[v2,1/4] memcg: Track exported dma-buffers

Message ID 20230123191728.2928839-2-tjmercier@google.com (mailing list archive)
State Not Applicable
Headers
Series Track exported dma-buffers with memcg |

Commit Message

T.J. Mercier Jan. 23, 2023, 7:17 p.m. UTC
  When a buffer is exported to userspace, use memcg to attribute the
buffer to the allocating cgroup until all buffer references are
released.

Unlike the dmabuf sysfs stats implementation, this memcg accounting
avoids contention over the kernfs_rwsem incurred when creating or
removing nodes.

Signed-off-by: T.J. Mercier <tjmercier@google.com>
---
 Documentation/admin-guide/cgroup-v2.rst |  4 +++
 drivers/dma-buf/dma-buf.c               | 13 +++++++++
 include/linux/dma-buf.h                 |  3 ++
 include/linux/memcontrol.h              | 38 +++++++++++++++++++++++++
 mm/memcontrol.c                         | 19 +++++++++++++
 5 files changed, 77 insertions(+)
  

Comments

Michal Hocko Jan. 24, 2023, 2:59 p.m. UTC | #1
On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> When a buffer is exported to userspace, use memcg to attribute the
> buffer to the allocating cgroup until all buffer references are
> released.

Is there any reason why this memory cannot be charged during the
allocation (__GFP_ACCOUNT used)?
Also you do charge and account the memory but underlying pages do not
know about their memcg (this is normally done with commit_charge for
user mapped pages). This would become a problem if the memory is
migrated for example. This also means that you have to maintain memcg
reference outside of the memcg proper which is not really nice either.
This mimicks tcp kmem limit implementation which I really have to say I
am not a great fan of and this pattern shouldn't be coppied.

Also you are not really saying anything about the oom behavior. With
this implementation the kernel will try to reclaim the memory and even
trigger the memcg oom killer if the request size is <= 8 pages. Is this
a desirable behavior?
  
T.J. Mercier Jan. 24, 2023, 6:55 p.m. UTC | #2
On Tue, Jan 24, 2023 at 7:00 AM Michal Hocko <mhocko@suse.com> wrote:
>
> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> > When a buffer is exported to userspace, use memcg to attribute the
> > buffer to the allocating cgroup until all buffer references are
> > released.
>
> Is there any reason why this memory cannot be charged during the
> allocation (__GFP_ACCOUNT used)?

My main motivation was to keep code changes away from exporters and
implement the accounting in one common spot for all of them. This is a
bit of a carryover from a previous approach [1] where there was some
objection to pushing off this work onto exporters and forcing them to
adapt, but __GFP_ACCOUNT does seem like a smaller burden than before
at least initially. However in order to support charge transfer
between cgroups with __GFP_ACCOUNT we'd need to be able to get at the
pages backing dmabuf objects, and the exporters are the ones with that
access. Meaning I think we'd have to add some additional dma_buf_ops
to achieve that, which was the objection from [1].

[1] https://lore.kernel.org/lkml/5cc27a05-8131-ce9b-dea1-5c75e994216d@amd.com/

>
> Also you do charge and account the memory but underlying pages do not
> know about their memcg (this is normally done with commit_charge for
> user mapped pages). This would become a problem if the memory is
> migrated for example.

Hmm, what problem do you see in this situation? If the backing pages
are to be migrated that requires the cooperation of the exporter,
which currently has no influence on how the cgroup charging is done
and that seems fine. (Unless you mean migrating the charge across
cgroups? In which case that's the next patch.)

> This also means that you have to maintain memcg
> reference outside of the memcg proper which is not really nice either.
> This mimicks tcp kmem limit implementation which I really have to say I
> am not a great fan of and this pattern shouldn't be coppied.
>
Ah, what can I say. This way looked simple to me. I think otherwise
we're back to making all exporters do more stuff for the accounting.

> Also you are not really saying anything about the oom behavior. With
> this implementation the kernel will try to reclaim the memory and even
> trigger the memcg oom killer if the request size is <= 8 pages. Is this
> a desirable behavior?

It will try to reclaim some memory, but not the dmabuf pages right?
Not *yet* anyway. This behavior sounds expected to me. I would only
expect it to be surprising for cgroups making heavy use of dmabufs
(that weren't accounted before) *and* with hard limits already very
close to actual usage. I remember Johannes mentioning that what counts
under memcg use is already a bit of a moving target.

> --
> Michal Hocko
> SUSE Labs
  
Shakeel Butt Jan. 24, 2023, 7:46 p.m. UTC | #3
On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> > When a buffer is exported to userspace, use memcg to attribute the
> > buffer to the allocating cgroup until all buffer references are
> > released.
> 
> Is there any reason why this memory cannot be charged during the
> allocation (__GFP_ACCOUNT used)?
> Also you do charge and account the memory but underlying pages do not
> know about their memcg (this is normally done with commit_charge for
> user mapped pages). This would become a problem if the memory is
> migrated for example.

I don't think this is movable memory.

> This also means that you have to maintain memcg
> reference outside of the memcg proper which is not really nice either.
> This mimicks tcp kmem limit implementation which I really have to say I
> am not a great fan of and this pattern shouldn't be coppied.
> 

I think we should keep the discussion on technical merits instead of
personal perference. To me using skmem like interface is totally fine
but the pros/cons need to be very explicit and the clear reasons to
select that option should be included.

To me there are two options:

1. Using skmem like interface as this patch series:

The main pros of this option is that it is very simple. Let me list down
the cons of this approach:

a. There is time window between the actual memory allocation/free and
the charge and uncharge and [un]charge happen when the whole memory is
allocated or freed. I think for the charge path that might not be a big
issue but on the uncharge, this can cause issues. The application and
the potential shrinkers have freed some of this dmabuf memory but until
the whole dmabuf is freed, the memcg uncharge will not happen. This can
consequences on reclaim and oom behavior of the application.

b. Due to the usage model i.e. a central daemon allocating the dmabuf
memory upfront, there is a requirement to have a memcg charge transfer
functionality to transfer the charge from the central daemon to the
client applications. This does introduce complexity and avenues of weird
reclaim and oom behavior.


2. Allocate and charge the memory on page fault by actual user

In this approach, the memory is not allocated upfront by the central
daemon but rather on the page fault by the client application and the
memcg charge happen at the same time.

The only cons I can think of is this approach is more involved and may
need some clever tricks to track the page on the free patch i.e. we to
decrement the dmabuf memcg stat on free path. Maybe a page flag.

The pros of this approach is there is no need have a charge transfer
functionality and the charge/uncharge being closely tied to the actual
memory allocation and free.

Personally I would prefer the second approach but I don't want to just
block this work if the dmabuf folks are ok with the cons mentioned of
the first approach.

thanks,
Shakeel
  
Michal Hocko Jan. 25, 2023, 11:52 a.m. UTC | #4
On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> > On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> > > When a buffer is exported to userspace, use memcg to attribute the
> > > buffer to the allocating cgroup until all buffer references are
> > > released.
> > 
> > Is there any reason why this memory cannot be charged during the
> > allocation (__GFP_ACCOUNT used)?
> > Also you do charge and account the memory but underlying pages do not
> > know about their memcg (this is normally done with commit_charge for
> > user mapped pages). This would become a problem if the memory is
> > migrated for example.
> 
> I don't think this is movable memory.
> 
> > This also means that you have to maintain memcg
> > reference outside of the memcg proper which is not really nice either.
> > This mimicks tcp kmem limit implementation which I really have to say I
> > am not a great fan of and this pattern shouldn't be coppied.
> > 
> 
> I think we should keep the discussion on technical merits instead of
> personal perference. To me using skmem like interface is totally fine
> but the pros/cons need to be very explicit and the clear reasons to
> select that option should be included.

