Skip to contentSkip to conversation

Fix memory ordering errors in queues

PATCH Applied Bob Pearson 3 participants 9 messages May 27, 2021 – Jun 3, 2021

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v3 cover Fix memory ordering errors in queues

These patches optimize the memory ordering in rxe_queue.h so that user space and not kernel space indices are protected for loads with smp_load_acquire() and stores with smp_store_release(). The original implementation of this did not apply to all index references which has recently caused test case errors traced to stale memory loads. These patches fix those errors and also protect kernel indices from malicious modification by user space.

Fixes d21a1240f516 ("RDMA/rxe: Use acquire/release for memory ordering")

---
v3:
Protected kernel index in shared queues from modification by user space. Pass queue type to allow compiler to optimize queue methods.
v2:
Add a way to distinguish between user and kernel indices.
v1:
Add missing smp_load_acquire() calls.
---
Bob Pearson (3):
RDMA/rxe: Add a type flag to rxe_queue structs
RDMA/rxe: Protect user space index loads/stores
RDMA/rxe: Protext kernel index from user space

 drivers/infiniband/sw/rxe/rxe_comp.c  |  31 ++-
 drivers/infiniband/sw/rxe/rxe_cq.c    |  32 ++-
 drivers/infiniband/sw/rxe/rxe_qp.c    |  22 ++-
 drivers/infiniband/sw/rxe/rxe_queue.c |  21 +-
 drivers/infiniband/sw/rxe/rxe_queue.h | 272 ++++++++++++++++++++------
 drivers/infiniband/sw/rxe/rxe_req.c   |  46 +++--
 drivers/infiniband/sw/rxe/rxe_resp.c  |  44 ++++-
 drivers/infiniband/sw/rxe/rxe_srq.c   |   5 +-
 drivers/infiniband/sw/rxe/rxe_verbs.c |  80 ++++++--
 drivers/infiniband/sw/rxe/rxe_verbs.h |   5 +-
 10 files changed, 423 insertions(+), 135 deletions(-)
Signature · 1 line
2.30.2
#1 lore raw

PATCH v3 1/3 RDMA/rxe: Add a type flag to rxe_queue structs

To create optimal code only want to use smp_load_acquire() and smp_store_release() for user indices in rxe_queue APIs since kernel indices are protected by locks which also act as memory barriers. By adding a type to the queues we can determine which indices need to be protected.

---
 drivers/infiniband/sw/rxe/rxe_cq.c    |  4 +++-
 drivers/infiniband/sw/rxe/rxe_qp.c    | 12 ++++++++----
 drivers/infiniband/sw/rxe/rxe_queue.c |  8 ++++----
 drivers/infiniband/sw/rxe/rxe_queue.h | 13 ++++++++++---
 drivers/infiniband/sw/rxe/rxe_srq.c   |  4 +++-
 5 files changed, 28 insertions(+), 13 deletions(-)
drivers/infiniband/sw/rxe/rxe_cq.c +3 -1
index b315ebf041ac..1d4d8a31bc12 100644
@@ -59,9 +59,11 @@ int rxe_cq_from_init(struct rxe_dev *rxe, struct rxe_cq *cq, int cqe,
 		     struct rxe_create_cq_resp __user *uresp)
 {
 	int err;
+	enum queue_type type;
 
+	type = uresp ? QUEUE_TYPE_TO_USER : QUEUE_TYPE_KERNEL;
 	cq->queue = rxe_queue_init(rxe, &cqe,
-				   sizeof(struct rxe_cqe));
+			sizeof(struct rxe_cqe), type);
 	if (!cq->queue) {
 		pr_warn("unable to create cq\n");
 		return -ENOMEM;
drivers/infiniband/sw/rxe/rxe_qp.c +8 -4
index 34ae957a315c..9bd6bf8f9bd9 100644
@@ -206,6 +206,7 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
 {
 	int err;
 	int wqe_size;
+	enum queue_type type;
 
 	err = sock_create_kern(&init_net, AF_INET, SOCK_DGRAM, 0, &qp->sk);
 	if (err < 0)
@@ -231,7 +232,9 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
 	qp->sq.max_inline = init->cap.max_inline_data = wqe_size;
 	wqe_size += sizeof(struct rxe_send_wqe);
 
-	qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr, wqe_size);
+	type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
+	qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr,
+				wqe_size, type);
 	if (!qp->sq.queue)
 		return -ENOMEM;
 
@@ -273,6 +276,7 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
 {
 	int err;
 	int wqe_size;
+	enum queue_type type;
 
 	if (!qp->srq) {
 		qp->rq.max_wr		= init->cap.max_recv_wr;
@@ -283,9 +287,9 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
 		pr_debug("qp#%d max_wr = %d, max_sge = %d, wqe_size = %d\n",
 			 qp_num(qp), qp->rq.max_wr, qp->rq.max_sge, wqe_size);
 
