_Module_AoT;
@@ -623,16 +617,15 @@ fail2:
     if (heap_size > 0)
         wasm_runtime_free(memory_inst->heap_handle);
 fail1:
-#ifndef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
+    if (mapped_mem)
+        wasm_munmap_linear_memory(mapped_mem, memory_data_size, map_size);
+    else
+#endif
+    {
         if (memory_inst->memory_data)
             wasm_runtime_free(memory_inst->memory_data);
-#else
+    }
-#ifdef BH_PLATFORM_WINDOWS
-    if (memory_inst->memory_data)
-        os_mem_decommit(p, memory_data_size);
-#endif
-    os_munmap(mapped_mem, map_size);
-#endif
     memory_inst->memory_data = NULL;
     return NULL;
 }

core/iwasm/common/wasm_runtime_common.c

@@ -6229,3 +6229,67 @@ wasm_runtime_set_linux_perf(bool flag)
     enable_linux_perf = flag;
 }
 #endif
+
+#ifdef WASM_LINEAR_MEMORY_MMAP
+void
+wasm_munmap_linear_memory(void *mapped_mem, uint64 commit_size, uint64 map_size)
+{
+#ifdef BH_PLATFORM_WINDOWS
+    os_mem_decommit(mapped_mem, commit_size);
+#else
+    (void)commit_size;
+#endif
+    os_munmap(mapped_mem, map_size);
+}
+
+void *
+wasm_mmap_linear_memory(uint64_t map_size, uint64 *io_memory_data_size,
+                        char *error_buf, uint32 error_buf_size)
+{
+    uint64 page_size = os_getpagesize();
+    void *mapped_mem = NULL;
+    uint64 memory_data_size;
+
+    bh_assert(io_memory_data_size);
+
+    memory_data_size =
+        (*io_memory_data_size + page_size - 1) & ~(page_size - 1);
+
+    if (memory_data_size > UINT32_MAX)
+        memory_data_size = UINT32_MAX;
+
+    if (!(mapped_mem = os_mmap(NULL, map_size, MMAP_PROT_NONE, MMAP_MAP_NONE,
+                               os_get_invalid_handle()))) {
+        set_error_buf(error_buf, error_buf_size, "mmap memory failed");
+        goto fail1;
+    }
+
+#ifdef BH_PLATFORM_WINDOWS
+    if (memory_data_size > 0
+        && !os_mem_commit(mapped_mem, memory_data_size,
+                          MMAP_PROT_READ | MMAP_PROT_WRITE)) {
+        set_error_buf(error_buf, error_buf_size, "commit memory failed");
+        os_munmap(mapped_mem, map_size);
+        goto fail1;
+    }
+#endif
+
+    if (os_mprotect(mapped_mem, memory_data_size,
+                    MMAP_PROT_READ | MMAP_PROT_WRITE)
+        != 0) {
+        set_error_buf(error_buf, error_buf_size, "mprotect memory failed");
+        goto fail2;
+    }
+
+    /* Newly allocated pages are filled with zero by the OS, we don't fill it
+     * again here */
+
+    *io_memory_data_size = memory_data_size;
+
+    return mapped_mem;
+fail2:
+    wasm_munmap_linear_memory(mapped_mem, memory_data_size, map_size);
+fail1:
+    return NULL;
+}
+#endif

core/iwasm/common/wasm_runtime_common.h

@@ -320,6 +320,11 @@ LOAD_I16(void *addr)
 #define SHARED_MEMORY_UNLOCK(memory) (void)0
 #endif
 
+#if defined(OS_ENABLE_HW_BOUND_CHECK) \
+    || (WASM_ENABLE_SHARED_MEMORY != 0 && WASM_ENABLE_SHARED_MEMORY_MMAP != 0)
+#define WASM_LINEAR_MEMORY_MMAP
+#endif
+
 typedef struct WASMModuleCommon {
     /* Module type, for module loaded from WASM bytecode binary,
        this field is Wasm_Module_Bytecode, and this structure should
@@ -1093,6 +1098,14 @@ wasm_runtime_quick_invoke_c_api_native(WASMModuleInstanceCommon *module_inst,
 void
 wasm_runtime_show_app_heap_corrupted_prompt();
 
