Re: [RFC PATCH v2] dmabuf-sync: Introduce buffer synchronization framework

On Tue, Jun 18, 2013 at 12:03:31AM +0900, Inki Dae wrote:

2013/6/17 Russell King - ARM Linux [off-list ref]
Exactly right. But that is not definitely my point. Could you please see
the below simple example?:
(Presume that CPU and DMA share a buffer and the buffer is mapped with user
space as cachable)

        handle1 = drm_gem_fd_to_handle(a dmabuf fd);  ----> 1
                 ...
        va1 = drm_gem_mmap(handle1);
        va2 = drm_gem_mmap(handle2);
        va3 = malloc(size);
                 ...

        while (conditions) {
                 memcpy(va1, some data, size);

Nooooooooooooooooooooooooooooooooooooooooooooo!

Well, the first thing to say here is that under the requirements of the
DMA API, the above is immediately invalid, because you're writing to a
buffer which under the terms of the DMA API is currently owned by the
DMA agent, *not* by the CPU.  You're supposed to call dma_sync_sg_for_cpu()
before you do that - but how is userspace supposed to know that requirement?
Why should userspace even _have_ to know these requirements of the DMA
API?

It's also entirely possible that drm_gem_fd_to_handle() (which indirectly
causes dma_map_sg() on the buffers scatterlist) followed by mmap'ing it
into userspace is a bug too, as it has the potential to touch caches or
stuff in ways that maybe the DMA or IOMMU may not expect - but I'm not
going to make too big a deal about that, because I don't think we have
anything that picky.

However, the first point above is the most important one, and exposing
the quirks of the DMA API to userland is certainly not a nice thing to be
doing.  This needs to be fixed - we can't go and enforce an API which is
deeply embedded within the kernel all the way out to userland.

What we need is something along the lines of:
(a) dma_buf_map_attachment() _not_ to map the scatterlist for DMA.
or
(b) drm_gem_prime_import() not to call dma_buf_map_attachment() at all.

and for the scatterlist to be mapped for DMA at the point where the DMA
operation is initiated, and unmapped at the point where the DMA operation
is complete.

So no, the problem is not that we need more APIs and code - we need the
existing kernel API fixed so that we don't go exposing userspace to the
requirements of the DMA API.  Unless we do that, we're going to end
up with a huge world of pain, where kernel architecture people need to
audit every damned DRM userspace implementation that happens to be run
on their platform, and that's not something arch people really can
afford to do.

Basically, I think the dma_buf stuff needs to be rewritten with the
requirements of the DMA API in the forefront of whosever mind is doing
the rewriting.

Note: the existing stuff does have the nice side effect of being able
to pass buffers which do not have a struct page * associated with them
through the dma_buf API - I think we can still preserve that by having
dma_buf provide a couple of new APIs to do the SG list map/sync/unmap,
but in any case we need to fix the existing API so that:

	dma_buf_map_attachment() becomes dma_buf_get_sg()
	dma_buf_unmap_attachment() becomes dma_buf_put_sg()

both getting rid of the DMA direction argument, and then we have four
new dma_buf calls:

	dma_buf_map_sg()
	dma_buf_unmap_sg()
	dma_buf_sync_sg_for_cpu()
	dma_buf_sync_sg_for_device()

which do the actual sg map/unmap via the DMA API *at the appropriate
time for DMA*.

So, the summary of this is - at the moment, I regard DRM Prime and dmabuf
to be utterly broken in design for architectures such as ARM where the
requirements of the DMA API have to be followed if you're going to have
a happy life.

On Mon, Jun 17, 2013 at 04:42:37PM +0100, Russell King - ARM Linux wrote:

On Tue, Jun 18, 2013 at 12:03:31AM +0900, Inki Dae wrote:

quoted

2013/6/17 Russell King - ARM Linux [off-list ref]
Exactly right. But that is not definitely my point. Could you please see
the below simple example?:
(Presume that CPU and DMA share a buffer and the buffer is mapped with user
space as cachable)

        handle1 = drm_gem_fd_to_handle(a dmabuf fd);  ----> 1
                 ...
        va1 = drm_gem_mmap(handle1);
        va2 = drm_gem_mmap(handle2);
        va3 = malloc(size);
                 ...

        while (conditions) {
                 memcpy(va1, some data, size);

Nooooooooooooooooooooooooooooooooooooooooooooo!

