[PATCH] ARM: mm: dma: Update coherent streaming apis with missing memory barrier
catalin.marinas at arm.com
Tue Apr 22 08:07:25 PDT 2014
On Tue, Apr 22, 2014 at 02:49:06PM +0100, Santosh Shilimkar wrote:
> On Tuesday 22 April 2014 06:28 AM, Will Deacon wrote:
> > On Mon, Apr 21, 2014 at 07:03:10PM +0100, Santosh Shilimkar wrote:
> >> ARM coherent CPU dma map APIS are assumed to be nops on cache coherent
> >> machines. While this is true, one still needs to ensure that no
> >> outstanding writes are pending in CPU write buffers. To take care
> >> of that, we at least need a memory barrier to commit those changes
> >> to main memory.
> >> Patch is trying to fix those cases. Without such a patch, you will
> >> end up patching device drivers to avoid the synchronisation issues.
> > Don't you only need these barriers if you're passing ownership of a CPU
> > buffer to a device? In that case, I would expect a subsequent writel to tell
> > the device about the new buffer, which includes the required __iowmb().
> > That's the reason for the relaxed accessors: to avoid this barrier when it's
> > not needed. Perhaps you're using the relaxed accessors where you actually
> > need the stronger ordering guarantees?
> I kind of guessed some one will bring up above point. Infact this is how
> mostly people have been living with the issue on coherent machines. On
> Keystone too, we did explicit barriers in respective drivers.
> I have added these barriers only on CPU to device streaming APIs because on
> other direction, the memory is already upto date from CPU's perspective.
> But if you look at the actual problem, its really responsibility of
> DMA streaming APIs which we are trying to push on to drivers. A device
> driver should be independent of whether it is running on a coherent or
> a non-coherent CPU.
> Lets take a example....
> MMC controller driver running on a non-coherent and coherent machine.
> Driver has below code sequence which is generic.
> 1. Prepare SG list
> 2. Perform CMO using DMA streaming API
> 3. Start DMA transfer...
The key here is how you start the DMA transfer. So far we assumed it's
done via an I/O operation like writel() and it has the right barriers.
If we have other ways for starting this (like writing the dma_addr in
some other memory descriptor), should we use explicit memory barriers?
More information about the linux-arm-kernel