Re: [PATCH 0/2] Fix double allocation in swiotlb_alloc()

[Petr Tesařík]

Hi Will,

On Fri, 26 Jan 2024 15:19:54 +0000
Will Deacon <will@xxxxxxxxxx> wrote:

> Hi folks,
> 
> These two patches fix a nasty double allocation problem in swiotlb_alloc()
> and add a diagnostic to help catch any similar issues in future. This was
> a royal pain to track down and I've had to make a bit of a leap at the
> correct alignment semantics (i.e. iotlb_align_mask vs alloc_align_mask).

Welcome to the club. I believe you had to re-discover what I described here:

  https://lore.kernel.org/linux-iommu/20231108101347.77cab795@xxxxxxxxxxxxxxxxxxxx/

The relevant part would be this:

  To sum it up, there are two types of alignment:

  1. specified by a device's min_align_mask; this says how many low
     bits of a buffer's physical address must be preserved,

  2. specified by allocation size and/or the alignment parameter;
     this says how many low bits in the first IO TLB slot's physical
     address must be zero.  

Fix for that has been sitting on my TODO list for too long. :-(

Petr T

> Without these changes, we've been observing random vsock hangs when
> communicating with virtual machines in Android.
> 
> Please have a look!
> 
> Cheers,
> 
> Will
> 
> Cc: iommu@xxxxxxxxxxxxxxx
> Cc: Christoph Hellwig <hch@xxxxxx>
> Cc: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
> Cc: Robin Murphy <robin.murphy@xxxxxxx>
> Cc: Petr Tesarik <petr.tesarik1@xxxxxxxxxxxxxxxxxxx>
> Cc: Dexuan Cui <decui@xxxxxxxxxxxxx>
> 
> --->8  
> 
> Will Deacon (2):
>   swiotlb: Fix allocation alignment requirement when searching slots
>   swiotlb: Enforce page alignment in swiotlb_alloc()
> 
>  kernel/dma/swiotlb.c | 32 ++++++++++++++++++++++----------
>  1 file changed, 22 insertions(+), 10 deletions(-)
>