I was able to reproduce it in my same test case with NCQ depth set to 1.
Of course, there were still some cache flushes going on there, so I'm not
certain that test really told us anything new. I'm rather doutbful that it's
related to reusing the same tag now, though, based on the tests I've been
doing. It may be something to do with switching between NCQ and non-NCQ
commands, maybe the controller isn't able to handle doing that too rapidly.
This patch seems to fix the problem - or at least it hasn't failed the tests that
I've run so far. It's not really ideal though, so I'd like to do some more
investigation/testing before proclaiming it as a fix. Experimentally, it
appears that 10 microseconds is not enough delay, but 20 seems to work better.
Hints from the peanut gallery remain welcome.
--- linux-2.6.20-git6edit/drivers/ata/sata_nv.c.before_hacking 2007-02-15 18:19:13.000000000 -0600
+++ linux-2.6.20-git6edit/drivers/ata/sata_nv.c 2007-02-15 22:36:02.000000000 -0600
@@ -219,6 +219,7 @@
void __iomem * gen_block;
void __iomem * notifier_clear_block;
u8 flags;
+ int last_issue_ncq;
};
struct nv_host_priv {
@@ -1260,6 +1261,7 @@
{
struct nv_adma_port_priv *pp = qc->ap->private_data;
void __iomem *mmio = pp->ctl_block;
+ int curr_ncq = (qc->tf.protocol == ATA_PROT_NCQ);
VPRINTK("ENTER\n");
@@ -1274,6 +1276,14 @@
/* write append register, command tag in lower 8 bits
and (number of cpbs to append -1) in top 8 bits */
wmb();
+
+ if(curr_ncq != pp->last_issue_ncq) {
+ /* Seems to need some delay before switching between NCQ and non-NCQ
+ commands, else we get command timeouts and such. */
+ udelay(20);
+ pp->last_issue_ncq = curr_ncq;
+ }
+
writew(qc->tag, mmio + NV_ADMA_APPEND);
DPRINTK("Issued tag %u\n",qc->tag);
-
