Fix EnableTerminationOnOutOfMemory for malloc

skia/ext/SkMemory_new_handler.cpp

Index: skia/ext/SkMemory_new_handler.cpp
diff --git a/skia/ext/SkMemory_new_handler.cpp b/skia/ext/SkMemory_new_handler.cpp
index 015521f313ac83b2797689d048557b495e3f9049..edf0f968bfed975ca431384ef5cbd1bea6026c35 100644
--- a/skia/ext/SkMemory_new_handler.cpp
+++ b/skia/ext/SkMemory_new_handler.cpp
@@ -60,13 +60,7 @@ static void* sk_malloc_nothrow(size_t size) {
     ignore_result(base::UncheckedMalloc(size, &result));
     return result;
 #else
-    // This is not really thread safe.  It only won't collide with itself, but we're totally

[cpu_(ooo_6.6-7.5), 2015/01/10 05:37:28]

this code is wrong. libraries shall not to touch global crt state, which btw the
std::set_new_handler and the _set_new_handler() are not the same so this code
was not doing anything: old_handler is 0 even if _set_new_handler is active
which means that OOM still crashes.

I mean other than grabbing/release 3 or 4 locks for each allocation.

[grt (UTC plus 2), 2015/01/12 14:29:14]

Should this use base::UncheckedMalloc on Win as well (ignoring for the moment
that it just calls through to malloc, and hence is really a checked malloc)
since that seems to be the intent here? That way this code will automatically do
what is intended if someone makes base::UncheckedMalloc work on Win.

-    // unprotected from races with other code that calls set_new_handler.
-    SkAutoMutexAcquire lock(gSkNewHandlerMutex);
-    std::new_handler old_handler = std::set_new_handler(NULL);
-    void* p = malloc(size);
-    std::set_new_handler(old_handler);
-    return p;
+    return malloc(size);
 #endif
 }
 
@@ -91,10 +85,6 @@ void* sk_calloc(size_t size) {
     ignore_result(base::UncheckedCalloc(size, 1, &result));
     return result;
 #else
-    SkAutoMutexAcquire lock(gSkNewHandlerMutex);
-    std::new_handler old_handler = std::set_new_handler(NULL);
-    void* p = calloc(size, 1);
-    std::set_new_handler(old_handler);
-    return p;
+    return calloc(size, 1);
 #endif
 }