Re: [PATCH 5/6] dma-mapping: Allow batched DMA sync operations if supported by the arch

[Leon Romanovsky <leon@kernel.org>]

On Tue, Dec 23, 2025 at 01:02:55PM +1300, Barry Song wrote:
> On Mon, Dec 22, 2025 at 9:49 PM Leon Romanovsky <leon@kernel.org> wrote:
> >
> > On Mon, Dec 22, 2025 at 03:24:58AM +0800, Barry Song wrote:
> > > On Sun, Dec 21, 2025 at 7:55 PM Leon Romanovsky <leon@kernel.org> wrote:
> > > [...]
> > > > > +
> > > >
> > > > I'm wondering why you don't implement this batch‑sync support inside the
> > > > arch_sync_dma_*() functions. Doing so would minimize changes to the generic
> > > > kernel/dma/* code and reduce the amount of #ifdef‑based spaghetti.
> > > >
> > >
> > > There are two cases: mapping an sg list and mapping a single
> > > buffer. The former can be batched with
> > > arch_sync_dma_*_batch_add() and flushed via
> > > arch_sync_dma_batch_flush(), while the latter requires all work to
> > > be done inside arch_sync_dma_*(). Therefore,
> > > arch_sync_dma_*() cannot always batch and flush.
> >
> > Probably in all cases you can call the _batch_ variant, followed by _flush_,
> > even when handling a single page. This keeps the code consistent across all
> > paths. On platforms that do not support _batch_, the _flush_ operation will be
> > a NOP anyway.
> 
> We have a lot of code outside kernel/dma that also calls
> arch_sync_dma_for_* such as arch/arm, arch/mips, drivers/xen,
> I guess we don’t want to modify so many things?

Aren't they using internal, arch specific, arch_sync_dma_for_* implementations?

> 
> for kernel/dma, we have two "single" callers only:
> kernel/dma/direct.h, kernel/dma/swiotlb.c.  and they looks quite
> straightforward:
> 
> static inline void dma_direct_sync_single_for_device(struct device *dev,
>                 dma_addr_t addr, size_t size, enum dma_data_direction dir)
> {
>         phys_addr_t paddr = dma_to_phys(dev, addr);
> 
>         swiotlb_sync_single_for_device(dev, paddr, size, dir);
> 
>         if (!dev_is_dma_coherent(dev))
>                 arch_sync_dma_for_device(paddr, size, dir);
> }
> 
> I guess moving to arch_sync_dma_for_device_batch + flush
> doesn’t really look much better, does it?
> 
> >
> > I would also rename arch_sync_dma_batch_flush() to arch_sync_dma_flush().
> 
> Sure.
> 
> >
> > You can also minimize changes in dma_direct_map_phys() too, by extending
> > it's signature to provide if flush is needed or not.
> 
> Yes. I have
> 
> static inline dma_addr_t __dma_direct_map_phys(struct device *dev,
>                 phys_addr_t phys, size_t size, enum dma_data_direction dir,
>                 unsigned long attrs, bool flush)

My suggestion is to use it directly, without wrappers.

> 
> and two wrappers:
> static inline dma_addr_t dma_direct_map_phys(struct device *dev,
>                 phys_addr_t phys, size_t size, enum dma_data_direction dir,
>                 unsigned long attrs)
> {
>         return __dma_direct_map_phys(dev, phys, size, dir, attrs, true);
> }
> 
> static inline dma_addr_t dma_direct_map_phys_batch_add(struct device *dev,
>                 phys_addr_t phys, size_t size, enum dma_data_direction dir,
>                 unsigned long attrs)
> {
>         return __dma_direct_map_phys(dev, phys, size, dir, attrs, false);
> }
> 
> If you prefer exposing "flush" directly in dma_direct_map_phys()
> and updating its callers with flush=true, I think that’s fine.

Yes

> 
> It could be also true for dma_direct_sync_single_for_device().
> 
> >
> > dma_direct_map_phys(....) -> dma_direct_map_phys(...., bool flush):
> >
> > static inline dma_addr_t dma_direct_map_phys(...., bool flush)
> > {
> >         ....
> >
> >         if (dma_addr != DMA_MAPPING_ERROR && !dev_is_dma_coherent(dev) &&
> >             !(attrs & (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_MMIO)))
> >         {
> >                 arch_sync_dma_for_device(phys, size, dir);
> >                 if (flush)
> >                         arch_sync_dma_flush();
> >         }
> > }
> >
> 
> Thanks
> Barry
>