[PATCH v2] RDMA/rxe: Use acquire/release for memory ordering

[PATCH v2] RDMA/rxe: Use acquire/release for memory ordering

[Bob Pearson]

Changed work and completion queues to use smp_load_acquire() and
smp_store_release() to synchronize between driver and users.
This commit goes with a matching series of commits in the rxe user
space provider.

v2: Addressed same issue for kernel ULPs which use rxe_post_send/recv().

Signed-off-by: Bob Pearson <rpearson@hpe.com>
---
 drivers/infiniband/sw/rxe/rxe_cq.c    |  5 --
 drivers/infiniband/sw/rxe/rxe_queue.h | 94 +++++++++++++++++----------
 drivers/infiniband/sw/rxe/rxe_verbs.c | 11 ----
 include/uapi/rdma/rdma_user_rxe.h     | 21 ++++++
 4 files changed, 81 insertions(+), 50 deletions(-)

diff --git a/drivers/infiniband/sw/rxe/rxe_cq.c b/drivers/infiniband/sw/rxe/rxe_cq.c
index 43394c3f29d4..b315ebf041ac 100644
--- a/drivers/infiniband/sw/rxe/rxe_cq.c
+++ b/drivers/infiniband/sw/rxe/rxe_cq.c
@@ -123,11 +123,6 @@ int rxe_cq_post(struct rxe_cq *cq, struct rxe_cqe *cqe, int solicited)
 
 	memcpy(producer_addr(cq->queue), cqe, sizeof(*cqe));
 
-	/* make sure all changes to the CQ are written before we update the
-	 * producer pointer
-	 */
-	smp_wmb();
-
 	advance_producer(cq->queue);
 	spin_unlock_irqrestore(&cq->cq_lock, flags);
 
diff --git a/drivers/infiniband/sw/rxe/rxe_queue.h b/drivers/infiniband/sw/rxe/rxe_queue.h
index 7d434a6837a7..2902ca7b288c 100644
--- a/drivers/infiniband/sw/rxe/rxe_queue.h
+++ b/drivers/infiniband/sw/rxe/rxe_queue.h
@@ -7,9 +7,11 @@
 #ifndef RXE_QUEUE_H
 #define RXE_QUEUE_H
 
+/* for definition of shared struct rxe_queue_buf */
+#include <uapi/rdma/rdma_user_rxe.h>
+
 /* implements a simple circular buffer that can optionally be
  * shared between user space and the kernel and can be resized
-
  * the requested element size is rounded up to a power of 2
  * and the number of elements in the buffer is also rounded
  * up to a power of 2. Since the queue is empty when the
@@ -17,28 +19,6 @@
  * of the queue is one less than the number of element slots
  */
 
-/* this data structure is shared between user space and kernel
- * space for those cases where the queue is shared. It contains
- * the producer and consumer indices. Is also contains a copy
- * of the queue size parameters for user space to use but the
- * kernel must use the parameters in the rxe_queue struct
- * this MUST MATCH the corresponding librxe struct
- * for performance reasons arrange to have producer and consumer
- * pointers in separate cache lines
- * the kernel should always mask the indices to avoid accessing
- * memory outside of the data area
- */
-struct rxe_queue_buf {
-	__u32			log2_elem_size;
-	__u32			index_mask;
-	__u32			pad_1[30];
-	__u32			producer_index;
-	__u32			pad_2[31];
-	__u32			consumer_index;
-	__u32			pad_3[31];
-	__u8			data[];
-};
-
 struct rxe_queue {
 	struct rxe_dev		*rxe;
 	struct rxe_queue_buf	*buf;
@@ -46,7 +26,7 @@ struct rxe_queue {
 	size_t			buf_size;
 	size_t			elem_size;
 	unsigned int		log2_elem_size;
-	unsigned int		index_mask;
+	u32			index_mask;
 };
 
 int do_mmap_info(struct rxe_dev *rxe, struct mminfo __user *outbuf,
@@ -76,26 +56,56 @@ static inline int next_index(struct rxe_queue *q, int index)
 
 static inline int queue_empty(struct rxe_queue *q)
 {
-	return ((q->buf->producer_index - q->buf->consumer_index)
-			& q->index_mask) == 0;
+	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);
+
+	return ((prod - cons) & q->index_mask) == 0;
 }
 
 static inline int queue_full(struct rxe_queue *q)
 {
-	return ((q->buf->producer_index + 1 - q->buf->consumer_index)
-			& q->index_mask) == 0;
+	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);
+
+	return ((prod + 1 - cons) & q->index_mask) == 0;
 }
 
 static inline void advance_producer(struct rxe_queue *q)
 {
-	q->buf->producer_index = (q->buf->producer_index + 1)
-			& q->index_mask;
+	u32 prod;
+
+	prod = (q->buf->producer_index + 1) & q->index_mask;
+
+	/* make sure all changes to queue complete before
+	 * changing producer index
+	 */
+	smp_store_release(&q->buf->producer_index, prod);
 }
 
 static inline void advance_consumer(struct rxe_queue *q)
 {
-	q->buf->consumer_index = (q->buf->consumer_index + 1)
-			& q->index_mask;
+	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);
 }
 
