Skip to contentSkip to conversation

nvmet-loop: avoid using mutex in IO hotpath

PATCH Superseded Nilay Shroff 4 participants 5 messages Nov 29, 2024 – Dec 9, 2024

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nvmet-loop: avoid using mutex in IO hotpath

Using mutex lock in IO hot path causes the kernel BUG sleeping while atomic. Shinichiro[1], first encountered this issue while running blktest nvme/052 shown below:

BUG: sleeping function called from invalid context at kernel/locking/mutex.c:585

in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 996, name: (udev-worker)
preempt_count: 0, expected: 0
RCU nest depth: 1, expected: 0
2 locks held by (udev-worker)/996:
#0: ffff8881004570c8 (mapping.invalidate_lock){.+.+}-{3:3}, at: page_cache_ra_unbounded+0x155/0x5c0 #1: ffffffff8607eaa0 (rcu_read_lock){....}-{1:2}, at: blk_mq_flush_plug_list+0xa75/0x1950

CPU: 2 UID: 0 PID: 996 Comm: (udev-worker) Not tainted 6.12.0-rc3+ #339
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-2.fc40 04/01/2014
Call Trace:
 <TASK>
 dump_stack_lvl+0x6a/0x90
 __might_resched.cold+0x1f7/0x23d
 ? __pfx___might_resched+0x10/0x10
 ? vsnprintf+0xdeb/0x18f0
 __mutex_lock+0xf4/0x1220
 ? nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
 ? __pfx_vsnprintf+0x10/0x10
 ? __pfx___mutex_lock+0x10/0x10
 ? snprintf+0xa5/0xe0
 ? xas_load+0x1ce/0x3f0
 ? nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
 nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
 ? __pfx_nvmet_subsys_nsid_exists+0x10/0x10 [nvmet]
 nvmet_req_find_ns+0x24e/0x300 [nvmet]
 nvmet_req_init+0x694/0xd40 [nvmet]
 ? blk_mq_start_request+0x11c/0x750
 ? nvme_setup_cmd+0x369/0x990 [nvme_core]
 nvme_loop_queue_rq+0x2a7/0x7a0 [nvme_loop]
 ? __pfx___lock_acquire+0x10/0x10
 ? __pfx_nvme_loop_queue_rq+0x10/0x10 [nvme_loop]
 __blk_mq_issue_directly+0xe2/0x1d0
 ? __pfx___blk_mq_issue_directly+0x10/0x10
 ? blk_mq_request_issue_directly+0xc2/0x140
 blk_mq_plug_issue_direct+0x13f/0x630
 ? lock_acquire+0x2d/0xc0
 ? blk_mq_flush_plug_list+0xa75/0x1950
 blk_mq_flush_plug_list+0xa9d/0x1950
 ? __pfx_blk_mq_flush_plug_list+0x10/0x10
 ? __pfx_mpage_readahead+0x10/0x10
 __blk_flush_plug+0x278/0x4d0
 ? __pfx___blk_flush_plug+0x10/0x10
 ? lock_release+0x460/0x7a0
 blk_finish_plug+0x4e/0x90
 read_pages+0x51b/0xbc0
 ? __pfx_read_pages+0x10/0x10
 ? lock_release+0x460/0x7a0
 page_cache_ra_unbounded+0x326/0x5c0
 force_page_cache_ra+0x1ea/0x2f0
 filemap_get_pages+0x59e/0x17b0
 ? __pfx_filemap_get_pages+0x10/0x10
 ? lock_is_held_type+0xd5/0x130
 ? __pfx___might_resched+0x10/0x10
 ? find_held_lock+0x2d/0x110
 filemap_read+0x317/0xb70
 ? up_write+0x1ba/0x510
 ? __pfx_filemap_read+0x10/0x10
 ? inode_security+0x54/0xf0
 ? selinux_file_permission+0x36d/0x420
 blkdev_read_iter+0x143/0x3b0
 vfs_read+0x6ac/0xa20
 ? __pfx_vfs_read+0x10/0x10
 ? __pfx_vm_mmap_pgoff+0x10/0x10
 ? __pfx___seccomp_filter+0x10/0x10
 ksys_read+0xf7/0x1d0
 ? __pfx_ksys_read+0x10/0x10
 do_syscall_64+0x93/0x180
 ? lockdep_hardirqs_on_prepare+0x16d/0x400
 ? do_syscall_64+0x9f/0x180
 ? lockdep_hardirqs_on+0x78/0x100
 ? do_syscall_64+0x9f/0x180
 ? lockdep_hardirqs_on_prepare+0x16d/0x400
 entry_SYSCALL_64_after_hwframe+0x76/0x7e
RIP: 0033:0x7f565bd1ce11
Code: 00 48 8b 15 09 90 0d 00 f7 d8 64 89 02 b8 ff ff ff ff eb bd e8 d0 ad 01 00 f3 0f 1e fa 80 3d 35 12 0e 00 00 74 13 31 c0 0f 05 <48> 3d 00 f0 ff ff 77 4f c3 66 0f 1f 44 00 00 55 48 89 e5 48 83 ec
RSP: 002b:00007ffd6e7a20c8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
RAX: ffffffffffffffda RBX: 0000000000001000 RCX: 00007f565bd1ce11
RDX: 0000000000001000 RSI: 00007f565babb000 RDI: 0000000000000014
RBP: 00007ffd6e7a2130 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000556000bfa610 R11: 0000000000000246 R12: 000000003ffff000
R13: 0000556000bfa5b0 R14: 0000000000000e00 R15: 0000556000c07328
 </TASK>

Apparently, the above issue is caused due to using mutex lock while we're in IO hot path. It's a regression caused with commit 505363957fad ("nvmet: fix nvme status code when namespace is disabled"). The mutex ->su_mutex is used to find whether a disabled nsid exists in the config group or not. This is to differentiate between a nsid that is disabled vs non-existent.

To mitigate the above issue, we've worked upon a fix[2] where we now insert nsid in subsys Xarray as soon as it's created under config group and later when that nsid is enabled, we add an Xarray mark on it and set ns->enabled to true. The Xarray mark is useful while we need to loop through all enabled namepsaces under a subsystem using xa_for_each_marked() API. If later a nsid is disabled then we clear Xarray mark from it and also set ns->enabled to false. It's only when nsid is deleted from the config group we delete it from the Xarray.

So with this change, now we could easily differentiate a nsid is disabled (i.e. Xarray entry for ns exists but ns->enabled is set to false) vs non- existent (i.e.Xarray entry for ns doesn't exist).

---
 drivers/nvme/target/admin-cmd.c |  13 ++--
 drivers/nvme/target/core.c      | 108 +++++++++++++++++++-------------
 drivers/nvme/target/nvmet.h     |   1 +
 drivers/nvme/target/pr.c        |  10 +--
 4 files changed, 79 insertions(+), 53 deletions(-)
drivers/nvme/target/admin-cmd.c +9 -4
index 934b401fbc2f..439fc96a82c8 100644
@@ -139,7 +139,8 @@ static u16 nvmet_get_smart_log_all(struct nvmet_req *req,
 	unsigned long idx;
 
 	ctrl = req->sq->ctrl;
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {
 		/* we don't have the right data for file backed ns */
 		if (!ns->bdev)
 			continue;
@@ -331,9 +332,11 @@ static u32 nvmet_format_ana_group(struct nvmet_req *req, u32 grpid,
 	u32 count = 0;
 
 	if (!(req->cmd->get_log_page.lsp & NVME_ANA_LOG_RGO)) {
-		xa_for_each(&ctrl->subsys->namespaces, idx, ns)
+		xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+				ns, NVMET_NS_ENABLED) {
 			if (ns->anagrpid == grpid)
 				desc->nsids[count++] = cpu_to_le32(ns->nsid);
+		}
 	}
 
 	desc->grpid = cpu_to_le32(grpid);
@@ -771,7 +774,8 @@ static void nvmet_execute_identify_endgrp_list(struct nvmet_req *req)
 		goto out;
 	}
 
