LU-9500 lnd: Don't Page Align remote_addr with FastReg

Trying to page align the remote_addr for IB_RDMA_WRITE work
requests is triggering "dump_cqe" errors from MOFED 4.x + mlx5.

This patch removes the address masking we were doing with FastReg
which was trying to page align remote_addr values. I am also
removing the setting of "mr->iova" with FastReg as this is being
done in the call to ib_map_mr_sg() and could cause problems.

Lustre-change: https://review.whamcloud.com/27149
Lustre-commit: 6c6341804133ea0a4d4535c621f28f61fe6c29ab

Signed-off-by: Doug Oucharek <doug.s.oucharek@intel.com>
Change-Id: If35baa467d8d60866f709b5feea7f619063c6da4
Reviewed-by: Gu Zheng <gzheng@ddn.com>
Reviewed-by: Dmitry Eremin <dmitry.eremin@intel.com>
Reviewed-by: Amir Shehata <amir.shehata@intel.com>
Signed-off-by: Minh Diep <minh.diep@intel.com>
Reviewed-on: https://review.whamcloud.com/28237
Tested-by: Jenkins
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: John L. Hammond <john.hammond@intel.com>

diff --git a/lnet/klnds/o2iblnd/o2iblnd.c b/lnet/klnds/o2iblnd/o2iblnd.c
index 298cd9f..110b6e6 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.c
+++ b/lnet/klnds/o2iblnd/o2iblnd.c
@@ -1848,7 +1848,7 @@ kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status)
 
 int
 kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx, kib_rdma_desc_t *rd,
-                   __u32 nob, __u64 iov, kib_fmr_t *fmr)
+                   __u32 nob, __u64 iov, kib_fmr_t *fmr, bool *is_fastreg)
 {
        kib_fmr_pool_t *fpo;
        __u64 *pages = tx->tx_pages;
@@ -1868,6 +1868,7 @@ again:
                if (fpo->fpo_is_fmr) {
                        struct ib_pool_fmr *pfmr;
 
+                       *is_fastreg = 0;
                        spin_unlock(&fps->fps_lock);
 
                        if (!tx_pages_mapped) {
@@ -1887,6 +1888,7 @@ again:
                        }
                        rc = PTR_ERR(pfmr);
                } else {
+                       *is_fastreg = 1;
                        if (!list_empty(&fpo->fast_reg.fpo_pool_list)) {
                                struct kib_fast_reg_descriptor *frd;
 #ifdef HAVE_IB_MAP_MR_SG
@@ -1939,8 +1941,6 @@ again:
                                        return n < 0 ? n : -EINVAL;
                                }
 
-                               mr->iova = iov;
-
                                wr = &frd->frd_fastreg_wr;
                                memset(wr, 0, sizeof(*wr));
 

diff --git a/lnet/klnds/o2iblnd/o2iblnd.h b/lnet/klnds/o2iblnd/o2iblnd.h
index be6e882..c4a4742 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd.h
+++ b/lnet/klnds/o2iblnd/o2iblnd.h
@@ -1187,7 +1187,7 @@ struct list_head *kiblnd_pool_alloc_node(kib_poolset_t *ps);
 
 int  kiblnd_fmr_pool_map(kib_fmr_poolset_t *fps, kib_tx_t *tx,
                         kib_rdma_desc_t *rd, __u32 nob, __u64 iov,
-                        kib_fmr_t *fmr);
+                        kib_fmr_t *fmr, bool *is_fastreg);
 void kiblnd_fmr_pool_unmap(kib_fmr_t *fmr, int status);
 
 int  kiblnd_tunables_setup(struct lnet_ni *ni);

diff --git a/lnet/klnds/o2iblnd/o2iblnd_cb.c b/lnet/klnds/o2iblnd/o2iblnd_cb.c
index 2ecccbf..17e2479 100644 (file)
--- a/lnet/klnds/o2iblnd/o2iblnd_cb.c
+++ b/lnet/klnds/o2iblnd/o2iblnd_cb.c
@@ -569,6 +569,7 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
        kib_fmr_poolset_t       *fps;
        int                     cpt;
        int                     rc;
+       bool                    is_fastreg = 0;
 
        LASSERT(tx->tx_pool != NULL);
        LASSERT(tx->tx_pool->tpo_pool.po_owner != NULL);
@@ -577,7 +578,7 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
        cpt = tx->tx_pool->tpo_pool.po_owner->ps_cpt;
 
        fps = net->ibn_fmr_ps[cpt];
-       rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr);
+       rc = kiblnd_fmr_pool_map(fps, tx, rd, nob, 0, &tx->fmr, &is_fastreg);
        if (rc != 0) {
                CERROR("Can't map %u pages: %d\n", nob, rc);
                return rc;
@@ -586,7 +587,8 @@ kiblnd_fmr_map_tx(kib_net_t *net, kib_tx_t *tx, kib_rdma_desc_t *rd, __u32 nob)
        /* If rd is not tx_rd, it's going to get sent to a peer_ni, who will need
         * the rkey */
        rd->rd_key = tx->fmr.fmr_key;
-       rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
+       if (!is_fastreg)
+               rd->rd_frags[0].rf_addr &= ~hdev->ibh_page_mask;
        rd->rd_frags[0].rf_nob   = nob;
        rd->rd_nfrags = 1;