Well, the first thing to say here is that under the requirements of the
DMA API, the above is immediately invalid, because you're writing to a
buffer which under the terms of the DMA API is currently owned by the
DMA agent, *not* by the CPU.  You're supposed to call dma_sync_sg_for_cpu()
before you do that - but how is userspace supposed to know that requirement?
Why should userspace even _have_ to know these requirements of the DMA
API?

It's also entirely possible that drm_gem_fd_to_handle() (which indirectly
causes dma_map_sg() on the buffers scatterlist) followed by mmap'ing it
into userspace is a bug too, as it has the potential to touch caches or
stuff in ways that maybe the DMA or IOMMU may not expect - but I'm not
going to make too big a deal about that, because I don't think we have
anything that picky.

However, the first point above is the most important one, and exposing
the quirks of the DMA API to userland is certainly not a nice thing to be
doing.  This needs to be fixed - we can't go and enforce an API which is
deeply embedded within the kernel all the way out to userland.

What we need is something along the lines of:
(a) dma_buf_map_attachment() _not_ to map the scatterlist for DMA.
or
(b) drm_gem_prime_import() not to call dma_buf_map_attachment() at all.

and for the scatterlist to be mapped for DMA at the point where the DMA
operation is initiated, and unmapped at the point where the DMA operation
is complete.

So no, the problem is not that we need more APIs and code - we need the
existing kernel API fixed so that we don't go exposing userspace to the
requirements of the DMA API.  Unless we do that, we're going to end
up with a huge world of pain, where kernel architecture people need to
audit every damned DRM userspace implementation that happens to be run
on their platform, and that's not something arch people really can
afford to do.

Basically, I think the dma_buf stuff needs to be rewritten with the
requirements of the DMA API in the forefront of whosever mind is doing
the rewriting.

Note: the existing stuff does have the nice side effect of being able
to pass buffers which do not have a struct page * associated with them
through the dma_buf API - I think we can still preserve that by having
dma_buf provide a couple of new APIs to do the SG list map/sync/unmap,
but in any case we need to fix the existing API so that:

	dma_buf_map_attachment() becomes dma_buf_get_sg()
	dma_buf_unmap_attachment() becomes dma_buf_put_sg()

both getting rid of the DMA direction argument, and then we have four
new dma_buf calls:

	dma_buf_map_sg()
	dma_buf_unmap_sg()
	dma_buf_sync_sg_for_cpu()
	dma_buf_sync_sg_for_device()

which do the actual sg map/unmap via the DMA API *at the appropriate
time for DMA*.

So, the summary of this is - at the moment, I regard DRM Prime and dmabuf
to be utterly broken in design for architectures such as ARM where the
requirements of the DMA API have to be followed if you're going to have
a happy life.

An addendum to the above:

I'll also point out that the whole thing of having random buffers mapped
into userspace, and doing DMA on them is _highly_ architecture specific.
I'm told that PARISC is one architecture where this does not work (because
DMA needs to have some kind of congruence factor programmed into it which
can be different between kernel and userspace mappings of the same
physical mappings.

I ran into this when trying to come up with a cross-arch DMA-API mmap API,
and PARISC ended up killing the idea off (the remains of that attempt is
the dma_mmap_* stuff in ARM's asm/dma-mapping.h.)  However, this was for
the DMA coherent stuff, not the streaming API, which is what _this_
DMA prime/dma_buf stuff is about.

What I'm saying is think _very_ _carefully_ about userspace having
interleaved access to random DMA buffers.  Arguably, DRM should _not_
allow this.

