Skip to contentSkip to conversation

libertas: reduce command retry time

PATCH In review Holger Schurig 4 participants 6 messages Apr 30, 2008 – May 2, 2008 Networking drivers (wireless)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 libertas: reduce command retry time

In the normal case, an unsuccessful command would be tried for 10*5 seconds, or 10*10 seconds in the worst case. This patch reduces this to 3*1 seconds, or 3*5 seconds in the worst case.

Also changed the needless del_timer_sync(&priv->command_timer); to del_timer().

---

The story behind this patch is the following: let the driver load, and let the hardware have a failure, e.g. because of a bad power supply to the device. Now try "pccardctl eject". It looks like this is now hanging, nothing happens. Only when all the retries are done does the "pccardctrl eject" do what it is supposed to do, that might have been in 50 or 100 seconds.

As I didn't have seen a case where it makes sense to retry a faulty command 10 times I changed this number.

Also, I've never seen a command (except scan/assoc) to take more than a few milliseconds. So giving them 5 seconds seemed excessive.

drivers/net/wireless/libertas/cmd.c +3 -3
Index: wireless-testing/drivers/net/wireless/libertas/cmd.c
===================================================================
@@ -1144,7 +1144,7 @@ static void lbs_submit_command(struct lb
 	struct cmd_header *cmd;
 	uint16_t cmdsize;
 	uint16_t command;
-	int timeo = 5 * HZ;
+	int timeo = HZ;
 	int ret;
 
 	lbs_deb_enter(LBS_DEB_HOST);
@@ -1162,7 +1162,7 @@ static void lbs_submit_command(struct lb
 	/* These commands take longer */
 	if (command == CMD_802_11_SCAN || command == CMD_802_11_ASSOCIATE ||
 	    command == CMD_802_11_AUTHENTICATE)
-		timeo = 10 * HZ;
+		timeo = 5 * HZ;
 
 	lbs_deb_cmd("DNLD_CMD: command 0x%04x, seq %d, size %d\n",
 		     command, le16_to_cpu(cmd->seqnum), cmdsize);
@@ -1174,7 +1174,7 @@ static void lbs_submit_command(struct lb
 		lbs_pr_info("DNLD_CMD: hw_host_to_card failed: %d\n", ret);
 		/* Let the timer kick in and retry, and potentially reset
 		   the whole thing if the condition persists */
-		timeo = HZ;
+		timeo = HZ/2;
 	}
 
 	/* Setup the timer after transmit command */
drivers/net/wireless/libertas/main.c +12 -10
Index: wireless-testing/drivers/net/wireless/libertas/main.c
===================================================================
@@ -749,15 +749,18 @@ static int lbs_thread(void *data)
 		if (priv->cmd_timed_out && priv->cur_cmd) {
 			struct cmd_ctrl_node *cmdnode = priv->cur_cmd;
 
-			if (++priv->nr_retries > 10) {
-				lbs_pr_info("Excessive timeouts submitting command %x\n",
-					    le16_to_cpu(cmdnode->cmdbuf->command));
+			if (++priv->nr_retries > 3) {
+				lbs_pr_info("excessive timeouts submitting "
+					"command 0x%04x\n",
+					le16_to_cpu(cmdnode->cmdbuf->command));
 				lbs_complete_command(priv, cmdnode, -ETIMEDOUT);
 				priv->nr_retries = 0;
 			} else {
 				priv->cur_cmd = NULL;
-				lbs_pr_info("requeueing command %x due to timeout (#%d)\n",
-					    le16_to_cpu(cmdnode->cmdbuf->command), priv->nr_retries);
+				lbs_pr_info("requeueing command 0x%04x due "
+					"to timeout (#%d)\n",
+					le16_to_cpu(cmdnode->cmdbuf->command),
+					priv->nr_retries);
 
 				/* Stick it back at the _top_ of the pending queue
 				   for immediate resubmission */
@@ -958,12 +961,11 @@ static void command_timer_fn(unsigned lo
 	lbs_deb_enter(LBS_DEB_CMD);
 	spin_lock_irqsave(&priv->driver_lock, flags);
 
-	if (!priv->cur_cmd) {
-		lbs_pr_info("Command timer expired; no pending command\n");
+	if (!priv->cur_cmd)
 		goto out;
-	}
 
-	lbs_pr_info("Command %x timed out\n", le16_to_cpu(priv->cur_cmd->cmdbuf->command));
+	lbs_pr_info("command 0x%04x timed out\n",
+		le16_to_cpu(priv->cur_cmd->cmdbuf->command));
 
 	priv->cmd_timed_out = 1;
 	wake_up_interruptible(&priv->waitq);
@@ -1278,7 +1280,7 @@ void lbs_stop_card(struct lbs_private *p
 		device_remove_file(&dev->dev, &dev_attr_lbs_mesh);
 
