Fix unbounded increase in stash_decayed.

Added an upper bound on how many pages we can decay during the current run.
Without this, decay could have unbounded increase in stashed, since other
threads could add new pages into the extents.

include/jemalloc/internal/extent_externs.h

@@ -36,7 +36,8 @@ extent_t *extents_alloc(tsdn_t *tsdn, arena_t *arena,
 void extents_dalloc(tsdn_t *tsdn, arena_t *arena,
     extent_hooks_t **r_extent_hooks, extents_t *extents, extent_t *extent);
 extent_t *extents_evict(tsdn_t *tsdn, arena_t *arena,
-    extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_min);
+    extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_min,
+    size_t npages_max);
 void extents_prefork(tsdn_t *tsdn, extents_t *extents);
 void extents_postfork_parent(tsdn_t *tsdn, extents_t *extents);
 void extents_postfork_child(tsdn_t *tsdn, extents_t *extents);

src/arena.c

@@ -62,7 +62,7 @@ const uint64_t h_steps[SMOOTHSTEP_NSTEPS] = {
 
 static void arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena,
     arena_decay_t *decay, extents_t *extents, bool all, size_t npages_limit,
-    bool is_background_thread);
+    size_t npages_decay_max, bool is_background_thread);
 static bool arena_decay_dirty(tsdn_t *tsdn, arena_t *arena,
     bool is_background_thread, bool all);
 static void arena_dalloc_bin_slab(tsdn_t *tsdn, arena_t *arena, extent_t *slab,
@@ -693,7 +693,8 @@ arena_decay_try_purge(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay,
     bool is_background_thread) {
 	if (current_npages > npages_limit) {
 		arena_decay_to_limit(tsdn, arena, decay, extents, false,
-		    npages_limit, is_background_thread);
+		    npages_limit, current_npages - npages_limit,
+		    is_background_thread);
 	}
 }
 
@@ -799,7 +800,8 @@ arena_maybe_decay(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay,
 	if (decay_ms <= 0) {
 		if (decay_ms == 0) {
 			arena_decay_to_limit(tsdn, arena, decay, extents, false,
-			    0, is_background_thread);
+			    0, extents_npages_get(extents),
+			    is_background_thread);
 		}
 		return false;
 	}
@@ -901,15 +903,16 @@ arena_muzzy_decay_ms_set(tsdn_t *tsdn, arena_t *arena,
 static size_t
 arena_stash_decayed(tsdn_t *tsdn, arena_t *arena,
     extent_hooks_t **r_extent_hooks, extents_t *extents, size_t npages_limit,
-    extent_list_t *decay_extents) {
+	size_t npages_decay_max, extent_list_t *decay_extents) {
 	witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn),
 	    WITNESS_RANK_CORE, 0);
 
 	/* Stash extents according to npages_limit. */
 	size_t nstashed = 0;
 	extent_t *extent;
-	while ((extent = extents_evict(tsdn, arena, r_extent_hooks, extents,
-	    npages_limit)) != NULL) {
+	while (nstashed < npages_decay_max &&
+	    (extent = extents_evict(tsdn, arena, r_extent_hooks, extents,
+	    npages_limit, npages_decay_max - nstashed)) != NULL) {
 		extent_list_append(decay_extents, extent);
 		nstashed += extent_size_get(extent) >> LG_PAGE;
 	}
@@ -983,12 +986,15 @@ arena_decay_stashed(tsdn_t *tsdn, arena_t *arena,
 }
 
 /*
- * npages_limit: Decay as many dirty extents as possible without violating the
- * invariant: (extents_npages_get(extents) >= npages_limit)
+ * npages_limit: Decay at most npages_decay_max pages without violating the
+ * invariant: (extents_npages_get(extents) >= npages_limit).  We need an upper
+ * bound on number of pages in order to prevent unbounded growth (namely in
+ * stashed), otherwise unbounded new pages could be added to extents during the
+ * current decay run, so that the purging thread never finishes.
  */
 static void
 arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay,
-    extents_t *extents, bool all, size_t npages_limit,
+    extents_t *extents, bool all, size_t npages_limit, size_t npages_decay_max,
     bool is_background_thread) {
 	witness_assert_depth_to_rank(tsdn_witness_tsdp_get(tsdn),
 	    WITNESS_RANK_CORE, 1);
@@ -1006,7 +1012,7 @@ arena_decay_to_limit(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay,
 	extent_list_init(&decay_extents);
 
 	size_t npurge = arena_stash_decayed(tsdn, arena, &extent_hooks, extents,
-	    npages_limit, &decay_extents);
+	    npages_limit, npages_decay_max, &decay_extents);
 	if (npurge != 0) {
 		UNUSED size_t npurged = arena_decay_stashed(tsdn, arena,
 		    &extent_hooks, decay, extents, all, &decay_extents,
@@ -1024,7 +1030,7 @@ arena_decay_impl(tsdn_t *tsdn, arena_t *arena, arena_decay_t *decay,
 	if (all) {
 		malloc_mutex_lock(tsdn, &decay->mtx);
 		arena_decay_to_limit(tsdn, arena, decay, extents, all, 0,
-		    is_background_thread);
+		    extents_npages_get(extents), is_background_thread);
 		malloc_mutex_unlock(tsdn, &decay->mtx);
 
 		return false;
@@ -1220,7 +1226,7 @@ arena_destroy_retained(tsdn_t *tsdn, arena_t *arena) {
 	extent_hooks_t *extent_hooks = extent_hooks_get(arena);
 	extent_t *extent;
 	while ((extent = extents_evict(tsdn, arena, &extent_hooks,
-	    &arena->extents_retained, 0)) != NULL) {
+	    &arena->extents_retained, 0, SIZE_MAX)) != NULL) {
 		extent_destroy_wrapper(tsdn, arena, &extent_hooks, extent);
 	}
 }

src/extent.c

@@ -472,7 +472,7 @@ extents_dalloc(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
 
 extent_t *
 extents_evict(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
-    extents_t *extents, size_t npages_min) {
+    extents_t *extents, size_t npages_min, size_t npages_max) {
 	rtree_ctx_t rtree_ctx_fallback;
 	rtree_ctx_t *rtree_ctx = tsdn_rtree_ctx(tsdn, &rtree_ctx_fallback);
 
@@ -493,7 +493,8 @@ extents_evict(tsdn_t *tsdn, arena_t *arena, extent_hooks_t **r_extent_hooks,
 		size_t npages = extent_size_get(extent) >> LG_PAGE;
 		size_t extents_npages = atomic_load_zu(&extents->npages,
 		    ATOMIC_RELAXED);
-		if (extents_npages - npages < npages_min) {
+		if (extents_npages - npages < npages_min ||
+		    npages > npages_max) {
 			extent = NULL;
 			goto label_return;
 		}