aoe: commands in retransmit queue use new destination on failure
authorEd Cashin <ecashin@coraid.com>
Tue, 18 Dec 2012 00:03:51 +0000 (16:03 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 18 Dec 2012 01:15:24 +0000 (17:15 -0800)
When one remote MAC address isn't working as a destination for AoE
commands, the frames used to track information associated with the AoE
commands are moved to a new aoetgt (defined by the tuple of {AoE major,
AoE minor, target MAC address}).

This patch makes sure that the frames on the queue for retransmits that
need to be done are updated to use the new destination, so that
retransmits will be sent through a working network path.

Without this change, packets on the retransmit queue will be needlessly
retransmitted to the unresponsive destination MAC, possibly causing
premature target failure before there's time for the retransmit timer to
run again, decide to retransmit again, and finally update the destination
to a working MAC address on the AoE target.

Signed-off-by: Ed Cashin <ecashin@coraid.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/block/aoe/aoe.h
drivers/block/aoe/aoecmd.c
drivers/block/aoe/aoedev.c

index 9fb68fc3b280ccefcdbf224126fc1b5858bc15c3..c253cca4985be261df09f9b02dcdb55c35e98e12 100644 (file)
@@ -108,6 +108,7 @@ struct frame {
        struct timeval sent;    /* high-res time packet was sent */
        u32 sent_jiffs;         /* low-res jiffies-based sent time */
        ulong waited;
+       ulong waited_total;
        struct aoetgt *t;               /* parent target I belong to */
        sector_t lba;
        struct sk_buff *skb;            /* command skb freed on module exit */
index a99220ad62628b117ed9f089a37034afdb479143..d9bc6ff56f49dcc044177c26833e625aaa6877c6 100644 (file)
@@ -352,6 +352,7 @@ aoecmd_ata_rw(struct aoedev *d)
        fhash(f);
        t->nout++;
        f->waited = 0;
+       f->waited_total = 0;
        f->buf = buf;
        f->bcnt = bcnt;
        f->lba = buf->sector;
@@ -556,46 +557,69 @@ ejectif(struct aoetgt *t, struct aoeif *ifp)
        dev_put(nd);
 }
 
+static struct frame *
+reassign_frame(struct list_head *pos)
+{
+       struct frame *f;
+       struct frame *nf;
+       struct sk_buff *skb;
+
+       f = list_entry(pos, struct frame, head);
+       nf = newframe(f->t->d);
+       if (!nf)
+               return NULL;
+
+       list_del(pos);
+
+       skb = nf->skb;
+       nf->skb = f->skb;
+       nf->buf = f->buf;
+       nf->bcnt = f->bcnt;
+       nf->lba = f->lba;
+       nf->bv = f->bv;
+       nf->bv_off = f->bv_off;
+       nf->waited = 0;
+       nf->waited_total = f->waited_total;
+       nf->sent = f->sent;
+       f->skb = skb;
+       aoe_freetframe(f);
+       f->t->nout--;
+       nf->t->nout++;
+
+       return nf;
+}
+
 static int
 sthtith(struct aoedev *d)
 {
        struct frame *f, *nf;
        struct list_head *nx, *pos, *head;
-       struct sk_buff *skb;
        struct aoetgt *ht = d->htgt;
        int i;
 
