RDMA/rxe: validate num_sge/cur_sge before indexing wqe->dma.sge[]

A user QP's send queue (qp->sq.queue) is a shared ring the userspace application writes to directly via mmap (vmalloc_user(), see rxe_queue.c). For such a QP, rxe_post_send() takes the qp->is_user branch and only schedules the requester task -- it never validates or copies the posted WQE:

	drivers/infiniband/sw/rxe/rxe_verbs.c:
		if (qp->is_user) {
			rxe_sched_task(&qp->send_task);
			...
		}

The requester then consumes the WQE in place straight out of that mmap'd ring:

	rxe_req.c:	wqe = req_next_wqe(qp);
	rxe_req.c:	err = copy_data(qp->pd, 0, &wqe->dma,
					payload_addr(pkt), payload,
					RXE_FROM_MR_OBJ);

copy_data() immediately indexes the per-WQE sge array with the attacker-controlled cur_sge field, before any bound is checked:

	rxe_mr.c:	struct rxe_sge *sge = &dma->sge[dma->cur_sge];
	rxe_mr.c:	...
	rxe_mr.c:	if (sge->length && (offset < sge->length)) {

dma->sge[] is a flex array whose real backing storage is exactly qp->sq.max_sge entries per WQE slot (see rxe_qp.c, wqe_size computed from max_sge at QP create time). Since a user QP's WQE bytes are entirely attacker-supplied, both wqe->dma.num_sge and wqe->dma.cur_sge can be set to arbitrary values independent of each other and of max_sge. Only the *kernel*-QP post path bounds num_sge:

	rxe_verbs.c: validate_send_wr()
		if (num_sge > sq->max_sge) {
			rxe_err_qp(qp, "num_sge > max_sge\n");

but that function is only reachable from rxe_post_one_send() for kernel-owned QPs; it is never consulted for a user QP's raw WQE.

The sibling receive path already has the equivalent guard, with the literal comment documenting exactly why it is required:

	rxe_resp.c: get_srq_wqe()
		/* don't trust user space data */
		if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
			...
			rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
			return RESPST_ERR_MALFORMED_WQE;
		}

The send/requester path has no analogous check, so a local, unprivileged user who can open /dev/infiniband/uverbs* and create a user QP on a soft-RoCE (rxe) link can hand-craft a WQE in the shared send queue with an out-of-range wqe->dma.cur_sge (or an oversized wqe->dma.num_sge) and ring the send doorbell. rxe_requester() then calls copy_data(), which dereferences &dma->sge[cur_sge] out of the bounds of the per-WQE sge array -- a vmalloc out-of-bounds *read* (confirmed via KASAN: "KASAN: vmalloc-out-of-bounds in copy_data"), reliably panicking the kernel (local DoS). sge->addr itself is still bounds-checked later by lookup_mr()/rxe_mr_copy(), so the primitive is an OOB read of sge metadata, not an arbitrary read/write primitive.

Fix this the same way get_srq_wqe() already does for SRQ entries:
bound both fields pulled from the (possibly user-mapped) send queue entry before they are ever used to index wqe->dma.sge[], right where the requester fetches the next WQE off the ring in rxe_requester(). num_sge is capped at qp->sq.max_sge (matching the sibling SRQ check and the kernel-QP validate_send_wr() check), and cur_sge is capped at qp->sq.max_sge directly, since that is the true per-WQE array capacity that copy_data() indexes into -- this also covers num_sge == 0 / cur_sge == 0 local-op and zero-payload WQEs, which remain valid.

This is a long-standing bug in the rxe (soft-RoCE) driver: the qp->is_user bypass in rxe_post_send() and the unbounded &dma->sge[dma->cur_sge] indexing in copy_data() have been present since the driver was introduced.

Runtime-verified on a v6.19 KASAN (CONFIG_KASAN_VMALLOC=y) stand: a reproducer that posts a user QP send WQE with an out-of-range cur_sge reliably tripped "KASAN: vmalloc-out-of-bounds in copy_data" (an out-of-bounds read) before this patch, and no longer triggers that report with the patch applied.

