Consider the following case -
1. A file gets FALLOCATE'd such that > "shard-lru-limit" number of
shards are created.
2. And then it is deleted after that.
The unique thing about FALLOCATE is that unlike WRITE, all of the
participant shards are resolved and created and fallocated in a single
batch. This means, in this case, after the first "shard-lru-limit"
number of shards are resolved and added to lru list, as part of
resolution of the remaining shards, some of the existing shards in lru
list will need to be evicted. So these evicted shards will be
inode_unlink()d as part of eviction. Now once the fop gets to the actual
FALLOCATE stage, the lru'd-out shards get added to fsync list.
2 things to note at this point:
i. the lru'd out shards are only part of fsync list, so each holds 1 ref
on base shard
ii. and the more recently used shards are part of both fsync and lru list.
So each of these shards holds 2 refs on base inode - one for being
part of fsync list, and the other for being part of lru list.
FALLOCATE completes successfully and then this very file is deleted, and
background shard deletion launched. Here's where the ref counts get mismatched.
First as part of inode_resolve()s during the deletion, the lru'd-out inodes
return NULL, because they are inode_unlink()'d by now. So these inodes need to
be freshly looked up. But as part of linking them in lookup_cbk (precisely in
shard_link_block_inode()), inode_link() returns the lru'd-out inode object.
And its inode ctx is still valid and ctx->base_inode valid from the last
time it was added to list.
But shard_common_lookup_shards_cbk() passes NULL in the place of base_pointer
to __shard_update_shards_inode_list(). This means, as part of adding the lru'd out
inode back to lru list, base inode is not ref'd since its NULL.
Whereas post unlinking this shard, during shard_unlink_block_inode(),
ctx->base_inode is accessible and is unref'd because the shard was found to be part
of LRU list, although the matching ref didn't occur. This at some point leads to
base_inode refcount becoming 0 and it getting destroyed and released back while some
of its associated shards are continuing to be unlinked in parallel and the client crashes
whenever it is accessed next.
Fix is to pass base shard correctly, if available, in shard_link_block_inode().
Also, the patch fixes the ret value check in tests/bugs/shard/shard-fallocate.c
Change-Id: Ibd0bc4c6952367608e10701473cbad3947d7559f
Updates: bz#1696136
Signed-off-by: Krutika Dhananjay <kdhananj@redhat.com>