{net,IB}/mlx5: Refactor page fault handling

* Update page fault event according to last specification.
* Separate code path for page fault EQ, completion EQ and async EQ.
* Move page fault handling work queue from mlx5_ib static variable
  into mlx5_core page fault EQ.
* Allocate memory to store ODP event dynamically as the
  events arrive, since in atomic context - use mempool.
* Make mlx5_ib page fault handler run in process context.
Signed-off-by: Artemy Kovalyov <artemyko@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/infiniband/hw/mlx5/main.c

@@ -3319,6 +3319,9 @@ static struct mlx5_interface mlx5_ib_interface = {
 	.add            = mlx5_ib_add,
 	.remove         = mlx5_ib_remove,
 	.event          = mlx5_ib_event,
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	.pfault		= mlx5_ib_pfault,
+#endif
 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
 };
 
@@ -3329,25 +3332,14 @@ static int __init mlx5_ib_init(void)
 	if (deprecated_prof_sel != 2)
 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
 
-	err = mlx5_ib_odp_init();
-	if (err)
-		return err;
-
 	err = mlx5_register_interface(&mlx5_ib_interface);
-	if (err)
-		goto clean_odp;
-
-	return err;
 
-clean_odp:
-	mlx5_ib_odp_cleanup();
 	return err;
 }
 
 static void __exit mlx5_ib_cleanup(void)
 {
 	mlx5_unregister_interface(&mlx5_ib_interface);
-	mlx5_ib_odp_cleanup();
 }
 
 module_init(mlx5_ib_init);

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -277,29 +277,6 @@ struct mlx5_ib_rwq_ind_table {
 	u32			rqtn;
 };
 
-/*
- * Connect-IB can trigger up to four concurrent pagefaults
- * per-QP.
- */
-enum mlx5_ib_pagefault_context {
-	MLX5_IB_PAGEFAULT_RESPONDER_READ,
-	MLX5_IB_PAGEFAULT_REQUESTOR_READ,
-	MLX5_IB_PAGEFAULT_RESPONDER_WRITE,
-	MLX5_IB_PAGEFAULT_REQUESTOR_WRITE,
-	MLX5_IB_PAGEFAULT_CONTEXTS
-};
-
-static inline enum mlx5_ib_pagefault_context
-	mlx5_ib_get_pagefault_context(struct mlx5_pagefault *pagefault)
-{
-	return pagefault->flags & (MLX5_PFAULT_REQUESTOR | MLX5_PFAULT_WRITE);
-}
-
-struct mlx5_ib_pfault {
-	struct work_struct	work;
-	struct mlx5_pagefault	mpfault;
-};
-
 struct mlx5_ib_ubuffer {
 	struct ib_umem	       *umem;
 	int			buf_size;
@@ -385,20 +362,6 @@ struct mlx5_ib_qp {
 	/* Store signature errors */
 	bool			signature_en;
 
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-	/*
-	 * A flag that is true for QP's that are in a state that doesn't
-	 * allow page faults, and shouldn't schedule any more faults.
-	 */
-	int                     disable_page_faults;
-	/*
-	 * The disable_page_faults_lock protects a QP's disable_page_faults
-	 * field, allowing for a thread to atomically check whether the QP
-	 * allows page faults, and if so schedule a page fault.
-	 */
-	spinlock_t              disable_page_faults_lock;
-	struct mlx5_ib_pfault	pagefaults[MLX5_IB_PAGEFAULT_CONTEXTS];
-#endif
 	struct list_head	qps_list;
 	struct list_head	cq_recv_list;
 	struct list_head	cq_send_list;
@@ -869,18 +832,13 @@ struct ib_rwq_ind_table *mlx5_ib_create_rwq_ind_table(struct ib_device *device,
 int mlx5_ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *wq_ind_table);
 
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-extern struct workqueue_struct *mlx5_ib_page_fault_wq;
-
 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev);
-void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp,
-			       struct mlx5_ib_pfault *pfault);
-void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp);
+void mlx5_ib_pfault(struct mlx5_core_dev *mdev, void *context,
+		    struct mlx5_pagefault *pfault);
 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev);
 void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev);
 int __init mlx5_ib_odp_init(void);
 void mlx5_ib_odp_cleanup(void);
-void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp);
-void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp);
 void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start,
 			      unsigned long end);
 #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
@@ -889,13 +847,10 @@ static inline void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
 	return;
 }
 
-static inline void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp)		{}
 static inline int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) { return 0; }
 static inline void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev)	{}
 static inline int mlx5_ib_odp_init(void) { return 0; }
 static inline void mlx5_ib_odp_cleanup(void)				{}
-static inline void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp) {}
-static inline void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp)  {}
 
 #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 

drivers/infiniband/hw/mlx5/odp.c

@@ -41,8 +41,6 @@
  * a pagefault. */
 #define MMU_NOTIFIER_TIMEOUT 1000
 
-struct workqueue_struct *mlx5_ib_page_fault_wq;
-
 void mlx5_ib_invalidate_range(struct ib_umem *umem, unsigned long start,
 			      unsigned long end)
 {
@@ -162,38 +160,38 @@ static struct mlx5_ib_mr *mlx5_ib_odp_find_mr_lkey(struct mlx5_ib_dev *dev,
 	return container_of(mmkey, struct mlx5_ib_mr, mmkey);
 }
 
-static void mlx5_ib_page_fault_resume(struct mlx5_ib_qp *qp,
-				      struct mlx5_ib_pfault *pfault,
+static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
+				      struct mlx5_pagefault *pfault,
 				      int error)
 {
-	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
-	u32 qpn = qp->trans_qp.base.mqp.qpn;
+	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
+		     pfault->wqe.wq_num : pfault->token;
 	int ret = mlx5_core_page_fault_resume(dev->mdev,
-					      qpn,
-					      pfault->mpfault.flags,
+					      pfault->token,
+					      wq_num,
+					      pfault->type,
 					      error);
 	if (ret)
-		pr_err("Failed to resolve the page fault on QP 0x%x\n", qpn);
+		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x\n",
+			    wq_num);
 }
 
 /*
- * Handle a single data segment in a page-fault WQE.
+ * Handle a single data segment in a page-fault WQE or RDMA region.
  *
- * Returns number of pages retrieved on success. The caller will continue to
+ * Returns number of pages retrieved on success. The caller may continue to
  * the next data segment.
  * Can return the following error codes:
  * -EAGAIN to designate a temporary error. The caller will abort handling the
  *  page fault and resolve it.
  * -EFAULT when there's an error mapping the requested pages. The caller will
- *  abort the page fault handling and possibly move the QP to an error state.
- * On other errors the QP should also be closed with an error.
+ *  abort the page fault handling.
  */
-static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
-					 struct mlx5_ib_pfault *pfault,
+static int pagefault_single_data_segment(struct mlx5_ib_dev *mib_dev,
 					 u32 key, u64 io_virt, size_t bcnt,
+					 u32 *bytes_committed,
 					 u32 *bytes_mapped)
 {
-	struct mlx5_ib_dev *mib_dev = to_mdev(qp->ibqp.pd->device);
 	int srcu_key;
 	unsigned int current_seq;
 	u64 start_idx;
@@ -219,12 +217,7 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
 			 key);
 		if (bytes_mapped)
 			*bytes_mapped +=
-				(bcnt - pfault->mpfault.bytes_committed);
-		goto srcu_unlock;
-	}
-	if (mr->ibmr.pd != qp->ibqp.pd) {
-		pr_err("Page-fault with different PDs for QP and MR.\n");
-		ret = -EFAULT;
+				(bcnt - *bytes_committed);
 		goto srcu_unlock;
 	}
 
@@ -240,8 +233,8 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
 	 * in all iterations (in iteration 2 and above,
 	 * bytes_committed == 0).
 	 */
-	io_virt += pfault->mpfault.bytes_committed;
-	bcnt -= pfault->mpfault.bytes_committed;
+	io_virt += *bytes_committed;
+	bcnt -= *bytes_committed;
 
 	start_idx = (io_virt - (mr->mmkey.iova & PAGE_MASK)) >> PAGE_SHIFT;
 
@@ -300,7 +293,7 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
 		}
 	}
 	srcu_read_unlock(&mib_dev->mr_srcu, srcu_key);
-	pfault->mpfault.bytes_committed = 0;
+	*bytes_committed = 0;
 	return ret ? ret : npages;
 }
 