+void
+wasm_munmap_linear_memory(void *mapped_mem, uint64 commit_size,
+                          uint64 map_size);
+
+void *
+wasm_mmap_linear_memory(uint64_t map_size, uint64 *io_memory_data_size,
+                        char *error_buf, uint32 error_buf_size);
+
 #if WASM_ENABLE_LOAD_CUSTOM_SECTION != 0
 void
 wasm_runtime_destroy_custom_sections(WASMCustomSection *section_list);

core/iwasm/interpreter/wasm_runtime.c

@@ -115,6 +115,9 @@ static void
 memories_deinstantiate(WASMModuleInstance *module_inst,
                        WASMMemoryInstance **memories, uint32 count)
 {
+#ifdef WASM_LINEAR_MEMORY_MMAP
+    uint64 map_size;
+#endif
     uint32 i;
     if (memories) {
         for (i = 0; i < count; i++) {
@@ -142,15 +145,21 @@ memories_deinstantiate(WASMModuleInstance *module_inst,
                 }
                 if (memories[i]->memory_data) {
 #ifndef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
+                    if (shared_memory_is_shared(memories[i])) {
+                        map_size = (uint64)memories[i]->num_bytes_per_page
+                                   * memories[i]->max_page_count;
+                        wasm_munmap_linear_memory(memories[i]->memory_data,
+                                                  map_size, map_size);
+                    }
+                    else
+#endif
                         wasm_runtime_free(memories[i]->memory_data);
 #else
-#ifdef BH_PLATFORM_WINDOWS
+                    map_size = (uint64)memories[i]->num_bytes_per_page
-                    os_mem_decommit(memories[i]->memory_data,
+                               * memories[i]->cur_page_count;
-                                    memories[i]->num_bytes_per_page
+                    wasm_munmap_linear_memory(memories[i]->memory_data,
-                                        * memories[i]->cur_page_count);
+                                              map_size, 8 * (uint64)BH_GB);
-#endif
-                    os_munmap((uint8 *)memories[i]->memory_data,
-                              8 * (uint64)BH_GB);
 #endif
                 }
             }
@@ -173,10 +182,9 @@ memory_instantiate(WASMModuleInstance *module_inst, WASMModuleInstance *parent,
     uint32 inc_page_count, aux_heap_base, global_idx;
     uint32 bytes_of_last_page, bytes_to_page_end;
     uint8 *global_addr;
-#ifdef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
-    uint8 *mapped_mem;
+    uint8 *mapped_mem = NULL;
-    uint64 map_size = 8 * (uint64)BH_GB;
+    uint64 map_size;
-    uint64 page_size = os_getpagesize();
 #endif
 
 #if WASM_ENABLE_SHARED_MEMORY != 0
@@ -295,18 +303,29 @@ memory_instantiate(WASMModuleInstance *module_inst, WASMModuleInstance *parent,
     (void)max_memory_data_size;
 
     bh_assert(memory != NULL);
+
 #ifndef OS_ENABLE_HW_BOUND_CHECK
 #if WASM_ENABLE_SHARED_MEMORY != 0
     if (is_shared_memory) {
         /* Allocate maximum memory size when memory is shared */
+#if WASM_ENABLE_SHARED_MEMORY_MMAP != 0
+        map_size = max_memory_data_size;
+        if (max_memory_data_size > 0
+            && !(memory->memory_data = mapped_mem =
+                     wasm_mmap_linear_memory(map_size, &max_memory_data_size,
+                                             error_buf, error_buf_size))) {
+            goto fail1;
+        }
+#else
         if (max_memory_data_size > 0
             && !(memory->memory_data = runtime_malloc(
                      max_memory_data_size, error_buf, error_buf_size))) {
             goto fail1;
         }
+#endif
     }
     else
-#endif
+#endif /* end of WASM_ENABLE_SHARED_MEMORY != 0 */
     {
         /* Allocate initial memory size when memory is not shared */
         if (memory_data_size > 0
@@ -316,42 +335,17 @@ memory_instantiate(WASMModuleInstance *module_inst, WASMModuleInstance *parent,
         }
     }
 #else  /* else of OS_ENABLE_HW_BOUND_CHECK */
-    memory_data_size = (memory_data_size + page_size - 1) & ~(page_size - 1);
-
     /* Totally 8G is mapped, the opcode load/store address range is 0 to 8G:
      *   ea = i + memarg.offset
      * both i and memarg.offset are u32 in range 0 to 4G
      * so the range of ea is 0 to 8G
      */
-    if (!(memory->memory_data = mapped_mem =
+    map_size = 8 * (uint64)BH_GB;
-              os_mmap(NULL, map_size, MMAP_PROT_NONE, MMAP_MAP_NONE,
+    if (!(memory->memory_data = mapped_mem = wasm_mmap_linear_memory(
-                      os_get_invalid_handle()))) {
+              map_size, &memory_data_size, error_buf, error_buf_size))) {
         set_error_buf(error_buf, error_buf_size, "mmap memory failed");
         goto fail1;
     }
-
-#ifdef BH_PLATFORM_WINDOWS
-    if (memory_data_size > 0
-        && !os_mem_commit(mapped_mem, memory_data_size,
-                          MMAP_PROT_READ | MMAP_PROT_WRITE)) {
-        set_error_buf(error_buf, error_buf_size, "commit memory failed");
-        os_munmap(mapped_mem, map_size);
-        goto fail1;
-    }
-#endif
-
-    if (os_mprotect(mapped_mem, memory_data_size,
-                    MMAP_PROT_READ | MMAP_PROT_WRITE)
-        != 0) {
-        set_error_buf(error_buf, error_buf_size, "mprotect memory failed");
-        goto fail2;
-    }
-
-    /* Newly allocated pages are filled with zero by the OS, we don't fill it
-     * again here */
-
-    if (memory_data_size > UINT32_MAX)
-        memory_data_size = UINT32_MAX;
 #endif /* end of OS_ENABLE_HW_BOUND_CHECK */
 
