Re: [PATCH 13/14] staging: android: ion: Do not sync CPU cache on map/unmap

From: Liam Mark
Date: Mon Jan 21 2019 - 15:11:24 EST


On Fri, 18 Jan 2019, Andrew F. Davis wrote:

> On 1/17/19 7:11 PM, Liam Mark wrote:
> > On Thu, 17 Jan 2019, Andrew F. Davis wrote:
> >
> >> On 1/16/19 4:54 PM, Liam Mark wrote:
> >>> On Wed, 16 Jan 2019, Andrew F. Davis wrote:
> >>>
> >>>> On 1/16/19 9:19 AM, Brian Starkey wrote:
> >>>>> Hi :-)
> >>>>>
> >>>>> On Tue, Jan 15, 2019 at 12:40:16PM -0600, Andrew F. Davis wrote:
> >>>>>> On 1/15/19 12:38 PM, Andrew F. Davis wrote:
> >>>>>>> On 1/15/19 11:45 AM, Liam Mark wrote:
> >>>>>>>> On Tue, 15 Jan 2019, Andrew F. Davis wrote:
> >>>>>>>>
> >>>>>>>>> On 1/14/19 11:13 AM, Liam Mark wrote:
> >>>>>>>>>> On Fri, 11 Jan 2019, Andrew F. Davis wrote:
> >>>>>>>>>>
> >>>>>>>>>>> Buffers may not be mapped from the CPU so skip cache maintenance here.
> >>>>>>>>>>> Accesses from the CPU to a cached heap should be bracketed with
> >>>>>>>>>>> {begin,end}_cpu_access calls so maintenance should not be needed anyway.
> >>>>>>>>>>>
> >>>>>>>>>>> Signed-off-by: Andrew F. Davis <afd@xxxxxx>
> >>>>>>>>>>> ---
> >>>>>>>>>>> drivers/staging/android/ion/ion.c | 7 ++++---
> >>>>>>>>>>> 1 file changed, 4 insertions(+), 3 deletions(-)
> >>>>>>>>>>>
> >>>>>>>>>>> diff --git a/drivers/staging/android/ion/ion.c b/drivers/staging/android/ion/ion.c
> >>>>>>>>>>> index 14e48f6eb734..09cb5a8e2b09 100644
> >>>>>>>>>>> --- a/drivers/staging/android/ion/ion.c
> >>>>>>>>>>> +++ b/drivers/staging/android/ion/ion.c
> >>>>>>>>>>> @@ -261,8 +261,8 @@ static struct sg_table *ion_map_dma_buf(struct dma_buf_attachment *attachment,
> >>>>>>>>>>>
> >>>>>>>>>>> table = a->table;
> >>>>>>>>>>>
> >>>>>>>>>>> - if (!dma_map_sg(attachment->dev, table->sgl, table->nents,
> >>>>>>>>>>> - direction))
> >>>>>>>>>>> + if (!dma_map_sg_attrs(attachment->dev, table->sgl, table->nents,
> >>>>>>>>>>> + direction, DMA_ATTR_SKIP_CPU_SYNC))
> >>>>>>>>>>
> >>>>>>>>>> Unfortunately I don't think you can do this for a couple reasons.
> >>>>>>>>>> You can't rely on {begin,end}_cpu_access calls to do cache maintenance.
> >>>>>>>>>> If the calls to {begin,end}_cpu_access were made before the call to
> >>>>>>>>>> dma_buf_attach then there won't have been a device attached so the calls
> >>>>>>>>>> to {begin,end}_cpu_access won't have done any cache maintenance.
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> That should be okay though, if you have no attachments (or all
> >>>>>>>>> attachments are IO-coherent) then there is no need for cache
> >>>>>>>>> maintenance. Unless you mean a sequence where a non-io-coherent device
> >>>>>>>>> is attached later after data has already been written. Does that
> >>>>>>>>> sequence need supporting?
> >>>>>>>>
> >>>>>>>> Yes, but also I think there are cases where CPU access can happen before
> >>>>>>>> in Android, but I will focus on later for now.
> >>>>>>>>
> >>>>>>>>> DMA-BUF doesn't have to allocate the backing
> >>>>>>>>> memory until map_dma_buf() time, and that should only happen after all
> >>>>>>>>> the devices have attached so it can know where to put the buffer. So we
> >>>>>>>>> shouldn't expect any CPU access to buffers before all the devices are
> >>>>>>>>> attached and mapped, right?
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>> Here is an example where CPU access can happen later in Android.
> >>>>>>>>
> >>>>>>>> Camera device records video -> software post processing -> video device
> >>>>>>>> (who does compression of raw data) and writes to a file
> >>>>>>>>
> >>>>>>>> In this example assume the buffer is cached and the devices are not
> >>>>>>>> IO-coherent (quite common).
> >>>>>>>>
> >>>>>>>
> >>>>>>> This is the start of the problem, having cached mappings of memory that
> >>>>>>> is also being accessed non-coherently is going to cause issues one way
> >>>>>>> or another. On top of the speculative cache fills that have to be
> >>>>>>> constantly fought back against with CMOs like below; some coherent
> >>>>>>> interconnects behave badly when you mix coherent and non-coherent access
> >>>>>>> (snoop filters get messed up).
> >>>>>>>
> >>>>>>> The solution is to either always have the addresses marked non-coherent
> >>>>>>> (like device memory, no-map carveouts), or if you really want to use
> >>>>>>> regular system memory allocated at runtime, then all cached mappings of
