Multipath was previously made to use blk_abort_queue() to allow for lower latency path deactivation (commit 224cb3e9). The call to blk_abort_queue has proven to be unsafe, and is now disabled by default, due to a race (between blk_abort_queue and scsi_request_fn) that can lead to list corruption, from: https://www.redhat.com/archives/dm-devel/2010-November/msg00085.html "the cmd gets blk_abort_queued/timedout run on it and the scsi eh somehow is able to complete and run scsi_queue_insert while scsi_request_fn is still trying to process the request." It is expected that this race will be fixed in the near-term so it makes little since to remove all associated code. Providing control over the call to blk_abort_queue() facilitates continued testing and future flexibility to opt-in to lower latency path deactivation. Opting to enable this feature will emit a warning for the time being. Add 'features' member to 'struct multipath' and introduce MPF_ABORT_QUEUE as the first feature flag. If "abort_queue_on_fail" is provided, via message or during mpath device configuration, the MPF_ABORT_QUEUE feature flag will be set and blk_abort_queue() will be called during path deactivation. The MPF_ABORT_QUEUE feature flag may be cleared using the "skip_abort_queue_on_fail" message. Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx> Cc: Mike Anderson <andmike@xxxxxxxxxxxxxxxxxx> Cc: Mike Christie <michaelc@xxxxxxxxxxx> --- drivers/md/dm-mpath.c | 46 +++++++++++++++++++++++++++++++++++++++++++--- 1 files changed, 43 insertions(+), 3 deletions(-) v5: - introduced abort_queue_on_fail() to perform m->lock locking around both set_bit and clear_bit - also added ability to set MPF_ABORT_QUEUE feature flag via "abort_queue_on_fail" message. v4: - revised the patch subject and header to emphaszie default to off - fixed missing m->lock locking when using test_bit() on m->features (MPF_ABORT_QUEUE is checked early in fail_path() to avoid queuing unnecessary work). - also added ability to clear MPF_ABORT_QUEUE feature flag via "skip_abort_queue_on_fail" message. v3: bumped target version and switched feature name from "abort_queue_on_failure" to "abort_queue_on_fail". diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c index 487ecda..a6f11c3 100644 --- a/drivers/md/dm-mpath.c +++ b/drivers/md/dm-mpath.c @@ -56,8 +56,14 @@ struct priority_group { struct list_head pgpaths; }; +/* + * Bits for the m->features + */ +#define MPF_ABORT_QUEUE 0 + /* Multipath context */ struct multipath { + unsigned long features; struct list_head list; struct dm_target *ti; @@ -414,6 +420,24 @@ static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path, return 0; } +static int abort_queue_on_fail(struct multipath *m, unsigned abort_queue_on_fail) +{ + unsigned long flags; + + spin_lock_irqsave(&m->lock, flags); + + if (abort_queue_on_fail) { + set_bit(MPF_ABORT_QUEUE, &m->features); + DMWARN("Enabling use of blk_abort_queue is unsafe."); + } else + clear_bit(MPF_ABORT_QUEUE, &m->features); + + spin_unlock_irqrestore(&m->lock, flags); + + return 0; +} + + /*----------------------------------------------------------------- * The multipath daemon is responsible for resubmitting queued ios. *---------------------------------------------------------------*/ @@ -813,6 +837,11 @@ static int parse_features(struct arg_set *as, struct multipath *m) continue; } + if (!strnicmp(param_name, MESG_STR("abort_queue_on_fail"))) { + r = abort_queue_on_fail(m, 1); + continue; + } + if (!strnicmp(param_name, MESG_STR("pg_init_retries")) && (argc >= 1)) { r = read_param(_params + 1, shift(as), @@ -995,7 +1024,9 @@ static int fail_path(struct pgpath *pgpath) pgpath->path.dev->name, m->nr_valid_paths); schedule_work(&m->trigger_event); - queue_work(kmultipathd, &pgpath->deactivate_path); + + if (test_bit(MPF_ABORT_QUEUE, &pgpath->pg->m->features)) + queue_work(kmultipathd, &pgpath->deactivate_path); out: spin_unlock_irqrestore(&m->lock, flags); @@ -1382,11 +1413,14 @@ static int multipath_status(struct dm_target *ti, status_type_t type, DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count); else { DMEMIT("%u ", m->queue_if_no_path + - (m->pg_init_retries > 0) * 2); + (m->pg_init_retries > 0) * 2 + + test_bit(MPF_ABORT_QUEUE, &m->features)); if (m->queue_if_no_path) DMEMIT("queue_if_no_path "); if (m->pg_init_retries) DMEMIT("pg_init_retries %u ", m->pg_init_retries); + if (test_bit(MPF_ABORT_QUEUE, &m->features)) + DMEMIT("abort_queue_on_fail "); } if (!m->hw_handler_name || type == STATUSTYPE_INFO) @@ -1490,6 +1524,12 @@ static int multipath_message(struct dm_target *ti, unsigned argc, char **argv) } else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) { r = queue_if_no_path(m, 0, 0); goto out; + } else if (!strnicmp(argv[0], MESG_STR("abort_queue_on_fail"))) { + r = abort_queue_on_fail(m, 1); + goto out; + } else if (!strnicmp(argv[0], MESG_STR("skip_abort_queue_on_fail"))) { + r = abort_queue_on_fail(m, 0); + goto out; } } @@ -1655,7 +1695,7 @@ out: *---------------------------------------------------------------*/ static struct target_type multipath_target = { .name = "multipath", - .version = {1, 1, 1}, + .version = {1, 2, 0}, .module = THIS_MODULE, .ctr = multipath_ctr, .dtr = multipath_dtr, -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