@@ -322,8 +315,9 @@ static int pagefault_single_data_segment(struct mlx5_ib_qp *qp,
  * Returns the number of pages loaded if positive, zero for an empty WQE, or a
  * negative error code.
  */
-static int pagefault_data_segments(struct mlx5_ib_qp *qp,
-				   struct mlx5_ib_pfault *pfault, void *wqe,
+static int pagefault_data_segments(struct mlx5_ib_dev *dev,
+				   struct mlx5_pagefault *pfault,
+				   struct mlx5_ib_qp *qp, void *wqe,
 				   void *wqe_end, u32 *bytes_mapped,
 				   u32 *total_wqe_bytes, int receive_queue)
 {
@@ -367,22 +361,23 @@ static int pagefault_data_segments(struct mlx5_ib_qp *qp,
 
 		if (!inline_segment && total_wqe_bytes) {
 			*total_wqe_bytes += bcnt - min_t(size_t, bcnt,
-					pfault->mpfault.bytes_committed);
+					pfault->bytes_committed);
 		}
 
 		/* A zero length data segment designates a length of 2GB. */
 		if (bcnt == 0)
 			bcnt = 1U << 31;
 
-		if (inline_segment || bcnt <= pfault->mpfault.bytes_committed) {
-			pfault->mpfault.bytes_committed -=
+		if (inline_segment || bcnt <= pfault->bytes_committed) {
+			pfault->bytes_committed -=
 				min_t(size_t, bcnt,
-				      pfault->mpfault.bytes_committed);
+				      pfault->bytes_committed);
 			continue;
 		}
 
-		ret = pagefault_single_data_segment(qp, pfault, key, io_virt,
-						    bcnt, bytes_mapped);
+		ret = pagefault_single_data_segment(dev, key, io_virt, bcnt,
+						    &pfault->bytes_committed,
+						    bytes_mapped);
 		if (ret < 0)
 			break;
 		npages += ret;
@@ -396,12 +391,11 @@ static int pagefault_data_segments(struct mlx5_ib_qp *qp,
  * scatter-gather list, and set wqe_end to the end of the WQE.
  */
 static int mlx5_ib_mr_initiator_pfault_handler(
-	struct mlx5_ib_qp *qp, struct mlx5_ib_pfault *pfault,
-	void **wqe, void **wqe_end, int wqe_length)
+	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
+	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
 {
-	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
 	struct mlx5_wqe_ctrl_seg *ctrl = *wqe;
-	u16 wqe_index = pfault->mpfault.wqe.wqe_index;
+	u16 wqe_index = pfault->wqe.wqe_index;
 	unsigned ds, opcode;
 #if defined(DEBUG)
 	u32 ctrl_wqe_index, ctrl_qpn;
@@ -502,10 +496,9 @@ static int mlx5_ib_mr_initiator_pfault_handler(
  * scatter-gather list, and set wqe_end to the end of the WQE.
  */
 static int mlx5_ib_mr_responder_pfault_handler(
-	struct mlx5_ib_qp *qp, struct mlx5_ib_pfault *pfault,
-	void **wqe, void **wqe_end, int wqe_length)
+	struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault,
+	struct mlx5_ib_qp *qp, void **wqe, void **wqe_end, int wqe_length)
 {
-	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
 	struct mlx5_ib_wq *wq = &qp->rq;
 	int wqe_size = 1 << wq->wqe_shift;
 
@@ -542,70 +535,83 @@ static int mlx5_ib_mr_responder_pfault_handler(
 	return 0;
 }
 
-static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_qp *qp,
-					  struct mlx5_ib_pfault *pfault)
+static struct mlx5_ib_qp *mlx5_ib_odp_find_qp(struct mlx5_ib_dev *dev,
+					      u32 wq_num)
+{
+	struct mlx5_core_qp *mqp = __mlx5_qp_lookup(dev->mdev, wq_num);
+
+	if (!mqp) {
+		mlx5_ib_err(dev, "QPN 0x%6x not found\n", wq_num);
+		return NULL;
+	}
+
+	return to_mibqp(mqp);
+}
+
+static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
+					  struct mlx5_pagefault *pfault)
 {
-	struct mlx5_ib_dev *dev = to_mdev(qp->ibqp.pd->device);
 	int ret;
 	void *wqe, *wqe_end;
 	u32 bytes_mapped, total_wqe_bytes;
 	char *buffer = NULL;
-	int resume_with_error = 0;
-	u16 wqe_index = pfault->mpfault.wqe.wqe_index;
-	int requestor = pfault->mpfault.flags & MLX5_PFAULT_REQUESTOR;
-	u32 qpn = qp->trans_qp.base.mqp.qpn;
+	int resume_with_error = 1;
+	u16 wqe_index = pfault->wqe.wqe_index;
+	int requestor = pfault->type & MLX5_PFAULT_REQUESTOR;
+	struct mlx5_ib_qp *qp;
 
 	buffer = (char *)__get_free_page(GFP_KERNEL);
 	if (!buffer) {
 		mlx5_ib_err(dev, "Error allocating memory for IO page fault handling.\n");
-		resume_with_error = 1;
 		goto resolve_page_fault;
 	}
 
+	qp = mlx5_ib_odp_find_qp(dev, pfault->wqe.wq_num);
+	if (!qp)
+		goto resolve_page_fault;
+
 	ret = mlx5_ib_read_user_wqe(qp, requestor, wqe_index, buffer,
 				    PAGE_SIZE, &qp->trans_qp.base);
 	if (ret < 0) {
-		mlx5_ib_err(dev, "Failed reading a WQE following page fault, error=%x, wqe_index=%x, qpn=%x\n",
-			    -ret, wqe_index, qpn);
-		resume_with_error = 1;
+		mlx5_ib_err(dev, "Failed reading a WQE following page fault, error=%d, wqe_index=%x, qpn=%x\n",
+			    ret, wqe_index, pfault->token);
 		goto resolve_page_fault;
 	}
 
 	wqe = buffer;
 	if (requestor)
-		ret = mlx5_ib_mr_initiator_pfault_handler(qp, pfault, &wqe,
+		ret = mlx5_ib_mr_initiator_pfault_handler(dev, pfault, qp, &wqe,
 							  &wqe_end, ret);
 	else
-		ret = mlx5_ib_mr_responder_pfault_handler(qp, pfault, &wqe,
+		ret = mlx5_ib_mr_responder_pfault_handler(dev, pfault, qp, &wqe,
 							  &wqe_end, ret);
-	if (ret < 0) {
-		resume_with_error = 1;
+	if (ret < 0)
 		goto resolve_page_fault;
-	}
 
 	if (wqe >= wqe_end) {
 		mlx5_ib_err(dev, "ODP fault on invalid WQE.\n");
-		resume_with_error = 1;
 		goto resolve_page_fault;
 	}
 
-	ret = pagefault_data_segments(qp, pfault, wqe, wqe_end, &bytes_mapped,
-				      &total_wqe_bytes, !requestor);
+	ret = pagefault_data_segments(dev, pfault, qp, wqe, wqe_end,
+				      &bytes_mapped, &total_wqe_bytes,
+				      !requestor);
 	if (ret == -EAGAIN) {
+		resume_with_error = 0;
 		goto resolve_page_fault;
 	} else if (ret < 0 || total_wqe_bytes > bytes_mapped) {
-		mlx5_ib_err(dev, "Error getting user pages for page fault. Error: 0x%x\n",
-			    -ret);
-		resume_with_error = 1;
+		if (ret != -ENOENT)
+			mlx5_ib_err(dev, "Error getting user pages for page fault. Error: %d\n",
+				    ret);
 		goto resolve_page_fault;
 	}
 
+	resume_with_error = 0;
 resolve_page_fault:
-	mlx5_ib_page_fault_resume(qp, pfault, resume_with_error);
-	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, flags: 0x%x\n",
-		    qpn, resume_with_error,
-		    pfault->mpfault.flags);
-
+	mlx5_ib_page_fault_resume(dev, pfault, resume_with_error);
+	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x resume_with_error=%d, type: 0x%x\n",
+		    pfault->token, resume_with_error,
+		    pfault->type);
 	free_page((unsigned long)buffer);
 }
 
@@ -615,15 +621,14 @@ static int pages_in_range(u64 address, u32 length)
 		(address & PAGE_MASK)) >> PAGE_SHIFT;
 }
 
