KASAN warning in bt_for_each

From: Evan Green
Date: Thu Feb 14 2019 - 13:05:03 EST

Greetings, block experts!

I'm trying to track down a KASAN warning I'm seeing in our downstream
4.19 kernel, and I could use a little help. The warning looks like

[ 224.564894] BUG: KASAN: use-after-free in bt_for_each+0x1ac/0x28c
[ 224.571195] Read of size 8 at addr ffffffc17c621340 by task fio/3851
[ 224.577745]
[ 224.579320] CPU: 2 PID: 3851 Comm: fio Tainted: G W
4.19.16 #398
[ 224.591905] Call trace:
[ 224.594465] dump_backtrace+0x0/0x2d8
[ 224.598263] show_stack+0x20/0x2c
[ 224.601708] __dump_stack+0x20/0x28
[ 224.605329] dump_stack+0xc8/0xec
[ 224.608769] print_address_description+0x74/0x240
[ 224.613632] kasan_report+0x258/0x274
[ 224.617431] __asan_report_load8_noabort+0x20/0x28
[ 224.622384] bt_for_each+0x1ac/0x28c
[ 224.626090] blk_mq_queue_tag_busy_iter+0x1d4/0x2f8
[ 224.631127] blk_mq_in_flight+0xa0/0x100
[ 224.635193] part_in_flight+0x54/0x19c
[ 224.639079] part_round_stats+0x138/0x1f0
[ 224.643230] blk_account_io_start+0x23c/0x54c
[ 224.647742] blk_mq_bio_to_request+0x24/0x2c
[ 224.652165] blk_mq_make_request+0x674/0xf90
[ 224.656581] generic_make_request+0x32c/0x7b4
[ 224.661090] submit_bio+0x230/0x464
[ 224.664702] __blkdev_direct_IO_simple+0x3d0/0x79c
[ 224.669656] blkdev_direct_IO+0x60/0x88
[ 224.673629] generic_file_read_iter+0x368/0xeb8
[ 224.678308] blkdev_read_iter+0xc0/0xd0
[ 224.682281] __vfs_read+0x1d8/0x2f0
[ 224.685900] vfs_read+0xd0/0x1bc
[ 224.689249] ksys_pread64+0x9c/0xec
[ 224.692869] __arm64_compat_sys_aarch32_pread64+0x60/0x68
[ 224.698441] el0_svc_common+0x120/0x1e0
[ 224.702409] el0_svc_compat_handler+0x64/0x80
[ 224.706919] el0_svc_compat+0x8/0x18
[ 224.710615]
[ 224.712177] The buggy address belongs to the page:
[ 224.717128] page:ffffffbf05f18840 count:0 mapcount:0
mapping:0000000000000000 index:0x0
[ 224.725377] flags: 0x4000000000000000()
[ 224.729361] raw: 4000000000000000 0000000000000000 ffffffff00000301
[ 224.737340] raw: 0000000000000000 0000000000000000 00000000ffffffff
[ 224.745318] page dumped because: kasan: bad access detected
[ 224.751070]
[ 224.752629] Memory state around the buggy address:
[ 224.757583] ffffffc17c621200: ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff ff
[ 224.765035] ffffffc17c621280: ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff ff
[ 224.772483] >ffffffc17c621300: ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff ff
[ 224.779921] ^
[ 224.785403] ffffffc17c621380: ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff ff
[ 224.792853] ffffffc17c621400: ff ff ff ff ff ff ff ff ff ff ff ff
ff ff ff ff
[ 224.800299]

What was I doing when I saw this?
I was running blktests block/005 (switch schedulers while doing IO) on
a loop device backed by a block device. I don't hit this on every run
of block/005, but with this script I can usually hit it in less than a

d=`losetup --show -f /dev/sda5` # Replace sda5 with your favorite
unused partition.
cd /usr/blktests
echo "TEST_DEVS=$d" > config
while true; do
./check block/005

I got totally lost in the blk-mq stuff, but my best hunch is related
to this comment in bt_iter:
* We can hit rq == NULL here, because the tagging functions
* test and set the bit before assining ->rqs[].
if (rq && rq->q == hctx->queue)
iter_data->fn(hctx, rq, iter_data->data, reserved);

I believe the KASAN warning is about touching rq->q. So my question
is: blk_mq_put_tag doesn't appear to clear out tags->rqs[tag] when
releasing the tag, so what's to prevent bt_iter from looking through a
stale rq, in between when someone has allocated that sbitmap tag (for
the second time), but not yet assigned their own rq in ->rqs[tag].
Maybe there's some synchronization further up the stack that I'm

I was hoping to prove this by nulling out tags->rqs[tag] in
blk_mq_put_tag, but apparently hctx->tags is sometimes not valid
anymore by that point? Horrible things happen when I do that.

So I moved sideways and tried a partial fix by clearing rqs[tag] right
away in blk_mq_get_tag. This of course doesn't close my suspected
window, but does make it a lot smaller, and indeed I saw my repro rate
drop to zero (at least letting it go for a few minutes).

Do you agree this might be an issue, and if so, is there a good place
I can null out tags->rqs[tag] before calling blk_mq_put_tag?