-		qp->rq.queue = rxe_queue_init(rxe,
-					      &qp->rq.max_wr,
-					      wqe_size);
+		type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
+		qp->rq.queue = rxe_queue_init(rxe, &qp->rq.max_wr,
+					wqe_size, type);
 		if (!qp->rq.queue)
 			return -ENOMEM;
 
drivers/infiniband/sw/rxe/rxe_queue.c +4 -4
index fa69241b1187..8f844d0b9e77 100644
@@ -52,9 +52,8 @@ inline void rxe_queue_reset(struct rxe_queue *q)
 	memset(q->buf->data, 0, q->buf_size - sizeof(struct rxe_queue_buf));
 }
 
-struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
-				 int *num_elem,
-				 unsigned int elem_size)
+struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe, int *num_elem,
+			unsigned int elem_size, enum queue_type type)
 {
 	struct rxe_queue *q;
 	size_t buf_size;
@@ -69,6 +68,7 @@ struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
 		goto err1;
 
 	q->rxe = rxe;
+	q->type = type;
 
 	/* used in resize, only need to copy used part of queue */
 	q->elem_size = elem_size;
@@ -136,7 +136,7 @@ int rxe_queue_resize(struct rxe_queue *q, unsigned int *num_elem_p,
 	int err;
 	unsigned long flags = 0, flags1;
 
-	new_q = rxe_queue_init(q->rxe, &num_elem, elem_size);
+	new_q = rxe_queue_init(q->rxe, &num_elem, elem_size, q->type);
 	if (!new_q)
 		return -ENOMEM;
 
drivers/infiniband/sw/rxe/rxe_queue.h +10 -3
index 2902ca7b288c..4512745419f8 100644
@@ -19,6 +19,13 @@
  * of the queue is one less than the number of element slots
  */
 
+/* type of queue */
+enum queue_type {
+	QUEUE_TYPE_KERNEL,
+	QUEUE_TYPE_TO_USER,
+	QUEUE_TYPE_FROM_USER,
+};
+
 struct rxe_queue {
 	struct rxe_dev		*rxe;
 	struct rxe_queue_buf	*buf;
@@ -27,6 +34,7 @@ struct rxe_queue {
 	size_t			elem_size;
 	unsigned int		log2_elem_size;
 	u32			index_mask;
+	enum queue_type		type;
 };
 
 int do_mmap_info(struct rxe_dev *rxe, struct mminfo __user *outbuf,
@@ -35,9 +43,8 @@ int do_mmap_info(struct rxe_dev *rxe, struct mminfo __user *outbuf,
 
 void rxe_queue_reset(struct rxe_queue *q);
 
-struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
-				 int *num_elem,
-				 unsigned int elem_size);
+struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe, int *num_elem,
+			unsigned int elem_size, enum queue_type type);
 
 int rxe_queue_resize(struct rxe_queue *q, unsigned int *num_elem_p,
 		     unsigned int elem_size, struct ib_udata *udata,
drivers/infiniband/sw/rxe/rxe_srq.c +3 -1
index 41b0d1e11baf..52c5593741ec 100644
@@ -78,6 +78,7 @@ int rxe_srq_from_init(struct rxe_dev *rxe, struct rxe_srq *srq,
 	int err;
 	int srq_wqe_size;
 	struct rxe_queue *q;
+	enum queue_type type;
 
 	srq->ibsrq.event_handler	= init->event_handler;
 	srq->ibsrq.srq_context		= init->srq_context;
@@ -91,8 +92,9 @@ int rxe_srq_from_init(struct rxe_dev *rxe, struct rxe_srq *srq,
 	spin_lock_init(&srq->rq.producer_lock);
 	spin_lock_init(&srq->rq.consumer_lock);
 
+	type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
 	q = rxe_queue_init(rxe, &srq->rq.max_wr,
-			   srq_wqe_size);
+			srq_wqe_size, type);
 	if (!q) {
 		pr_warn("unable to allocate queue for srq\n");
 		return -ENOMEM;
Signature · 1 line
2.30.2
maintainer
#2 lore raw

Re: [PATCH for-next v3 1/3] RDMA/rxe: Add a type flag to rxe_queue structs

Quoted 21 lines · Bob Pearson

To create optimal code only want to use smp_load_acquire() and smp_store_release() for user indices in rxe_queue APIs since kernel indices are protected by locks which also act as memory barriers. By adding a type to the queues we can determine which indices need to be protected.

