HPA: Use dirtiest-first purging.

This seems to be practically beneficial, despite some pathological corner cases.

include/jemalloc/internal/hpdata.h

@@ -6,42 +6,6 @@
 #include "jemalloc/internal/ql.h"
 #include "jemalloc/internal/typed_list.h"
 
-/*
- * How badly we want to purge some region of memory.  This is a temporary
- * definition; it gets deleted in the next commit (where we adopt a more
- * explicit dirtiest-first policy that only considers hugification status).
- */
-enum hpdata_purge_level_e {
-	/*
-	 * The level number is important -- we use it as indices into an array
-	 * of size 2 (one for each purge level).
-	 */
-
-	/* "Regular" candidates for purging. */
-	hpdata_purge_level_default = 0,
-
-	/*
-	 * Candidates for purging, but as a last resort.  Practically,
-	 * nonpreferred corresponds to hugified regions that are below the
-	 * hugification threshold but have not yet reached the dehugification
-	 * threshold, while strongly nonpreferred candidates are those which are
-	 * above the hugification threshold.
-	 */
-	hpdata_purge_level_nonpreferred = 1,
-	hpdata_purge_level_strongly_nonpreferred = 2,
-
-	/* Don't purge, no matter what. */
-	hpdata_purge_level_never = 2,
-
-	/*
-	 * How big an array has to be to accomodate all purge levels.  This
-	 * relies on the fact that we don't actually keep unpurgable hpdatas in
-	 * a container.
-	 */
-	hpdata_purge_level_count = hpdata_purge_level_never
-};
-typedef enum hpdata_purge_level_e hpdata_purge_level_t;
-
 /*
  * The metadata representation we use for extents in hugepages.  While the PAC
  * uses the edata_t to represent both active and inactive extents, the HP only
@@ -87,9 +51,13 @@ struct hpdata_s {
 	bool h_alloc_allowed;
 	bool h_in_psset_alloc_container;
 
-	/* The same, but with purging. */
-	uint8_t h_purge_level;
-	uint8_t h_purge_container_level;
+	/*
+	 * The same, but with purging.  There's no corresponding
+	 * h_in_psset_purge_container, because the psset (currently) always
+	 * removes hpdatas from their containers during updates (to implement
+	 * LRU for purging).
+	 */
+	bool h_purge_allowed;
 
 	/* And with hugifying. */
 	bool h_hugify_allowed;
@@ -200,26 +168,15 @@ hpdata_in_psset_alloc_container_set(hpdata_t *hpdata, bool in_container) {
 	hpdata->h_in_psset_alloc_container = in_container;
 }
 
-static inline hpdata_purge_level_t
-hpdata_purge_level_get(const hpdata_t *hpdata) {
-	return (hpdata_purge_level_t)hpdata->h_purge_level;
+static inline bool
+hpdata_purge_allowed_get(const hpdata_t *hpdata) {
+	return hpdata->h_purge_allowed;
 }
 
 static inline void
-hpdata_purge_level_set(hpdata_t *hpdata, hpdata_purge_level_t level) {
-	assert(level == hpdata_purge_level_never || !hpdata->h_mid_purge);
-	hpdata->h_purge_level = (uint8_t)level;
-}
-
-static inline hpdata_purge_level_t
-hpdata_purge_container_level_get(const hpdata_t *hpdata) {
-	return (hpdata_purge_level_t)hpdata->h_purge_container_level;
-}
-
-static inline void
-hpdata_purge_container_level_set(hpdata_t *hpdata, hpdata_purge_level_t level) {
-	assert(level != hpdata->h_purge_container_level);
-	hpdata->h_purge_container_level = level;
+hpdata_purge_allowed_set(hpdata_t *hpdata, bool purge_allowed) {
+	assert(purge_allowed == false || !hpdata->h_mid_purge);
+	hpdata->h_purge_allowed = purge_allowed;
 }
 
 static inline bool
@@ -357,12 +314,7 @@ hpdata_consistent(hpdata_t *hpdata) {
 		return false;
 	}
 	if (hpdata_changing_state_get(hpdata)
-	    && ((hpdata->h_purge_level != hpdata_purge_level_never)
-	    || hpdata->h_hugify_allowed)) {
-		return false;
-	}
-	if (hpdata_purge_level_get(hpdata)
-	    != hpdata_purge_container_level_get(hpdata)) {
+	    && ((hpdata->h_purge_allowed) || hpdata->h_hugify_allowed)) {
 		return false;
 	}
 	if (hpdata_hugify_allowed_get(hpdata)

