Merge branch 'dev'

jemalloc/ChangeLog

@@ -6,6 +6,15 @@ found in the git revision history:
     http://www.canonware.com/cgi-bin/gitweb.cgi?p=jemalloc.git
     git://canonware.com/jemalloc.git
 
+* 2.0.1
+
+  Bug fixes:
+  - Fix a race condition in heap profiling that could cause undefined behavior
+    if opt.prof_accum were disabled.
+  - Add missing mutex unlocks for some OOM error paths in the heap profiling
+    code.
+  - Fix a compilation error for non-C99 builds.
+
 * 2.0.0
 
   This version focuses on the experimental *allocm() API, and on improved
@@ -13,35 +22,34 @@ found in the git revision history:
   improvements are also included.
 
   New features:
-    - Implement the experimental {,r,s,d}allocm() API, which provides a superset
-      of the functionality available via malloc(), calloc(), posix_memalign(),
-      realloc(), malloc_usable_size(), and free().  These functions can be used
-      to allocate/reallocate aligned zeroed memory, ask for optional extra
-      memory during reallocation, prevent object movement during reallocation,
-      etc.
-    - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
-      more human-readable, and more flexible.  For example:
-        JEMALLOC_OPTIONS=AJP
-      is now:
-        MALLOC_CONF=abort:true,fill:true,stats_print:true
-    - Port to Apple OS X.  Sponsored by Mozilla.
-    - Make it possible for the application to control thread-->arena mappings
-      via the "thread.arena" mallctl.
-    - Add compile-time support for all TLS-related functionality via pthreads
-      TSD.  This is mainly of interest for OS X, which does not support TLS, but
-      has a TSD implementation with similar performance.
-    - Override memalign() and valloc() if they are provided by the system.
-    - Add the "arenas.purge" mallctl, which can be used to synchronously purge
-      all dirty unused pages.
-    - Make cumulative heap profiling data optional, so that it is possible to
-      limit the amount of memory consumed by heap profiling data structures.
-    - Add per thread allocation counters that can be accessed via the
-      "thread.allocated" and "thread.deallocated" mallctls.
+  - Implement the experimental {,r,s,d}allocm() API, which provides a superset
+    of the functionality available via malloc(), calloc(), posix_memalign(),
+    realloc(), malloc_usable_size(), and free().  These functions can be used to
+    allocate/reallocate aligned zeroed memory, ask for optional extra memory
+    during reallocation, prevent object movement during reallocation, etc.
+  - Replace JEMALLOC_OPTIONS/JEMALLOC_PROF_PREFIX with MALLOC_CONF, which is
+    more human-readable, and more flexible.  For example:
+      JEMALLOC_OPTIONS=AJP
+    is now:
+      MALLOC_CONF=abort:true,fill:true,stats_print:true
+  - Port to Apple OS X.  Sponsored by Mozilla.
+  - Make it possible for the application to control thread-->arena mappings via
+    the "thread.arena" mallctl.
+  - Add compile-time support for all TLS-related functionality via pthreads TSD.
+    This is mainly of interest for OS X, which does not support TLS, but has a
+    TSD implementation with similar performance.
+  - Override memalign() and valloc() if they are provided by the system.
+  - Add the "arenas.purge" mallctl, which can be used to synchronously purge all
+    dirty unused pages.
+  - Make cumulative heap profiling data optional, so that it is possible to
+    limit the amount of memory consumed by heap profiling data structures.
+  - Add per thread allocation counters that can be accessed via the
+    "thread.allocated" and "thread.deallocated" mallctls.
 
   Incompatible changes:
-    - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
-    - Increase default backtrace depth from 4 to 128 for heap profiling.
-    - Disable interval-based profile dumps by default.
+  - Remove JEMALLOC_OPTIONS and malloc_options (see MALLOC_CONF above).
+  - Increase default backtrace depth from 4 to 128 for heap profiling.
+  - Disable interval-based profile dumps by default.
 