---
 drivers/infiniband/sw/rxe/rxe_cq.c    |  4 +++-
 drivers/infiniband/sw/rxe/rxe_qp.c    | 12 ++++++++----
 drivers/infiniband/sw/rxe/rxe_queue.c |  8 ++++----
 drivers/infiniband/sw/rxe/rxe_queue.h | 13 ++++++++++---
 drivers/infiniband/sw/rxe/rxe_srq.c   |  4 +++-
 5 files changed, 28 insertions(+), 13 deletions(-)
drivers/infiniband/sw/rxe/rxe_cq.c +0 -0
index b315ebf041ac..1d4d8a31bc12 100644

drivers/infiniband/sw/rxe/rxe_cq.c int rxe_cq_from_init(struct rxe_dev *rxe, struct rxe_cq *cq, int cqe, · line 64

Quoted 6 lines
drivers/infiniband/sw/rxe/rxe_cq.c +2 -0
@@ -59,9 +59,11 @@ int rxe_cq_from_init(struct rxe_dev *rxe, struct rxe_cq *cq, int cqe,
                     struct rxe_create_cq_resp __user *uresp)
 {
        int err;
+       enum queue_type type;
 
+       type = uresp ? QUEUE_TYPE_TO_USER : QUEUE_TYPE_KERNEL;

When is QUEUE_TYPE_TO_USER used? and when is QUEUE_TYPE_FROM_USER is used?

Zhu Yanjun

Quoted 147 lines
drivers/infiniband/sw/rxe/rxe_cq.c +1 -1
        cq->queue = rxe_queue_init(rxe, &cqe,
-                                  sizeof(struct rxe_cqe));
+                       sizeof(struct rxe_cqe), type);
        if (!cq->queue) {
                pr_warn("unable to create cq\n");
                return -ENOMEM;
drivers/infiniband/sw/rxe/rxe_qp.c +8 -4
index 34ae957a315c..9bd6bf8f9bd9 100644
@@ -206,6 +206,7 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
 {
        int err;
        int wqe_size;
+       enum queue_type type;
 
        err = sock_create_kern(&init_net, AF_INET, SOCK_DGRAM, 0, &qp->sk);
        if (err < 0)
@@ -231,7 +232,9 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
        qp->sq.max_inline = init->cap.max_inline_data = wqe_size;
        wqe_size += sizeof(struct rxe_send_wqe);
 
-       qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr, wqe_size);
+       type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
+       qp->sq.queue = rxe_queue_init(rxe, &qp->sq.max_wr,
+                               wqe_size, type);
        if (!qp->sq.queue)
                return -ENOMEM;
 
@@ -273,6 +276,7 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
 {
        int err;
        int wqe_size;
+       enum queue_type type;
 
        if (!qp->srq) {
                qp->rq.max_wr           = init->cap.max_recv_wr;
@@ -283,9 +287,9 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
                pr_debug("qp#%d max_wr = %d, max_sge = %d, wqe_size = %d\n",
                         qp_num(qp), qp->rq.max_wr, qp->rq.max_sge, wqe_size);
 
-               qp->rq.queue = rxe_queue_init(rxe,
-                                             &qp->rq.max_wr,
-                                             wqe_size);
+               type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
+               qp->rq.queue = rxe_queue_init(rxe, &qp->rq.max_wr,
+                                       wqe_size, type);
                if (!qp->rq.queue)
                        return -ENOMEM;
 
drivers/infiniband/sw/rxe/rxe_queue.c +4 -4
index fa69241b1187..8f844d0b9e77 100644
@@ -52,9 +52,8 @@ inline void rxe_queue_reset(struct rxe_queue *q)
        memset(q->buf->data, 0, q->buf_size - sizeof(struct rxe_queue_buf));
 }
 
-struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
-                                int *num_elem,
-                                unsigned int elem_size)
+struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe, int *num_elem,
+                       unsigned int elem_size, enum queue_type type)
 {
        struct rxe_queue *q;
        size_t buf_size;
@@ -69,6 +68,7 @@ struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
                goto err1;
 
        q->rxe = rxe;
+       q->type = type;
 
        /* used in resize, only need to copy used part of queue */
        q->elem_size = elem_size;
@@ -136,7 +136,7 @@ int rxe_queue_resize(struct rxe_queue *q, unsigned int *num_elem_p,
        int err;
        unsigned long flags = 0, flags1;
 
-       new_q = rxe_queue_init(q->rxe, &num_elem, elem_size);
+       new_q = rxe_queue_init(q->rxe, &num_elem, elem_size, q->type);
        if (!new_q)
                return -ENOMEM;
 
drivers/infiniband/sw/rxe/rxe_queue.h +10 -3
index 2902ca7b288c..4512745419f8 100644
@@ -19,6 +19,13 @@
  * of the queue is one less than the number of element slots
  */
 
+/* type of queue */
+enum queue_type {
+       QUEUE_TYPE_KERNEL,
+       QUEUE_TYPE_TO_USER,
+       QUEUE_TYPE_FROM_USER,
+};
+
 struct rxe_queue {
        struct rxe_dev          *rxe;
        struct rxe_queue_buf    *buf;
@@ -27,6 +34,7 @@ struct rxe_queue {
        size_t                  elem_size;
        unsigned int            log2_elem_size;
        u32                     index_mask;
+       enum queue_type         type;
 };
 
 int do_mmap_info(struct rxe_dev *rxe, struct mminfo __user *outbuf,
@@ -35,9 +43,8 @@ int do_mmap_info(struct rxe_dev *rxe, struct mminfo __user *outbuf,
 
 void rxe_queue_reset(struct rxe_queue *q);
 
-struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe,
-                                int *num_elem,
-                                unsigned int elem_size);
+struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe, int *num_elem,
+                       unsigned int elem_size, enum queue_type type);
 
 int rxe_queue_resize(struct rxe_queue *q, unsigned int *num_elem_p,
                     unsigned int elem_size, struct ib_udata *udata,
drivers/infiniband/sw/rxe/rxe_srq.c +3 -2
index 41b0d1e11baf..52c5593741ec 100644
@@ -78,6 +78,7 @@ int rxe_srq_from_init(struct rxe_dev *rxe, struct rxe_srq *srq,
        int err;
        int srq_wqe_size;
        struct rxe_queue *q;
+       enum queue_type type;
 
        srq->ibsrq.event_handler        = init->event_handler;
        srq->ibsrq.srq_context          = init->srq_context;
@@ -91,8 +92,9 @@ int rxe_srq_from_init(struct rxe_dev *rxe, struct rxe_srq *srq,
        spin_lock_init(&srq->rq.producer_lock);
        spin_lock_init(&srq->rq.consumer_lock);
 
+       type = uresp ? QUEUE_TYPE_FROM_USER : QUEUE_TYPE_KERNEL;
        q = rxe_queue_init(rxe, &srq->rq.max_wr,
-                          srq_wqe_size);
+                       srq_wqe_size, type);
        if (!q) {
                pr_warn("unable to allocate queue for srq\n");
                return -ENOMEM;
--

2.30.2

#3 lore raw

Re: [PATCH for-next v3 1/3] RDMA/rxe: Add a type flag to rxe_queue structs

Quoted 26 lines · Zhu Yanjun
Quoted 25 lines · Bob Pearson

To create optimal code only want to use smp_load_acquire() and smp_store_release() for user indices in rxe_queue APIs since kernel indices are protected by locks which also act as memory barriers. By adding a type to the queues we can determine which indices need to be protected.

---
  drivers/infiniband/sw/rxe/rxe_cq.c    |  4 +++-
  drivers/infiniband/sw/rxe/rxe_qp.c    | 12 ++++++++----
  drivers/infiniband/sw/rxe/rxe_queue.c |  8 ++++----
  drivers/infiniband/sw/rxe/rxe_queue.h | 13 ++++++++++---
  drivers/infiniband/sw/rxe/rxe_srq.c   |  4 +++-
  5 files changed, 28 insertions(+), 13 deletions(-)
drivers/infiniband/sw/rxe/rxe_cq.c +1 -0
index b315ebf041ac..1d4d8a31bc12 100644
@@ -59,9 +59,11 @@ int rxe_cq_from_init(struct rxe_dev *rxe, struct rxe_cq *cq, int cqe,
                      struct rxe_create_cq_resp __user *uresp)
  {
         int err;
+       enum queue_type type;
 
Quoted 4 lines

drivers/infiniband/sw/rxe/rxe_cq.c int rxe_cq_from_init(struct rxe_dev *rxe, struct rxe_cq *cq, int cqe, · line 64

Quoted 1 line
drivers/infiniband/sw/rxe/rxe_cq.c +1 -0
+       type = uresp ? QUEUE_TYPE_TO_USER : QUEUE_TYPE_KERNEL;

When is QUEUE_TYPE_TO_USER used? and when is QUEUE_TYPE_FROM_USER is used?

Zhu Yanjun

QUEUE_TYPE_FROM_USER is used for user space send, recv and shared receive work queues. QUEUE_TYPE_TO_USER is used for user space completion queues. QUEUE_TYPE_KERNEL is used for kernel verbs consumers for any queue.

Bob

Quoted 1 line
#4 lore raw

PATCH v3 2/3 RDMA/rxe: Protect user space index loads/stores

In reply to Bob Pearson

Modify the queue APIs to protect all user space index loads with smp_load_acquire() and all user space index stores with smp_store_release(). Base this on the types of the queues which can be one of ..KERNEL, ..FROM_USER, ..TO_USER. Kernel space indices are protected by locks which also provide memory barriers.