include/jemalloc/internal/psset.h

@@ -20,6 +20,14 @@
  */
 #define PSSET_NPSIZES 64
 
+/*
+ * We keep two purge lists per page size class; one for hugified hpdatas (at
+ * index 2*pszind), and one for the non-hugified hpdatas (at index 2*pszind +
+ * 1).  This lets us implement a preference for purging non-hugified hpdatas
+ * among similarly-dirty ones.
+ */
+#define PSSET_NPURGE_LISTS (2 * PSSET_NPSIZES)
+
 typedef struct psset_bin_stats_s psset_bin_stats_t;
 struct psset_bin_stats_s {
 	/* How many pageslabs are in this bin? */
@@ -71,7 +79,9 @@ struct psset_s {
 	 */
 	hpdata_empty_list_t empty;
 	/* Slabs which are available to be purged, ordered by purge level. */
-	hpdata_purge_list_t to_purge[hpdata_purge_level_count];
+	hpdata_purge_list_t to_purge[PSSET_NPURGE_LISTS];
+	/* Bitmap for which set bits correspond to non-empty purge lists. */
+	fb_group_t purge_bitmap[FB_NGROUPS(PSSET_NPURGE_LISTS)];
 	/* Slabs which are available to be hugified. */
 	hpdata_hugify_list_t to_hugify;
 };

src/hpa.c

@@ -195,7 +195,7 @@ hpa_update_purge_hugify_eligibility(tsdn_t *tsdn, hpa_shard_t *shard,
     hpdata_t *ps) {
 	malloc_mutex_assert_owner(tsdn, &shard->mtx);
 	if (hpdata_changing_state_get(ps)) {
-		hpdata_purge_level_set(ps, hpdata_purge_level_never);
+		hpdata_purge_allowed_set(ps, false);
 		hpdata_hugify_allowed_set(ps, false);
 		return;
 	}
@@ -221,23 +221,7 @@ hpa_update_purge_hugify_eligibility(tsdn_t *tsdn, hpa_shard_t *shard,
 	 * allocator's end at all; we just try to pack allocations in a
 	 * hugepage-friendly manner and let the OS hugify in the background.
 	 */
-	if (hpdata_ndirty_get(ps) > 0) {
-		if (hpdata_huge_get(ps)) {
-			if (hpa_good_hugification_candidate(shard, ps)) {
-				hpdata_purge_level_set(ps,
-				    hpdata_purge_level_strongly_nonpreferred);
-			} else if (hpdata_ndirty_get(ps) * PAGE
-			    >= shard->opts.dehugification_threshold) {
-				hpdata_purge_level_set(ps,
-				    hpdata_purge_level_nonpreferred);
-			} else {
-				hpdata_purge_level_set(ps,
-				    hpdata_purge_level_default);
-			}
-		} else {
-			hpdata_purge_level_set(ps, hpdata_purge_level_default);
-		}
-	}
+	hpdata_purge_allowed_set(ps, hpdata_ndirty_get(ps) > 0);
 	if (hpa_good_hugification_candidate(shard, ps)
 	    && !hpdata_huge_get(ps)) {
 		hpdata_hugify_allowed_set(ps, true);
@@ -317,7 +301,7 @@ hpa_try_purge(tsdn_t *tsdn, hpa_shard_t *shard) {
 	if (to_purge == NULL) {
 		return false;
 	}
-	assert(hpdata_purge_level_get(to_purge) != hpdata_purge_level_never);
+	assert(hpdata_purge_allowed_get(to_purge));
 	assert(!hpdata_changing_state_get(to_purge));
 
 	/*
@@ -328,7 +312,7 @@ hpa_try_purge(tsdn_t *tsdn, hpa_shard_t *shard) {
 	psset_update_begin(&shard->psset, to_purge);
 	assert(hpdata_alloc_allowed_get(to_purge));
 	hpdata_mid_purge_set(to_purge, true);
-	hpdata_purge_level_set(to_purge, hpdata_purge_level_never);
+	hpdata_purge_allowed_set(to_purge, false);
 	hpdata_hugify_allowed_set(to_purge, false);
 	/*
 	 * Unlike with hugification (where concurrent
@@ -413,7 +397,7 @@ hpa_try_hugify(tsdn_t *tsdn, hpa_shard_t *shard) {
 	 */
 	psset_update_begin(&shard->psset, to_hugify);
 	hpdata_mid_hugify_set(to_hugify, true);
-	hpdata_purge_level_set(to_hugify, hpdata_purge_level_never);
+	hpdata_purge_allowed_set(to_hugify, false);
 	hpdata_hugify_allowed_set(to_hugify, false);
 	assert(hpdata_alloc_allowed_get(to_hugify));
 	psset_update_end(&shard->psset, to_hugify);