-static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_qp *qp,
-					   struct mlx5_ib_pfault *pfault)
+static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
+					   struct mlx5_pagefault *pfault)
 {
-	struct mlx5_pagefault *mpfault = &pfault->mpfault;
 	u64 address;
 	u32 length;
-	u32 prefetch_len = mpfault->bytes_committed;
+	u32 prefetch_len = pfault->bytes_committed;
 	int prefetch_activated = 0;
-	u32 rkey = mpfault->rdma.r_key;
+	u32 rkey = pfault->rdma.r_key;
 	int ret;
 
 	/* The RDMA responder handler handles the page fault in two parts.
@@ -632,38 +637,40 @@ static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_qp *qp,
 	 * prefetches more pages. The second operation cannot use the pfault
 	 * context and therefore uses the dummy_pfault context allocated on
 	 * the stack */
-	struct mlx5_ib_pfault dummy_pfault = {};
+	pfault->rdma.rdma_va += pfault->bytes_committed;
+	pfault->rdma.rdma_op_len -= min(pfault->bytes_committed,
+					 pfault->rdma.rdma_op_len);
+	pfault->bytes_committed = 0;
 
-	dummy_pfault.mpfault.bytes_committed = 0;
-
-	mpfault->rdma.rdma_va += mpfault->bytes_committed;
-	mpfault->rdma.rdma_op_len -= min(mpfault->bytes_committed,
-					 mpfault->rdma.rdma_op_len);
-	mpfault->bytes_committed = 0;
-
-	address = mpfault->rdma.rdma_va;
-	length  = mpfault->rdma.rdma_op_len;
+	address = pfault->rdma.rdma_va;
+	length  = pfault->rdma.rdma_op_len;
 
 	/* For some operations, the hardware cannot tell the exact message
 	 * length, and in those cases it reports zero. Use prefetch
 	 * logic. */
 	if (length == 0) {
 		prefetch_activated = 1;
-		length = mpfault->rdma.packet_size;
+		length = pfault->rdma.packet_size;
 		prefetch_len = min(MAX_PREFETCH_LEN, prefetch_len);
 	}
 
-	ret = pagefault_single_data_segment(qp, pfault, rkey, address, length,
-					    NULL);
+	ret = pagefault_single_data_segment(dev, rkey, address, length,
+					    &pfault->bytes_committed, NULL);
 	if (ret == -EAGAIN) {
 		/* We're racing with an invalidation, don't prefetch */
 		prefetch_activated = 0;
 	} else if (ret < 0 || pages_in_range(address, length) > ret) {
-		mlx5_ib_page_fault_resume(qp, pfault, 1);
+		mlx5_ib_page_fault_resume(dev, pfault, 1);
+		if (ret != -ENOENT)
+			mlx5_ib_warn(dev, "PAGE FAULT error %d. QP 0x%x, type: 0x%x\n",
+				     ret, pfault->token, pfault->type);
 		return;
 	}
 
-	mlx5_ib_page_fault_resume(qp, pfault, 0);
+	mlx5_ib_page_fault_resume(dev, pfault, 0);
+	mlx5_ib_dbg(dev, "PAGE FAULT completed. QP 0x%x, type: 0x%x, prefetch_activated: %d\n",
+		    pfault->token, pfault->type,
+		    prefetch_activated);
 
 	/* At this point, there might be a new pagefault already arriving in
 	 * the eq, switch to the dummy pagefault for the rest of the
@@ -671,112 +678,39 @@ static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_qp *qp,
 	 * work-queue is being fenced. */
 
 	if (prefetch_activated) {
-		ret = pagefault_single_data_segment(qp, &dummy_pfault, rkey,
-						    address,
+		u32 bytes_committed = 0;
+
+		ret = pagefault_single_data_segment(dev, rkey, address,
 						    prefetch_len,
-						    NULL);
+						    &bytes_committed, NULL);
 		if (ret < 0) {
-			pr_warn("Prefetch failed (ret = %d, prefetch_activated = %d) for QPN %d, address: 0x%.16llx, length = 0x%.16x\n",
-				ret, prefetch_activated,
-				qp->ibqp.qp_num, address, prefetch_len);
+			mlx5_ib_warn(dev, "Prefetch failed. ret: %d, QP 0x%x, address: 0x%.16llx, length = 0x%.16x\n",
+				     ret, pfault->token, address,
+				     prefetch_len);
 		}
 	}
 }
 
-void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp,
-			       struct mlx5_ib_pfault *pfault)
+void mlx5_ib_pfault(struct mlx5_core_dev *mdev, void *context,
+		    struct mlx5_pagefault *pfault)
 {
-	u8 event_subtype = pfault->mpfault.event_subtype;
+	struct mlx5_ib_dev *dev = context;
+	u8 event_subtype = pfault->event_subtype;
 
 	switch (event_subtype) {
 	case MLX5_PFAULT_SUBTYPE_WQE:
-		mlx5_ib_mr_wqe_pfault_handler(qp, pfault);
+		mlx5_ib_mr_wqe_pfault_handler(dev, pfault);
 		break;
 	case MLX5_PFAULT_SUBTYPE_RDMA:
-		mlx5_ib_mr_rdma_pfault_handler(qp, pfault);
+		mlx5_ib_mr_rdma_pfault_handler(dev, pfault);
 		break;
 	default:
-		pr_warn("Invalid page fault event subtype: 0x%x\n",
-			event_subtype);
-		mlx5_ib_page_fault_resume(qp, pfault, 1);
-		break;
+		mlx5_ib_err(dev, "Invalid page fault event subtype: 0x%x\n",
+			    event_subtype);
+		mlx5_ib_page_fault_resume(dev, pfault, 1);
 	}
 }
 
-static void mlx5_ib_qp_pfault_action(struct work_struct *work)
-{
-	struct mlx5_ib_pfault *pfault = container_of(work,
-						     struct mlx5_ib_pfault,
-						     work);
-	enum mlx5_ib_pagefault_context context =
-		mlx5_ib_get_pagefault_context(&pfault->mpfault);
-	struct mlx5_ib_qp *qp = container_of(pfault, struct mlx5_ib_qp,
-					     pagefaults[context]);
-	mlx5_ib_mr_pfault_handler(qp, pfault);
-}
-
-void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp)
-{
-	unsigned long flags;
-
-	spin_lock_irqsave(&qp->disable_page_faults_lock, flags);
-	qp->disable_page_faults = 1;
-	spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags);
-
-	/*
-	 * Note that at this point, we are guarenteed that no more
-	 * work queue elements will be posted to the work queue with
-	 * the QP we are closing.
-	 */
-	flush_workqueue(mlx5_ib_page_fault_wq);
-}
-
-void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp)
-{
-	unsigned long flags;
-
-	spin_lock_irqsave(&qp->disable_page_faults_lock, flags);
-	qp->disable_page_faults = 0;
-	spin_unlock_irqrestore(&qp->disable_page_faults_lock, flags);
-}
-
-static void mlx5_ib_pfault_handler(struct mlx5_core_qp *qp,
-				   struct mlx5_pagefault *pfault)
-{
-	/*
-	 * Note that we will only get one fault event per QP per context
-	 * (responder/initiator, read/write), until we resolve the page fault
-	 * with the mlx5_ib_page_fault_resume command. Since this function is
-	 * called from within the work element, there is no risk of missing
-	 * events.
-	 */
-	struct mlx5_ib_qp *mibqp = to_mibqp(qp);
-	enum mlx5_ib_pagefault_context context =
-		mlx5_ib_get_pagefault_context(pfault);
-	struct mlx5_ib_pfault *qp_pfault = &mibqp->pagefaults[context];
-
-	qp_pfault->mpfault = *pfault;
-
-	/* No need to stop interrupts here since we are in an interrupt */
-	spin_lock(&mibqp->disable_page_faults_lock);
-	if (!mibqp->disable_page_faults)
-		queue_work(mlx5_ib_page_fault_wq, &qp_pfault->work);
-	spin_unlock(&mibqp->disable_page_faults_lock);
-}
-
-void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp)
-{
-	int i;
-
-	qp->disable_page_faults = 1;
-	spin_lock_init(&qp->disable_page_faults_lock);
-
-	qp->trans_qp.base.mqp.pfault_handler = mlx5_ib_pfault_handler;
-
-	for (i = 0; i < MLX5_IB_PAGEFAULT_CONTEXTS; ++i)
-		INIT_WORK(&qp->pagefaults[i].work, mlx5_ib_qp_pfault_action);
-}
-
 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev)
 {
 	int ret;
@@ -793,17 +727,3 @@ void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev)
 	cleanup_srcu_struct(&ibdev->mr_srcu);
 }
 
