Re: igb_poll - device driver failed to check map error

[Christoph Paasch <christoph.paasch@uclouvain.be>]

On Thursday 14 March 2013 19:18:18 Alexander Duyck wrote:
> On 03/12/2013 02:31 AM, Christoph Paasch wrote:
> > Hello,
> > 
> > I'm seeing a warning while booting my machine when DMA_API_DEBUG is set:
> > 
> > [   36.402824] ------------[ cut here ]------------
> > [   36.458070] WARNING: at
> > /home/cpaasch/builder/net-next/lib/dma-debug.c:934
> > check_unmap+0x648/0x702()
> > [   36.567377] Hardware name: ProLiant DL165 G7
> > [   36.618452] igb 0000:04:00.0: DMA-API: device driver failed to check
> > map
> > error[device address=0x0000000233d9b232] [size=154 bytes] [mapped as
> > single] [   36.776640] Modules linked in:
> > [   36.815446] Pid: 0, comm: swapper/7 Not tainted 3.9.0-rc1-mptcp+ #101
> > [   36.892515] Call Trace:
> > [   36.921745]  <IRQ>  [<ffffffff8102ad7f>] warn_slowpath_common+0x80/0x9a
> > [   37.001023]  [<ffffffff8102ae2d>] warn_slowpath_fmt+0x41/0x43
> > [   37.069771]  [<ffffffff811db17f>] check_unmap+0x648/0x702
> > [   37.134363]  [<ffffffff811db3e9>] debug_dma_unmap_page+0x50/0x52
> > [   37.206234]  [<ffffffff8136676a>] igb_poll+0x144/0xf7c
> > [   37.267706]  [<ffffffff8104dd19>] ? sched_clock_cpu+0x46/0xd1
> > [   37.336456]  [<ffffffff814458ce>] net_rx_action+0xa7/0x1d0
> > [   37.402085]  [<ffffffff81030b65>] __do_softirq+0xb4/0x16f
> > [   37.466673]  [<ffffffff81030c90>] irq_exit+0x40/0x87
> > [   37.526067]  [<ffffffff81002db1>] do_IRQ+0x98/0xaf
> > [   37.583378]  [<ffffffff815210aa>] common_interrupt+0x6a/0x6a
> > [   37.651086]  <EOI>  [<ffffffff8105d4be>] ?
> > __tick_nohz_idle_enter+0x116/0x31f
> > [   37.736595]  [<ffffffff81008a04>] ? default_idle+0x24/0x39
> > [   37.802224]  [<ffffffff81008c62>] cpu_idle+0x68/0xa4
> > [   37.861616]  [<ffffffff81519f78>] start_secondary+0x1a9/0x1ad
> > [   37.930364] ---[ end trace 01b5bb0fd75a464c ]---
> > 
> > 
> > It happens shortly after mounting the NFS-root filesystem.
> > 
> > I tried to understand what is going on, but I am now at my wit's end.
> > 
> > By adding some print-statements, here is what I found out (not sure if
> > this is anyhow helpful):
> > 
> > The difference between tx_buffer->time_stamp and the current 'jiffies' is
> > up to 2000 jiffies (HZ==1000) at the first time the above warning happens
> > (this seems too much for me). From then on, I see my print 3-4 times
> > appear but without such a big difference between the timestamps
> > (difference around 1 and 2 jiffies).
> > 
> > Some other stuff, I printed:
> > tx_buffer->skb: ffff880235054c80
> > tx_buffer->bytecount: 154
> > tx_buffer->gso_segs: 1
> > tx_buffer->protocol: 8
> > tx_buffer->tx_flags 0x20
> > 
> > 
> > One last thing:
> > Am I right that after each call to dma_map_single/page a call to
> > dma_mapping_error is needed? If that's the case, I have some patches that
> > add this statement at missing places in the e1000, e1000e and ixgb
> > driver. But these patches do not fix my above problem.
> > 
> > 
> > Thanks for your help,
> > Christoph
> 
> Christoph,
> 
> One thing that might be useful would be to reproduce this with a
> standard 3.9-rc kernel instead of one using the multipath TCP patches.
> This will help us to verify that the issue is reproducible with a stock
> kernel and is not related to any ongoing work you may have only in your
> tree.

Hello,

this is on a clean net-next kernel without any MPTCP-code.

I bisected it down to  787314c35fbb (Merge tag 'iommu-updates-v3.8' of 
git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu), which simply 
introduces the debug_dma_mapping_error-checks.

Am I right with the missing calls to dma_mapping_error in e1000, e1000e and 
ixgb?

Cheers,
Christoph



-- 
IP Networking Lab --- http://inl.info.ucl.ac.be
MultiPath TCP in the Linux Kernel --- http://multipath-tcp.org
UCLouvain
--