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

[Christoph Hellwig]

[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.

> 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.

> +	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.