block: make sure a big bio is split into at most 256 bvecs
authorMing Lei <ming.lei@canonical.com>
Tue, 23 Aug 2016 13:49:45 +0000 (21:49 +0800)
committerJens Axboe <axboe@fb.com>
Wed, 24 Aug 2016 14:17:24 +0000 (08:17 -0600)
After arbitrary bio size was introduced, the incoming bio may
be very big. We have to split the bio into small bios so that
each holds at most BIO_MAX_PAGES bvecs for safety reason, such
as bio_clone().

This patch fixes the following kernel crash:

> [  172.660142] BUG: unable to handle kernel NULL pointer dereference at 0000000000000028
> [  172.660229] IP: [<ffffffff811e53b4>] bio_trim+0xf/0x2a
> [  172.660289] PGD 7faf3e067 PUD 7f9279067 PMD 0
> [  172.660399] Oops: 0000 [#1] SMP
> [...]
> [  172.664780] Call Trace:
> [  172.664813]  [<ffffffffa007f3be>] ? raid1_make_request+0x2e8/0xad7 [raid1]
> [  172.664846]  [<ffffffff811f07da>] ? blk_queue_split+0x377/0x3d4
> [  172.664880]  [<ffffffffa005fb5f>] ? md_make_request+0xf6/0x1e9 [md_mod]
> [  172.664912]  [<ffffffff811eb860>] ? generic_make_request+0xb5/0x155
> [  172.664947]  [<ffffffffa0445c89>] ? prio_io+0x85/0x95 [bcache]
> [  172.664981]  [<ffffffffa0448252>] ? register_cache_set+0x355/0x8d0 [bcache]
> [  172.665016]  [<ffffffffa04497d3>] ? register_bcache+0x1006/0x1174 [bcache]

The issue can be reproduced by the following steps:
- create one raid1 over two virtio-blk
- build bcache device over the above raid1 and another cache device
and bucket size is set as 2Mbytes
- set cache mode as writeback
- run random write over ext4 on the bcache device

Fixes: 54efd50(block: make generic_make_request handle arbitrarily sized bios)
Reported-by: Sebastian Roesner <sroesner-kernelorg@roesner-online.de>
Reported-by: Eric Wheeler <bcache@lists.ewheeler.net>
Cc: stable@vger.kernel.org (4.3+)
Cc: Shaohua Li <shli@fb.com>
Acked-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-merge.c

index 72627e3..2642e5f 100644 (file)
@@ -94,9 +94,31 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
        bool do_split = true;
        struct bio *new = NULL;
        const unsigned max_sectors = get_max_io_size(q, bio);
+       unsigned bvecs = 0;
 
        bio_for_each_segment(bv, bio, iter) {
                /*
+                * With arbitrary bio size, the incoming bio may be very
+                * big. We have to split the bio into small bios so that
+                * each holds at most BIO_MAX_PAGES bvecs because
+                * bio_clone() can fail to allocate big bvecs.
+                *
+                * It should have been better to apply the limit per
+                * request queue in which bio_clone() is involved,
+                * instead of globally. The biggest blocker is the
+                * bio_clone() in bio bounce.
+                *
+                * If bio is splitted by this reason, we should have
+                * allowed to continue bios merging, but don't do
+                * that now for making the change simple.
+                *
+                * TODO: deal with bio bounce's bio_clone() gracefully
+                * and convert the global limit into per-queue limit.
+                */
+               if (bvecs++ >= BIO_MAX_PAGES)
+                       goto split;
+
+               /*
                 * If the queue doesn't support SG gaps and adding this
                 * offset would create a gap, disallow it.
                 */