Illumos 5694 - traverse_prefetcher does not prefetch enough

5694 traverse_prefetcher does not prefetch enough
Reviewed by: Matthew Ahrens <[email protected]>
Reviewed by: Alex Reece <[email protected]>
Reviewed by: Christopher Siden <[email protected]>
Reviewed by: Josef 'Jeff' Sipek <[email protected]>
Reviewed by: Bayard Bell <[email protected]>
Approved by: Garrett D'Amore <[email protected]>

References:
  https://www.illumos.org/issues/5694
  https://github.com/illumos/illumos-gate/commit/34d7ce05

Ported-by: Chris Dunlop <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #3230

1 files changed, 11 insertions, 12 deletions

diff --git a/module/zfs/dmu_traverse.c b/module/zfs/dmu_traverse.c
index a8481aee6..9280a89b2 100644
--- a/module/zfs/dmu_traverse.c
+++ b/module/zfs/dmu_traverse.c
@@ -38,13 +38,12 @@
 #include <sys/callb.h>
 #include <sys/zfeature.h>
 
-int zfs_pd_blks_max = 100;
+int32_t zfs_pd_bytes_max = 50 * 1024 * 1024;	/* 50MB */
 
 typedef struct prefetch_data {
 	kmutex_t pd_mtx;
 	kcondvar_t pd_cv;
-	int pd_blks_max;
-	int pd_blks_fetched;
+	int32_t pd_bytes_fetched;
 	int pd_flags;
 	boolean_t pd_cancel;
 	boolean_t pd_exited;
@@ -251,11 +250,12 @@ traverse_visitbp(traverse_data_t *td, const dnode_phys_t *dnp,
 	}
 
 	if (pd != NULL && !pd->pd_exited && prefetch_needed(pd, bp)) {
+		uint64_t size = BP_GET_LSIZE(bp);
 		mutex_enter(&pd->pd_mtx);
-		ASSERT(pd->pd_blks_fetched >= 0);
-		while (pd->pd_blks_fetched == 0 && !pd->pd_exited)
+		ASSERT(pd->pd_bytes_fetched >= 0);
+		while (pd->pd_bytes_fetched < size && !pd->pd_exited)
 			cv_wait(&pd->pd_cv, &pd->pd_mtx);
-		pd->pd_blks_fetched--;
+		pd->pd_bytes_fetched -= size;
 		cv_broadcast(&pd->pd_cv);
 		mutex_exit(&pd->pd_mtx);
 	}
@@ -452,7 +452,7 @@ traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 	prefetch_data_t *pfd = arg;
 	uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
 
-	ASSERT(pfd->pd_blks_fetched >= 0);
+	ASSERT(pfd->pd_bytes_fetched >= 0);
 	if (pfd->pd_cancel)
 		return (SET_ERROR(EINTR));
 
@@ -460,9 +460,9 @@ traverse_prefetcher(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
 		return (0);
 
 	mutex_enter(&pfd->pd_mtx);
-	while (!pfd->pd_cancel && pfd->pd_blks_fetched >= pfd->pd_blks_max)
+	while (!pfd->pd_cancel && pfd->pd_bytes_fetched >= zfs_pd_bytes_max)
 		cv_wait(&pfd->pd_cv, &pfd->pd_mtx);
-	pfd->pd_blks_fetched++;
+	pfd->pd_bytes_fetched += BP_GET_LSIZE(bp);
 	cv_broadcast(&pfd->pd_cv);
 	mutex_exit(&pfd->pd_mtx);
 
@@ -531,7 +531,6 @@ traverse_impl(spa_t *spa, dsl_dataset_t *ds, uint64_t objset, blkptr_t *rootbp,
 	td->td_flags = flags;
 	td->td_paused = B_FALSE;
 
-	pd->pd_blks_max = zfs_pd_blks_max;
 	pd->pd_flags = flags;
 	mutex_init(&pd->pd_mtx, NULL, MUTEX_DEFAULT, NULL);
 	cv_init(&pd->pd_cv, NULL, CV_DEFAULT, NULL);
@@ -661,6 +660,6 @@ traverse_pool(spa_t *spa, uint64_t txg_start, int flags,
 EXPORT_SYMBOL(traverse_dataset);
 EXPORT_SYMBOL(traverse_pool);
 
-module_param(zfs_pd_blks_max, int, 0644);
-MODULE_PARM_DESC(zfs_pd_blks_max, "Max number of blocks to prefetch");
+module_param(zfs_pd_bytes_max, int, 0644);
+MODULE_PARM_DESC(zfs_pd_bytes_max, "Max number of bytes to prefetch");
 #endif