Skip to contentSkip to conversation

looking for help interpreting softlockup/stack trace

DISCUSSION Chris Friesen 6 participants 11 messages Aug 5, 2008 – Aug 7, 2008 MIPS

lore
Share & tools
lore

Chris Friesen

author
#0 lore raw

I've run into an interesting issue with an Octeon-based board, where it just seems to hang. I suspect we're hitting some kind of locking bug, and I'm trying to track it down. If it matters, the kernel is quite old (heavily patched 2.6.14) and I've got no chance of upgrading it. (The usual embedded scenario.)

I've added some scheduler instrumentation, as well as adding a stack dump to the output of the softlockup code.

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

Any insights are greatly appreciated...

Chris

cpu9: jiffies: 4295366417, hrtime: 552015654907, 500 ms without calling schedule() since scheduler requested
cpu1: jiffies: 4295366418, hrtime: 552016450325, 500 ms without calling schedule() since scheduler requested
cpu3: jiffies: 4295366442, hrtime: 552034584838, 500 ms without calling schedule() since scheduler requested
cpu9: jiffies: 4295367258, hrtime: 552646827014, 841 ms between scheduler_tick() calls

BUG: soft lockup detected on CPU#0!
Cpu 0
$ 0   : 0000000000000000 0000000000000000 ffffffff80000000 ffffffff00000000
$ 4   : a8000000db3ce808 0000000000000084 a8000000d34cb1ea a8000000d34cb1ea
$ 8   : 0000000045000084 0000000000000002 0000000000000001 ffffffff8165e1e0
$12   : 0000000000000000 ffffffff812dedac 0000500000008000 0000100000000000
$16   : a8000000d3a40b68 a8000000db3ce580 a8000000d3a40b00 ffffffff8151e448
$20   : a8000000db3ce808 ffffffff81593808 a8000000d3a40b00 0000000000000000
$24   : 0000000000000030 c000000001ae1820
$28   : ffffffff81590000 ffffffff81593680 ffffffff81593810 c000000001b4ab9c
Hi    : 0000000000000000
Lo    : 0000000000000000
epc   : ffffffff8151e44c _read_lock+0x4/0x20     Tainted: P
ra    : c000000001b4ab9c $LBB378+0x0/0x18 [vnb]
Status: 1000ffe3    KX SX UX KERNEL EXL IE
Cause : 40808000
PrId  : 000d0301
Call Trace:
 [<ffffffff811a251c>] softlockup_tick+0x12c/0x158
 [<ffffffff811126dc>] timer_interrupt+0xd4/0x4c8
 [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
 [<ffffffff81101d50>] octeon_main_timer_interrupt+0x78/0x90
 [<ffffffff811a2c24>] handle_IRQ_event+0x45c/0x1528
 [<ffffffff811a3dc4>] __do_IRQ+0xd4/0x230
 [<ffffffff8151e448>] _read_lock+0x0/0x20
 [<ffffffff8110c4a8>] do_IRQ+0x440/0x1520
 [<ffffffff8139263c>] neigh_lookup+0xb4/0x128
 [<ffffffff8151e448>] _read_lock+0x0/0x20
 [<ffffffff8110a430>] ret_from_irq+0x0/0x10
 [<ffffffff8114fbdc>] __wake_up_common+0x6c/0xb8
 [<ffffffff812dedac>] memset_partial+0x48/0x60
 [<ffffffff8151e448>] _read_lock+0x0/0x20
 [<c000000001ae1820>] $Ltext0+0x0/0x4 [vnb]
 [<ffffffff8151e618>] _read_unlock+0x0/0x20
 [<c000000001b4ab9c>] $LBB378+0x0/0x18 [vnb]
 [<ffffffff8151e44c>] _read_lock+0x4/0x20
 [<ffffffff81407fcc>] arp_rcv+0x18c/0x258
 [<ffffffff8138b84c>] netif_dispatch_skb+0x18c/0x2d0
 [<c000000000861208>] $LBE370+0x0/0x10 [cavium_ethernet]
 [<c000000001ae2acc>] $L396+0x0/0x1c [vnb]
 [<c000000001b47ba8>] $LCFI35+0x8c/0x9c [vnb]
 [<c000000001adb36c>] $L1193+0x2c/0x48 [vnb]
 [<c000000001b3dc5c>] $L1047+0x28/0xd0 [vnb]
 [<c000000001b3d6f0>] $LBB344+0x2c/0x34 [vnb]
 [<ffffffff811476a8>] activate_task+0x98/0x108
 [<c000000001adb4cc>] $L1215+0x2c/0x30 [vnb]
 [<c000000001adb7c0>] $L1273+0x0/0x18 [vnb]
 [<c000000001ae18a8>] $L28+0x24/0x4c [vnb]
 [<ffffffff8138bc90>] netif_receive_skb+0x300/0x730
 [<ffffffff8138130c>] __alloc_skb+0x84/0x1a8
 [<c000000000861970>] $L620+0x20/0x28 [cavium_ethernet]
 [<ffffffff811734e8>] process_timeout+0x0/0xb08
 [<ffffffff81111cc0>] do_gettimeofday+0xa8/0x220
 [<ffffffff81169b2c>] tasklet_action+0x684/0x1950
 [<ffffffff81187b9c>] hrtimer_run_queues+0x84/0x1a0
 [<c0000000008613d8>] cvm_oct_tasklet_rx+0x0/0x4 [cavium_ethernet]
 [<ffffffff81171ce4>] run_timer_softirq+0x20c/0xce0
 [<c0000000008613d8>] cvm_oct_tasklet_rx+0x0/0x4 [cavium_ethernet]
 [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
 [<ffffffff81167afc>] __do_softirq+0x56c/0x1818
 [<ffffffff81168e40>] do_softirq+0x98/0xb8
 [<ffffffff8110c874>] do_IRQ+0x80c/0x1520
 [<ffffffff8110dd98>] kernel_thread+0xa0/0xbf8
 [<ffffffff8110a430>] ret_from_irq+0x0/0x10
 [<ffffffff8151bef4>] schedule+0xd4c/0x1a98
 [<ffffffff81109ec0>] r4k_wait+0x0/0x10
 [<ffffffff813859b0>] datagram_poll+0x0/0x140
 [<ffffffff8110d974>] cpu_idle+0x44/0x60
 [<ffffffff81109ec8>] r4k_wait+0x8/0x10
 [<ffffffff81611e70>] start_kernel+0x460/0x4f8
 [<ffffffff81611e68>] start_kernel+0x458/0x4f8
 [<ffffffff81611000>] kernel_entry+0x0/0x4c
#1 lore raw
Quoted 10 lines · Chris Friesem

I've run into an interesting issue with an Octeon-based board, where it
just seems to hang. I suspect we're hitting some kind of locking bug,
and I'm trying to track it down. If it matters, the kernel is quite old
(heavily patched 2.6.14) and I've got no chance of upgrading it. (The
usual embedded scenario.)

Quoted 7 lines

I've added some scheduler instrumentation, as well as adding a stack dump to the output of the softlockup code.

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

Hi Chris

The contents of RA may or may not be useful. RA is used to hold the return address when some branch instructions are executed. The contents of RA may provide some kind of clue about what the
processor was doing previously. Often, RA can give you the parent function, but not always.

Here it looks it does give the parent function; it is apparently inside a kernel module called "vnb", which called _read_lock and got stuck there.

Mind you, these Linux MIPS stack traces are not completely trustworthy; the routine just walks the stack one word at a time and prints out anything that looks like the address of code. So the stack trace may include stale stack data from previous call chains that have already returned. It also might not include the full call chain, because the return address is not always stored on the stack. For instance, here it looks like RA is pointing into a kernel module. But you don't actually see this in the fake stack trace. Nowhere in the stack trace do you see _read_lock being called by c000000001b4ab9c.

I would proceed by inspecting this vnb module and its use of locks.

Quoted 33 lines

Any insights are greatly appreciated...

