Try to keep all pages from extents read from disk in the cache. (#10558)
Markos Fountoulakis committed
Feb 24, 2021 at 16:25 UTC
7387e05943ead3d08fe08da2acbb2826bc0c34a4
3 files changed
+93
-10
database/engine/pagecache.c
+68
@@ -699,6 +699,74 @@ void pg_cache_get_filtered_info_prev(struct rrdengine_instance *ctx, struct pg_c
699
}
700
uv_rwlock_rdunlock(&page_index->lock);
701
}
702
+
703
+/**
704
+ * Searches for an unallocated page without triggering disk I/O. Attempts to reserve the page and get a reference.
705
+ * @param ctx DB context
706
+ * @param id lookup by UUID
707
+ * @param start_time exact starting time in usec
708
+ * @param ret_page_indexp Sets the page index pointer (*ret_page_indexp) for the given UUID.
709
+ * @return the page descriptor or NULL on failure. It can fail if:
710
+ * 1. The page is already allocated to the page cache.
711
+ * 2. It did not succeed to get a reference.
712
+ * 3. It did not succeed to reserve a spot in the page cache.
713
+ */
714
+struct rrdeng_page_descr *pg_cache_lookup_unpopulated_and_lock(struct rrdengine_instance *ctx, uuid_t *id,
715
+ usec_t start_time)
716
+{
717
+ struct page_cache *pg_cache = &ctx->pg_cache;
718
+ struct rrdeng_page_descr *descr = NULL;
719
+ struct page_cache_descr *pg_cache_descr = NULL;
720
+ unsigned long flags;
721
+ Pvoid_t *PValue;
722
+ struct pg_cache_page_index *page_index = NULL;
723
+ Word_t Index;
724
+
725
+ uv_rwlock_rdlock(&pg_cache->metrics_index.lock);
726
+ PValue = JudyHSGet(pg_cache->metrics_index.JudyHS_array, id, sizeof(uuid_t));
727
+ if (likely(NULL != PValue)) {
728
+ page_index = *PValue;
729
+ }
730
+ uv_rwlock_rdunlock(&pg_cache->metrics_index.lock);
731
+
732
+ if ((NULL == PValue) || !pg_cache_try_reserve_pages(ctx, 1)) {
733
+ /* Failed to find page or failed to reserve a spot in the cache */
734
+ return NULL;
735
+ }
736
+
737
+ uv_rwlock_rdlock(&page_index->lock);
738
+ Index = (Word_t)(start_time / USEC_PER_SEC);
739
+ PValue = JudyLGet(page_index->JudyL_array, Index, PJE0);
740
+ if (likely(NULL != PValue)) {
741
+ descr = *PValue;
742
+ }
743
+ if (NULL == PValue || 0 == descr->page_length) {
744
+ /* Failed to find non-empty page */
745
+ uv_rwlock_rdunlock(&page_index->lock);
746
+
747
+ pg_cache_release_pages(ctx, 1);
748
+ return NULL;
749
+ }
750
+
751
+ rrdeng_page_descr_mutex_lock(ctx, descr);
752
+ pg_cache_descr = descr->pg_cache_descr;
753
+ flags = pg_cache_descr->flags;
754
+ uv_rwlock_rdunlock(&page_index->lock);
755
+
756
+ if ((flags & RRD_PAGE_POPULATED) || !pg_cache_try_get_unsafe(descr, 1)) {
757
+ /* Failed to get reference or page is already populated */
758
+ rrdeng_page_descr_mutex_unlock(ctx, descr);
759
+
760
+ pg_cache_release_pages(ctx, 1);
761
+ return NULL;
762
+ }
763
+ /* success */
764
+ rrdeng_page_descr_mutex_unlock(ctx, descr);
765
+ rrd_stat_atomic_add(&ctx->stats.pg_cache_misses, 1);
766
+
767
+ return descr;
768
+}
769
+
770
/**
771
* Searches for pages in a time range and triggers disk I/O if necessary and possible.
772
* Does not get a reference.
database/engine/pagecache.h
+2
@@ -172,6 +172,8 @@ extern usec_t pg_cache_oldest_time_in_range(struct rrdengine_instance *ctx, uuid
172
extern void pg_cache_get_filtered_info_prev(struct rrdengine_instance *ctx, struct pg_cache_page_index *page_index,
173
usec_t point_in_time, pg_cache_page_info_filter_t *filter,
174
struct rrdeng_page_info *page_info);
175
+extern struct rrdeng_page_descr *pg_cache_lookup_unpopulated_and_lock(struct rrdengine_instance *ctx, uuid_t *id,
176
+ usec_t start_time);
177
extern unsigned
178
pg_cache_preload(struct rrdengine_instance *ctx, uuid_t *id, usec_t start_time, usec_t end_time,
179
struct rrdeng_page_info **page_info_arrayp, struct pg_cache_page_index **ret_page_indexp);
database/engine/rrdengine.c
+23
-10
@@ -305,19 +305,32 @@ after_crc_check:
305
}
306
}
307
308
- for (i = 0 ; i < xt_io_descr->descr_count; ++i) {
309
- page = mallocz(RRDENG_BLOCK_SIZE);
310
- descr = xt_io_descr->descr_array[i];
311
- for (j = 0, page_offset = 0; j < count; ++j) {
308
+ for (i = 0, page_offset = 0; i < count; page_offset += header->descr[i++].page_length) {
309
+ uint8_t is_prefetched_page;
310
+ descr = NULL;
311
+ for (j = 0 ; j < xt_io_descr->descr_count; ++j) {
312
+ struct rrdeng_page_descr *descrj;
313
+
314
+ descrj = xt_io_descr->descr_array[j];
315
/* care, we don't hold the descriptor mutex */
313
- if (!uuid_compare(*(uuid_t *) header->descr[j].uuid, *descr->id) &&
314
- header->descr[j].page_length == descr->page_length &&
315
- header->descr[j].start_time == descr->start_time &&
316
- header->descr[j].end_time == descr->end_time) {
316
+ if (!uuid_compare(*(uuid_t *) header->descr[i].uuid, *descrj->id) &&
317
+ header->descr[i].page_length == descrj->page_length &&
318
+ header->descr[i].start_time == descrj->start_time &&
319
+ header->descr[i].end_time == descrj->end_time) {
320
+ descr = descrj;
321
break;
322
}
319
- page_offset += header->descr[j].page_length;
323
}
324
+ is_prefetched_page = 0;
325
+ if (!descr) { /* This extent page has not been requested. Try populating it for locality (best effort). */
326
+ descr = pg_cache_lookup_unpopulated_and_lock(ctx, (uuid_t *)header->descr[i].uuid,
327
+ header->descr[i].start_time);
328
+ if (!descr)
329
+ continue; /* Failed to reserve a suitable page */
330
+ is_prefetched_page = 1;
331
+ }
332
+ page = mallocz(RRDENG_BLOCK_SIZE);
333
+
334
/* care, we don't hold the descriptor mutex */
335
if (have_read_error) {
336
/* Applications should make sure NULL values match 0 as does SN_EMPTY_SLOT */
@@ -334,7 +347,7 @@ after_crc_check:
347
pg_cache_descr->flags &= ~RRD_PAGE_READ_PENDING;
348
rrdeng_page_descr_mutex_unlock(ctx, descr);
349
pg_cache_replaceQ_insert(ctx, descr);
337
- if (xt_io_descr->release_descr) {
350
+ if (xt_io_descr->release_descr || is_prefetched_page) {
351
pg_cache_put(ctx, descr);
352
} else {
353
debug(D_RRDENGINE, "%s: Waking up waiters.", __func__);