I do agree with that. I didn't want sound to be personal wrt tcp kmem
accounting but the overall code maintenance cost is higher because
of how tcp take on accounting differs from anything else in the memcg
proper. I would prefer to not grow another example like that.

> To me there are two options:
> 
> 1. Using skmem like interface as this patch series:
> 
> The main pros of this option is that it is very simple. Let me list down
> the cons of this approach:
> 
> a. There is time window between the actual memory allocation/free and
> the charge and uncharge and [un]charge happen when the whole memory is
> allocated or freed. I think for the charge path that might not be a big
> issue but on the uncharge, this can cause issues. The application and
> the potential shrinkers have freed some of this dmabuf memory but until
> the whole dmabuf is freed, the memcg uncharge will not happen. This can
> consequences on reclaim and oom behavior of the application.
> 
> b. Due to the usage model i.e. a central daemon allocating the dmabuf
> memory upfront, there is a requirement to have a memcg charge transfer
> functionality to transfer the charge from the central daemon to the
> client applications. This does introduce complexity and avenues of weird
> reclaim and oom behavior.
> 
> 
> 2. Allocate and charge the memory on page fault by actual user
> 
> In this approach, the memory is not allocated upfront by the central
> daemon but rather on the page fault by the client application and the
> memcg charge happen at the same time.
> 
> The only cons I can think of is this approach is more involved and may
> need some clever tricks to track the page on the free patch i.e. we to
> decrement the dmabuf memcg stat on free path. Maybe a page flag.
> 
> The pros of this approach is there is no need have a charge transfer
> functionality and the charge/uncharge being closely tied to the actual
> memory allocation and free.
> 
> Personally I would prefer the second approach but I don't want to just
> block this work if the dmabuf folks are ok with the cons mentioned of
> the first approach.

I am not familiar with dmabuf internals to judge complexity on their end
but I fully agree that charge-when-used is much more easier to reason
about and it should have less subtle surprises.
  
Michal Hocko Jan. 25, 2023, 12:05 p.m. UTC | #5
On Tue 24-01-23 10:55:21, T.J. Mercier wrote:
> On Tue, Jan 24, 2023 at 7:00 AM Michal Hocko <mhocko@suse.com> wrote:
> >
> > On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> > > When a buffer is exported to userspace, use memcg to attribute the
> > > buffer to the allocating cgroup until all buffer references are
> > > released.
> >
> > Is there any reason why this memory cannot be charged during the
> > allocation (__GFP_ACCOUNT used)?
> 
> My main motivation was to keep code changes away from exporters and
> implement the accounting in one common spot for all of them. This is a
> bit of a carryover from a previous approach [1] where there was some
> objection to pushing off this work onto exporters and forcing them to
> adapt, but __GFP_ACCOUNT does seem like a smaller burden than before
> at least initially. However in order to support charge transfer
> between cgroups with __GFP_ACCOUNT we'd need to be able to get at the
> pages backing dmabuf objects, and the exporters are the ones with that
> access. Meaning I think we'd have to add some additional dma_buf_ops
> to achieve that, which was the objection from [1].
> 
> [1] https://lore.kernel.org/lkml/5cc27a05-8131-ce9b-dea1-5c75e994216d@amd.com/
> 
> >
> > Also you do charge and account the memory but underlying pages do not
> > know about their memcg (this is normally done with commit_charge for
> > user mapped pages). This would become a problem if the memory is
> > migrated for example.
> 
> Hmm, what problem do you see in this situation? If the backing pages
> are to be migrated that requires the cooperation of the exporter,
> which currently has no influence on how the cgroup charging is done
> and that seems fine. (Unless you mean migrating the charge across
> cgroups? In which case that's the next patch.)

My main concern was that page migration could lose the external tracking
without some additional steps on the dmabuf front.

> > This also means that you have to maintain memcg
> > reference outside of the memcg proper which is not really nice either.
> > This mimicks tcp kmem limit implementation which I really have to say I
> > am not a great fan of and this pattern shouldn't be coppied.
> >
> Ah, what can I say. This way looked simple to me. I think otherwise
> we're back to making all exporters do more stuff for the accounting.
> 
> > Also you are not really saying anything about the oom behavior. With
> > this implementation the kernel will try to reclaim the memory and even
> > trigger the memcg oom killer if the request size is <= 8 pages. Is this
> > a desirable behavior?
> 
> It will try to reclaim some memory, but not the dmabuf pages right?
> Not *yet* anyway. This behavior sounds expected to me.

Yes, we have discussed that shrinkers will follow up later which is
fine. The question is how much reclaim actually makes sense at this
stage. Charging interface usually copes with sizes resulting from
allocation requests (so usually 1<<order based). I can imagine that a
batch charge like implemented here could easily be 100s of MBs and it is
much harder to define reclaim targets for. At least that is something
the memcg charging hasn't really considered yet.  Maybe the existing
try_charge implementation can cope with that just fine but it would be
really great to have the expected behavior described.

E.g. should be memcg OOM killer be invoked? Should reclaim really target
regular memory at all costs or just a lightweight memory reclaim is
preferred (is the dmabuf charge failure an expensive operation wrt.
memory refault due to reclaim).
  
Tvrtko Ursulin Jan. 25, 2023, 5:30 p.m. UTC | #6
Hi,

On 25/01/2023 11:52, Michal Hocko wrote:
> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>> buffer to the allocating cgroup until all buffer references are
>>>> released.
>>>
>>> Is there any reason why this memory cannot be charged during the
>>> allocation (__GFP_ACCOUNT used)?
>>> Also you do charge and account the memory but underlying pages do not
>>> know about their memcg (this is normally done with commit_charge for
>>> user mapped pages). This would become a problem if the memory is
>>> migrated for example.
>>
>> I don't think this is movable memory.
>>
>>> This also means that you have to maintain memcg
>>> reference outside of the memcg proper which is not really nice either.
>>> This mimicks tcp kmem limit implementation which I really have to say I
>>> am not a great fan of and this pattern shouldn't be coppied.
>>>
>>
>> I think we should keep the discussion on technical merits instead of
>> personal perference. To me using skmem like interface is totally fine
>> but the pros/cons need to be very explicit and the clear reasons to
>> select that option should be included.
> 
> I do agree with that. I didn't want sound to be personal wrt tcp kmem
> accounting but the overall code maintenance cost is higher because
> of how tcp take on accounting differs from anything else in the memcg
> proper. I would prefer to not grow another example like that.
> 
>> To me there are two options:
>>
>> 1. Using skmem like interface as this patch series:
>>
>> The main pros of this option is that it is very simple. Let me list down
>> the cons of this approach:
>>
>> a. There is time window between the actual memory allocation/free and
>> the charge and uncharge and [un]charge happen when the whole memory is
>> allocated or freed. I think for the charge path that might not be a big
>> issue but on the uncharge, this can cause issues. The application and
>> the potential shrinkers have freed some of this dmabuf memory but until
>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
>> consequences on reclaim and oom behavior of the application.
>>
>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
>> memory upfront, there is a requirement to have a memcg charge transfer
>> functionality to transfer the charge from the central daemon to the
>> client applications. This does introduce complexity and avenues of weird
>> reclaim and oom behavior.
>>
>>
>> 2. Allocate and charge the memory on page fault by actual user
>>
>> In this approach, the memory is not allocated upfront by the central
>> daemon but rather on the page fault by the client application and the
>> memcg charge happen at the same time.
>>
>> The only cons I can think of is this approach is more involved and may
>> need some clever tricks to track the page on the free patch i.e. we to
>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>
>> The pros of this approach is there is no need have a charge transfer
>> functionality and the charge/uncharge being closely tied to the actual
>> memory allocation and free.
>>
>> Personally I would prefer the second approach but I don't want to just
>> block this work if the dmabuf folks are ok with the cons mentioned of
>> the first approach.
> 
> I am not familiar with dmabuf internals to judge complexity on their end
> but I fully agree that charge-when-used is much more easier to reason
> about and it should have less subtle surprises.

Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I 
am missing something, but in principle yes, I agree that the 2nd option 
(charge the user, not exporter) should be preferred. Thing being that at 
export time there may not be any backing store allocated, plus if the 
series is restricting the charge transfer to just Android clients then 
it seems it has the potential to miss many other use cases. At least 
needs to outline a description on how the feature will be useful outside 
Android.

Also stepping back for a moment - is a new memory category really 
needed, versus perhaps attempting to charge the actual backing store 
memory to the correct client? (There might have been many past 
discussions on this so it's okay to point me towards something in the 
archives.)

Regards,

Tvrtko
  
T.J. Mercier Jan. 25, 2023, 8:04 p.m. UTC | #7
On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> Hi,
>
> On 25/01/2023 11:52, Michal Hocko wrote:
> > On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
> >> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> >>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> >>>> When a buffer is exported to userspace, use memcg to attribute the
> >>>> buffer to the allocating cgroup until all buffer references are
> >>>> released.
> >>>
> >>> Is there any reason why this memory cannot be charged during the
> >>> allocation (__GFP_ACCOUNT used)?
> >>> Also you do charge and account the memory but underlying pages do not
> >>> know about their memcg (this is normally done with commit_charge for
> >>> user mapped pages). This would become a problem if the memory is
> >>> migrated for example.
> >>
> >> I don't think this is movable memory.
> >>
> >>> This also means that you have to maintain memcg
> >>> reference outside of the memcg proper which is not really nice either.
> >>> This mimicks tcp kmem limit implementation which I really have to say I
> >>> am not a great fan of and this pattern shouldn't be coppied.
> >>>
> >>
> >> I think we should keep the discussion on technical merits instead of
> >> personal perference. To me using skmem like interface is totally fine
> >> but the pros/cons need to be very explicit and the clear reasons to
> >> select that option should be included.
> >
> > I do agree with that. I didn't want sound to be personal wrt tcp kmem
> > accounting but the overall code maintenance cost is higher because
> > of how tcp take on accounting differs from anything else in the memcg
> > proper. I would prefer to not grow another example like that.
> >
> >> To me there are two options:
> >>
> >> 1. Using skmem like interface as this patch series:
> >>
> >> The main pros of this option is that it is very simple. Let me list down
> >> the cons of this approach:
> >>
> >> a. There is time window between the actual memory allocation/free and
> >> the charge and uncharge and [un]charge happen when the whole memory is
> >> allocated or freed. I think for the charge path that might not be a big
> >> issue but on the uncharge, this can cause issues. The application and
> >> the potential shrinkers have freed some of this dmabuf memory but until
> >> the whole dmabuf is freed, the memcg uncharge will not happen. This can
> >> consequences on reclaim and oom behavior of the application.
> >>
> >> b. Due to the usage model i.e. a central daemon allocating the dmabuf
> >> memory upfront, there is a requirement to have a memcg charge transfer
> >> functionality to transfer the charge from the central daemon to the
> >> client applications. This does introduce complexity and avenues of weird
> >> reclaim and oom behavior.
> >>
> >>
> >> 2. Allocate and charge the memory on page fault by actual user
> >>
> >> In this approach, the memory is not allocated upfront by the central
> >> daemon but rather on the page fault by the client application and the
> >> memcg charge happen at the same time.
> >>
> >> The only cons I can think of is this approach is more involved and may
> >> need some clever tricks to track the page on the free patch i.e. we to
> >> decrement the dmabuf memcg stat on free path. Maybe a page flag.
> >>
> >> The pros of this approach is there is no need have a charge transfer
> >> functionality and the charge/uncharge being closely tied to the actual
> >> memory allocation and free.
> >>
> >> Personally I would prefer the second approach but I don't want to just
> >> block this work if the dmabuf folks are ok with the cons mentioned of
> >> the first approach.
> >
> > I am not familiar with dmabuf internals to judge complexity on their end
> > but I fully agree that charge-when-used is much more easier to reason
> > about and it should have less subtle surprises.
>
> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
> am missing something, but in principle yes, I agree that the 2nd option
> (charge the user, not exporter) should be preferred. Thing being that at
> export time there may not be any backing store allocated, plus if the
> series is restricting the charge transfer to just Android clients then
> it seems it has the potential to miss many other use cases. At least
> needs to outline a description on how the feature will be useful outside
> Android.
>
There is no restriction like that. It's available to anybody who wants
to call dma_buf_charge_transfer if they actually have a need for that,
which I don't really expect to be common since most users/owners of
the buffers will be the ones causing the export in the first place.
It's just not like that on Android with the extra allocator process in
the middle most of the time.

> Also stepping back for a moment - is a new memory category really
> needed, versus perhaps attempting to charge the actual backing store
> memory to the correct client? (There might have been many past
> discussions on this so it's okay to point me towards something in the
> archives.)
>
Well the dmabuf counter for the stat file is really just a subcategory
of memory that is charged. Its existence is not related to getting the
charge attributed to the right process/cgroup. We do want to know how
much of the memory attributed to a process is for dmabufs, which is
the main point of this series.

> Regards,
>
> Tvrtko
  
T.J. Mercier Jan. 25, 2023, 8:04 p.m. UTC | #8
On Wed, Jan 25, 2023 at 4:05 AM Michal Hocko <mhocko@suse.com> wrote:
>
> On Tue 24-01-23 10:55:21, T.J. Mercier wrote:
> > On Tue, Jan 24, 2023 at 7:00 AM Michal Hocko <mhocko@suse.com> wrote:
> > >
> > > On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> > > > When a buffer is exported to userspace, use memcg to attribute the
> > > > buffer to the allocating cgroup until all buffer references are
> > > > released.
> > >
> > > Is there any reason why this memory cannot be charged during the
> > > allocation (__GFP_ACCOUNT used)?
> >
> > My main motivation was to keep code changes away from exporters and
> > implement the accounting in one common spot for all of them. This is a
> > bit of a carryover from a previous approach [1] where there was some
> > objection to pushing off this work onto exporters and forcing them to
> > adapt, but __GFP_ACCOUNT does seem like a smaller burden than before
> > at least initially. However in order to support charge transfer
> > between cgroups with __GFP_ACCOUNT we'd need to be able to get at the
> > pages backing dmabuf objects, and the exporters are the ones with that
> > access. Meaning I think we'd have to add some additional dma_buf_ops
> > to achieve that, which was the objection from [1].
> >
> > [1] https://lore.kernel.org/lkml/5cc27a05-8131-ce9b-dea1-5c75e994216d@amd.com/
> >
> > >
> > > Also you do charge and account the memory but underlying pages do not
> > > know about their memcg (this is normally done with commit_charge for
> > > user mapped pages). This would become a problem if the memory is
> > > migrated for example.
> >
> > Hmm, what problem do you see in this situation? If the backing pages
> > are to be migrated that requires the cooperation of the exporter,
> > which currently has no influence on how the cgroup charging is done
> > and that seems fine. (Unless you mean migrating the charge across
> > cgroups? In which case that's the next patch.)
>
> My main concern was that page migration could lose the external tracking
> without some additional steps on the dmabuf front.
>
I see, yes that would be true if an exporter moves data around between
system memory and VRAM for example. (I think TTM does this sort of
thing, but not sure if that's actually within a single dma buffer.)
VRAM feels like it maybe doesn't belong in memcg, yet it would still
be charged there under this series right now. I don't really see a way
around this except to involve the exporters directly in the accounting
(or don't attempt to distinguish between types of memory).

