[SCSI] fcoe: Out of order tx frames was causing several check condition SCSI status
authorVasu Dev <vasu.dev@intel.com>
Fri, 27 Feb 2009 18:56:27 +0000 (10:56 -0800)
committerJames Bottomley <James.Bottomley@HansenPartnership.com>
Tue, 10 Mar 2009 14:09:21 +0000 (09:09 -0500)
frames followed by these errors in log.

[sdp] Result: hostbyte=DID_OK driverbyte=DRIVER_SENSE,SUGGEST_OK
[sdp] Sense Key : Aborted Command [current]
[sdp] Add. Sense: Data phase error

This was causing some test apps to exit due to write failure under heavy
load.

This was due to a race around adding and removing tx frame skb in
fcoe_pending_queue, Chris Leech helped me to find that brief unlocking
period when pulling skb from fcoe_pending_queue in various contexts
(fcoe_watchdog and fcoe_xmit) and then adding skb back into fcoe_pending_queue
up on a failed fcoe_start_io could change skb/tx frame order in
fcoe_pending_queue. Thanks Chris.

This patch allows only single context to pull skb from fcoe_pending_queue
at any time to prevent above described ordering issue/race by use of
fcoe_pending_queue_active flag.

This patch simplified fcoe_watchdog with modified fcoe_check_wait_queue by
use of FCOE_LOW_QUEUE_DEPTH instead previously used several conditionals
to clear and set lp->qfull.

I think FCOE_MAX_QUEUE_DEPTH with FCOE_LOW_QUEUE_DEPTH  will work better
in re/setting lp->qfull and these could be fine tuned for performance.

Signed-off-by: Vasu Dev <vasu.dev@intel.com>
Signed-off-by: Robert Love <robert.w.love@intel.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
drivers/scsi/fcoe/fcoe_sw.c
drivers/scsi/fcoe/libfcoe.c
include/scsi/libfcoe.h

index 37d359d..da210eb 100644 (file)
@@ -188,6 +188,7 @@ static int fcoe_sw_netdev_config(struct fc_lport *lp, struct net_device *netdev)
 
 
        skb_queue_head_init(&fc->fcoe_pending_queue);
+       fc->fcoe_pending_queue_active = 0;
 
        /* setup Source Mac Address */
        memcpy(fc->ctl_src_addr, fc->real_dev->dev_addr,
index 7265e09..14dd8a0 100644 (file)
@@ -49,6 +49,7 @@
 static int debug_fcoe;
 
 #define FCOE_MAX_QUEUE_DEPTH  256
+#define FCOE_LOW_QUEUE_DEPTH  32
 
 /* destination address mode */
 #define FCOE_GW_ADDR_MODE          0x00
@@ -723,21 +724,12 @@ static void fcoe_recv_flogi(struct fcoe_softc *fc, struct fc_frame *fp, u8 *sa)
  */
 void fcoe_watchdog(ulong vp)
 {
-       struct fc_lport *lp;
        struct fcoe_softc *fc;
-       int qfilled = 0;
 
        read_lock(&fcoe_hostlist_lock);
        list_for_each_entry(fc, &fcoe_hostlist, list) {
-               lp = fc->lp;
-               if (lp) {
-                       if (fc->fcoe_pending_queue.qlen > FCOE_MAX_QUEUE_DEPTH)
-                               qfilled = 1;
-                       if (fcoe_check_wait_queue(lp) <  FCOE_MAX_QUEUE_DEPTH) {
-                               if (qfilled)
-                                       lp->qfull = 0;
-                       }
-               }
+               if (fc->lp)
+                       fcoe_check_wait_queue(fc->lp);
        }
        read_unlock(&fcoe_hostlist_lock);
 
@@ -753,8 +745,8 @@ void fcoe_watchdog(ulong vp)
  *
  * This empties the wait_queue, dequeue the head of the wait_queue queue
  * and calls fcoe_start_io() for each packet, if all skb have been
- * transmitted, return 0 if a error occurs, then restore wait_queue and
- * try again later.
+ * transmitted, return qlen or -1 if a error occurs, then restore
+ * wait_queue and  try again later.
  *
  * The wait_queue is used when the skb transmit fails. skb will go
  * in the wait_queue which will be emptied by the time function OR
@@ -764,33 +756,38 @@ void fcoe_watchdog(ulong vp)
  */
 static int fcoe_check_wait_queue(struct fc_lport *lp)
 {
-       int rc;
        struct sk_buff *skb;
        struct fcoe_softc *fc;
+       int rc = -1;
 
        fc = lport_priv(lp);
        spin_lock_bh(&fc->fcoe_pending_queue.lock);
 
-       /*
-        * if interface pending queue full then set qfull in lport.
-        */
-       if (fc->fcoe_pending_queue.qlen > FCOE_MAX_QUEUE_DEPTH)
-               lp->qfull = 1;
+       if (fc->fcoe_pending_queue_active)
+               goto out;
+       fc->fcoe_pending_queue_active = 1;
        if (fc->fcoe_pending_queue.qlen) {
                while ((skb = __skb_dequeue(&fc->fcoe_pending_queue)) != NULL) {
                        spin_unlock_bh(&fc->fcoe_pending_queue.lock);
                        rc = fcoe_start_io(skb);
-                       if (rc) {
+                       if (rc)
                                fcoe_insert_wait_queue_head(lp, skb);
-                               return rc;
-                       }
                        spin_lock_bh(&fc->fcoe_pending_queue.lock);
+                       if (rc)
+                               break;
                }
-               if (fc->fcoe_pending_queue.qlen < FCOE_MAX_QUEUE_DEPTH)
+               /*
+                * if interface pending queue is below FCOE_LOW_QUEUE_DEPTH
+                * then clear qfull flag.
+                */
+               if (fc->fcoe_pending_queue.qlen < FCOE_LOW_QUEUE_DEPTH)
                        lp->qfull = 0;
        }
+       fc->fcoe_pending_queue_active = 0;
+       rc = fc->fcoe_pending_queue.qlen;
+out:
        spin_unlock_bh(&fc->fcoe_pending_queue.lock);
-       return fc->fcoe_pending_queue.qlen;
+       return rc;
 }
 
 /**
index f43d383..941818f 100644 (file)
@@ -46,6 +46,7 @@ struct fcoe_softc {
        struct net_device *phys_dev;            /* device with ethtool_ops */
        struct packet_type  fcoe_packet_type;
        struct sk_buff_head fcoe_pending_queue;
+       u8      fcoe_pending_queue_active;
 
        u8 dest_addr[ETH_ALEN];
        u8 ctl_src_addr[ETH_ALEN];