Skip to contentSkip to conversation

netpoll/netconsole minor tweaks

PATCH Chris Wright 5 participants 10 messages Sep 17, 2003 – Sep 19, 2003

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Hi Matt,

Here's a couple small tweaks. The first is to netpoll_setup. The settle time was too short for my e100, and the system would hang. The second is to netconsole so that it registers a console with CON_PRINTBUFFER. This helps debugging early bootup issues where you want to capture data from before netconsole is initialized. Perhaps it should be a param to netconsole?

thanks,
-chris

net/core/netpoll.c renamed from net/core/netpoll.c.wait_fix +1 -1
@@ -526,7 +526,7 @@ int netpoll_setup(struct netpoll *np)
 		rtnl_shunlock();
 
 		/* Give driver a chance to settle */
-		jiff = jiffies + 2*HZ;
+		jiff = jiffies + 4*HZ;
 		while (time_before(jiffies, jiff))
 			;
 	}
drivers/net/netconsole.c renamed from drivers/net/netconsole.c.print_buf +1 -1
@@ -95,7 +95,7 @@
 }
 
 static struct console netconsole = {
-	.flags = CON_ENABLED,
+	.flags = CON_ENABLED | CON_PRINTBUFFER,
 	.write = write_msg
 };
#1 lore raw
Quoted 4 lines · Chris Wright

Hi Matt,

Here's a couple small tweaks. The first is to netpoll_setup. The settle time was too short for my e100, and the system would hang.

This probably ought to be a command line arg, my tg3 apparently doesn't need it at all. One second is what's in the nfs-root stuff, perhaps they need to share an arg, I changed it to two because my tlan was dropping stuff (but not hanging - that may be a driver bug). Ideally, I'd like to find a way to wait until the damn thing is up before sending.

net/ipv4/ipconfig.c:
/* Define the friendly delay before and after opening net devices */

 #define CONF_PRE_OPEN           (HZ/2)  /* Before opening: 1/2 second */
 #define CONF_POST_OPEN          (1*HZ)  /* After opening: 1 second */

Anyway, I'm still struggling with getting stuff working on my Opteron box, care to take a stab at it?

Quoted 5 lines

The second
is to netconsole so that it registers a console with CON_PRINTBUFFER. This helps debugging early bootup issues where you want to capture data from before netconsole is initialized. Perhaps it should be a param to netconsole?

I think this can probably be unconditional. Merged.

Signature · 1 line
Matt Mackall : http://www.selenic.com : of or relating to the moon
#2 lore raw
Quoted 10 lines · * Matt Mackall ([email protected]) wrote:
Quoted 2 lines · Chris Wright

Here's a couple small tweaks. The first is to netpoll_setup. The settle time was too short for my e100, and the system would hang.

This probably ought to be a command line arg, my tg3 apparently doesn't need it at all. One second is what's in the nfs-root stuff, perhaps they need to share an arg, I changed it to two because my tlan was dropping stuff (but not hanging - that may be a driver bug). Ideally, I'd like to find a way to wait until the damn thing is up before sending.

Yeah, I agree, that seems best.

Quoted 7 lines

net/ipv4/ipconfig.c:
/* Define the friendly delay before and after opening net devices */

 #define CONF_PRE_OPEN           (HZ/2)  /* Before opening: 1/2 second */
 #define CONF_POST_OPEN          (1*HZ)  /* After opening: 1 second */

Anyway, I'm still struggling with getting stuff working on my Opteron box, care to take a stab at it?

Sure, I'll poke at it tomorrow, unless you get to it first.

Quoted 7 lines
Quoted 5 lines

The second
is to netconsole so that it registers a console with CON_PRINTBUFFER. This helps debugging early bootup issues where you want to capture data from before netconsole is initialized. Perhaps it should be a param to netconsole?

I think this can probably be unconditional. Merged.

Ok, cool.
thanks,
-chris

Signature · 1 line
Linux Security Modules     http://lsm.immunix.org     http://lsm.bkbits.net
#3 lore raw
In reply to Chris Wright

net/core/netpoll.c int netpoll_setup(struct netpoll *np) · line 530

Quoted 4 lines · Chris Wright
net/core/netpoll.c renamed from net/core/netpoll.c.wait_fix +1 -1
 		/* Give driver a chance to settle */
-		jiff = jiffies + 2*HZ;
+		jiff = jiffies + 4*HZ;
 		while (time_before(jiffies, jiff))

<pedant>

Quoted 1 line

;

should be cpu_relax();
</pedant>

#4 lore raw
Quoted 8 lines · * Alan Cox ([email protected]) wrote:

net/core/netpoll.c int netpoll_setup(struct netpoll *np) · line 530

Quoted 4 lines · Chris Wright
net/core/netpoll.c renamed from net/core/netpoll.c.wait_fix +1 -1
 		/* Give driver a chance to settle */
-		jiff = jiffies + 2*HZ;
+		jiff = jiffies + 4*HZ;
 		while (time_before(jiffies, jiff))

<pedant>
should be cpu_relax();
</pedant>

Hrm, there's many spots that aren't using it. What's the benefit, less power consumption? Is it worth a patch to convert other things over?

thanks,
-chris

Signature · 1 line
Linux Security Modules     http://lsm.immunix.org     http://lsm.bkbits.net
#5 lore raw
Quoted 3 lines

Hrm, there's many spots that aren't using it. What's the benefit, less power consumption?