> > > This also means that you have to maintain memcg
> > > reference outside of the memcg proper which is not really nice either.
> > > This mimicks tcp kmem limit implementation which I really have to say I
> > > am not a great fan of and this pattern shouldn't be coppied.
> > >
> > Ah, what can I say. This way looked simple to me. I think otherwise
> > we're back to making all exporters do more stuff for the accounting.
> >
> > > Also you are not really saying anything about the oom behavior. With
> > > this implementation the kernel will try to reclaim the memory and even
> > > trigger the memcg oom killer if the request size is <= 8 pages. Is this
> > > a desirable behavior?
> >
> > It will try to reclaim some memory, but not the dmabuf pages right?
> > Not *yet* anyway. This behavior sounds expected to me.
>
> Yes, we have discussed that shrinkers will follow up later which is
> fine. The question is how much reclaim actually makes sense at this
> stage. Charging interface usually copes with sizes resulting from
> allocation requests (so usually 1<<order based). I can imagine that a
> batch charge like implemented here could easily be 100s of MBs and it is
> much harder to define reclaim targets for. At least that is something
> the memcg charging hasn't really considered yet.  Maybe the existing
> try_charge implementation can cope with that just fine but it would be
> really great to have the expected behavior described.
>
> E.g. should be memcg OOM killer be invoked? Should reclaim really target
> regular memory at all costs or just a lightweight memory reclaim is
> preferred (is the dmabuf charge failure an expensive operation wrt.
> memory refault due to reclaim).

Ah, in my experience very large individual buffers like that are rare.
Cumulative system-wide usage might reach 100s of megs or more spread
across many buffers. On my phone the majority of buffer sizes are 4
pages or less, but there are a few that reach into the tens of megs.
But now I see your point. I still think that where a memcg limit is
exceeded and we can't reclaim enough as a result of a new dmabuf
allocation, we should see a memcg OOM kill. Sounds like you are
looking for that to be written down, so I'll try to find a place for
that.

Part of the motivation for this accounting is to eventually have a
well defined limit for applications to know how much more they can
allocate. So where buffer size or number of buffers is a flexible
variable, I'd like to see an application checking this limit before
making a large request in an effort to avoid reclaim in the first
place. Where there is heavy memory pressure and multiple competing
apps, the status-quo today is a kill for us anyways (typically LMKD).




> --
> Michal Hocko
> SUSE Labs
  
Tvrtko Ursulin Jan. 31, 2023, 2 p.m. UTC | #9
On 25/01/2023 20:04, T.J. Mercier wrote:
> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>>
>>
>> Hi,
>>
>> On 25/01/2023 11:52, Michal Hocko wrote:
>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>>>> buffer to the allocating cgroup until all buffer references are
>>>>>> released.
>>>>>
>>>>> Is there any reason why this memory cannot be charged during the
>>>>> allocation (__GFP_ACCOUNT used)?
>>>>> Also you do charge and account the memory but underlying pages do not
>>>>> know about their memcg (this is normally done with commit_charge for
>>>>> user mapped pages). This would become a problem if the memory is
>>>>> migrated for example.
>>>>
>>>> I don't think this is movable memory.
>>>>
>>>>> This also means that you have to maintain memcg
>>>>> reference outside of the memcg proper which is not really nice either.
>>>>> This mimicks tcp kmem limit implementation which I really have to say I
>>>>> am not a great fan of and this pattern shouldn't be coppied.
>>>>>
>>>>
>>>> I think we should keep the discussion on technical merits instead of
>>>> personal perference. To me using skmem like interface is totally fine
>>>> but the pros/cons need to be very explicit and the clear reasons to
>>>> select that option should be included.
>>>
>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
>>> accounting but the overall code maintenance cost is higher because
>>> of how tcp take on accounting differs from anything else in the memcg
>>> proper. I would prefer to not grow another example like that.
>>>
>>>> To me there are two options:
>>>>
>>>> 1. Using skmem like interface as this patch series:
>>>>
>>>> The main pros of this option is that it is very simple. Let me list down
>>>> the cons of this approach:
>>>>
>>>> a. There is time window between the actual memory allocation/free and
>>>> the charge and uncharge and [un]charge happen when the whole memory is
>>>> allocated or freed. I think for the charge path that might not be a big
>>>> issue but on the uncharge, this can cause issues. The application and
>>>> the potential shrinkers have freed some of this dmabuf memory but until
>>>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
>>>> consequences on reclaim and oom behavior of the application.
>>>>
>>>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
>>>> memory upfront, there is a requirement to have a memcg charge transfer
>>>> functionality to transfer the charge from the central daemon to the
>>>> client applications. This does introduce complexity and avenues of weird
>>>> reclaim and oom behavior.
>>>>
>>>>
>>>> 2. Allocate and charge the memory on page fault by actual user
>>>>
>>>> In this approach, the memory is not allocated upfront by the central
>>>> daemon but rather on the page fault by the client application and the
>>>> memcg charge happen at the same time.
>>>>
>>>> The only cons I can think of is this approach is more involved and may
>>>> need some clever tricks to track the page on the free patch i.e. we to
>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>>>
>>>> The pros of this approach is there is no need have a charge transfer
>>>> functionality and the charge/uncharge being closely tied to the actual
>>>> memory allocation and free.
>>>>
>>>> Personally I would prefer the second approach but I don't want to just
>>>> block this work if the dmabuf folks are ok with the cons mentioned of
>>>> the first approach.
>>>
>>> I am not familiar with dmabuf internals to judge complexity on their end
>>> but I fully agree that charge-when-used is much more easier to reason
>>> about and it should have less subtle surprises.
>>
>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
>> am missing something, but in principle yes, I agree that the 2nd option
>> (charge the user, not exporter) should be preferred. Thing being that at
>> export time there may not be any backing store allocated, plus if the
>> series is restricting the charge transfer to just Android clients then
>> it seems it has the potential to miss many other use cases. At least
>> needs to outline a description on how the feature will be useful outside
>> Android.
>>
> There is no restriction like that. It's available to anybody who wants
> to call dma_buf_charge_transfer if they actually have a need for that,
> which I don't really expect to be common since most users/owners of
> the buffers will be the ones causing the export in the first place.
> It's just not like that on Android with the extra allocator process in
> the middle most of the time.

Yeah I used the wrong term "restrict", apologies. What I meant was, if 
the idea was to allow spotting memory leaks, with the charge transfer 
being optional and in the series only wired up for Android Binder, then 
it obviously only fully works for that one case. So a step back..

.. For instance, it is not feasible to transfer the charge when dmabuf 
is attached, or imported? That would attribute the usage to the 
user/importer so give better visibility on who is actually causing the 
memory leak.

Further more, if above is feasible, then could it also be implemented in 
the common layer so it would automatically cover all drivers?

>> Also stepping back for a moment - is a new memory category really
>> needed, versus perhaps attempting to charge the actual backing store
>> memory to the correct client? (There might have been many past
>> discussions on this so it's okay to point me towards something in the
>> archives.)
>>
> Well the dmabuf counter for the stat file is really just a subcategory
> of memory that is charged. Its existence is not related to getting the
> charge attributed to the right process/cgroup. We do want to know how
> much of the memory attributed to a process is for dmabufs, which is
> the main point of this series.

