svcrdma: Port to new memory registration API
authorSagi Grimberg <sagig@mellanox.com>
Tue, 13 Oct 2015 16:11:36 +0000 (19:11 +0300)
committerDoug Ledford <dledford@redhat.com>
Thu, 29 Oct 2015 02:27:18 +0000 (22:27 -0400)
Instead of maintaining a fastreg page list, keep an sg table
and convert an array of pages to a sg list. Then call ib_map_mr_sg
and construct ib_reg_wr.

Signed-off-by: Sagi Grimberg <sagig@mellanox.com>
Acked-by: Christoph Hellwig <hch@lst.de>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Selvin Xavier <selvin.xavier@avagotech.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
include/linux/sunrpc/svc_rdma.h
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c
net/sunrpc/xprtrdma/svc_rdma_transport.c

index 7ccc961f33e933fc25eae43dd74fbd888920d2d7..1e4438ea2380dd07e79f3d8011445900edea7437 100644 (file)
@@ -105,11 +105,9 @@ struct svc_rdma_chunk_sge {
 };
 struct svc_rdma_fastreg_mr {
        struct ib_mr *mr;
-       void *kva;
-       struct ib_fast_reg_page_list *page_list;
-       int page_list_len;
+       struct scatterlist *sg;
+       int sg_nents;
        unsigned long access_flags;
-       unsigned long map_len;
        enum dma_data_direction direction;
        struct list_head frmr_list;
 };
index 7be42d0da19e4db58b46b72a91704b8ab509071a..cb0991345816e4ab2c2c99c6242726bef7e6c98a 100644 (file)
@@ -220,12 +220,12 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
 {
        struct ib_rdma_wr read_wr;
        struct ib_send_wr inv_wr;
-       struct ib_fast_reg_wr fastreg_wr;
+       struct ib_reg_wr reg_wr;
        u8 key;
-       int pages_needed = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
+       int nents = PAGE_ALIGN(*page_offset + rs_length) >> PAGE_SHIFT;
        struct svc_rdma_op_ctxt *ctxt = svc_rdma_get_context(xprt);
        struct svc_rdma_fastreg_mr *frmr = svc_rdma_get_frmr(xprt);
-       int ret, read, pno;
+       int ret, read, pno, dma_nents, n;
        u32 pg_off = *page_offset;
        u32 pg_no = *page_no;
 
@@ -234,16 +234,14 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
 
        ctxt->direction = DMA_FROM_DEVICE;
        ctxt->frmr = frmr;
-       pages_needed = min_t(int, pages_needed, xprt->sc_frmr_pg_list_len);
-       read = min_t(int, pages_needed << PAGE_SHIFT, rs_length);
+       nents = min_t(unsigned int, nents, xprt->sc_frmr_pg_list_len);
+       read = min_t(int, nents << PAGE_SHIFT, rs_length);
 
-       frmr->kva = page_address(rqstp->rq_arg.pages[pg_no]);
        frmr->direction = DMA_FROM_DEVICE;
        frmr->access_flags = (IB_ACCESS_LOCAL_WRITE|IB_ACCESS_REMOTE_WRITE);
-       frmr->map_len = pages_needed << PAGE_SHIFT;
-       frmr->page_list_len = pages_needed;
+       frmr->sg_nents = nents;
 
-       for (pno = 0; pno < pages_needed; pno++) {
+       for (pno = 0; pno < nents; pno++) {
                int len = min_t(int, rs_length, PAGE_SIZE - pg_off);
 
                head->arg.pages[pg_no] = rqstp->rq_arg.pages[pg_no];
@@ -251,17 +249,12 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
                head->arg.len += len;
                if (!pg_off)
                        head->count++;
+
+               sg_set_page(&frmr->sg[pno], rqstp->rq_arg.pages[pg_no],
+                           len, pg_off);
+
                rqstp->rq_respages = &rqstp->rq_arg.pages[pg_no+1];
                rqstp->rq_next_page = rqstp->rq_respages + 1;
-               frmr->page_list->page_list[pno] =
-                       ib_dma_map_page(xprt->sc_cm_id->device,
-                                       head->arg.pages[pg_no], 0,
-                                       PAGE_SIZE, DMA_FROM_DEVICE);
-               ret = ib_dma_mapping_error(xprt->sc_cm_id->device,
-                                          frmr->page_list->page_list[pno]);
-               if (ret)
-                       goto err;
-               atomic_inc(&xprt->sc_dma_used);
 