On Mon, Jun 17, 2013 at 04:42:37PM +0100, Russell King - ARM Linux wrote:

On Tue, Jun 18, 2013 at 12:03:31AM +0900, Inki Dae wrote:

quoted

2013/6/17 Russell King - ARM Linux [off-list ref]
Exactly right. But that is not definitely my point. Could you please see
the below simple example?:
(Presume that CPU and DMA share a buffer and the buffer is mapped with user
space as cachable)

        handle1 = drm_gem_fd_to_handle(a dmabuf fd);  ----> 1
                 ...
        va1 = drm_gem_mmap(handle1);
        va2 = drm_gem_mmap(handle2);
        va3 = malloc(size);
                 ...

        while (conditions) {
                 memcpy(va1, some data, size);

Nooooooooooooooooooooooooooooooooooooooooooooo!

Well, the first thing to say here is that under the requirements of the
DMA API, the above is immediately invalid, because you're writing to a
buffer which under the terms of the DMA API is currently owned by the
DMA agent, *not* by the CPU.  You're supposed to call dma_sync_sg_for_cpu()
before you do that - but how is userspace supposed to know that requirement?
Why should userspace even _have_ to know these requirements of the DMA
API?

It's also entirely possible that drm_gem_fd_to_handle() (which indirectly
causes dma_map_sg() on the buffers scatterlist) followed by mmap'ing it
into userspace is a bug too, as it has the potential to touch caches or
stuff in ways that maybe the DMA or IOMMU may not expect - but I'm not
going to make too big a deal about that, because I don't think we have
anything that picky.

However, the first point above is the most important one, and exposing
the quirks of the DMA API to userland is certainly not a nice thing to be
doing.  This needs to be fixed - we can't go and enforce an API which is
deeply embedded within the kernel all the way out to userland.

What we need is something along the lines of:
(a) dma_buf_map_attachment() _not_ to map the scatterlist for DMA.
or
(b) drm_gem_prime_import() not to call dma_buf_map_attachment() at all.

I strongly vote for (b) (plus adding a dma_buf_map_sync interface to allow
syncing to other devices/cpu whithout dropping the dma mappings). At least
that's been the idea behind things, but currently all (x86-based) drm
drivers cut corners here massively.

Aside: The entire reason behind hiding the dma map step in the dma-buf
attachment is to allow drivers to expose crazy iommus (e.g. the tiler unit
on omap) to other devices. So imo (a) isn't the right choice.

and for the scatterlist to be mapped for DMA at the point where the DMA
operation is initiated, and unmapped at the point where the DMA operation
is complete.

So no, the problem is not that we need more APIs and code - we need the
existing kernel API fixed so that we don't go exposing userspace to the
requirements of the DMA API.  Unless we do that, we're going to end
up with a huge world of pain, where kernel architecture people need to
audit every damned DRM userspace implementation that happens to be run
on their platform, and that's not something arch people really can
afford to do.

Basically, I think the dma_buf stuff needs to be rewritten with the
requirements of the DMA API in the forefront of whosever mind is doing
the rewriting.

Note: the existing stuff does have the nice side effect of being able
to pass buffers which do not have a struct page * associated with them
through the dma_buf API - I think we can still preserve that by having
dma_buf provide a couple of new APIs to do the SG list map/sync/unmap,
but in any case we need to fix the existing API so that:

dma_buf_map_attachment() becomes dma_buf_get_sg()
dma_buf_unmap_attachment() becomes dma_buf_put_sg()

both getting rid of the DMA direction argument, and then we have four
new dma_buf calls:

dma_buf_map_sg()
dma_buf_unmap_sg()
dma_buf_sync_sg_for_cpu()
dma_buf_sync_sg_for_device()

which do the actual sg map/unmap via the DMA API *at the appropriate
time for DMA*.

Hm, my idea was to just add a dma_buf_sync_attchment for the device side
syncing, since the cpu access stuff is already bracketed with the
begin/end cpu access stuff. We might need a sync_for_cpu or so for mmap,
but imo mmap support for dma_buf is a bit insane anyway, so I don't care
too much about it.

Since such dma mappings would be really longstanding in most cases anyway
drivers could just map with BIDIRECTIONAL and do all the real flushing
with the new sync stuff.

Another piece of fun will be integration with the fencing stuff Maarten is
doing. I'm not sure whether we should integrate that into the dma-buf
interface (for simple users who don't want to deal with the full
complexity at least).

So, the summary of this is - at the moment, I regard DRM Prime and dmabuf
to be utterly broken in design for architectures such as ARM where the
requirements of the DMA API have to be followed if you're going to have
a happy life.