Then I am probably missing something because the statement how proposal 
is not intended to charge to the right process, but wants to know how 
much dmabuf "size" is attributed to a process, confuses me due a seeming 
contradiction. And the fact it would not be externally observable how 
much of the stats is accurate and how much is not (without knowing the 
implementation detail of which drivers implement charge transfer and 
when). Maybe I completely misunderstood the use case.

Regards,

Tvrtko
  
T.J. Mercier Feb. 1, 2023, 1:49 a.m. UTC | #10
On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 25/01/2023 20:04, T.J. Mercier wrote:
> > On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
> > <tvrtko.ursulin@linux.intel.com> wrote:
> >>
> >>
> >> Hi,
> >>
> >> On 25/01/2023 11:52, Michal Hocko wrote:
> >>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
> >>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> >>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> >>>>>> When a buffer is exported to userspace, use memcg to attribute the
> >>>>>> buffer to the allocating cgroup until all buffer references are
> >>>>>> released.
> >>>>>
> >>>>> Is there any reason why this memory cannot be charged during the
> >>>>> allocation (__GFP_ACCOUNT used)?
> >>>>> Also you do charge and account the memory but underlying pages do not
> >>>>> know about their memcg (this is normally done with commit_charge for
> >>>>> user mapped pages). This would become a problem if the memory is
> >>>>> migrated for example.
> >>>>
> >>>> I don't think this is movable memory.
> >>>>
> >>>>> This also means that you have to maintain memcg
> >>>>> reference outside of the memcg proper which is not really nice either.
> >>>>> This mimicks tcp kmem limit implementation which I really have to say I
> >>>>> am not a great fan of and this pattern shouldn't be coppied.
> >>>>>
> >>>>
> >>>> I think we should keep the discussion on technical merits instead of
> >>>> personal perference. To me using skmem like interface is totally fine
> >>>> but the pros/cons need to be very explicit and the clear reasons to
> >>>> select that option should be included.
> >>>
> >>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
> >>> accounting but the overall code maintenance cost is higher because
> >>> of how tcp take on accounting differs from anything else in the memcg
> >>> proper. I would prefer to not grow another example like that.
> >>>
> >>>> To me there are two options:
> >>>>
> >>>> 1. Using skmem like interface as this patch series:
> >>>>
> >>>> The main pros of this option is that it is very simple. Let me list down
> >>>> the cons of this approach:
> >>>>
> >>>> a. There is time window between the actual memory allocation/free and
> >>>> the charge and uncharge and [un]charge happen when the whole memory is
> >>>> allocated or freed. I think for the charge path that might not be a big
> >>>> issue but on the uncharge, this can cause issues. The application and
> >>>> the potential shrinkers have freed some of this dmabuf memory but until
> >>>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
> >>>> consequences on reclaim and oom behavior of the application.
> >>>>
> >>>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
> >>>> memory upfront, there is a requirement to have a memcg charge transfer
> >>>> functionality to transfer the charge from the central daemon to the
> >>>> client applications. This does introduce complexity and avenues of weird
> >>>> reclaim and oom behavior.
> >>>>
> >>>>
> >>>> 2. Allocate and charge the memory on page fault by actual user
> >>>>
> >>>> In this approach, the memory is not allocated upfront by the central
> >>>> daemon but rather on the page fault by the client application and the
> >>>> memcg charge happen at the same time.
> >>>>
> >>>> The only cons I can think of is this approach is more involved and may
> >>>> need some clever tricks to track the page on the free patch i.e. we to
> >>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
> >>>>
> >>>> The pros of this approach is there is no need have a charge transfer
> >>>> functionality and the charge/uncharge being closely tied to the actual
> >>>> memory allocation and free.
> >>>>
> >>>> Personally I would prefer the second approach but I don't want to just
> >>>> block this work if the dmabuf folks are ok with the cons mentioned of
> >>>> the first approach.
> >>>
> >>> I am not familiar with dmabuf internals to judge complexity on their end
> >>> but I fully agree that charge-when-used is much more easier to reason
> >>> about and it should have less subtle surprises.
> >>
> >> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
> >> am missing something, but in principle yes, I agree that the 2nd option
> >> (charge the user, not exporter) should be preferred. Thing being that at
> >> export time there may not be any backing store allocated, plus if the
> >> series is restricting the charge transfer to just Android clients then
> >> it seems it has the potential to miss many other use cases. At least
> >> needs to outline a description on how the feature will be useful outside
> >> Android.
> >>
> > There is no restriction like that. It's available to anybody who wants
> > to call dma_buf_charge_transfer if they actually have a need for that,
> > which I don't really expect to be common since most users/owners of
> > the buffers will be the ones causing the export in the first place.
> > It's just not like that on Android with the extra allocator process in
> > the middle most of the time.
>
> Yeah I used the wrong term "restrict", apologies. What I meant was, if
> the idea was to allow spotting memory leaks, with the charge transfer
> being optional and in the series only wired up for Android Binder, then
> it obviously only fully works for that one case. So a step back..
>
Oh, spotting kernel memory leaks is a side-benefit of accounting
kernel-only buffers in the root cgroup. The primary goal is to
attribute buffers to applications that originated them (via
per-application cgroups) simply for accounting purposes. Buffers are
using memory on the system, and we want to know who created them and
how much memory is used. That information is/will no longer available
with the recent deprecation of the dmabuf sysfs statistics.

> .. For instance, it is not feasible to transfer the charge when dmabuf
> is attached, or imported? That would attribute the usage to the
> user/importer so give better visibility on who is actually causing the
> memory leak.
>
Instead of accounting at export, we could account at attach. That just
turns out not to be very useful when the majority of our
heap-allocated buffers don't have attachments at any particular point
in time. :\ But again it's less about leaks and more about knowing
which buffers exist in the first place.

> Further more, if above is feasible, then could it also be implemented in
> the common layer so it would automatically cover all drivers?
>
Which common layer code specifically? The dmabuf interface appears to
be the most central/common place to me.

> >> Also stepping back for a moment - is a new memory category really
> >> needed, versus perhaps attempting to charge the actual backing store
> >> memory to the correct client? (There might have been many past
> >> discussions on this so it's okay to point me towards something in the
> >> archives.)
> >>
> > Well the dmabuf counter for the stat file is really just a subcategory
> > of memory that is charged. Its existence is not related to getting the
> > charge attributed to the right process/cgroup. We do want to know how
> > much of the memory attributed to a process is for dmabufs, which is
> > the main point of this series.
>
> Then I am probably missing something because the statement how proposal
> is not intended to charge to the right process, but wants to know how
> much dmabuf "size" is attributed to a process, confuses me due a seeming
> contradiction. And the fact it would not be externally observable how
> much of the stats is accurate and how much is not (without knowing the
> implementation detail of which drivers implement charge transfer and
> when). Maybe I completely misunderstood the use case.
>
Hmm, did I clear this up above or no? The current proposal is for the
process causing the export of a buffer to be charged for it,
regardless of whatever happens afterwards. (Unless that process is
like gralloc on Android, in which case the charge is transferred from
gralloc to whoever called gralloc to allocate the buffer on their
behalf.)

