zfetch: Don't issue new streams when old have not completed

The current dmu_zfetch code implicitly assumes that I/Os complete
within min_sec_reap seconds. With async dmu and a readonly workload
(and thus no exponential backoff in operations from the "write
throttle") such as L2ARC rebuild it is possible to saturate the drives
with I/O requests. These are then effectively compounded with prefetch
requests.

This change reference counts streams and prevents them from being
recycled after their min_sec_reap timeout if they still have
outstanding I/Os.

Reviewed-by: Alexander Motin <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Matt Macy <[email protected]>
Closes #10900 

1 files changed, 6 insertions, 1 deletions

diff --git a/include/sys/dbuf.h b/include/sys/dbuf.h
index 04338b2c4..ca2154e12 100644
--- a/include/sys/dbuf.h
+++ b/include/sys/dbuf.h
@@ -309,6 +309,8 @@ typedef struct dbuf_hash_table {
 	kmutex_t hash_mutexes[DBUF_MUTEXES];
 } dbuf_hash_table_t;
 
+typedef void (*dbuf_prefetch_fn)(void *, boolean_t);
+
 uint64_t dbuf_whichblock(const struct dnode *di, const int64_t level,
     const uint64_t offset);
 
@@ -324,7 +326,10 @@ int dbuf_hold_impl(struct dnode *dn, uint8_t level, uint64_t blkid,
     boolean_t fail_sparse, boolean_t fail_uncached,
     void *tag, dmu_buf_impl_t **dbp);
 
-void dbuf_prefetch(struct dnode *dn, int64_t level, uint64_t blkid,
+int dbuf_prefetch_impl(struct dnode *dn, int64_t level, uint64_t blkid,
+    zio_priority_t prio, arc_flags_t aflags, dbuf_prefetch_fn cb,
+    void *arg);
+int dbuf_prefetch(struct dnode *dn, int64_t level, uint64_t blkid,
     zio_priority_t prio, arc_flags_t aflags);
 
 void dbuf_add_ref(dmu_buf_impl_t *db, void *tag);