-int __init mlx5_ib_odp_init(void)
-{
-	mlx5_ib_page_fault_wq = alloc_ordered_workqueue("mlx5_ib_page_faults",
-							WQ_MEM_RECLAIM);
-	if (!mlx5_ib_page_fault_wq)
-		return -ENOMEM;
-
-	return 0;
-}
-
-void mlx5_ib_odp_cleanup(void)
-{
-	destroy_workqueue(mlx5_ib_page_fault_wq);
-}

drivers/infiniband/hw/mlx5/qp.c

@@ -1526,9 +1526,6 @@ static int create_qp_common(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 	       &qp->raw_packet_qp.rq.base :
 	       &qp->trans_qp.base;
 
-	if (init_attr->qp_type != IB_QPT_RAW_PACKET)
-		mlx5_ib_odp_create_qp(qp);
-
 	mutex_init(&qp->mutex);
 	spin_lock_init(&qp->sq.lock);
 	spin_lock_init(&qp->rq.lock);
@@ -1923,7 +1920,6 @@ static void destroy_qp_common(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
 
 	if (qp->state != IB_QPS_RESET) {
 		if (qp->ibqp.qp_type != IB_QPT_RAW_PACKET) {
-			mlx5_ib_qp_disable_pagefaults(qp);
 			err = mlx5_core_qp_modify(dev->mdev,
 						  MLX5_CMD_OP_2RST_QP, 0,
 						  NULL, &base->mqp);
@@ -2823,16 +2819,6 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 	if (mlx5_st < 0)
 		goto out;
 
-	/* If moving to a reset or error state, we must disable page faults on
-	 * this QP and flush all current page faults. Otherwise a stale page
-	 * fault may attempt to work on this QP after it is reset and moved
-	 * again to RTS, and may cause the driver and the device to get out of
-	 * sync. */
-	if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
-	    (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR) &&
-	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
-		mlx5_ib_qp_disable_pagefaults(qp);
-
 	if (mlx5_cur >= MLX5_QP_NUM_STATE || mlx5_new >= MLX5_QP_NUM_STATE ||
 	    !optab[mlx5_cur][mlx5_new])
 		goto out;
@@ -2864,10 +2850,6 @@ static int __mlx5_ib_modify_qp(struct ib_qp *ibqp,
 	if (err)
 		goto out;
 
-	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT &&
-	    (qp->ibqp.qp_type != IB_QPT_RAW_PACKET))
-		mlx5_ib_qp_enable_pagefaults(qp);
-
 	qp->state = new_state;
 
 	if (attr_mask & IB_QP_ACCESS_FLAGS)
@@ -4533,14 +4515,6 @@ int mlx5_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr,
 		return mlx5_ib_gsi_query_qp(ibqp, qp_attr, qp_attr_mask,
 					    qp_init_attr);
 
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-	/*
-	 * Wait for any outstanding page faults, in case the user frees memory
-	 * based upon this query's result.
-	 */
-	flush_workqueue(mlx5_ib_page_fault_wq);
-#endif
-
 	mutex_lock(&qp->mutex);
 
 	if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET) {

drivers/net/ethernet/mellanox/mlx5/core/dev.c

@@ -71,6 +71,16 @@ void mlx5_add_device(struct mlx5_interface *intf, struct mlx5_priv *priv)
 	if (dev_ctx->context) {
 		spin_lock_irq(&priv->ctx_lock);
 		list_add_tail(&dev_ctx->list, &priv->ctx_list);
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+		if (dev_ctx->intf->pfault) {
+			if (priv->pfault) {
+				mlx5_core_err(dev, "multiple page fault handlers not supported");
+			} else {
+				priv->pfault_ctx = dev_ctx->context;
+				priv->pfault = dev_ctx->intf->pfault;
+			}
+		}
+#endif
 		spin_unlock_irq(&priv->ctx_lock);
 	} else {
 		kfree(dev_ctx);
@@ -97,6 +107,15 @@ void mlx5_remove_device(struct mlx5_interface *intf, struct mlx5_priv *priv)
 	if (!dev_ctx)
 		return;
 
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	spin_lock_irq(&priv->ctx_lock);
+	if (priv->pfault == dev_ctx->intf->pfault)
+		priv->pfault = NULL;
+	spin_unlock_irq(&priv->ctx_lock);
+
+	synchronize_srcu(&priv->pfault_srcu);
+#endif
+
 	spin_lock_irq(&priv->ctx_lock);
 	list_del(&dev_ctx->list);
 	spin_unlock_irq(&priv->ctx_lock);
@@ -329,6 +348,20 @@ void mlx5_core_event(struct mlx5_core_dev *dev, enum mlx5_dev_event event,
 	spin_unlock_irqrestore(&priv->ctx_lock, flags);
 }
 
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+void mlx5_core_page_fault(struct mlx5_core_dev *dev,
+			  struct mlx5_pagefault *pfault)
+{
+	struct mlx5_priv *priv = &dev->priv;
+	int srcu_idx;
+
+	srcu_idx = srcu_read_lock(&priv->pfault_srcu);
+	if (priv->pfault)
+		priv->pfault(dev, priv->pfault_ctx, pfault);
+	srcu_read_unlock(&priv->pfault_srcu, srcu_idx);
+}
+#endif
+
 void mlx5_dev_list_lock(void)
 {
 	mutex_lock(&mlx5_intf_mutex);

drivers/net/ethernet/mellanox/mlx5/core/eq.c

@@ -54,6 +54,7 @@ enum {
 	MLX5_NUM_SPARE_EQE	= 0x80,
 	MLX5_NUM_ASYNC_EQE	= 0x100,
 	MLX5_NUM_CMD_EQE	= 32,
+	MLX5_NUM_PF_DRAIN	= 64,
 };
 
 enum {
@@ -188,10 +189,193 @@ static void eq_update_ci(struct mlx5_eq *eq, int arm)
 	mb();
 }
 
