imrcpan/freebsd
1
0 Fork 0
mirror of https://github.com/freebsd/freebsd-src.git synced 2026-10-28 11:25:21 +01:00
Code Issues Packages Projects Releases Wiki Activity

RAIDZ: Fix parity regeneration/check condition

Browse Source 
Profiling RAIDZ1/dRAID1 resilver I've noticed that they calculate
the parity twice for most of blocks: first to reconstruct the data
column and then to "verify" the parity column.  Same time it is
obvious that parity generated from data reconstructed from the
parity will be identical to the original.  The code even had this
condition, but it was overridden by ZIO_FLAG_RESILVER check.

I think the ZIO_FLAG_RESILVER condition is not right.  Instead we
should check for parity_errors > 0, when we failed to read some
parity columns that we'll need to rewrite.  It should not matter
if we are resilvering or just doing self healing on regular read.

Profiling shows this saving ~16% of ZFS CPU time when resilvering
RAIDZ1.  RAIDZ2+ are out of luck, unless two disks are replaced
same time, since there is still a parity to verify.

Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Closes #18707
This commit is contained in:
Alexander Motin authored and GitHub committed 2026-06-29 15:23:23 -07:00
1 parent 19d69a9193
commit 0d1f3b1c60
1 file changed
+13 -5
Download Patch File Download Diff File Expand all files Collapse all files
module/zfs/vdev_raidz.c
+13 -5
View File
@@ -3153,6 +3153,7 @@ vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
zio_flag_t add_flags = 0;
ASSERT3U(zio->io_type, ==, ZIO_TYPE_READ);
ASSERT0(zio->io_error);
for (int c = 0; c < rr->rr_cols; c++) {
raidz_col_t *rc = &rr->rr_col[c];
@@ -3178,12 +3179,19 @@ vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
* reconstruction, confirm that the other parity disks produced
* correct data.
*
* Note that we also regenerate parity when resilvering so we
* can write it out to failed devices later.
* We also regenerate parity to write it back to any failed parity
* columns. However, if all available parity was consumed by
* reconstruction (parity_verify is false), regenerating parity is
* a mathematical identity -- the result is guaranteed to equal the
* input that was used for reconstruction, whether correct or
* corrupted. In that case the only reason to regenerate is to
* write back a failed parity column, so skip regeneration when no
* parity column failed or the pool is read-only.
*/
boolean_t parity_verify = (parity_errors + parity_untried) <
(rr->rr_firstdatacol - data_errors);
if (parity_verify || (zio->io_flags & ZIO_FLAG_RESILVER)) {
if (parity_verify || (parity_errors > 0 &&
spa_writeable(zio->io_spa))) {
int n = raidz_parity_verify(zio, rr);
/*
* In, Reed-Solomon encoding, if we have ndata+1 columns and
@@ -3208,7 +3216,7 @@ vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
unexpected_errors += n;
}
if (zio->io_error == 0 && spa_writeable(zio->io_spa) &&
if (spa_writeable(zio->io_spa) &&
(unexpected_errors > 0 || (zio->io_flags & ZIO_FLAG_RESILVER))) {
/*
* Use the good data we have in hand to repair damaged children.
@@ -3261,7 +3269,7 @@ vdev_raidz_io_done_verified(zio_t *zio, raidz_row_t *rr)
* necessary, but since expansion is paused during scrub/resilver, at
* most a single row will have a shadow location.
*/
if (zio->io_error == 0 && spa_writeable(zio->io_spa) &&
if (spa_writeable(zio->io_spa) &&
(zio->io_flags & (ZIO_FLAG_RESILVER | ZIO_FLAG_SCRUB))) {
for (int c = 0; c < rr->rr_cols; c++) {
raidz_col_t *rc = &rr->rr_col[c];