Re: [RFC v1 3/4] swiotlb: Allow dynamic allocation of bounce buffers

[Juerg Haefliger <juergh@proton.me>]

[-- Attachment #1: Type: text/plain, Size: 5229 bytes --]

On Tue, 28 Mar 2023 09:54:35 +0200
Petr Tesarik <petrtesarik@huaweicloud.com> wrote:

> On 3/28/2023 6:07 AM, Christoph Hellwig wrote:
> > [adding Alex as he has been interested in this in the past]
> >
> > On Mon, Mar 20, 2023 at 01:28:15PM +0100, Petr Tesarik wrote:
> >> Second, on the Raspberry Pi 4, swiotlb is used by dma-buf for pages
> >> moved from the rendering GPU (v3d driver), which can access all
> >> memory, to the display output (vc4 driver), which is connected to a
> >> bus with an address limit of 1 GiB and no IOMMU. These buffers can
> >> be large (several megabytes) and cannot be handled by SWIOTLB,
> >> because they exceed maximum segment size of 256 KiB. Such mapping
> >> failures can be easily reproduced on a Raspberry Pi4: Starting
> >> GNOME remote desktop results in a flood of kernel messages like
> >> these:
> >
> > Shouldn't we make sure dma-buf allocates the buffers for the most
> > restricted devices, and more importantly does something like a dma
> > coherent allocation instead of a dynamic mapping of random memory?
> >
> > While a larger swiotlb works around this I don't think this fixes the root
> > cause.
>
> I tend to agree here. However, it's the DMABUF design itself that causes
> some trouble. The buffer is allocated by the v3d driver, which does not
> have the restriction, so the DMA API typically allocates an address
> somewhere near the 4G boundary. Userspace then exports the buffer, sends
> it to another process as a file descriptor and imports it into the vc4
> driver, which requires DMA below 1G. In the beginning, v3d had no idea
> that the buffer would be exported to userspace, much less that it would
> be later imported into vc4.
>
> Anyway, I suspected that the buffers need not be imported into the vc4
> driver (also hinted by Eric Anholt in a 2018 blog post [1]), and it
> seems I was right. I encountered the issue with Ubuntu 22.10; I
> installed latest openSUSE Tumbleweed yesterday, and I was not able to
> reproduce the issue there, most likely because the Mesa drivers have
> been fixed meanwhile. This makes the specific case of the Raspberry Pi 4
> drivers moot. The issue may still affect other combinations of drivers,
> but I don't have any other real-world example ATM.

I'm only seeing this problem with Wayland, no issue when switching Ubuntu to
X. It seems Tumbleweed is using X by default.

...Juerg


> [1] https://anholt.github.io/twivc4/2018/02/12/twiv/
>
> >> 1. The value is limited to ULONG_MAX, which is too little both for
> >>    physical addresses (e.g. x86 PAE or 32-bit ARM LPAE) and DMA
> >>    addresses (e.g. Xen guests on 32-bit ARM).
> >>
> >> 2. Since buffers are currently allocated with page granularity, a
> >>    PFN can be used instead. However, some values are reserved by
> >>    the maple tree implementation. Liam suggests to use
> >>    xa_mk_value() in that case, but that reduces the usable range by
> >>    half. Luckily, 31 bits are still enough to hold a PFN on all
> >>    32-bit platforms.
> >>
> >> 3. Software IO TLB is used from interrupt context. The maple tree
> >>    implementation is not IRQ-safe (MT_FLAGS_LOCK_IRQ does nothing
> >>    AFAICS). Instead, I use an external lock, spin_lock_irqsave() and
> >>    spin_unlock_irqrestore().
> >>
> >> Note that bounce buffers are never allocated dynamically if the
> >> software IO TLB is in fact a DMA restricted pool, which is intended
> >> to be stay in its designated location in physical memory.
> >
> > I'm a little worried about all that because it causes quite a bit
> > of overhead even for callers that don't end up going into the
> > dynamic range or do not use swiotlb at all.  I don't really have a
> > good answer here except for the usual avoid bounce buffering whenever
> > you can that might not always be easy to do.
>
> I'm also worried about all this overhead. OTOH I was not able to confirm
> it, because the difference between two successive fio test runs on an
> unmodified kernel was bigger than the difference between a vanilla and a
> patched kernel, except the maximum completion latency, which OTOH
> affected less than 0.01% of all requests.
>
> BTW my testing also suggests that the streaming DMA API is quite
> inefficient, because UAS performance _improved_ with swiotlb=force.
> Sure, this should probably be addressed in the UAS and/or xHCI driver,
> but what I mean is that moving away from swiotlb may even cause
> performance regressions, which is counter-intuitive. At least I would
> _not_ have expected it.
>
> >> +	gfp = (attrs & DMA_ATTR_MAY_SLEEP) ? GFP_KERNEL : GFP_NOWAIT;
> >> +	slot = kmalloc(sizeof(*slot), gfp | __GFP_NOWARN);
> >> +	if (!slot)
> >> +		goto err;
> >> +
> >> +	slot->orig_addr = orig_addr;
> >> +	slot->alloc_size = alloc_size;
> >> +	slot->page = dma_direct_alloc_pages(dev, PAGE_ALIGN(alloc_size),
> >> +					    &slot->dma_addr, dir,
> >> +					    gfp | __GFP_NOWARN);
> >> +	if (!slot->page)
> >> +		goto err_free_slot;
> >
> > Without GFP_NOIO allocations this will deadlock eventually.
>
> Ah, that would affect the non-sleeping case (GFP_KERNEL), right?
>
> Petr T
>


[-- Attachment #2: attachment.sig --]
[-- Type: application/pgp-signature, Size: 849 bytes --]