Skip to contentSkip to conversation

Use ODP MRs for kernel ULPs

PATCH In review Leon Romanovsky 5 participants 24 messages Jan 15, 2020 – Jan 18, 2020

Subsystems (15)
lore
Share & tools
lore
#0 lore raw

cover Use ODP MRs for kernel ULPs

From: Leon Romanovsky <[email protected]>

Hi,

The following series extends MR creation routines to allow creation of user MRs through kernel ULPs as a proxy. The immediate use case is to allow RDS to work over FS-DAX, which requires ODP (on-demand-paging) MRs to be created and such MRs were not possible to create prior this series.

The first part of this patchset extends RDMA to have special verb ib_reg_user_mr(). The common use case that uses this function is a userspace application that allocates memory for HCA access but the responsibility to register the memory at the HCA is on an kernel ULP. This ULP that acts as an agent for the userspace application.

The second part provides advise MR functionality for ULPs. This is integral part of ODP flows and used to trigger pagefaults in advance to prepare memory before running working set.

The third part is actual user of those in-kernel APIs.

Thanks

Hans Westgaard Ry (3):
net/rds: Detect need of On-Demand-Paging memory registration net/rds: Handle ODP mr registration/unregistration net/rds: Use prefetch for On-Demand-Paging MR

Jason Gunthorpe (1):
RDMA/mlx5: Fix handling of IOVA != user_va in ODP paths

Leon Romanovsky (1):
RDMA/mlx5: Don't fake udata for kernel path

Moni Shoua (5):
IB: Allow calls to ib_umem_get from kernel ULPs IB/core: Introduce ib_reg_user_mr
IB/core: Add interface to advise_mr for kernel users IB/mlx5: Add ODP WQE handlers for kernel QPs IB/mlx5: Mask out unsupported ODP capabilities for kernel QPs

 drivers/infiniband/core/umem.c                |  27 +--
 drivers/infiniband/core/umem_odp.c            |  29 +--
 drivers/infiniband/core/verbs.c               |  41 +++++
 drivers/infiniband/hw/bnxt_re/ib_verbs.c      |  12 +-
 drivers/infiniband/hw/cxgb4/mem.c             |   2 +-
 drivers/infiniband/hw/efa/efa_verbs.c         |   4 +-
 drivers/infiniband/hw/hns/hns_roce_cq.c       |   2 +-
 drivers/infiniband/hw/hns/hns_roce_db.c       |   3 +-
 drivers/infiniband/hw/hns/hns_roce_mr.c       |   4 +-
 drivers/infiniband/hw/hns/hns_roce_qp.c       |   2 +-
 drivers/infiniband/hw/hns/hns_roce_srq.c      |   5 +-
 drivers/infiniband/hw/i40iw/i40iw_verbs.c     |   5 +-
 drivers/infiniband/hw/mlx4/cq.c               |   2 +-
 drivers/infiniband/hw/mlx4/doorbell.c         |   3 +-
 drivers/infiniband/hw/mlx4/mr.c               |   8 +-
 drivers/infiniband/hw/mlx4/qp.c               |   5 +-
 drivers/infiniband/hw/mlx4/srq.c              |   3 +-
 drivers/infiniband/hw/mlx5/cq.c               |   6 +-
 drivers/infiniband/hw/mlx5/devx.c             |   2 +-
 drivers/infiniband/hw/mlx5/doorbell.c         |   3 +-
 drivers/infiniband/hw/mlx5/main.c             |  51 ++++--
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |  12 +-
 drivers/infiniband/hw/mlx5/mr.c               |  20 +--
 drivers/infiniband/hw/mlx5/odp.c              |  33 ++--
 drivers/infiniband/hw/mlx5/qp.c               | 167 +++++++++++-------
 drivers/infiniband/hw/mlx5/srq.c              |   2 +-
 drivers/infiniband/hw/mthca/mthca_provider.c  |   2 +-
 drivers/infiniband/hw/ocrdma/ocrdma_verbs.c   |   2 +-
 drivers/infiniband/hw/qedr/verbs.c            |   9 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c  |   2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c  |   2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c  |   7 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_srq.c |   2 +-
 drivers/infiniband/sw/rdmavt/mr.c             |   2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c            |   2 +-
 include/rdma/ib_umem.h                        |   4 +-
 include/rdma/ib_umem_odp.h                    |   6 +-
 include/rdma/ib_verbs.h                       |   9 +
 net/rds/ib.c                                  |   7 +
 net/rds/ib.h                                  |   3 +-
 net/rds/ib_mr.h                               |   7 +-
 net/rds/ib_rdma.c                             |  83 ++++++++-
 net/rds/ib_send.c                             |  44 +++--
 net/rds/rdma.c                                | 156 +++++++++++-----
 net/rds/rds.h                                 |  13 +-
 45 files changed, 559 insertions(+), 256 deletions(-)

--
2.20.1

#1 lore raw

PATCH 1/10 IB: Allow calls to ib_umem_get from kernel ULPs

From: Moni Shoua <[email protected]>

So far the assumption was that ib_umem_get() and ib_umem_odp_get() are called from flows that start in UVERBS and therefore has a user context. This assumption restricts flows that are initiated by ULPs and need the service that ib_umem_get() provides.

This patch changes ib_umem_get() and ib_umem_odp_get() to get IB device directly by relying on the fact that both UVERBS and ULPs sets that field correctly.

---
 drivers/infiniband/core/umem.c                | 27 ++++++-----------
 drivers/infiniband/core/umem_odp.c            | 29 +++++--------------
 drivers/infiniband/hw/bnxt_re/ib_verbs.c      | 12 ++++----
 drivers/infiniband/hw/cxgb4/mem.c             |  2 +-
 drivers/infiniband/hw/efa/efa_verbs.c         |  2 +-
 drivers/infiniband/hw/hns/hns_roce_cq.c       |  2 +-
 drivers/infiniband/hw/hns/hns_roce_db.c       |  3 +-
 drivers/infiniband/hw/hns/hns_roce_mr.c       |  4 +--
 drivers/infiniband/hw/hns/hns_roce_qp.c       |  2 +-
 drivers/infiniband/hw/hns/hns_roce_srq.c      |  5 ++--
 drivers/infiniband/hw/i40iw/i40iw_verbs.c     |  2 +-
 drivers/infiniband/hw/mlx4/cq.c               |  2 +-
 drivers/infiniband/hw/mlx4/doorbell.c         |  3 +-
 drivers/infiniband/hw/mlx4/mr.c               |  8 ++---
 drivers/infiniband/hw/mlx4/qp.c               |  5 ++--
 drivers/infiniband/hw/mlx4/srq.c              |  3 +-
 drivers/infiniband/hw/mlx5/cq.c               |  6 ++--
 drivers/infiniband/hw/mlx5/devx.c             |  2 +-
 drivers/infiniband/hw/mlx5/doorbell.c         |  3 +-
 drivers/infiniband/hw/mlx5/mr.c               | 18 +++++-------
 drivers/infiniband/hw/mlx5/odp.c              |  2 +-
 drivers/infiniband/hw/mlx5/qp.c               |  4 +--
 drivers/infiniband/hw/mlx5/srq.c              |  2 +-
 drivers/infiniband/hw/mthca/mthca_provider.c  |  2 +-
 drivers/infiniband/hw/ocrdma/ocrdma_verbs.c   |  2 +-
 drivers/infiniband/hw/qedr/verbs.c            |  9 +++---
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c  |  2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c  |  2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c  |  7 +++--
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_srq.c |  2 +-
 drivers/infiniband/sw/rdmavt/mr.c             |  2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c            |  2 +-
 include/rdma/ib_umem.h                        |  4 +--
 include/rdma/ib_umem_odp.h                    |  6 ++--
 34 files changed, 85 insertions(+), 103 deletions(-)
drivers/infiniband/core/umem.c +9 -18
index 7a3b99597ead..146f98fbf22b 100644
@@ -181,15 +181,14 @@ EXPORT_SYMBOL(ib_umem_find_best_pgsz);
 /**
  * ib_umem_get - Pin and DMA map userspace memory.
  *
- * @udata: userspace context to pin memory for
+ * @device: IB device to connect UMEM
  * @addr: userspace virtual address to start at
  * @size: length of region to pin
  * @access: IB_ACCESS_xxx flags for memory being pinned
  */
-struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
+struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access)
 {
-	struct ib_ucontext *context;
 	struct ib_umem *umem;
 	struct page **page_list;
 	unsigned long lock_limit;
@@ -201,14 +200,6 @@ struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
 	struct scatterlist *sg;
 	unsigned int gup_flags = FOLL_WRITE;
 
-	if (!udata)
-		return ERR_PTR(-EIO);
-
-	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
-			  ->context;
-	if (!context)
-		return ERR_PTR(-EIO);
-
 	/*
 	 * If the combination of the addr and size requested for this memory
 	 * region causes an integer overflow, return error.
@@ -226,7 +217,7 @@ struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
 	umem = kzalloc(sizeof(*umem), GFP_KERNEL);
 	if (!umem)
 		return ERR_PTR(-ENOMEM);
-	umem->ibdev = context->device;
+	umem->ibdev      = device;
 	umem->length     = size;
 	umem->address    = addr;
 	umem->writable   = ib_access_writable(access);
@@ -281,7 +272,7 @@ struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
 		npages   -= ret;
 
 		sg = ib_umem_add_sg_table(sg, page_list, ret,
-			dma_get_max_seg_size(context->device->dma_device),
+			dma_get_max_seg_size(device->dma_device),
 			&umem->sg_nents);
 
 		up_read(&mm->mmap_sem);
@@ -289,10 +280,10 @@ struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
 
 	sg_mark_end(sg);
 
-	umem->nmap = ib_dma_map_sg(context->device,
-				  umem->sg_head.sgl,
-				  umem->sg_nents,
-				  DMA_BIDIRECTIONAL);
+	umem->nmap = ib_dma_map_sg(device,
+				   umem->sg_head.sgl,
+				   umem->sg_nents,
+				   DMA_BIDIRECTIONAL);
 
 	if (!umem->nmap) {
 		ret = -ENOMEM;
@@ -303,7 +294,7 @@ struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
 	goto out;
 
 umem_release:
-	__ib_umem_release(context->device, umem, 0);
+	__ib_umem_release(device, umem, 0);
 vma:
 	atomic64_sub(ib_umem_num_pages(umem), &mm->pinned_vm);
 out:
drivers/infiniband/core/umem_odp.c +7 -22
index b9baf7d0a5cb..cb3b17a7b9b0 100644
@@ -110,15 +110,12 @@ static inline int ib_init_umem_odp(struct ib_umem_odp *umem_odp,
  * They exist only to hold the per_mm reference to help the driver create
  * children umems.
  *
- * @udata: udata from the syscall being used to create the umem
+ * @device: IB device to create UMEM
  * @access: ib_reg_mr access flags
  */
-struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_udata *udata,
+struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 					       int access)
 {
-	struct ib_ucontext *context =
-		container_of(udata, struct uverbs_attr_bundle, driver_udata)
-			->context;
 	struct ib_umem *umem;
 	struct ib_umem_odp *umem_odp;
 	int ret;
@@ -126,14 +123,11 @@ struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_udata *udata,
 	if (access & IB_ACCESS_HUGETLB)
 		return ERR_PTR(-EINVAL);
 
-	if (!context)
-		return ERR_PTR(-EIO);
-
 	umem_odp = kzalloc(sizeof(*umem_odp), GFP_KERNEL);
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 	umem = &umem_odp->umem;
-	umem->ibdev = context->device;
+	umem->ibdev = device;
 	umem->writable = ib_access_writable(access);
 	umem->owning_mm = current->mm;
 	umem_odp->is_implicit_odp = 1;
@@ -201,7 +195,7 @@ EXPORT_SYMBOL(ib_umem_odp_alloc_child);
 /**
  * ib_umem_odp_get - Create a umem_odp for a userspace va
  *
- * @udata: userspace context to pin memory for
+ * @device: IB device struct to get UMEM
  * @addr: userspace virtual address to start at
  * @size: length of region to pin
  * @access: IB_ACCESS_xxx flags for memory being pinned
@@ -210,23 +204,14 @@ EXPORT_SYMBOL(ib_umem_odp_alloc_child);
  * pinning, instead, stores the mm for future page fault handling in
  * conjunction with MMU notifiers.
  */
-struct ib_umem_odp *ib_umem_odp_get(struct ib_udata *udata, unsigned long addr,
-				    size_t size, int access,
+struct ib_umem_odp *ib_umem_odp_get(struct ib_device *device,
+				    unsigned long addr, size_t size, int access,
 				    const struct mmu_interval_notifier_ops *ops)
 {
 	struct ib_umem_odp *umem_odp;
-	struct ib_ucontext *context;
 	struct mm_struct *mm;
 	int ret;
 
-	if (!udata)
-		return ERR_PTR(-EIO);
-
-	context = container_of(udata, struct uverbs_attr_bundle, driver_udata)
-			  ->context;
-	if (!context)
-		return ERR_PTR(-EIO);
-
 	if (WARN_ON_ONCE(!(access & IB_ACCESS_ON_DEMAND)))
 		return ERR_PTR(-EINVAL);
 
@@ -234,7 +219,7 @@ struct ib_umem_odp *ib_umem_odp_get(struct ib_udata *udata, unsigned long addr,
 	if (!umem_odp)
 		return ERR_PTR(-ENOMEM);
 
-	umem_odp->umem.ibdev = context->device;
+	umem_odp->umem.ibdev = device;
 	umem_odp->umem.length = size;
 	umem_odp->umem.address = addr;
 	umem_odp->umem.writable = ib_access_writable(access);
drivers/infiniband/hw/bnxt_re/ib_verbs.c +7 -5
index 9b6ca15a183c..077f184e47e5 100644
@@ -837,7 +837,8 @@ static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
 		bytes += (qplib_qp->sq.max_wqe * psn_sz);
 	}
 	bytes = PAGE_ALIGN(bytes);
-	umem = ib_umem_get(udata, ureq.qpsva, bytes, IB_ACCESS_LOCAL_WRITE);
+	umem = ib_umem_get(&rdev->ibdev, ureq.qpsva, bytes,
+			   IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(umem))
 		return PTR_ERR(umem);
 