-static int mlx5_eq_int(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+static void eqe_pf_action(struct work_struct *work)
+{
+	struct mlx5_pagefault *pfault = container_of(work,
+						     struct mlx5_pagefault,
+						     work);
+	struct mlx5_eq *eq = pfault->eq;
+
+	mlx5_core_page_fault(eq->dev, pfault);
+	mempool_free(pfault, eq->pf_ctx.pool);
+}
+
+static void eq_pf_process(struct mlx5_eq *eq)
+{
+	struct mlx5_core_dev *dev = eq->dev;
+	struct mlx5_eqe_page_fault *pf_eqe;
+	struct mlx5_pagefault *pfault;
+	struct mlx5_eqe *eqe;
+	int set_ci = 0;
+
+	while ((eqe = next_eqe_sw(eq))) {
+		pfault = mempool_alloc(eq->pf_ctx.pool, GFP_ATOMIC);
+		if (!pfault) {
+			schedule_work(&eq->pf_ctx.work);
+			break;
+		}
+
+		dma_rmb();
+		pf_eqe = &eqe->data.page_fault;
+		pfault->event_subtype = eqe->sub_type;
+		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
+
+		mlx5_core_dbg(dev,
+			      "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
+			      eqe->sub_type, pfault->bytes_committed);
+
+		switch (eqe->sub_type) {
+		case MLX5_PFAULT_SUBTYPE_RDMA:
+			/* RDMA based event */
+			pfault->type =
+				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
+			pfault->token =
+				be32_to_cpu(pf_eqe->rdma.pftype_token) &
+				MLX5_24BIT_MASK;
+			pfault->rdma.r_key =
+				be32_to_cpu(pf_eqe->rdma.r_key);
+			pfault->rdma.packet_size =
+				be16_to_cpu(pf_eqe->rdma.packet_length);
+			pfault->rdma.rdma_op_len =
+				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
+			pfault->rdma.rdma_va =
+				be64_to_cpu(pf_eqe->rdma.rdma_va);
+			mlx5_core_dbg(dev,
+				      "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
+				      pfault->type, pfault->token,
+				      pfault->rdma.r_key);
+			mlx5_core_dbg(dev,
+				      "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
+				      pfault->rdma.rdma_op_len,
+				      pfault->rdma.rdma_va);
+			break;
+
+		case MLX5_PFAULT_SUBTYPE_WQE:
+			/* WQE based event */
+			pfault->type =
+				be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24;
+			pfault->token =
+				be32_to_cpu(pf_eqe->wqe.token);
+			pfault->wqe.wq_num =
+				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
+				MLX5_24BIT_MASK;
+			pfault->wqe.wqe_index =
+				be16_to_cpu(pf_eqe->wqe.wqe_index);
+			pfault->wqe.packet_size =
+				be16_to_cpu(pf_eqe->wqe.packet_length);
+			mlx5_core_dbg(dev,
+				      "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
+				      pfault->type, pfault->token,
+				      pfault->wqe.wq_num,
+				      pfault->wqe.wqe_index);
+			break;
+
+		default:
+			mlx5_core_warn(dev,
+				       "Unsupported page fault event sub-type: 0x%02hhx\n",
+				       eqe->sub_type);
+			/* Unsupported page faults should still be
+			 * resolved by the page fault handler
+			 */
+		}
+
+		pfault->eq = eq;
+		INIT_WORK(&pfault->work, eqe_pf_action);
+		queue_work(eq->pf_ctx.wq, &pfault->work);
+
+		++eq->cons_index;
+		++set_ci;
+
+		if (unlikely(set_ci >= MLX5_NUM_SPARE_EQE)) {
+			eq_update_ci(eq, 0);
+			set_ci = 0;
+		}
+	}
+
+	eq_update_ci(eq, 1);
+}
+
+static irqreturn_t mlx5_eq_pf_int(int irq, void *eq_ptr)
+{
+	struct mlx5_eq *eq = eq_ptr;
+	unsigned long flags;
+
+	if (spin_trylock_irqsave(&eq->pf_ctx.lock, flags)) {
+		eq_pf_process(eq);
+		spin_unlock_irqrestore(&eq->pf_ctx.lock, flags);
+	} else {
+		schedule_work(&eq->pf_ctx.work);
+	}
+
+	return IRQ_HANDLED;
+}
+
+/* mempool_refill() was proposed but unfortunately wasn't accepted
+ * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
+ * Chip workaround.
+ */
+static void mempool_refill(mempool_t *pool)
+{
+	while (pool->curr_nr < pool->min_nr)
+		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
+}
+
+static void eq_pf_action(struct work_struct *work)
+{
+	struct mlx5_eq *eq = container_of(work, struct mlx5_eq, pf_ctx.work);
+
+	mempool_refill(eq->pf_ctx.pool);
+
+	spin_lock_irq(&eq->pf_ctx.lock);
+	eq_pf_process(eq);
+	spin_unlock_irq(&eq->pf_ctx.lock);
+}
+
+static int init_pf_ctx(struct mlx5_eq_pagefault *pf_ctx, const char *name)
 {
+	spin_lock_init(&pf_ctx->lock);
+	INIT_WORK(&pf_ctx->work, eq_pf_action);
+
+	pf_ctx->wq = alloc_ordered_workqueue(name,
+					     WQ_MEM_RECLAIM);
+	if (!pf_ctx->wq)
+		return -ENOMEM;
+
+	pf_ctx->pool = mempool_create_kmalloc_pool
+		(MLX5_NUM_PF_DRAIN, sizeof(struct mlx5_pagefault));
+	if (!pf_ctx->pool)
+		goto err_wq;
+
+	return 0;
+err_wq:
+	destroy_workqueue(pf_ctx->wq);
+	return -ENOMEM;
+}
+
+int mlx5_core_page_fault_resume(struct mlx5_core_dev *dev, u32 token,
+				u32 wq_num, u8 type, int error)
+{
+	u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = {0};
+	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)]   = {0};
+
+	MLX5_SET(page_fault_resume_in, in, opcode,
+		 MLX5_CMD_OP_PAGE_FAULT_RESUME);
+	MLX5_SET(page_fault_resume_in, in, error, !!error);
+	MLX5_SET(page_fault_resume_in, in, page_fault_type, type);
+	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
+	MLX5_SET(page_fault_resume_in, in, token, token);
+
+	return mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
+}
+EXPORT_SYMBOL_GPL(mlx5_core_page_fault_resume);
+#endif
+
+static irqreturn_t mlx5_eq_int(int irq, void *eq_ptr)
+{
+	struct mlx5_eq *eq = eq_ptr;
+	struct mlx5_core_dev *dev = eq->dev;
 	struct mlx5_eqe *eqe;
-	int eqes_found = 0;
 	int set_ci = 0;
 	u32 cqn = -1;
 	u32 rsn;
@@ -276,12 +460,6 @@ static int mlx5_eq_int(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
 			}
 			break;
 
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-		case MLX5_EVENT_TYPE_PAGE_FAULT:
-			mlx5_eq_pagefault(dev, eqe);
-			break;
-#endif
-
 #ifdef CONFIG_MLX5_CORE_EN
 		case MLX5_EVENT_TYPE_NIC_VPORT_CHANGE:
 			mlx5_eswitch_vport_event(dev->priv.eswitch, eqe);
@@ -299,7 +477,6 @@ static int mlx5_eq_int(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
 		}
 
 		++eq->cons_index;
-		eqes_found = 1;
 		++set_ci;
 
 		/* The HCA will think the queue has overflowed if we
@@ -319,17 +496,6 @@ static int mlx5_eq_int(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
 	if (cqn != -1)
 		tasklet_schedule(&eq->tasklet_ctx.task);
 
-	return eqes_found;
-}
-
-static irqreturn_t mlx5_msix_handler(int irq, void *eq_ptr)
-{
-	struct mlx5_eq *eq = eq_ptr;
-	struct mlx5_core_dev *dev = eq->dev;
-
-	mlx5_eq_int(dev, eq);
-
-	/* MSI-X vectors always belong to us */
 	return IRQ_HANDLED;
 }
 
@@ -345,22 +511,32 @@ static void init_eq_buf(struct mlx5_eq *eq)
 }
 
 int mlx5_create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq, u8 vecidx,
-		       int nent, u64 mask, const char *name, struct mlx5_uar *uar)
+		       int nent, u64 mask, const char *name,
+		       struct mlx5_uar *uar, enum mlx5_eq_type type)
 {
 	u32 out[MLX5_ST_SZ_DW(create_eq_out)] = {0};
 	struct mlx5_priv *priv = &dev->priv;
+	irq_handler_t handler;
 	__be64 *pas;
 	void *eqc;
 	int inlen;
 	u32 *in;
 	int err;
 
+	eq->type = type;
 	eq->nent = roundup_pow_of_two(nent + MLX5_NUM_SPARE_EQE);
 	eq->cons_index = 0;
 	err = mlx5_buf_alloc(dev, eq->nent * MLX5_EQE_SIZE, &eq->buf);
 	if (err)
 		return err;
 
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	if (type == MLX5_EQ_TYPE_PF)
+		handler = mlx5_eq_pf_int;
+	else
+#endif
+		handler = mlx5_eq_int;
+
 	init_eq_buf(eq);
 
 	inlen = MLX5_ST_SZ_BYTES(create_eq_in) +
@@ -396,7 +572,7 @@ int mlx5_create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq, u8 vecidx,
 	eq->irqn = priv->msix_arr[vecidx].vector;
 	eq->dev = dev;
 	eq->doorbell = uar->map + MLX5_EQ_DOORBEL_OFFSET;
-	err = request_irq(eq->irqn, mlx5_msix_handler, 0,
+	err = request_irq(eq->irqn, handler, 0,
 			  priv->irq_info[vecidx].name, eq);
 	if (err)
 		goto err_eq;
@@ -405,11 +581,20 @@ int mlx5_create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq, u8 vecidx,
 	if (err)
 		goto err_irq;
 