---
 drivers/infiniband/sw/rxe/rxe_queue.h | 168 ++++++++++++++++++--------
 1 file changed, 117 insertions(+), 51 deletions(-)
drivers/infiniband/sw/rxe/rxe_queue.h +117 -51
index 4512745419f8..6e705e09d357 100644
@@ -66,12 +66,22 @@ static inline int queue_empty(struct rxe_queue *q)
 	u32 prod;
 	u32 cons;
 
-	/* make sure all changes to queue complete before
-	 * testing queue empty
-	 */
-	prod = smp_load_acquire(&q->buf->producer_index);
-	/* same */
-	cons = smp_load_acquire(&q->buf->consumer_index);
+	switch (q->type) {
+	case QUEUE_TYPE_FROM_USER:
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+		cons = q->buf->consumer_index;
+		break;
+	case QUEUE_TYPE_TO_USER:
+		prod = q->buf->producer_index;
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+		break;
+	case QUEUE_TYPE_KERNEL:
+		prod = q->buf->producer_index;
+		cons = q->buf->consumer_index;
+		break;
+	}
 
 	return ((prod - cons) & q->index_mask) == 0;
 }
@@ -81,95 +91,151 @@ static inline int queue_full(struct rxe_queue *q)
 	u32 prod;
 	u32 cons;
 
-	/* make sure all changes to queue complete before
-	 * testing queue full
-	 */
-	prod = smp_load_acquire(&q->buf->producer_index);
-	/* same */
-	cons = smp_load_acquire(&q->buf->consumer_index);
+	switch (q->type) {
+	case QUEUE_TYPE_FROM_USER:
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+		cons = q->buf->consumer_index;
+		break;
+	case QUEUE_TYPE_TO_USER:
+		prod = q->buf->producer_index;
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+		break;
+	case QUEUE_TYPE_KERNEL:
+		prod = q->buf->producer_index;
+		cons = q->buf->consumer_index;
+		break;
+	}
 
 	return ((prod + 1 - cons) & q->index_mask) == 0;
 }
 
-static inline void advance_producer(struct rxe_queue *q)
+static inline unsigned int queue_count(const struct rxe_queue *q)
 {
 	u32 prod;
+	u32 cons;
 
-	prod = (q->buf->producer_index + 1) & q->index_mask;
+	switch (q->type) {
+	case QUEUE_TYPE_FROM_USER:
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+		cons = q->buf->consumer_index;
+		break;
+	case QUEUE_TYPE_TO_USER:
+		prod = q->buf->producer_index;
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+		break;
+	case QUEUE_TYPE_KERNEL:
+		prod = q->buf->producer_index;
+		cons = q->buf->consumer_index;
+		break;
+	}
+
+	return (prod - cons) & q->index_mask;
+}
+
+static inline void advance_producer(struct rxe_queue *q)
+{
+	u32 prod;
 
-	/* make sure all changes to queue complete before
-	 * changing producer index
-	 */
-	smp_store_release(&q->buf->producer_index, prod);
+	if (q->type == QUEUE_TYPE_FROM_USER) {
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+		prod = (prod + 1) & q->index_mask;
+		/* same */
+		smp_store_release(&q->buf->producer_index, prod);
+	} else {
+		prod = q->buf->producer_index;
+		q->buf->producer_index = (prod + 1) & q->index_mask;
+	}
 }
 
 static inline void advance_consumer(struct rxe_queue *q)
 {
 	u32 cons;
 
-	cons = (q->buf->consumer_index + 1) & q->index_mask;
-
-	/* make sure all changes to queue complete before
-	 * changing consumer index
-	 */
-	smp_store_release(&q->buf->consumer_index, cons);
+	if (q->type == QUEUE_TYPE_TO_USER) {
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+		cons = (cons + 1) & q->index_mask;
+		/* same */
+		smp_store_release(&q->buf->consumer_index, cons);
+	} else {
+		cons = q->buf->consumer_index;
+		q->buf->consumer_index = (cons + 1) & q->index_mask;
+	}
 }
 
 static inline void *producer_addr(struct rxe_queue *q)
 {
-	return q->buf->data + ((q->buf->producer_index & q->index_mask)
-				<< q->log2_elem_size);
+	u32 prod;
+
+	if (q->type == QUEUE_TYPE_FROM_USER)
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+	else
+		prod = q->buf->producer_index;
+
+	return q->buf->data + ((prod & q->index_mask) << q->log2_elem_size);
 }
 
 static inline void *consumer_addr(struct rxe_queue *q)
 {
-	return q->buf->data + ((q->buf->consumer_index & q->index_mask)
-				<< q->log2_elem_size);
+	u32 cons;
+
+	if (q->type == QUEUE_TYPE_TO_USER)
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+	else
+		cons = q->buf->consumer_index;
+
+	return q->buf->data + ((cons & q->index_mask) << q->log2_elem_size);
 }
 
 static inline unsigned int producer_index(struct rxe_queue *q)
 {
-	u32 index;
+	u32 prod;
+
+	if (q->type == QUEUE_TYPE_FROM_USER)
+		/* protect user space index */
+		prod = smp_load_acquire(&q->buf->producer_index);
+	else
+		prod = q->buf->producer_index;
 
