[Nouveau] [PATCH 4/4] drm/nouveau: introduce CPU cache flushing macro

[Alexandre Courbot]

On Mon, Jun 9, 2014 at 7:41 PM, Alexandre Courbot <gnurou at gmail.com> wrote:
> On Mon, May 19, 2014 at 6:22 PM, Lucas Stach <l.stach at pengutronix.de> wrote:
>> Am Montag, den 19.05.2014, 11:02 +0200 schrieb Thierry Reding:
>>> On Mon, May 19, 2014 at 04:10:58PM +0900, Alexandre Courbot wrote:
>>> > Some architectures (e.g. ARM) need the CPU buffers to be explicitely
>>> > flushed for a memory write to take effect. Not doing so results in
>>> > synchronization issues, especially after writing to BOs.
>>>
>>> It seems to me that the above is generally true for all architectures,
>>> not just ARM.
>>>
>> No, on PCI coherent arches, like x86 and some PowerPCs, the GPU will
>> snoop the CPU caches and therefore an explicit cache flush is not
>> required.
>>
>>> Also: s/explicitely/explicitly/
>>>
>>> > This patch introduces a macro that flushes the caches on ARM and
>>> > translates to a no-op on other architectures, and uses it when
>>> > writing to in-memory BOs. It will also be useful for implementations of
>>> > instmem that access shared memory directly instead of going through
>>> > PRAMIN.
>>>
>>> Presumably instmem can access shared memory on all architectures, so
>>> this doesn't seem like a property of the architecture but rather of the
>>> memory pool backing the instmem.
>>>
>>> In that case I wonder if this shouldn't be moved into an operation that
>>> is implemented by the backing memory pool and be a noop where the cache
>>> doesn't need explicit flushing.
>>>
>>> > diff --git a/drivers/gpu/drm/nouveau/core/os.h b/drivers/gpu/drm/nouveau/core/os.h
>>> > index d0ced94ca54c..274b4460bb03 100644
>>> > --- a/drivers/gpu/drm/nouveau/core/os.h
>>> > +++ b/drivers/gpu/drm/nouveau/core/os.h
>>> > @@ -38,4 +38,21 @@
>>> >  #endif /* def __BIG_ENDIAN else */
>>> >  #endif /* !ioread32_native */
>>> >
>>> > +#if defined(__arm__)
>>> > +
>>> > +#define nv_cpu_cache_flush_area(va, size)  \
>>> > +do {                                               \
>>> > +   phys_addr_t pa = virt_to_phys(va);      \
>>> > +   __cpuc_flush_dcache_area(va, size);     \
>>> > +   outer_flush_range(pa, pa + size);       \
>>> > +} while (0)
>>>
>>> Couldn't this be a static inline function?
>>>
>>> > diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> [...]
>>> > index 0886f47e5244..b9c9729c5733 100644
>>> > --- a/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> > +++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> > @@ -437,8 +437,10 @@ nouveau_bo_wr16(struct nouveau_bo *nvbo, unsigned index, u16 val)
>>> >     mem = &mem[index];
>>> >     if (is_iomem)
>>> >             iowrite16_native(val, (void __force __iomem *)mem);
>>> > -   else
>>> > +   else {
>>> >             *mem = val;
>>> > +           nv_cpu_cache_flush_area(mem, 2);
>>> > +   }
>>> >  }
>>> >
>>> >  u32
>>> > @@ -461,8 +463,10 @@ nouveau_bo_wr32(struct nouveau_bo *nvbo, unsigned index, u32 val)
>>> >     mem = &mem[index];
>>> >     if (is_iomem)
>>> >             iowrite32_native(val, (void __force __iomem *)mem);
>>> > -   else
>>> > +   else {
>>> >             *mem = val;
>>> > +           nv_cpu_cache_flush_area(mem, 4);
>>> > +   }
>>>
>>> This looks rather like a sledgehammer to me. Effectively this turns nvbo
>>> into an uncached buffer. With additional overhead of constantly flushing
>>> caches. Wouldn't it make more sense to locate the places where these are
>>> called and flush the cache after all the writes have completed?
>>>
>> I don't think the explicit flushing for those things makes sense. I
>> think it is a lot more effective to just map the BOs write-combined on
>> PCI non-coherent arches. This way any writes will be buffered. Reads
>> will be slow, but I don't think nouveau is reading back a lot from those
>> buffers.
>> Using the write-combining buffer doesn't need any additional
>> synchronization as it will get flushed on pushbuf kickoff anyways.
>
> I tried to go that way, and something interesting happened.
>
> What I did: remove this patch and instead set the following caching
> parameters for the TTM_PL_TT case in nouveau_bo_init_mem_type():
>
>     man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC;
>     man->default_caching = TTM_PL_FLAG_WC;
>
> What happened: no runtime errors as what happened when caching is
> enabled. However, many of the vertex and texture buffers seem to be
> partially corrupted. In glmark2 the 3d models had many vertices (but
> not all) at the wrong position. Note that not all the scenes ended up
> being corrupted - in particular, when two consecutive scenes used the
> same model, the second instance would be uncorrupted.
>
> Forcing the caching to TTM_PL_FLAG_UNCACHED led to the same result.
> What is interesting is that while data like vertices and textures got
> corrupted, pushbuffers and shader programs seem to be just fine, as I
> could not see any runtime error.

An interesting fact: if I change ttm_bo_kmap_ttm() such as kernel
mappings of BOs are always performed write-combined, and leave the
TTM_PL_TT default caching to TTM_PL_FLAG_CACHED so user-space mappings
remain cached, the corruptions just vanish. It seems to be the fact of
setting user-space mappings to anything non-cached that leads to this
puzzling behavior. Certainly some subtlety of ARM mappings are getting
over my head here.

If we need to implement different policies for kernel and user-space
mappings, this might complicate things a bit, especially since support
needs to be in TTM and not only Nouveau. I will submit a RFC tomorrow
if I don't hear better ideas by then.

Alex.