Skip to contentSkip to conversation

nvmet: don't abort the preempting host itself

PATCH In review Guixin Liu 3 participants 4 messages Nov 20

Subsystems (2)
lore
Share & tools
lore

Guixin Liu

author
#0 lore raw

PATCH 1/1 nvmet: don't abort the preempting host itself

When a host issues a Reservation Acquire with the Preempt and Abort action and its own reservation key matches the Preempt Reservation Key (PRKEY), the host unregisters itself as the specification requires:
all paths in NVMe Base Specification section 8.1.25.7 that unregister registrants by PRKEY (no reservation held, All Registrants types with a non-zero PRKEY, and a PRKEY that does not match the holder's key) apply to every registrant whose key matches, without excluding the host that issued the command.

But nvmet also adds the issuing host's controllers to the abort set. The abort flow kills the per-controller ref of every controller in the set and waits for their counts to drop to zero before completing the preempt command. The preempt command itself holds a reference on its own controller's per-controller ref until it completes, so the abort flow ends up waiting for the very command it is supposed to complete afterwards. The namespace's pr_sem is never released and all further reservation commands on the namespace hang.

The abort of commands as a side effect of preempting a reservation is best effort, so skip the issuing host's controllers and just unregister the registrant.

Fixes 5a47c2080a73 ("nvmet: support reservation feature")
---
 drivers/nvme/target/pr.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)
drivers/nvme/target/pr.c +8 -1
index 8c818f24815b..ed530d7362f5 100644
@@ -502,7 +502,14 @@ static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, u64 prkey,
 		if (reg->rkey == prkey) {
 			status = NVME_SC_SUCCESS;
 			uuid_copy(&hostid, &reg->hostid);
-			if (abort)
+			/*
+			 * Aborting is best effort (NVMe Base Specification
+			 * section 8.1.25.7): skip the issuing host, as the
+			 * abort flow would otherwise wait for the preempt
+			 * command itself, which holds a reference on its
+			 * controller's per-controller ref until it completes.
+			 */
+			if (abort && !uuid_equal(&hostid, send_hostid))
 				nvmet_pr_set_ctrl_to_abort(req, &hostid);
 			nvmet_pr_unregister_one(pr, reg);
 			if (!uuid_equal(&hostid, send_hostid))
Signature · 1 line
2.43.7
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue to consider: - [Medium] nvmet: nvmet_pr_unreg_all_host_by_prkey() skips abort on false pc_ref premise --

Thank you for your contribution! Sashiko AI review found 1 potential issue to consider:
- [Medium] nvmet: nvmet_pr_unreg_all_host_by_prkey() skips abort on false pc_ref premise
--

commit 07f50eedfd5f5e5d148cd9fefe65d629fc37077e
Author: Guixin Liu <[email protected]>

nvmet: don't abort the preempting host itself

This commit skips calling nvmet_pr_set_ctrl_to_abort() for the issuing host in nvmet_pr_unreg_all_host_by_prkey() when handling a Reservation Acquire command with the Preempt and Abort action.

Quoted 10 lines
drivers/nvme/target/pr.c +1 -1
index 8c818f24815bd..ed530d7362f5b 100644
@@ -502,7 +502,14 @@ static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, u64 prkey,
 		if (reg->rkey == prkey) {
 			status = NVME_SC_SUCCESS;
 			uuid_copy(&hostid, &reg->hostid);
-			if (abort)
+			/*

drivers/nvme/target/pr.c static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, u64 prkey, · line 515

Quoted 10 lines
drivers/nvme/target/pr.c +7 -0
+			 * Aborting is best effort (NVMe Base Specification
+			 * section 8.1.25.7): skip the issuing host, as the
+			 * abort flow would otherwise wait for the preempt
+			 * command itself, which holds a reference on its
+			 * controller's per-controller ref until it completes.
+			 */
+			if (abort && !uuid_equal(&hostid, send_hostid))
 				nvmet_pr_set_ctrl_to_abort(req, &hostid);
 			nvmet_pr_unregister_one(pr, reg);
 			if (!uuid_equal(&hostid, send_hostid))

[Severity: Medium]
Does a Reservation Acquire command actually hold a reference on pc_ref?

