Re: [PATCH v2 1/3] swiotlb: Fix allocation alignment requirement when searching slots

[Robin Murphy]

On 01/02/2024 12:46 pm, Will Deacon wrote:
> Hey Robin,
>
> Cheers for having a look.
>
> On Wed, Jan 31, 2024 at 03:54:03PM +0000, Robin Murphy wrote:
> > On 31/01/2024 12:25 pm, Will Deacon wrote:
> > > Commit bbb73a103fbb ("swiotlb: fix a braino in the alignment check fix"),
> > > which was a fix for commit 0eee5ae10256 ("swiotlb: fix slot alignment
> > > checks"), causes a functional regression with vsock in a virtual machine
> > > using bouncing via a restricted DMA SWIOTLB pool.
> > >
> > > When virtio allocates the virtqueues for the vsock device using
> > > dma_alloc_coherent(), the SWIOTLB search fails to take into account the
> > > 8KiB buffer size and returns page-unaligned allocations if 'area->index'
> > > was left unaligned by a previous allocation from the buffer:
> >
> > Hmm, but isn't this fundamentally swiotlb_alloc()'s fault for assuming it's
> > going to get a page-aligned address back despite asking for 0 alignment in
> > the first place? I'm not sure SWIOTLB has ever promised implicit
> > size-alignment, so it feels somewhat misplaced to be messing with the
> > algorithm before fixing the obvious issue in the caller :/
>
> It's hard to tell which guarantees are intentional here given that this
> interface is all internal to swiotlb.c, but the 'alloc_align_mask'
> parameter didn't even exist prior to e81e99bacc9f ("swiotlb: Support
> aligned swiotlb buffers") and practically the implementation has ensured
> page-aligned allocations for buffers >= PAGE_SIZE prior to 0eee5ae10256
> ("swiotlb: fix slot alignment checks") by virtue of aligning the search
> index to the stride.
>
> In any case, this patch is required because the current state of
> swiotlb_search_pool_area() conflates the DMA alignment mask, the
> allocation alignment mask and the stride so that even if a non-zero
> 'alloc_align_mask' is passed in, it won't necessarily be honoured.

Sure, I didn't mean to suggest there wasn't anything to fix here - if 
the existing code was intending to align to PAGE_SIZE even for a 
alloc_align_mask=0 and failing then that's clearly its own bug - I'm 
mostly being confused by the example of returning an unsuitably-aligned 
address for an 8KB dma_alloc_coherent() 75% of the time, if the end 
result of this fix is that we'll *still* return an incorrectly-aligned 
buffer for that same request 50% of the time (which just happens to be 
less fatal), since there are two separate bugs in that path.

Cheers,
Robin.

> For example, I just gave it a spin with only patch #3 and then this log:
>
> > >    # Final address in brackets is the SWIOTLB address returned to the caller
> > >    | virtio-pci 0000:00:07.0: orig_addr 0x0 alloc_size 0x2000, iotlb_align_mask 0x800 stride 0x2: got slot 1645-1649/7168 (0x98326800)
>
> Becomes:
>
>    | virtio-pci 0000:00:07.0: alloc_size 0x2000, iotlb_align_mask 0x1800 stride 0x4: got slot 1645-1649/7168 (0x98326800)
>
> So even though the stride is correct, we still end up with a 2KiB aligned
> allocation.
>
> Cheers,
>
> Will