process_prepared_discard_passdown_pt1() should cleanup
dm_thin_new_mapping in cases of error.
dm_pool_inc_data_range() can fail trying to get a block reference:
metadata operation 'dm_pool_inc_data_range' failed: error = -61
When dm_pool_inc_data_range() fails, dm thin aborts current metadata
transaction and marks pool as PM_READ_ONLY. Memory for thin mapping
is released as well. However, current thin mapping will be queued
onto next stage as part of queue_passdown_pt2() or passdown_endio().
This dangling thin mapping memory when processed and accessed in
next stage will lead to device mapper crashing.
Code flow without fix:
-> process_prepared_discard_passdown_pt1(m)
-> dm_thin_remove_range()
-> discard passdown
--> passdown_endio(m) queues m onto next stage
-> dm_pool_inc_data_range() fails, frees memory m
but does not remove it from next stage queue
-> process_prepared_discard_passdown_pt2(m)
-> processes freed memory m and crashes
One such stack:
Call Trace:
[<ffffffffa037a46f>] dm_cell_release_no_holder+0x2f/0x70 [dm_bio_prison]
[<ffffffffa039b6dc>] cell_defer_no_holder+0x3c/0x80 [dm_thin_pool]
[<ffffffffa039b88b>] process_prepared_discard_passdown_pt2+0x4b/0x90 [dm_thin_pool]
[<ffffffffa0399611>] process_prepared+0x81/0xa0 [dm_thin_pool]
[<ffffffffa039e735>] do_worker+0xc5/0x820 [dm_thin_pool]
[<ffffffff8152bf54>] ? __schedule+0x244/0x680
[<ffffffff81087e72>] ? pwq_activate_delayed_work+0x42/0xb0
[<ffffffff81089f53>] process_one_work+0x153/0x3f0
[<ffffffff8108a71b>] worker_thread+0x12b/0x4b0
[<ffffffff8108a5f0>] ? rescuer_thread+0x350/0x350
[<ffffffff8108fd6a>] kthread+0xca/0xe0
[<ffffffff8108fca0>] ? kthread_park+0x60/0x60
[<ffffffff81530b45>] ret_from_fork+0x25/0x30
The fix is to first take the block ref count for discarded block and
then do a passdown discard of this block. If block ref count fails,
then bail out aborting current metadata transaction, mark pool as
PM_READ_ONLY and also free current thin mapping memory (existing error
handling code) without queueing this thin mapping onto next stage of
processing. If block ref count succeeds, then passdown discard of this
block. Discard callback of passdown_endio() will queue this thin mapping
onto next stage of processing.
Code flow with fix:
-> process_prepared_discard_passdown_pt1(m)
-> dm_thin_remove_range()
-> dm_pool_inc_data_range()
--> if fails, free memory m and bail out
-> discard passdown
--> passdown_endio(m) queues m onto next stage
Cc: stable <stable@vger.kernel.org> # v4.9+
Reviewed-by: Eduardo Valentin <eduval@amazon.com>
Reviewed-by: Cristian Gafton <gafton@amazon.com>
Reviewed-by: Anchal Agarwal <anchalag@amazon.com>
Signed-off-by: Vallish Vaidyeshwara <vallish@amazon.com>
Reviewed-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>