   Bug fixes:
   - Remove bad assertions in fork handler functions.  These assertions could

jemalloc/src/arena.c

@@ -304,8 +304,10 @@ arena_run_reg_dalloc(arena_run_t *run, void *ptr)
 static inline void
 arena_chunk_validate_zeroed(arena_chunk_t *chunk, size_t run_ind)
 {
+	size_t i;
 	size_t *p = (size_t *)((uintptr_t)chunk + (run_ind << PAGE_SHIFT));
-	for (size_t i = 0; i < PAGE_SIZE / sizeof(size_t); i++)
+
+	for (i = 0; i < PAGE_SIZE / sizeof(size_t); i++)
 		assert(p[i] == 0);
 }
 #endif

jemalloc/src/prof.c

@@ -255,6 +255,7 @@ prof_backtrace(prof_bt_t *bt, unsigned nignore, unsigned max)
 	} else								\
 		return;
 
+	assert(nignore <= 3);
 	assert(max <= (1U << opt_lg_prof_bt_max));
 
 	BT_FRAME(0)
@@ -398,7 +399,7 @@ prof_backtrace(prof_bt_t *bt, unsigned nignore, unsigned max)
 	BT_FRAME(126)
 	BT_FRAME(127)
 
-	/* Extras to compensate for NIGNORE. */
+	/* Extras to compensate for nignore. */
 	BT_FRAME(128)
 	BT_FRAME(129)
 	BT_FRAME(130)
@@ -496,8 +497,10 @@ prof_lookup(prof_bt_t *bt)
 			    opt_lg_prof_tcmax));
 			/* Allocate and partially initialize a new cnt. */
 			ret.v = imalloc(sizeof(prof_thr_cnt_t));
-			if (ret.p == NULL)
+			if (ret.p == NULL) {
+				malloc_mutex_unlock(&ctx.p->lock);
 				return (NULL);
+			}
 			ql_elm_new(ret.p, cnts_link);
 			ql_elm_new(ret.p, lru_link);
 		}
@@ -506,6 +509,7 @@ prof_lookup(prof_bt_t *bt)
 		ret.p->epoch = 0;
 		memset(&ret.p->cnts, 0, sizeof(prof_cnt_t));
 		if (ckh_insert(&prof_tdata->bt2cnt, btkey.v, ret.v)) {
+			malloc_mutex_unlock(&ctx.p->lock);
 			idalloc(ret.v);
 			return (NULL);
 		}
@@ -625,11 +629,14 @@ prof_ctx_destroy(prof_ctx_t *ctx)
 	/*
 	 * Check that ctx is still unused by any thread cache before destroying
 	 * it.  prof_lookup() interlocks bt2ctx_mtx and ctx->lock in order to
-	 * avoid a race condition with this function.
+	 * avoid a race condition with this function, and prof_ctx_merge()
+	 * artificially raises ctx->cnt_merged.curobjs in order to avoid a race
+	 * between the main body of prof_ctx_merge() and entry into this
+	 * function.
 	 */
 	prof_enter();
 	malloc_mutex_lock(&ctx->lock);
-	if (ql_first(&ctx->cnts_ql) == NULL && ctx->cnt_merged.curobjs == 0) {
+	if (ql_first(&ctx->cnts_ql) == NULL && ctx->cnt_merged.curobjs == 1) {
 		assert(ctx->cnt_merged.curbytes == 0);
 		assert(ctx->cnt_merged.accumobjs == 0);
 		assert(ctx->cnt_merged.accumbytes == 0);
@@ -642,6 +649,8 @@ prof_ctx_destroy(prof_ctx_t *ctx)
 		malloc_mutex_destroy(&ctx->lock);
 		idalloc(ctx);
 	} else {
+		/* Compensate for increment in prof_ctx_merge(). */
+		ctx->cnt_merged.curobjs--;
 		malloc_mutex_unlock(&ctx->lock);
 		prof_leave();
 	}
@@ -660,9 +669,23 @@ prof_ctx_merge(prof_ctx_t *ctx, prof_thr_cnt_t *cnt)
 	ctx->cnt_merged.accumbytes += cnt->cnts.accumbytes;
 	ql_remove(&ctx->cnts_ql, cnt, cnts_link);
 	if (opt_prof_accum == false && ql_first(&ctx->cnts_ql) == NULL &&
-	    ctx->cnt_merged.curobjs == 0)
+	    ctx->cnt_merged.curobjs == 0) {
+		/*
+		 * Artificially raise ctx->cnt_merged.curobjs in order to keep
+		 * another thread from winning the race to destroy ctx while
+		 * this one has ctx->lock dropped.  Without this, it would be
+		 * possible for another thread to:
+		 *
+		 * 1) Sample an allocation associated with ctx.
+		 * 2) Deallocate the sampled object.
+		 * 3) Successfully prof_ctx_destroy(ctx).
+		 *
+		 * The result would be that ctx no longer exists by the time
+		 * this thread accesses it in prof_ctx_destroy().
+		 */
+		ctx->cnt_merged.curobjs++;
 		destroy = true;
-	else
+	} else
 		destroy = false;
 	malloc_mutex_unlock(&ctx->lock);
 	if (destroy)