Re: [Xen-devel] [PATCH 03 of 38] swiotlb: allow architectures tooverride swiotlb pool allocation

[Keir Fraser <keir.fraser@eu.citrix.com>]

On 17/11/08 08:04, "Chris Lalancette" <clalance@redhat.com> wrote:

>> Could you be more specific?  The swiotlb allocation should be machine
>> contiguous and so there's no stradding required, but I think I'm missing
>> your point.
> 
> In general, I think you are right; swiotlb should be machine contiguous, so it
> works in the normal case.  The range_straddles_page_boundary function takes
> care
> of a corner case, where you can run into swiotlb exhaustion when you really
> shouldn't.  As I understand it, it comes about because it is possible to get a
> swiotlb request with two pages that just happen to be machine contiguous, but
> were *not* allocated through xen_create_contiguous_region (and hence weren't
> marked in the contiguous_bitmap as such).  In this case, you split the request
> into two separate requests, and this can more easily lead to exhaustion.
> range_straddles_page_boundary works around this by checking whether any two
> pages coming through the swiotlb layer are machine contiguous, and if they
> are,
> not splitting the request.

A more specific problem solved by range_straddle_page_boundary() in our
2.6.18 kernel was that the block layer would do bio merging because it
checked that pages really were contiguous, and then swiotlb (without
r_s_p_b) would decide that the pages weren't contiguous (because the
contiguity was random luck rather than explicitly requested) and hence do
bounce buffering. Result was that sufficiently aggressive I/O would exhaust
swiotlb resources and crash the kernel.

In the 2.6.18 port we actually got rid of contiguous_bitmap[] entirely.

 -- Keir