                /* adjust offset and wrap to next page if needed */
                pg_off += len;
@@ -277,28 +270,42 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
        else
                clear_bit(RDMACTXT_F_LAST_CTXT, &ctxt->flags);
 
+       dma_nents = ib_dma_map_sg(xprt->sc_cm_id->device,
+                                 frmr->sg, frmr->sg_nents,
+                                 frmr->direction);
+       if (!dma_nents) {
+               pr_err("svcrdma: failed to dma map sg %p\n",
+                      frmr->sg);
+               return -ENOMEM;
+       }
+       atomic_inc(&xprt->sc_dma_used);
+
+       n = ib_map_mr_sg(frmr->mr, frmr->sg, frmr->sg_nents, PAGE_SIZE);
+       if (unlikely(n != frmr->sg_nents)) {
+               pr_err("svcrdma: failed to map mr %p (%d/%d elements)\n",
+                      frmr->mr, n, frmr->sg_nents);
+               return n < 0 ? n : -EINVAL;
+       }
+
        /* Bump the key */
        key = (u8)(frmr->mr->lkey & 0x000000FF);
        ib_update_fast_reg_key(frmr->mr, ++key);
 
-       ctxt->sge[0].addr = (unsigned long)frmr->kva + *page_offset;
+       ctxt->sge[0].addr = frmr->mr->iova;
        ctxt->sge[0].lkey = frmr->mr->lkey;
-       ctxt->sge[0].length = read;
+       ctxt->sge[0].length = frmr->mr->length;
        ctxt->count = 1;
        ctxt->read_hdr = head;
 
-       /* Prepare FASTREG WR */
-       memset(&fastreg_wr, 0, sizeof(fastreg_wr));
-       fastreg_wr.wr.opcode = IB_WR_FAST_REG_MR;
-       fastreg_wr.wr.send_flags = IB_SEND_SIGNALED;
-       fastreg_wr.iova_start = (unsigned long)frmr->kva;
-       fastreg_wr.page_list = frmr->page_list;
-       fastreg_wr.page_list_len = frmr->page_list_len;
-       fastreg_wr.page_shift = PAGE_SHIFT;
-       fastreg_wr.length = frmr->map_len;
-       fastreg_wr.access_flags = frmr->access_flags;
-       fastreg_wr.rkey = frmr->mr->lkey;
-       fastreg_wr.wr.next = &read_wr.wr;
+       /* Prepare REG WR */
+       reg_wr.wr.opcode = IB_WR_REG_MR;
+       reg_wr.wr.wr_id = 0;
+       reg_wr.wr.send_flags = IB_SEND_SIGNALED;
+       reg_wr.wr.num_sge = 0;
+       reg_wr.mr = frmr->mr;
+       reg_wr.key = frmr->mr->lkey;
+       reg_wr.access = frmr->access_flags;
+       reg_wr.wr.next = &read_wr.wr;
 
        /* Prepare RDMA_READ */
        memset(&read_wr, 0, sizeof(read_wr));
@@ -324,7 +331,7 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
        ctxt->wr_op = read_wr.wr.opcode;
 
