Skip to contentSkip to conversation

Regression in 2fdac010 drivers/net/ethernet/via/via-velocity.c: update napi implementation

BUG Ruslan N. Marchenko 3 participants 6 messages Sep 30, 2013 – Oct 1, 2013

lore
Share & tools
lore
#0 lore raw

Hi Julia et al.,

With this commit VIA NANO board get CPU lockup in random places on any network activity. The error is random, most of the times dead locked CPU on boot.

After loading via-velocity module it hangs for a while and then spits kernel mesages about hanged tasks from varios places - like ata, rcu_preempt or network stack. Also, if you boot the box with network disconnected - it works, until you plug the network cable.

On some kernel builds (like ubuntu stock) behaviour varies, sometimes the box itself works, but network stack dies after multiple ICMPv6 skb errors, so box ends up with network disconnected. Reloading network susbsys helps to recover for some time (depending on how fast eth led is blinking).

Reverting the patch allows me to boot without problems even from latest Linus trunk and send this mail.

Regards,
Ruslan

#1 lore raw
Quoted 9 lines · Ruslan N. Marchenko

Hi Julia et al.,

With this commit VIA NANO board get CPU lockup in random places on any network activity. The error is random, most of the times dead locked CPU on boot.

After loading via-velocity module it hangs for a while and then spits kernel mesages about hanged tasks from varios places - like ata, rcu_preempt or network stack. Also, if you boot the box with network disconnected - it works, until you plug the network cable.

Quoted 5 lines

On some kernel builds (like ubuntu stock) behaviour varies, sometimes the box itself works, but network stack dies after multiple ICMPv6 skb errors, so box ends up with network disconnected. Reloading network susbsys helps to recover for some time (depending on how fast eth led is blinking).

Reverting the patch allows me to boot without problems even from latest Linus trunk and send this mail.

There has already been a discussion about this, and a patch has already been proposed. It has to do with lock managament. I will look for the email.

julia

#3 lore raw
In reply to Ruslan N. Marchenko
Quoted 9 lines · Ruslan N. Marchenko

Hi Julia et al.,

With this commit VIA NANO board get CPU lockup in random places on any network activity. The error is random, most of the times dead locked CPU on boot.

After loading via-velocity module it hangs for a while and then spits kernel mesages about hanged tasks from varios places - like ata, rcu_preempt or network stack. Also, if you boot the box with network disconnected - it works, until you plug the network cable.

Quoted 5 lines

On some kernel builds (like ubuntu stock) behaviour varies, sometimes the box itself works, but network stack dies after multiple ICMPv6 skb errors, so box ends up with network disconnected. Reloading network susbsys helps to recover for some time (depending on how fast eth led is blinking).

Reverting the patch allows me to boot without problems even from latest Linus trunk and send this mail.

You can find the discussion and improved patch in this thread:

http://www.spinics.net/lists/netdev/msg251287.html
Bug - regression - Via velocity interface coming up freezes kernel
From: Dirk Kraft <dirk.kraft@xxxxxxxxx>
Date: Sun, 22 Sep 2013 19:28:52 +0200

julia

Quoted 4 lines

Regards,
Ruslan

#4 lore raw
Quoted 17 lines · Julia Lawall
Quoted 9 lines · Ruslan N. Marchenko

Hi Julia et al.,

With this commit VIA NANO board get CPU lockup in random places on any network activity. The error is random, most of the times dead locked CPU on boot.

After loading via-velocity module it hangs for a while and then spits kernel mesages about hanged tasks from varios places - like ata, rcu_preempt or network stack. Also, if you boot the box with network disconnected - it works, until you plug the network cable.

Quoted 5 lines

On some kernel builds (like ubuntu stock) behaviour varies, sometimes the box itself works, but network stack dies after multiple ICMPv6 skb errors, so box ends up with network disconnected. Reloading network susbsys helps to recover for some time (depending on how fast eth led is blinking).

Reverting the patch allows me to boot without problems even from latest Linus trunk and send this mail.

Quoted 8 lines

You can find the discussion and improved patch in this thread:

http://www.spinics.net/lists/netdev/msg251287.html
Bug - regression - Via velocity interface coming up freezes kernel
From: Dirk Kraft <dirk.kraft@xxxxxxxxx>
Date: Sun, 22 Sep 2013 19:28:52 +0200

julia

Oops, sorry, somehow I missed it even though 've been searching for via-velocity regression. Will give a try to new proposed patch as well.

Thanks,
Ruslan

#5 lore raw
Quoted 16 lines · Ruslan N. Marchenko
Quoted 15 lines · Julia Lawall
Quoted 5 lines · Ruslan N. Marchenko

Hi Julia et al.,

With this commit VIA NANO board get CPU lockup in random places on any network activity. .
Reverting the patch allows me to boot without problems even from latest Linus trunk and send this mail.

You can find the discussion and improved patch in this thread:

http://www.spinics.net/lists/netdev/msg251287.html
Bug - regression - Via velocity interface coming up freezes kernel
From: Dirk Kraft <dirk.kraft@xxxxxxxxx>
Date: Sun, 22 Sep 2013 19:28:52 +0200

Quoted 5 lines
Quoted 1 line

julia

Oops, sorry, somehow I missed it even though 've been searching for via-velocity regression. Will give a try to new proposed patch as well.

and short confirmation adding patch from Francois Romieu works on current -rc3+

Regards,
Ruslan

#2 lore raw
In reply to Julia Lawall

Julia Lawall <[email protected]> :
[...]

Quoted 3 lines

There has already been a discussion about this, and a patch has already been proposed. It has to do with lock managament. I will look for the email.

The underlying problem has to do with disabled irq. netif_receive_skb assumes irq to be enabled. Current via-velocity poll() method should narrow its (spinlocked) irq disabled section.

What I've done should not require much analysis (aka "what could race with the rx bh processing in a napi driver ?") and avoids a more intrusive lockless napi design.

Signature · 1 line
Ueimor