Merge git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[cascardo/linux.git] / drivers / infiniband / hw / ipath / ipath_cq.c
index 049221b..3e9241b 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;
@@ -75,7 +76,20 @@ void ipath_cq_enter(struct ipath_cq *cq, struct ib_wc *entry, int solicited)
                }
                return;
        }
-       wc->queue[head] = *entry;
+       wc->queue[head].wr_id = entry->wr_id;
+       wc->queue[head].status = entry->status;
+       wc->queue[head].opcode = entry->opcode;
+       wc->queue[head].vendor_err = entry->vendor_err;
+       wc->queue[head].byte_len = entry->byte_len;
+       wc->queue[head].imm_data = (__u32 __force)entry->imm_data;
+       wc->queue[head].qp_num = entry->qp->qp_num;
+       wc->queue[head].src_qp = entry->src_qp;
+       wc->queue[head].wc_flags = entry->wc_flags;
+       wc->queue[head].pkey_index = entry->pkey_index;
+       wc->queue[head].slid = entry->slid;
+       wc->queue[head].sl = entry->sl;
+       wc->queue[head].dlid_path_bits = entry->dlid_path_bits;
+       wc->queue[head].port_num = entry->port_num;
        wc->head = next;
 
        if (cq->notify == IB_CQ_NEXT_COMP ||
@@ -109,21 +123,48 @@ 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)
+               struct ipath_qp *qp;
+
+               if (tail == wc->head)
                        break;
-               *entry = wc->queue[wc->tail];
-               if (wc->tail >= cq->ibcq.cqe)
-                       wc->tail = 0;
+
+               qp = ipath_lookup_qpn(&to_idev(cq->ibcq.device)->qp_table,
+                                     wc->queue[tail].qp_num);
+               entry->qp = &qp->ibqp;
+               if (atomic_dec_and_test(&qp->refcount))
+                       wake_up(&qp->wait);
+
+               entry->wr_id = wc->queue[tail].wr_id;
+               entry->status = wc->queue[tail].status;
+               entry->opcode = wc->queue[tail].opcode;
+               entry->vendor_err = wc->queue[tail].vendor_err;
+               entry->byte_len = wc->queue[tail].byte_len;
+               entry->imm_data = wc->queue[tail].imm_data;
+               entry->src_qp = wc->queue[tail].src_qp;
+               entry->wc_flags = wc->queue[tail].wc_flags;
+               entry->pkey_index = wc->queue[tail].pkey_index;
+               entry->slid = wc->queue[tail].slid;
+               entry->sl = wc->queue[tail].sl;
+               entry->dlid_path_bits = wc->queue[tail].dlid_path_bits;
+               entry->port_num = wc->queue[tail].port_num;
+               if (tail >= cq->ibcq.cqe)
+                       tail = 0;
                else
-                       wc->tail++;
+                       tail++;
        }
+       wc->tail = tail;
 
        spin_unlock_irqrestore(&cq->lock, flags);
 
@@ -163,7 +204,7 @@ static void send_complete(unsigned long data)
  *
  * Called by ib_create_cq() in the generic verbs code.
  */
-struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries,
+struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries, int comp_vector,
                              struct ib_ucontext *context,
                              struct ib_udata *udata)
 {
@@ -177,11 +218,6 @@ struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries,
                goto done;
        }
 
-       if (dev->n_cqs_allocated == ib_ipath_max_cqs) {
-               ret = ERR_PTR(-ENOMEM);
-               goto done;
-       }
-
        /* Allocate the completion queue structure. */
        cq = kmalloc(sizeof(*cq), GFP_KERNEL);
        if (!cq) {
@@ -207,36 +243,40 @@ struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries,
         * See ipath_mmap() for details.
         */
        if (udata && udata->outlen >= sizeof(__u64)) {
-               struct ipath_mmap_info *ip;
-               __u64 offset = (__u64) wc;
                int err;
+               u32 s = sizeof *wc + sizeof(struct ib_wc) * entries;
 
-               err = ib_copy_to_udata(udata, &offset, sizeof(offset));
-               if (err) {
-                       ret = ERR_PTR(err);
+               cq->ip = ipath_create_mmap_info(dev, s, context, wc);
+               if (!cq->ip) {
+                       ret = ERR_PTR(-ENOMEM);
                        goto bail_wc;
                }
 
-               /* Allocate info for ipath_mmap(). */
-               ip = kmalloc(sizeof(*ip), GFP_KERNEL);
-               if (!ip) {
-                       ret = ERR_PTR(-ENOMEM);
-                       goto bail_wc;
+               err = ib_copy_to_udata(udata, &cq->ip->offset,
+                                      sizeof(cq->ip->offset));
+               if (err) {
+                       ret = ERR_PTR(err);
+                       goto bail_ip;
                }
-               cq->ip = ip;
-               ip->context = context;
-               ip->obj = wc;
-               kref_init(&ip->ref);
-               ip->mmap_cnt = 0;
-               ip->size = PAGE_ALIGN(sizeof(*wc) +
-                                     sizeof(struct ib_wc) * entries);
-               spin_lock_irq(&dev->pending_lock);
-               ip->next = dev->pending_mmaps;
-               dev->pending_mmaps = ip;
-               spin_unlock_irq(&dev->pending_lock);
        } else
                cq->ip = NULL;
 
+       spin_lock(&dev->n_cqs_lock);
+       if (dev->n_cqs_allocated == ib_ipath_max_cqs) {
+               spin_unlock(&dev->n_cqs_lock);
+               ret = ERR_PTR(-ENOMEM);
+               goto bail_ip;
+       }
+
+       dev->n_cqs_allocated++;
+       spin_unlock(&dev->n_cqs_lock);
+
+       if (cq->ip) {
+               spin_lock_irq(&dev->pending_lock);
+               list_add(&cq->ip->pending_mmaps, &dev->pending_mmaps);
+               spin_unlock_irq(&dev->pending_lock);
+       }
+
        /*
         * ib_create_cq() will initialize cq->ibcq except for cq->ibcq.cqe.
         * The number of entries should be >= the number requested or return
@@ -253,15 +293,14 @@ struct ib_cq *ipath_create_cq(struct ib_device *ibdev, int entries,
 
        ret = &cq->ibcq;
 
-       dev->n_cqs_allocated++;
        goto done;
 
+bail_ip:
+       kfree(cq->ip);
 bail_wc:
        vfree(wc);
-
 bail_cq:
        kfree(cq);
-
 done:
        return ret;
 }
@@ -280,7 +319,9 @@ int ipath_destroy_cq(struct ib_cq *ibcq)
        struct ipath_cq *cq = to_icq(ibcq);
 
        tasklet_kill(&cq->comptask);
+       spin_lock(&dev->n_cqs_lock);
        dev->n_cqs_allocated--;
+       spin_unlock(&dev->n_cqs_lock);
        if (cq->ip)
                kref_put(&cq->ip->ref, ipath_release_mmap_info);
        else
@@ -293,17 +334,18 @@ int ipath_destroy_cq(struct ib_cq *ibcq)
 /**
  * ipath_req_notify_cq - change the notification type for a completion queue
  * @ibcq: the completion queue
- * @notify: the type of notification to request
+ * @notify_flags: the type of notification to request
  *
  * Returns 0 for success.
  *
  * This may be called from interrupt context.  Also called by
  * ib_req_notify_cq() in the generic verbs code.
  */
-int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
+int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags notify_flags)
 {
        struct ipath_cq *cq = to_icq(ibcq);
        unsigned long flags;
+       int ret = 0;
 
        spin_lock_irqsave(&cq->lock, flags);
        /*
@@ -311,15 +353,27 @@ int ipath_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify notify)
         * any other transitions (see C11-31 and C11-32 in ch. 11.4.2.2).
         */
        if (cq->notify != IB_CQ_NEXT_COMP)
-               cq->notify = notify;
+               cq->notify = notify_flags & IB_CQ_SOLICITED_MASK;
+
+       if ((notify_flags & IB_CQ_REPORT_MISSED_EVENTS) &&
+           cq->queue->head != cq->queue->tail)
+               ret = 1;
+
        spin_unlock_irqrestore(&cq->lock, flags);
-       return 0;
+
+       return ret;
 }
 
+/**
+ * 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;
@@ -355,6 +409,7 @@ 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;
@@ -389,13 +444,12 @@ int ipath_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
        if (cq->ip) {
                struct ipath_ibdev *dev = to_idev(ibcq->device);
                struct ipath_mmap_info *ip = cq->ip;
+               u32 s = sizeof *wc + sizeof(struct ib_wc) * cqe;
 
-               ip->obj = wc;
-               ip->size = PAGE_ALIGN(sizeof(*wc) +
-                                     sizeof(struct ib_wc) * cqe);
+               ipath_update_mmap_info(dev, ip, s, wc);
                spin_lock_irq(&dev->pending_lock);
-               ip->next = dev->pending_mmaps;
-               dev->pending_mmaps = ip;
+               if (list_empty(&ip->pending_mmaps))
+                       list_add(&ip->pending_mmaps, &dev->pending_mmaps);
                spin_unlock_irq(&dev->pending_lock);
        }