I've recently tracked down a problem where the IOMMU fills up.  The
machine is a 2 cpu x86_64 (Opteron) with 6G of RAM and is doing lots of
network activity on a 32-bit bus mastering PCI card at the time.  The
following message gets written to the console and the system ends up in
a partly-functional state.

    PCI-DMA: Out of IOMMU space for 53248 bytes at device 0000:03:05.0
    end_request: I/O error, dev sda, sector 84087213

Since sda holds the root filesystem and swap, that explains the loss of
functionality.

The problem is caused by a race condition in free_iommu in pci-gart.c.
It looks like it has been optimized for the case of single page
allocations to avoid taking iommu_bitmap_lock, however this isn't valid
on an SMP machine since the atomic instruction in clear_bit can be
inserted into the middle of one of the non-atomic (read, modify, write)
instructions in __clear_bit_string or set_bit_string.   The effect of
this is that the value written by the non-atomic instruction overwrites
the value written by the atomic instruction and so that IOMMU entry
doesn't get freed.  After running for some time with heavy I/O, the
IOMMU gets full and the system grinds to a halt.

I've attached a patch against 2.6.16.9 to remove the unsafe optimization.

Cheers, Neil.

The IOMMU boot messages were:
    Checking aperture...
    CPU 0: aperture @ f2000000 size 32 MB
    Aperture from northbridge cpu 0 too small (32 MB)
    No AGP bridge found
    Your BIOS doesn't leave a aperture memory hole
    Please enable the IOMMU option in the BIOS setup
    This costs you 64 MB of RAM
    Mapping aperture over 65536 KB of RAM @ 4000000
...
    PCI-DMA: Disabling AGP.
    PCI-DMA: aperture base @ 4000000 size 65536 KB
    PCI-DMA: Reserving 64MB of IOMMU area in the AGP aperture