diff options
| author | Tom Caputi <[email protected]> | 2019-09-11 14:16:48 -0400 |
|---|---|---|
| committer | Brian Behlendorf <[email protected]> | 2019-09-11 11:16:48 -0700 |
| commit | 5815f7ac30e108fcbf4c6487328c28d818e9e014 (patch) | |
| tree | 7ef5339fe09e98b3fdb7ef768a78fa43a1d0c68b /module/zfs/dsl_scan.c | |
| parent | 5b51c15861a0c066b25bfa4741a6df16d0c63884 (diff) | |
Fix stalled txg with repeated noop scans
Currently, the DSL scan code figures out when it should suspend
processing and allow a txg to continue by calling the function
dsl_scan_check_suspend(). Unfortunately, this function only
allows the scan to suspend at a level 0 block. In the event that
the system is scanning a bunch of empty snapshots or a resilver
is running with a high enough scn_cur_min_txg, the scan will
stop processing each dataset at the root level, deciding it
has nothing left to do. This means that the check_suspend
function is never called and the txg remains stuck until a
dataset is found that has data to scan.
This patch fixes the problem by allowing scans to suspend at
the root level of the objset. For backwards compatibility, we
use the bookmark <objsetid, 0, 0, 0> when we suspend here so
that older versions of the code will work as intended.
Reviewed-by: Matt Ahrens <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Tom Caputi <[email protected]>
Closes #9300
Diffstat (limited to 'module/zfs/dsl_scan.c')
| -rw-r--r-- | module/zfs/dsl_scan.c | 15 |
1 files changed, 12 insertions, 3 deletions
diff --git a/module/zfs/dsl_scan.c b/module/zfs/dsl_scan.c index 160bc8ff3..ec71a6d91 100644 --- a/module/zfs/dsl_scan.c +++ b/module/zfs/dsl_scan.c @@ -1272,8 +1272,8 @@ dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb) if (!ZB_IS_ZERO(&scn->scn_phys.scn_bookmark)) return (B_FALSE); /* we're resuming */ - /* We only know how to resume from level-0 blocks. */ - if (zb && zb->zb_level != 0) + /* We only know how to resume from level-0 and objset blocks. */ + if (zb && (zb->zb_level != 0 && zb->zb_level != ZB_ROOT_LEVEL)) return (B_FALSE); /* @@ -1304,7 +1304,16 @@ dsl_scan_check_suspend(dsl_scan_t *scn, const zbookmark_phys_t *zb) NSEC2SEC(sync_time_ns) >= zfs_txg_timeout)) || spa_shutting_down(scn->scn_dp->dp_spa) || (zfs_scan_strict_mem_lim && dsl_scan_should_clear(scn))) { - if (zb) { + if (zb && zb->zb_level == ZB_ROOT_LEVEL) { + dprintf("suspending at first available bookmark " + "%llx/%llx/%llx/%llx\n", + (longlong_t)zb->zb_objset, + (longlong_t)zb->zb_object, + (longlong_t)zb->zb_level, + (longlong_t)zb->zb_blkid); + SET_BOOKMARK(&scn->scn_phys.scn_bookmark, + zb->zb_objset, 0, 0, 0); + } else if (zb != NULL) { dprintf("suspending at bookmark %llx/%llx/%llx/%llx\n", (longlong_t)zb->zb_objset, (longlong_t)zb->zb_object, |