> Regards,
>
> Tvrtko
  
Tvrtko Ursulin Feb. 1, 2023, 2:23 p.m. UTC | #11
On 01/02/2023 01:49, T.J. Mercier wrote:
> On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>>
>>
>> On 25/01/2023 20:04, T.J. Mercier wrote:
>>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>
>>>>
>>>> Hi,
>>>>
>>>> On 25/01/2023 11:52, Michal Hocko wrote:
>>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>>>>>> buffer to the allocating cgroup until all buffer references are
>>>>>>>> released.
>>>>>>>
>>>>>>> Is there any reason why this memory cannot be charged during the
>>>>>>> allocation (__GFP_ACCOUNT used)?
>>>>>>> Also you do charge and account the memory but underlying pages do not
>>>>>>> know about their memcg (this is normally done with commit_charge for
>>>>>>> user mapped pages). This would become a problem if the memory is
>>>>>>> migrated for example.
>>>>>>
>>>>>> I don't think this is movable memory.
>>>>>>
>>>>>>> This also means that you have to maintain memcg
>>>>>>> reference outside of the memcg proper which is not really nice either.
>>>>>>> This mimicks tcp kmem limit implementation which I really have to say I
>>>>>>> am not a great fan of and this pattern shouldn't be coppied.
>>>>>>>
>>>>>>
>>>>>> I think we should keep the discussion on technical merits instead of
>>>>>> personal perference. To me using skmem like interface is totally fine
>>>>>> but the pros/cons need to be very explicit and the clear reasons to
>>>>>> select that option should be included.
>>>>>
>>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
>>>>> accounting but the overall code maintenance cost is higher because
>>>>> of how tcp take on accounting differs from anything else in the memcg
>>>>> proper. I would prefer to not grow another example like that.
>>>>>
>>>>>> To me there are two options:
>>>>>>
>>>>>> 1. Using skmem like interface as this patch series:
>>>>>>
>>>>>> The main pros of this option is that it is very simple. Let me list down
>>>>>> the cons of this approach:
>>>>>>
>>>>>> a. There is time window between the actual memory allocation/free and
>>>>>> the charge and uncharge and [un]charge happen when the whole memory is
>>>>>> allocated or freed. I think for the charge path that might not be a big
>>>>>> issue but on the uncharge, this can cause issues. The application and
>>>>>> the potential shrinkers have freed some of this dmabuf memory but until
>>>>>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
>>>>>> consequences on reclaim and oom behavior of the application.
>>>>>>
>>>>>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
>>>>>> memory upfront, there is a requirement to have a memcg charge transfer
>>>>>> functionality to transfer the charge from the central daemon to the
>>>>>> client applications. This does introduce complexity and avenues of weird
>>>>>> reclaim and oom behavior.
>>>>>>
>>>>>>
>>>>>> 2. Allocate and charge the memory on page fault by actual user
>>>>>>
>>>>>> In this approach, the memory is not allocated upfront by the central
>>>>>> daemon but rather on the page fault by the client application and the
>>>>>> memcg charge happen at the same time.
>>>>>>
>>>>>> The only cons I can think of is this approach is more involved and may
>>>>>> need some clever tricks to track the page on the free patch i.e. we to
>>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>>>>>
>>>>>> The pros of this approach is there is no need have a charge transfer
>>>>>> functionality and the charge/uncharge being closely tied to the actual
>>>>>> memory allocation and free.
>>>>>>
>>>>>> Personally I would prefer the second approach but I don't want to just
>>>>>> block this work if the dmabuf folks are ok with the cons mentioned of
>>>>>> the first approach.
>>>>>
>>>>> I am not familiar with dmabuf internals to judge complexity on their end
>>>>> but I fully agree that charge-when-used is much more easier to reason
>>>>> about and it should have less subtle surprises.
>>>>
>>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
>>>> am missing something, but in principle yes, I agree that the 2nd option
>>>> (charge the user, not exporter) should be preferred. Thing being that at
>>>> export time there may not be any backing store allocated, plus if the
>>>> series is restricting the charge transfer to just Android clients then
>>>> it seems it has the potential to miss many other use cases. At least
>>>> needs to outline a description on how the feature will be useful outside
>>>> Android.
>>>>
>>> There is no restriction like that. It's available to anybody who wants
>>> to call dma_buf_charge_transfer if they actually have a need for that,
>>> which I don't really expect to be common since most users/owners of
>>> the buffers will be the ones causing the export in the first place.
>>> It's just not like that on Android with the extra allocator process in
>>> the middle most of the time.
>>
>> Yeah I used the wrong term "restrict", apologies. What I meant was, if
>> the idea was to allow spotting memory leaks, with the charge transfer
>> being optional and in the series only wired up for Android Binder, then
>> it obviously only fully works for that one case. So a step back..
>>
> Oh, spotting kernel memory leaks is a side-benefit of accounting
> kernel-only buffers in the root cgroup. The primary goal is to
> attribute buffers to applications that originated them (via
> per-application cgroups) simply for accounting purposes. Buffers are
> using memory on the system, and we want to know who created them and
> how much memory is used. That information is/will no longer available
> with the recent deprecation of the dmabuf sysfs statistics.
> 
>> .. For instance, it is not feasible to transfer the charge when dmabuf
>> is attached, or imported? That would attribute the usage to the
>> user/importer so give better visibility on who is actually causing the
>> memory leak.
>>
> Instead of accounting at export, we could account at attach. That just
> turns out not to be very useful when the majority of our
> heap-allocated buffers don't have attachments at any particular point
> in time. :\ But again it's less about leaks and more about knowing
> which buffers exist in the first place.
> 
>> Further more, if above is feasible, then could it also be implemented in
>> the common layer so it would automatically cover all drivers?
>>
> Which common layer code specifically? The dmabuf interface appears to
> be the most central/common place to me.

Yes, I meant dma_buf_attach / detach. More below.
>>>> Also stepping back for a moment - is a new memory category really
>>>> needed, versus perhaps attempting to charge the actual backing store
>>>> memory to the correct client? (There might have been many past
>>>> discussions on this so it's okay to point me towards something in the
>>>> archives.)
>>>>
>>> Well the dmabuf counter for the stat file is really just a subcategory
>>> of memory that is charged. Its existence is not related to getting the
>>> charge attributed to the right process/cgroup. We do want to know how
>>> much of the memory attributed to a process is for dmabufs, which is
>>> the main point of this series.
>>
>> Then I am probably missing something because the statement how proposal
>> is not intended to charge to the right process, but wants to know how
>> much dmabuf "size" is attributed to a process, confuses me due a seeming
>> contradiction. And the fact it would not be externally observable how
>> much of the stats is accurate and how much is not (without knowing the
>> implementation detail of which drivers implement charge transfer and
>> when). Maybe I completely misunderstood the use case.
>>
> Hmm, did I clear this up above or no? The current proposal is for the
> process causing the export of a buffer to be charged for it,
> regardless of whatever happens afterwards. (Unless that process is
> like gralloc on Android, in which case the charge is transferred from
> gralloc to whoever called gralloc to allocate the buffer on their
> behalf.)

Main problem for me is that charging at export time has no relation to memory used. But I am not familiar with the memcg counters to know if any other counter sets that same precedent. If all other are about real memory use then IMO this does not fit that well. I mean specifically this:

+	  dmabuf (npn)
+		Amount of memory used for exported DMA buffers allocated by the cgroup.
+		Stays with the allocating cgroup regardless of how the buffer is shared.
+

I think that "Amount of memory used for exported..." is not correct. As implemented it is more akin the virtual address space size in the cpu space - it can have no relation to the actual usage since backing store is not allocated until the attachment is made.

Then also this:

@@ -446,6 +447,8 @@ struct dma_buf {
  		struct dma_buf *dmabuf;
  	} *sysfs_entry;
  #endif
+	/* The cgroup to which this buffer is currently attributed */
+	struct mem_cgroup *memcg;
  };

Does not conceptually fit in my mind. Dmabufs are not associated with one cgroup at a time.