Fixes 8700e3e7c485 ("Soft RoCE driver")

Cc [email protected]

Signed-off-by Ibrahim Hashimov <[email protected]>

Assisted-by AuditCode-AI:2026.07

---
  drivers/infiniband/sw/rxe/rxe_req.c | 16 ++++++++++++++++
  1 file changed, 16 insertions(+)

drivers/infiniband/sw/rxe/rxe_req.c +0 -0

index 12d03f390b09..9fb2c49fb503 100644
@@ -701,6 +701,22 @@ int rxe_requester(struct rxe_qp *qp)
  	if (unlikely(!wqe))
  		goto exit;
 

drivers/infiniband/sw/rxe/rxe_req.c +6 -0

+		     wqe->dma.cur_sge >= qp->sq.max_sge)) {
+		rxe_dbg_qp(qp, "invalid num_sge/cur_sge in send wqe\n");
+		wqe->status = IB_WC_LOC_QP_OP_ERR;
+		goto err;
+	}
+
  	if (rxe_wqe_is_fenced(qp, wqe)) {
  		qp->req.wait_fence = 1;
  		goto exit;

From this function,
static int rxe_init_sq(struct rxe_qp *qp, struct ibv_qp_init_attr *init_attr)
...

     qp->sq.max_sge = init_attr->cap.max_send_sge;

...
qp->sq.max_sge is also from the user space application. It is possible that qp->sq.max_sge is 0.
Then this makes rxe_requester will return error and exit.

A user QP's send queue (qp->sq.queue) is a shared ring the userspace application writes to directly via mmap (vmalloc_user(), see rxe_queue.c). For such a QP, rxe_post_send() takes the qp->is_user branch and only schedules the requester task -- it never validates or copies the posted WQE:

	drivers/infiniband/sw/rxe/rxe_verbs.c:
		if (qp->is_user) {
			rxe_sched_task(&qp->send_task);
			...
		}

The requester then consumes the WQE in place straight out of that mmap'd ring:

	rxe_req.c:	wqe = req_next_wqe(qp);
	rxe_req.c:	err = copy_data(qp->pd, 0, &wqe->dma,
					payload_addr(pkt), payload,
					RXE_FROM_MR_OBJ);

copy_data() indexes the per-WQE sge array with the attacker-controlled cur_sge field and dereferences it (once there is payload to copy):

	rxe_mr.c:	struct rxe_sge *sge = &dma->sge[dma->cur_sge];
	rxe_mr.c:	...
	rxe_mr.c:	if (sge->length && (offset < sge->length)) {

dma->sge[] is a flex array whose real backing storage is exactly qp->sq.max_sge entries per WQE slot (see rxe_qp.c, wqe_size computed from max_sge at QP create time). Since a user QP's WQE bytes are entirely attacker-supplied, both wqe->dma.num_sge and wqe->dma.cur_sge can be set to arbitrary values independent of each other and of max_sge. Only the *kernel*-QP post path bounds num_sge:

	rxe_verbs.c: validate_send_wr()
		if (num_sge > sq->max_sge) {
			rxe_err_qp(qp, "num_sge > max_sge\n");

but that function is only reachable from rxe_post_one_send() for kernel-owned QPs; it is never consulted for a user QP's raw WQE.

The sibling receive path already has the equivalent guard, with the literal comment documenting exactly why it is required:

	rxe_resp.c: get_srq_wqe()
		/* don't trust user space data */
		if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
			...
			rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
			return RESPST_ERR_MALFORMED_WQE;
		}

The send/requester path has no analogous check, so a local, unprivileged user who can open /dev/infiniband/uverbs* and create a user QP on a soft-RoCE (rxe) link can hand-craft a WQE in the shared send queue with an out-of-range wqe->dma.cur_sge (or an oversized wqe->dma.num_sge) and ring the send doorbell. rxe_requester() then calls copy_data(), which dereferences &dma->sge[cur_sge] out of the bounds of the per-WQE sge array -- a vmalloc out-of-bounds *read* (confirmed via KASAN: "KASAN: vmalloc-out-of-bounds in copy_data"), reliably panicking the kernel (local DoS). sge->addr itself is still bounds-checked later by lookup_mr()/rxe_mr_copy(), so the primitive is an OOB read of sge metadata, not an arbitrary read/write primitive.