> >>>>>>> it need to be dropped, even the kernel logical address (area as painful
> >>>>>>> as that would be).
> >>>>>
> >>>>> Ouch :-( I wasn't aware about these potential interconnect issues. How
> >>>>> "real" is that? It seems that we aren't really hitting that today on
> >>>>> real devices.
> >>>>>
> >>>>
> >>>> Sadly there is at least one real device like this now (TI AM654). We
> >>>> spent some time working with the ARM interconnect spec designers to see
> >>>> if this was allowed behavior, final conclusion was mixing coherent and
> >>>> non-coherent accesses is never a good idea.. So we have been working to
> >>>> try to minimize any cases of mixed attributes [0], if a region is
> >>>> coherent then everyone in the system needs to treat it as such and
> >>>> vice-versa, even clever CMO that work on other systems wont save you
> >>>> here. :(
> >>>>
> >>>> [0] https://github.com/ARM-software/arm-trusted-firmware/pull/1553
> >>>>
> >>>>
> >>>>>>>
> >>>>>>>> ION buffer is allocated.
> >>>>>>>>
> >>>>>>>> //Camera device records video
> >>>>>>>> dma_buf_attach
> >>>>>>>> dma_map_attachment (buffer needs to be cleaned)
> >>>>>>>
> >>>>>>> Why does the buffer need to be cleaned here? I just got through reading
> >>>>>>> the thread linked by Laura in the other reply. I do like +Brian's
> >>>>>>
> >>>>>> Actually +Brian this time :)
> >>>>>>
> >>>>>>> suggestion of tracking if the buffer has had CPU access since the last
> >>>>>>> time and only flushing the cache if it has. As unmapped heaps never get
> >>>>>>> CPU mapped this would never be the case for unmapped heaps, it solves my
> >>>>>>> problem.
> >>>>>>>
> >>>>>>>> [camera device writes to buffer]
> >>>>>>>> dma_buf_unmap_attachment (buffer needs to be invalidated)
> >>>>>>>
> >>>>>>> It doesn't know there will be any further CPU access, it could get freed
> >>>>>>> after this for all we know, the invalidate can be saved until the CPU
> >>>>>>> requests access again.
> >>>>>
> >>>>> We don't have any API to allow the invalidate to happen on CPU access
> >>>>> if all devices already detached. We need a struct device pointer to
> >>>>> give to the DMA API, otherwise on arm64 there'll be no invalidate.
> >>>>>
> >>>>> I had a chat with a few people internally after the previous
> >>>>> discussion with Liam. One suggestion was to use
> >>>>> DMA_ATTR_SKIP_CPU_SYNC in unmap_dma_buf, but only if there's at least
> >>>>> one other device attached (guarantees that we can do an invalidate in
> >>>>> the future if begin_cpu_access is called). If the last device
> >>>>> detaches, do a sync then.
> >>>>>
> >>>>> Conversely, in map_dma_buf, we would track if there was any CPU access
> >>>>> and use/skip the sync appropriately.
> >>>>>
> >>>>
> >>>> Now that I think this all through I agree this patch is probably wrong.
> >>>> The real fix needs to be better handling in the dma_map_sg() to deal
> >>>> with the case of the memory not being mapped (what I'm dealing with for
> >>>> unmapped heaps), and for cases when the memory in question is not cached
> >>>> (Liam's issue I think). For both these cases the dma_map_sg() does the
> >>>> wrong thing.
> >>>>
> >>>>> I did start poking the code to check out how that would look, but then
> >>>>> Christmas happened and I'm still catching back up.
> >>>>>
> >>>>>>>
> >>>>>>>> dma_buf_detach (device cannot stay attached because it is being sent down
> >>>>>>>> the pipeline and Camera doesn't know the end of the use case)
> >>>>>>>>
> >>>>>>>
> >>>>>>> This seems like a broken use-case, I understand the desire to keep
> >>>>>>> everything as modular as possible and separate the steps, but at this
> >>>>>>> point no one owns this buffers backing memory, not the CPU or any
> >>>>>>> device. I would go as far as to say DMA-BUF should be free now to
> >>>>>>> de-allocate the backing storage if it wants, that way it could get ready
> >>>>>>> for the next attachment, which may change the required backing memory
> >>>>>>> completely.
> >>>>>>>
> >>>>>>> All devices should attach before the first mapping, and only let go
> >>>>>>> after the task is complete, otherwise this buffers data needs copied off
> >>>>>>> to a different location or the CPU needs to take ownership in-between.
> >>>>>>>
> >>>>>
> >>>>> Yeah.. that's certainly the theory. Are there any DMA-BUF
> >>>>> implementations which actually do that? I hear it quoted a lot,
