Hi, Jeff.

Jeff Garzik wrote:

Tejun Heo wrote:

quoted

 sata_sil Mod15Write workaround was broken by the following commit by
Albert Lee.

Commit: 21b1ed74ee3667dcabcba92e486988ea9119a085
[PATCH] libata: Prevent the interrupt handler from completing a 
command twice

 This commit clears ATA_QCFLAG_ACTIVE in ata_qc_complete() and doesn't
handle IRQ if ATA_QCFLAG_ACTIVE is cleared on entry to interrupt
routine.  As m15w workaround executes single command multiple times,
the flag is cleared after the first chunk completion and the following
interrupt gets ignored resulting in "nobody cared" interrupt error.

 The following changes are made in m15w workaround to fix this.

 * Moved clearing of ATA_QCFLAG_ACTIVE before invoking ->complete_fn,
   so that ->complete_fn can mangle with the flag.  This doesn't affect
   any users.
 * Added setting ATA_QCFLAG_ACTIVE in m15w chunk completion function.

quoted

diff --git a/drivers/scsi/libata-core.c b/drivers/scsi/libata-core.c
--- a/drivers/scsi/libata-core.c
+++ b/drivers/scsi/libata-core.c

@@ -3087,8 +3087,8 @@ void ata_qc_complete(struct ata_queued_c
         ata_sg_clean(qc);
 
     /* call completion callback */
-    rc = qc->complete_fn(qc, drv_stat);
     qc->flags &= ~ATA_QCFLAG_ACTIVE;
+    rc = qc->complete_fn(qc, drv_stat);
 
     /* if callback indicates not to complete command (non-zero),
      * return immediately

I'm leaning towards applying latest Albert's ATA_QCFLAG_ACTIVE fix, 
which does the same thing your patch did.

For various complex tasks in the SCSI translation layer (SAT), we may 
wish to issue multiple ATA commands, before signalling completion.  As 
your mod15write patch is an example of this, it helps point out what 
parts of libata need work in order to accomplish this.

Any comments before I apply Albert's patch?

  I think the following patch I've posted deal with the same problem.

http://marc.theaimsgroup.com/?l=linux-ide&m=112454734102242&w=2

  Which, I think, is a better way to fix it.  It's dependant on the 
previous patchset you've just NAK'ed, so it cannot be applied but please 
take a look.  The base of the problem is that we run both 
ata_qc_complete and EH concurrently.

-- 
tejun