-	INIT_LIST_HEAD(&eq->tasklet_ctx.list);
-	INIT_LIST_HEAD(&eq->tasklet_ctx.process_list);
-	spin_lock_init(&eq->tasklet_ctx.lock);
-	tasklet_init(&eq->tasklet_ctx.task, mlx5_cq_tasklet_cb,
-		     (unsigned long)&eq->tasklet_ctx);
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	if (type == MLX5_EQ_TYPE_PF) {
+		err = init_pf_ctx(&eq->pf_ctx, name);
+		if (err)
+			goto err_irq;
+	} else
+#endif
+	{
+		INIT_LIST_HEAD(&eq->tasklet_ctx.list);
+		INIT_LIST_HEAD(&eq->tasklet_ctx.process_list);
+		spin_lock_init(&eq->tasklet_ctx.lock);
+		tasklet_init(&eq->tasklet_ctx.task, mlx5_cq_tasklet_cb,
+			     (unsigned long)&eq->tasklet_ctx);
+	}
 
 	/* EQs are created in ARMED state
 	 */
@@ -444,7 +629,16 @@ int mlx5_destroy_unmap_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq)
 		mlx5_core_warn(dev, "failed to destroy a previously created eq: eqn %d\n",
 			       eq->eqn);
 	synchronize_irq(eq->irqn);
-	tasklet_disable(&eq->tasklet_ctx.task);
+
+	if (eq->type == MLX5_EQ_TYPE_COMP) {
+		tasklet_disable(&eq->tasklet_ctx.task);
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	} else if (eq->type == MLX5_EQ_TYPE_PF) {
+		cancel_work_sync(&eq->pf_ctx.work);
+		destroy_workqueue(eq->pf_ctx.wq);
+		mempool_destroy(eq->pf_ctx.pool);
+#endif
+	}
 	mlx5_buf_free(dev, &eq->buf);
 
 	return err;
@@ -479,8 +673,6 @@ int mlx5_start_eqs(struct mlx5_core_dev *dev)
 	u64 async_event_mask = MLX5_ASYNC_EVENT_MASK;
 	int err;
 
-	if (MLX5_CAP_GEN(dev, pg))
-		async_event_mask |= (1ull << MLX5_EVENT_TYPE_PAGE_FAULT);
 
 	if (MLX5_CAP_GEN(dev, port_type) == MLX5_CAP_PORT_TYPE_ETH &&
 	    MLX5_CAP_GEN(dev, vport_group_manager) &&
@@ -494,7 +686,8 @@ int mlx5_start_eqs(struct mlx5_core_dev *dev)
 
 	err = mlx5_create_map_eq(dev, &table->cmd_eq, MLX5_EQ_VEC_CMD,
 				 MLX5_NUM_CMD_EQE, 1ull << MLX5_EVENT_TYPE_CMD,
-				 "mlx5_cmd_eq", &dev->priv.uuari.uars[0]);
+				 "mlx5_cmd_eq", &dev->priv.uuari.uars[0],
+				 MLX5_EQ_TYPE_ASYNC);
 	if (err) {
 		mlx5_core_warn(dev, "failed to create cmd EQ %d\n", err);
 		return err;
@@ -504,7 +697,8 @@ int mlx5_start_eqs(struct mlx5_core_dev *dev)
 
 	err = mlx5_create_map_eq(dev, &table->async_eq, MLX5_EQ_VEC_ASYNC,
 				 MLX5_NUM_ASYNC_EQE, async_event_mask,
-				 "mlx5_async_eq", &dev->priv.uuari.uars[0]);
+				 "mlx5_async_eq", &dev->priv.uuari.uars[0],
+				 MLX5_EQ_TYPE_ASYNC);
 	if (err) {
 		mlx5_core_warn(dev, "failed to create async EQ %d\n", err);
 		goto err1;
@@ -514,13 +708,35 @@ int mlx5_start_eqs(struct mlx5_core_dev *dev)
 				 MLX5_EQ_VEC_PAGES,
 				 /* TODO: sriov max_vf + */ 1,
 				 1 << MLX5_EVENT_TYPE_PAGE_REQUEST, "mlx5_pages_eq",
-				 &dev->priv.uuari.uars[0]);
+				 &dev->priv.uuari.uars[0],
+				 MLX5_EQ_TYPE_ASYNC);
 	if (err) {
 		mlx5_core_warn(dev, "failed to create pages EQ %d\n", err);
 		goto err2;
 	}
 
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	if (MLX5_CAP_GEN(dev, pg)) {
+		err = mlx5_create_map_eq(dev, &table->pfault_eq,
+					 MLX5_EQ_VEC_PFAULT,
+					 MLX5_NUM_ASYNC_EQE,
+					 1 << MLX5_EVENT_TYPE_PAGE_FAULT,
+					 "mlx5_page_fault_eq",
+					 &dev->priv.uuari.uars[0],
+					 MLX5_EQ_TYPE_PF);
+		if (err) {
+			mlx5_core_warn(dev, "failed to create page fault EQ %d\n",
+				       err);
+			goto err3;
+		}
+	}
+
 	return err;
+err3:
+	mlx5_destroy_unmap_eq(dev, &table->pages_eq);
+#else
+	return err;
+#endif
 
 err2:
 	mlx5_destroy_unmap_eq(dev, &table->async_eq);
@@ -536,6 +752,14 @@ int mlx5_stop_eqs(struct mlx5_core_dev *dev)
 	struct mlx5_eq_table *table = &dev->priv.eq_table;
 	int err;
 
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	if (MLX5_CAP_GEN(dev, pg)) {
+		err = mlx5_destroy_unmap_eq(dev, &table->pfault_eq);
+		if (err)
+			return err;
+	}
+#endif
+
 	err = mlx5_destroy_unmap_eq(dev, &table->pages_eq);
 	if (err)
 		return err;

drivers/net/ethernet/mellanox/mlx5/core/main.c

@@ -753,7 +753,8 @@ static int alloc_comp_eqs(struct mlx5_core_dev *dev)
 		snprintf(name, MLX5_MAX_IRQ_NAME, "mlx5_comp%d", i);
 		err = mlx5_create_map_eq(dev, eq,
 					 i + MLX5_EQ_VEC_COMP_BASE, nent, 0,
-					 name, &dev->priv.uuari.uars[0]);
+					 name, &dev->priv.uuari.uars[0],
+					 MLX5_EQ_TYPE_COMP);
 		if (err) {
 			kfree(eq);
 			goto clean;
@@ -1295,10 +1296,19 @@ static int init_one(struct pci_dev *pdev,
 	spin_lock_init(&priv->ctx_lock);
 	mutex_init(&dev->pci_status_mutex);
 	mutex_init(&dev->intf_state_mutex);
+
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	err = init_srcu_struct(&priv->pfault_srcu);
+	if (err) {
+		dev_err(&pdev->dev, "init_srcu_struct failed with error code %d\n",
+			err);
+		goto clean_dev;
+	}
+#endif
 	err = mlx5_pci_init(dev, priv);
 	if (err) {
 		dev_err(&pdev->dev, "mlx5_pci_init failed with error code %d\n", err);
-		goto clean_dev;
+		goto clean_srcu;
 	}
 
 	err = mlx5_health_init(dev);
@@ -1332,7 +1342,11 @@ static int init_one(struct pci_dev *pdev,
 	mlx5_health_cleanup(dev);
 close_pci:
 	mlx5_pci_close(dev, priv);
+clean_srcu:
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	cleanup_srcu_struct(&priv->pfault_srcu);
 clean_dev:
+#endif
 	pci_set_drvdata(pdev, NULL);
 	devlink_free(devlink);
 
@@ -1357,6 +1371,9 @@ static void remove_one(struct pci_dev *pdev)
 	mlx5_pagealloc_cleanup(dev);
 	mlx5_health_cleanup(dev);
 	mlx5_pci_close(dev, priv);
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	cleanup_srcu_struct(&priv->pfault_srcu);
+#endif
 	pci_set_drvdata(pdev, NULL);
 	devlink_free(devlink);
 }

drivers/net/ethernet/mellanox/mlx5/core/mlx5_core.h

@@ -86,6 +86,8 @@ int mlx5_cmd_init_hca(struct mlx5_core_dev *dev);
 int mlx5_cmd_teardown_hca(struct mlx5_core_dev *dev);
 void mlx5_core_event(struct mlx5_core_dev *dev, enum mlx5_dev_event event,
 		     unsigned long param);
+void mlx5_core_page_fault(struct mlx5_core_dev *dev,
+			  struct mlx5_pagefault *pfault);
 void mlx5_port_module_event(struct mlx5_core_dev *dev, struct mlx5_eqe *eqe);
 void mlx5_enter_error_state(struct mlx5_core_dev *dev);
 void mlx5_disable_device(struct mlx5_core_dev *dev);

drivers/net/ethernet/mellanox/mlx5/core/qp.c

@@ -143,95 +143,6 @@ void mlx5_rsc_event(struct mlx5_core_dev *dev, u32 rsn, int event_type)
 	mlx5_core_put_rsc(common);
 }
 