-	/* make sure all changes to queue
-	 * complete before getting producer index
-	 */
-	index = smp_load_acquire(&q->buf->producer_index);
-	index &= q->index_mask;
+	prod &= q->index_mask;
 
-	return index;
+	return prod;
 }
 
 static inline unsigned int consumer_index(struct rxe_queue *q)
 {
-	u32 index;
+	u32 cons;
 
-	/* make sure all changes to queue
-	 * complete before getting consumer index
-	 */
-	index = smp_load_acquire(&q->buf->consumer_index);
-	index &= q->index_mask;
+	if (q->type == QUEUE_TYPE_TO_USER)
+		/* protect user space index */
+		cons = smp_load_acquire(&q->buf->consumer_index);
+	else
+		cons = q->buf->consumer_index;
 
-	return index;
+	cons &= q->index_mask;
+
+	return cons;
 }
 
-static inline void *addr_from_index(struct rxe_queue *q, unsigned int index)
+static inline void *addr_from_index(struct rxe_queue *q,
+				unsigned int index)
 {
 	return q->buf->data + ((index & q->index_mask)
 				<< q->buf->log2_elem_size);
 }
 
 static inline unsigned int index_from_addr(const struct rxe_queue *q,
-					   const void *addr)
+				const void *addr)
 {
 	return (((u8 *)addr - q->buf->data) >> q->log2_elem_size)
-		& q->index_mask;
-}
-
-static inline unsigned int queue_count(const struct rxe_queue *q)
-{
-	return (q->buf->producer_index - q->buf->consumer_index)
-		& q->index_mask;
+				& q->index_mask;
 }
 
 static inline void *queue_head(struct rxe_queue *q)
Signature · 1 line
2.30.2
#5 lore raw

PATCH v3 3/3 RDMA/rxe: Protext kernel index from user space

In reply to Bob Pearson

In order to prevent user space from modifying the index that belongs to the kernel for shared queues let the kernel use a local copy of the index and copy any new values of that index to the shared rxe_queue_bus struct.

This adds more switch statements which decrease the peformance of the queue API. Move the type into the parameter list for these functions so that the compiler can optimize out the switch statements when the explicit type is known. Modify all the calls in the driver on performance paths to pass in the explicit queue type.

---
 drivers/infiniband/sw/rxe/rxe_comp.c  |  31 ++++--
 drivers/infiniband/sw/rxe/rxe_cq.c    |  28 ++++-
 drivers/infiniband/sw/rxe/rxe_qp.c    |  10 +-
 drivers/infiniband/sw/rxe/rxe_queue.c |  13 +--
 drivers/infiniband/sw/rxe/rxe_queue.h | 145 +++++++++++++++++++-------
 drivers/infiniband/sw/rxe/rxe_req.c   |  46 ++++++--
 drivers/infiniband/sw/rxe/rxe_resp.c  |  44 ++++++--
 drivers/infiniband/sw/rxe/rxe_srq.c   |   1 +
 drivers/infiniband/sw/rxe/rxe_verbs.c |  80 +++++++++++---
 drivers/infiniband/sw/rxe/rxe_verbs.h |   5 +-
 10 files changed, 305 insertions(+), 98 deletions(-)
drivers/infiniband/sw/rxe/rxe_comp.c +21 -10
index 207aa7ef52c4..d4ceb81a96df 100644
@@ -142,7 +142,10 @@ static inline enum comp_state get_wqe(struct rxe_qp *qp,
 	/* we come here whether or not we found a response packet to see if
 	 * there are any posted WQEs
 	 */
-	wqe = queue_head(qp->sq.queue);
+	if (qp->is_user)
+		wqe = queue_head(qp->sq.queue, QUEUE_TYPE_FROM_USER);
+	else
+		wqe = queue_head(qp->sq.queue, QUEUE_TYPE_KERNEL);
 	*wqe_p = wqe;
 
 	/* no WQE or requester has not started it yet */
@@ -415,16 +418,23 @@ static void do_complete(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 {
 	struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 	struct rxe_cqe cqe;
+	bool post;
+
+	/* do we need to post a completion */
+	post = ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
+			(wqe->wr.send_flags & IB_SEND_SIGNALED) ||
+			wqe->status != IB_WC_SUCCESS);
 
-	if ((qp->sq_sig_type == IB_SIGNAL_ALL_WR) ||
-	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
-	    wqe->status != IB_WC_SUCCESS) {
+	if (post)
 		make_send_cqe(qp, wqe, &cqe);
-		advance_consumer(qp->sq.queue);
+
+	if (qp->is_user)
+		advance_consumer(qp->sq.queue, QUEUE_TYPE_FROM_USER);
+	else
+		advance_consumer(qp->sq.queue, QUEUE_TYPE_KERNEL);
+
+	if (post)
 		rxe_cq_post(qp->scq, &cqe, 0);
-	} else {
-		advance_consumer(qp->sq.queue);
-	}
 
 	if (wqe->wr.opcode == IB_WR_SEND ||
 	    wqe->wr.opcode == IB_WR_SEND_WITH_IMM ||
@@ -512,6 +522,7 @@ static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
 {
 	struct sk_buff *skb;
 	struct rxe_send_wqe *wqe;
+	struct rxe_queue *q = qp->sq.queue;
 
 	while ((skb = skb_dequeue(&qp->resp_pkts))) {
 		rxe_drop_ref(qp);
@@ -519,12 +530,12 @@ static void rxe_drain_resp_pkts(struct rxe_qp *qp, bool notify)
 		ib_device_put(qp->ibqp.device);
 	}
 