     memory->module_type = Wasm_Module_Bytecode;
@@ -398,15 +392,15 @@ fail3:
     if (heap_size > 0)
         wasm_runtime_free(memory->heap_handle);
 fail2:
-#ifndef OS_ENABLE_HW_BOUND_CHECK
+#ifdef WASM_LINEAR_MEMORY_MMAP
+    if (mapped_mem)
+        wasm_munmap_linear_memory(mapped_mem, memory_data_size, map_size);
+    else
+#endif
+    {
         if (memory->memory_data)
             wasm_runtime_free(memory->memory_data);
-#else
+    }
-#ifdef BH_PLATFORM_WINDOWS
-    os_mem_decommit(mapped_mem, memory_data_size);
-#endif
-    os_munmap(mapped_mem, map_size);
-#endif
 fail1:
     return NULL;
 }

core/shared/platform/android/platform_internal.h

@@ -80,8 +80,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -101,6 +99,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64/RISCV64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 typedef long int __syscall_slong_t;
 
 #if __ANDROID_API__ < 19

core/shared/platform/cosmopolitan/platform_internal.h

@@ -98,8 +98,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -119,6 +117,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64/RISCV64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 #ifdef __cplusplus
 }
 #endif

core/shared/platform/darwin/platform_internal.h

@@ -82,8 +82,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -103,6 +101,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64/RISCV64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 #if WASM_DISABLE_WAKEUP_BLOCKING_OP == 0
 #define OS_ENABLE_WAKEUP_BLOCKING_OP
 #endif

core/shared/platform/freebsd/platform_internal.h

@@ -85,8 +85,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -106,6 +104,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64/RISCV64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 #if WASM_DISABLE_WAKEUP_BLOCKING_OP == 0
 #define OS_ENABLE_WAKEUP_BLOCKING_OP
 #endif

core/shared/platform/linux/platform_internal.h

@@ -95,8 +95,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -116,6 +114,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64/RISCV64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 #if WASM_DISABLE_WAKEUP_BLOCKING_OP == 0
 #define OS_ENABLE_WAKEUP_BLOCKING_OP
 #endif

core/shared/platform/vxworks/platform_internal.h

@@ -79,8 +79,6 @@ typedef jmp_buf korp_jmpbuf;
 #define os_longjmp longjmp
 #define os_alloca alloca
 
-#define os_getpagesize getpagesize
-
 typedef void (*os_signal_handler)(void *sig_addr);
 
 int
@@ -100,6 +98,8 @@ os_sigreturn();
 #endif /* end of BUILD_TARGET_X86_64/AMD_64/AARCH64 */
 #endif /* end of WASM_DISABLE_HW_BOUND_CHECK */
 
+#define os_getpagesize getpagesize
+
 static inline os_file_handle
 os_get_invalid_handle()
 {