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {
 		if (ns->nsid <= min_endgid)
 			continue;
 
@@ -814,7 +818,8 @@ static void nvmet_execute_identify_nslist(struct nvmet_req *req, bool match_css)
 		goto out;
 	}
 
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {
 		if (ns->nsid <= min_nsid)
 			continue;
 		if (match_css && req->ns->csi != req->cmd->identify.csi)
drivers/nvme/target/core.c +63 -45

191 lines not on this page · raw message

drivers/nvme/target/nvmet.h +1 -0
index 58328b35dc96..0fc0abf1150c 100644
@@ -24,6 +24,7 @@
 
 #define NVMET_DEFAULT_VS		NVME_VS(2, 1, 0)
 
+#define NVMET_NS_ENABLED		XA_MARK_1
 #define NVMET_ASYNC_EVENTS		4
 #define NVMET_ERROR_LOG_SLOTS		128
 #define NVMET_NO_ERROR_LOC		((u16)-1)
drivers/nvme/target/pr.c +6 -4
index 25a02b50d9f3..50acdb603572 100644
@@ -60,7 +60,8 @@ u16 nvmet_set_feat_resv_notif_mask(struct nvmet_req *req, u32 mask)
 		goto success;
 	}
 
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {
 		if (ns->pr.enable)
 			WRITE_ONCE(ns->pr.notify_mask, mask);
 	}
@@ -1057,7 +1058,7 @@ int nvmet_ctrl_init_pr(struct nvmet_ctrl *ctrl)
 	 * nvmet_pr_init_ns(), see more details in nvmet_ns_enable().
 	 * So just check ns->pr.enable.
 	 */
-	xa_for_each(&subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&subsys->namespaces, idx, ns, NVMET_NS_ENABLED) {
 		if (ns->pr.enable) {
 			ret = nvmet_pr_alloc_and_insert_pc_ref(ns, ctrl->cntlid,
 							&ctrl->hostid);
@@ -1068,7 +1069,7 @@ int nvmet_ctrl_init_pr(struct nvmet_ctrl *ctrl)
 	return 0;
 
 free_per_ctrl_refs:
-	xa_for_each(&subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&subsys->namespaces, idx, ns, NVMET_NS_ENABLED) {
 		if (ns->pr.enable) {
 			pc_ref = xa_erase(&ns->pr_per_ctrl_refs, ctrl->cntlid);
 			if (pc_ref)
@@ -1088,7 +1089,8 @@ void nvmet_ctrl_destroy_pr(struct nvmet_ctrl *ctrl)
 	kfifo_free(&ctrl->pr_log_mgr.log_queue);
 	mutex_destroy(&ctrl->pr_log_mgr.lock);
 
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {
 		if (ns->pr.enable) {
 			pc_ref = xa_erase(&ns->pr_per_ctrl_refs, ctrl->cntlid);
 			if (pc_ref)
Signature · 1 line
2.45.2
Reviewed-by on v1 1/1
#1 lore raw
Quoted 113 lines · Nilay Shroff

Using mutex lock in IO hot path causes the kernel BUG sleeping while atomic. Shinichiro[1], first encountered this issue while running blktest nvme/052 shown below:

BUG: sleeping function called from invalid context at kernel/locking/mutex.c:585

in_atomic(): 0, irqs_disabled(): 0, non_block: 0, pid: 996, name: (udev-worker)
preempt_count: 0, expected: 0
RCU nest depth: 1, expected: 0
2 locks held by (udev-worker)/996:
#0: ffff8881004570c8 (mapping.invalidate_lock){.+.+}-{3:3}, at: page_cache_ra_unbounded+0x155/0x5c0 #1: ffffffff8607eaa0 (rcu_read_lock){....}-{1:2}, at: blk_mq_flush_plug_list+0xa75/0x1950