+       /* look through the active and pending retransmit frames */
        for (i = 0; i < NFACTIVE; i++) {
                head = &d->factive[i];
                list_for_each_safe(pos, nx, head) {
                        f = list_entry(pos, struct frame, head);
                        if (f->t != ht)
                                continue;
-
-                       nf = newframe(d);
+                       nf = reassign_frame(pos);
                        if (!nf)
                                return 0;
-
-                       /* remove frame from active list */
-                       list_del(pos);
-
-                       /* reassign all pertinent bits to new outbound frame */
-                       skb = nf->skb;
-                       nf->skb = f->skb;
-                       nf->buf = f->buf;
-                       nf->bcnt = f->bcnt;
-                       nf->lba = f->lba;
-                       nf->bv = f->bv;
-                       nf->bv_off = f->bv_off;
-                       nf->waited = 0;
-                       nf->sent_jiffs = f->sent_jiffs;
-                       f->skb = skb;
-                       aoe_freetframe(f);
-                       ht->nout--;
-                       nf->t->nout++;
                        resend(d, nf);
                }
        }
+       head = &d->rexmitq;
+       list_for_each_safe(pos, nx, head) {
+               f = list_entry(pos, struct frame, head);
+               if (f->t != ht)
+                       continue;
+               nf = reassign_frame(pos);
+               if (!nf)
+                       return 0;
+               resend(d, nf);
+       }
        /* We've cleaned up the outstanding so take away his
         * interfaces so he won't be used.  We should remove him from
         * the target array here, but cleaning up a target is
@@ -612,6 +636,7 @@ rexmit_deferred(struct aoedev *d)
        struct aoetgt *t;
        struct frame *f;
        struct list_head *pos, *nx, *head;
+       int since;
 
        head = &d->rexmitq;
        list_for_each_safe(pos, nx, head) {
@@ -621,6 +646,9 @@ rexmit_deferred(struct aoedev *d)
                        continue;
                list_del(pos);
                t->nout++;
+               since = tsince_hr(f);
+               f->waited += since;
+               f->waited_total += since;
                resend(d, f);
        }
 }
@@ -637,6 +665,7 @@ rexmit_timer(ulong vp)
        register long timeout;
        ulong flags, n;
        int i;
+       int since;
 
        d = (struct aoedev *) vp;
 
@@ -669,7 +698,8 @@ rexmit_timer(ulong vp)
        while (!list_empty(&flist)) {
                pos = flist.next;
                f = list_entry(pos, struct frame, head);
-               n = f->waited += tsince_hr(f);
+               since = tsince_hr(f);
+               n = f->waited_total + since;
                n /= USEC_PER_SEC;
                if (n > aoe_deadsecs) {
                        /* Waited too long.  Device failure.
@@ -1301,6 +1331,7 @@ aoecmd_ata_id(struct aoedev *d)
        fhash(f);
        t->nout++;
        f->waited = 0;
+       f->waited_total = 0;
 
        /* set up ata header */
        ah->scnt = 1;
index 3c3aef2081069b14dd55da7bf26da8707566eb6e..91f7c99eb2a922b8dae749667aaf82960664b843 100644 (file)
@@ -170,30 +170,40 @@ aoe_failip(struct aoedev *d)
                aoe_end_request(d, rq, 0);
 }
 
+static void
+downdev_frame(struct list_head *pos)
+{
+       struct frame *f;
+
+       f = list_entry(pos, struct frame, head);
+       list_del(pos);
+       if (f->buf) {
+               f->buf->nframesout--;
+               aoe_failbuf(f->t->d, f->buf);
+       }
+       aoe_freetframe(f);
+}
+
 void
 aoedev_downdev(struct aoedev *d)
 {
        struct aoetgt *t, **tt, **te;
-       struct frame *f;
        struct list_head *head, *pos, *nx;
        struct request *rq;
        int i;
 
        d->flags &= ~DEVFL_UP;
 
-       /* clean out active buffers */
+       /* clean out active and to-be-retransmitted buffers */
        for (i = 0; i < NFACTIVE; i++) {
                head = &d->factive[i];
-               list_for_each_safe(pos, nx, head) {
-                       f = list_entry(pos, struct frame, head);
-                       list_del(pos);
-                       if (f->buf) {
-                               f->buf->nframesout--;
-                               aoe_failbuf(d, f->buf);
-                       }
-                       aoe_freetframe(f);
-               }
+               list_for_each_safe(pos, nx, head)
+                       downdev_frame(pos);
        }
+       head = &d->rexmitq;
+       list_for_each_safe(pos, nx, head)
+               downdev_frame(pos);
+
        /* reset window dressings */
        tt = d->targets;
        te = tt + NTARGETS;