Agreed. Unfortunately there are not many real drivers shipping in
upstream, and x86 is _very_ forgiving about all this stuff.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

On Tue, Jun 18, 2013 at 09:00:16AM +0200, Daniel Vetter wrote:

On Mon, Jun 17, 2013 at 04:42:37PM +0100, Russell King - ARM Linux wrote:

quoted

What we need is something along the lines of:
(a) dma_buf_map_attachment() _not_ to map the scatterlist for DMA.
or
(b) drm_gem_prime_import() not to call dma_buf_map_attachment() at all.

I strongly vote for (b) (plus adding a dma_buf_map_sync interface to allow
syncing to other devices/cpu whithout dropping the dma mappings). At least
that's been the idea behind things, but currently all (x86-based) drm
drivers cut corners here massively.

Aside: The entire reason behind hiding the dma map step in the dma-buf
attachment is to allow drivers to expose crazy iommus (e.g. the tiler unit
on omap) to other devices. So imo (a) isn't the right choice.

That's why I also talk below about adding the dma_buf_map/sync callbacks
below, so that a dma_buf implementation can do whatever is necessary with
the sg at the point of use.

However, you are correct that this should be unnecessary, as DRM should
only be calling dma_buf_map_attachment() when it needs to know about the
memory behind the object.  The problem is that people will cache that
information - it also may be expensive to setup for the dma_buf - as it
involves counting the number of pages, memory allocation, and maybe
obtaining the set of pages from shmem.

With (a) plus the callbacks below, it means that step is only performed
once, and then the DMA API part is entirely separate - we can have our
cake and eat it in that regard.

quoted

Note: the existing stuff does have the nice side effect of being able
to pass buffers which do not have a struct page * associated with them
through the dma_buf API - I think we can still preserve that by having
dma_buf provide a couple of new APIs to do the SG list map/sync/unmap,
but in any case we need to fix the existing API so that:

dma_buf_map_attachment() becomes dma_buf_get_sg()
dma_buf_unmap_attachment() becomes dma_buf_put_sg()

both getting rid of the DMA direction argument, and then we have four
new dma_buf calls:

dma_buf_map_sg()
dma_buf_unmap_sg()
dma_buf_sync_sg_for_cpu()
dma_buf_sync_sg_for_device()

which do the actual sg map/unmap via the DMA API *at the appropriate
time for DMA*.

Hm, my idea was to just add a dma_buf_sync_attchment for the device side
syncing, since the cpu access stuff is already bracketed with the
begin/end cpu access stuff. We might need a sync_for_cpu or so for mmap,
but imo mmap support for dma_buf is a bit insane anyway, so I don't care
too much about it.

Since such dma mappings would be really longstanding in most cases anyway
drivers could just map with BIDIRECTIONAL and do all the real flushing
with the new sync stuff.

Note that the DMA API debug doesn't allow you to change the direction
argument on an existing mapping (neither should it, again this is
documented in the DMA API stuff in Documentation/).  This is where you
would need the complete set of four functions I mention above which
reflect the functionality of the DMA API.

The dma_buf implementation doesn't _have_ to implement them if they
are no-ops - for example, your dma_buf sg array contains DMA pointers,
and the memory is already coherent in some way (for example, it's a
separate chunk of memory which isn't part of system RAM.)

On Tue, Jun 18, 2013 at 12:46 PM, Russell King - ARM Linux
[off-list ref] wrote:

quoted

quoted

Note: the existing stuff does have the nice side effect of being able
to pass buffers which do not have a struct page * associated with them
through the dma_buf API - I think we can still preserve that by having
dma_buf provide a couple of new APIs to do the SG list map/sync/unmap,
but in any case we need to fix the existing API so that:

dma_buf_map_attachment() becomes dma_buf_get_sg()
dma_buf_unmap_attachment() becomes dma_buf_put_sg()

both getting rid of the DMA direction argument, and then we have four
new dma_buf calls:

dma_buf_map_sg()
dma_buf_unmap_sg()
dma_buf_sync_sg_for_cpu()
dma_buf_sync_sg_for_device()

which do the actual sg map/unmap via the DMA API *at the appropriate
time for DMA*.