src/hpdata.c

@@ -24,8 +24,7 @@ hpdata_init(hpdata_t *hpdata, void *addr, uint64_t age) {
 	hpdata->h_huge = false;
 	hpdata->h_alloc_allowed = true;
 	hpdata->h_in_psset_alloc_container = false;
-	hpdata->h_purge_level = hpdata_purge_level_never;
-	hpdata->h_purge_container_level = hpdata_purge_level_never;
+	hpdata->h_purge_allowed = false;
 	hpdata->h_hugify_allowed = false;
 	hpdata->h_in_psset_hugify_container = false;
 	hpdata->h_mid_purge = false;

src/psset.c

@@ -14,9 +14,10 @@ psset_init(psset_t *psset) {
 	memset(&psset->merged_stats, 0, sizeof(psset->merged_stats));
 	memset(&psset->stats, 0, sizeof(psset->stats));
 	hpdata_empty_list_init(&psset->empty);
-	for (int i = 0; i < hpdata_purge_level_count; i++) {
+	for (int i = 0; i < PSSET_NPURGE_LISTS; i++) {
 		hpdata_purge_list_init(&psset->to_purge[i]);
 	}
+	fb_init(psset->purge_bitmap, PSSET_NPURGE_LISTS);
 	hpdata_hugify_list_init(&psset->to_hugify);
 }
 
@@ -195,6 +196,51 @@ psset_alloc_container_remove(psset_t *psset, hpdata_t *ps) {
 	}
 }
 
+static size_t
+psset_purge_list_ind(hpdata_t *ps) {
+	size_t ndirty = hpdata_ndirty_get(ps);
+	/* Shouldn't have something with no dirty pages purgeable. */
+	assert(ndirty > 0);
+	pszind_t pind = sz_psz2ind(sz_psz_quantize_floor(ndirty << LG_PAGE));
+	/*
+	 * Higher indices correspond to lists we'd like to purge earlier;
+	 * increment the index for the nonhugified hpdatas first, so that we'll
+	 * pick them before picking hugified ones.
+	 */
+	return (size_t)pind * 2 + (hpdata_huge_get(ps) ? 0 : 1);
+}
+
+static void
+psset_maybe_remove_purge_list(psset_t *psset, hpdata_t *ps) {
+	/*
+	 * Remove the hpdata from its purge list (if it's in one).  Even if it's
+	 * going to stay in the same one, by appending it during
+	 * psset_update_end, we move it to the end of its queue, so that we
+	 * purge LRU within a given dirtiness bucket.
+	 */
+	if (hpdata_purge_allowed_get(ps)) {
+		size_t ind = psset_purge_list_ind(ps);
+		hpdata_purge_list_t *purge_list = &psset->to_purge[ind];
+		hpdata_purge_list_remove(purge_list, ps);
+		if (hpdata_purge_list_empty(purge_list)) {
+			fb_unset(psset->purge_bitmap, PSSET_NPURGE_LISTS, ind);
+		}
+	}
+}
+
+static void
+psset_maybe_insert_purge_list(psset_t *psset, hpdata_t *ps) {
+	if (hpdata_purge_allowed_get(ps)) {
+		size_t ind = psset_purge_list_ind(ps);
+		hpdata_purge_list_t *purge_list = &psset->to_purge[ind];
+		if (hpdata_purge_list_empty(purge_list)) {
+			fb_set(psset->purge_bitmap, PSSET_NPURGE_LISTS, ind);
+		}
+		hpdata_purge_list_append(purge_list, ps);
+	}
+
+}
+
 void
 psset_update_begin(psset_t *psset, hpdata_t *ps) {
 	hpdata_assert_consistent(ps);
@@ -210,10 +256,11 @@ psset_update_begin(psset_t *psset, hpdata_t *ps) {
 		assert(hpdata_alloc_allowed_get(ps));
 		psset_alloc_container_remove(psset, ps);
 	}
+	psset_maybe_remove_purge_list(psset, ps);
 	/*
-	 * We don't update presence in the purge list or hugify list; we try to
-	 * keep those FIFO, even in the presence of other metadata updates.
-	 * We'll update presence at the end of the metadata update if necessary.
+	 * We don't update presence in the hugify list; we try to keep it FIFO,
+	 * even in the presence of other metadata updates.  We'll update
+	 * presence at the end of the metadata update if necessary.
 	 */
 }
 