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-void mlx5_eq_pagefault(struct mlx5_core_dev *dev, struct mlx5_eqe *eqe)
-{
-	struct mlx5_eqe_page_fault *pf_eqe = &eqe->data.page_fault;
-	int qpn = be32_to_cpu(pf_eqe->flags_qpn) & MLX5_QPN_MASK;
-	struct mlx5_core_rsc_common *common = mlx5_get_rsc(dev, qpn);
-	struct mlx5_core_qp *qp =
-		container_of(common, struct mlx5_core_qp, common);
-	struct mlx5_pagefault pfault;
-
-	if (!qp) {
-		mlx5_core_warn(dev, "ODP event for non-existent QP %06x\n",
-			       qpn);
-		return;
-	}
-
-	pfault.event_subtype = eqe->sub_type;
-	pfault.flags = (be32_to_cpu(pf_eqe->flags_qpn) >> MLX5_QPN_BITS) &
-		(MLX5_PFAULT_REQUESTOR | MLX5_PFAULT_WRITE | MLX5_PFAULT_RDMA);
-	pfault.bytes_committed = be32_to_cpu(
-		pf_eqe->bytes_committed);
-
-	mlx5_core_dbg(dev,
-		      "PAGE_FAULT: subtype: 0x%02x, flags: 0x%02x,\n",
-		      eqe->sub_type, pfault.flags);
-
-	switch (eqe->sub_type) {
-	case MLX5_PFAULT_SUBTYPE_RDMA:
-		/* RDMA based event */
-		pfault.rdma.r_key =
-			be32_to_cpu(pf_eqe->rdma.r_key);
-		pfault.rdma.packet_size =
-			be16_to_cpu(pf_eqe->rdma.packet_length);
-		pfault.rdma.rdma_op_len =
-			be32_to_cpu(pf_eqe->rdma.rdma_op_len);
-		pfault.rdma.rdma_va =
-			be64_to_cpu(pf_eqe->rdma.rdma_va);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: qpn: 0x%06x, r_key: 0x%08x,\n",
-			      qpn, pfault.rdma.r_key);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: rdma_op_len: 0x%08x,\n",
-			      pfault.rdma.rdma_op_len);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: rdma_va: 0x%016llx,\n",
-			      pfault.rdma.rdma_va);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: bytes_committed: 0x%06x\n",
-			      pfault.bytes_committed);
-		break;
-
-	case MLX5_PFAULT_SUBTYPE_WQE:
-		/* WQE based event */
-		pfault.wqe.wqe_index =
-			be16_to_cpu(pf_eqe->wqe.wqe_index);
-		pfault.wqe.packet_size =
-			be16_to_cpu(pf_eqe->wqe.packet_length);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: qpn: 0x%06x, wqe_index: 0x%04x,\n",
-			      qpn, pfault.wqe.wqe_index);
-		mlx5_core_dbg(dev,
-			      "PAGE_FAULT: bytes_committed: 0x%06x\n",
-			      pfault.bytes_committed);
-		break;
-
-	default:
-		mlx5_core_warn(dev,
-			       "Unsupported page fault event sub-type: 0x%02hhx, QP %06x\n",
-			       eqe->sub_type, qpn);
-		/* Unsupported page faults should still be resolved by the
-		 * page fault handler
-		 */
-	}
-
-	if (qp->pfault_handler) {
-		qp->pfault_handler(qp, &pfault);
-	} else {
-		mlx5_core_err(dev,
-			      "ODP event for QP %08x, without a fault handler in QP\n",
-			      qpn);
-		/* Page fault will remain unresolved. QP will hang until it is
-		 * destroyed
-		 */
-	}
-
-	mlx5_core_put_rsc(common);
-}
-#endif
-
 static int create_qprqsq_common(struct mlx5_core_dev *dev,
 				struct mlx5_core_qp *qp,
 				int rsc_type)
@@ -506,25 +417,6 @@ int mlx5_core_xrcd_dealloc(struct mlx5_core_dev *dev, u32 xrcdn)
 }
 EXPORT_SYMBOL_GPL(mlx5_core_xrcd_dealloc);
 
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-int mlx5_core_page_fault_resume(struct mlx5_core_dev *dev, u32 qpn,
-				u8 flags, int error)
-{
-	u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = {0};
-	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)]   = {0};
-
-	MLX5_SET(page_fault_resume_in, in, opcode,
-		 MLX5_CMD_OP_PAGE_FAULT_RESUME);
-	MLX5_SET(page_fault_resume_in, in, wq_number, qpn);
-	MLX5_SET(page_fault_resume_in, in, page_fault_type, flags);
-	if (error)
-		MLX5_SET(page_fault_resume_in, in, error, 1);
-
-	return mlx5_cmd_exec(dev, in, sizeof(in), out, sizeof(out));
-}
-EXPORT_SYMBOL_GPL(mlx5_core_page_fault_resume);
-#endif
-
 int mlx5_core_create_rq_tracked(struct mlx5_core_dev *dev, u32 *in, int inlen,
 				struct mlx5_core_qp *rq)
 {

include/linux/mlx5/device.h

@@ -534,7 +534,9 @@ struct mlx5_eqe_page_fault {
 			__be16  wqe_index;
 			u16	reserved2;
 			__be16  packet_length;
-			u8	reserved3[12];
+			__be32  token;
+			u8	reserved4[8];
+			__be32  pftype_wq;
 		} __packed wqe;
 		struct {
 			__be32  r_key;
@@ -542,9 +544,9 @@ struct mlx5_eqe_page_fault {
 			__be16  packet_length;
 			__be32  rdma_op_len;
 			__be64  rdma_va;
+			__be32  pftype_token;
 		} __packed rdma;
 	} __packed;
-	__be32 flags_qpn;
 } __packed;
 
 struct mlx5_eqe_vport_change {

include/linux/mlx5/driver.h

@@ -42,6 +42,7 @@
 #include <linux/vmalloc.h>
 #include <linux/radix-tree.h>
 #include <linux/workqueue.h>
+#include <linux/mempool.h>
 #include <linux/interrupt.h>
 
 #include <linux/mlx5/device.h>
@@ -83,6 +84,7 @@ enum {
 	MLX5_EQ_VEC_PAGES	 = 0,
 	MLX5_EQ_VEC_CMD		 = 1,
 	MLX5_EQ_VEC_ASYNC	 = 2,
+	MLX5_EQ_VEC_PFAULT	 = 3,
 	MLX5_EQ_VEC_COMP_BASE,
 };
 
@@ -178,6 +180,14 @@ enum mlx5_port_status {
 	MLX5_PORT_DOWN      = 2,
 };
 
+enum mlx5_eq_type {
+	MLX5_EQ_TYPE_COMP,
+	MLX5_EQ_TYPE_ASYNC,
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	MLX5_EQ_TYPE_PF,
+#endif
+};
+
 struct mlx5_uuar_info {
 	struct mlx5_uar	       *uars;
 	int			num_uars;
@@ -333,6 +343,14 @@ struct mlx5_eq_tasklet {
 	spinlock_t lock;
 };
 
+struct mlx5_eq_pagefault {
+	struct work_struct       work;
+	/* Pagefaults lock */
+	spinlock_t		 lock;
+	struct workqueue_struct *wq;
+	mempool_t		*pool;
+};
+
 struct mlx5_eq {
 	struct mlx5_core_dev   *dev;
 	__be32 __iomem	       *doorbell;
@@ -346,7 +364,13 @@ struct mlx5_eq {
 	struct list_head	list;
 	int			index;
 	struct mlx5_rsc_debug	*dbg;
-	struct mlx5_eq_tasklet	tasklet_ctx;
+	enum mlx5_eq_type	type;
+	union {
+		struct mlx5_eq_tasklet   tasklet_ctx;
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+		struct mlx5_eq_pagefault pf_ctx;
+#endif
+	};
 };
 
 struct mlx5_core_psv {
@@ -377,6 +401,8 @@ struct mlx5_core_mkey {
 	u32			pd;
 };
 
