Revert re-addition of purge_lock.

Linux kernels have been capable of concurrent page table access since
2.6.27, so this hack is not necessary for modern kernels.

jemalloc/include/jemalloc/internal/arena.h

@@ -318,9 +318,12 @@ struct arena_s {
 	size_t			ndirty;
 
 	/*
-	 * True if pages are currently being purged by a thread.
+	 * Approximate number of pages being purged.  It is possible for
+	 * multiple threads to purge dirty pages concurrently, and they use
+	 * npurgatory to indicate the total number of pages all threads are
+	 * attempting to purge.
 	 */
-	bool			purgatory;
+	size_t			npurgatory;
 
 	/*
 	 * Size/address-ordered trees of this arena's available runs.  The trees

jemalloc/src/arena.c

@@ -23,9 +23,6 @@ size_t		sspace_max;
 size_t		lg_mspace;
 size_t		mspace_mask;
 
-/* Used to prevent threads from concurrently calling madvise(2). */
-static malloc_mutex_t  purge_lock;
-
 /*
  * const_small_size2bin is a static constant lookup table that in the common
  * case can be used as-is for small_size2bin.  For dynamically linked programs,
@@ -587,9 +584,10 @@ arena_maybe_purge(arena_t *arena)
 {
 
 	/* Enforce opt_lg_dirty_mult. */
-	if (opt_lg_dirty_mult >= 0 && arena->purgatory == false &&
+	if (opt_lg_dirty_mult >= 0 && arena->ndirty > arena->npurgatory &&
-	    arena->ndirty > chunk_npages && (arena->nactive >>
+	    (arena->ndirty - arena->npurgatory) > chunk_npages &&
-	    opt_lg_dirty_mult) < arena->ndirty)
+	    (arena->nactive >> opt_lg_dirty_mult) < (arena->ndirty -
+	    arena->npurgatory))
 		arena_purge(arena);
 }
 
@@ -759,32 +757,23 @@ arena_purge(arena_t *arena)
 	}
 	assert(ndirty == arena->ndirty);
 #endif
+	assert(arena->ndirty > arena->npurgatory);
 	assert(arena->ndirty > chunk_npages);
 	assert((arena->nactive >> opt_lg_dirty_mult) < arena->ndirty);
 
-	/*
-	 * Only allow one thread at a time to purge dirty pages.  madvise(2)
-	 * causes the kernel to modify virtual memory data structures that are
-	 * typically protected by a lock, and purging isn't important enough to
-	 * suffer lock contention in the kernel.  The result of failing to
-	 * acquire purge_lock here is that this arena will operate with ndirty
-	 * above the threshold until some dirty pages are re-used, or the
-	 * creation of more dirty pages causes this function to be called
-	 * again.
-	 */
-	if (malloc_mutex_trylock(&purge_lock))
-		return;
-
 #ifdef JEMALLOC_STATS
 	arena->stats.npurge++;
 #endif
 
 	/*
 	 * Compute the minimum number of pages that this thread should try to
-	 * purge.
+	 * purge, and add the result to arena->npurgatory.  This will keep
+	 * multiple threads from racing to reduce ndirty below the threshold.
 	 */
-	npurgatory = arena->ndirty - (arena->nactive >> opt_lg_dirty_mult);
+	npurgatory = (arena->ndirty - arena->npurgatory) - (arena->nactive >>
-	arena->purgatory = true;
+	    opt_lg_dirty_mult);
+	arena->npurgatory += npurgatory;
+
 	while (npurgatory > 0) {
 		/* Get next chunk with dirty pages. */
 		chunk = ql_first(&arena->chunks_dirty);
@@ -792,9 +781,11 @@ arena_purge(arena_t *arena)
 			/*
 			 * This thread was unable to purge as many pages as
 			 * originally intended, due to races with other threads
-			 * that re-used dirty pages.
+			 * that either did some of the purging work, or re-used
+			 * dirty pages.
 			 */
-			goto RETURN;
+			arena->npurgatory -= npurgatory;
+			return;
 		}
 		while (chunk->ndirty == 0) {
 			ql_remove(&arena->chunks_dirty, chunk, link_dirty);
@@ -802,20 +793,38 @@ arena_purge(arena_t *arena)
 			chunk = ql_first(&arena->chunks_dirty);
 			if (chunk == NULL) {
 				/* Same logic as for above. */
-				goto RETURN;
+				arena->npurgatory -= npurgatory;
+				return;
 			}
 		}
 
-		if (chunk->ndirty >= npurgatory) {
+		if (chunk->ndirty > npurgatory) {
-			/* This thread will purge all the pages in chunk. */
+			/*
-			npurgatory = 0;
+			 * This thread will, at a minimum, purge all the dirty
-		} else
+			 * pages in chunk, so set npurgatory to reflect this
-			npurgatory -= chunk->ndirty;
+			 * thread's commitment to purge the pages.  This tends
+			 * to reduce the chances of the following scenario:
+			 *
+			 * 1) This thread sets arena->npurgatory such that
+			 *    (arena->ndirty - arena->npurgatory) is at the
+			 *    threshold.
+			 * 2) This thread drops arena->lock.
+			 * 3) Another thread causes one or more pages to be
+			 *    dirtied, and immediately determines that it must
+			 *    purge dirty pages.
+			 *
+			 * If this scenario *does* play out, that's okay,
+			 * because all of the purging work being done really
+			 * needs to happen.
+			 */
+			arena->npurgatory += chunk->ndirty - npurgatory;
+			npurgatory = chunk->ndirty;
+		}
+
+		arena->npurgatory -= chunk->ndirty;
+		npurgatory -= chunk->ndirty;
 		arena_chunk_purge(arena, chunk);
 	}
-RETURN:
-	arena->purgatory = false;
-	malloc_mutex_unlock(&purge_lock);
 }
 
 static void
@@ -2137,7 +2146,7 @@ arena_new(arena_t *arena, unsigned ind)
 
 	arena->nactive = 0;
 	arena->ndirty = 0;
-	arena->purgatory = false;
+	arena->npurgatory = 0;
 
 	arena_avail_tree_new(&arena->runs_avail_clean);
 	arena_avail_tree_new(&arena->runs_avail_dirty);
@@ -2433,8 +2442,5 @@ arena_boot(void)
 	    ((header_size & PAGE_MASK) != 0);
 	arena_maxclass = chunksize - (arena_chunk_header_npages << PAGE_SHIFT);
 
-	if (malloc_mutex_init(&purge_lock))
-		return (true);
-
 	return (false);
 }