Chris

cpu9: jiffies: 4295366417, hrtime: 552015654907, 500 ms without calling
schedule() since scheduler requested
cpu1: jiffies: 4295366418, hrtime: 552016450325, 500 ms without calling
schedule() since scheduler requested
cpu3: jiffies: 4295366442, hrtime: 552034584838, 500 ms without calling
schedule() since scheduler requested
cpu9: jiffies: 4295367258, hrtime: 552646827014, 841 ms between scheduler_tick() calls BUG: soft lockup detected on CPU#0! Cpu 0

$ 0   : 0000000000000000 0000000000000000 ffffffff80000000
ffffffff00000000 $ 4   : a8000000db3ce808 0000000000000084
a8000000d34cb1ea a8000000d34cb1ea $ 8   : 0000000045000084
0000000000000002 0000000000000001 ffffffff8165e1e0 $12   :
0000000000000000 ffffffff812dedac 0000500000008000 0000100000000000
$16   : a8000000d3a40b68 a8000000db3ce580 a8000000d3a40b00
ffffffff8151e448 $20   : a8000000db3ce808 ffffffff81593808
a8000000d3a40b00 0000000000000000 $24   : 0000000000000030
c000000001ae1820 $28   : ffffffff81590000 ffffffff81593680
ffffffff81593810 c000000001b4ab9c Hi    : 0000000000000000 
Lo    : 0000000000000000
epc   : ffffffff8151e44c _read_lock+0x4/0x20     Tainted: P
ra    : c000000001b4ab9c $LBB378+0x0/0x18 [vnb]
Status: 1000ffe3    KX SX UX KERNEL EXL IE
Cause : 40808000
PrId  : 000d0301
Quoted 8 lines
Call Trace:
  [<ffffffff811a251c>] softlockup_tick+0x12c/0x158
  [<ffffffff811126dc>] timer_interrupt+0xd4/0x4c8
  [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
  [<ffffffff81101d50>] octeon_main_timer_interrupt+0x78/0x90
  [<ffffffff811a2c24>] handle_IRQ_event+0x45c/0x1528
  [<ffffffff811a3dc4>] __do_IRQ+0xd4/0x230
  [<ffffffff8151e448>] _read_lock+0x0/0x20

This read lock is almost certainly a red herring; it's stuff left over on the stack that hasn't been overwritten by new activation chains.

Quoted 1 line

[<ffffffff8110c4a8>] do_IRQ+0x440/0x1520

This is the timer interrupt going off, which ends up detecting the lockup. So the problem is above here.

Quoted 1 line

[<ffffffff8139263c>] neigh_lookup+0xb4/0x128

This may also be garbage. Basically, the stack trace is not that useful, except that it provides some additional circumstantial evidence implicating the vnb module.

author
#2 lore raw
Quoted 8 lines · Kaz Kylheku
Quoted 4 lines · Chris Friesem

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

Quoted 10 lines

Mind you, these Linux MIPS stack traces are not completely trustworthy; the routine just walks the stack one word at a time and prints out anything that looks like the address of code. So the stack trace may include stale stack data from previous call chains that have already returned. It also might not include the full call chain, because the return address is not always stored on the stack. For instance, here it looks like RA is pointing into a kernel module. But you don't actually see this in the fake stack trace. Nowhere in the stack trace do you see _read_lock being called by c000000001b4ab9c.

Right...but it does look like it calls _read_unlock().

Quoted 1 line

I would proceed by inspecting this vnb module and its use of locks.

Yeah...I'll do that.

Quoted 1 line
Quoted 1 line

scheduler_tick() calls BUG: soft lockup detected on CPU#0!

One thing I've noticed, all the softlockups are on cpu0 even though other cpus are also complaining about scheduling delays. Is there something special about cpu0?

Quoted 10 lines
Quoted 8 lines
Call Trace:
 [<ffffffff811a251c>] softlockup_tick+0x12c/0x158
 [<ffffffff811126dc>] timer_interrupt+0xd4/0x4c8
 [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
 [<ffffffff81101d50>] octeon_main_timer_interrupt+0x78/0x90
 [<ffffffff811a2c24>] handle_IRQ_event+0x45c/0x1528
 [<ffffffff811a3dc4>] __do_IRQ+0xd4/0x230
 [<ffffffff8151e448>] _read_lock+0x0/0x20
Quoted 3 lines

This read lock is almost certainly a red herring; it's stuff left over on the stack that hasn't been overwritten by new activation chains.