+#define MLX5_24BIT_MASK		((1 << 24) - 1)
+
 enum mlx5_res_type {
 	MLX5_RES_QP	= MLX5_EVENT_QUEUE_TYPE_QP,
 	MLX5_RES_RQ	= MLX5_EVENT_QUEUE_TYPE_RQ,
@@ -411,6 +437,9 @@ struct mlx5_eq_table {
 	struct mlx5_eq		pages_eq;
 	struct mlx5_eq		async_eq;
 	struct mlx5_eq		cmd_eq;
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	struct mlx5_eq		pfault_eq;
+#endif
 	int			num_comp_vectors;
 	/* protect EQs list
 	 */
@@ -497,6 +526,7 @@ struct mlx5_fc_stats {
 
 struct mlx5_eswitch;
 struct mlx5_lag;
+struct mlx5_pagefault;
 
 struct mlx5_rl_entry {
 	u32                     rate;
@@ -601,6 +631,14 @@ struct mlx5_priv {
 	struct mlx5_rl_table            rl_table;
 
 	struct mlx5_port_module_event_stats  pme_stats;
+
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	void		      (*pfault)(struct mlx5_core_dev *dev,
+					void *context,
+					struct mlx5_pagefault *pfault);
+	void		       *pfault_ctx;
+	struct srcu_struct      pfault_srcu;
+#endif
 };
 
 enum mlx5_device_state {
@@ -619,6 +657,50 @@ enum mlx5_pci_status {
 	MLX5_PCI_STATUS_ENABLED,
 };
 
+enum mlx5_pagefault_type_flags {
+	MLX5_PFAULT_REQUESTOR = 1 << 0,
+	MLX5_PFAULT_WRITE     = 1 << 1,
+	MLX5_PFAULT_RDMA      = 1 << 2,
+};
+
+/* Contains the details of a pagefault. */
+struct mlx5_pagefault {
+	u32			bytes_committed;
+	u32			token;
+	u8			event_subtype;
+	u8			type;
+	union {
+		/* Initiator or send message responder pagefault details. */
+		struct {
+			/* Received packet size, only valid for responders. */
+			u32	packet_size;
+			/*
+			 * Number of resource holding WQE, depends on type.
+			 */
+			u32	wq_num;
+			/*
+			 * WQE index. Refers to either the send queue or
+			 * receive queue, according to event_subtype.
+			 */
+			u16	wqe_index;
+		} wqe;
+		/* RDMA responder pagefault details */
+		struct {
+			u32	r_key;
+			/*
+			 * Received packet size, minimal size page fault
+			 * resolution required for forward progress.
+			 */
+			u32	packet_size;
+			u32	rdma_op_len;
+			u64	rdma_va;
+		} rdma;
+	};
+
+	struct mlx5_eq	       *eq;
+	struct work_struct	work;
+};
+
 struct mlx5_td {
 	struct list_head tirs_list;
 	u32              tdn;
@@ -879,15 +961,13 @@ void mlx5_fill_page_array(struct mlx5_buf *buf, __be64 *pas);
 void mlx5_fill_page_frag_array(struct mlx5_frag_buf *frag_buf, __be64 *pas);
 void mlx5_cq_completion(struct mlx5_core_dev *dev, u32 cqn);
 void mlx5_rsc_event(struct mlx5_core_dev *dev, u32 rsn, int event_type);
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-void mlx5_eq_pagefault(struct mlx5_core_dev *dev, struct mlx5_eqe *eqe);
-#endif
 void mlx5_srq_event(struct mlx5_core_dev *dev, u32 srqn, int event_type);
 struct mlx5_core_srq *mlx5_core_get_srq(struct mlx5_core_dev *dev, u32 srqn);
 void mlx5_cmd_comp_handler(struct mlx5_core_dev *dev, u64 vec);
 void mlx5_cq_event(struct mlx5_core_dev *dev, u32 cqn, int event_type);
 int mlx5_create_map_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq, u8 vecidx,
-		       int nent, u64 mask, const char *name, struct mlx5_uar *uar);
+		       int nent, u64 mask, const char *name,
+		       struct mlx5_uar *uar, enum mlx5_eq_type type);
 int mlx5_destroy_unmap_eq(struct mlx5_core_dev *dev, struct mlx5_eq *eq);
 int mlx5_start_eqs(struct mlx5_core_dev *dev);
 int mlx5_stop_eqs(struct mlx5_core_dev *dev);
@@ -926,6 +1006,10 @@ int mlx5_query_odp_caps(struct mlx5_core_dev *dev,
 			struct mlx5_odp_caps *odp_caps);
 int mlx5_core_query_ib_ppcnt(struct mlx5_core_dev *dev,
 			     u8 port_num, void *out, size_t sz);
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+int mlx5_core_page_fault_resume(struct mlx5_core_dev *dev, u32 token,
+				u32 wq_num, u8 type, int error);
+#endif
 
 int mlx5_init_rl_table(struct mlx5_core_dev *dev);
 void mlx5_cleanup_rl_table(struct mlx5_core_dev *dev);
@@ -974,6 +1058,9 @@ struct mlx5_interface {
 	void			(*detach)(struct mlx5_core_dev *dev, void *context);
 	void			(*event)(struct mlx5_core_dev *dev, void *context,
 					 enum mlx5_dev_event event, unsigned long param);
+	void			(*pfault)(struct mlx5_core_dev *dev,
+					  void *context,
+					  struct mlx5_pagefault *pfault);
 	void *                  (*get_dev)(void *context);
 	int			protocol;
 	struct list_head	list;

include/linux/mlx5/qp.h

@@ -50,9 +50,6 @@
 #define MLX5_BSF_APPTAG_ESCAPE	0x1
 #define MLX5_BSF_APPREF_ESCAPE	0x2
 
-#define MLX5_QPN_BITS		24
-#define MLX5_QPN_MASK		((1 << MLX5_QPN_BITS) - 1)
-
 enum mlx5_qp_optpar {
 	MLX5_QP_OPTPAR_ALT_ADDR_PATH		= 1 << 0,
 	MLX5_QP_OPTPAR_RRE			= 1 << 1,
@@ -418,46 +415,9 @@ struct mlx5_stride_block_ctrl_seg {
 	__be16		num_entries;
 };
 
-enum mlx5_pagefault_flags {
-	MLX5_PFAULT_REQUESTOR = 1 << 0,
-	MLX5_PFAULT_WRITE     = 1 << 1,
-	MLX5_PFAULT_RDMA      = 1 << 2,
-};
-
-/* Contains the details of a pagefault. */
-struct mlx5_pagefault {
-	u32			bytes_committed;
-	u8			event_subtype;
-	enum mlx5_pagefault_flags flags;
-	union {
-		/* Initiator or send message responder pagefault details. */
-		struct {
-			/* Received packet size, only valid for responders. */
-			u32	packet_size;
-			/*
-			 * WQE index. Refers to either the send queue or
-			 * receive queue, according to event_subtype.
-			 */
-			u16	wqe_index;
-		} wqe;
-		/* RDMA responder pagefault details */
-		struct {
-			u32	r_key;
-			/*
-			 * Received packet size, minimal size page fault
-			 * resolution required for forward progress.
-			 */
-			u32	packet_size;
-			u32	rdma_op_len;
-			u64	rdma_va;
-		} rdma;
-	};
-};
-
 struct mlx5_core_qp {
 	struct mlx5_core_rsc_common	common; /* must be first */
 	void (*event)		(struct mlx5_core_qp *, int);
-	void (*pfault_handler)(struct mlx5_core_qp *, struct mlx5_pagefault *);
 	int			qpn;
 	struct mlx5_rsc_debug	*dbg;
 	int			pid;
@@ -557,10 +517,6 @@ void mlx5_init_qp_table(struct mlx5_core_dev *dev);
 void mlx5_cleanup_qp_table(struct mlx5_core_dev *dev);
 int mlx5_debug_qp_add(struct mlx5_core_dev *dev, struct mlx5_core_qp *qp);
 void mlx5_debug_qp_remove(struct mlx5_core_dev *dev, struct mlx5_core_qp *qp);
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-int mlx5_core_page_fault_resume(struct mlx5_core_dev *dev, u32 qpn,
-				u8 context, int error);
-#endif
 int mlx5_core_create_rq_tracked(struct mlx5_core_dev *dev, u32 *in, int inlen,
 				struct mlx5_core_qp *rq);
 void mlx5_core_destroy_rq_tracked(struct mlx5_core_dev *dev,