Improve zfs receive performance with lightweight write

The performance of `zfs receive` can be bottlenecked on the CPU consumed
by the `receive_writer` thread, especially when receiving streams with
small compressed block sizes.  Much of the CPU is spent creating and
destroying dbuf's and arc buf's, one for each `WRITE` record in the send
stream.

This commit introduces the concept of "lightweight writes", which allows
`zfs receive` to write to the DMU by providing an ABD, and instantiating
only a new type of `dbuf_dirty_record_t`.  The dbuf and arc buf for this
"dirty leaf block" are not instantiated.

Because there is no dbuf with the dirty data, this mechanism doesn't
support reading from "lightweight-dirty" blocks (they would see the
on-disk state rather than the dirty data).  Since the dedup-receive code
has been removed, `zfs receive` is write-only, so this works fine.

Because there are no arc bufs for the received data, the received data
is no longer cached in the ARC.

Testing a receive of a stream with average compressed block size of 4KB,
this commit improves performance by 50%, while also reducing CPU usage
by 50% of a CPU.  On a per-block basis, CPU consumed by receive_writer()
and dbuf_evict() is now 1/7th (14%) of what it was.

Baseline: 450MB/s, CPU in receive_writer() 40% + dbuf_evict() 35%
New: 670MB/s, CPU in receive_writer() 17% + dbuf_evict() 0%

The code is also restructured in a few ways:

Added a `dr_dnode` field to the dbuf_dirty_record_t.  This simplifies
some existing code that no longer needs `DB_DNODE_ENTER()` and related
routines.  The new field is needed by the lightweight-type dirty record.

To ensure that the `dr_dnode` field remains valid until the dirty record
is freed, we have to ensure that the `dnode_move()` doesn't relocate the
dnode_t.  To do this we keep a hold on the dnode until it's zio's have
completed.  This is already done by the user-accounting code
(`userquota_updates_task()`), this commit extends that so that it always
keeps the dnode hold until zio completion (see `dnode_rele_task()`).

`dn_dirty_txg` was previously zeroed when the dnode was synced.  This
was not necessary, since its meaning can be "when was this dnode last
dirtied".  This change simplifies the new `dnode_rele_task()` code.

Removed some dead code related to `DRR_WRITE_BYREF` (dedup receive).

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Paul Dagnelie <[email protected]>
Reviewed-by: George Wilson <[email protected]>
Signed-off-by: Matthew Ahrens <[email protected]>
Closes #11105 

1 files changed, 2 insertions, 4 deletions

diff --git a/module/zfs/dsl_dataset.c b/module/zfs/dsl_dataset.c
index 1fcd83db7..de60c3358 100644
--- a/module/zfs/dsl_dataset.c
+++ b/module/zfs/dsl_dataset.c
@@ -2267,10 +2267,8 @@ dsl_dataset_sync_done(dsl_dataset_t *ds, dmu_tx_t *tx)
 
 	dsl_bookmark_sync_done(ds, tx);
 
-	if (os->os_synced_dnodes != NULL) {
-		multilist_destroy(os->os_synced_dnodes);
-		os->os_synced_dnodes = NULL;
-	}
+	multilist_destroy(os->os_synced_dnodes);
+	os->os_synced_dnodes = NULL;
 
 	if (os->os_encrypted)
 		os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_FALSE;