dn_bonus evicted too early
o Fix bonus buffer eviction regression introduced by the fixes for
illumos issue 5056.
o Remove assert that can trigger due to a long standing race between
draining references on dbufs and processing of freed dnodes in the
syncer. References are guaranteed to drain, and the introduction
of db_pending_evict ensures proper dbuf eviction when references
finally do drain.
The bonus buffer associated with a dnode is expected to remain resident
until: the dnode is evicted via dnode_buf_pageout(), the dnode is
freed in dnode_sync_free(), or the objset containing the dnode is
evicted via dmu_objset_evict(). However, since bonus buffers (and DMU
buffers in general) can have draining references when these events occur,
dbuf_rele_and_unlock() has logic to ensure that once these late references
are released affected buffers will be evicted.
dbuf_rele_and_unlock() currently checks for a dbuf for an evicting
objset via the os->os_evicting flag, and detects buffers for a freed
dnode by testing dn->dn_type and dn->dn_free_txg fields. Unfortunately,
the free'd dnode test can fire prematurely - anytime after the dnode is
scheduled to be freed via dnode_free() until the free is committed in
dnode_sync_free(). If all references to the bonus buffer are dropped
within this window, the bonus buffer will be evicted and code in
dnode_sync() that relies on the bonus buffer will fail.
Additionally, the "free'd dnode test" isn't applied to normal buffers
(buffers that are not the bonus buffer) and there is no mechanism to
guarantee eviction in the dnode_buf_pageout() case (the dnode is not
being freed nor is the objset being evicted).
Replace the two existing deferred eviction mechanisms with a per-dbuf
flag, db_pending_evict. This is set when explicit eviction is requested
via either dnode_evict_dbufs() or dnode_evict_bonus(). These actions
only occur after it is safe for dnode buffers to be evicted (e.g. the
bonus buffer will not be referenced again).
Add comments for boolean fields in dmu_buf_impl_t.
Add the db_pending_evict field.
Rename db_immediate_evict to db_user_immediate_evict to avoid
confusion between dbuf user state eviction and deferred eviction
of a dbuf.
Consistently use TRUE/FALSE for boolean fields in
Simplify pending eviction logic to use the new db_pending_evict
flag in all cases.
Remove objset_t's os_evicting field. This same functionality
is now provided by db_pending_evict.
In dnode_evict_dbufs() and dnode_evict_bonus(), mark dbufs
with draining references (dbufs that can't be evicted inline)
In dnode_sync_free(), remove ASSERT() that a dnode being free'd
has no active dbufs. This is usually the case, but is not
guaranteed due to draining references. (e.g. The deadlist for a
deleted dataset may still be open if another thread referenced
the dataset just before it was freed and the dsl_dataset_t hasn't
been released or is still being evicted).
Updated by Electric Monk about 4 years ago
- Status changed from New to Closed
- % Done changed from 0 to 100
commit d2058105c61ec61df3a2dd3f839fed8c3fe7bfd6 Author: Justin T. Gibbs <gibbs@FreeBSD.org> Date: 2015-10-11T16:16:29.000Z 6267 dn_bonus evicted too early Reviewed by: Richard Yao <email@example.com> Reviewed by: Xin LI <firstname.lastname@example.org> Reviewed by: Matthew Ahrens <email@example.com> Approved by: Richard Lowe <firstname.lastname@example.org>