@@ -231,33 +278,7 @@ psset_update_end(psset_t *psset, hpdata_t *ps) {
 	if (hpdata_alloc_allowed_get(ps)) {
 		psset_alloc_container_insert(psset, ps);
 	}
-
-	if (hpdata_purge_level_get(ps) == hpdata_purge_level_never
-	    && hpdata_purge_container_level_get(ps)
-	    != hpdata_purge_level_never) {
-		/* In some purge container, but shouldn't be in any. */
-		hpdata_purge_list_remove(
-		    &psset->to_purge[hpdata_purge_container_level_get(ps)],
-		    ps);
-		hpdata_purge_container_level_set(ps, hpdata_purge_level_never);
-	} else if (hpdata_purge_level_get(ps) != hpdata_purge_level_never
-	    && hpdata_purge_container_level_get(ps)
-	    == hpdata_purge_level_never) {
-		/* Not in any purge container, but should be in one. */
-		hpdata_purge_list_append(
-		    &psset->to_purge[hpdata_purge_level_get(ps)], ps);
-		hpdata_purge_container_level_set(ps,
-		    hpdata_purge_level_get(ps));
-	} else if (hpdata_purge_level_get(ps)
-	    != hpdata_purge_container_level_get(ps)) {
-		/* Should switch containers. */
-		hpdata_purge_list_remove(
-		    &psset->to_purge[hpdata_purge_container_level_get(ps)], ps);
-		hpdata_purge_list_append(
-		    &psset->to_purge[hpdata_purge_level_get(ps)], ps);
-		hpdata_purge_container_level_set(ps,
-		    hpdata_purge_level_get(ps));
-	}
+	psset_maybe_insert_purge_list(psset, ps);
 
 	if (hpdata_hugify_allowed_get(ps)
 	    && !hpdata_in_psset_hugify_container_get(ps)) {
@@ -294,13 +315,16 @@ psset_pick_alloc(psset_t *psset, size_t size) {
 
 hpdata_t *
 psset_pick_purge(psset_t *psset) {
-	for (int i = 0; i < hpdata_purge_level_count; i++) {
-		hpdata_t *ps = hpdata_purge_list_first(&psset->to_purge[i]);
-		if (ps != NULL) {
-			return ps;
-		}
+	ssize_t ind_ssz = fb_fls(psset->purge_bitmap, PSSET_NPURGE_LISTS,
+	    PSSET_NPURGE_LISTS - 1);
+	if (ind_ssz < 0) {
+		return NULL;
 	}
-	return NULL;
+	pszind_t ind = (pszind_t)ind_ssz;
+	assert(ind < PSSET_NPSIZES);
+	hpdata_t *ps = hpdata_purge_list_first(&psset->to_purge[ind]);
+	assert(ps != NULL);
+	return ps;
 }
 
 hpdata_t *
@@ -316,14 +340,7 @@ psset_insert(psset_t *psset, hpdata_t *ps) {
 	if (hpdata_alloc_allowed_get(ps)) {
 		psset_alloc_container_insert(psset, ps);
 	}
-	assert(
-	    hpdata_purge_container_level_get(ps) == hpdata_purge_level_never);
-	if (hpdata_purge_level_get(ps) != hpdata_purge_level_never) {
-		hpdata_purge_container_level_set(ps,
-		    hpdata_purge_level_get(ps));
-		hpdata_purge_list_append(
-		    &psset->to_purge[hpdata_purge_level_get(ps)], ps);
-	}
+	psset_maybe_insert_purge_list(psset, ps);
 
 	if (hpdata_hugify_allowed_get(ps)) {
 		hpdata_in_psset_hugify_container_set(ps, true);
@@ -339,11 +356,7 @@ psset_remove(psset_t *psset, hpdata_t *ps) {
 	if (hpdata_in_psset_alloc_container_get(ps)) {
 		psset_alloc_container_remove(psset, ps);
 	}
-	if (hpdata_purge_container_level_get(ps) != hpdata_purge_level_never) {
-		hpdata_purge_list_remove(
-		    &psset->to_purge[hpdata_purge_container_level_get(ps)], ps);
-		hpdata_purge_container_level_set(ps, hpdata_purge_level_never);
-	}
+	psset_maybe_remove_purge_list(psset, ps);
 	if (hpdata_in_psset_hugify_container_get(ps)) {
 		hpdata_in_psset_hugify_container_set(ps, false);
 		hpdata_hugify_list_remove(&psset->to_hugify, ps);

