Skip to contentSkip to conversation

nvmet-pci: validate endpoint queue IDs

PATCH In review Michael Bommarito 2 participants 5 messages Jul 10

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover nvmet-pci: validate endpoint queue IDs

A PCI root-complex host can crash an NVMe PCI endpoint target with malformed queue IDs. The endpoint transport allocates its SQ/CQ arrays using ctrl->nr_queues, which is capped by endpoint interrupt capacity, but the common target admin validation only checks queue IDs against subsys->max_qid. A host can therefore submit Create/Delete SQ/CQ commands with qids that pass the common checks yet index past the smaller endpoint transport arrays.

Patch 1 rejects queue IDs outside ctrl->nr_queues before the endpoint SQ/CQ arrays are indexed. Patch 2 adds same-translation-unit KUnit/KASAN
coverage: a valid queue ID that must still be accepted and the out-of-range Create/Delete SQ/CQ cases that must now be rejected.

Reproduced with the KUnit/KASAN test: the stock Create CQ path faults in nvmet_pci_epf_create_cq() after nvmet_check_io_cqid() accepts qid 2 with max_qid 8 and nr_queues 2; patched rejects the malformed cases while the benign control still passes.

Michael Bommarito (2):
nvmet-pci: validate queue IDs against endpoint queues
nvmet-pci: add KUnit coverage for endpoint queue IDs

 drivers/nvme/target/Kconfig   |  11 +++
 drivers/nvme/target/pci-epf.c | 151 ++++++++++++++++++++++++++++++++--
 2 files changed, 157 insertions(+), 5 deletions(-)

--
2.53.0

#1 lore raw

PATCH 1/2 nvmet-pci: validate queue IDs against endpoint queues

The NVMe PCI endpoint transport allocates SQ/CQ arrays using ctrl->nr_queues, which is capped by endpoint interrupt capacity. Common target admin validation only checks queue IDs against subsys->max_qid, so a root-complex host can submit Create/Delete SQ/CQ commands with qids that pass the common checks but index past the smaller endpoint transport arrays.

Impact: A PCI root-complex host can crash an NVMe PCI endpoint target with malformed queue IDs.

Reject queue IDs that are outside ctrl->nr_queues before indexing the endpoint SQ/CQ arrays.

Fixes 0faa0fe6f90e ("nvmet: New NVMe PCI endpoint function target driver")
Assisted-by Codex:gpt-5-5-xhigh

---

I reproduced this with a same-translation-unit KUnit/KASAN test. The stock Create CQ path faults in nvmet_pci_epf_create_cq() after nvmet_check_io_cqid() accepts qid 2 with max_qid 8 and nr_queues 2. The patched checks reject malformed Create/Delete SQ/CQ cases while the benign control still passes.

 drivers/nvme/target/pci-epf.c | 31 ++++++++++++++++++++++++++-----
 1 file changed, 26 insertions(+), 5 deletions(-)
drivers/nvme/target/pci-epf.c +26 -5
index 4e9db96ebfecd..5bddda09c0538 100644
@@ -1267,10 +1267,15 @@ static u16 nvmet_pci_epf_create_cq(struct nvmet_ctrl *tctrl,
 		u16 cqid, u16 flags, u16 qsize, u64 pci_addr, u16 vector)
 {
 	struct nvmet_pci_epf_ctrl *ctrl = tctrl->drvdata;
-	struct nvmet_pci_epf_queue *cq = &ctrl->cq[cqid];
+	struct nvmet_pci_epf_queue *cq;
 	u16 status;
 	int ret;
 
+	if (cqid >= ctrl->nr_queues)
+		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+
+	cq = &ctrl->cq[cqid];
+
 	if (test_bit(NVMET_PCI_EPF_Q_LIVE, &cq->flags))
 		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
 
@@ -1348,7 +1353,12 @@ static u16 nvmet_pci_epf_create_cq(struct nvmet_ctrl *tctrl,
 static u16 nvmet_pci_epf_delete_cq(struct nvmet_ctrl *tctrl, u16 cqid)
 {
 	struct nvmet_pci_epf_ctrl *ctrl = tctrl->drvdata;
-	struct nvmet_pci_epf_queue *cq = &ctrl->cq[cqid];
+	struct nvmet_pci_epf_queue *cq;
+
+	if (cqid >= ctrl->nr_queues)
+		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+
+	cq = &ctrl->cq[cqid];
 
 	if (!test_and_clear_bit(NVMET_PCI_EPF_Q_LIVE, &cq->flags))
 		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
@@ -1367,10 +1377,16 @@ static u16 nvmet_pci_epf_create_sq(struct nvmet_ctrl *tctrl,
 		u16 sqid, u16 cqid, u16 flags, u16 qsize, u64 pci_addr)
 {
 	struct nvmet_pci_epf_ctrl *ctrl = tctrl->drvdata;
-	struct nvmet_pci_epf_queue *sq = &ctrl->sq[sqid];
-	struct nvmet_pci_epf_queue *cq = &ctrl->cq[cqid];
+	struct nvmet_pci_epf_queue *sq;
+	struct nvmet_pci_epf_queue *cq;
 	u16 status;
 