        /* Post the chain */
-       ret = svc_rdma_send(xprt, &fastreg_wr.wr);
+       ret = svc_rdma_send(xprt, &reg_wr.wr);
        if (ret) {
                pr_err("svcrdma: Error %d posting RDMA_READ\n", ret);
                set_bit(XPT_CLOSE, &xprt->sc_xprt.xpt_flags);
@@ -338,7 +345,8 @@ int rdma_read_chunk_frmr(struct svcxprt_rdma *xprt,
        atomic_inc(&rdma_stat_read);
        return ret;
  err:
-       svc_rdma_unmap_dma(ctxt);
+       ib_dma_unmap_sg(xprt->sc_cm_id->device,
+                       frmr->sg, frmr->sg_nents, frmr->direction);
        svc_rdma_put_context(ctxt, 0);
        svc_rdma_put_frmr(xprt, frmr);
        return ret;
index 4a41122d586f05900229217aab75c3acbeab02b1..a266e870d870e8b2cf0d938c3009460b46fddc7b 100644 (file)
@@ -732,7 +732,7 @@ static struct svc_xprt *svc_rdma_create(struct svc_serv *serv,
 static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
 {
        struct ib_mr *mr;
-       struct ib_fast_reg_page_list *pl;
+       struct scatterlist *sg;
        struct svc_rdma_fastreg_mr *frmr;
        u32 num_sg;
 
@@ -745,13 +745,14 @@ static struct svc_rdma_fastreg_mr *rdma_alloc_frmr(struct svcxprt_rdma *xprt)
        if (IS_ERR(mr))
                goto err_free_frmr;
 
-       pl = ib_alloc_fast_reg_page_list(xprt->sc_cm_id->device,
-                                        num_sg);
-       if (IS_ERR(pl))
+       sg = kcalloc(RPCSVC_MAXPAGES, sizeof(*sg), GFP_KERNEL);
+       if (!sg)
                goto err_free_mr;
 
+       sg_init_table(sg, RPCSVC_MAXPAGES);
+
        frmr->mr = mr;
-       frmr->page_list = pl;
+       frmr->sg = sg;
        INIT_LIST_HEAD(&frmr->frmr_list);
        return frmr;
 
@@ -771,8 +772,8 @@ static void rdma_dealloc_frmr_q(struct svcxprt_rdma *xprt)
                frmr = list_entry(xprt->sc_frmr_q.next,
                                  struct svc_rdma_fastreg_mr, frmr_list);
                list_del_init(&frmr->frmr_list);
+               kfree(frmr->sg);
                ib_dereg_mr(frmr->mr);
-               ib_free_fast_reg_page_list(frmr->page_list);
                kfree(frmr);
        }
 }
@@ -786,8 +787,7 @@ struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
                frmr = list_entry(rdma->sc_frmr_q.next,
                                  struct svc_rdma_fastreg_mr, frmr_list);
                list_del_init(&frmr->frmr_list);
-               frmr->map_len = 0;
-               frmr->page_list_len = 0;
+               frmr->sg_nents = 0;
        }
        spin_unlock_bh(&rdma->sc_frmr_q_lock);
        if (frmr)
@@ -796,25 +796,13 @@ struct svc_rdma_fastreg_mr *svc_rdma_get_frmr(struct svcxprt_rdma *rdma)
        return rdma_alloc_frmr(rdma);
 }
 
-static void frmr_unmap_dma(struct svcxprt_rdma *xprt,
-                          struct svc_rdma_fastreg_mr *frmr)
-{
-       int page_no;
-       for (page_no = 0; page_no < frmr->page_list_len; page_no++) {
-               dma_addr_t addr = frmr->page_list->page_list[page_no];
-               if (ib_dma_mapping_error(frmr->mr->device, addr))
-                       continue;
-               atomic_dec(&xprt->sc_dma_used);
-               ib_dma_unmap_page(frmr->mr->device, addr, PAGE_SIZE,
-                                 frmr->direction);
-       }
-}
-
 void svc_rdma_put_frmr(struct svcxprt_rdma *rdma,
                       struct svc_rdma_fastreg_mr *frmr)
 {
        if (frmr) {
-               frmr_unmap_dma(rdma, frmr);
+               ib_dma_unmap_sg(rdma->sc_cm_id->device,
+                               frmr->sg, frmr->sg_nents, frmr->direction);
+               atomic_dec(&rdma->sc_dma_used);
                spin_lock_bh(&rdma->sc_frmr_q_lock);
                WARN_ON_ONCE(!list_empty(&frmr->frmr_list));
                list_add(&frmr->frmr_list, &rdma->sc_frmr_q);