IB/ipath: Fix races with ib_resize_cq()
authorBryan O'Sullivan <bos@pathscale.com>
Thu, 28 Sep 2006 16:00:23 +0000 (09:00 -0700)
committerRoland Dreier <rolandd@cisco.com>
Thu, 28 Sep 2006 18:17:12 +0000 (11:17 -0700)
The resize CQ function changes the memory used to store the queue.
Other routines need to honor the lock before accessing the pointer
to the queue and verify that the head and tail are in range.

Signed-off-by: Bryan O'Sullivan <bryan.osullivan@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_cq.c

index 00440d5..87462e0 100644 (file)
@@ -46,7 +46,7 @@
  */
 void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
 {
-       struct ipath_cq_wc *wc = cq->queue;
+       struct ipath_cq_wc *wc;
        unsigned long flags;
        u32 head;
        u32 next;
@@ -57,6 +57,7 @@ void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
         * Note that the head pointer might be writable by user processes.
         * Take care to verify it is a sane value.
         */
+       wc = cq->queue;
        head = wc->head;
        if (head >= (unsigned) cq->ibcq.cqe) {
                head = cq->ibcq.cqe;
@@ -109,21 +110,27 @@ void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
 int ipath_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *entry)
 {
        struct ipath_cq *cq = to_icq(ibcq);
-       struct ipath_cq_wc *wc = cq->queue;
+       struct ipath_cq_wc *wc;
        unsigned long flags;
        int npolled;
+       u32 tail;
 
        spin_lock_irqsave(&cq->lock, flags);
 
+       wc = cq->queue;
+       tail = wc->tail;
+       if (tail > (u32) cq->ibcq.cqe)
+               tail = (u32) cq->ibcq.cqe;
        for (npolled = 0; npolled < num_entries; ++npolled, ++entry) {
-               if (wc->tail == wc->head)
+               if (tail == wc->head)
                        break;
-               *entry = wc->queue[wc->tail];
-               if (wc->tail >= cq->ibcq.cqe)
-                       wc->tail = 0;
+               *entry = wc->queue[tail];
+               if (tail >= cq->ibcq.cqe)
+                       tail = 0;
                else
-                       wc->tail++;
+                       tail++;
        }
+       wc->tail = tail;
 
        spin_unlock_irqrestore(&cq->lock, flags);
 
@@ -322,10 +329,16 @@ int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
        return 0;
 }
 
+/**
+ * ipath_resize_cq - change the size of the CQ
+ * @ibcq: the completion queue
+ *
+ * Returns 0 for success.
+ */
 int ipath_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
 {
        struct ipath_cq *cq = to_icq(ibcq);
-       struct ipath_cq_wc *old_wc = cq->queue;
+       struct ipath_cq_wc *old_wc;
        struct ipath_cq_wc *wc;
        u32 head, tail, n;
        int ret;
@@ -361,6 +374,7 @@ int ipath_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
         * Make sure head and tail are sane since they
         * might be user writable.
         */
+       old_wc = cq->queue;
        head = old_wc->head;
        if (head > (u32) cq->ibcq.cqe)
                head = (u32) cq->ibcq.cqe;