+	if (sqid >= ctrl->nr_queues || cqid >= ctrl->nr_queues)
+		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+
+	sq = &ctrl->sq[sqid];
+	cq = &ctrl->cq[cqid];
+
 	if (test_bit(NVMET_PCI_EPF_Q_LIVE, &sq->flags))
 		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
 
@@ -1419,7 +1435,12 @@ static u16 nvmet_pci_epf_create_sq(struct nvmet_ctrl *tctrl,
 static u16 nvmet_pci_epf_delete_sq(struct nvmet_ctrl *tctrl, u16 sqid)
 {
 	struct nvmet_pci_epf_ctrl *ctrl = tctrl->drvdata;
-	struct nvmet_pci_epf_queue *sq = &ctrl->sq[sqid];
+	struct nvmet_pci_epf_queue *sq;
+
+	if (sqid >= ctrl->nr_queues)
+		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+
+	sq = &ctrl->sq[sqid];
 
 	if (!test_and_clear_bit(NVMET_PCI_EPF_Q_LIVE, &sq->flags))
 		return NVME_SC_QID_INVALID | NVME_STATUS_DNR;
Signature · 1 line
2.53.0
Reviewed-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] nvmet-pci: validate queue IDs against endpoint queues

Quoted 10 lines · Michael Bommarito

The NVMe PCI endpoint transport allocates SQ/CQ arrays using ctrl->nr_queues, which is capped by endpoint interrupt capacity. Common target admin validation only checks queue IDs against subsys->max_qid, so a root-complex host can submit Create/Delete SQ/CQ commands with qids that pass the common checks but index past the smaller endpoint transport arrays.

Impact: A PCI root-complex host can crash an NVMe PCI endpoint target with malformed queue IDs.

Quoted 7 lines

Reject queue IDs that are outside ctrl->nr_queues before indexing the endpoint SQ/CQ arrays.

Fixes 0faa0fe6f90e ("nvmet: New NVMe PCI endpoint function target driver")
Assisted-by Codex:gpt-5-5-xhigh

Looks good.

Signature · 2 lines
Damien Le Moal
Western Digital Research
#3 lore raw

PATCH 2/2 nvmet-pci: add KUnit coverage for endpoint queue IDs

In reply to Michael Bommarito

Add KUnit coverage for the PCI endpoint target queue-id boundary. The tests model the case where target-core max_qid is larger than the endpoint transport's ctrl->nr_queues, confirm the common qid check accepts the malformed id, and verify the endpoint callbacks reject out-of-range Create/Delete SQ/CQ requests before indexing transport-private arrays.

This covers the regression fixed by the preceding patch.

Assisted-by Codex:gpt-5-5-xhigh
---
 drivers/nvme/target/Kconfig   |  11 ++++
 drivers/nvme/target/pci-epf.c | 120 ++++++++++++++++++++++++++++++++++
 2 files changed, 131 insertions(+)
drivers/nvme/target/Kconfig +11 -0
index 4904097dfd490..ea64bbe9882c5 100644
@@ -127,3 +127,14 @@ config NVME_TARGET_PCI_EPF
 	  capable PCI controller.
 
 	  If unsure, say N.
+
+config NVMET_PCI_EPF_KUNIT_TEST
+	bool "NVMe PCI endpoint target KUnit tests" if !KUNIT_ALL_TESTS
+	depends on KUNIT
+	depends on NVME_TARGET_PCI_EPF=y
+	default KUNIT_ALL_TESTS
+	help
+	  KUnit tests for the NVMe PCI endpoint target transport.
+	  These tests exercise transport-private queue ID checks for
+	  Create/Delete SQ/CQ commands when target-core max_qid is larger than
+	  the endpoint controller's available queue arrays.
drivers/nvme/target/pci-epf.c +120 -0

129 lines not on this page · raw message

Signature · 1 line
2.53.0
Reviewed-by on v1 2/2
#4 lore raw

Re: [PATCH 2/2] nvmet-pci: add KUnit coverage for endpoint queue IDs

Quoted 10 lines · Michael Bommarito

Add KUnit coverage for the PCI endpoint target queue-id boundary. The tests model the case where target-core max_qid is larger than the endpoint transport's ctrl->nr_queues, confirm the common qid check accepts the malformed id, and verify the endpoint callbacks reject out-of-range Create/Delete SQ/CQ requests before indexing transport-private arrays.

This covers the regression fixed by the preceding patch.

Assisted-by Codex:gpt-5-5-xhigh

I am not very familiar with kunit tests, but looks OK.

Signature · 2 lines
Damien Le Moal
Western Digital Research