> >>>>> because that's what the docs say - but if the reality doesn't match
> >>>>> it, maybe we should change the docs.
> >>>>>
> >>>>
> >>>> Do you mean on the userspace side? I'm not sure, seems like Android
> >>>> might be doing this wrong from what I can gather. From kernel side if
> >>>> you mean the "de-allocate the backing storage", we will have some cases
> >>>> like this soon, so I want to make sure userspace is not abusing DMA-BUF
> >>>> in ways not specified in the documentation. Changing the docs to force
> >>>> the backing memory to always be allocated breaks the central goal in
> >>>> having attach/map in DMA-BUF separate.
> >>>>
> >>>>>>>> //buffer is send down the pipeline
> >>>>>>>>
> >>>>>>>> // Usersapce software post processing occurs
> >>>>>>>> mmap buffer
> >>>>>>>
> >>>>>>> Perhaps the invalidate should happen here in mmap.
> >>>>>>>
> >>>>>>>> DMA_BUF_IOCTL_SYNC IOCT with flags DMA_BUF_SYNC_START // No CMO since no
> >>>>>>>> devices attached to buffer
> >>>>>>>
> >>>>>>> And that should be okay, mmap does the sync, and if no devices are
> >>>>>>> attached nothing could have changed the underlying memory in the
> >>>>>>> mean-time, DMA_BUF_SYNC_START can safely be a no-op as they are.
> >>>>>
> >>>>> Yeah, that's true - so long as you did an invalidate in unmap_dma_buf.
> >>>>> Liam was saying that it's too painful for them to do that every time a
> >>>>> device unmaps - when in many cases (device->device, no CPU) it's not
> >>>>> needed.
> >>>>
> >>>> Invalidates are painless, at least compared to a real cache flush, just
> >>>> set the invalid bit vs actually writing out lines. I thought the issue
> >>>> was on the map side.
> >>>>
> >>>
> >>> Invalidates aren't painless for us because we have a coherent system cache
> >>> so clean lines get written out.
> >>
> >> That seems very broken, why would clean lines ever need to be written
> >> out, that defeats the whole point of having the invalidate separate from
> >> clean. How do you deal with stale cache lines? I guess in your case this
> >> is what forces you to have to use uncached memory for DMA-able memory.
> >>
> >
> > My understanding is that our ARM invalidate is a clean + invalidate, I had
> > concerns about the clean lines being written to the the system cache as
> > part of the 'clean', but the following 'invalidate' would take care of
> > actually invalidating the lines (so nothign broken).
> > But i am probably wrong on this and it is probably smart enough not to the
> > writing of the clean lines.
> >
>
> You are correct that for a lot of ARM cores "invalidate" is always a
> "clean + invalidate". At first I thought this was kinda silly as there
> is now no way to mark a dirty line invalid without it getting written
> out first, but if you think about it any dirty cache-line can be written
> out (cleaned) at anytime anyway, so this doesn't actually change system
> behavior. You should just not write to memory (make the line dirty)
> anything you don't want eventually written out.
>
> Point two, it's not just smart enough to not write-out clean lines, it
> is guaranteed not to write them out by the spec. Otherwise since
> cache-lines can be randomly filled if those same clean lines got written
> out on invalidate operations there would be no way to maintain coherency
> and things would be written over top each other all over the place.
>
> > But regardless, targets supporting a coherent system cache is a legitamate
> > configuration and an invalidate on this configuration does have to go to
> > the bus to invalidate the system cache (which isn't free) so I dont' think
> > you can make the assumption that invalidates are cheap so that it is okay
> > to do them (even if they are not needed) on every dma unmap.
> >
>
> Very true, CMOs need to be broadcast to other coherent masters on a
> coherent interconnect (and the interconnect itself if it has a cache as
> well (L3)), so not 100% free, but almost, just the infinitesimal cost of
> the cache tag check in hardware. If there are no non-coherent devices
> attached then the CMOs are no-ops, if there are then the data needs to
> be written out either way, doing it every access like is done with
> uncached memory (- any write combining) will blow away any saving made
> from the one less CMO. Either way you lose with uncached mappings of
> memory. If I'm wrong I would love to know.
>