CPU: 2 UID: 0 PID: 996 Comm: (udev-worker) Not tainted 6.12.0-rc3+ #339
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-2.fc40 04/01/2014
Call Trace:
  <TASK>
  dump_stack_lvl+0x6a/0x90
  __might_resched.cold+0x1f7/0x23d
  ? __pfx___might_resched+0x10/0x10
  ? vsnprintf+0xdeb/0x18f0
  __mutex_lock+0xf4/0x1220
  ? nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
  ? __pfx_vsnprintf+0x10/0x10
  ? __pfx___mutex_lock+0x10/0x10
  ? snprintf+0xa5/0xe0
  ? xas_load+0x1ce/0x3f0
  ? nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
  nvmet_subsys_nsid_exists+0xb9/0x150 [nvmet]
  ? __pfx_nvmet_subsys_nsid_exists+0x10/0x10 [nvmet]
  nvmet_req_find_ns+0x24e/0x300 [nvmet]
  nvmet_req_init+0x694/0xd40 [nvmet]
  ? blk_mq_start_request+0x11c/0x750
  ? nvme_setup_cmd+0x369/0x990 [nvme_core]
  nvme_loop_queue_rq+0x2a7/0x7a0 [nvme_loop]
  ? __pfx___lock_acquire+0x10/0x10
  ? __pfx_nvme_loop_queue_rq+0x10/0x10 [nvme_loop]
  __blk_mq_issue_directly+0xe2/0x1d0
  ? __pfx___blk_mq_issue_directly+0x10/0x10
  ? blk_mq_request_issue_directly+0xc2/0x140
  blk_mq_plug_issue_direct+0x13f/0x630
  ? lock_acquire+0x2d/0xc0
  ? blk_mq_flush_plug_list+0xa75/0x1950
  blk_mq_flush_plug_list+0xa9d/0x1950
  ? __pfx_blk_mq_flush_plug_list+0x10/0x10
  ? __pfx_mpage_readahead+0x10/0x10
  __blk_flush_plug+0x278/0x4d0
  ? __pfx___blk_flush_plug+0x10/0x10
  ? lock_release+0x460/0x7a0
  blk_finish_plug+0x4e/0x90
  read_pages+0x51b/0xbc0
  ? __pfx_read_pages+0x10/0x10
  ? lock_release+0x460/0x7a0
  page_cache_ra_unbounded+0x326/0x5c0
  force_page_cache_ra+0x1ea/0x2f0
  filemap_get_pages+0x59e/0x17b0
  ? __pfx_filemap_get_pages+0x10/0x10
  ? lock_is_held_type+0xd5/0x130
  ? __pfx___might_resched+0x10/0x10
  ? find_held_lock+0x2d/0x110
  filemap_read+0x317/0xb70
  ? up_write+0x1ba/0x510
  ? __pfx_filemap_read+0x10/0x10
  ? inode_security+0x54/0xf0
  ? selinux_file_permission+0x36d/0x420
  blkdev_read_iter+0x143/0x3b0
  vfs_read+0x6ac/0xa20
  ? __pfx_vfs_read+0x10/0x10
  ? __pfx_vm_mmap_pgoff+0x10/0x10
  ? __pfx___seccomp_filter+0x10/0x10
  ksys_read+0xf7/0x1d0
  ? __pfx_ksys_read+0x10/0x10
  do_syscall_64+0x93/0x180
  ? lockdep_hardirqs_on_prepare+0x16d/0x400
  ? do_syscall_64+0x9f/0x180
  ? lockdep_hardirqs_on+0x78/0x100
  ? do_syscall_64+0x9f/0x180
  ? lockdep_hardirqs_on_prepare+0x16d/0x400
  entry_SYSCALL_64_after_hwframe+0x76/0x7e
RIP: 0033:0x7f565bd1ce11
Code: 00 48 8b 15 09 90 0d 00 f7 d8 64 89 02 b8 ff ff ff ff eb bd e8 d0 ad 01 00 f3 0f 1e fa 80 3d 35 12 0e 00 00 74 13 31 c0 0f 05 <48> 3d 00 f0 ff ff 77 4f c3 66 0f 1f 44 00 00 55 48 89 e5 48 83 ec
RSP: 002b:00007ffd6e7a20c8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000
RAX: ffffffffffffffda RBX: 0000000000001000 RCX: 00007f565bd1ce11
RDX: 0000000000001000 RSI: 00007f565babb000 RDI: 0000000000000014
RBP: 00007ffd6e7a2130 R08: 00000000ffffffff R09: 0000000000000000
R10: 0000556000bfa610 R11: 0000000000000246 R12: 000000003ffff000
R13: 0000556000bfa5b0 R14: 0000000000000e00 R15: 0000556000c07328
  </TASK>

