RDMA/cxgb4: Avoid false GTS CIDX_INC overflows

The T4 IQ hw design assumes CIDX_INC credits will be returned on a
regular basis and always before the CIDX counter crosses over the PIDX
counter.  For RDMA CQs, however, returning CIDX_INC credits is only
needed and desired when and if the CQ is armed for notification.  This
can lead to a GTS write returning credits that causes the HW to reject
the credit update because it causes CIDX to pass PIDX.  Once this
happens, the CIDX/PIDX counters get out of whack and an application
can miss a notification and get stuck blocked awaiting a notification.

To avoid this, we allocate the HW IQ 2x times the requested size.
This seems to avoid the false overflow failures.  If we see more
issues with this, then we'll have to add code in the poll path to
return credits periodically like when the amount reaches 1/2 the queue
depth).  I would like to avoid this as it adds a PCI write transaction
for applications that never arm the CQ (like most MPIs).

Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

1 files changed, 20 insertions, 5 deletions

diff --git a/drivers/infiniband/hw/cxgb4/cq.c b/drivers/infiniband/hw/cxgb4/cq.c
index e1317f581168..fac5c6e68011 100644
--- a/drivers/infiniband/hw/cxgb4/cq.c
+++ b/drivers/infiniband/hw/cxgb4/cq.c
@@ -764,7 +764,7 @@ struct ib_cq *c4iw_create_cq(struct ib_device *ibdev, int entries,
         struct c4iw_create_cq_resp uresp;
         struct c4iw_ucontext *ucontext = NULL;
         int ret;
-        size_t memsize;
+        size_t memsize, hwentries;
         struct c4iw_mm_entry *mm, *mm2;
 
         PDBG("%s ib_dev %p entries %d\n", __func__, ibdev, entries);
@@ -788,14 +788,29 @@ struct ib_cq *c4iw_create_cq(struct ib_device *ibdev, int entries,
          * entries must be multiple of 16 for HW.
          */
         entries = roundup(entries, 16);
-        memsize = entries * sizeof *chp->cq.queue;
+
+        /*
+         * Make actual HW queue 2x to avoid cdix_inc overflows.
+         */
+        hwentries = entries * 2;
+
+        /*
+         * Make HW queue at least 64 entries so GTS updates aren't too
+         * frequent.
+         */
+        if (hwentries < 64)
+                hwentries = 64;
+
+        memsize = hwentries * sizeof *chp->cq.queue;
 
         /*
          * memsize must be a multiple of the page size if its a user cq.
          */
-        if (ucontext)
+        if (ucontext) {
                 memsize = roundup(memsize, PAGE_SIZE);
-        chp->cq.size = entries;
+                hwentries = memsize / sizeof *chp->cq.queue;
+        }
+        chp->cq.size = hwentries;
         chp->cq.memsize = memsize;
 
         ret = create_cq(&rhp->rdev, &chp->cq,
@@ -805,7 +820,7 @@ struct ib_cq *c4iw_create_cq(struct ib_device *ibdev, int entries,
 
         chp->rhp = rhp;
         chp->cq.size--;                         /* status page */
-        chp->ibcq.cqe = chp->cq.size - 1;
+        chp->ibcq.cqe = entries - 2;
         spin_lock_init(&chp->lock);
         atomic_set(&chp->refcnt, 1);
         init_waitqueue_head(&chp->wait);