[PATCH] ibm_emac: fix graceful stop timeout handling
This patch fixes graceful stop timeout handling in PPC4xx EMAC driver.
Currently, when we stop TX/RX channels we just do some number of loops
without relying on actual spent time. This has finally bitten me on
one of our systems (heavy network traffic during start up, RX channel
is stopped several times to configure multicast list).
Graceful channel stop can take up to 1 frame time, so I've added
device specific timeout counter which depends on current link speed
and calls to udelay() to really wait required amount of time before
giving up.
Signed-off-by: Eugene Surovegin <ebs@ebshome.net>
Signed-off-by: Jeff Garzik <jgarzik@pobox.com>
diff --git a/drivers/net/ibm_emac/ibm_emac_core.c b/drivers/net/ibm_emac/ibm_emac_core.c
index eb7d694..1da8a66 100644
--- a/drivers/net/ibm_emac/ibm_emac_core.c
+++ b/drivers/net/ibm_emac/ibm_emac_core.c
@@ -65,7 +65,7 @@
*/
#define DRV_NAME "emac"
-#define DRV_VERSION "3.53"
+#define DRV_VERSION "3.54"
#define DRV_DESC "PPC 4xx OCP EMAC driver"
MODULE_DESCRIPTION(DRV_DESC);
@@ -158,6 +158,14 @@
#define PHY_POLL_LINK_ON HZ
#define PHY_POLL_LINK_OFF (HZ / 5)
+/* Graceful stop timeouts in us.
+ * We should allow up to 1 frame time (full-duplex, ignoring collisions)
+ */
+#define STOP_TIMEOUT_10 1230
+#define STOP_TIMEOUT_100 124
+#define STOP_TIMEOUT_1000 13
+#define STOP_TIMEOUT_1000_JUMBO 73
+
/* Please, keep in sync with struct ibm_emac_stats/ibm_emac_error_stats */
static const char emac_stats_keys[EMAC_ETHTOOL_STATS_COUNT][ETH_GSTRING_LEN] = {
"rx_packets", "rx_bytes", "tx_packets", "tx_bytes", "rx_packets_csum",
@@ -222,10 +230,12 @@
r = in_be32(&p->mr0);
if (r & EMAC_MR0_TXE) {
- int n = 300;
+ int n = dev->stop_timeout;
out_be32(&p->mr0, r & ~EMAC_MR0_TXE);
- while (!(in_be32(&p->mr0) & EMAC_MR0_TXI) && n)
+ while (!(in_be32(&p->mr0) & EMAC_MR0_TXI) && n) {
+ udelay(1);
--n;
+ }
if (unlikely(!n))
emac_report_timeout_error(dev, "TX disable timeout");
}
@@ -248,9 +258,11 @@
if (!(r & EMAC_MR0_RXE)) {
if (unlikely(!(r & EMAC_MR0_RXI))) {
/* Wait if previous async disable is still in progress */
- int n = 100;
- while (!(r = in_be32(&p->mr0) & EMAC_MR0_RXI) && n)
+ int n = dev->stop_timeout;
+ while (!(r = in_be32(&p->mr0) & EMAC_MR0_RXI) && n) {
+ udelay(1);
--n;
+ }
if (unlikely(!n))
emac_report_timeout_error(dev,
"RX disable timeout");
@@ -273,10 +285,12 @@
r = in_be32(&p->mr0);
if (r & EMAC_MR0_RXE) {
- int n = 300;
+ int n = dev->stop_timeout;
out_be32(&p->mr0, r & ~EMAC_MR0_RXE);
- while (!(in_be32(&p->mr0) & EMAC_MR0_RXI) && n)
+ while (!(in_be32(&p->mr0) & EMAC_MR0_RXI) && n) {
+ udelay(1);
--n;
+ }
if (unlikely(!n))
emac_report_timeout_error(dev, "RX disable timeout");
}
@@ -395,6 +409,7 @@
r = EMAC_MR1_BASE(emac_opb_mhz()) | EMAC_MR1_VLE | EMAC_MR1_IST;
if (dev->phy.duplex == DUPLEX_FULL)
r |= EMAC_MR1_FDE;
+ dev->stop_timeout = STOP_TIMEOUT_10;
switch (dev->phy.speed) {
case SPEED_1000:
if (emac_phy_gpcs(dev->phy.mode)) {
@@ -409,12 +424,16 @@
r |= EMAC_MR1_MF_1000;
r |= EMAC_MR1_RFS_16K;
gige = 1;
-
- if (dev->ndev->mtu > ETH_DATA_LEN)
+
+ if (dev->ndev->mtu > ETH_DATA_LEN) {
r |= EMAC_MR1_JPSM;
+ dev->stop_timeout = STOP_TIMEOUT_1000_JUMBO;
+ } else
+ dev->stop_timeout = STOP_TIMEOUT_1000;
break;
case SPEED_100:
r |= EMAC_MR1_MF_100;
+ dev->stop_timeout = STOP_TIMEOUT_100;
/* Fall through */
default:
r |= EMAC_MR1_RFS_4K;
@@ -2048,6 +2067,7 @@
dev->phy.duplex = DUPLEX_FULL;
dev->phy.autoneg = AUTONEG_DISABLE;
dev->phy.pause = dev->phy.asym_pause = 0;
+ dev->stop_timeout = STOP_TIMEOUT_100;
init_timer(&dev->link_timer);
dev->link_timer.function = emac_link_timer;
dev->link_timer.data = (unsigned long)dev;
diff --git a/drivers/net/ibm_emac/ibm_emac_core.h b/drivers/net/ibm_emac/ibm_emac_core.h
index e9b44d0..911abba 100644
--- a/drivers/net/ibm_emac/ibm_emac_core.h
+++ b/drivers/net/ibm_emac/ibm_emac_core.h
@@ -189,6 +189,8 @@
struct timer_list link_timer;
int reset_failed;
+ int stop_timeout; /* in us */
+
struct ibm_emac_error_stats estats;
struct net_device_stats nstats;