-	while ((wqe = queue_head(qp->sq.queue))) {
+	while ((wqe = queue_head(q, q->type))) {
 		if (notify) {
 			wqe->status = IB_WC_WR_FLUSH_ERR;
 			do_complete(qp, wqe);
 		} else {
-			advance_consumer(qp->sq.queue);
+			advance_consumer(q, q->type);
 		}
 	}
 }
drivers/infiniband/sw/rxe/rxe_cq.c +24 -4
index 1d4d8a31bc12..aef288f164fd 100644
@@ -25,7 +25,11 @@ int rxe_cq_chk_attr(struct rxe_dev *rxe, struct rxe_cq *cq,
 	}
 
 	if (cq) {
-		count = queue_count(cq->queue);
+		if (cq->is_user)
+			count = queue_count(cq->queue, QUEUE_TYPE_TO_USER);
+		else
+			count = queue_count(cq->queue, QUEUE_TYPE_KERNEL);
+
 		if (cqe < count) {
 			pr_warn("cqe(%d) < current # elements in queue (%d)",
 				cqe, count);
@@ -108,10 +112,17 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)
 {
 	struct ib_event ev;
 	unsigned long flags;
+	int full;
+	void *addr;
 
 	spin_lock_irqsave(&cq->cq_lock, flags);
 
-	if (unlikely(queue_full(cq->queue))) {
+	if (cq->is_user)
+		full = queue_full(cq->queue, QUEUE_TYPE_TO_USER);
+	else
+		full = queue_full(cq->queue, QUEUE_TYPE_KERNEL);
+
+	if (unlikely(full)) {
 		spin_unlock_irqrestore(&cq->cq_lock, flags);
 		if (cq->ibcq.event_handler) {
 			ev.device = cq->ibcq.device;
@@ -123,9 +134,18 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)
 		return -EBUSY;
 	}
 
-	memcpy(producer_addr(cq->queue), cqe, sizeof(*cqe));
+	if (cq->is_user)
+		addr = producer_addr(cq->queue, QUEUE_TYPE_TO_USER);
+	else
+		addr = producer_addr(cq->queue, QUEUE_TYPE_KERNEL);
+
+	memcpy(addr, cqe, sizeof(*cqe));
+
+	if (cq->is_user)
+		advance_producer(cq->queue, QUEUE_TYPE_TO_USER);
+	else
+		advance_producer(cq->queue, QUEUE_TYPE_KERNEL);
 
-	advance_producer(cq->queue);
 	spin_unlock_irqrestore(&cq->cq_lock, flags);
 
 	if ((cq->notify == IB_CQ_NEXT_COMP) ||
drivers/infiniband/sw/rxe/rxe_qp.c +9 -1
index 9bd6bf8f9bd9..a9256862464b 100644
@@ -248,7 +248,13 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
 		return err;
 	}
 
-	qp->req.wqe_index	= producer_index(qp->sq.queue);
+	if (qp->is_user)
+		qp->req.wqe_index = producer_index(qp->sq.queue,
+						QUEUE_TYPE_FROM_USER);
+	else
+		qp->req.wqe_index = producer_index(qp->sq.queue,
+						QUEUE_TYPE_KERNEL);
+
 	qp->req.state		= QP_STATE_RESET;
 	qp->req.opcode		= -1;
 	qp->comp.opcode		= -1;
@@ -306,6 +312,8 @@ static int rxe_qp_init_resp(struct rxe_dev *rxe, struct rxe_qp *qp,
 	spin_lock_init(&qp->rq.producer_lock);
 	spin_lock_init(&qp->rq.consumer_lock);
 
+	qp->rq.is_user = qp->is_user;
+
 	skb_queue_head_init(&qp->resp_pkts);
 
 	rxe_init_task(rxe, &qp->resp.task, qp,
drivers/infiniband/sw/rxe/rxe_queue.c +7 -6
index 8f844d0b9e77..85b812586ed4 100644
@@ -111,14 +111,15 @@ struct rxe_queue *rxe_queue_init(struct rxe_dev *rxe, int *num_elem,
 static int resize_finish(struct rxe_queue *q, struct rxe_queue *new_q,
 			 unsigned int num_elem)
 {
-	if (!queue_empty(q) && (num_elem < queue_count(q)))
+	if (!queue_empty(q, q->type) && (num_elem < queue_count(q, q->type)))
 		return -EINVAL;
 
