Re: [RFC PATCH] x86/pci-dma: add "any" keyword to swiotlb= kernel parameter

[Aakarsh Jain]

Thanks Robin, Marek for the review.

On 7/9/2026 5:58 PM, Robin Murphy wrote:
> On 08/07/2026 12:42 pm, Aakarsh Jain wrote:
> > CoCo guests (AMD SEV-SNP, Intel TDX) require large swiotlb pools for
> > streaming DMA workloads such as high-speed NIC and AI accelerator
> > inference. The existing swiotlb pool allocator restricts placement to
> > low memory (below 4GB by default), capping usable pool size at ~1GB even
> > when a larger pool is requested via swiotlb=<nslabs>.
> >
> > The SWIOTLB_ANY flag already exists to lift this restriction, and
> > swiotlb_init_remap() already handles it correctly via the flags
> > parameter (see CONFIG_SWIOTLB_DYNAMIC path: io_tlb_default_mem.phys_limit
> > is set to virt_to_phys(high_memory-1) when SWIOTLB_ANY is set).
> >
> > However, there is no way to set SWIOTLB_ANY from the command line. The
> > only existing mechanism was via arch-specific code (e.g. powerpc SVM sets
> > SWIOTLB_ANY in pci_iommu_init). x86 CoCo guests have no such path.
> >
> > After Aneesh  series ("dma-mapping: Track shared DMA state through
> > direct, pool and swiotlb paths", https://patchwork.kernel.org/project/ 
> > linux-arm-kernel/cover/20260701054926.825925-1-aneesh.kumar@xxxxxxxxxx/)
> > removes SWIOTLB_FORCE, x86 pci_swiotlb_detect() leaves 
> > x86_swiotlb_flags = 0 for
> > CoCo guests. The pool falls back to low memory and caps at ~1GB:
>
> This is entirely irrelevant; SWIOTLB_FORCE has no impact on allocation 
> behaviour anyway. x86 wasn't passing SWIOTLB_ANY before and it still 
> isn't, although there doesn't seem to be any particular reason why the 
> well-reviewed patch for that hasn't been picked up:
>
> https://lore.kernel.org/lkml/20260625012616.2992535-1-jun.miao@xxxxxxxxx/
You are right my commit message incorrectly tied SWIOTLB_FORCE removal 
to pool placement. SWIOTLB_FORCE only controls whether bounce buffering 
is forced at runtime, it has no bearing on where the pool is allocated. 
Placement is governed solely by SWIOTLB_ANY. I'll drop that paragraph 
entirely.
I see both you and Marek are pointing to this patch, which fixes the 
issue architecturally. I'm happy to help get that landed instead of 
adding a user facing knob. My only hesitation is the 32-bit concern below:
auto-enabling SWIOTLB_ANY for *all* CoCo guests carries the same 
breakage risk for limited-mask devices, which was flagged here:
https://lore.kernel.org/lkml/0a08cea7-62a2-4b33-8d8b-dd14b3c74235@xxxxxxx/

> >    Without "any": pool at 0x35a9c000 (~900MB, below 4GB boundary)
> >    With    "any": pool at 0x1df9c00000 (~120GB, anywhere in RAM)
> >    [Tested on AMD SEV-SNP guest, swiotlb=4194304]
> >
> > Add "any" as a new keyword to the swiotlb= kernel parameter. This is an
> > explicit, opt-in mechanism that sets SWIOTLB_ANY for the default pool at
> > boot time, without touching any arch-specific code.
> >
> > Devices with 32-bit DMA masks are not affected, they still use the normal
> > low-memory bounce buffer path. The "any" option is only meaningful for
> > workloads where all active DMA devices have 64-bit masks.
>
> That doesn't make any sense - if a user passes this option then any 
> devices with DMA addressing limitations definitely *are* going to be 
> affected, and quite likely broken altogether. If anything, the 
> documentation should be even more explicit that this should only be used 
> if you do know for sure that no devices have DMA addressing limitations.
You are correct here too. My statement that "32-bit devices are not 
affected" is wrong. There is a single default pool, once it moves above 
4GB via "any", a device with a <64-bit DMA mask has no reachable bounce 
buffer and will break.
The opt-in nature only means the admin accepted that risk, it does not 
make the risk go away. If this option survives at all, the documentation 
must state the opposite of what I wrote: use it ONLY when you are 
certain no device in the system has DMA addressing limitations.