Wouldn't it be part of the do_IRQ call below?

Quoted 5 lines
Quoted 1 line

[<ffffffff8110c4a8>] do_IRQ+0x440/0x1520

This is the timer interrupt going off, which ends up detecting the lockup. So the problem is above here.

I assume you mean "above" in terms of chronological order, since in terms of the call trace the problem would be below.

Quoted 7 lines
Quoted 1 line

[<ffffffff8139263c>] neigh_lookup+0xb4/0x128

This may also be garbage. Basically, the stack trace is not that useful, except that it provides some additional circumstantial evidence implicating the vnb module.

It's unfortunate that the call trace isn't accurate.

Thanks for the help,

Chris

#3 lore raw
In reply to Chris Friesen
Quoted 6 lines · Chris Friesen

I've run into an interesting issue with an Octeon-based board, where it just seems to hang. I suspect we're hitting some kind of locking bug, and I'm trying to track it down. If it matters, the kernel is quite old (heavily patched 2.6.14) and I've got no chance of upgrading it. (The usual embedded scenario.)

Quoted 7 lines

I've added some scheduler instrumentation, as well as adding a stack dump to the output of the softlockup code.

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

Quoted 1 line

Any insights are greatly appreciated...

You may also try lockdep; it gives much more detailed information though it's more heavyweight.

Ralf

author
#4 lore raw
Quoted 2 lines · Ralf Baechle

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

What determines which function names get compiled into the object file, and which ones get these numerical symbols?

Quoted 2 lines

You may also try lockdep; it gives much more detailed information though it's more heavyweight.

I would love to be able to use lockdep...however I'm stuck on an old kernel.

Chris

#5 lore raw
In reply to Ralf Baechle
Quoted 11 lines · Ralf Baechle
Quoted 5 lines · Chris Friesen

I've run into an interesting issue with an Octeon-based board, where it just seems to hang. I suspect we're hitting some kind of locking bug, and I'm trying to track it down. If it matters, the kernel is quite old (heavily patched 2.6.14) and I've got no chance of upgrading it. (The usual embedded scenario.)

I've added some scheduler instrumentation, as well as adding a stack dump to the output of the softlockup code.

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

I have thought about eliminating these internal labels when the module's symbols are read. Would this make any sense?

David Daney

#6 lore raw
Quoted 10 lines · David Daney
Quoted 8 lines
Quoted 4 lines

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

I have thought about eliminating these internal labels when the module's symbols are read. Would this make any sense?

I think so. Maybe that could even be done when the module is linked. I don't think there are ever any relocations against these local symbols. Thiemo?

Ralf

#7 lore raw
Quoted 11 lines · Ralf Baechle
Quoted 9 lines · David Daney
Quoted 8 lines
Quoted 4 lines

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

Quoted 5 lines
Quoted 1 line

I have thought about eliminating these internal labels when the module's symbols are read. Would this make any sense?

I think so. Maybe that could even be done when the module is linked. I don't think there are ever any relocations against these local symbols. Thiemo?

AFAIR all such symbols are branch labels or debug info labels, the module loader should never use them. (But I don't know if kgdb uses them.)

Thiemo

#8 lore raw
In reply to Ralf Baechle
Quoted 12 lines · Ralf Baechle
Quoted 5 lines · David Daney
Quoted 4 lines
Quoted 1 line

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

$LBB378 is an internal symbol. The value of RA may not be very informative if it was overwritten by a random subroutine call.

I have thought about eliminating these internal labels when the module's symbols are read. Would this make any sense?

I think so. Maybe that could even be done when the module is linked. I don't think there are ever any relocations against these local symbols. Thiemo?

I take it back. Most of the $LC... symbols (typically string constants) are needed for relocations. Currently I cannot find in my builds any $LB... symbols. I wonder if gcc-4.3 eliminates these.

I think the reason they show up in stack traces is that they are data pointers that have been stored on the stack that are misinterpreted as function return addresses.

I wonder what would happen if we compiled all the code with -funwind-tables and had a small DWARF2 unwinder ala GCC's C++ exception handling mechanism. That would allow exact stack traces with no runtime overhead.