So if you would place tracking into dma_buf_attach/detach you would be able to charge to correct cgroup regardless of a driver and since by contract at this stage there is backing store, the reflected memory usage counter would be truthful.

But then you state a problem, that majority of the time there are no attachments in your setup, and you also say the proposal is not so much about leaks but more about knowing what is exported.

In this case you could additionally track that via dma_buf_getfile / dma_buf_file_release as a separate category like dmabuf-exported? But again, I personally don't know if such "may not really be using memory" counters fit in memcg.

(Hm you'd probably still need dmabuf->export_memcg to store who was the original caller of dma_buf_getfile, in case last reference is dropped from a different process/context. Even dmabuf->attach_memcg for attach/detach to work correctly for the same reason.)

Regards,

Tvrtko
  
Tvrtko Ursulin Feb. 1, 2023, 2:52 p.m. UTC | #12
On 01/02/2023 14:23, Tvrtko Ursulin wrote:
> 
> On 01/02/2023 01:49, T.J. Mercier wrote:
>> On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>
>>>
>>> On 25/01/2023 20:04, T.J. Mercier wrote:
>>>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
>>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>>
>>>>>
>>>>> Hi,
>>>>>
>>>>> On 25/01/2023 11:52, Michal Hocko wrote:
>>>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>>>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>>>>>>> buffer to the allocating cgroup until all buffer references are
>>>>>>>>> released.
>>>>>>>>
>>>>>>>> Is there any reason why this memory cannot be charged during the
>>>>>>>> allocation (__GFP_ACCOUNT used)?
>>>>>>>> Also you do charge and account the memory but underlying pages 
>>>>>>>> do not
>>>>>>>> know about their memcg (this is normally done with commit_charge 
>>>>>>>> for
>>>>>>>> user mapped pages). This would become a problem if the memory is
>>>>>>>> migrated for example.
>>>>>>>
>>>>>>> I don't think this is movable memory.
>>>>>>>
>>>>>>>> This also means that you have to maintain memcg
>>>>>>>> reference outside of the memcg proper which is not really nice 
>>>>>>>> either.
>>>>>>>> This mimicks tcp kmem limit implementation which I really have 
>>>>>>>> to say I
>>>>>>>> am not a great fan of and this pattern shouldn't be coppied.
>>>>>>>>
>>>>>>>
>>>>>>> I think we should keep the discussion on technical merits instead of
>>>>>>> personal perference. To me using skmem like interface is totally 
>>>>>>> fine
>>>>>>> but the pros/cons need to be very explicit and the clear reasons to
>>>>>>> select that option should be included.
>>>>>>
>>>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
>>>>>> accounting but the overall code maintenance cost is higher because
>>>>>> of how tcp take on accounting differs from anything else in the memcg
>>>>>> proper. I would prefer to not grow another example like that.
>>>>>>
>>>>>>> To me there are two options:
>>>>>>>
>>>>>>> 1. Using skmem like interface as this patch series:
>>>>>>>
>>>>>>> The main pros of this option is that it is very simple. Let me 
>>>>>>> list down
>>>>>>> the cons of this approach:
>>>>>>>
>>>>>>> a. There is time window between the actual memory allocation/free 
>>>>>>> and
>>>>>>> the charge and uncharge and [un]charge happen when the whole 
>>>>>>> memory is
>>>>>>> allocated or freed. I think for the charge path that might not be 
>>>>>>> a big
>>>>>>> issue but on the uncharge, this can cause issues. The application 
>>>>>>> and
>>>>>>> the potential shrinkers have freed some of this dmabuf memory but 
>>>>>>> until
>>>>>>> the whole dmabuf is freed, the memcg uncharge will not happen. 
>>>>>>> This can
>>>>>>> consequences on reclaim and oom behavior of the application.
>>>>>>>
>>>>>>> b. Due to the usage model i.e. a central daemon allocating the 
>>>>>>> dmabuf
>>>>>>> memory upfront, there is a requirement to have a memcg charge 
>>>>>>> transfer
>>>>>>> functionality to transfer the charge from the central daemon to the
>>>>>>> client applications. This does introduce complexity and avenues 
>>>>>>> of weird
>>>>>>> reclaim and oom behavior.
>>>>>>>
>>>>>>>
>>>>>>> 2. Allocate and charge the memory on page fault by actual user
>>>>>>>
>>>>>>> In this approach, the memory is not allocated upfront by the central
>>>>>>> daemon but rather on the page fault by the client application and 
>>>>>>> the
>>>>>>> memcg charge happen at the same time.
>>>>>>>
>>>>>>> The only cons I can think of is this approach is more involved 
>>>>>>> and may
>>>>>>> need some clever tricks to track the page on the free patch i.e. 
>>>>>>> we to
>>>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>>>>>>
>>>>>>> The pros of this approach is there is no need have a charge transfer
>>>>>>> functionality and the charge/uncharge being closely tied to the 
>>>>>>> actual
>>>>>>> memory allocation and free.
>>>>>>>
>>>>>>> Personally I would prefer the second approach but I don't want to 
>>>>>>> just
>>>>>>> block this work if the dmabuf folks are ok with the cons 
>>>>>>> mentioned of
>>>>>>> the first approach.
>>>>>>
>>>>>> I am not familiar with dmabuf internals to judge complexity on 
>>>>>> their end
>>>>>> but I fully agree that charge-when-used is much more easier to reason
>>>>>> about and it should have less subtle surprises.
>>>>>
>>>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so 
>>>>> maybe I
>>>>> am missing something, but in principle yes, I agree that the 2nd 
>>>>> option
>>>>> (charge the user, not exporter) should be preferred. Thing being 
>>>>> that at
>>>>> export time there may not be any backing store allocated, plus if the
>>>>> series is restricting the charge transfer to just Android clients then
>>>>> it seems it has the potential to miss many other use cases. At least
>>>>> needs to outline a description on how the feature will be useful 
>>>>> outside
>>>>> Android.
>>>>>
>>>> There is no restriction like that. It's available to anybody who wants
>>>> to call dma_buf_charge_transfer if they actually have a need for that,
>>>> which I don't really expect to be common since most users/owners of
>>>> the buffers will be the ones causing the export in the first place.
>>>> It's just not like that on Android with the extra allocator process in
>>>> the middle most of the time.
>>>
>>> Yeah I used the wrong term "restrict", apologies. What I meant was, if
>>> the idea was to allow spotting memory leaks, with the charge transfer
>>> being optional and in the series only wired up for Android Binder, then
>>> it obviously only fully works for that one case. So a step back..
>>>
>> Oh, spotting kernel memory leaks is a side-benefit of accounting
>> kernel-only buffers in the root cgroup. The primary goal is to
>> attribute buffers to applications that originated them (via
>> per-application cgroups) simply for accounting purposes. Buffers are
>> using memory on the system, and we want to know who created them and
>> how much memory is used. That information is/will no longer available
>> with the recent deprecation of the dmabuf sysfs statistics.
>>
>>> .. For instance, it is not feasible to transfer the charge when dmabuf
>>> is attached, or imported? That would attribute the usage to the
>>> user/importer so give better visibility on who is actually causing the
>>> memory leak.
>>>
>> Instead of accounting at export, we could account at attach. That just
>> turns out not to be very useful when the majority of our
>> heap-allocated buffers don't have attachments at any particular point
>> in time. :\ But again it's less about leaks and more about knowing
>> which buffers exist in the first place.
>>
>>> Further more, if above is feasible, then could it also be implemented in
>>> the common layer so it would automatically cover all drivers?
>>>
>> Which common layer code specifically? The dmabuf interface appears to
>> be the most central/common place to me.
> 
> Yes, I meant dma_buf_attach / detach. More below.
>>>>> Also stepping back for a moment - is a new memory category really
>>>>> needed, versus perhaps attempting to charge the actual backing store
>>>>> memory to the correct client? (There might have been many past
>>>>> discussions on this so it's okay to point me towards something in the
>>>>> archives.)
>>>>>
>>>> Well the dmabuf counter for the stat file is really just a subcategory
>>>> of memory that is charged. Its existence is not related to getting the
>>>> charge attributed to the right process/cgroup. We do want to know how
>>>> much of the memory attributed to a process is for dmabufs, which is
>>>> the main point of this series.
>>>
>>> Then I am probably missing something because the statement how proposal
>>> is not intended to charge to the right process, but wants to know how
>>> much dmabuf "size" is attributed to a process, confuses me due a seeming
>>> contradiction. And the fact it would not be externally observable how
>>> much of the stats is accurate and how much is not (without knowing the
>>> implementation detail of which drivers implement charge transfer and
>>> when). Maybe I completely misunderstood the use case.
>>>
>> Hmm, did I clear this up above or no? The current proposal is for the
>> process causing the export of a buffer to be charged for it,
>> regardless of whatever happens afterwards. (Unless that process is
>> like gralloc on Android, in which case the charge is transferred from
>> gralloc to whoever called gralloc to allocate the buffer on their
>> behalf.)
> 
> Main problem for me is that charging at export time has no relation to 
> memory used. But I am not familiar with the memcg counters to know if 
> any other counter sets that same precedent. If all other are about real 
> memory use then IMO this does not fit that well. I mean specifically this:
> 
> +      dmabuf (npn)
> +        Amount of memory used for exported DMA buffers allocated by the 
> cgroup.
> +        Stays with the allocating cgroup regardless of how the buffer 
> is shared.
> +
> 
> I think that "Amount of memory used for exported..." is not correct. As 
> implemented it is more akin the virtual address space size in the cpu 
> space - it can have no relation to the actual usage since backing store 
> is not allocated until the attachment is made.
> 
> Then also this:
> 
> @@ -446,6 +447,8 @@ struct dma_buf {
>           struct dma_buf *dmabuf;
>       } *sysfs_entry;
>   #endif
> +    /* The cgroup to which this buffer is currently attributed */
> +    struct mem_cgroup *memcg;
>   };
> 
> Does not conceptually fit in my mind. Dmabufs are not associated with 
> one cgroup at a time.
> 
> So if you would place tracking into dma_buf_attach/detach you would be 
> able to charge to correct cgroup regardless of a driver and since by 
> contract at this stage there is backing store, the reflected memory 
> usage counter would be truthful.
> 
> But then you state a problem, that majority of the time there are no 
> attachments in your setup, and you also say the proposal is not so much 
> about leaks but more about knowing what is exported.
> 
> In this case you could additionally track that via dma_buf_getfile / 
> dma_buf_file_release as a separate category like dmabuf-exported? But 
> again, I personally don't know if such "may not really be using memory" 
> counters fit in memcg.
> 
> (Hm you'd probably still need dmabuf->export_memcg to store who was the 
> original caller of dma_buf_getfile, in case last reference is dropped 
> from a different process/context. Even dmabuf->attach_memcg for 
> attach/detach to work correctly for the same reason.)

