@@ -4793,21 +4793,26 @@ void swap_buf_le16(u16 *buf, unsigned int buf_words)
static struct ata_queued_cmd *ata_qc_new(struct ata_port *ap)
{
struct ata_queued_cmd *qc = NULL;
- unsigned int i;
+ unsigned int i, tag;
/* no command while frozen */
if (unlikely(ap->pflags & ATA_PFLAG_FROZEN))
return NULL;
- /* the last tag is reserved for internal command. */
- for (i = 0; i < ATA_MAX_QUEUE - 1; i++)
- if (!test_and_set_bit(i, &ap->qc_allocated)) {
- qc = __ata_qc_from_tag(ap, i);
+ for (i = 0; i < ATA_MAX_QUEUE; i++) {
+ tag = (i + ap->tag_order) & (ATA_MAX_QUEUE - 1);
+
+ /* the last tag is reserved for internal command. */
+ if (tag == ATA_TAG_INTERNAL)
+ continue;
+
+ if (!test_and_set_bit(tag, &ap->qc_allocated)) {
+ qc = __ata_qc_from_tag(ap, tag);
+ qc->tag = tag;
+ ap->tag_order = (tag + 1) & (ATA_MAX_QUEUE - 1);
break;
}
-
- if (qc)
- qc->tag = i;
+ }
return qc;
}
@@ -822,6 +822,7 @@ struct ata_port {
unsigned long qc_allocated;
unsigned int qc_active;
int nr_active_links; /* #links with active qcs */
+ unsigned int tag_order; /* next tag hw expects */
struct ata_link link; /* host default link */
struct ata_link *slave_link; /* see ata_slave_link_init() */
The AHCI spec allows implementations to issue commands in tag order rather than FIFO order: 5.3.2.12 P:SelectCmd HBA sets pSlotLoc = (pSlotLoc + 1) mod (CAP.NCS + 1) or HBA selects the command to issue that has had the PxCI bit set to '1' longer than any other command pending to be issued. The result is that commands posted sequentially (time-wise) may play out of sequence when issued by hardware. This behavior has likely been hidden by drives that arrange for commands to complete in issue order. However, it appears recent drives (two from different vendors that we have found so far) inflict out-of-order completions as a matter of course. So, we need to take care to maintain ordered submission, otherwise we risk triggering a drive to fall out of sequential-io automation and back to random-io processing, which incurs large latency and degrades throughput. This issue was found in simple benchmarks where QD=2 seq-write performance was 30-50% *greater* than QD=32 seq-write performance. Tagging for -stable and making the change globally since it has a low risk-to-reward ratio. Also, word is that recent versions of an unnamed OS also does it this way now. So, drives in the field are already experienced with this tag ordering scheme. Cc: <stable@vger.kernel.org> Cc: Dave Jiang <dave.jiang@intel.com> Cc: Ed Ciechanowski <ed.ciechanowski@intel.com> Signed-off-by: Dan Williams <dan.j.williams@intel.com> --- drivers/ata/libata-core.c | 21 +++++++++++++-------- include/linux/libata.h | 1 + 2 files changed, 14 insertions(+), 8 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-ide" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html