Fix this the same way get_srq_wqe() already does for SRQ entries:
bound the fields pulled from the (possibly user-mapped) send queue entry before they are used to index wqe->dma.sge[], right where the requester fetches the next WQE off the ring in rxe_requester(). num_sge is capped at qp->sq.max_sge (matching the sibling SRQ check and the kernel-QP validate_send_wr() check). cur_sge is bounded only when the WQE actually carries payload (wqe->dma.resid): copy_data() dereferences dma->sge[cur_sge] only after its own length == 0 early return, so a zero-payload WQE never touches the sge array and must not be rejected -- notably that is the only kind of WQE a max_sge == 0 QP can post (qp->sq.max_sge is itself derived from user-supplied max_send_sge / max_inline_data and may be 0). Gating on resid rather than num_sge is
deliberate: payload is wqe->dma.resid, which is independent of num_sge, so a WQE with num_sge == 0 but a large resid and an out-of-range cur_sge would still reach the sge dereference.

This is a long-standing bug in the rxe (soft-RoCE) driver: the qp->is_user bypass in rxe_post_send() and the unbounded &dma->sge[dma->cur_sge] indexing in copy_data() have been present since the driver was introduced.

Runtime-verified on a v6.19 KASAN (CONFIG_KASAN_VMALLOC=y) stand: a reproducer that posts a user QP send WQE with an out-of-range cur_sge reliably tripped "KASAN: vmalloc-out-of-bounds in copy_data" (an out-of-bounds read) before this patch, and no longer triggers that report with the patch applied.

Fixes 8700e3e7c485 ("Soft RoCE driver")

Cc [email protected]

Signed-off-by Ibrahim Hashimov <[email protected]>

Assisted-by AuditCode-AI:2026.07

---
v2: address Zhu Yanjun's review of v1

     (https://lore.kernel.org/linux-rdma/[email protected]/):
     qp->sq.max_sge can legitimately be 0, and v1's unconditional
     "cur_sge >= max_sge" check then wrongly rejected a valid zero-payload
     WQE. Gate the cur_sge bound on wqe->dma.resid instead (copy_data()
     dereferences dma->sge[] only when there is payload), so zero-payload
     WQEs -- the only kind a max_sge == 0 QP can post -- are accepted while
     the out-of-range cur_sge OOB is still rejected. Commit message fixed.

  drivers/infiniband/sw/rxe/rxe_req.c | 22 ++++++++++++++++++++++
  1 file changed, 22 insertions(+)

drivers/infiniband/sw/rxe/rxe_req.c +22 -0

index 12d03f390b09..363c56a1edbb 100644
@@ -701,6 +701,28 @@ int rxe_requester(struct rxe_qp *qp)
  	if (unlikely(!wqe))
  		goto exit;
 
+	/*
+	 * Don't trust user space data: for a user QP, qp->sq.queue is a
+	 * raw ring the application writes directly, so this WQE's num_sge
+	 * and cur_sge are attacker-controlled. copy_data() dereferences
+	 * dma->sge[cur_sge] without bounding the initial cur_sge against
+	 * the per-WQE sge array, whose capacity is qp->sq.max_sge (the
+	 * loop there only bounds subsequent increments, against num_sge).
+	 * Bound num_sge to that capacity, the way get_srq_wqe() and
+	 * validate_send_wr() already do, and bound cur_sge only when the
+	 * WQE actually carries payload (dma.resid): copy_data() returns
+	 * early on a zero-length copy before it ever touches dma->sge[],
+	 * so a zero-payload WQE -- the only valid WQE on a max_sge == 0
+	 * QP -- must not be rejected here.
+	 */
+	if (unlikely(wqe->dma.num_sge > qp->sq.max_sge ||
+		     (wqe->dma.resid &&
+		      wqe->dma.cur_sge >= qp->sq.max_sge))) {
+		rxe_dbg_qp(qp, "invalid num_sge/cur_sge in send wqe\n");
+		wqe->status = IB_WC_LOC_QP_OP_ERR;
+		goto err;
+	}
+
  	if (rxe_wqe_is_fenced(qp, wqe)) {
  		qp->req.wait_fence = 1;
  		goto exit;