@@ -850,7 +851,7 @@ static int bnxt_re_init_user_qp(struct bnxt_re_dev *rdev, struct bnxt_re_pd *pd,
 	if (!qp->qplib_qp.srq) {
 		bytes = (qplib_qp->rq.max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
 		bytes = PAGE_ALIGN(bytes);
-		umem = ib_umem_get(udata, ureq.qprva, bytes,
+		umem = ib_umem_get(&rdev->ibdev, ureq.qprva, bytes,
 				   IB_ACCESS_LOCAL_WRITE);
 		if (IS_ERR(umem))
 			goto rqfail;
@@ -1304,7 +1305,8 @@ static int bnxt_re_init_user_srq(struct bnxt_re_dev *rdev,
 
 	bytes = (qplib_srq->max_wqe * BNXT_QPLIB_MAX_RQE_ENTRY_SIZE);
 	bytes = PAGE_ALIGN(bytes);
-	umem = ib_umem_get(udata, ureq.srqva, bytes, IB_ACCESS_LOCAL_WRITE);
+	umem = ib_umem_get(&rdev->ibdev, ureq.srqva, bytes,
+			   IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(umem))
 		return PTR_ERR(umem);
 
@@ -2545,7 +2547,7 @@ int bnxt_re_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 			goto fail;
 		}
 
-		cq->umem = ib_umem_get(udata, req.cq_va,
+		cq->umem = ib_umem_get(&rdev->ibdev, req.cq_va,
 				       entries * sizeof(struct cq_base),
 				       IB_ACCESS_LOCAL_WRITE);
 		if (IS_ERR(cq->umem)) {
@@ -3512,7 +3514,7 @@ struct ib_mr *bnxt_re_reg_user_mr(struct ib_pd *ib_pd, u64 start, u64 length,
 	/* The fixed portion of the rkey is the same as the lkey */
 	mr->ib_mr.rkey = mr->qplib_mr.rkey;
 
-	umem = ib_umem_get(udata, start, length, mr_access_flags);
+	umem = ib_umem_get(&rdev->ibdev, start, length, mr_access_flags);
 	if (IS_ERR(umem)) {
 		dev_err(rdev_to_dev(rdev), "Failed to get umem");
 		rc = -EFAULT;
drivers/infiniband/hw/cxgb4/mem.c +1 -1
index fe3a7e8561df..962dc97a8ff2 100644
@@ -543,7 +543,7 @@ struct ib_mr *c4iw_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	mhp->rhp = rhp;
 
-	mhp->umem = ib_umem_get(udata, start, length, acc);
+	mhp->umem = ib_umem_get(pd->device, start, length, acc);
 	if (IS_ERR(mhp->umem))
 		goto err_free_skb;
 
drivers/infiniband/hw/efa/efa_verbs.c +1 -1
index b6b936cb5c53..7e05033a650f 100644
@@ -1385,7 +1385,7 @@ struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
 		goto err_out;
 	}
 
-	mr->umem = ib_umem_get(udata, start, length, access_flags);
+	mr->umem = ib_umem_get(ibpd->device, start, length, access_flags);
 	if (IS_ERR(mr->umem)) {
 		err = PTR_ERR(mr->umem);
 		ibdev_dbg(&dev->ibdev,
drivers/infiniband/hw/hns/hns_roce_cq.c +1 -1
index 61f53a85767b..5ffe4c996ed3 100644
@@ -163,7 +163,7 @@ static int get_cq_umem(struct hns_roce_dev *hr_dev, struct hns_roce_cq *hr_cq,
 	u32 npages;
 	int ret;
 
-	*umem = ib_umem_get(udata, ucmd.buf_addr, buf->size,
+	*umem = ib_umem_get(&hr_dev->ib_dev, ucmd.buf_addr, buf->size,
 			    IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(*umem))
 		return PTR_ERR(*umem);
drivers/infiniband/hw/hns/hns_roce_db.c +2 -1
index 10af6958ab69..bff6abdccfb0 100644
@@ -31,7 +31,8 @@ int hns_roce_db_map_user(struct hns_roce_ucontext *context,
 
 	refcount_set(&page->refcount, 1);
 	page->user_virt = page_addr;
-	page->umem = ib_umem_get(udata, page_addr, PAGE_SIZE, 0);
+	page->umem = ib_umem_get(context->ibucontext.device, page_addr,
+				 PAGE_SIZE, 0);
 	if (IS_ERR(page->umem)) {
 		ret = PTR_ERR(page->umem);
 		kfree(page);
drivers/infiniband/hw/hns/hns_roce_mr.c +2 -2
index 95765560c1cf..b9898e71655a 100644
@@ -1145,7 +1145,7 @@ struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (!mr)
 		return ERR_PTR(-ENOMEM);
 
-	mr->umem = ib_umem_get(udata, start, length, access_flags);
+	mr->umem = ib_umem_get(pd->device, start, length, access_flags);
 	if (IS_ERR(mr->umem)) {
 		ret = PTR_ERR(mr->umem);
 		goto err_free;
@@ -1230,7 +1230,7 @@ static int rereg_mr_trans(struct ib_mr *ibmr, int flags,
 	}
 	ib_umem_release(mr->umem);
 
-	mr->umem = ib_umem_get(udata, start, length, mr_access_flags);
+	mr->umem = ib_umem_get(ibmr->device, start, length, mr_access_flags);
 	if (IS_ERR(mr->umem)) {
 		ret = PTR_ERR(mr->umem);
 		mr->umem = NULL;
drivers/infiniband/hw/hns/hns_roce_qp.c +1 -1
index 7c8de1e0d48a..3257ad11be48 100644
@@ -763,7 +763,7 @@ static int hns_roce_create_qp_common(struct hns_roce_dev *hr_dev,
 			goto err_alloc_rq_inline_buf;
 		}
 
-		hr_qp->umem = ib_umem_get(udata, ucmd.buf_addr,
+		hr_qp->umem = ib_umem_get(ib_pd->device, ucmd.buf_addr,
 					  hr_qp->buff_size, 0);
 		if (IS_ERR(hr_qp->umem)) {
 			dev_err(dev, "ib_umem_get error for create qp\n");
drivers/infiniband/hw/hns/hns_roce_srq.c +3 -2
index 7113ebfdb4f0..c6d5f06f9cde 100644
@@ -186,7 +186,8 @@ static int create_user_srq(struct hns_roce_srq *srq, struct ib_udata *udata,
 	if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
 		return -EFAULT;
 
-	srq->umem = ib_umem_get(udata, ucmd.buf_addr, srq_buf_size, 0);
+	srq->umem =
+		ib_umem_get(srq->ibsrq.device, ucmd.buf_addr, srq_buf_size, 0);
 	if (IS_ERR(srq->umem))
 		return PTR_ERR(srq->umem);
 
@@ -205,7 +206,7 @@ static int create_user_srq(struct hns_roce_srq *srq, struct ib_udata *udata,
 		goto err_user_srq_mtt;
 
 	/* config index queue BA */
-	srq->idx_que.umem = ib_umem_get(udata, ucmd.que_addr,
+	srq->idx_que.umem = ib_umem_get(srq->ibsrq.device, ucmd.que_addr,
 					srq->idx_que.buf_size, 0);
 	if (IS_ERR(srq->idx_que.umem)) {
 		dev_err(hr_dev->dev, "ib_umem_get error for index queue\n");
drivers/infiniband/hw/i40iw/i40iw_verbs.c +1 -1
index 86375947bc67..e75787ddc941 100644
@@ -1763,7 +1763,7 @@ static struct ib_mr *i40iw_reg_user_mr(struct ib_pd *pd,
 
 	if (length > I40IW_MAX_MR_SIZE)
 		return ERR_PTR(-EINVAL);
-	region = ib_umem_get(udata, start, length, acc);
+	region = ib_umem_get(pd->device, start, length, acc);
 	if (IS_ERR(region))
 		return (struct ib_mr *)region;
 
drivers/infiniband/hw/mlx4/cq.c +1 -1
index 72eeb9a85bc5..f8b936b76dcd 100644
@@ -144,7 +144,7 @@ static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_udata *udata,
 	int shift;
 	int n;
 
-	*umem = ib_umem_get(udata, buf_addr, cqe * cqe_size,
+	*umem = ib_umem_get(&dev->ib_dev, buf_addr, cqe * cqe_size,
 			    IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(*umem))
 		return PTR_ERR(*umem);
drivers/infiniband/hw/mlx4/doorbell.c +2 -1
index 714f9df5bf39..d41f03ccb0e1 100644
@@ -64,7 +64,8 @@ int mlx4_ib_db_map_user(struct ib_udata *udata, unsigned long virt,
 
 	page->user_virt = (virt & PAGE_MASK);
 	page->refcnt    = 0;
-	page->umem = ib_umem_get(udata, virt & PAGE_MASK, PAGE_SIZE, 0);
+	page->umem = ib_umem_get(context->ibucontext.device, virt & PAGE_MASK,
+				 PAGE_SIZE, 0);
 	if (IS_ERR(page->umem)) {
 		err = PTR_ERR(page->umem);
 		kfree(page);
drivers/infiniband/hw/mlx4/mr.c +4 -4
index dfa17bcdcdbc..b0121c90c561 100644
@@ -367,7 +367,7 @@ int mlx4_ib_umem_calc_optimal_mtt_size(struct ib_umem *umem, u64 start_va,
 	return block_shift;
 }
 
-static struct ib_umem *mlx4_get_umem_mr(struct ib_udata *udata, u64 start,
+static struct ib_umem *mlx4_get_umem_mr(struct ib_device *device, u64 start,
 					u64 length, int access_flags)
 {
 	/*
@@ -398,7 +398,7 @@ static struct ib_umem *mlx4_get_umem_mr(struct ib_udata *udata, u64 start,
 		up_read(&current->mm->mmap_sem);
 	}
 
-	return ib_umem_get(udata, start, length, access_flags);
+	return ib_umem_get(device, start, length, access_flags);
 }
 
 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
@@ -415,7 +415,7 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (!mr)
 		return ERR_PTR(-ENOMEM);
 
-	mr->umem = mlx4_get_umem_mr(udata, start, length, access_flags);
+	mr->umem = mlx4_get_umem_mr(pd->device, start, length, access_flags);
 	if (IS_ERR(mr->umem)) {
 		err = PTR_ERR(mr->umem);
 		goto err_free;
@@ -504,7 +504,7 @@ int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags,
 
 		mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr);
 		ib_umem_release(mmr->umem);
-		mmr->umem = mlx4_get_umem_mr(udata, start, length,
+		mmr->umem = mlx4_get_umem_mr(mr->device, start, length,
 					     mr_access_flags);
 		if (IS_ERR(mmr->umem)) {
 			err = PTR_ERR(mmr->umem);
drivers/infiniband/hw/mlx4/qp.c +3 -2
index 8d240bc92b6b..26425dd2d960 100644
@@ -916,7 +916,7 @@ static int create_rq(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
 	qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
 		       (qp->sq.wqe_cnt << qp->sq.wqe_shift);
 
-	qp->umem = ib_umem_get(udata, wq.buf_addr, qp->buf_size, 0);
+	qp->umem = ib_umem_get(pd->device, wq.buf_addr, qp->buf_size, 0);
 	if (IS_ERR(qp->umem)) {
 		err = PTR_ERR(qp->umem);
 		goto err;
@@ -1110,7 +1110,8 @@ static int create_qp_common(struct ib_pd *pd, struct ib_qp_init_attr *init_attr,
 		if (err)
 			goto err;
 
-		qp->umem = ib_umem_get(udata, ucmd.buf_addr, qp->buf_size, 0);
+		qp->umem =
+			ib_umem_get(pd->device, ucmd.buf_addr, qp->buf_size, 0);
 		if (IS_ERR(qp->umem)) {
 			err = PTR_ERR(qp->umem);
 			goto err;
drivers/infiniband/hw/mlx4/srq.c +2 -1
index 8dcf6e3d9ae2..8f9d5035142d 100644
@@ -110,7 +110,8 @@ int mlx4_ib_create_srq(struct ib_srq *ib_srq,
 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
 			return -EFAULT;
 
-		srq->umem = ib_umem_get(udata, ucmd.buf_addr, buf_size, 0);
+		srq->umem =
+			ib_umem_get(ib_srq->device, ucmd.buf_addr, buf_size, 0);
 		if (IS_ERR(srq->umem))
 			return PTR_ERR(srq->umem);
 
drivers/infiniband/hw/mlx5/cq.c +3 -3
index dd8d24ee8e1d..367a71bc5f4b 100644
@@ -708,8 +708,8 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 	*cqe_size = ucmd.cqe_size;
 
 	cq->buf.umem =
-		ib_umem_get(udata, ucmd.buf_addr, entries * ucmd.cqe_size,
-			    IB_ACCESS_LOCAL_WRITE);
+		ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
+			    entries * ucmd.cqe_size, IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(cq->buf.umem)) {
 		err = PTR_ERR(cq->buf.umem);
 		return err;
@@ -1108,7 +1108,7 @@ static int resize_user(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
 	if (ucmd.cqe_size && SIZE_MAX / ucmd.cqe_size <= entries - 1)
 		return -EINVAL;
 
-	umem = ib_umem_get(udata, ucmd.buf_addr,
+	umem = ib_umem_get(&dev->ib_dev, ucmd.buf_addr,
 			   (size_t)ucmd.cqe_size * entries,
 			   IB_ACCESS_LOCAL_WRITE);
 	if (IS_ERR(umem)) {
drivers/infiniband/hw/mlx5/devx.c +1 -1
index 7bb91d24d394..d7efc9f6daf0 100644
@@ -2129,7 +2129,7 @@ static int devx_umem_get(struct mlx5_ib_dev *dev, struct ib_ucontext *ucontext,
 	if (err)
 		return err;
 
-	obj->umem = ib_umem_get(&attrs->driver_udata, addr, size, access);
+	obj->umem = ib_umem_get(&dev->ib_dev, addr, size, access);
 	if (IS_ERR(obj->umem))
 		return PTR_ERR(obj->umem);
 
drivers/infiniband/hw/mlx5/doorbell.c +2 -1
index 12737c509aa2..61475b571531 100644
@@ -64,7 +64,8 @@ int mlx5_ib_db_map_user(struct mlx5_ib_ucontext *context,
 
 	page->user_virt = (virt & PAGE_MASK);
 	page->refcnt    = 0;
-	page->umem = ib_umem_get(udata, virt & PAGE_MASK, PAGE_SIZE, 0);
+	page->umem = ib_umem_get(context->ibucontext.device, virt & PAGE_MASK,
+				 PAGE_SIZE, 0);
 	if (IS_ERR(page->umem)) {
 		err = PTR_ERR(page->umem);
 		kfree(page);
drivers/infiniband/hw/mlx5/mr.c +8 -10
index 1fcae0735e01..1913e88522ec 100644
@@ -746,10 +746,9 @@ static int mr_cache_max_order(struct mlx5_ib_dev *dev)
 	return MLX5_MAX_UMR_SHIFT;
 }
 
-static int mr_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
-		       u64 start, u64 length, int access_flags,
-		       struct ib_umem **umem, int *npages, int *page_shift,
-		       int *ncont, int *order)
+static int mr_umem_get(struct mlx5_ib_dev *dev, u64 start, u64 length,
+		       int access_flags, struct ib_umem **umem, int *npages,
+		       int *page_shift, int *ncont, int *order)
 {
 	struct ib_umem *u;
 
@@ -758,7 +757,7 @@ static int mr_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 	if (access_flags & IB_ACCESS_ON_DEMAND) {
 		struct ib_umem_odp *odp;
 
-		odp = ib_umem_odp_get(udata, start, length, access_flags,
+		odp = ib_umem_odp_get(&dev->ib_dev, start, length, access_flags,
 				      &mlx5_mn_ops);
 		if (IS_ERR(odp)) {
 			mlx5_ib_dbg(dev, "umem get failed (%ld)\n",
@@ -774,7 +773,7 @@ static int mr_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 		if (order)
 			*order = ilog2(roundup_pow_of_two(*ncont));
 	} else {
-		u = ib_umem_get(udata, start, length, access_flags);
+		u = ib_umem_get(&dev->ib_dev, start, length, access_flags);
 		if (IS_ERR(u)) {
 			mlx5_ib_dbg(dev, "umem get failed (%ld)\n", PTR_ERR(u));
 			return PTR_ERR(u);
@@ -1260,7 +1259,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 		return &mr->ibmr;
 	}
 
-	err = mr_umem_get(dev, udata, start, length, access_flags, &umem,
+	err = mr_umem_get(dev, start, length, access_flags, &umem,
 			  &npages, &page_shift, &ncont, &order);
 
 	if (err < 0)
@@ -1427,9 +1426,8 @@ int mlx5_ib_rereg_user_mr(struct ib_mr *ib_mr, int flags, u64 start,
 		flags |= IB_MR_REREG_TRANS;
 		ib_umem_release(mr->umem);
 		mr->umem = NULL;
-		err = mr_umem_get(dev, udata, addr, len, access_flags,
-				  &mr->umem, &npages, &page_shift, &ncont,
-				  &order);
+		err = mr_umem_get(dev, addr, len, access_flags, &mr->umem,
+				  &npages, &page_shift, &ncont, &order);
 		if (err)
 			goto err;
 	}
drivers/infiniband/hw/mlx5/odp.c +1 -1
index 409dffb5b0c0..14265175a2d8 100644
@@ -535,7 +535,7 @@ struct mlx5_ib_mr *mlx5_ib_alloc_implicit_mr(struct mlx5_ib_pd *pd,
 	struct mlx5_ib_mr *imr;
 	int err;
 
-	umem_odp = ib_umem_odp_alloc_implicit(udata, access_flags);
+	umem_odp = ib_umem_odp_alloc_implicit(&dev->ib_dev, access_flags);
 	if (IS_ERR(umem_odp))
 		return ERR_CAST(umem_odp);
 
drivers/infiniband/hw/mlx5/qp.c +2 -2
index 425efa1349b0..a102bae6d74b 100644
@@ -749,7 +749,7 @@ static int mlx5_ib_umem_get(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 {
 	int err;
 
-	*umem = ib_umem_get(udata, addr, size, 0);
+	*umem = ib_umem_get(&dev->ib_dev, addr, size, 0);
 	if (IS_ERR(*umem)) {
 		mlx5_ib_dbg(dev, "umem_get failed\n");
 		return PTR_ERR(*umem);
@@ -806,7 +806,7 @@ static int create_user_rq(struct mlx5_ib_dev *dev, struct ib_pd *pd,
 	if (!ucmd->buf_addr)
 		return -EINVAL;
 
-	rwq->umem = ib_umem_get(udata, ucmd->buf_addr, rwq->buf_size, 0);
+	rwq->umem = ib_umem_get(&dev->ib_dev, ucmd->buf_addr, rwq->buf_size, 0);
 	if (IS_ERR(rwq->umem)) {
 		mlx5_ib_dbg(dev, "umem_get failed\n");
 		err = PTR_ERR(rwq->umem);
drivers/infiniband/hw/mlx5/srq.c +1 -1
index 62939df3c692..b1a8a9175040 100644
@@ -80,7 +80,7 @@ static int create_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq,
 
 	srq->wq_sig = !!(ucmd.flags & MLX5_SRQ_FLAG_SIGNATURE);
 
-	srq->umem = ib_umem_get(udata, ucmd.buf_addr, buf_size, 0);
+	srq->umem = ib_umem_get(pd->device, ucmd.buf_addr, buf_size, 0);
 	if (IS_ERR(srq->umem)) {
 		mlx5_ib_dbg(dev, "failed umem get, size %d\n", buf_size);
 		err = PTR_ERR(srq->umem);
drivers/infiniband/hw/mthca/mthca_provider.c +1 -1
index 33002530fee7..ac19d57803b5 100644
@@ -880,7 +880,7 @@ static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (!mr)
 		return ERR_PTR(-ENOMEM);
 
-	mr->umem = ib_umem_get(udata, start, length, acc);
+	mr->umem = ib_umem_get(pd->device, start, length, acc);
 	if (IS_ERR(mr->umem)) {
 		err = PTR_ERR(mr->umem);
 		goto err;
drivers/infiniband/hw/ocrdma/ocrdma_verbs.c +1 -1
index 9bc1ca6f6f9e..d47ea675734b 100644
@@ -869,7 +869,7 @@ struct ib_mr *ocrdma_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 	if (!mr)
 		return ERR_PTR(status);
-	mr->umem = ib_umem_get(udata, start, len, acc);
+	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
 	if (IS_ERR(mr->umem)) {
 		status = -EFAULT;
 		goto umem_err;
drivers/infiniband/hw/qedr/verbs.c +4 -5
index 115f187f0c86..484b555150e0 100644
@@ -783,7 +783,7 @@ static inline int qedr_init_user_queue(struct ib_udata *udata,
 
 	q->buf_addr = buf_addr;
 	q->buf_len = buf_len;
-	q->umem = ib_umem_get(udata, q->buf_addr, q->buf_len, access);
+	q->umem = ib_umem_get(&dev->ibdev, q->buf_addr, q->buf_len, access);
 	if (IS_ERR(q->umem)) {
 		DP_ERR(dev, "create user queue: failed ib_umem_get, got %ld\n",
 		       PTR_ERR(q->umem));
@@ -1426,9 +1426,8 @@ static int qedr_init_srq_user_params(struct ib_udata *udata,
 	if (rc)
 		return rc;
 
-	srq->prod_umem =
-		ib_umem_get(udata, ureq->prod_pair_addr,
-			    sizeof(struct rdma_srq_producers), access);
+	srq->prod_umem = ib_umem_get(srq->ibsrq.device, ureq->prod_pair_addr,
+				     sizeof(struct rdma_srq_producers), access);
 	if (IS_ERR(srq->prod_umem)) {
 		qedr_free_pbl(srq->dev, &srq->usrq.pbl_info, srq->usrq.pbl_tbl);
 		ib_umem_release(srq->usrq.umem);
@@ -2850,7 +2849,7 @@ struct ib_mr *qedr_reg_user_mr(struct ib_pd *ibpd, u64 start, u64 len,
 
 	mr->type = QEDR_MR_USER;
 
-	mr->umem = ib_umem_get(udata, start, len, acc);
+	mr->umem = ib_umem_get(ibpd->device, start, len, acc);
 	if (IS_ERR(mr->umem)) {
 		rc = -EFAULT;
 		goto err0;
drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c +1 -1
index a26a4fd86bf4..4f6cc0de7ef9 100644
@@ -135,7 +135,7 @@ int pvrdma_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 			goto err_cq;
 		}
 
-		cq->umem = ib_umem_get(udata, ucmd.buf_addr, ucmd.buf_size,
+		cq->umem = ib_umem_get(ibdev, ucmd.buf_addr, ucmd.buf_size,
 				       IB_ACCESS_LOCAL_WRITE);
 		if (IS_ERR(cq->umem)) {
 			ret = PTR_ERR(cq->umem);
drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c +1 -1
index c61e665ff261..b039f1f00e05 100644
@@ -126,7 +126,7 @@ struct ib_mr *pvrdma_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 		return ERR_PTR(-EINVAL);
 	}
 
-	umem = ib_umem_get(udata, start, length, access_flags);
+	umem = ib_umem_get(pd->device, start, length, access_flags);
 	if (IS_ERR(umem)) {
 		dev_warn(&dev->pdev->dev,
 			 "could not get umem for mem region\n");
drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c +4 -3
index f15809c28f67..9de1281f9a3b 100644
@@ -276,8 +276,9 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 
 			if (!is_srq) {
 				/* set qp->sq.wqe_cnt, shift, buf_size.. */
-				qp->rumem = ib_umem_get(udata, ucmd.rbuf_addr,
-							ucmd.rbuf_size, 0);
+				qp->rumem =
+					ib_umem_get(pd->device, ucmd.rbuf_addr,
+						    ucmd.rbuf_size, 0);
 				if (IS_ERR(qp->rumem)) {
 					ret = PTR_ERR(qp->rumem);
 					goto err_qp;
@@ -288,7 +289,7 @@ struct ib_qp *pvrdma_create_qp(struct ib_pd *pd,
 				qp->srq = to_vsrq(init_attr->srq);
 			}
 
-			qp->sumem = ib_umem_get(udata, ucmd.sbuf_addr,
+			qp->sumem = ib_umem_get(pd->device, ucmd.sbuf_addr,
 						ucmd.sbuf_size, 0);
 			if (IS_ERR(qp->sumem)) {
 				if (!is_srq)
drivers/infiniband/hw/vmw_pvrdma/pvrdma_srq.c +1 -1
index 98c8be71d91d..d330decfb80a 100644
@@ -146,7 +146,7 @@ int pvrdma_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init_attr,
 		goto err_srq;
 	}
 
-	srq->umem = ib_umem_get(udata, ucmd.buf_addr, ucmd.buf_size, 0);
+	srq->umem = ib_umem_get(ibsrq->device, ucmd.buf_addr, ucmd.buf_size, 0);
 	if (IS_ERR(srq->umem)) {
 		ret = PTR_ERR(srq->umem);
 		goto err_srq;
drivers/infiniband/sw/rdmavt/mr.c +1 -1
index b9a76bf74857..72f6534fbb52 100644
@@ -390,7 +390,7 @@ struct ib_mr *rvt_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	if (length == 0)
 		return ERR_PTR(-EINVAL);
 
-	umem = ib_umem_get(udata, start, length, mr_access_flags);
+	umem = ib_umem_get(pd->device, start, length, mr_access_flags);
 	if (IS_ERR(umem))
 		return (void *)umem;
 
drivers/infiniband/sw/rxe/rxe_mr.c +1 -1
index 35a2baf2f364..e83c7b518bfa 100644
@@ -169,7 +169,7 @@ int rxe_mem_init_user(struct rxe_pd *pd, u64 start,
 	void			*vaddr;
 	int err;
 
-	umem = ib_umem_get(udata, start, length, access);
+	umem = ib_umem_get(pd->ibpd.device, start, length, access);
 	if (IS_ERR(umem)) {
 		pr_warn("err %d from rxe_umem_get\n",
 			(int)PTR_ERR(umem));
include/rdma/ib_umem.h +2 -2
index 753f54e17e0a..e3518fd6b95b 100644
@@ -69,7 +69,7 @@ static inline size_t ib_umem_num_pages(struct ib_umem *umem)
 
 #ifdef CONFIG_INFINIBAND_USER_MEM
 
-struct ib_umem *ib_umem_get(struct ib_udata *udata, unsigned long addr,
+struct ib_umem *ib_umem_get(struct ib_device *device, unsigned long addr,
 			    size_t size, int access);
 void ib_umem_release(struct ib_umem *umem);
 int ib_umem_page_count(struct ib_umem *umem);
@@ -83,7 +83,7 @@ unsigned long ib_umem_find_best_pgsz(struct ib_umem *umem,
 
 #include <linux/err.h>
 
-static inline struct ib_umem *ib_umem_get(struct ib_udata *udata,
+static inline struct ib_umem *ib_umem_get(struct ib_device *device,
 					  unsigned long addr, size_t size,
 					  int access)
 {
include/rdma/ib_umem_odp.h +3 -3
index 81429acc8257..64314ff76612 100644
@@ -114,9 +114,9 @@ static inline size_t ib_umem_odp_num_pages(struct ib_umem_odp *umem_odp)
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 
 struct ib_umem_odp *
-ib_umem_odp_get(struct ib_udata *udata, unsigned long addr, size_t size,
+ib_umem_odp_get(struct ib_device *device, unsigned long addr, size_t size,
 		int access, const struct mmu_interval_notifier_ops *ops);
-struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_udata *udata,
+struct ib_umem_odp *ib_umem_odp_alloc_implicit(struct ib_device *device,
 					       int access);
 struct ib_umem_odp *
 ib_umem_odp_alloc_child(struct ib_umem_odp *root_umem, unsigned long addr,
@@ -134,7 +134,7 @@ void ib_umem_odp_unmap_dma_pages(struct ib_umem_odp *umem_odp, u64 start_offset,
 #else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 
 static inline struct ib_umem_odp *
-ib_umem_odp_get(struct ib_udata *udata, unsigned long addr, size_t size,
+ib_umem_odp_get(struct ib_device *device, unsigned long addr, size_t size,
 		int access, const struct mmu_interval_notifier_ops *ops)
 {
 	return ERR_PTR(-EINVAL);

--
2.20.1

#3 lore raw

PATCH 3/10 IB/core: Add interface to advise_mr for kernel users

In reply to Leon Romanovsky

From: Moni Shoua <[email protected]>

Allow ULPs to call advise_mr, so they can control ODP regions in the same way as user space applications.

---
 drivers/infiniband/core/verbs.c | 11 +++++++++++
 include/rdma/ib_verbs.h         |  3 +++
 2 files changed, 14 insertions(+)
drivers/infiniband/core/verbs.c +11 -0
index 23d9911f7365..3ebae3b65c28 100644
@@ -2023,6 +2023,17 @@ struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 }
 EXPORT_SYMBOL(ib_reg_user_mr);
 
+int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
+		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
+{
+	if (!pd->device->ops.advise_mr)
+		return -EOPNOTSUPP;
+
+	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
+					 NULL);
+}
+EXPORT_SYMBOL(ib_advise_mr);
+
 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
 {
 	struct ib_pd *pd = mr->pd;
include/rdma/ib_verbs.h +3 -0
index 1aeb92609279..1f779fad3a1e 100644
@@ -4172,6 +4172,9 @@ static inline void ib_dma_free_coherent(struct ib_device *dev,
 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 			     u64 virt_addr, int mr_access_flags);
 
+/* ib_advise_mr -  give an advice about an address range in a memory region */
+int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
+		 u32 flags, struct ib_sge *sg_list, u32 num_sge);
 /**
  * ib_dereg_mr_user - Deregisters a memory region and removes it from the
  *   HCA translation table.

--
2.20.1

#4 lore raw

PATCH 4/10 IB/mlx5: Add ODP WQE handlers for kernel QPs

In reply to Leon Romanovsky

From: Moni Shoua <[email protected]>

One of the steps in ODP page fault handler for WQEs is to read a WQE from a QP send queue or receive queue buffer at a specific index.

Since the implementation of this buffer is different between kernel and user QP the implementation of the handler needs to be aware of that and handle it in a different way.

ODP for kernel MRs is currently supported only for RDMA_READ and RDMA_WRITE operations so change the handler to
- read a WQE from a kernel QP send queue
- fail if access to receive queue or shared receive queue is required for a kernel QP

---
 drivers/infiniband/hw/mlx5/mlx5_ib.h |  12 +-
 drivers/infiniband/hw/mlx5/odp.c     |  12 +-
 drivers/infiniband/hw/mlx5/qp.c      | 163 +++++++++++++++++----------
 3 files changed, 117 insertions(+), 70 deletions(-)
drivers/infiniband/hw/mlx5/mlx5_ib.h +6 -6
index aa14d3c8abd9..7b019bd4de4b 100644
@@ -1170,12 +1170,12 @@ int mlx5_ib_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 		      const struct ib_send_wr **bad_wr);
 int mlx5_ib_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
 		      const struct ib_recv_wr **bad_wr);
-int mlx5_ib_read_user_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
-			     int buflen, size_t *bc);
-int mlx5_ib_read_user_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
-			     int buflen, size_t *bc);
-int mlx5_ib_read_user_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index,
-			      void *buffer, int buflen, size_t *bc);
+int mlx5_ib_read_wqe_sq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
+			size_t buflen, size_t *bc);
+int mlx5_ib_read_wqe_rq(struct mlx5_ib_qp *qp, int wqe_index, void *buffer,
+			size_t buflen, size_t *bc);
+int mlx5_ib_read_wqe_srq(struct mlx5_ib_srq *srq, int wqe_index, void *buffer,
+			 size_t buflen, size_t *bc);
 int mlx5_ib_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 		      struct ib_udata *udata);
 void mlx5_ib_destroy_cq(struct ib_cq *cq, struct ib_udata *udata);
drivers/infiniband/hw/mlx5/odp.c +6 -6
index 14265175a2d8..879ed9ac0af9 100644
@@ -1275,15 +1275,15 @@ static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
 	wqe = wqe_start;
 	qp = (res->res == MLX5_RES_QP) ? res_to_qp(res) : NULL;
 	if (qp && sq) {
-		ret = mlx5_ib_read_user_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
-					       &bytes_copied);
+		ret = mlx5_ib_read_wqe_sq(qp, wqe_index, wqe, PAGE_SIZE,
+					  &bytes_copied);
 		if (ret)
 			goto read_user;
 		ret = mlx5_ib_mr_initiator_pfault_handler(
 			dev, pfault, qp, &wqe, &wqe_end, bytes_copied);
 	} else if (qp && !sq) {
-		ret = mlx5_ib_read_user_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
-					       &bytes_copied);
+		ret = mlx5_ib_read_wqe_rq(qp, wqe_index, wqe, PAGE_SIZE,
+					  &bytes_copied);
 		if (ret)
 			goto read_user;
 		ret = mlx5_ib_mr_responder_pfault_handler_rq(
@@ -1291,8 +1291,8 @@ static void mlx5_ib_mr_wqe_pfault_handler(struct mlx5_ib_dev *dev,
 	} else if (!qp) {
 		struct mlx5_ib_srq *srq = res_to_srq(res);
 
-		ret = mlx5_ib_read_user_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
-						&bytes_copied);
+		ret = mlx5_ib_read_wqe_srq(srq, wqe_index, wqe, PAGE_SIZE,
+					   &bytes_copied);
 		if (ret)
 			goto read_user;
 		ret = mlx5_ib_mr_responder_pfault_handler_srq(
drivers/infiniband/hw/mlx5/qp.c +105 -58

217 lines not on this page · raw message

--
2.20.1

#2 lore raw

PATCH 2/10 IB/core: Introduce ib_reg_user_mr

In reply to Leon Romanovsky

From: Moni Shoua <[email protected]>

Add ib_reg_user_mr() for kernel ULPs to register user MRs.

The common use case that uses this function is a userspace application that allocates memory for HCA access but the responsibility to register the memory at the HCA is on an kernel ULP. This ULP that acts as an agent for the userspace application.

This function is intended to be used without a user context so vendor drivers need to be aware of calling reg_user_mr() device operation with udata equal to NULL.

Among all drivers, i40iw is the only driver which relies on presence of udata, so check udata existence for that driver.

---
 drivers/infiniband/core/verbs.c           | 30 +++++++++++++++++++++++
 drivers/infiniband/hw/efa/efa_verbs.c     |  2 +-
 drivers/infiniband/hw/i40iw/i40iw_verbs.c |  3 +++
 include/rdma/ib_verbs.h                   |  6 +++++
 4 files changed, 40 insertions(+), 1 deletion(-)
drivers/infiniband/core/verbs.c +30 -0
index 78b27aff2846..23d9911f7365 100644
@@ -1993,6 +1993,36 @@ EXPORT_SYMBOL(ib_resize_cq);
 
 /* Memory regions */
 
+struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+			     u64 virt_addr, int access_flags)
+{
+	struct ib_mr *mr;
+
+	if (access_flags & IB_ACCESS_ON_DEMAND) {
+		if (!(pd->device->attrs.device_cap_flags &
+		      IB_DEVICE_ON_DEMAND_PAGING)) {
+			pr_debug("ODP support not available\n");
+			return ERR_PTR(-EINVAL);
+		}
+	}
+
+	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
+					 access_flags, NULL);
+
+	if (IS_ERR(mr))
+		return mr;
+
+	mr->device = pd->device;
+	mr->pd = pd;
+	mr->dm = NULL;
+	atomic_inc(&pd->usecnt);
+	mr->res.type = RDMA_RESTRACK_MR;
+	rdma_restrack_kadd(&mr->res);
+
+	return mr;
+}
+EXPORT_SYMBOL(ib_reg_user_mr);
+
 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
 {
 	struct ib_pd *pd = mr->pd;
drivers/infiniband/hw/efa/efa_verbs.c +1 -1
index 7e05033a650f..74c5ed32c7c5 100644
@@ -1358,7 +1358,7 @@ struct ib_mr *efa_reg_mr(struct ib_pd *ibpd, u64 start, u64 length,
 	int inline_size;
 	int err;
 
-	if (udata->inlen &&
+	if (udata && udata->inlen &&
 	    !ib_is_udata_cleared(udata, 0, sizeof(udata->inlen))) {
 		ibdev_dbg(&dev->ibdev,
 			  "Incompatible ABI params, udata not cleared\n");
drivers/infiniband/hw/i40iw/i40iw_verbs.c +3 -0
index e75787ddc941..4072cc68aa10 100644
@@ -1758,6 +1758,9 @@ static struct ib_mr *i40iw_reg_user_mr(struct ib_pd *pd,
 	int ret;
 	int pg_shift;
 
+	if (!udata)
+		return ERR_PTR(-EOPNOTSUPP);
+
 	if (iwdev->closing)
 		return ERR_PTR(-ENODEV);
 
include/rdma/ib_verbs.h +6 -0
index 6506df9f31ae..1aeb92609279 100644
@@ -4166,6 +4166,12 @@ static inline void ib_dma_free_coherent(struct ib_device *dev,
 	dma_free_coherent(dev->dma_device, size, cpu_addr, dma_handle);
 }
 
+/* ib_reg_user_mr - register a memory region for virtual addresses from kernel
+ * space. This function should be called when 'current' is the owning MM.
+ */
+struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
+			     u64 virt_addr, int mr_access_flags);
+
 /**
  * ib_dereg_mr_user - Deregisters a memory region and removes it from the
  *   HCA translation table.

--
2.20.1

#6 lore raw

PATCH 6/10 IB/mlx5: Mask out unsupported ODP capabilities for kernel QPs

In reply to Leon Romanovsky

From: Moni Shoua <[email protected]>

The ODP handler for WQEs in RQ or SRQ is not implented for kernel QPs. Therefore don't report support in these if query comes from a kernel user.

---
 drivers/infiniband/hw/mlx5/main.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)
drivers/infiniband/hw/mlx5/main.c +17 -0
index e34531d5c806..01fc09f3ddd3 100644
@@ -1015,6 +1015,23 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		if (dev->odp_caps.general_caps & IB_ODP_SUPPORT)
 			props->device_cap_flags |= IB_DEVICE_ON_DEMAND_PAGING;
 		props->odp_caps = dev->odp_caps;
+		if (!uhw) {
+			/* ODP for kernel QPs is not implemented for receive
+			 * WQEs and SRQ WQEs
+			 */
+			props->odp_caps.per_transport_caps.rc_odp_caps &=
+				~(IB_ODP_SUPPORT_READ |
+				  IB_ODP_SUPPORT_SRQ_RECV);
+			props->odp_caps.per_transport_caps.uc_odp_caps &=
+				~(IB_ODP_SUPPORT_READ |
+				  IB_ODP_SUPPORT_SRQ_RECV);
+			props->odp_caps.per_transport_caps.ud_odp_caps &=
+				~(IB_ODP_SUPPORT_READ |
+				  IB_ODP_SUPPORT_SRQ_RECV);
+			props->odp_caps.per_transport_caps.xrc_odp_caps &=
+				~(IB_ODP_SUPPORT_READ |
+				  IB_ODP_SUPPORT_SRQ_RECV);
+		}
 	}
 
 	if (MLX5_CAP_GEN(mdev, cd))

--
2.20.1

#5 lore raw

PATCH 5/10 RDMA/mlx5: Don't fake udata for kernel path

In reply to Leon Romanovsky

From: Leon Romanovsky <[email protected]>

Kernel paths must not set udata and provide NULL pointer, instead of faking zeroed udata struct.

---
 drivers/infiniband/hw/mlx5/main.c | 34 +++++++++++++++----------------
 1 file changed, 16 insertions(+), 18 deletions(-)
drivers/infiniband/hw/mlx5/main.c +16 -18
index 5d41a2c69400..e34531d5c806 100644
@@ -815,6 +815,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 				struct ib_device_attr *props,
 				struct ib_udata *uhw)
 {
+	size_t uhw_outlen = (uhw) ? uhw->outlen : 0;
 	struct mlx5_ib_dev *dev = to_mdev(ibdev);
 	struct mlx5_core_dev *mdev = dev->mdev;
 	int err = -ENOMEM;
@@ -828,12 +829,12 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 	u64 max_tso;
 
 	resp_len = sizeof(resp.comp_mask) + sizeof(resp.response_length);
-	if (uhw->outlen && uhw->outlen < resp_len)
+	if (uhw_outlen && uhw_outlen < resp_len)
 		return -EINVAL;
 
 	resp.response_length = resp_len;
 
-	if (uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
+	if (uhw && uhw->inlen && !ib_is_udata_cleared(uhw, 0, uhw->inlen))
 		return -EINVAL;
 
 	memset(props, 0, sizeof(*props));
@@ -897,7 +898,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 			props->raw_packet_caps |=
 				IB_RAW_PACKET_CAP_CVLAN_STRIPPING;
 
-		if (field_avail(typeof(resp), tso_caps, uhw->outlen)) {
+		if (field_avail(typeof(resp), tso_caps, uhw_outlen)) {
 			max_tso = MLX5_CAP_ETH(mdev, max_lso_cap);
 			if (max_tso) {
 				resp.tso_caps.max_tso = 1 << max_tso;
@@ -907,7 +908,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 			}
 		}
 
-		if (field_avail(typeof(resp), rss_caps, uhw->outlen)) {
+		if (field_avail(typeof(resp), rss_caps, uhw_outlen)) {
 			resp.rss_caps.rx_hash_function =
 						MLX5_RX_HASH_FUNC_TOEPLITZ;
 			resp.rss_caps.rx_hash_fields_mask =
@@ -927,9 +928,9 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 			resp.response_length += sizeof(resp.rss_caps);
 		}
 	} else {
-		if (field_avail(typeof(resp), tso_caps, uhw->outlen))
+		if (field_avail(typeof(resp), tso_caps, uhw_outlen))
 			resp.response_length += sizeof(resp.tso_caps);
-		if (field_avail(typeof(resp), rss_caps, uhw->outlen))
+		if (field_avail(typeof(resp), rss_caps, uhw_outlen))
 			resp.response_length += sizeof(resp.rss_caps);
 	}
 
@@ -1054,7 +1055,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 						MLX5_MAX_CQ_PERIOD;
 	}
 
-	if (field_avail(typeof(resp), cqe_comp_caps, uhw->outlen)) {
+	if (field_avail(typeof(resp), cqe_comp_caps, uhw_outlen)) {
 		resp.response_length += sizeof(resp.cqe_comp_caps);
 
 		if (MLX5_CAP_GEN(dev->mdev, cqe_compression)) {
@@ -1072,7 +1073,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		}
 	}
 
-	if (field_avail(typeof(resp), packet_pacing_caps, uhw->outlen) &&
+	if (field_avail(typeof(resp), packet_pacing_caps, uhw_outlen) &&
 	    raw_support) {
 		if (MLX5_CAP_QOS(mdev, packet_pacing) &&
 		    MLX5_CAP_GEN(mdev, qos)) {
@@ -1091,7 +1092,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 	}
 
 	if (field_avail(typeof(resp), mlx5_ib_support_multi_pkt_send_wqes,
-			uhw->outlen)) {
+			uhw_outlen)) {
 		if (MLX5_CAP_ETH(mdev, multi_pkt_send_wqe))
 			resp.mlx5_ib_support_multi_pkt_send_wqes =
 				MLX5_IB_ALLOW_MPW;
@@ -1104,7 +1105,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 			sizeof(resp.mlx5_ib_support_multi_pkt_send_wqes);
 	}
 
-	if (field_avail(typeof(resp), flags, uhw->outlen)) {
+	if (field_avail(typeof(resp), flags, uhw_outlen)) {
 		resp.response_length += sizeof(resp.flags);
 
 		if (MLX5_CAP_GEN(mdev, cqe_compression_128))
@@ -1120,8 +1121,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		resp.flags |= MLX5_IB_QUERY_DEV_RESP_FLAGS_SCAT2CQE_DCT;
 	}
 
-	if (field_avail(typeof(resp), sw_parsing_caps,
-			uhw->outlen)) {
+	if (field_avail(typeof(resp), sw_parsing_caps, uhw_outlen)) {
 		resp.response_length += sizeof(resp.sw_parsing_caps);
 		if (MLX5_CAP_ETH(mdev, swp)) {
 			resp.sw_parsing_caps.sw_parsing_offloads |=
@@ -1141,7 +1141,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		}
 	}
 
-	if (field_avail(typeof(resp), striding_rq_caps, uhw->outlen) &&
+	if (field_avail(typeof(resp), striding_rq_caps, uhw_outlen) &&
 	    raw_support) {
 		resp.response_length += sizeof(resp.striding_rq_caps);
 		if (MLX5_CAP_GEN(mdev, striding_rq)) {
@@ -1164,8 +1164,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 		}
 	}
 
-	if (field_avail(typeof(resp), tunnel_offloads_caps,
-			uhw->outlen)) {
+	if (field_avail(typeof(resp), tunnel_offloads_caps, uhw_outlen)) {
 		resp.response_length += sizeof(resp.tunnel_offloads_caps);
 		if (MLX5_CAP_ETH(mdev, tunnel_stateless_vxlan))
 			resp.tunnel_offloads_caps |=
@@ -1186,7 +1185,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
 				MLX5_IB_TUNNELED_OFFLOADS_MPLS_UDP;
 	}
 
-	if (uhw->outlen) {
+	if (uhw_outlen) {
 		err = ib_copy_to_udata(uhw, &resp, resp.response_length);
 
 		if (err)
@@ -4790,7 +4789,6 @@ static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
 	struct ib_device_attr *dprops = NULL;
 	struct ib_port_attr *pprops = NULL;
 	int err = -ENOMEM;
-	struct ib_udata uhw = {.inlen = 0, .outlen = 0};
 
 	pprops = kzalloc(sizeof(*pprops), GFP_KERNEL);
 	if (!pprops)
@@ -4800,7 +4798,7 @@ static int __get_port_caps(struct mlx5_ib_dev *dev, u8 port)
 	if (!dprops)
 		goto out;
 
-	err = mlx5_ib_query_device(&dev->ib_dev, dprops, &uhw);
+	err = mlx5_ib_query_device(&dev->ib_dev, dprops, NULL);
 	if (err) {
 		mlx5_ib_warn(dev, "query_device failed %d\n", err);
 		goto out;

--
2.20.1

#8 lore raw

PATCH 8/10 net/rds: Detect need of On-Demand-Paging memory registration

In reply to Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

Add code to check if memory intended for RDMA is FS-DAX-memory. RDS will fail with error code EOPNOTSUPP if FS-DAX-memory is detected.

---
 net/rds/rdma.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
net/rds/rdma.c +4 -2
index 916f5ec373d8..eb23c38ce2b3 100644
@@ -156,11 +156,13 @@ void rds_rdma_drop_keys(struct rds_sock *rs)
 static int rds_pin_pages(unsigned long user_addr, unsigned int nr_pages,
 			struct page **pages, int write)
 {
+	unsigned int gup_flags = FOLL_LONGTERM;
 	int ret;
 
-	ret = get_user_pages_fast(user_addr, nr_pages, write ? FOLL_WRITE : 0,
-				  pages);
+	if (write)
+		gup_flags |= FOLL_WRITE;
 
+	ret = get_user_pages_fast(user_addr, nr_pages, gup_flags, pages);
 	if (ret >= 0 && ret < nr_pages) {
 		while (ret--)
 			put_page(pages[ret]);

--
2.20.1

#10 lore raw

PATCH 9/10 net/rds: Handle ODP mr registration/unregistration

In reply to Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr.

---
 net/rds/ib.c      |   7 +++
 net/rds/ib.h      |   3 +-
 net/rds/ib_mr.h   |   7 ++-
 net/rds/ib_rdma.c |  74 ++++++++++++++++++++++-
 net/rds/ib_send.c |  44 ++++++++++----
 net/rds/rdma.c    | 150 ++++++++++++++++++++++++++++++++++------------
 net/rds/rds.h     |  13 +++-
 7 files changed, 242 insertions(+), 56 deletions(-)
net/rds/ib.c +7 -0
index 3fd5f40189bd..a792d8a3872a 100644
@@ -156,6 +156,13 @@ static void rds_ib_add_one(struct ib_device *device)
 	has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
 		   device->ops.map_phys_fmr && device->ops.unmap_fmr);
 	rds_ibdev->use_fastreg = (has_fr && !has_fmr);
+	rds_ibdev->odp_capable =
+		!!(device->attrs.device_cap_flags &
+		   IB_DEVICE_ON_DEMAND_PAGING) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_WRITE) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_READ);
 
 	rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 	rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
net/rds/ib.h +2 -1
index 6e6f24753998..0296f1f7acda 100644
@@ -247,7 +247,8 @@ struct rds_ib_device {
 	struct ib_device	*dev;
 	struct ib_pd		*pd;
 	struct dma_pool		*rid_hdrs_pool; /* RDS headers DMA pool */
-	bool                    use_fastreg;
+	u8			use_fastreg:1;
+	u8			odp_capable:1;
 
 	unsigned int		max_mrs;
 	struct rds_ib_mr_pool	*mr_1m_pool;
net/rds/ib_mr.h +6 -1
index 9045a8c0edff..0c8252d7fe2b 100644
@@ -67,6 +67,7 @@ struct rds_ib_frmr {
 
 /* This is stored as mr->r_trans_private. */
 struct rds_ib_mr {
+	struct delayed_work		work;
 	struct rds_ib_device		*device;
 	struct rds_ib_mr_pool		*pool;
 	struct rds_ib_connection	*ic;
@@ -81,9 +82,11 @@ struct rds_ib_mr {
 	unsigned int			sg_len;
 	int				sg_dma_len;
 
+	u8				odp:1;
 	union {
 		struct rds_ib_fmr	fmr;
 		struct rds_ib_frmr	frmr;
+		struct ib_mr		*mr;
 	} u;
 };
 
@@ -122,12 +125,14 @@ void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		    struct rds_sock *rs, u32 *key_ret,
-		    struct rds_connection *conn);
+		    struct rds_connection *conn, u64 start, u64 length,
+		    int need_odp);
 void rds_ib_sync_mr(void *trans_private, int dir);
 void rds_ib_free_mr(void *trans_private, int invalidate);
 void rds_ib_flush_mrs(void);
 int rds_ib_mr_init(void);
 void rds_ib_mr_exit(void);
+u32 rds_ib_get_lkey(void *trans_private);
 
 void __rds_ib_teardown_mr(struct rds_ib_mr *);
 void rds_ib_teardown_mr(struct rds_ib_mr *);
net/rds/ib_rdma.c +73 -1

111 lines not on this page · raw message

net/rds/ib_send.c +31 -13

83 lines not on this page · raw message

net/rds/rdma.c +112 -38

243 lines not on this page · raw message

net/rds/rds.h +11 -2
index 53e86911773a..e4a603523083 100644
@@ -40,7 +40,6 @@
 #ifdef ATOMIC64_INIT
 #define KERNEL_HAS_ATOMIC64
 #endif
-
 #ifdef RDS_DEBUG
 #define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
 #else
@@ -478,6 +477,9 @@ struct rds_message {
 			struct rds_notifier	*op_notifier;
 
 			struct rds_mr		*op_rdma_mr;
+
+			u64			op_odp_addr;
+			struct rds_mr		*op_odp_mr;
 		} rdma;
 		struct rm_data_op {
 			unsigned int		op_active:1;
@@ -573,7 +575,8 @@ struct rds_transport {
 	void (*exit)(void);
 	void *(*get_mr)(struct scatterlist *sg, unsigned long nr_sg,
 			struct rds_sock *rs, u32 *key_ret,
-			struct rds_connection *conn);
+			struct rds_connection *conn,
+			u64 start, u64 length, int need_odp);
 	void (*sync_mr)(void *trans_private, int direction);
 	void (*free_mr)(void *trans_private, int invalidate);
 	void (*flush_mrs)(void);
@@ -956,6 +959,12 @@ static inline bool rds_destroy_pending(struct rds_connection *conn)
 	       (conn->c_trans->t_unloading && conn->c_trans->t_unloading(conn));
 }
 
+enum {
+	ODP_NOT_NEEDED,
+	ODP_ZEROBASED,
+	ODP_VIRTUAL
+};
+
 /* stats.c */
 DECLARE_PER_CPU_SHARED_ALIGNED(struct rds_statistics, rds_stats);
 #define rds_stats_inc_which(which, member) do {		\

--
2.20.1

#17 lore raw

PATCH 10/10 net/rds: Use prefetch for On-Demand-Paging MR

In reply to Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

Try prefetching pages when using On-Demand-Paging MR using ib_advise_mr.

---
 net/rds/ib_rdma.c | 9 +++++++++
 1 file changed, 9 insertions(+)
net/rds/ib_rdma.c +9 -0
index 5a02b313ec50..5268a76e5db7 100644
@@ -574,6 +574,7 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 			(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
 			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
 			 IB_ACCESS_ON_DEMAND);
+		struct ib_sge sge = {};
 		struct ib_mr *ib_mr;
 
 		if (!rds_ibdev->odp_capable) {
@@ -601,6 +602,14 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		}
 		ibmr->u.mr = ib_mr;
 		ibmr->odp = 1;
+
+		sge.addr = virt_addr;
+		sge.length = length;
+		sge.lkey = ib_mr->lkey;
+
+		ib_advise_mr(rds_ibdev->pd,
+			     IB_UVERBS_ADVISE_MR_ADVICE_PREFETCH_WRITE,
+			     IB_UVERBS_ADVISE_MR_FLAG_FLUSH, &sge, 1);
 		return ibmr;
 	}
 

--
2.20.1

#7 lore raw

PATCH 7/10 RDMA/mlx5: Fix handling of IOVA != user_va in ODP paths

In reply to Leon Romanovsky

From: Jason Gunthorpe <[email protected]>

Till recently it was not possible for userspace to specify a different IOVA, but with the new ibv_reg_mr_iova() library call this can be done.

To compute the user_va we must compute:
user_va = (iova - iova_start) + user_va_start

while being cautious of overflow and other math problems.

The iova is not reliably stored in the mmkey when the MR is created. Only the cached creation path (the common one) set it, so it must also be set when creating uncached MRs.

Fix the weird use of iova when computing the starting page index in the MR. In the normal case, when iova == umem.address:
iova & (~(BIT(page_shift) - 1)) == ALIGN_DOWN(umem.address, odp->page_size) == ib_umem_start(odp)

And when iova is different using it in math with a user_va is wrong.

Finally, do not allow an implicit ODP to be created with a non-zero IOVA as we have no support for that.

Fixes 7bdf65d411c1 ("IB/mlx5: Handle page faults")
---
 drivers/infiniband/hw/mlx5/mr.c  |  2 ++
 drivers/infiniband/hw/mlx5/odp.c | 19 +++++++++++++------
 2 files changed, 15 insertions(+), 6 deletions(-)
drivers/infiniband/hw/mlx5/mr.c +2 -0
index 1913e88522ec..6fa0a83c19de 100644
@@ -1249,6 +1249,8 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	if (IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && !start &&
 	    length == U64_MAX) {
+		if (virt_addr != start)
+			return ERR_PTR(-EINVAL);
 		if (!(access_flags & IB_ACCESS_ON_DEMAND) ||
 		    !(dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
 			return ERR_PTR(-EINVAL);
drivers/infiniband/hw/mlx5/odp.c +13 -6
index 879ed9ac0af9..4216814ba871 100644
@@ -662,11 +662,10 @@ static int pagefault_real_mr(struct mlx5_ib_mr *mr, struct ib_umem_odp *odp,
 	bool downgrade = flags & MLX5_PF_FLAGS_DOWNGRADE;
 	unsigned long current_seq;
 	u64 access_mask;
-	u64 start_idx, page_mask;
+	u64 start_idx;
 
 	page_shift = odp->page_shift;
-	page_mask = ~(BIT(page_shift) - 1);
-	start_idx = (user_va - (mr->mmkey.iova & page_mask)) >> page_shift;
+	start_idx = (user_va - ib_umem_start(odp)) >> page_shift;
 	access_mask = ODP_READ_ALLOWED_BIT;
 
 	if (odp->umem.writable && !downgrade)
@@ -805,11 +804,19 @@ static int pagefault_mr(struct mlx5_ib_mr *mr, u64 io_virt, size_t bcnt,
 {
 	struct ib_umem_odp *odp = to_ib_umem_odp(mr->umem);
 
+	if (unlikely(io_virt < mr->mmkey.iova))
+		return -EFAULT;
+
 	if (!odp->is_implicit_odp) {
-		if (unlikely(io_virt < ib_umem_start(odp) ||
-			     ib_umem_end(odp) - io_virt < bcnt))
+		u64 user_va;
+
+		if (check_add_overflow(io_virt - mr->mmkey.iova,
+				       (u64)odp->umem.address, &user_va))
+			return -EFAULT;
+		if (unlikely(user_va >= ib_umem_end(odp) ||
+			     ib_umem_end(odp) - user_va < bcnt))
 			return -EFAULT;
-		return pagefault_real_mr(mr, odp, io_virt, bcnt, bytes_mapped,
+		return pagefault_real_mr(mr, odp, user_va, bcnt, bytes_mapped,
 					 flags);
 	}
 	return pagefault_implicit_mr(mr, odp, io_virt, bcnt, bytes_mapped,

--
2.20.1

Acked-by on v1 8/10
#9 lore raw

Re: [PATCH mlx5-next 08/10] net/rds: Detect need of On-Demand-Paging memory registration

In reply to Leon Romanovsky
Quoted 8 lines · Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

Add code to check if memory intended for RDMA is FS-DAX-memory. RDS will fail with error code EOPNOTSUPP if FS-DAX-memory is detected.

---

Acked-by on v1 9/10
#11 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

In reply to Leon Romanovsky
Quoted 8 lines · Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr.

---

Have already reviewed this patchset on internal list. Couple of minor nits below o.w patch looks good to me.

[...]

Quoted 21 lines
net/rds/ib_rdma.c +7 -0
index c8c1e3ae8d84..5a02b313ec50 100644
@@ -37,8 +37,15 @@
 
  #include "rds_single_path.h"
  #include "ib_mr.h"
+#include "rds.h"
 
  struct workqueue_struct *rds_ib_mr_wq;
+struct rds_ib_dereg_odp_mr {
+	struct work_struct work;
+	struct ib_mr *mr;
+};
+
+static void rds_ib_odp_mr_worker(struct work_struct *work);
 
  static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
  {

net/rds/ib_rdma.c void rds_ib_sync_mr(void *trans_private, int direction) · line 224

Quoted 5 lines
net/rds/ib_rdma.c +2 -0
@@ -213,6 +220,8 @@ void rds_ib_sync_mr(void *trans_private, int direction)
  	struct rds_ib_mr *ibmr = trans_private;
  	struct rds_ib_device *rds_ibdev = ibmr->device;
 
+	if (ibmr->odp)
+		return;

Add a new line here.

Quoted 3 lines
net/rds/ib_rdma.c +0 -0
  	switch (direction) {
  	case DMA_FROM_DEVICE:
  		ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,

[...]

Quoted 22 lines
net/rds/rdma.c +3 -2
index eb23c38ce2b3..3c6afdda709b 100644
@@ -177,13 +177,14 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
  			  struct rds_conn_path *cp)
  {
  	struct rds_mr *mr = NULL, *found;
+	struct scatterlist *sg = NULL;
  	unsigned int nr_pages;
  	struct page **pages = NULL;
-	struct scatterlist *sg;
  	void *trans_private;
  	unsigned long flags;
  	rds_rdma_cookie_t cookie;
-	unsigned int nents;
+	unsigned int nents = 0;
+	int need_odp = 0;
  	long i;
  	int ret;
 

net/rds/rdma.c static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args, · line 199

Quoted 3 lines
net/rds/rdma.c +0 -0
@@ -196,6 +197,20 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
  		ret = -EOPNOTSUPP;
  		goto out;
  	}

New line pls

#19 lore raw
In reply to Leon Romanovsky
Quoted 80 lines · Leon Romanovsky

From: Leon Romanovsky <[email protected]>

Hi,

The following series extends MR creation routines to allow creation of user MRs through kernel ULPs as a proxy. The immediate use case is to allow RDS to work over FS-DAX, which requires ODP (on-demand-paging) MRs to be created and such MRs were not possible to create prior this series.

The first part of this patchset extends RDMA to have special verb ib_reg_user_mr(). The common use case that uses this function is a userspace application that allocates memory for HCA access but the responsibility to register the memory at the HCA is on an kernel ULP. This ULP that acts as an agent for the userspace application.

The second part provides advise MR functionality for ULPs. This is integral part of ODP flows and used to trigger pagefaults in advance to prepare memory before running working set.

The third part is actual user of those in-kernel APIs.

Thanks

Hans Westgaard Ry (3):
net/rds: Detect need of On-Demand-Paging memory registration net/rds: Handle ODP mr registration/unregistration net/rds: Use prefetch for On-Demand-Paging MR

Jason Gunthorpe (1):
RDMA/mlx5: Fix handling of IOVA != user_va in ODP paths

Leon Romanovsky (1):
RDMA/mlx5: Don't fake udata for kernel path

Moni Shoua (5):
IB: Allow calls to ib_umem_get from kernel ULPs IB/core: Introduce ib_reg_user_mr
IB/core: Add interface to advise_mr for kernel users IB/mlx5: Add ODP WQE handlers for kernel QPs IB/mlx5: Mask out unsupported ODP capabilities for kernel QPs

 drivers/infiniband/core/umem.c                |  27 +--
 drivers/infiniband/core/umem_odp.c            |  29 +--
 drivers/infiniband/core/verbs.c               |  41 +++++
 drivers/infiniband/hw/bnxt_re/ib_verbs.c      |  12 +-
 drivers/infiniband/hw/cxgb4/mem.c             |   2 +-
 drivers/infiniband/hw/efa/efa_verbs.c         |   4 +-
 drivers/infiniband/hw/hns/hns_roce_cq.c       |   2 +-
 drivers/infiniband/hw/hns/hns_roce_db.c       |   3 +-
 drivers/infiniband/hw/hns/hns_roce_mr.c       |   4 +-
 drivers/infiniband/hw/hns/hns_roce_qp.c       |   2 +-
 drivers/infiniband/hw/hns/hns_roce_srq.c      |   5 +-
 drivers/infiniband/hw/i40iw/i40iw_verbs.c     |   5 +-
 drivers/infiniband/hw/mlx4/cq.c               |   2 +-
 drivers/infiniband/hw/mlx4/doorbell.c         |   3 +-
 drivers/infiniband/hw/mlx4/mr.c               |   8 +-
 drivers/infiniband/hw/mlx4/qp.c               |   5 +-
 drivers/infiniband/hw/mlx4/srq.c              |   3 +-
 drivers/infiniband/hw/mlx5/cq.c               |   6 +-
 drivers/infiniband/hw/mlx5/devx.c             |   2 +-
 drivers/infiniband/hw/mlx5/doorbell.c         |   3 +-
 drivers/infiniband/hw/mlx5/main.c             |  51 ++++--
 drivers/infiniband/hw/mlx5/mlx5_ib.h          |  12 +-
 drivers/infiniband/hw/mlx5/mr.c               |  20 +--
 drivers/infiniband/hw/mlx5/odp.c              |  33 ++--
 drivers/infiniband/hw/mlx5/qp.c               | 167 +++++++++++-------
 drivers/infiniband/hw/mlx5/srq.c              |   2 +-
 drivers/infiniband/hw/mthca/mthca_provider.c  |   2 +-
 drivers/infiniband/hw/ocrdma/ocrdma_verbs.c   |   2 +-
 drivers/infiniband/hw/qedr/verbs.c            |   9 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_cq.c  |   2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_mr.c  |   2 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_qp.c  |   7 +-
 drivers/infiniband/hw/vmw_pvrdma/pvrdma_srq.c |   2 +-
 drivers/infiniband/sw/rdmavt/mr.c             |   2 +-
 drivers/infiniband/sw/rxe/rxe_mr.c            |   2 +-
 include/rdma/ib_umem.h                        |   4 +-
 include/rdma/ib_umem_odp.h                    |   6 +-
 include/rdma/ib_verbs.h                       |   9 +
Quoted 8 lines
 net/rds/ib.c                                  |   7 +
 net/rds/ib.h                                  |   3 +-
 net/rds/ib_mr.h                               |   7 +-
 net/rds/ib_rdma.c                             |  83 ++++++++-
 net/rds/ib_send.c                             |  44 +++--
 net/rds/rdma.c                                | 156 +++++++++++-----
 net/rds/rds.h                                 |  13 +-
 45 files changed, 559 insertions(+), 256 deletions(-)

Thanks Santosh for your review.

David,
Is it ok to route those patches through RDMA tree given the fact that we are touching a lot of files in drivers/infiniband/* ?

There is no conflict between netdev and RDMA versions of RDS, but to be on safe side, I'll put all this code to mlx5-next tree.

Thanks

Quoted 4 lines

--
2.20.1

#12 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

In reply to [email protected]
Quoted 6 lines · On Wed, Jan 15, 2020 at 01:51:23PM -0800, [email protected] wrote:
Quoted 5 lines · Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr.

Quoted 8 lines
Quoted 3 lines

---

Have already reviewed this patchset on internal list. Couple of minor nits below o.w patch looks good to me.

Thanks Santosh, Once, I'll figure the apply path for this series, I will add extra lines while applying the patches.

#13 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

Quoted 11 lines · Leon Romanovsky
Quoted 10 lines · On Wed, Jan 15, 2020 at 01:51:23PM -0800, [email protected] wrote:
Quoted 8 lines · Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr.

---

Quoted 8 lines
Quoted 4 lines

Have already reviewed this patchset on internal list. Couple of minor nits below o.w patch looks good to me.

Thanks Santosh, Once, I'll figure the apply path for this series, I will add extra lines while applying the patches.

Sure. Thanks for picking it up !!

Regards,
Santosh

#20 lore raw
In reply to Leon Romanovsky
Quoted 10 lines · Leon Romanovsky
Quoted 1 line
 45 files changed, 559 insertions(+), 256 deletions(-)

Thanks Santosh for your review.

David,
Is it ok to route those patches through RDMA tree given the fact that we are touching a lot of files in drivers/infiniband/* ?

There is no conflict between netdev and RDMA versions of RDS, but to be on safe side, I'll put all this code to mlx5-next tree.

Er, lets not contaminate the mlx5-next with this..

It looks like it applies clean to -rc6 so if it has to be in both trees a clean PR against -rc5/6 is the way to do it.

Santos, do you anticipate more RDS patches this cycle?

Jason

#21 lore raw
Quoted 9 lines · Jason Gunthorpe
Quoted 8 lines · Leon Romanovsky
Quoted 1 line
 45 files changed, 559 insertions(+), 256 deletions(-)

Thanks Santosh for your review.

David,
Is it ok to route those patches through RDMA tree given the fact that we are touching a lot of files in drivers/infiniband/* ?

Quoted 7 lines
Quoted 2 lines

There is no conflict between netdev and RDMA versions of RDS, but to be on safe side, I'll put all this code to mlx5-next tree.

Er, lets not contaminate the mlx5-next with this..

It looks like it applies clean to -rc6 so if it has to be in both trees a clean PR against -rc5/6 is the way to do it.

Yes, it applies cleanly.

Quoted 4 lines

Santos, do you anticipate more RDS patches this cycle?

Jason

#22 lore raw
In reply to Jason Gunthorpe
Quoted 12 lines · Jason Gunthorpe
Quoted 10 lines · Leon Romanovsky
Quoted 1 line
  45 files changed, 559 insertions(+), 256 deletions(-)

Thanks Santosh for your review.

David,
Is it ok to route those patches through RDMA tree given the fact that we are touching a lot of files in drivers/infiniband/* ?

There is no conflict between netdev and RDMA versions of RDS, but to be on safe side, I'll put all this code to mlx5-next tree.

Quoted 7 lines

Er, lets not contaminate the mlx5-next with this..

It looks like it applies clean to -rc6 so if it has to be in both trees a clean PR against -rc5/6 is the way to do it.

Santos, do you anticipate more RDS patches this cycle?

Not for upcoming merge window afaik.

#23 lore raw
Quoted 15 lines · On Thu, Jan 16, 2020 at 11:34:18AM -0800, [email protected] wrote:
Quoted 14 lines · Jason Gunthorpe
Quoted 10 lines · Leon Romanovsky
Quoted 1 line
  45 files changed, 559 insertions(+), 256 deletions(-)

Thanks Santosh for your review.

David,
Is it ok to route those patches through RDMA tree given the fact that we are touching a lot of files in drivers/infiniband/* ?

There is no conflict between netdev and RDMA versions of RDS, but to be on safe side, I'll put all this code to mlx5-next tree.

Er, lets not contaminate the mlx5-next with this..

Quoted 7 lines
Quoted 5 lines

It looks like it applies clean to -rc6 so if it has to be in both trees a clean PR against -rc5/6 is the way to do it.

Santos, do you anticipate more RDS patches this cycle?

Not for upcoming merge window afaik.

In this case DaveM, will you ack and we can take it through RDMA?

The RDMA pieces look OK to me, like Santos I have reviewed many versions of this already..

Thanks,
Jason

#14 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

In reply to Leon Romanovsky
Thank you for the patch! Yet something to improve:

Hi Leon,

Thank you for the patch! Yet something to improve:

[auto build test ERROR on rdma/for-next]
[also build test ERROR on net-next/master net/master linus/master v5.5-rc6 next-20200117] [cannot apply to sparc-next/master]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system. BTW, we also suggest to use '--base' option to specify the base tree in git format-patch, please see https://stackoverflow.com/a/37406982]

url:    https://github.com/0day-ci/linux/commits/Leon-Romanovsky/Use-ODP-MRs-for-kernel-ULPs/20200116-234233
base:   https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git for-next
config: c6x-allyesconfig (attached as .config)
compiler: c6x-elf-gcc (GCC) 7.5.0

reproduce:
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
GCC_VERSION=7.5.0 make.cross ARCH=c6x

If you fix the issue, kindly add following tag

All errors (new ones prefixed by >>):

net///rds/rdma.c: In function 'rds_cmsg_rdma_args':

>> net///rds/rdma.c:767:8: error: 'struct scatterlist' has no member named 'dma_length'; did you mean 'length'?
       sg->dma_length = sg->length;
           ^~~~~~~~~~
           length

vim +767 net///rds/rdma.c

600
601 /*
602 * The application asks for a RDMA transfer.
603 * Extract all arguments and set up the rdma_op
604 */
605 int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
606 struct cmsghdr *cmsg,
607 struct rds_iov_vector *vec)
608 {
609 struct rds_rdma_args *args;
610 struct rm_rdma_op *op = &rm->rdma;
611 int nr_pages;
612 unsigned int nr_bytes;
613 struct page **pages = NULL;
614 struct rds_iovec *iovs;
615 unsigned int i, j;
616 int ret = 0;
617 bool odp_supported = true;
618
619 if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
620 || rm->rdma.op_active)
621 return -EINVAL;
622
623 args = CMSG_DATA(cmsg);
624
625 if (ipv6_addr_any(&rs->rs_bound_addr)) {
626 ret = -ENOTCONN; /* XXX not a great errno */
627 goto out_ret;
628 }
629
630 if (args->nr_local > UIO_MAXIOV) {
631 ret = -EMSGSIZE;
632 goto out_ret;
633 }
634
635 if (vec->len != args->nr_local) {
636 ret = -EINVAL;
637 goto out_ret;
638 }
639 /* odp-mr is not supported for multiple requests within one message */
640 if (args->nr_local != 1)
641 odp_supported = false;
642
643 iovs = vec->iov;
644
645 nr_pages = rds_rdma_pages(iovs, args->nr_local);
646 if (nr_pages < 0) {
647 ret = -EINVAL;
648 goto out_ret;
649 }
650
651 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_KERNEL);
652 if (!pages) {
653 ret = -ENOMEM;
654 goto out_ret;
655 }
656
657 op->op_write = !!(args->flags & RDS_RDMA_READWRITE);
658 op->op_fence = !!(args->flags & RDS_RDMA_FENCE);
659 op->op_notify = !!(args->flags & RDS_RDMA_NOTIFY_ME);
660 op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
661 op->op_active = 1;
662 op->op_recverr = rs->rs_recverr;
663 op->op_odp_mr = NULL;
664
665 WARN_ON(!nr_pages);
666 op->op_sg = rds_message_alloc_sgs(rm, nr_pages, &ret);
667 if (!op->op_sg)
668 goto out_pages;
669
670 if (op->op_notify || op->op_recverr) {
671 /* We allocate an uninitialized notifier here, because
672 * we don't want to do that in the completion handler. We
673 * would have to use GFP_ATOMIC there, and don't want to deal
674 * with failed allocations.
675 */
676 op->op_notifier = kmalloc(sizeof(struct rds_notifier), GFP_KERNEL);
677 if (!op->op_notifier) {
678 ret = -ENOMEM;
679 goto out_pages;
680 }
681 op->op_notifier->n_user_token = args->user_token;
682 op->op_notifier->n_status = RDS_RDMA_SUCCESS;
683 }
684
685 /* The cookie contains the R_Key of the remote memory region, and
686 * optionally an offset into it. This is how we implement RDMA into
687 * unaligned memory.
688 * When setting up the RDMA, we need to add that offset to the
689 * destination address (which is really an offset into the MR)
690 * FIXME: We may want to move this into ib_rdma.c
691 */
692 op->op_rkey = rds_rdma_cookie_key(args->cookie);
693 op->op_remote_addr = args->remote_vec.addr + rds_rdma_cookie_offset(args->cookie);
694
695 nr_bytes = 0;
696
697 rdsdebug("RDS: rdma prepare nr_local %llu rva %llx rkey %x\n",

   698		       (unsigned long long)args->nr_local,
   699		       (unsigned long long)args->remote_vec.addr,
   700		       op->op_rkey);
   701	
   702		for (i = 0; i < args->nr_local; i++) {
   703			struct rds_iovec *iov = &iovs[i];
   704			/* don't need to check, rds_rdma_pages() verified nr will be +nonzero */
   705			unsigned int nr = rds_pages_in_vec(iov);
   706	
   707			rs->rs_user_addr = iov->addr;
   708			rs->rs_user_bytes = iov->bytes;
   709	
   710			/* If it's a WRITE operation, we want to pin the pages for reading.
   711			 * If it's a READ operation, we need to pin the pages for writing.
   712			 */
   713			ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
   714			if ((!odp_supported && ret <= 0) ||
   715			    (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
   716				goto out_pages;
   717	
   718			if (ret == -EOPNOTSUPP) {
   719				struct rds_mr *local_odp_mr;
   720	
   721				if (!rs->rs_transport->get_mr) {
   722					ret = -EOPNOTSUPP;
   723					goto out_pages;
   724				}
   725				local_odp_mr =
   726					kzalloc(sizeof(*local_odp_mr), GFP_KERNEL);
   727				if (!local_odp_mr) {
   728					ret = -ENOMEM;
   729					goto out_pages;
   730				}
   731				RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
   732				refcount_set(&local_odp_mr->r_refcount, 1);
   733				local_odp_mr->r_trans = rs->rs_transport;
   734				local_odp_mr->r_sock = rs;
   735				local_odp_mr->r_trans_private =
   736					rs->rs_transport->get_mr(
   737						NULL, 0, rs, &local_odp_mr->r_key, NULL,
   738						iov->addr, iov->bytes, ODP_VIRTUAL);
   739				if (IS_ERR(local_odp_mr->r_trans_private)) {
   740					ret = IS_ERR(local_odp_mr->r_trans_private);
   741					rdsdebug("get_mr ret %d %p\"", ret,
   742						 local_odp_mr->r_trans_private);
   743					kfree(local_odp_mr);
   744					ret = -EOPNOTSUPP;
   745					goto out_pages;
   746				}
   747				rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
   748					 local_odp_mr, local_odp_mr->r_trans_private);
   749				op->op_odp_mr = local_odp_mr;
   750				op->op_odp_addr = iov->addr;
   751			}
   752	
   753			rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
   754				 nr_bytes, nr, iov->bytes, iov->addr);
   755	
   756			nr_bytes += iov->bytes;
   757	
   758			for (j = 0; j < nr; j++) {
   759				unsigned int offset = iov->addr & ~PAGE_MASK;
   760				struct scatterlist *sg;
   761	
   762				sg = &op->op_sg[op->op_nents + j];
   763				sg_set_page(sg, pages[j],
   764						min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
   765						offset);
   766	
 > 767				sg->dma_length = sg->length;

768 rdsdebug("RDS: sg->offset %x sg->len %x iov->addr %llx iov->bytes %llu\n",
769 sg->offset, sg->length, iov->addr, iov->bytes);
770
771 iov->addr += sg->length;
772 iov->bytes -= sg->length;
773 }
774
775 op->op_nents += nr;
776 }
777
778 if (nr_bytes > args->remote_vec.bytes) {
779 rdsdebug("RDS nr_bytes %u remote_bytes %u do not match\n",
780 nr_bytes,
781 (unsigned int) args->remote_vec.bytes);
782 ret = -EINVAL;
783 goto out_pages;
784 }
785 op->op_bytes = nr_bytes;
786 ret = 0;
787
788 out_pages:
789 kfree(pages);
790 out_ret:
791 if (ret)
792 rds_rdma_free_op(op);
793 else
794 rds_stats_inc(s_send_rdma);
795
796 return ret;
797 }
798

---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org Intel Corporation

#15 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

Quoted 26 lines · kbuild test robot

Hi Leon,

Thank you for the patch! Yet something to improve:

[auto build test ERROR on rdma/for-next]
[also build test ERROR on net-next/master net/master linus/master v5.5-rc6 next-20200117] [cannot apply to sparc-next/master]
[if your patch is applied to the wrong git tree, please drop us a note to help improve the system. BTW, we also suggest to use '--base' option to specify the base tree in git format-patch, please see https://stackoverflow.com/a/37406982]

url:    https://github.com/0day-ci/linux/commits/Leon-Romanovsky/Use-ODP-MRs-for-kernel-ULPs/20200116-234233
base:   https://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma.git for-next
config: c6x-allyesconfig (attached as .config)
compiler: c6x-elf-gcc (GCC) 7.5.0

reproduce:
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# save the attached .config to linux build tree
GCC_VERSION=7.5.0 make.cross ARCH=c6x

If you fix the issue, kindly add following tag

All errors (new ones prefixed by >>):

Quoted 7 lines

net///rds/rdma.c: In function 'rds_cmsg_rdma_args':

Quoted 1 line
Quoted 1 line

net///rds/rdma.c:767:8: error: 'struct scatterlist' has no member named 'dma_length'; did you mean 'length'?

       sg->dma_length = sg->length;
           ^~~~~~~~~~
           length

vim +767 net///rds/rdma.c

Thanks this line should be "sg_dma_len(sg) = sg->length;"

Thanks

#16 lore raw

Re: [PATCH mlx5-next 09/10] net/rds: Handle ODP mr registration/unregistration

In reply to Leon Romanovsky
Quoted 536 lines · Leon Romanovsky

From: Hans Westgaard Ry <[email protected]>

On-Demand-Paging MRs are registered using ib_reg_user_mr and unregistered with ib_dereg_mr.

---
 net/rds/ib.c      |   7 +++
 net/rds/ib.h      |   3 +-
 net/rds/ib_mr.h   |   7 ++-
 net/rds/ib_rdma.c |  74 ++++++++++++++++++++++-
 net/rds/ib_send.c |  44 ++++++++++----
 net/rds/rdma.c    | 150 ++++++++++++++++++++++++++++++++++------------
 net/rds/rds.h     |  13 +++-
 7 files changed, 242 insertions(+), 56 deletions(-)
net/rds/ib.c +7 -0
index 3fd5f40189bd..a792d8a3872a 100644
@@ -156,6 +156,13 @@ static void rds_ib_add_one(struct ib_device *device)
 	has_fmr = (device->ops.alloc_fmr && device->ops.dealloc_fmr &&
 		   device->ops.map_phys_fmr && device->ops.unmap_fmr);
 	rds_ibdev->use_fastreg = (has_fr && !has_fmr);
+	rds_ibdev->odp_capable =
+		!!(device->attrs.device_cap_flags &
+		   IB_DEVICE_ON_DEMAND_PAGING) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_WRITE) &&
+		!!(device->attrs.odp_caps.per_transport_caps.rc_odp_caps &
+		   IB_ODP_SUPPORT_READ);
 
 	rds_ibdev->fmr_max_remaps = device->attrs.max_map_per_fmr?: 32;
 	rds_ibdev->max_1m_mrs = device->attrs.max_mr ?
net/rds/ib.h +2 -1
index 6e6f24753998..0296f1f7acda 100644
@@ -247,7 +247,8 @@ struct rds_ib_device {
 	struct ib_device	*dev;
 	struct ib_pd		*pd;
 	struct dma_pool		*rid_hdrs_pool; /* RDS headers DMA pool */
-	bool                    use_fastreg;
+	u8			use_fastreg:1;
+	u8			odp_capable:1;
 
 	unsigned int		max_mrs;
 	struct rds_ib_mr_pool	*mr_1m_pool;
net/rds/ib_mr.h +6 -1
index 9045a8c0edff..0c8252d7fe2b 100644
@@ -67,6 +67,7 @@ struct rds_ib_frmr {
 
 /* This is stored as mr->r_trans_private. */
 struct rds_ib_mr {
+	struct delayed_work		work;
 	struct rds_ib_device		*device;
 	struct rds_ib_mr_pool		*pool;
 	struct rds_ib_connection	*ic;
@@ -81,9 +82,11 @@ struct rds_ib_mr {
 	unsigned int			sg_len;
 	int				sg_dma_len;
 
+	u8				odp:1;
 	union {
 		struct rds_ib_fmr	fmr;
 		struct rds_ib_frmr	frmr;
+		struct ib_mr		*mr;
 	} u;
 };
 
@@ -122,12 +125,14 @@ void rds6_ib_get_mr_info(struct rds_ib_device *rds_ibdev,
 void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		    struct rds_sock *rs, u32 *key_ret,
-		    struct rds_connection *conn);
+		    struct rds_connection *conn, u64 start, u64 length,
+		    int need_odp);
 void rds_ib_sync_mr(void *trans_private, int dir);
 void rds_ib_free_mr(void *trans_private, int invalidate);
 void rds_ib_flush_mrs(void);
 int rds_ib_mr_init(void);
 void rds_ib_mr_exit(void);
+u32 rds_ib_get_lkey(void *trans_private);
 
 void __rds_ib_teardown_mr(struct rds_ib_mr *);
 void rds_ib_teardown_mr(struct rds_ib_mr *);
net/rds/ib_rdma.c +73 -1
index c8c1e3ae8d84..5a02b313ec50 100644
@@ -37,8 +37,15 @@
 
 #include "rds_single_path.h"
 #include "ib_mr.h"
+#include "rds.h"
 
 struct workqueue_struct *rds_ib_mr_wq;
+struct rds_ib_dereg_odp_mr {
+	struct work_struct work;
+	struct ib_mr *mr;
+};
+
+static void rds_ib_odp_mr_worker(struct work_struct *work);
 
 static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
 {
@@ -213,6 +220,8 @@ void rds_ib_sync_mr(void *trans_private, int direction)
 	struct rds_ib_mr *ibmr = trans_private;
 	struct rds_ib_device *rds_ibdev = ibmr->device;
 
+	if (ibmr->odp)
+		return;
 	switch (direction) {
 	case DMA_FROM_DEVICE:
 		ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
@@ -482,6 +491,16 @@ void rds_ib_free_mr(void *trans_private, int invalidate)
 
 	rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
 
+	if (ibmr->odp) {
+		/* A MR created and marked as use_once. We use delayed work,
+		 * because there is a change that we are in interrupt and can't
+		 * call to ib_dereg_mr() directly.
+		 */
+		INIT_DELAYED_WORK(&ibmr->work, rds_ib_odp_mr_worker);
+		queue_delayed_work(rds_ib_mr_wq, &ibmr->work, 0);
+		return;
+	}
+
 	/* Return it to the pool's free list */
 	if (rds_ibdev->use_fastreg)
 		rds_ib_free_frmr_list(ibmr);
@@ -526,9 +545,17 @@ void rds_ib_flush_mrs(void)
 	up_read(&rds_ib_devices_lock);
 }
 
+u32 rds_ib_get_lkey(void *trans_private)
+{
+	struct rds_ib_mr *ibmr = trans_private;
+
+	return ibmr->u.mr->lkey;
+}
+
 void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		    struct rds_sock *rs, u32 *key_ret,
-		    struct rds_connection *conn)
+		    struct rds_connection *conn,
+		    u64 start, u64 length, int need_odp)
 {
 	struct rds_ib_device *rds_ibdev;
 	struct rds_ib_mr *ibmr = NULL;
@@ -541,6 +568,42 @@ void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
 		goto out;
 	}
 
+	if (need_odp == ODP_ZEROBASED || need_odp == ODP_VIRTUAL) {
+		u64 virt_addr = need_odp == ODP_ZEROBASED ? 0 : start;
+		int access_flags =
+			(IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
+			 IB_ACCESS_REMOTE_WRITE | IB_ACCESS_REMOTE_ATOMIC |
+			 IB_ACCESS_ON_DEMAND);
+		struct ib_mr *ib_mr;
+
+		if (!rds_ibdev->odp_capable) {
+			ret = -EOPNOTSUPP;
+			goto out;
+		}
+
+		ib_mr = ib_reg_user_mr(rds_ibdev->pd, start, length, virt_addr,
+				       access_flags);
+
+		if (IS_ERR(ib_mr)) {
+			rdsdebug("rds_ib_get_user_mr returned %d\n",
+				 IS_ERR(ib_mr));
+			ret = PTR_ERR(ib_mr);
+			goto out;
+		}
+		if (key_ret)
+			*key_ret = ib_mr->rkey;
+
+		ibmr = kzalloc(sizeof(*ibmr), GFP_KERNEL);
+		if (!ibmr) {
+			ib_dereg_mr(ib_mr);
+			ret = -ENOMEM;
+			goto out;
+		}
+		ibmr->u.mr = ib_mr;
+		ibmr->odp = 1;
+		return ibmr;
+	}
+
 	if (conn)
 		ic = conn->c_transport_data;
 
@@ -629,3 +692,12 @@ void rds_ib_mr_exit(void)
 {
 	destroy_workqueue(rds_ib_mr_wq);
 }
+
+static void rds_ib_odp_mr_worker(struct work_struct  *work)
+{
+	struct rds_ib_mr *ibmr;
+
+	ibmr = container_of(work, struct rds_ib_mr, work.work);
+	ib_dereg_mr(ibmr->u.mr);
+	kfree(ibmr);
+}
net/rds/ib_send.c +31 -13
index d1cc1d7778d8..dfe778220657 100644
@@ -39,6 +39,7 @@
 #include "rds_single_path.h"
 #include "rds.h"
 #include "ib.h"
+#include "ib_mr.h"
 
 /*
  * Convert IB-specific error message to RDS error message and call core
@@ -635,6 +636,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 		send->s_sge[0].addr = ic->i_send_hdrs_dma[pos];
 
 		send->s_sge[0].length = sizeof(struct rds_header);
+		send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
 
 		memcpy(ic->i_send_hdrs[pos], &rm->m_inc.i_hdr,
 		       sizeof(struct rds_header));
@@ -650,6 +652,7 @@ int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
 			send->s_sge[1].addr = sg_dma_address(scat);
 			send->s_sge[1].addr += rm->data.op_dmaoff;
 			send->s_sge[1].length = len;
+			send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
 
 			bytes_sent += len;
 			rm->data.op_dmaoff += len;
@@ -858,20 +861,29 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 	int ret;
 	int num_sge;
 	int nr_sig = 0;
+	u64 odp_addr = op->op_odp_addr;
+	u32 odp_lkey = 0;
 
 	/* map the op the first time we see it */
-	if (!op->op_mapped) {
-		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
-					     op->op_sg, op->op_nents, (op->op_write) ?
-					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
-		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
-		if (op->op_count == 0) {
-			rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
-			ret = -ENOMEM; /* XXX ? */
-			goto out;
+	if (!op->op_odp_mr) {
+		if (!op->op_mapped) {
+			op->op_count =
+				ib_dma_map_sg(ic->i_cm_id->device, op->op_sg,
+					      op->op_nents,
+					      (op->op_write) ? DMA_TO_DEVICE :
+							       DMA_FROM_DEVICE);
+			rdsdebug("ic %p mapping op %p: %d\n", ic, op,
+				 op->op_count);
+			if (op->op_count == 0) {
+				rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
+				ret = -ENOMEM; /* XXX ? */
+				goto out;
+			}
+			op->op_mapped = 1;
 		}
-
-		op->op_mapped = 1;
+	} else {
+		op->op_count = op->op_nents;
+		odp_lkey = rds_ib_get_lkey(op->op_odp_mr->r_trans_private);
 	}
 
 	/*
@@ -923,14 +935,20 @@ int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
 		for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
 		     scat != &op->op_sg[op->op_count]; j++) {
 			len = sg_dma_len(scat);
-			send->s_sge[j].addr = sg_dma_address(scat);
+			if (!op->op_odp_mr) {
+				send->s_sge[j].addr = sg_dma_address(scat);
+				send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
+			} else {
+				send->s_sge[j].addr = odp_addr;
+				send->s_sge[j].lkey = odp_lkey;
+			}
 			send->s_sge[j].length = len;
-			send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
 
 			sent += len;
 			rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
 
 			remote_addr += len;
+			odp_addr += len;
 			scat++;
 		}
 
net/rds/rdma.c +108 -37
index eb23c38ce2b3..3c6afdda709b 100644
@@ -177,13 +177,14 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 			  struct rds_conn_path *cp)
 {
 	struct rds_mr *mr = NULL, *found;
+	struct scatterlist *sg = NULL;
 	unsigned int nr_pages;
 	struct page **pages = NULL;
-	struct scatterlist *sg;
 	void *trans_private;
 	unsigned long flags;
 	rds_rdma_cookie_t cookie;
-	unsigned int nents;
+	unsigned int nents = 0;
+	int need_odp = 0;
 	long i;
 	int ret;
 
@@ -196,6 +197,20 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 		ret = -EOPNOTSUPP;
 		goto out;
 	}
+	/* If the combination of the addr and size requested for this memory
+	 * region causes an integer overflow, return error.
+	 */
+	if (((args->vec.addr + args->vec.bytes) < args->vec.addr) ||
+	    PAGE_ALIGN(args->vec.addr + args->vec.bytes) <
+		    (args->vec.addr + args->vec.bytes)) {
+		ret = -EINVAL;
+		goto out;
+	}
+
+	if (!can_do_mlock()) {
+		ret = -EPERM;
+		goto out;
+	}
 
 	nr_pages = rds_pages_in_vec(&args->vec);
 	if (nr_pages == 0) {
@@ -250,36 +265,44 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 	 * the zero page.
 	 */
 	ret = rds_pin_pages(args->vec.addr, nr_pages, pages, 1);
-	if (ret < 0)
+	if (ret == -EOPNOTSUPP) {
+		need_odp = 1;
+	} else if (ret <= 0) {
 		goto out;
+	} else {
+		nents = ret;
+		sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
+		if (!sg) {
+			ret = -ENOMEM;
+			goto out;
+		}
+		WARN_ON(!nents);
+		sg_init_table(sg, nents);
 
-	nents = ret;
-	sg = kcalloc(nents, sizeof(*sg), GFP_KERNEL);
-	if (!sg) {
-		ret = -ENOMEM;
-		goto out;
-	}
-	WARN_ON(!nents);
-	sg_init_table(sg, nents);
-
-	/* Stick all pages into the scatterlist */
-	for (i = 0 ; i < nents; i++)
-		sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
-
-	rdsdebug("RDS: trans_private nents is %u\n", nents);
+		/* Stick all pages into the scatterlist */
+		for (i = 0 ; i < nents; i++)
+			sg_set_page(&sg[i], pages[i], PAGE_SIZE, 0);
 
+		rdsdebug("RDS: trans_private nents is %u\n", nents);
+	}
 	/* Obtain a transport specific MR. If this succeeds, the
 	 * s/g list is now owned by the MR.
 	 * Note that dma_map() implies that pending writes are
 	 * flushed to RAM, so no dma_sync is needed here. */
-	trans_private = rs->rs_transport->get_mr(sg, nents, rs,
-						 &mr->r_key,
-						 cp ? cp->cp_conn : NULL);
+	trans_private = rs->rs_transport->get_mr(
+		sg, nents, rs, &mr->r_key, cp ? cp->cp_conn : NULL,
+		args->vec.addr, args->vec.bytes,
+		need_odp ? ODP_ZEROBASED : ODP_NOT_NEEDED);
 
 	if (IS_ERR(trans_private)) {
-		for (i = 0 ; i < nents; i++)
-			put_page(sg_page(&sg[i]));
-		kfree(sg);
+		/* In ODP case, we don't GUP pages, so don't need
+		 * to release anything.
+		 */
+		if (!need_odp) {
+			for (i = 0 ; i < nents; i++)
+				put_page(sg_page(&sg[i]));
+			kfree(sg);
+		}
 		ret = PTR_ERR(trans_private);
 		goto out;
 	}
@@ -293,7 +316,11 @@ static int __rds_rdma_map(struct rds_sock *rs, struct rds_get_mr_args *args,
 	 * map page aligned regions. So we keep the offset, and build
 	 * a 64bit cookie containing <R_Key, offset> and pass that
 	 * around. */
-	cookie = rds_rdma_make_cookie(mr->r_key, args->vec.addr & ~PAGE_MASK);
+	if (need_odp)
+		cookie = rds_rdma_make_cookie(mr->r_key, 0);
+	else
+		cookie = rds_rdma_make_cookie(mr->r_key,
+					      args->vec.addr & ~PAGE_MASK);
 	if (cookie_ret)
 		*cookie_ret = cookie;
 
@@ -458,22 +485,26 @@ void rds_rdma_free_op(struct rm_rdma_op *ro)
 {
 	unsigned int i;
 
-	for (i = 0; i < ro->op_nents; i++) {
-		struct page *page = sg_page(&ro->op_sg[i]);
-
-		/* Mark page dirty if it was possibly modified, which
-		 * is the case for a RDMA_READ which copies from remote
-		 * to local memory */
-		if (!ro->op_write) {
-			WARN_ON(!page->mapping && irqs_disabled());
-			set_page_dirty(page);
+	if (ro->op_odp_mr) {
+		rds_mr_put(ro->op_odp_mr);
+	} else {
+		for (i = 0; i < ro->op_nents; i++) {
+			struct page *page = sg_page(&ro->op_sg[i]);
+
+			/* Mark page dirty if it was possibly modified, which
+			 * is the case for a RDMA_READ which copies from remote
+			 * to local memory
+			 */
+			if (!ro->op_write)
+				set_page_dirty(page);
+			put_page(page);
 		}
-		put_page(page);
 	}
 
 	kfree(ro->op_notifier);
 	ro->op_notifier = NULL;
 	ro->op_active = 0;
+	ro->op_odp_mr = NULL;
 }
 
 void rds_atomic_free_op(struct rm_atomic_op *ao)
@@ -583,6 +614,7 @@ int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 	struct rds_iovec *iovs;
 	unsigned int i, j;
 	int ret = 0;
+	bool odp_supported = true;
 
 	if (cmsg->cmsg_len < CMSG_LEN(sizeof(struct rds_rdma_args))
 	    || rm->rdma.op_active)
@@ -604,6 +636,9 @@ int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 		ret = -EINVAL;
 		goto out_ret;
 	}
+	/* odp-mr is not supported for multiple requests within one message */
+	if (args->nr_local != 1)
+		odp_supported = false;
 
 	iovs = vec->iov;
 
@@ -625,6 +660,8 @@ int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 	op->op_silent = !!(args->flags & RDS_RDMA_SILENT);
 	op->op_active = 1;
 	op->op_recverr = rs->rs_recverr;
+	op->op_odp_mr = NULL;
+
 	WARN_ON(!nr_pages);
 	op->op_sg = rds_message_alloc_sgs(rm, nr_pages, &ret);
 	if (!op->op_sg)
@@ -674,10 +711,44 @@ int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 		 * If it's a READ operation, we need to pin the pages for writing.
 		 */
 		ret = rds_pin_pages(iov->addr, nr, pages, !op->op_write);
-		if (ret < 0)
+		if ((!odp_supported && ret <= 0) ||
+		    (odp_supported && ret <= 0 && ret != -EOPNOTSUPP))
 			goto out_pages;
-		else
-			ret = 0;
+
+		if (ret == -EOPNOTSUPP) {
+			struct rds_mr *local_odp_mr;
+
+			if (!rs->rs_transport->get_mr) {
+				ret = -EOPNOTSUPP;
+				goto out_pages;
+			}
+			local_odp_mr =
+				kzalloc(sizeof(*local_odp_mr), GFP_KERNEL);
+			if (!local_odp_mr) {
+				ret = -ENOMEM;
+				goto out_pages;
+			}
+			RB_CLEAR_NODE(&local_odp_mr->r_rb_node);
+			refcount_set(&local_odp_mr->r_refcount, 1);
+			local_odp_mr->r_trans = rs->rs_transport;
+			local_odp_mr->r_sock = rs;
+			local_odp_mr->r_trans_private =
+				rs->rs_transport->get_mr(
+					NULL, 0, rs, &local_odp_mr->r_key, NULL,
+					iov->addr, iov->bytes, ODP_VIRTUAL);
+			if (IS_ERR(local_odp_mr->r_trans_private)) {
+				ret = IS_ERR(local_odp_mr->r_trans_private);
+				rdsdebug("get_mr ret %d %p\"", ret,
+					 local_odp_mr->r_trans_private);
+				kfree(local_odp_mr);
+				ret = -EOPNOTSUPP;
+				goto out_pages;
+			}
+			rdsdebug("Need odp; local_odp_mr %p trans_private %p\n",
+				 local_odp_mr, local_odp_mr->r_trans_private);
+			op->op_odp_mr = local_odp_mr;
+			op->op_odp_addr = iov->addr;
+		}
 
 		rdsdebug("RDS: nr_bytes %u nr %u iov->bytes %llu iov->addr %llx\n",
 			 nr_bytes, nr, iov->bytes, iov->addr);

net/rds/rdma.c int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm, · line 767

Quoted 4 lines
net/rds/rdma.c +1 -0
@@ -693,6 +764,7 @@ int rds_cmsg_rdma_args(struct rds_sock *rs, struct rds_message *rm,
 					min_t(unsigned int, iov->bytes, PAGE_SIZE - offset),
 					offset);
 
+			sg->dma_length = sg->length;

This line should be "sg_dma_len(sg) = sg->length;".

Thanks