[ath9k-devel] Running out of SWIOTLB space

[Juergen Kreileder <jk@blackdown.de>]

Luis R. Rodriguez wrote:
> On Mon, Aug 04, 2008 at 03:09:31PM -0700, Juergen Kreileder wrote:
>> Robert Hancock wrote:
>>> Tejun Heo wrote:
>>>> Juergen Kreileder wrote:
>>>>> Robert Hancock wrote:
>>>>>> Juergen Kreileder wrote:
>>>>>>> Hi,
>>>>>>>
>>>>>>> I get lots of errors like this with 2.6.27-rc1 on my
>>>>>>> Macbook Pro (3rd generation), 2.6.26 seems to work fine:
>>>>>>>
>>>>>>> [  907.524509] DMA: Out of SW-IOMMU space for 45056 bytes at device 0000:00:1f.2
>>>>>>> [  907.524783] ata3.00: exception Emask 0x0 SAct 0x0 SErr 0x0 action 0x0
>>>>>>> [  907.524796] ata3.00: cmd 35/00:50:08:91:5c/00:01:0f:00:00/e0 tag 0 dma 172032 out
>>>>>>> [  907.524798]          res 50/00:00:ff:90:5c/00:00:0f:00:00/ef Emask 0x40 (internal error)
>>>>>>> [  907.524805] ata3.00: status: { DRDY }
>>>>>>> [  907.646590] ata3.00: configured for UDMA/133
>>>>>>> [  907.646624] ata3: EH complete
>>>>>>> Any suggestions?
>>>>>> You can try increasing the swiotlb size by booting with swiotlb=65536
>>>>>> (for 128MB) or swiotlb=131072 (for 256MB), and see if that makes the
>>>>>> problem go away.. It's unclear why you'd be running out of SWIOTLB space
>>>>>> now, though, if 2.6.26 worked fine..
>>>>> Still happening, although 256M make it a bit harder to trigger.
>>>> Hmm... sounds like mapping is leaking.  Eh... iommu leak debug seems to
>>>> support only AMD GART IOMMU.  How long does it take to reproduce the
>>>> bug?  Once it happens, it never recovers, right?  Does the problem
>>>> happen with minimal configuration w/ only libata enabled (no network, no
>>>> sound, no usb...)?
>>> Yes, that would be a good test.. Easiest way might be booting with
>>> init=/bin/sh and running some processes that do lots of disk access..
>> After a bit more testing I'm pretty sure the problem is caused by an
>> external driver, namely the new ath9k driver.
>>
>> Sorry for the noise on the wrong lists.
> 
> Juergen,
> 
> can you elaborate a bit on the details of when this happens? I may be
> able to get a Power Book to try to reproduce. Are you testing ath9k
> based on ath9k.git or my-wireless-testing.git? What was the last ath9k
> patch you had?

I'm using a 3rd gen Macbook Pro with 2.6.27-rc1 x86-64.  The ath9k driver
was pulled from ath9k.git on Saturday IIRC.

$ dmesg | grep -i -e "ath9k\|atheros"
[   13.331818] ath9k: 0.1
[   13.331861] ath9k 0000:0b:00.0: PCI INT A -> GSI 16 (level, low) -> IRQ 16
[   13.331882] ath9k 0000:0b:00.0: setting latency timer to 64
[   13.462713] phy0: Selected rate control algorithm 'ath9k_rate_control'
[   13.564127] phy0: Atheros 5416: mem=0xffffc200000c0000, irq=16

Apparently it doesn't require much ath9k activity to trigger the problem.
I only had the module loaded but not connected to any network, the disk
driver freaks out pretty soon then.  Without the ath9k module I'm unable
to reproduce any SW-IOMMU problems.
 
> Please note ath9k is now part of wireless-testing.git.

I'll give that a shot tomorrow.

I did a short test with the current version of ath9k.git this evening.
I got an oops from sched-powersave but the driver actually tried to
connect to an network.  It timed out after entering the passphrase though.
Then the machine ran out of SW-IOMMU space again (reported against the
ath9k device this time).


	Juergen

-- 
Juergen Kreileder, Blackdown Java-Linux Team
http://blog.blackdown.de/