 static inline void *producer_addr(struct rxe_queue *q)
@@ -112,12 +122,28 @@ static inline void *consumer_addr(struct rxe_queue *q)
 
 static inline unsigned int producer_index(struct rxe_queue *q)
 {
-	return q->buf->producer_index;
+	u32 index;
+
+	/* make sure all changes to queue
+	 * complete before getting producer index
+	 */
+	index = smp_load_acquire(&q->buf->producer_index);
+	index &= q->index_mask;
+
+	return index;
 }
 
 static inline unsigned int consumer_index(struct rxe_queue *q)
 {
-	return q->buf->consumer_index;
+	u32 index;
+
+	/* make sure all changes to queue
+	 * complete before getting consumer index
+	 */
+	index = smp_load_acquire(&q->buf->consumer_index);
+	index &= q->index_mask;
+
+	return index;
 }
 
 static inline void *addr_from_index(struct rxe_queue *q, unsigned int index)
diff --git a/drivers/infiniband/sw/rxe/rxe_verbs.c b/drivers/infiniband/sw/rxe/rxe_verbs.c
index 2fbea2b2d72a..a031514e2f41 100644
--- a/drivers/infiniband/sw/rxe/rxe_verbs.c
+++ b/drivers/infiniband/sw/rxe/rxe_verbs.c
@@ -244,11 +244,6 @@ static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
 	recv_wqe->dma.cur_sge		= 0;
 	recv_wqe->dma.sge_offset	= 0;
 
-	/* make sure all changes to the work queue are written before we
-	 * update the producer pointer
-	 */
-	smp_wmb();
-
 	advance_producer(rq->queue);
 	return 0;
 
@@ -633,12 +628,6 @@ static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 	if (unlikely(err))
 		goto err1;
 
-	/*
-	 * make sure all changes to the work queue are
-	 * written before we update the producer pointer
-	 */
-	smp_wmb();
-
 	advance_producer(sq->queue);
 	spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
 
diff --git a/include/uapi/rdma/rdma_user_rxe.h b/include/uapi/rdma/rdma_user_rxe.h
index e591d8c1f3cf..068433e2229d 100644
--- a/include/uapi/rdma/rdma_user_rxe.h
+++ b/include/uapi/rdma/rdma_user_rxe.h
@@ -181,4 +181,25 @@ struct rxe_modify_srq_cmd {
 	__aligned_u64 mmap_info_addr;
 };
 
+/* This data structure is stored at the base of work and
+ * completion queues shared between user space and kernel space.
+ * It contains the producer and consumer indices. Is also
+ * contains a copy of the queue size parameters for user space
+ * to use but the kernel must use the parameters in the
+ * rxe_queue struct. For performance reasons arrange to have
+ * producer and consumer indices in separate cache lines
+ * the kernel should always mask the indices to avoid accessing
+ * memory outside of the data area
+ */
+struct rxe_queue_buf {
+	__u32			log2_elem_size;
+	__u32			index_mask;
+	__u32			pad_1[30];
+	__u32			producer_index;
+	__u32			pad_2[31];
+	__u32			consumer_index;
+	__u32			pad_3[31];
+	__u8			data[];
+};
+
 #endif /* RDMA_USER_RXE_H */
-- 
2.27.0

Re: [PATCH v2] RDMA/rxe: Use acquire/release for memory ordering

[Jason Gunthorpe]

On Thu, Dec 10, 2020 at 11:42:59AM -0600, Bob Pearson wrote:

> @@ -76,26 +56,56 @@ static inline int next_index(struct rxe_queue *q, int index)
>  
>  static inline int queue_empty(struct rxe_queue *q)
>  {
> -	return ((q->buf->producer_index - q->buf->consumer_index)
> -			& q->index_mask) == 0;
> +	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);

This is not so sensible. The one written by the kernel should be just
a normal load and there must be some lock held here to protect that

The one written by user space should be just READ_ONCE

acquire only has meaning if you go on to read additional data based on
the result of the acquire - then acquire ensures the additional reads
can't cross writes that occured before the matching release.

Jason

Re: [PATCH v2] RDMA/rxe: Use acquire/release for memory ordering

[Jason Gunthorpe]

On Thu, Dec 10, 2020 at 11:42:59AM -0600, Bob Pearson wrote:
> Changed work and completion queues to use smp_load_acquire() and
> smp_store_release() to synchronize between driver and users.
> This commit goes with a matching series of commits in the rxe user
> space provider.
> 
> v2: Addressed same issue for kernel ULPs which use rxe_post_send/recv().
> 
> Signed-off-by: Bob Pearson <rpearson@hpe.com>
> ---
>  drivers/infiniband/sw/rxe/rxe_cq.c    |  5 --
>  drivers/infiniband/sw/rxe/rxe_queue.h | 94 +++++++++++++++++----------
>  drivers/infiniband/sw/rxe/rxe_verbs.c | 11 ----
>  include/uapi/rdma/rdma_user_rxe.h     | 21 ++++++
>  4 files changed, 81 insertions(+), 50 deletions(-)

This really is a lot better than what was here, the extra barriers on
empty/full can be fine tuned later

So applied to for-next

Thanks,
Jason