Hello,
request queue quiescing has been applied in lots of block drivers and
block core from different/un-related code paths. So far, both quiesce
and unquiesce changes the queue state unconditionally. This way has
caused trouble, such as, driver is quiescing queue for its own purpose,
however block core's queue quiesce may come because of elevator switch,
updating nr_requests or other queue attributes, then un-expected
unquiesce may come too early.
It has been observed kernel panic when running stress test on dm-mpath
suspend and updating nr_requests.
Fix the issue by supporting concurrent queue quiescing. But nvme has very
complicated uses on quiesce/unquiesce, which two may not be called in
pairing, so switch to this way in patch 1~4, and patch 5 provides
nested queue quiesce.
V4:
- one small patch style change as suggested by Christoph, only patch 6/6
is touched
V3:
- add patch 5/6 to clear NVME_CTRL_ADMIN_Q_STOPPED for nvme-loop
after reallocating admin queue
- take Bart's suggestion to add warning in blk_mq_unquiesce_queue()
& update commit log
V2:
- replace mutex with atomic ops for supporting paring quiesce &
unquiesce
Ming Lei (6):
nvme: add APIs for stopping/starting admin queue
nvme: apply nvme API to quiesce/unquiesce admin queue
nvme: prepare for pairing quiescing and unquiescing
nvme: paring quiesce/unquiesce
nvme: loop: clear NVME_CTRL_ADMIN_Q_STOPPED after admin queue is
reallocated
blk-mq: support concurrent queue quiesce/unquiesce
block/blk-mq.c | 22 ++++++++++--
drivers/nvme/host/core.c | 70 ++++++++++++++++++++++++++------------
drivers/nvme/host/fc.c | 8 ++---
drivers/nvme/host/nvme.h | 4 +++
drivers/nvme/host/pci.c | 8 ++---
drivers/nvme/host/rdma.c | 14 ++++----
drivers/nvme/host/tcp.c | 16 ++++-----
drivers/nvme/target/loop.c | 6 ++--
include/linux/blkdev.h | 2 ++
9 files changed, 100 insertions(+), 50 deletions(-)
--
2.31.1