I would need to think about this more before replying.

> >>> And these invalidates can occur on fairly large buffers.
> >>>
> >>> That is why we haven't went with using cached ION memory and "tracking CPU
> >>> access" because it only solves half the problem, ie there isn't a way to
> >>> safely skip the invalidate (because we can't read the future).
> >>> Our solution was to go with uncached ION memory (when possible), but as
> >>> you can see in other discussions upstream support for uncached memory has
> >>> its own issues.
> >>>
> >>
> >> Sounds like you need to fix upstream support then, finding a way to drop
> >> all cacheable mappings of memory you want to make uncached mappings for
> >> seems to be the only solution.
> >>
> >
> > I think we can probably agree that there woudln't be a good way to remove
> > cached mappings without causing an unacceptable performance degradation
> > since it would fragment all the nice 1GB kernel mappings we have.
> >
> > So I am trying to find an alternative solution.
> >
>
> I'm not sure there is a better solution. How hard is this solution to
> implement anyway? The kernel already has to make gaps and cut up that
> nice 1GB mapping when you make a reserved memory space in the lowmem
> area, so all the logic is probably already implemented. Just need to
> allow it to be hooked into from Ion when doing doing the uncached mappings.
>

Even before attempting to implement it may be best to first run an
experiment where block mappings are disabled for the kernel mappings in
order to simulate a system where a lot of memory is allocated with
uncached mappings (such as could happen when using ION with Android). Then
try running any important benchmarks, my expectation is
that the performance impact of losing these kernel mappings will be
unacceptable.

It was a couple years ago, but I remember when I played around with the
kernel block mappings it was impacting some of our benchmarks.

> >>>>>
> >>>>>>>
> >>>>>>>> [CPU reads/writes to the buffer]
> >>>>>>>> DMA_BUF_IOCTL_SYNC IOCTL with flags DMA_BUF_SYNC_END // No CMO since no
> >>>>>>>> devices attached to buffer
> >>>>>>>> munmap buffer
> >>>>>>>>
> >>>>>>>> //buffer is send down the pipeline
> >>>>>>>> // Buffer is send to video device (who does compression of raw data) and
> >>>>>>>> writes to a file
> >>>>>>>> dma_buf_attach
> >>>>>>>> dma_map_attachment (buffer needs to be cleaned)
> >>>>>>>> [video device writes to buffer]
> >>>>>>>> dma_buf_unmap_attachment
> >>>>>>>> dma_buf_detach (device cannot stay attached because it is being sent down
> >>>>>>>> the pipeline and Video doesn't know the end of the use case)
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>
> >>>>>>>>>> Also ION no longer provides DMA ready memory, so if you are not doing CPU
> >>>>>>>>>> access then there is no requirement (that I am aware of) for you to call
> >>>>>>>>>> {begin,end}_cpu_access before passing the buffer to the device and if this
> >>>>>>>>>> buffer is cached and your device is not IO-coherent then the cache maintenance
> >>>>>>>>>> in ion_map_dma_buf and ion_unmap_dma_buf is required.
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> If I am not doing any CPU access then why do I need CPU cache
> >>>>>>>>> maintenance on the buffer?
> >>>>>>>>>
> >>>>>>>>
> >>>>>>>> Because ION no longer provides DMA ready memory.
> >>>>>>>> Take the above example.
> >>>>>>>>
> >>>>>>>> ION allocates memory from buddy allocator and requests zeroing.
> >>>>>>>> Zeros are written to the cache.
> >>>>>>>>
> >>>>>>>> You pass the buffer to the camera device which is not IO-coherent.
> >>>>>>>> The camera devices writes directly to the buffer in DDR.
> >>>>>>>> Since you didn't clean the buffer a dirty cache line (one of the zeros) is
> >>>>>>>> evicted from the cache, this zero overwrites data the camera device has
> >>>>>>>> written which corrupts your data.
> >>>>>>>>
> >>>>>>>
> >>>>>>> The zeroing *is* a CPU access, therefor it should handle the needed CMO
> >>>>>>> for CPU access at the time of zeroing.
> >>>>>>>
> >>>>>
> >>>>> Actually that should be at the point of the first non-coherent device
> >>>>> mapping the buffer right? No point in doing CMO if the future accesses
> >>>>> are coherent.
> >>>>
> >>>> I see your point, as long as the zeroing is guaranteed to be the first
> >>>> access to this buffer then it should be safe.
> >>>>
> >>>> Andrew
> >>>>
> >>>>>
> >>>>> Cheers,
> >>>>> -Brian
> >>>>>
> >>>>>>> Andrew
> >>>>>>>
> >>>>>>>> Liam
> >>>>>>>>
> >>>>>>>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> >>>>>>>> a Linux Foundation Collaborative Project
> >>>>>>>>
> >>>>
> >>>
> >>> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> >>> a Linux Foundation Collaborative Project
> >>>
> >>
> >
> > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
> > a Linux Foundation Collaborative Project
> >
>

Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project