> As a general SWIOTLB-behaviour-debugging option for orthogonality with 
> force/noforce I'm not opposed to the idea, but it is definitely not 
> something that real CoCo use-cases should rely on - if the general 
> consensus if that CoCo environments want a different setup by default 
> then the arch/CoCo code should be taking care of that.

So my takeaway, this patch, if kept, is at best a debugging knob for 
orthogonality with force/noforce, and not something real CoCo 
deployments should rely on. I'm fine positioning it that way, or 
dropping it.
Which brings me to the actual motivation behind posting this patch.
what should be recommended approach for large streaming DMA workloads in 
CoCo guests(high-speed NIC / AI accelerator inference)?
Is the intended direction to move such drivers to the coherent DMA path 
rather than growing or relocating the SWIOTLB pool?

Thanks,
Aakarsh


> Thanks,
> Robin.
>
> > Signed-off-by: Aakarsh Jain <aakarsh.jain@xxxxxxxxxxxxxxxx>
> > ---
> >   Documentation/admin-guide/kernel-parameters.txt | 5 ++++-
> >   kernel/dma/swiotlb.c                            | 5 +++++
> >   2 files changed, 9 insertions(+), 1 deletion(-)
> >
> > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/ 
> > Documentation/admin-guide/kernel-parameters.txt
> > index b5493a7f8f22..8a1fccbd9b25 100644
> > --- a/Documentation/admin-guide/kernel-parameters.txt
> > +++ b/Documentation/admin-guide/kernel-parameters.txt
> > @@ -7477,7 +7477,7 @@ Kernel parameters
> >               Execution Facility on pSeries.
> >       swiotlb=    [ARM,PPC,MIPS,X86,S390,EARLY]
> > -            Format: { <int> [,<int>] | force | noforce }
> > +            Format: { <int> [,<int>] | force | noforce | any}
> >               <int> -- Number of I/O TLB slabs
> >               <int> -- Second integer after comma. Number of swiotlb
> >                    areas with their own lock. Will be rounded up
> > @@ -7485,6 +7485,9 @@ Kernel parameters
> >               force -- force using of bounce buffers even if they
> >                        wouldn't be automatically used by the kernel
> >               noforce -- Never use bounce buffers (for debugging)
> > +            any --  Allow the swiotlb pool to be placed anywhere in
> > +                system RAM, lifting the default low-memory (4GB)
> > +                restriction.
> >       switches=    [HW,M68k,EARLY]
> > diff --git a/kernel/dma/swiotlb.c b/kernel/dma/swiotlb.c
> > index 1abd3e6146f4..34773ae7c770 100644
> > --- a/kernel/dma/swiotlb.c
> > +++ b/kernel/dma/swiotlb.c
> > @@ -80,6 +80,7 @@ struct io_tlb_slot {
> >   static bool swiotlb_force_bounce;
> >   static bool swiotlb_force_disable;
> > +static unsigned int swiotlb_param_flags __initdata;
> >   #ifdef CONFIG_SWIOTLB_DYNAMIC
> > @@ -198,6 +199,8 @@ setup_io_tlb_npages(char *str)
> >           swiotlb_force_bounce = true;
> >       else if (!strcmp(str, "noforce"))
> >           swiotlb_force_disable = true;
> > +    else if (!strcmp(str, "any"))
> > +        swiotlb_param_flags |= SWIOTLB_ANY;
> >       return 0;
> >   }
> > @@ -445,6 +448,8 @@ int swiotlb_init_late(size_t size, gfp_t gfp_mask,
> >       io_tlb_default_mem.force_bounce = swiotlb_force_bounce;
> > +    flags |= swiotlb_param_flags;
> > +
> >   #ifdef CONFIG_SWIOTLB_DYNAMIC
> >       if (!remap)
> >           io_tlb_default_mem.can_grow = true;