 	/* Flush pending command nodes */
-	del_timer_sync(&priv->command_timer);
+	del_timer(&priv->command_timer);
 	spin_lock_irqsave(&priv->driver_lock, flags);
 	list_for_each_entry(cmdnode, &priv->cmdpendingq, list) {
 		cmdnode->result = -ENOENT;
#1 lore raw
Quoted 8 lines · Holger Schurig

In the normal case, an unsuccessful command would be tried for 10*5 seconds, or 10*10 seconds in the worst case. This patch reduces this to 3*1 seconds, or 3*5 seconds in the worst case.

Also changed the needless del_timer_sync(&priv->command_timer); to del_timer().

Sounds fine to me... Woodhouse?

Dan

Quoted 101 lines

---

The story behind this patch is the following: let the driver load, and let the hardware have a failure, e.g. because of a bad power supply to the device. Now try "pccardctl eject". It looks like this is now hanging, nothing happens. Only when all the retries are done does the "pccardctrl eject" do what it is supposed to do, that might have been in 50 or 100 seconds.

As I didn't have seen a case where it makes sense to retry a faulty command 10 times I changed this number.

Also, I've never seen a command (except scan/assoc) to take more than a few milliseconds. So giving them 5 seconds seemed excessive.

Index: wireless-testing/drivers/net/wireless/libertas/cmd.c
===================================================================

drivers/net/wireless/libertas/cmd.c +3 -3
@@ -1144,7 +1144,7 @@ static void lbs_submit_command(struct lb
 	struct cmd_header *cmd;
 	uint16_t cmdsize;
 	uint16_t command;
-	int timeo = 5 * HZ;
+	int timeo = HZ;
 	int ret;
 
 	lbs_deb_enter(LBS_DEB_HOST);
@@ -1162,7 +1162,7 @@ static void lbs_submit_command(struct lb
 	/* These commands take longer */
 	if (command == CMD_802_11_SCAN || command == CMD_802_11_ASSOCIATE ||
 	    command == CMD_802_11_AUTHENTICATE)
-		timeo = 10 * HZ;
+		timeo = 5 * HZ;
 
 	lbs_deb_cmd("DNLD_CMD: command 0x%04x, seq %d, size %d\n",
 		     command, le16_to_cpu(cmd->seqnum), cmdsize);
@@ -1174,7 +1174,7 @@ static void lbs_submit_command(struct lb
 		lbs_pr_info("DNLD_CMD: hw_host_to_card failed: %d\n", ret);
 		/* Let the timer kick in and retry, and potentially reset
 		   the whole thing if the condition persists */
-		timeo = HZ;
+		timeo = HZ/2;
 	}
 
 	/* Setup the timer after transmit command */

Index: wireless-testing/drivers/net/wireless/libertas/main.c
===================================================================

drivers/net/wireless/libertas/main.c +12 -10
@@ -749,15 +749,18 @@ static int lbs_thread(void *data)
 		if (priv->cmd_timed_out && priv->cur_cmd) {
 			struct cmd_ctrl_node *cmdnode = priv->cur_cmd;
 
-			if (++priv->nr_retries > 10) {
-				lbs_pr_info("Excessive timeouts submitting command %x\n",
-					    le16_to_cpu(cmdnode->cmdbuf->command));
+			if (++priv->nr_retries > 3) {
+				lbs_pr_info("excessive timeouts submitting "
+					"command 0x%04x\n",
+					le16_to_cpu(cmdnode->cmdbuf->command));
 				lbs_complete_command(priv, cmdnode, -ETIMEDOUT);
 				priv->nr_retries = 0;
 			} else {
 				priv->cur_cmd = NULL;
-				lbs_pr_info("requeueing command %x due to timeout (#%d)\n",
-					    le16_to_cpu(cmdnode->cmdbuf->command), priv->nr_retries);
+				lbs_pr_info("requeueing command 0x%04x due "
+					"to timeout (#%d)\n",
+					le16_to_cpu(cmdnode->cmdbuf->command),
+					priv->nr_retries);
 
 				/* Stick it back at the _top_ of the pending queue
 				   for immediate resubmission */
@@ -958,12 +961,11 @@ static void command_timer_fn(unsigned lo
 	lbs_deb_enter(LBS_DEB_CMD);
 	spin_lock_irqsave(&priv->driver_lock, flags);
 
