dn_dirty_ctx is always set to the highest txg that has ever dirtied the
dnode. It is set in dbuf_dirty() when a data or metadnode dbuf is
dirtied, and never cleared.
[analysis of bug #15526 and fix#15571 below, for future readers]
The previous dirty check was:
for (int i = 0; i < TXG_SIZE; i++) {
if (multilist_link_active(&dn->dn_dirty_link[i])
[dnode is dirty]
However, this check is not "is the dnode dirty?" but rather, "is the
dnode on a list?".
There is a gap in dmu_objset_sync_dnodes() where the dnode is moved from
os_dirty_dnodes to os_synced_dnodes, before dnode_sync() is called to
write out the dirty dbufs. So, there is a moment when the dnode is not
on a list, and so the check fails.
It doesn't matter that the dirty check takes dn_mtx, because that lock
isn't used for dn_dirty_link. The os_dirty_dnodes sublist lock is held
in dmu_objset_sync_dnodes(), but trying to take that would mean possibly
waiting until everything on that sublist has been synced.
The correct fix has to check something that positively asserts the dnode
is dirty, rather than an implementation detail. dn_dirty_txg (via
DNODE_IS_DIRTY()) is that - its a normal bit of dnode state, under the
dn_mtx lock, and unambiguously indicates whether or not there's changes
pending.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
531af601a1