Less power consumption, and on HT/SMT CPU's it's a "yield" to the other half/halves.

Quoted 3 lines

Is it worth a patch to convert other things over?

yes

#6 lore raw
Quoted 2 lines · * Arjan van de Ven ([email protected]) wrote:

Less power consumption, and on HT/SMT CPU's it's a "yield" to the other half/halves.

I see, thanks.

Quoted 2 lines
Quoted 1 line

Is it worth a patch to convert other things over?

yes

OK, I'll spin one up, thanks,
-chris

Signature · 1 line
Linux Security Modules     http://lsm.immunix.org     http://lsm.bkbits.net
#7 lore raw
In reply to Chris Wright
Quoted 6 lines · Chris Wright
Quoted 3 lines

<pedant>
should be cpu_relax();
</pedant>

Hrm, there's many spots that aren't using it. What's the benefit, less power consumption? Is it worth a patch to convert other things over?

It speeds up hyperthreading CPUs by letting them know that this paticular thread is waiting for stuff (sched_yield for silicon)

#8 lore raw
Quoted 11 lines · Alan Cox
Quoted 6 lines · Chris Wright
Quoted 3 lines

<pedant>
should be cpu_relax();
</pedant>

Hrm, there's many spots that aren't using it. What's the benefit, less power consumption? Is it worth a patch to convert other things over?

It speeds up hyperthreading CPUs by letting them know that this paticular thread is waiting for stuff (sched_yield for silicon)

Does that mean that it would be benneficial to do something like this in for example eepro100.c ??

diff -up linux-2.6.0-test5-orig/drivers/net/eepro100.c linux-2.6.0-test5/drivers/net/eepro100.c

drivers/net/eepro100.c 2003-09-19 01:03:19.000000000 +0200 renamed from drivers/net/eepro100.c 2003-09-08 21:50:09.000000000 +0200 +3 -1
@@ -913,9 +913,11 @@ static void do_slow_command(struct net_d
 
        for (wait = 0; wait <= 100; wait++)
                if (inb(cmd_ioaddr) == 0) return;
-       for (; wait <= 20000; wait++)
+       for (; wait <= 20000; wait++) {
                if (inb(cmd_ioaddr) == 0) return;
                else udelay(1);
+               cpu_relax();
+       }
        printk(KERN_ERR "Command %4.4x was not accepted after %d polls!"
               "  Current status %8.8x.\n",
               cmd, wait, inl(dev->base_addr + SCBStatus));

or maybe even take it to the extreme like

diff -up linux-2.6.0-test5-orig/drivers/net/eepro100.c linux-2.6.0-test5/drivers/net/eepro100.c

drivers/net/eepro100.c 2003-09-19 01:07:50.000000000 +0200 renamed from drivers/net/eepro100.c 2003-09-08 21:50:09.000000000 +0200 +3 -2
@@ -902,20 +902,25 @@ static void do_slow_command(struct net_d
 {
        long cmd_ioaddr = dev->base_addr + SCBCmd;
        int wait = 0;
-       do
+       do {
                if (inb(cmd_ioaddr) == 0) break;
-       while(++wait <= 200);
+               cpu_relax();
+       } while(++wait <= 200);
        if (wait > 100)
                printk(KERN_ERR "Command %4.4x never accepted (%d

polls)!\n",

                       inb(cmd_ioaddr), wait);

        outb(cmd, cmd_ioaddr);

- for (wait = 0; wait <= 100; wait++)

+       for (wait = 0; wait <= 100; wait++) {
                if (inb(cmd_ioaddr) == 0) return;
-       for (; wait <= 20000; wait++)
+               cpu_relax();
+       }
+       for (; wait <= 20000; wait++) {
                if (inb(cmd_ioaddr) == 0) return;
                else udelay(1);
+               cpu_relax();
+       }
        printk(KERN_ERR "Command %4.4x was not accepted after %d polls!"
               "  Current status %8.8x.\n",
               cmd, wait, inl(dev->base_addr + SCBStatus));

How short/long waits are we talking about before it is benneficial to be calling cpu_relax() ?

Kind regards,

Jesper Juhl <[email protected]>

#9 lore raw

drivers/net/eepro100.c 2003-09-19 01:03:19.000000000 +0200 static void do_slow_command(struct net_d · line 920

Quoted 8 lines · Jesper Juhl

00; wait++)

                if (inb(cmd_ioaddr) == 0) return;
drivers/net/eepro100.c 2003-09-19 01:03:19.000000000 +0200 renamed from drivers/net/eepro100.c 2003-09-08 21:50:09.000000000 +0200 +3 -1
-       for (; wait <= 20000; wait++)
+       for (; wait <= 20000; wait++) {
                if (inb(cmd_ioaddr) == 0) return;
                else udelay(1);
+               cpu_relax();
+       }

udelay() should have cpu_relax() inside it already basically..