Re: [PATCH] I/O space write barrier

[Jeremy Higdon]

On Thu, Sep 30, 2004 at 11:21:39PM +0200, Guennadi Liakhovetski wrote:
> 
> A pretty obvious note: instead of repeating this nice but pretty lengthy 
> comment 3 times in the same file, wouldn't it be better to write at 
> further locations something like
> 
> 	/* Enforce IO-ordering. See comment in <function> for details. */
> 
> Also helps if you later have to modify the comment, or move it, or add 
> more mmiowb()s, or do some other modifications.
> 
> Thanks
> Guennadi
> ---
> Guennadi Liakhovetski

Suggestion applied.  Now that the mmiowb is in the mm patch, this
patch to qla1280 should be ready for inclusion therein also.

signed-off-by: Jeremy Higdon <jeremy@xxxxxxx>


--- linux-2.6.9-rc4.orig/drivers/scsi/qla1280.c	2004-10-15 17:21:21.000000000 -0700
+++ linux-2.6.9-rc4/drivers/scsi/qla1280.c	2004-10-15 17:23:08.000000000 -0700
@@ -3409,7 +3409,8 @@
 	sp->flags |= SRB_SENT;
 	ha->actthreads++;
 	WRT_REG_WORD(&reg->mailbox4, ha->req_ring_index);
-	(void) RD_REG_WORD(&reg->mailbox4); /* PCI posted write flush */
+	/* Enforce mmio write ordering; see comment in qla1280_isp_cmd(). */
+	mmiowb();
 
  out:
 	if (status)
@@ -3677,7 +3678,8 @@
 	sp->flags |= SRB_SENT;
 	ha->actthreads++;
 	WRT_REG_WORD(&reg->mailbox4, ha->req_ring_index);
-	(void) RD_REG_WORD(&reg->mailbox4); /* PCI posted write flush */
+	/* Enforce mmio write ordering; see comment in qla1280_isp_cmd(). */
+	mmiowb();
 
 out:
 	if (status)
@@ -3787,9 +3789,21 @@
 	} else
 		ha->request_ring_ptr++;
 
-	/* Set chip new ring index. */
+	/*
+	 * Update request index to mailbox4 (Request Queue In).
+	 * The mmiowb() ensures that this write is ordered with writes by other
+	 * CPUs.  Without the mmiowb(), it is possible for the following:
+	 *    CPUA posts write of index 5 to mailbox4
+	 *    CPUA releases host lock
+	 *    CPUB acquires host lock
+	 *    CPUB posts write of index 6 to mailbox4
+	 *    On PCI bus, order reverses and write of 6 posts, then index 5,
+	 *       causing chip to issue full queue of stale commands
+	 * The mmiowb() prevents future writes from crossing the barrier.
+	 * See Documentation/DocBook/deviceiobook.tmpl for more information.
+	 */
 	WRT_REG_WORD(&reg->mailbox4, ha->req_ring_index);
-	(void) RD_REG_WORD(&reg->mailbox4); /* PCI posted write flush */
+	mmiowb();
 
 	LEAVE("qla1280_isp_cmd");
 }
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/