Quoted 97 lines · Ibrahim Hashimov

A user QP's send queue (qp->sq.queue) is a shared ring the userspace application writes to directly via mmap (vmalloc_user(), see rxe_queue.c). For such a QP, rxe_post_send() takes the qp->is_user branch and only schedules the requester task -- it never validates or copies the posted WQE:

	drivers/infiniband/sw/rxe/rxe_verbs.c:
		if (qp->is_user) {
			rxe_sched_task(&qp->send_task);
			...
		}

The requester then consumes the WQE in place straight out of that mmap'd ring:

	rxe_req.c:	wqe = req_next_wqe(qp);
	rxe_req.c:	err = copy_data(qp->pd, 0, &wqe->dma,
					payload_addr(pkt), payload,
					RXE_FROM_MR_OBJ);

copy_data() indexes the per-WQE sge array with the attacker-controlled cur_sge field and dereferences it (once there is payload to copy):

	rxe_mr.c:	struct rxe_sge *sge = &dma->sge[dma->cur_sge];
	rxe_mr.c:	...
	rxe_mr.c:	if (sge->length && (offset < sge->length)) {

dma->sge[] is a flex array whose real backing storage is exactly qp->sq.max_sge entries per WQE slot (see rxe_qp.c, wqe_size computed from max_sge at QP create time). Since a user QP's WQE bytes are entirely attacker-supplied, both wqe->dma.num_sge and wqe->dma.cur_sge can be set to arbitrary values independent of each other and of max_sge. Only the *kernel*-QP post path bounds num_sge:

	rxe_verbs.c: validate_send_wr()
		if (num_sge > sq->max_sge) {
			rxe_err_qp(qp, "num_sge > max_sge\n");

but that function is only reachable from rxe_post_one_send() for kernel-owned QPs; it is never consulted for a user QP's raw WQE.

The sibling receive path already has the equivalent guard, with the literal comment documenting exactly why it is required:

	rxe_resp.c: get_srq_wqe()
		/* don't trust user space data */
		if (unlikely(wqe->dma.num_sge > srq->rq.max_sge)) {
			...
			rxe_dbg_qp(qp, "invalid num_sge in SRQ entry\n");
			return RESPST_ERR_MALFORMED_WQE;
		}

The send/requester path has no analogous check, so a local, unprivileged user who can open /dev/infiniband/uverbs* and create a user QP on a soft-RoCE (rxe) link can hand-craft a WQE in the shared send queue with an out-of-range wqe->dma.cur_sge (or an oversized wqe->dma.num_sge) and ring the send doorbell. rxe_requester() then calls copy_data(), which dereferences &dma->sge[cur_sge] out of the bounds of the per-WQE sge array -- a vmalloc out-of-bounds *read* (confirmed via KASAN: "KASAN: vmalloc-out-of-bounds in copy_data"), reliably panicking the kernel (local DoS). sge->addr itself is still bounds-checked later by lookup_mr()/rxe_mr_copy(), so the primitive is an OOB read of sge metadata, not an arbitrary read/write primitive.