David Daney

#9 lore raw
In reply to Chris Friesen

I assume this is happening under high network load. Some early Octeon ethernet drivers had a problem where they could starve the rest of the system processing incoming packets. The message you are getting is the kernel warning you that userspace hasn't been given any processing time. What is probably happening is the cavium_ethernet driver is spending all its time in a receive tasklet getting packets and then dropping them. This has been fixed in a later cavium SDK, but it looks like you are running an ancient kernel. A newer kernel is available on the cavium support site.

The cavium ethernet driver doesn't use the standard NAPI interface since it doesn't support multicore receive for a single port.

Chad

Quoted 99 lines · Chris Friesen

I've run into an interesting issue with an Octeon-based board, where it just seems to hang. I suspect we're hitting some kind of locking bug, and I'm trying to track it down. If it matters, the kernel is quite old (heavily patched 2.6.14) and I've got no chance of upgrading it. (The usual embedded scenario.)

I've added some scheduler instrumentation, as well as adding a stack dump to the output of the softlockup code.

In the trace below, is "epc" the program counter at the time of the timer interrupt? How does "ra" fit into this, given that the function whose address it contains isn't seen in the stack trace until quite a ways down?

Any insights are greatly appreciated...

Chris

cpu9: jiffies: 4295366417, hrtime: 552015654907, 500 ms without calling schedule() since scheduler requested
cpu1: jiffies: 4295366418, hrtime: 552016450325, 500 ms without calling schedule() since scheduler requested
cpu3: jiffies: 4295366442, hrtime: 552034584838, 500 ms without calling schedule() since scheduler requested
cpu9: jiffies: 4295367258, hrtime: 552646827014, 841 ms between scheduler_tick() calls

