Skip to content
Commit 18a25da8 authored by NeilBrown's avatar NeilBrown Committed by Mike Snitzer
Browse files

dm: ensure bio submission follows a depth-first tree walk



A dm device can, in general, represent a tree of targets, each of which
handles a sub-range of the range of blocks handled by the parent.

The bio sequencing managed by generic_make_request() requires that bios
are generated and handled in a depth-first manner.  Each call to a
make_request_fn() may submit bios to a single member device, and may
submit bios for a reduced region of the same device as the
make_request_fn.

In particular, any bios submitted to member devices must be expected to
be processed in order, so a later one must never wait for an earlier
one.

This ordering is usually achieved by using bio_split() to reduce a bio
to a size that can be completely handled by one target, and resubmitting
the remainder to the originating device. bio_queue_split() shows the
canonical approach.

dm doesn't follow this approach, largely because it has needed to split
bios since long before bio_split() was available.  It currently can
submit bios to separate targets within the one dm_make_request() call.
Dependencies between these targets, as can happen with dm-snap, can
cause deadlocks if either bios gets stuck behind the other in the queues
managed by generic_make_request().  This requires the 'rescue'
functionality provided by dm_offload_{start,end}.

Some of this requirement can be removed by changing the order of bio
submission to follow the canonical approach.  That is, if dm finds that
it needs to split a bio, the remainder should be sent to
generic_make_request() rather than being handled immediately.  This
delays the handling until the first part is completely processed, so the
deadlock problems do not occur.

__split_and_process_bio() can be called both from dm_make_request() and
from dm_wq_work().  When called from dm_wq_work() the current approach
is perfectly satisfactory as each bio will be processed immediately.
When called from dm_make_request(), current->bio_list will be non-NULL,
and in this case it is best to create a separate "clone" bio for the
remainder.

When we use bio_clone_bioset() to split off the front part of a bio
and chain the two together and submit the remainder to
generic_make_request(), it is important that the newly allocated
bio is used as the head to be processed immediately, and the original
bio gets "bio_advance()"d and sent to generic_make_request() as the
remainder.  Otherwise, if the newly allocated bio is used as the
remainder, and if it then needs to be split again, then the next
bio_clone_bioset() call will be made while holding a reference a bio
(result of the first clone) from the same bioset.  This can potentially
exhaust the bioset mempool and result in a memory allocation deadlock.

Note that there is no race caused by reassigning cio.io->bio after already
calling __map_bio().  This bio will only be dereferenced again after
dec_pending() has found io->io_count to be zero, and this cannot happen
before the dec_pending() call at the end of __split_and_process_bio().

To provide the clone bio when splitting, we use q->bio_split.  This
was previously being freed by bio-based dm to avoid having excess
rescuer threads.  As bio_split bio sets no longer create rescuer
threads, there is little cost and much gain from restoring the
q->bio_split bio set.

Signed-off-by: default avatarNeilBrown <neilb@suse.com>
Signed-off-by: default avatarMike Snitzer <snitzer@redhat.com>
parent c110a4b6
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment