IB/core: Introduce new fast registration API
authorSagi Grimberg <sagig@mellanox.com>
Tue, 13 Oct 2015 16:11:24 +0000 (19:11 +0300)
committerDoug Ledford <dledford@redhat.com>
Thu, 29 Oct 2015 02:27:17 +0000 (22:27 -0400)
The new fast registration  verb ib_map_mr_sg receives a scatterlist
and converts it to a page list under the verbs API thus hiding
the specific HW mapping details away from the consumer.

The provider drivers are provided with a generic helper ib_sg_to_pages
that converts a scatterlist into a vector of page addresses. The
drivers can still perform any HW specific page address setting
by passing a set_page function pointer which will be invoked for
each page address. This allows drivers to avoid keeping a shadow
page vectors and convert them to HW specific translations by doing
extra copies.

This API will allow ULPs to remove the duplicated code of constructing
a page vector from a given sg list.

The send work request ib_reg_wr also shrinks as it will contain only
mr, key and access flags in addition.

Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Tested-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/core/verbs.c
include/rdma/ib_verbs.h

index 46d97f09fbae7d696db1ade865bae77256e1e1b7..5673df06121337144e4c2114cfc0fb6ad528524d 100644 (file)
@@ -1496,3 +1496,110 @@ int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
                mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
 }
 EXPORT_SYMBOL(ib_check_mr_status);
+
+/**
+ * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
+ *     and set it the memory region.
+ * @mr:            memory region
+ * @sg:            dma mapped scatterlist
+ * @sg_nents:      number of entries in sg
+ * @page_size:     page vector desired page size
+ *
+ * Constraints:
+ * - The first sg element is allowed to have an offset.
+ * - Each sg element must be aligned to page_size (or physically
+ *   contiguous to the previous element). In case an sg element has a
+ *   non contiguous offset, the mapping prefix will not include it.
+ * - The last sg element is allowed to have length less than page_size.
+ * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
+ *   then only max_num_sg entries will be mapped.
+ *
+ * Returns the number of sg elements that were mapped to the memory region.
+ *
+ * After this completes successfully, the  memory region
+ * is ready for registration.
+ */
+int ib_map_mr_sg(struct ib_mr *mr,
+                struct scatterlist *sg,
+                int sg_nents,
+                unsigned int page_size)
+{
+       if (unlikely(!mr->device->map_mr_sg))
+               return -ENOSYS;
+
+       mr->page_size = page_size;
+
+       return mr->device->map_mr_sg(mr, sg, sg_nents);
+}
+EXPORT_SYMBOL(ib_map_mr_sg);
+
+/**
+ * ib_sg_to_pages() - Convert the largest prefix of a sg list
+ *     to a page vector
+ * @mr:            memory region
+ * @sgl:           dma mapped scatterlist
+ * @sg_nents:      number of entries in sg
+ * @set_page:      driver page assignment function pointer
+ *
+ * Core service helper for drivers to covert the largest
+ * prefix of given sg list to a page vector. The sg list
+ * prefix converted is the prefix that meet the requirements
+ * of ib_map_mr_sg.
+ *
+ * Returns the number of sg elements that were assigned to
+ * a page vector.
+ */
+int ib_sg_to_pages(struct ib_mr *mr,
+                  struct scatterlist *sgl,
+                  int sg_nents,
+                  int (*set_page)(struct ib_mr *, u64))
+{
+       struct scatterlist *sg;
+       u64 last_end_dma_addr = 0, last_page_addr = 0;
+       unsigned int last_page_off = 0;
+       u64 page_mask = ~((u64)mr->page_size - 1);
+       int i;
+
+       mr->iova = sg_dma_address(&sgl[0]);
+       mr->length = 0;
+
+       for_each_sg(sgl, sg, sg_nents, i) {
+               u64 dma_addr = sg_dma_address(sg);
+               unsigned int dma_len = sg_dma_len(sg);
+               u64 end_dma_addr = dma_addr + dma_len;
+               u64 page_addr = dma_addr & page_mask;
+
+               if (i && page_addr != dma_addr) {
+                       if (last_end_dma_addr != dma_addr) {
+                               /* gap */
+                               goto done;
+
+                       } else if (last_page_off + dma_len <= mr->page_size) {
+                               /* chunk this fragment with the last */
+                               mr->length += dma_len;
+                               last_end_dma_addr += dma_len;
+                               last_page_off += dma_len;
+                               continue;
+                       } else {
+                               /* map starting from the next page */
+                               page_addr = last_page_addr + mr->page_size;
+                               dma_len -= mr->page_size - last_page_off;
+                       }
+               }
+
+               do {
+                       if (unlikely(set_page(mr, page_addr)))
+                               goto done;
+                       page_addr += mr->page_size;
+               } while (page_addr < end_dma_addr);
+
+               mr->length += dma_len;
+               last_end_dma_addr = end_dma_addr;
+               last_page_addr = end_dma_addr & page_mask;
+               last_page_off = end_dma_addr & ~page_mask;
+       }
+
+done:
+       return i;
+}
+EXPORT_SYMBOL(ib_sg_to_pages);
index 85103aff909bf38631d521ac9e8d9d3061984cc9..065d37c13aa6a67165a8cd084072b123483e45a6 100644 (file)
@@ -738,6 +738,7 @@ enum ib_wc_opcode {
        IB_WC_LSO,
        IB_WC_LOCAL_INV,
        IB_WC_FAST_REG_MR,
+       IB_WC_REG_MR,
        IB_WC_MASKED_COMP_SWAP,
        IB_WC_MASKED_FETCH_ADD,
 /*
@@ -1025,6 +1026,7 @@ enum ib_wr_opcode {
        IB_WR_RDMA_READ_WITH_INV,
        IB_WR_LOCAL_INV,
        IB_WR_FAST_REG_MR,
+       IB_WR_REG_MR,
        IB_WR_MASKED_ATOMIC_CMP_AND_SWP,
        IB_WR_MASKED_ATOMIC_FETCH_AND_ADD,
        IB_WR_BIND_MW,
@@ -1157,6 +1159,18 @@ static inline struct ib_fast_reg_wr *fast_reg_wr(struct ib_send_wr *wr)
        return container_of(wr, struct ib_fast_reg_wr, wr);
 }
 
