IB/srp: Make receive buffer handling more robust
authorBart Van Assche <bvanassche@acm.org>
Fri, 30 Jul 2010 10:59:05 +0000 (10:59 +0000)
committerRoland Dreier <rolandd@cisco.com>
Wed, 4 Aug 2010 18:47:39 +0000 (11:47 -0700)
The current strategy in ib_srp for posting receive buffers is:

 * Post one buffer after channel establishment.
 * Post one buffer before sending an SRP_CMD or SRP_TSK_MGMT to the target.

As a result, only the first non-SRP_RSP information unit from the
target will be processed.  If that first information unit is an
SRP_T_LOGOUT, it will be processed.  On the other hand, if the
initiator receives an SRP_CRED_REQ or SRP_AER_REQ before it receives a
SRP_T_LOGOUT, the SRP_T_LOGOUT won't be processed.

We can fix this inconsistency by changing the strategy for posting
receive buffers to:

 * Post all receive buffers after channel establishment.
 * After a receive buffer has been consumed and processed, post it again.

A side effect is that the ib_post_recv() call is moved out of the SCSI
command processing path.  Since __srp_post_recv() is not called
directly any more, get rid of it and move the code directly into
srp_post_recv().  Also, move srp_post_recv() up in the file to avoid a
forward declaration.

Signed-off-by: Bart Van Assche <bart.vanassche@gmail.com>
Acked-by: David Dillow <dave@thedillows.org>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/ulp/srp/ib_srp.c

index 4675def..ffdd2d1 100644 (file)
@@ -811,6 +811,38 @@ static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
        return len;
 }
 
+static int srp_post_recv(struct srp_target_port *target)
+{
+       unsigned long flags;
+       struct srp_iu *iu;
+       struct ib_sge list;
+       struct ib_recv_wr wr, *bad_wr;
+       unsigned int next;
+       int ret;
+
+       spin_lock_irqsave(target->scsi_host->host_lock, flags);
+
+       next     = target->rx_head & (SRP_RQ_SIZE - 1);
+       wr.wr_id = next;
+       iu       = target->rx_ring[next];
+
+       list.addr   = iu->dma;
+       list.length = iu->size;
+       list.lkey   = target->srp_host->srp_dev->mr->lkey;
+
+       wr.next     = NULL;
+       wr.sg_list  = &list;
+       wr.num_sge  = 1;
+
+       ret = ib_post_recv(target->qp, &wr, &bad_wr);
+       if (!ret)
+               ++target->rx_head;
+
+       spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
+
+       return ret;
+}
+
 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
 {
        struct srp_request *req;
@@ -868,6 +900,7 @@ static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
 {
        struct ib_device *dev;
        struct srp_iu *iu;
+       int res;
        u8 opcode;
 
        iu = target->rx_ring[wc->wr_id];
@@ -904,6 +937,11 @@ static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
 
        ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
                                      DMA_FROM_DEVICE);
+
+       res = srp_post_recv(target);
+       if (res != 0)
+               shost_printk(KERN_ERR, target->scsi_host,
+                            PFX "Recv failed with error code %d\n", res);
 }
 
 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
@@ -943,45 +981,6 @@ static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
        }
 }
 
-static int __srp_post_recv(struct srp_target_port *target)
-{
-       struct srp_iu *iu;
-       struct ib_sge list;
-       struct ib_recv_wr wr, *bad_wr;
-       unsigned int next;
-       int ret;
-
-       next     = target->rx_head & (SRP_RQ_SIZE - 1);
-       wr.wr_id = next;
-       iu       = target->rx_ring[next];
-
-       list.addr   = iu->dma;
-       list.length = iu->size;
-       list.lkey   = target->srp_host->srp_dev->mr->lkey;
-
-       wr.next     = NULL;
-       wr.sg_list  = &list;
-       wr.num_sge  = 1;
-
-       ret = ib_post_recv(target->qp, &wr, &bad_wr);
-       if (!ret)
-               ++target->rx_head;
-
-       return ret;
-}
-
-static int srp_post_recv(struct srp_target_port *target)
-{
-       unsigned long flags;
-       int ret;
-
-       spin_lock_irqsave(target->scsi_host->host_lock, flags);
-       ret = __srp_post_recv(target);
-       spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
-
-       return ret;
-}
-
 /*
  * Must be called with target->scsi_host->host_lock held to protect
  * req_lim and tx_head.  Lock cannot be dropped between call here and
@@ -1091,11 +1090,6 @@ static int srp_queuecommand(struct scsi_cmnd *scmnd,
                goto err;
        }
 
-       if (__srp_post_recv(target)) {
-               shost_printk(KERN_ERR, target->scsi_host, PFX "Recv failed\n");
-               goto err_unmap;
-       }
-
        ib_dma_sync_single_for_device(dev, iu->dma, srp_max_iu_len,
                                      DMA_TO_DEVICE);
 
@@ -1238,6 +1232,7 @@ static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
        int attr_mask = 0;
        int comp = 0;
        int opcode = 0;
+       int i;
 
        switch (event->event) {
        case IB_CM_REQ_ERROR:
@@ -1287,7 +1282,11 @@ static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
                if (target->status)
                        break;
 
-               target->status = srp_post_recv(target);
+               for (i = 0; i < SRP_RQ_SIZE; i++) {
+                       target->status = srp_post_recv(target);
+                       if (target->status)
+                               break;
+               }
                if (target->status)
                        break;