-	if (!priv->cur_cmd) {
-		lbs_pr_info("Command timer expired; no pending command\n");
+	if (!priv->cur_cmd)
 		goto out;
-	}
 
-	lbs_pr_info("Command %x timed out\n", le16_to_cpu(priv->cur_cmd->cmdbuf->command));
+	lbs_pr_info("command 0x%04x timed out\n",
+		le16_to_cpu(priv->cur_cmd->cmdbuf->command));
 
 	priv->cmd_timed_out = 1;
 	wake_up_interruptible(&priv->waitq);
@@ -1278,7 +1280,7 @@ void lbs_stop_card(struct lbs_private *p
 		device_remove_file(&dev->dev, &dev_attr_lbs_mesh);
 
 	/* Flush pending command nodes */
-	del_timer_sync(&priv->command_timer);
+	del_timer(&priv->command_timer);
 	spin_lock_irqsave(&priv->driver_lock, flags);
 	list_for_each_entry(cmdnode, &priv->cmdpendingq, list) {
 		cmdnode->result = -ENOENT;
#2 lore raw
Quoted 11 lines · Dan Williams
Quoted 8 lines · Holger Schurig

In the normal case, an unsuccessful command would be tried for 10*5 seconds, or 10*10 seconds in the worst case. This patch reduces this to 3*1 seconds, or 3*5 seconds in the worst case.

Also changed the needless del_timer_sync(&priv->command_timer); to del_timer().

Sounds fine to me... Woodhouse?

I think I've seen it take a few of the 5-second timeouts. The firmware on this thing really is a steaming pile of shit. I wouldn't like to decrease the maximum so drastically.

Also, that del_timer_sync() happens in the module unload path, doesn't it? I'd want to see why you think it isn't needed.

Signature · 1 line
dwmw2
#3 lore raw
Quoted 1 line

I think I've seen it take a few of the 5-second timeouts.

Do you you know for which commands this has been?

Quoted 1 line

I wouldn't like to decrease the maximum so drastically.

How far would you go?

Or: maybe there is another way to make "pccardctl eject" work reasonably fast in the case of a retrying command. Any idea?

BTW, it's easy to force this kind of error, in cmdresp.c:

if (resp->seqnum==5) {
lbs_deb_cmd("forced FAIL\n"); spin_unlock_irqrestore(&priv->driver_lock, flags); ret = -1;
goto out;
}

Quoted 2 lines

Also, that del_timer_sync() happens in the module unload path, doesn't it? I'd want to see why you think it isn't needed.

I thought that del_timer_sync() waits until the timer has been executed? If yes, then that's the wrong thing here, I want to unload quickly. Makes no sense to wait till the command_timer finally fires and re-sends the failed command.

#4 lore raw
Quoted 5 lines · Holger Schurig
Quoted 1 line

I think I've seen it take a few of the 5-second timeouts.

Do you you know for which commands this has been?

Quoted 7 lines
Quoted 1 line

I wouldn't like to decrease the maximum so drastically.

How far would you go?

Or: maybe there is another way to make "pccardctl eject" work reasonably fast in the case of a retrying command. Any idea?

Does it set priv->surpriseremoved? We should honour that.

Quoted 7 lines
Quoted 2 lines

Also, that del_timer_sync() happens in the module unload path, doesn't it? I'd want to see why you think it isn't needed.

I thought that del_timer_sync() waits until the timer has been executed? If yes, then that's the wrong thing here, I want to unload quickly. Makes no sense to wait till the command_timer finally fires and re-sends the failed command.

No, it just waits until the timer has been properly removed. If the timer was set for some time in the future, it doesn't wait for it to happen naturally.

Signature · 1 line
dwmw2
#5 lore raw
Quoted 5 lines
Quoted 3 lines

Or: maybe there is another way to make "pccardctl eject" work reasonably fast in the case of a retrying command. Any idea?

Does it set priv->surpriseremoved? We should honour that.

Unfortunately not. Manually ejecting the card sets this (via ISR). Then the driver very quickly disables the card, retrying commands or not. It's just "pccardctl eject" that can, under circumstances, take ages. Unfortunately this circumstances happen in my (embedded) hardware setup, e.g. when suspending/resuming.

At resume time, I simply do "pccardctl eject" / "pccardctl insert" to get the driver correctly initialized again and associating. Because doing "real" suspend/resume is currently impossible for me with my knowledge of the firmware.

Quoted 3 lines

No, it just waits until the timer has been properly removed. If the timer was set for some time in the future, it doesn't wait for it to happen naturally.

Oh, this is what I thought would happen. In this case, it {c/sh}ould stay at del_timer_sync().