intel-iommu: Fix tiny theoretical race in write-buffer flush.
authorDavid Woodhouse <David.Woodhouse@intel.com>
Sun, 10 May 2009 19:18:18 +0000 (20:18 +0100)
committerDavid Woodhouse <David.Woodhouse@intel.com>
Sun, 10 May 2009 19:18:18 +0000 (20:18 +0100)
commit462b60f6ccc685f7e8aa04ff430e6b4ffedf629f
tree5f034b5c4993311147f78285a4ffab37def99ed6
parent1f0ef2aa18802a8ce7eb5a5164aaaf4d59073801
intel-iommu: Fix tiny theoretical race in write-buffer flush.

In iommu_flush_write_buffer() we read iommu->gcmd before taking the
register_lock, and then we mask in the WBF bit and write it to the
register.

There is a tiny chance that something else could have _changed_
iommu->gcmd before we take the lock, but after we read it. So we could
be undoing that change.

Never actually going to have happened in practice, since nothing else
changes that register at runtime -- aside from the write-buffer flush
it's only ever touched at startup for enabling translation, etc.

But worth fixing anyway.

Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
drivers/pci/intel-iommu.c