+struct ib_reg_wr {
+       struct ib_send_wr       wr;
+       struct ib_mr            *mr;
+       u32                     key;
+       int                     access;
+};
+
+static inline struct ib_reg_wr *reg_wr(struct ib_send_wr *wr)
+{
+       return container_of(wr, struct ib_reg_wr, wr);
+}
+
 struct ib_bind_mw_wr {
        struct ib_send_wr       wr;
        struct ib_mw            *mw;
@@ -1369,6 +1383,9 @@ struct ib_mr {
        struct ib_uobject *uobject;
        u32                lkey;
        u32                rkey;
+       u64                iova;
+       u32                length;
+       unsigned int       page_size;
        atomic_t           usecnt; /* count number of MWs */
 };
 
@@ -1753,6 +1770,9 @@ struct ib_device {
        struct ib_mr *             (*alloc_mr)(struct ib_pd *pd,
                                               enum ib_mr_type mr_type,
                                               u32 max_num_sg);
+       int                        (*map_mr_sg)(struct ib_mr *mr,
+                                               struct scatterlist *sg,
+                                               int sg_nents);
        struct ib_fast_reg_page_list * (*alloc_fast_reg_page_list)(struct ib_device *device,
                                                                   int page_list_len);
        void                       (*free_fast_reg_page_list)(struct ib_fast_reg_page_list *page_list);
@@ -3059,4 +3079,28 @@ struct net_device *ib_get_net_dev_by_params(struct ib_device *dev, u8 port,
                                            u16 pkey, const union ib_gid *gid,
                                            const struct sockaddr *addr);
 
+int ib_map_mr_sg(struct ib_mr *mr,
+                struct scatterlist *sg,
+                int sg_nents,
+                unsigned int page_size);
+
+static inline int
+ib_map_mr_sg_zbva(struct ib_mr *mr,
+                 struct scatterlist *sg,
+                 int sg_nents,
+                 unsigned int page_size)
+{
+       int n;
+
+       n = ib_map_mr_sg(mr, sg, sg_nents, page_size);
+       mr->iova = 0;
+
+       return n;
+}
+
+int ib_sg_to_pages(struct ib_mr *mr,
+                  struct scatterlist *sgl,
+                  int sg_nents,
+                  int (*set_page)(struct ib_mr *, u64));
+
 #endif /* IB_VERBS_H */