nvme: fix a possible deadlock when passthru commands sent to a multipath device
authorSagi Grimberg <sagi@grimberg.me>
Wed, 31 Jul 2019 18:00:26 +0000 (11:00 -0700)
committerSagi Grimberg <sagi@grimberg.me>
Thu, 1 Aug 2019 00:59:01 +0000 (17:59 -0700)
When the user issues a command with side effects, we will end up freezing
the namespace request queue when updating disk info (and the same for
the corresponding mpath disk node).

However, we are not freezing the mpath node request queue,
which means that mpath I/O can still come in and block on blk_queue_enter
(called from nvme_ns_head_make_request -> direct_make_request).

This is a deadlock, because blk_queue_enter will block until the inner
namespace request queue is unfroze, but that process is blocked because
the namespace revalidation is trying to update the mpath disk info
and freeze its request queue (which will never complete because
of the I/O that is blocked on blk_queue_enter).

Fix this by freezing all the subsystem nsheads request queues before
executing the passthru command. Given that these commands are infrequent
we should not worry about this temporary I/O freeze to keep things sane.

Here is the matching hang traces:
--
[ 374.465002] INFO: task systemd-udevd:17994 blocked for more than 122 seconds.
[ 374.472975] Not tainted 5.2.0-rc3-mpdebug+ #42
[ 374.478522] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 374.487274] systemd-udevd D 0 17994 1 0x00000000
[ 374.493407] Call Trace:
[ 374.496145] __schedule+0x2ef/0x620
[ 374.500047] schedule+0x38/0xa0
[ 374.503569] blk_queue_enter+0x139/0x220
[ 374.507959] ? remove_wait_queue+0x60/0x60
[ 374.512540] direct_make_request+0x60/0x130
[ 374.517219] nvme_ns_head_make_request+0x11d/0x420 [nvme_core]
[ 374.523740] ? generic_make_request_checks+0x307/0x6f0
[ 374.529484] generic_make_request+0x10d/0x2e0
[ 374.534356] submit_bio+0x75/0x140
[ 374.538163] ? guard_bio_eod+0x32/0xe0
[ 374.542361] submit_bh_wbc+0x171/0x1b0
[ 374.546553] block_read_full_page+0x1ed/0x330
[ 374.551426] ? check_disk_change+0x70/0x70
[ 374.556008] ? scan_shadow_nodes+0x30/0x30
[ 374.560588] blkdev_readpage+0x18/0x20
[ 374.564783] do_read_cache_page+0x301/0x860
[ 374.569463] ? blkdev_writepages+0x10/0x10
[ 374.574037] ? prep_new_page+0x88/0x130
[ 374.578329] ? get_page_from_freelist+0xa2f/0x1280
[ 374.583688] ? __alloc_pages_nodemask+0x179/0x320
[ 374.588947] read_cache_page+0x12/0x20
[ 374.593142] read_dev_sector+0x2d/0xd0
[ 374.597337] read_lba+0x104/0x1f0
[ 374.601046] find_valid_gpt+0xfa/0x720
[ 374.605243] ? string_nocheck+0x58/0x70
[ 374.609534] ? find_valid_gpt+0x720/0x720
[ 374.614016] efi_partition+0x89/0x430
[ 374.618113] ? string+0x48/0x60
[ 374.621632] ? snprintf+0x49/0x70
[ 374.625339] ? find_valid_gpt+0x720/0x720
[ 374.629828] check_partition+0x116/0x210
[ 374.634214] rescan_partitions+0xb6/0x360
[ 374.638699] __blkdev_reread_part+0x64/0x70
[ 374.643377] blkdev_reread_part+0x23/0x40
[ 374.647860] blkdev_ioctl+0x48c/0x990
[ 374.651956] block_ioctl+0x41/0x50
[ 374.655766] do_vfs_ioctl+0xa7/0x600
[ 374.659766] ? locks_lock_inode_wait+0xb1/0x150
[ 374.664832] ksys_ioctl+0x67/0x90
[ 374.668539] __x64_sys_ioctl+0x1a/0x20
[ 374.672732] do_syscall_64+0x5a/0x1c0
[ 374.676828] entry_SYSCALL_64_after_hwframe+0x44/0xa9

[ 374.738474] INFO: task nvmeadm:49141 blocked for more than 123 seconds.
[ 374.745871] Not tainted 5.2.0-rc3-mpdebug+ #42
[ 374.751419] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 374.760170] nvmeadm D 0 49141 36333 0x00004080
[ 374.766301] Call Trace:
[ 374.769038] __schedule+0x2ef/0x620
[ 374.772939] schedule+0x38/0xa0
[ 374.776452] blk_mq_freeze_queue_wait+0x59/0x100
[ 374.781614] ? remove_wait_queue+0x60/0x60
[ 374.786192] blk_mq_freeze_queue+0x1a/0x20
[ 374.790773] nvme_update_disk_info.isra.57+0x5f/0x350 [nvme_core]
[ 374.797582] ? nvme_identify_ns.isra.50+0x71/0xc0 [nvme_core]
[ 374.804006] __nvme_revalidate_disk+0xe5/0x110 [nvme_core]
[ 374.810139] nvme_revalidate_disk+0xa6/0x120 [nvme_core]
[ 374.816078] ? nvme_submit_user_cmd+0x11e/0x320 [nvme_core]
[ 374.822299] nvme_user_cmd+0x264/0x370 [nvme_core]
[ 374.827661] nvme_dev_ioctl+0x112/0x1d0 [nvme_core]
[ 374.833114] do_vfs_ioctl+0xa7/0x600
[ 374.837117] ? __audit_syscall_entry+0xdd/0x130
[ 374.842184] ksys_ioctl+0x67/0x90
[ 374.845891] __x64_sys_ioctl+0x1a/0x20
[ 374.850082] do_syscall_64+0x5a/0x1c0
[ 374.854178] entry_SYSCALL_64_after_hwframe+0x44/0xa9
--