Apparently, the above issue is caused due to using mutex lock while we're in IO hot path. It's a regression caused with commit 505363957fad ("nvmet: fix nvme status code when namespace is disabled"). The mutex ->su_mutex is used to find whether a disabled nsid exists in the config group or not. This is to differentiate between a nsid that is disabled vs non-existent.

To mitigate the above issue, we've worked upon a fix[2] where we now insert nsid in subsys Xarray as soon as it's created under config group and later when that nsid is enabled, we add an Xarray mark on it and set ns->enabled to true. The Xarray mark is useful while we need to loop through all enabled namepsaces under a subsystem using xa_for_each_marked() API. If later a nsid is disabled then we clear Xarray mark from it and also set ns->enabled to false. It's only when nsid is deleted from the config group we delete it from the Xarray.

So with this change, now we could easily differentiate a nsid is disabled (i.e. Xarray entry for ns exists but ns->enabled is set to false) vs non- existent (i.e.Xarray entry for ns doesn't exist).

Quoted 7 lines
---
  drivers/nvme/target/admin-cmd.c |  13 ++--
  drivers/nvme/target/core.c      | 108 +++++++++++++++++++-------------
  drivers/nvme/target/nvmet.h     |   1 +
  drivers/nvme/target/pr.c        |  10 +--
  4 files changed, 79 insertions(+), 53 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
#3 lore raw
In reply to Nilay Shroff

drivers/nvme/target/admin-cmd.c static u16 nvmet_get_smart_log_all(struct nvmet_req *req, · line 143

Quoted 3 lines · Nilay Shroff
drivers/nvme/target/admin-cmd.c +2 -1
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {

Given that this is duplicated quite a bit, maybe add a nvmet_for_each_enabled_ns wrapper for it? And also a nvmet_for_each_ns one that iterates all namespaces?

Otherwise this looks good, and I really like the use of the xarray mark to simplify things!

#4 lore raw
Quoted 8 lines · Christoph Hellwig

drivers/nvme/target/admin-cmd.c static u16 nvmet_get_smart_log_all(struct nvmet_req *req, · line 143

Quoted 3 lines · Nilay Shroff
drivers/nvme/target/admin-cmd.c +2 -1
-	xa_for_each(&ctrl->subsys->namespaces, idx, ns) {
+	xa_for_each_marked(&ctrl->subsys->namespaces, idx,
+			ns, NVMET_NS_ENABLED) {

Given that this is duplicated quite a bit, maybe add a nvmet_for_each_enabled_ns wrapper for it? And also a nvmet_for_each_ns one that iterates all namespaces?

Yeah sure, I will add those wrappers and spin a new patch.

Quoted 3 lines

Otherwise this looks good, and I really like the use of the xarray mark to simplify things!

Thanks,
--Nilay