BUG: soft lockup detected on CPU#0!
Cpu 0
$ 0   : 0000000000000000 0000000000000000 ffffffff80000000 ffffffff00000000
$ 4   : a8000000db3ce808 0000000000000084 a8000000d34cb1ea a8000000d34cb1ea
$ 8   : 0000000045000084 0000000000000002 0000000000000001 ffffffff8165e1e0
$12   : 0000000000000000 ffffffff812dedac 0000500000008000 0000100000000000
$16   : a8000000d3a40b68 a8000000db3ce580 a8000000d3a40b00 ffffffff8151e448
$20   : a8000000db3ce808 ffffffff81593808 a8000000d3a40b00 0000000000000000
$24   : 0000000000000030 c000000001ae1820
$28   : ffffffff81590000 ffffffff81593680 ffffffff81593810 c000000001b4ab9c
Hi    : 0000000000000000
Lo    : 0000000000000000
epc   : ffffffff8151e44c _read_lock+0x4/0x20     Tainted: P
ra    : c000000001b4ab9c $LBB378+0x0/0x18 [vnb]
Status: 1000ffe3    KX SX UX KERNEL EXL IE
Cause : 40808000
PrId  : 000d0301
Call Trace:
  [<ffffffff811a251c>] softlockup_tick+0x12c/0x158
  [<ffffffff811126dc>] timer_interrupt+0xd4/0x4c8
  [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
  [<ffffffff81101d50>] octeon_main_timer_interrupt+0x78/0x90
  [<ffffffff811a2c24>] handle_IRQ_event+0x45c/0x1528
  [<ffffffff811a3dc4>] __do_IRQ+0xd4/0x230
  [<ffffffff8151e448>] _read_lock+0x0/0x20
  [<ffffffff8110c4a8>] do_IRQ+0x440/0x1520
  [<ffffffff8139263c>] neigh_lookup+0xb4/0x128
  [<ffffffff8151e448>] _read_lock+0x0/0x20
  [<ffffffff8110a430>] ret_from_irq+0x0/0x10
  [<ffffffff8114fbdc>] __wake_up_common+0x6c/0xb8
  [<ffffffff812dedac>] memset_partial+0x48/0x60
  [<ffffffff8151e448>] _read_lock+0x0/0x20
  [<c000000001ae1820>] $Ltext0+0x0/0x4 [vnb]
  [<ffffffff8151e618>] _read_unlock+0x0/0x20
  [<c000000001b4ab9c>] $LBB378+0x0/0x18 [vnb]
  [<ffffffff8151e44c>] _read_lock+0x4/0x20
  [<ffffffff81407fcc>] arp_rcv+0x18c/0x258
  [<ffffffff8138b84c>] netif_dispatch_skb+0x18c/0x2d0
  [<c000000000861208>] $LBE370+0x0/0x10 [cavium_ethernet]
  [<c000000001ae2acc>] $L396+0x0/0x1c [vnb]
  [<c000000001b47ba8>] $LCFI35+0x8c/0x9c [vnb]
  [<c000000001adb36c>] $L1193+0x2c/0x48 [vnb]
  [<c000000001b3dc5c>] $L1047+0x28/0xd0 [vnb]
  [<c000000001b3d6f0>] $LBB344+0x2c/0x34 [vnb]
  [<ffffffff811476a8>] activate_task+0x98/0x108
  [<c000000001adb4cc>] $L1215+0x2c/0x30 [vnb]
  [<c000000001adb7c0>] $L1273+0x0/0x18 [vnb]
  [<c000000001ae18a8>] $L28+0x24/0x4c [vnb]
  [<ffffffff8138bc90>] netif_receive_skb+0x300/0x730
  [<ffffffff8138130c>] __alloc_skb+0x84/0x1a8
  [<c000000000861970>] $L620+0x20/0x28 [cavium_ethernet]
  [<ffffffff811734e8>] process_timeout+0x0/0xb08
  [<ffffffff81111cc0>] do_gettimeofday+0xa8/0x220
  [<ffffffff81169b2c>] tasklet_action+0x684/0x1950
  [<ffffffff81187b9c>] hrtimer_run_queues+0x84/0x1a0
  [<c0000000008613d8>] cvm_oct_tasklet_rx+0x0/0x4 [cavium_ethernet]
  [<ffffffff81171ce4>] run_timer_softirq+0x20c/0xce0
  [<c0000000008613d8>] cvm_oct_tasklet_rx+0x0/0x4 [cavium_ethernet]
  [<ffffffff81112890>] timer_interrupt+0x288/0x4c8
  [<ffffffff81167afc>] __do_softirq+0x56c/0x1818
  [<ffffffff81168e40>] do_softirq+0x98/0xb8
  [<ffffffff8110c874>] do_IRQ+0x80c/0x1520
  [<ffffffff8110dd98>] kernel_thread+0xa0/0xbf8
  [<ffffffff8110a430>] ret_from_irq+0x0/0x10
  [<ffffffff8151bef4>] schedule+0xd4c/0x1a98
  [<ffffffff81109ec0>] r4k_wait+0x0/0x10
  [<ffffffff813859b0>] datagram_poll+0x0/0x140
  [<ffffffff8110d974>] cpu_idle+0x44/0x60
  [<ffffffff81109ec8>] r4k_wait+0x8/0x10
  [<ffffffff81611e70>] start_kernel+0x460/0x4f8
  [<ffffffff81611e68>] start_kernel+0x458/0x4f8
  [<ffffffff81611000>] kernel_entry+0x0/0x4c
Signature · 2 lines
Chad Reese <[email protected]>
Cavium Networks
author
#10 lore raw
Quoted 9 lines · Chad Reese

I assume this is happening under high network load. Some early Octeon ethernet drivers had a problem where they could starve the rest of the system processing incoming packets. The message you are getting is the kernel warning you that userspace hasn't been given any processing time. What is probably happening is the cavium_ethernet driver is spending all its time in a receive tasklet getting packets and then dropping them. This has been fixed in a later cavium SDK, but it looks like you are running an ancient kernel. A newer kernel is available on the cavium support site.

It seems unlikely it's due to cpu load. It's only receiving around 30000 packets/sec total, and before the problem shows up "top" shows some cores at around 25% and others totally idle. The only symptom of the problem is that the system simply stops responding.

You wouldn't happen to know which versions fixed the problem you described above, would you? Upgrading the kernel isn't an option, but it wouldn't be the first time we've had to backport things.

Quoted 2 lines

The cavium ethernet driver doesn't use the standard NAPI interface since it doesn't support multicore receive for a single port.

Ah...I was wondering about that.

Chris