Reported-by: James Puthukattukaran <james.puthukattukaran@oracle.com>
Tested-by: James Puthukattukaran <james.puthukattukaran@oracle.com>
Reviewed-by: Keith Busch <kbusch@kernel.org>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
drivers/nvme/host/core.c
drivers/nvme/host/multipath.c
drivers/nvme/host/nvme.h

index e35f16b60fc9e7d093836d7f6ef0ddbd3bdd87f3..88b8dfd7928aa60895e283ab2b28b456c32ca5ec 100644 (file)
@@ -1286,6 +1286,9 @@ static u32 nvme_passthru_start(struct nvme_ctrl *ctrl, struct nvme_ns *ns,
         */
        if (effects & (NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_CSE_MASK)) {
                mutex_lock(&ctrl->scan_lock);
+               mutex_lock(&ctrl->subsys->lock);
+               nvme_mpath_start_freeze(ctrl->subsys);
+               nvme_mpath_wait_freeze(ctrl->subsys);
                nvme_start_freeze(ctrl);
                nvme_wait_freeze(ctrl);
        }
@@ -1316,6 +1319,8 @@ static void nvme_passthru_end(struct nvme_ctrl *ctrl, u32 effects)
                nvme_update_formats(ctrl);
        if (effects & (NVME_CMD_EFFECTS_LBCC | NVME_CMD_EFFECTS_CSE_MASK)) {
                nvme_unfreeze(ctrl);
+               nvme_mpath_unfreeze(ctrl->subsys);
+               mutex_unlock(&ctrl->subsys->lock);
                mutex_unlock(&ctrl->scan_lock);
        }
        if (effects & NVME_CMD_EFFECTS_CCC)
index 4f0d0d12744e0e94dbc1e518ca0c1935648c1bac..b34dcb2288e700e6f45795ed80c65a0be497b8cf 100644 (file)
@@ -12,6 +12,36 @@ module_param(multipath, bool, 0444);
 MODULE_PARM_DESC(multipath,
        "turn on native support for multiple controllers per subsystem");
 
+void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
+{
+       struct nvme_ns_head *h;
+
+       lockdep_assert_held(&subsys->lock);
+       list_for_each_entry(h, &subsys->nsheads, entry)
+               if (h->disk)
+                       blk_mq_unfreeze_queue(h->disk->queue);
+}
+
+void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
+{
+       struct nvme_ns_head *h;
+
+       lockdep_assert_held(&subsys->lock);
+       list_for_each_entry(h, &subsys->nsheads, entry)
+               if (h->disk)
+                       blk_mq_freeze_queue_wait(h->disk->queue);
+}
+
+void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
+{
+       struct nvme_ns_head *h;
+
+       lockdep_assert_held(&subsys->lock);
+       list_for_each_entry(h, &subsys->nsheads, entry)
+               if (h->disk)
+                       blk_freeze_queue_start(h->disk->queue);
+}
+
 /*
  * If multipathing is enabled we need to always use the subsystem instance
  * number for numbering our devices to avoid conflicts between subsystems that
index 26b563f9985b5a9bf4163a18646e9fee68a2ae17..6b4fb67124c690021c12d0e5851ad230a20c88d5 100644 (file)
@@ -490,6 +490,9 @@ static inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
        return ctrl->ana_log_buf != NULL;
 }
 
+void nvme_mpath_unfreeze(struct nvme_subsystem *subsys);
+void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys);
+void nvme_mpath_start_freeze(struct nvme_subsystem *subsys);
 void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
                        struct nvme_ctrl *ctrl, int *flags);
 void nvme_failover_req(struct request *req);
@@ -568,6 +571,15 @@ static inline void nvme_mpath_uninit(struct nvme_ctrl *ctrl)
 static inline void nvme_mpath_stop(struct nvme_ctrl *ctrl)
 {
 }
+static inline void nvme_mpath_unfreeze(struct nvme_subsystem *subsys)
+{
+}
+static inline void nvme_mpath_wait_freeze(struct nvme_subsystem *subsys)
+{
+}
+static inline void nvme_mpath_start_freeze(struct nvme_subsystem *subsys)
+{
+}
 #endif /* CONFIG_NVME_MULTIPATH */
 
 #ifdef CONFIG_NVM