[libata] Disable R_OK (Early ACK) on SII 3726 PMP
In 2009, While running "cache read" performance test of drives behind
SII PMP we encountered a "all 5 drives" timeout on more than 30% of the
machines under test. This patch reduces the rate by a factor of about 70.
Low enough that we didn't care to further investigate the issue.
Performance impact with any sort of "normal" use was ~2%+ CPU and less
than 1% throughput degradation. Worst case impact (cached read) was
6% IOPS reduction. This is with NCQ off (q=1) but I believe FIS based
switching enabled in the SATA driver.
The patch disables "Early ACK" in the 3726 port multiplier.
"Early ACK" is issued when device sends a FIS to the host (via PMP)
and the PMP sends an ACK immediately back to the device - well before
the host gets the response. Under worst case IOPs load (cached read
test) and more than 2 PMPs connected to a 4-port SATA controller,
I suspect the time to service all of the PMPs is exceeding the PMPs
ability to keep track of outstanding FIS it owes the Host. Reducing
the number of PMPs to 2 (or 1) reduces the frequency by several orders
of magnitude. Kudos to Gwendal for initial debugging of this issue.
[Any errors in the description are mine, not his.]
Patch is currently in production on Google servers.
Signed-off-by: Grant Grundler <grundler@google.com>
Signed-off-by: Gwendal Grignou <gwendal@google.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jeff Garzik <jgarzik@redhat.com>
diff --git a/drivers/ata/libata-pmp.c b/drivers/ata/libata-pmp.c
index 00305f4..224faab 100644
--- a/drivers/ata/libata-pmp.c
+++ b/drivers/ata/libata-pmp.c
@@ -231,10 +231,14 @@
return "<unknown>";
}
+#define PMP_GSCR_SII_POL 129
+
static int sata_pmp_configure(struct ata_device *dev, int print_info)
{
struct ata_port *ap = dev->link->ap;
u32 *gscr = dev->gscr;
+ u16 vendor = sata_pmp_gscr_vendor(gscr);
+ u16 devid = sata_pmp_gscr_devid(gscr);
unsigned int err_mask = 0;
const char *reason;
int nr_ports, rc;
@@ -260,12 +264,34 @@
goto fail;
}
+ /* Disable sending Early R_OK.
+ * With "cached read" HDD testing and multiple ports busy on a SATA
+ * host controller, 3726 PMP will very rarely drop a deferred
+ * R_OK that was intended for the host. Symptom will be all
+ * 5 drives under test will timeout, get reset, and recover.
+ */
+ if (vendor == 0x1095 && devid == 0x3726) {
+ u32 reg;
+
+ err_mask = sata_pmp_read(&ap->link, PMP_GSCR_SII_POL, ®);
+ if (err_mask) {
+ rc = -EIO;
+ reason = "failed to read Sil3726 Private Register";
+ goto fail;
+ }
+ reg &= ~0x1;
+ err_mask = sata_pmp_write(&ap->link, PMP_GSCR_SII_POL, reg);
+ if (err_mask) {
+ rc = -EIO;
+ reason = "failed to write Sil3726 Private Register";
+ goto fail;
+ }
+ }
+
if (print_info) {
ata_dev_printk(dev, KERN_INFO, "Port Multiplier %s, "
"0x%04x:0x%04x r%d, %d ports, feat 0x%x/0x%x\n",
- sata_pmp_spec_rev_str(gscr),
- sata_pmp_gscr_vendor(gscr),
- sata_pmp_gscr_devid(gscr),
+ sata_pmp_spec_rev_str(gscr), vendor, devid,
sata_pmp_gscr_rev(gscr),
nr_ports, gscr[SATA_PMP_GSCR_FEAT_EN],
gscr[SATA_PMP_GSCR_FEAT]);