On Wed, 22 Nov 2017, NeilBrown wrote:

On Tue, Nov 21 2017, Mikulas Patocka wrote:

quoted

On Tue, 21 Nov 2017, Mike Snitzer wrote:

quoted

On Tue, Nov 21 2017 at  4:23pm -0500,
Mikulas Patocka [off-list ref] wrote:

quoted

This is not correct:

   2206 static void dm_wq_work(struct work_struct *work)
   2207 {
   2208         struct mapped_device *md = container_of(work, struct mapped_device, work);
   2209         struct bio *bio;
   2210         int srcu_idx;
   2211         struct dm_table *map;
   2212
   2213         if (!bio_list_empty(&md->rescued)) {
   2214                 struct bio_list list;
   2215                 spin_lock_irq(&md->deferred_lock);
   2216                 list = md->rescued;
   2217                 bio_list_init(&md->rescued);
   2218                 spin_unlock_irq(&md->deferred_lock);
   2219                 while ((bio = bio_list_pop(&list)))
   2220                         generic_make_request(bio);
   2221         }
   2222
   2223         map = dm_get_live_table(md, &srcu_idx);
   2224
   2225         while (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
   2226                 spin_lock_irq(&md->deferred_lock);
   2227                 bio = bio_list_pop(&md->deferred);
   2228                 spin_unlock_irq(&md->deferred_lock);
   2229
   2230                 if (!bio)
   2231                         break;
   2232
   2233                 if (dm_request_based(md))
   2234                         generic_make_request(bio);
   2235                 else
   2236                         __split_and_process_bio(md, map, bio);
   2237         }
   2238
   2239         dm_put_live_table(md, srcu_idx);
   2240 }

You can see that if we are in dm_wq_work in __split_and_process_bio, we 
will not process md->rescued list.

Can you elaborate further?  We cannot be "in dm_wq_work in
__split_and_process_bio" simultaneously.  Do you mean as a side-effect
of scheduling away from __split_and_process_bio?

The more detail you can share the better.

Suppose this scenario:

* dm_wq_work calls __split_and_process_bio
* __split_and_process_bio eventually reaches the function snapshot_map
* snapshot_map attempts to take the snapshot lock

* the snapshot lock could be released only if some bios submitted by the 
snapshot driver to the underlying device complete
* the bios submitted to the underlying device were already offloaded by 
some other task and they are waiting on the list md->rescued
* the bios waiting on md->rescued are not processed, because dm_wq_work is 
blocked in snapshot_map (called from __split_and_process_bio)

Yes, I think you are right. 

I think the solution is to get rid of the dm_offload() infrastructure
and make it not necessary.
i.e. discard my patches
    dm: prepare to discontinue use of BIOSET_NEED_RESCUER
and
    dm: revise 'rescue' strategy for bio-based bioset allocations

And build on "dm: ensure bio submission follows a depth-first tree walk"
which was written after those and already makes dm_offload() less
important.

Since that "depth-first" patch, every request to the dm device, after
the initial splitting, allocates just one dm_target_io structure, and
makes just one __map_bio() call, and so will behave exactly the way
generic_make_request() expects and copes with - thus avoiding awkward
dependencies and deadlocks.  Except....

a/ If any target defines ->num_write_bios() to return >1,
   __clone_and_map_data_bio() will make multiple calls to alloc_tio()
   and __map_bio(), which might need rescuing.
   But no target defines num_write_bios, and none have since it was
   removed from dm-cache 4.5 years ago.
   Can we discard num_write_bios??

b/ If any target sets any of num_{flush,discard,write_same,write_zeroes}_bios
   to a value > 1, then __send_duplicate_bios() will also make multiple
   calls to alloc_tio() and __map_bio().
   Some do.
     dm-cache-target:  flush=2
     dm-snap: flush=2
     dm-stripe: discard, write_same, write_zeroes all set to 'stripes'.

These will only be a problem if the second (or subsequent) alloc_tio()
blocks waiting for an earlier allocation to complete.  This will only
be a problem if multiple threads are each trying to allocate multiple
dm_target_io from the same bioset at the same time.
This is rare and should be easier to address than the current
dm_offload() approach. 
One possibility would be to copy the approach taken by
crypt_alloc_buffer() which needs to allocate multiple entries from a
mempool.
It first tries the with GFP_NOWAIT.  If that fails it take a mutex and
tries with GFP_NOIO.  This mean only one thread will try to allocate
multiple bios at once, so there can be no deadlock.

Below are two RFC patches.  The first removes num_write_bios.
The second is incomplete and makes a stab are allocating multiple bios
at once safely.
A third would be needed to remove dm_offload() etc... but I cannot quite
fit that in today - must be off.

Thanks,
NeilBrown

Another problem is this:

struct bio *b = bio_clone_bioset(bio, GFP_NOIO, md->queue->bio_split);
bio_advance(b, (bio_sectors(b) - ci.sector_count) << 9);
bio_chain(b, bio);

What if it blocks because the bioset is exhausted?

The code basically builds a chain of bios of unlimited length (suppose for 
example a case when we are splitting on every sector boundary, so there 
will be one bio for every sector in the original bio), it could exhaust 
the bioset easily.

It would be better to use mechanism from md-raid that chains all the 
sub-bios to the same master bio and doesn't create long chains of bios:

        if (max_sectors < bio_sectors(bio)) {
                struct bio *split = bio_split(bio, max_sectors,
                                              gfp, conf->bio_split);
                bio_chain(split, bio);
                generic_make_request(bio);
                bio = split;
                r1_bio->master_bio = bio;
                r1_bio->sectors = max_sectors;
        }

Mikulas