-	while (!queue_empty(q)) {
-		memcpy(producer_addr(new_q), consumer_addr(q),
-		       new_q->elem_size);
-		advance_producer(new_q);
-		advance_consumer(q);
+	while (!queue_empty(q, q->type)) {
+		memcpy(producer_addr(new_q, new_q->type),
+					consumer_addr(q, q->type),
+					new_q->elem_size);
+		advance_producer(new_q, new_q->type);
+		advance_consumer(q, q->type);
 	}
 
 	swap(*q, *new_q);
drivers/infiniband/sw/rxe/rxe_queue.h +107 -38

275 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_req.c +35 -11

95 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_resp.c +34 -10

93 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_srq.c +1 -0
index 52c5593741ec..610c98d24b5c 100644
@@ -86,6 +86,7 @@ int rxe_srq_from_init(struct rxe_dev *rxe, struct rxe_srq *srq,
 	srq->srq_num		= srq->pelem.index;
 	srq->rq.max_wr		= init->attr.max_wr;
 	srq->rq.max_sge		= init->attr.max_sge;
+	srq->rq.is_user		= srq->is_user;
 
 	srq_wqe_size		= rcv_wqe_size(srq->rq.max_sge);
 
drivers/infiniband/sw/rxe/rxe_verbs.c +63 -17

157 lines not on this page · raw message

drivers/infiniband/sw/rxe/rxe_verbs.h +4 -1
index cf8cae64f7df..959a3260fcab 100644
@@ -77,6 +77,7 @@ enum wqe_state {
 };
 
 struct rxe_sq {
+	bool			is_user;
 	int			max_wr;
 	int			max_sge;
 	int			max_inline;
@@ -85,6 +86,7 @@ struct rxe_sq {
 };
 
 struct rxe_rq {
+	bool			is_user;
 	int			max_wr;
 	int			max_sge;
 	spinlock_t		producer_lock; /* guard queue producer */
@@ -98,6 +100,7 @@ struct rxe_srq {
 	struct rxe_pd		*pd;
 	struct rxe_rq		rq;
 	u32			srq_num;
+	bool			is_user;
 
 	int			limit;
 	int			error;
@@ -212,7 +215,7 @@ struct rxe_qp {
 	struct ib_qp_attr	attr;
 	unsigned int		valid;
 	unsigned int		mtu;
-	int			is_user;
+	bool			is_user;
 
 	struct rxe_pd		*pd;
 	struct rxe_srq		*srq;
Signature · 1 line
2.30.2
maintainer
#6 lore raw
In reply to Bob Pearson
Quoted 7 lines · Bob Pearson

These patches optimize the memory ordering in rxe_queue.h so that user space and not kernel space indices are protected for loads with smp_load_acquire() and stores with smp_store_release(). The original implementation of this did not apply to all index references which has recently caused test case errors traced to stale memory loads. These patches fix those errors and also protect kernel indices from malicious modification by user space.

I didn't read it carefully, but I think this captures the basic solution

Jason

maintainer
#7 lore raw
Quoted 10 lines · Jason Gunthorpe
Quoted 7 lines · Bob Pearson

These patches optimize the memory ordering in rxe_queue.h so that user space and not kernel space indices are protected for loads with smp_load_acquire() and stores with smp_store_release(). The original implementation of this did not apply to all index references which has recently caused test case errors traced to stale memory loads. These patches fix those errors and also protect kernel indices from malicious modification by user space.

Quoted 2 lines

I didn't read it carefully, but I think this captures the basic solution

I made tests with rdma-core.

With or without changes of kernel-headers/rdma/rdma_user_rxe.h, the previously mentioned 17 errors disappear.

So this patch series takes effect.

Zhu Yanjun

Quoted 2 lines

Jason

maintainer Applied
#8 lore raw
In reply to Bob Pearson
Quoted 11 lines · Bob Pearson

These patches optimize the memory ordering in rxe_queue.h so that user space and not kernel space indices are protected for loads with smp_load_acquire() and stores with smp_store_release(). The original implementation of this did not apply to all index references which has recently caused test case errors traced to stale memory loads. These patches fix those errors and also protect kernel indices from malicious modification by user space.

Fixes d21a1240f516 ("RDMA/rxe: Use acquire/release for memory ordering")
Quoted 13 lines

---
v3:
Protected kernel index in shared queues from modification by user space. Pass queue type to allow compiler to optimize queue methods.
v2:
Add a way to distinguish between user and kernel indices.
v1:
Add missing smp_load_acquire() calls.
---
Bob Pearson (3):
RDMA/rxe: Add a type flag to rxe_queue structs
RDMA/rxe: Protect user space index loads/stores
RDMA/rxe: Protext kernel index from user space

Applied to for-next, thanks

Jason