sif: rq: Do not clear the rq_sw until the completion of flush_rq

Orabug: 23754857

The rq can be invalidated from reset_qp or flush_rq. Nevertheless,
the rq_sw data structure has been reset after rq is invalidated in
reset_qp regardless of the completion of the flush_rq. Thus, move
the rq synchronization to reset_qp, and place the synchronization
in between of reset_qp and flush_rq. After invalidating and
reseting the rq, no flush rq is required as both head and tail
have been reset to 0.

This commit creates another atomic_t variable for the synchronization
between reset rq and flush_rq.

Signed-off-by: Wei Lin Guay <wei.lin.guay@oracle.com>
Reviewed-by: Knut Omang <knut.omang@oracle.com>

diff --git a/drivers/infiniband/hw/sif/sif_qp.c b/drivers/infiniband/hw/sif/sif_qp.c
index 53b578e092e80245e75418bee5156204aef10085..41ef47699f941edb2afa8a4e5ccfb3a196f8c59b 100644 (file)
--- a/drivers/infiniband/hw/sif/sif_qp.c
+++ b/drivers/infiniband/hw/sif/sif_qp.c
@@ -2347,9 +2347,23 @@ failed:
                if (ret)
                        return ret;
 
+               /* Make sure the in-progress rq flush has
+                * completed before reset the rq tail
+                * and head.
+                */
+               if (atomic_dec_and_test(&rq->flush_in_progress))
+                       complete(&rq->can_reset);
+               wait_for_completion(&rq->can_reset);
+
                /* Reset pointers */
                memset(rq_sw, 0, sizeof(*rq_sw));
                set_psif_rq_hw__head_indx(&rq->d, 0);
+
+               /* reset the flush_in_progress, if the qp is reset
+                * and the qp can be reused again.
+                * Thus, reset the flush_in_progress to 1.
+                */
+               atomic_set(&rq->flush_in_progress, 1);
        }
 
        mb();

diff --git a/drivers/infiniband/hw/sif/sif_rq.c b/drivers/infiniband/hw/sif/sif_rq.c
index e3154ae35319ed57b5bdbc5e606d9483ac8019e7..0ff96a96866ce329a3f2265666c2d04fc515ac53 100644 (file)
--- a/drivers/infiniband/hw/sif/sif_rq.c
+++ b/drivers/infiniband/hw/sif/sif_rq.c
@@ -153,8 +153,9 @@ int alloc_rq(struct sif_dev *sdev, struct sif_pd *pd,
        }
 
        rq->sg_entries = sg_entries;
-       init_completion(&rq->can_destroy);
+       init_completion(&rq->can_reset);
        atomic_set(&rq->refcnt, 1);
+       atomic_set(&rq->flush_in_progress, 1);
 
        /* Initialize hw part of descriptor */
        memset(&lrq_hw, 0, sizeof(lrq_hw));
@@ -359,7 +360,7 @@ static void sif_flush_rq(struct work_struct *work)
        if (len == 0)
                goto error;
 
-       if (atomic_add_unless(&rq->refcnt, 1, 0)) {
+       if (atomic_add_unless(&rq->flush_in_progress, 1, 0)) {
                sif_log(sdev, SIF_INFO_V, "flushing %d entries out of %d/%d entries remaining",
                        len, atomic_read(&rq_sw->length), rq->entries);
 
@@ -569,8 +570,8 @@ flush_rq_again:
                if (test_and_clear_bit(FLUSH_RQ_IN_FLIGHT, &rq_sw->flags))
                        goto flush_rq_again;
 free_rq_error:
-               if (atomic_dec_and_test(&rq->refcnt))
-                       complete(&rq->can_destroy);
+               if (atomic_dec_and_test(&rq->flush_in_progress))
+                       complete(&rq->can_reset);
        }
 error:
        clear_bit(FLUSH_RQ_IN_PROGRESS, &rq_sw->flags);
@@ -581,25 +582,24 @@ done:
 
 int free_rq(struct sif_dev *sdev, int rq_idx)
 {
-       struct sif_rq *rq;
+       struct sif_rq *rq = get_sif_rq(sdev, rq_idx);
+       struct sif_rq_sw *rq_sw = get_sif_rq_sw(sdev, rq_idx);
 
-       rq = get_sif_rq(sdev, rq_idx);
        sif_log(sdev, SIF_RQ, "entry %d", rq_idx);
 
-       if (!rq->is_srq) {
-               if (atomic_dec_and_test(&rq->refcnt))
-                       complete(&rq->can_destroy);
-               wait_for_completion(&rq->can_destroy);
-               goto clean_rq;
-       }
-
        if (!atomic_dec_and_test(&rq->refcnt)) {
                sif_log(sdev, SIF_RQ, "rq %d still in use - ref.cnt %d",
                        rq_idx, atomic_read(&rq->refcnt));
                return -EBUSY;
        }
 
-clean_rq:
+       /* Reset rq pointers, for srq and the error path in create_qp.
+        * This also means that rq_sw will be reset twice in the
+        * happy path for !srq.
+        */
+       memset(rq_sw, 0, sizeof(*rq_sw));
+       set_psif_rq_hw__head_indx(&rq->d, 0);
+
        sif_release_rq(sdev, rq->index);
        return 0;
 }

diff --git a/drivers/infiniband/hw/sif/sif_rq.h b/drivers/infiniband/hw/sif/sif_rq.h
index f7aea2cb5a180679d8a6644a7e259e5d9855e200..70c83cd24fbdd6b0c22c08011ae9697c4dae44f5 100644 (file)
--- a/drivers/infiniband/hw/sif/sif_rq.h
+++ b/drivers/infiniband/hw/sif/sif_rq.h
@@ -26,7 +26,8 @@ struct sif_rq {
        bool is_srq; /* Set if this is a shared receive queue */
        int xrc_domain; /* If != 0: This is an XRC SRQ member of this domain idx */
        atomic_t refcnt; /* Ref.count for usage as a shared receive queue */
-       struct completion can_destroy; /* use refcnt to synchronization in !srq case */
+       atomic_t flush_in_progress; /* flush in progress synchronization */
+       struct completion can_reset; /* use flush_in_progress to synchronization reset and flush */
        u16 entries;      /* Allocated entries */
        u16 entries_user; /* Entries reported to user (entries -1 if max) */
        u32 sg_entries; /* Max receive scatter/gather configured for this rq */