Hm, my idea was to just add a dma_buf_sync_attchment for the device side
syncing, since the cpu access stuff is already bracketed with the
begin/end cpu access stuff. We might need a sync_for_cpu or so for mmap,
but imo mmap support for dma_buf is a bit insane anyway, so I don't care
too much about it.

Since such dma mappings would be really longstanding in most cases anyway
drivers could just map with BIDIRECTIONAL and do all the real flushing
with the new sync stuff.

Note that the DMA API debug doesn't allow you to change the direction
argument on an existing mapping (neither should it, again this is
documented in the DMA API stuff in Documentation/).  This is where you
would need the complete set of four functions I mention above which
reflect the functionality of the DMA API.

[Been travelling a bit, hence the delay.]

Just a quick question on your assertion that we need all four
functions: Since we already have begin/end_cpu_access functions
(intention here was to allow the dma_buf exporter to ensure the memory
is pinned, e.g. for swapable gem objects, but also allow cpu cache
flushing if required) do we still need the sync_sg_for_cpu? At least
with i915 as the exporter we currently hide the cflushing behind our
begin_cpu_access callback. For device dma we currently punt on it due
to lack of the dma_buf_sync_sg_for_device interface.

Aside: I know that i915 doing the clflushing dance itself is a bit
ugly, but thus far we've been the only guys on the x86 block with
non-coherent dma. But it sounds like a bunch of other blocks on Atom
SoCs have similar needs, so I guess it would make sense to move that
into the dma layer.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

On Tue, Jun 25, 2013 at 11:23:21AM +0200, Daniel Vetter wrote:

Just a quick question on your assertion that we need all four
functions: Since we already have begin/end_cpu_access functions
(intention here was to allow the dma_buf exporter to ensure the memory
is pinned, e.g. for swapable gem objects, but also allow cpu cache
flushing if required) do we still need the sync_sg_for_cpu?

Possibly not, but let's first look at that kind of scenario:

- It attaches to a dma_buf using dma_buf_attach()
- it maps the DMA buffer using dma_buf_map_attachment().  This does
  dma_map_sg().  This possibly incurs a cache flush depending on the
  coherence properties of the architecture.
- it then calls dma_buf_begin_cpu_access().  This presumably does a
  dma_sync_sg_for_cpu().  This possibly incurs another cache flush
  depending on the coherence properties of the architecture.
- then, presumably, it calls dma_buf_kmap_atomic() or dma_buf_kmap()
  to gain a pointer to the buffer.
- at some point, dma_buf_kunmap_atomic() or dma_buf_kunmap() gets
  called.  On certain cache architectures, kunmap_atomic() results in
  a cache flush.
- dma_buf_end_cpu_access() gets called, calling through presumably to
  dma_sync_sg_for_device().  Possibly incurs a cache flush.
- dma_buf_unmap_attachment()... another cache flush.

Out of those all of those five cache flushes, the only cache flush which is
really necessary out of all those would be the one in kunmap_atomic().  The
rest of them are completely irrelevant to the CPU access provided that there
is no DMA access by the device.  What's more is that you can't say "well,
the architecture should optimize them!" to which I'd respond "how does the
architecture know at dma_map_sg() time that there isn't going to be a
DMA access?"

Now, if we say that it's not necessary to call dma_buf_map_attachment()..
dma_buf_unmap_attachment() around the calls to dma_buf_begin_cpu_access()..
dma_buf_end_cpu_access(), then how does dma_buf_begin_cpu_access() know
whether the DMA buffer has been dma_map_sg()'d?  It's invalid to call
dma_sync_sg_for_cpu() on a buffer which has not been dma_map_sg()'d.
Bear in mind that dma_buf_begin_cpu_access() is passed only the
struct dma_buf and not the struct dma_buf_attachment *attach, there's
no hope for the begin_cpu_access() method to know which user of the
dma_buf is asking for this call, so you can't track any state there.

Thank you for pointing this out, I'm less impressed with this dma_buf
API now that I've looked deeper at those two interfaces, and even more
convinced it needs to be redesigned! :P