test/unit/psset.c

@@ -540,6 +540,89 @@ TEST_BEGIN(test_insert_remove) {
 }
 TEST_END
 
+TEST_BEGIN(test_purge_prefers_nonhuge) {
+	/*
+	 * All else being equal, we should prefer purging non-huge pages over
+	 * huge ones.
+	 */
+
+	/* Nothing magic about this constant. */
+	enum {
+		NHP = 23,
+	};
+	hpdata_t *hpdata;
+
+	psset_t psset;
+	psset_init(&psset);
+
+	hpdata_t hpdata_huge[NHP];
+	uintptr_t huge_begin = (uintptr_t)&hpdata_huge[0];
+	uintptr_t huge_end = (uintptr_t)&hpdata_huge[NHP];
+	hpdata_t hpdata_nonhuge[NHP];
+	uintptr_t nonhuge_begin = (uintptr_t)&hpdata_nonhuge[0];
+	uintptr_t nonhuge_end = (uintptr_t)&hpdata_nonhuge[NHP];
+
+	for (size_t i = 0; i < NHP; i++) {
+		hpdata_init(&hpdata_huge[i], (void *)((10 + i) * HUGEPAGE),
+		    123 + i);
+		psset_insert(&psset, &hpdata_huge[i]);
+
+		hpdata_init(&hpdata_nonhuge[i],
+		    (void *)((10 + NHP + i) * HUGEPAGE),
+		    456 + i);
+		psset_insert(&psset, &hpdata_nonhuge[i]);
+
+	}
+	for (int i = 0; i < 2 * NHP; i++) {
+		hpdata = psset_pick_alloc(&psset, HUGEPAGE * 3 / 4);
+		psset_update_begin(&psset, hpdata);
+		void *ptr;
+		ptr = hpdata_reserve_alloc(hpdata, HUGEPAGE * 3 / 4);
+		/* Ignore the first alloc, which will stick around. */
+		(void)ptr;
+		/*
+		 * The second alloc is to dirty the pages; free it immediately
+		 * after allocating.
+		 */
+		ptr = hpdata_reserve_alloc(hpdata, HUGEPAGE / 4);
+		hpdata_unreserve(hpdata, ptr, HUGEPAGE / 4);
+
+		if (huge_begin <= (uintptr_t)hpdata
+		    && (uintptr_t)hpdata < huge_end) {
+			hpdata_hugify(hpdata);
+		}
+
+		hpdata_purge_allowed_set(hpdata, true);
+		psset_update_end(&psset, hpdata);
+	}
+
+	/*
+	 * We've got a bunch of 1/8th dirty hpdatas.  It should give us all the
+	 * non-huge ones to purge, then all the huge ones, then refuse to purge
+	 * further.
+	 */
+	for (int i = 0; i < NHP; i++) {
+		hpdata = psset_pick_purge(&psset);
+		assert_true(nonhuge_begin <= (uintptr_t)hpdata
+		    && (uintptr_t)hpdata < nonhuge_end, "");
+		psset_update_begin(&psset, hpdata);
+		test_psset_fake_purge(hpdata);
+		hpdata_purge_allowed_set(hpdata, false);
+		psset_update_end(&psset, hpdata);
+	}
+	for (int i = 0; i < NHP; i++) {
+		hpdata = psset_pick_purge(&psset);
+		expect_true(huge_begin <= (uintptr_t)hpdata
+		    && (uintptr_t)hpdata < huge_end, "");
+		psset_update_begin(&psset, hpdata);
+		hpdata_dehugify(hpdata);
+		test_psset_fake_purge(hpdata);
+		hpdata_purge_allowed_set(hpdata, false);
+		psset_update_end(&psset, hpdata);
+	}
+}
+TEST_END
+
 int
 main(void) {
 	return test_no_reentrancy(
@@ -550,5 +633,6 @@ main(void) {
 	    test_multi_pageslab,
 	    test_stats,
 	    test_oldest_fit,
-	    test_insert_remove);
+	    test_insert_remove,
+	    test_purge_prefers_nonhuge);
 }