Messages in this thread | | | Date | Sat, 28 Jan 2006 09:40:22 +1300 | Subject | Re: iommu_alloc failure and panic | From | Olof Johansson <> |
| |
Hi,
On Fri, Jan 27, 2006 at 08:57:40AM -0800, Mark Haverkamp wrote: > While running a disk test (bonnie++) I have been seeing iommu_alloc > failed messages in the syslog leading to a panic. The machine that I
Hmm. The IOMMU code tries to be clever and not write on a cacheline that's already in use, to avoid invalidating cached entries on the I/O bridges.
Since a cache line is 128 bytes (16 entries), this means it advances that much on each allocation, and tries to allocate a new entry in the next block of 16. Essentially it is fragmenting the allocation space on purpose.
So far I haven't seen anyone else have problems with this. I'm suspecting that the SCSI probe code might map something per disk (or similar), such that there's alot of small allocations being done, each using up part of a line. once the end of the allocation space is reached, the allocator wraps to the beginning and starts walking again.
Since this greatly reduces the chance of allocating anything 16 pages or more, a part of the table (25%) is set aside and used for the large allocations. If an allocation in one section of the table (large/small) doesn't succeed, the other half is also searched.
Each allocation in your report is 10 entries, so that means that at least 7 must be taken on each line, since the allocation won't go to the large area by default.
Can you try this, just to confirm that this is what we're seeing?
1. In iommu_setparms_lpar() in arch/powerpc/platforms/pseries/iommu.c, can you try changing it_blocksize from 16 to 1?
Thanks,
Olof
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |