| Message ID | 20250319105518.468941-1-nilay@linux.ibm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | fix locking issues with blk-wbt parameters update | expand |
On Wed, 19 Mar 2025 16:23:44 +0530, Nilay Shroff wrote: > This patchset contains two patches. > > The first patch fixes a missed release of q->elevator_lock which was > mistakenly omitted in one of the return code path of ioc_qos_write. > > The second patch fixes the locdep splat reported due to the incorrect > locking order between q->elevator_lock and q->rq_qos_mutex. The commit > 245618f8e45f ("block: protect wbt_lat_usec using q->elevator_lock") > introduced q->elevator_lock to protect updates to blk-wbt parameters > when writing to the sysfs attribute wbt_lat_usec and the cgroup attribute > io.cost.qos. However, writes to these attributes also acquire q->rq_qos_ > mutex, creating a potential circular dependency if the locking order is > not correctly followed. This patch ensures the correct locking sequence > to prevent such issues. Unfortunately, blktests currently lacks a test > case for writes to these attributes, which might have caught this issue > earlier. I plan to submit a blktest to cover these cases. > > [...] Applied, thanks! [1/2] block: release q->elevator_lock in ioc_qos_write commit: 89ed5fa3b5419f04452051fbcb6d3e5b801cdb1b [2/2] block: correct locking order for protecting blk-wbt parameters commit: 9730763f4756e32520cb86778331465e8d063a8f Best regards,
Hi, This patchset contains two patches. The first patch fixes a missed release of q->elevator_lock which was mistakenly omitted in one of the return code path of ioc_qos_write. The second patch fixes the locdep splat reported due to the incorrect locking order between q->elevator_lock and q->rq_qos_mutex. The commit 245618f8e45f ("block: protect wbt_lat_usec using q->elevator_lock") introduced q->elevator_lock to protect updates to blk-wbt parameters when writing to the sysfs attribute wbt_lat_usec and the cgroup attribute io.cost.qos. However, writes to these attributes also acquire q->rq_qos_ mutex, creating a potential circular dependency if the locking order is not correctly followed. This patch ensures the correct locking sequence to prevent such issues. Unfortunately, blktests currently lacks a test case for writes to these attributes, which might have caught this issue earlier. I plan to submit a blktest to cover these cases. Nilay Shroff (2): block: release q->elevator_lock in ioc_qos_write block: correct locking order for protecting blk-wbt parameters block/blk-cgroup.c | 51 ++++++++++++++++++++++++++++++++++++++++++++++ block/blk-cgroup.h | 2 ++ block/blk-iocost.c | 17 +++++----------- 3 files changed, 58 insertions(+), 12 deletions(-)