Fix this the same way get_srq_wqe() already does for SRQ entries:
bound the fields pulled from the (possibly user-mapped) send queue entry before they are used to index wqe->dma.sge[], right where the requester fetches the next WQE off the ring in rxe_requester(). num_sge is capped at qp->sq.max_sge (matching the sibling SRQ check and the kernel-QP validate_send_wr() check). cur_sge is bounded only when the WQE actually carries payload (wqe->dma.resid): copy_data() dereferences dma->sge[cur_sge] only after its own length == 0 early return, so a zero-payload WQE never touches the sge array and must not be rejected -- notably that is the only kind of WQE a max_sge == 0 QP can post (qp->sq.max_sge is itself derived from user-supplied max_send_sge / max_inline_data and may be 0). Gating on resid rather than num_sge is
deliberate: payload is wqe->dma.resid, which is independent of num_sge, so a WQE with num_sge == 0 but a large resid and an out-of-range cur_sge would still reach the sge dereference.

This is a long-standing bug in the rxe (soft-RoCE) driver: the qp->is_user bypass in rxe_post_send() and the unbounded &dma->sge[dma->cur_sge] indexing in copy_data() have been present since the driver was introduced.

Runtime-verified on a v6.19 KASAN (CONFIG_KASAN_VMALLOC=y) stand: a reproducer that posts a user QP send WQE with an out-of-range cur_sge reliably tripped "KASAN: vmalloc-out-of-bounds in copy_data" (an out-of-bounds read) before this patch, and no longer triggers that report with the patch applied.

Fixes 8700e3e7c485 ("Soft RoCE driver")

Cc [email protected]

Signed-off-by Ibrahim Hashimov <[email protected]>

Assisted-by AuditCode-AI:2026.07

---
v2: address Zhu Yanjun's review of v1

Quoted 7 lines

     (https://lore.kernel.org/linux-rdma/[email protected]/):
     qp->sq.max_sge can legitimately be 0, and v1's unconditional
     "cur_sge >= max_sge" check then wrongly rejected a valid zero-payload
     WQE. Gate the cur_sge bound on wqe->dma.resid instead (copy_data()
     dereferences dma->sge[] only when there is payload), so zero-payload
     WQEs -- the only kind a max_sge == 0 QP can post -- are accepted while
     the out-of-range cur_sge OOB is still rejected. Commit message fixed.

Thanks a lot. I am fine with this. Please Leon and Jason comment on this.

Reviewed-by Zhu Yanjun <[email protected]>

Zhu Yanjun

Quoted 36 lines

  drivers/infiniband/sw/rxe/rxe_req.c | 22 ++++++++++++++++++++++
  1 file changed, 22 insertions(+)

drivers/infiniband/sw/rxe/rxe_req.c +22 -0

index 12d03f390b09..363c56a1edbb 100644
@@ -701,6 +701,28 @@ int rxe_requester(struct rxe_qp *qp)
  	if (unlikely(!wqe))
  		goto exit;
+	/*
+	 * Don't trust user space data: for a user QP, qp->sq.queue is a
+	 * raw ring the application writes directly, so this WQE's num_sge
+	 * and cur_sge are attacker-controlled. copy_data() dereferences
+	 * dma->sge[cur_sge] without bounding the initial cur_sge against
+	 * the per-WQE sge array, whose capacity is qp->sq.max_sge (the
+	 * loop there only bounds subsequent increments, against num_sge).
+	 * Bound num_sge to that capacity, the way get_srq_wqe() and
+	 * validate_send_wr() already do, and bound cur_sge only when the
+	 * WQE actually carries payload (dma.resid): copy_data() returns
+	 * early on a zero-length copy before it ever touches dma->sge[],
+	 * so a zero-payload WQE -- the only valid WQE on a max_sge == 0
+	 * QP -- must not be rejected here.
+	 */
+	if (unlikely(wqe->dma.num_sge > qp->sq.max_sge ||
+		     (wqe->dma.resid &&
+		      wqe->dma.cur_sge >= qp->sq.max_sge))) {
+		rxe_dbg_qp(qp, "invalid num_sge/cur_sge in send wqe\n");
+		wqe->status = IB_WC_LOC_QP_OP_ERR;
+		goto err;
+	}
+
  	if (rxe_wqe_is_fenced(qp, wqe)) {
  		qp->req.wait_fence = 1;
  		goto exit;