In nvmet_parse_io_cmd(), nvmet_parse_pr_cmd() returns 0 for reservation opcodes and returns early before reaching nvmet_pr_get_ns_pc_ref():

drivers/nvme/target/core.c:nvmet_parse_io_cmd() {
	...
	if (req->ns->pr.enable) {
		ret = nvmet_parse_pr_cmd(req);
		if (!ret)
			return ret;
	}
	...
	if (req->ns->pr.enable) {
		ret = nvmet_pr_check_cmd_access(req);
		if (ret)
			return ret;

		ret = nvmet_pr_get_ns_pc_ref(req);
	}
	...
}

Because that is the only caller of nvmet_pr_get_ns_pc_ref(), the preempt command itself never holds a reference on pc_ref->ref and cannot cause the self-deadlock described here. Skipping nvmet_pr_set_ctrl_to_abort() for send_hostid instead stops nvmet from waiting for in-flight I/O commands on the newly unregistered issuing host's controllers to drain before completing the Preempt and Abort command.

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
author
#2 lore raw
Quoted 59 lines · 在 2026/11/20 16:18, [email protected] 写道:

Thank you for your contribution! Sashiko AI review found 1 potential issue to consider:
- [Medium] nvmet: nvmet_pr_unreg_all_host_by_prkey() skips abort on false pc_ref premise
--

commit 07f50eedfd5f5e5d148cd9fefe65d629fc37077e
Author: Guixin Liu <[email protected]>

nvmet: don't abort the preempting host itself

This commit skips calling nvmet_pr_set_ctrl_to_abort() for the issuing host in nvmet_pr_unreg_all_host_by_prkey() when handling a Reservation Acquire command with the Preempt and Abort action.

Quoted 10 lines
drivers/nvme/target/pr.c +1 -1
index 8c818f24815bd..ed530d7362f5b 100644
@@ -502,7 +502,14 @@ static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, u64 prkey,
 		if (reg->rkey == prkey) {
 			status = NVME_SC_SUCCESS;
 			uuid_copy(&hostid, &reg->hostid);
-			if (abort)
+			/*

drivers/nvme/target/pr.c static u16 nvmet_pr_unreg_all_host_by_prkey(struct nvmet_req *req, u64 prkey, · line 515

Quoted 10 lines
drivers/nvme/target/pr.c +7 -0
+			 * Aborting is best effort (NVMe Base Specification
+			 * section 8.1.25.7): skip the issuing host, as the
+			 * abort flow would otherwise wait for the preempt
+			 * command itself, which holds a reference on its
+			 * controller's per-controller ref until it completes.
+			 */
+			if (abort && !uuid_equal(&hostid, send_hostid))
 				nvmet_pr_set_ctrl_to_abort(req, &hostid);
 			nvmet_pr_unregister_one(pr, reg);
 			if (!uuid_equal(&hostid, send_hostid))

[Severity: Medium]
Does a Reservation Acquire command actually hold a reference on pc_ref?

In nvmet_parse_io_cmd(), nvmet_parse_pr_cmd() returns 0 for reservation opcodes and returns early before reaching nvmet_pr_get_ns_pc_ref():

drivers/nvme/target/core.c:nvmet_parse_io_cmd() {
	...
	if (req->ns->pr.enable) {
		ret = nvmet_parse_pr_cmd(req);
		if (!ret)
			return ret;
	}
	...
	if (req->ns->pr.enable) {
		ret = nvmet_pr_check_cmd_access(req);
		if (ret)
			return ret;

		ret = nvmet_pr_get_ns_pc_ref(req);
	}
	...
}
Quoted 6 lines

Because that is the only caller of nvmet_pr_get_ns_pc_ref(), the preempt command itself never holds a reference on pc_ref->ref and cannot cause the self-deadlock described here. Skipping nvmet_pr_set_ctrl_to_abort() for send_hostid instead stops nvmet from waiting for in-flight I/O commands on the newly unregistered issuing host's controllers to drain before completing the Preempt and Abort command.

Sashiko is right, Please ignore this patch.

Quoted 1 line