]> git.baikalelectronics.ru Git - kernel.git/commit
blk-wbt: call rq_qos_add() after wb_normal is initialized
authorYu Kuai <yukuai3@huawei.com>
Tue, 13 Sep 2022 10:57:49 +0000 (18:57 +0800)
committerJens Axboe <axboe@kernel.dk>
Wed, 21 Sep 2022 14:36:13 +0000 (08:36 -0600)
commit8c5035dfbb9475b67c82b3fdb7351236525bf52b
treeabcac6d89b9e248e7714ef76b1878b06ea33e346
parentf7de4886fe8f008ac4f7beeacd9bedb6b91241f5
blk-wbt: call rq_qos_add() after wb_normal is initialized

Our test found a problem that wbt inflight counter is negative, which
will cause io hang(noted that this problem doesn't exist in mainline):

t1: device create t2: issue io
add_disk
 blk_register_queue
  wbt_enable_default
   wbt_init
    rq_qos_add
    // wb_normal is still 0
/*
 * in mainline, disk can't be opened before
 * bdev_add(), however, in old kernels, disk
 * can be opened before blk_register_queue().
 */
blkdev_issue_flush
                        // disk size is 0, however, it's not checked
                         submit_bio_wait
                          submit_bio
                           blk_mq_submit_bio
                            rq_qos_throttle
                             wbt_wait
      bio_to_wbt_flags
                               rwb_enabled
       // wb_normal is 0, inflight is not increased

    wbt_queue_depth_changed(&rwb->rqos);
     wbt_update_limits
     // wb_normal is initialized
                            rq_qos_track
                             wbt_track
                              rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
      // wb_normal is not 0,wbt_flags will be set
t3: io completion
blk_mq_free_request
 rq_qos_done
  wbt_done
   wbt_is_tracked
   // return true
   __wbt_done
    wbt_rqw_done
     atomic_dec_return(&rqw->inflight);
     // inflight is decreased

commit 8235b5c1e8c1 ("block: call bdev_add later in device_add_disk") can
avoid this problem, however it's better to fix this problem in wbt:

1) Lower kernel can't backport this patch due to lots of refactor.
2) Root cause is that wbt call rq_qos_add() before wb_normal is
initialized.

Fixes: e34cbd307477 ("blk-wbt: add general throttling mechanism")
Cc: <stable@vger.kernel.org>
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Link: https://lore.kernel.org/r/20220913105749.3086243-1-yukuai1@huaweicloud.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-wbt.c