Or to work around the "may not really be using memory" problem with the 
exported tracking, perhaps you could record dmabuf->export_memcg at 
dma_buf_export time, but only charge against it at dma_buf_getfile time. 
Assuming it is possible to keep references to those memcg's over the 
dmabuf lifetime without any issues.

That way we could have dmabuf-exported and dmabuf-imported memcg 
categories which would better correlate with real memory usage. I say 
better, because I don't think it would still be perfect since individual 
drivers are allowed to hold onto the backing store post detach and that 
is invisible to dmabuf API. But that probably is a different problem.

Regards,

Tvrtko
  
T.J. Mercier Feb. 2, 2023, 11:43 p.m. UTC | #13
On Wed, Feb 1, 2023 at 6:23 AM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 01/02/2023 01:49, T.J. Mercier wrote:
> > On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
> > <tvrtko.ursulin@linux.intel.com> wrote:
> >>
> >>
> >> On 25/01/2023 20:04, T.J. Mercier wrote:
> >>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
> >>> <tvrtko.ursulin@linux.intel.com> wrote:
> >>>>
> >>>>
> >>>> Hi,
> >>>>
> >>>> On 25/01/2023 11:52, Michal Hocko wrote:
> >>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
> >>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> >>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> >>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
> >>>>>>>> buffer to the allocating cgroup until all buffer references are
> >>>>>>>> released.
> >>>>>>>
> >>>>>>> Is there any reason why this memory cannot be charged during the
> >>>>>>> allocation (__GFP_ACCOUNT used)?
> >>>>>>> Also you do charge and account the memory but underlying pages do not
> >>>>>>> know about their memcg (this is normally done with commit_charge for
> >>>>>>> user mapped pages). This would become a problem if the memory is
> >>>>>>> migrated for example.
> >>>>>>
> >>>>>> I don't think this is movable memory.
> >>>>>>
> >>>>>>> This also means that you have to maintain memcg
> >>>>>>> reference outside of the memcg proper which is not really nice either.
> >>>>>>> This mimicks tcp kmem limit implementation which I really have to say I
> >>>>>>> am not a great fan of and this pattern shouldn't be coppied.
> >>>>>>>
> >>>>>>
> >>>>>> I think we should keep the discussion on technical merits instead of
> >>>>>> personal perference. To me using skmem like interface is totally fine
> >>>>>> but the pros/cons need to be very explicit and the clear reasons to
> >>>>>> select that option should be included.
> >>>>>
> >>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
> >>>>> accounting but the overall code maintenance cost is higher because
> >>>>> of how tcp take on accounting differs from anything else in the memcg
> >>>>> proper. I would prefer to not grow another example like that.
> >>>>>
> >>>>>> To me there are two options:
> >>>>>>
> >>>>>> 1. Using skmem like interface as this patch series:
> >>>>>>
> >>>>>> The main pros of this option is that it is very simple. Let me list down
> >>>>>> the cons of this approach:
> >>>>>>
> >>>>>> a. There is time window between the actual memory allocation/free and
> >>>>>> the charge and uncharge and [un]charge happen when the whole memory is
> >>>>>> allocated or freed. I think for the charge path that might not be a big
> >>>>>> issue but on the uncharge, this can cause issues. The application and
> >>>>>> the potential shrinkers have freed some of this dmabuf memory but until
> >>>>>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
> >>>>>> consequences on reclaim and oom behavior of the application.
> >>>>>>
> >>>>>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
> >>>>>> memory upfront, there is a requirement to have a memcg charge transfer
> >>>>>> functionality to transfer the charge from the central daemon to the
> >>>>>> client applications. This does introduce complexity and avenues of weird
> >>>>>> reclaim and oom behavior.
> >>>>>>
> >>>>>>
> >>>>>> 2. Allocate and charge the memory on page fault by actual user
> >>>>>>
> >>>>>> In this approach, the memory is not allocated upfront by the central
> >>>>>> daemon but rather on the page fault by the client application and the
> >>>>>> memcg charge happen at the same time.
> >>>>>>
> >>>>>> The only cons I can think of is this approach is more involved and may
> >>>>>> need some clever tricks to track the page on the free patch i.e. we to
> >>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
> >>>>>>
> >>>>>> The pros of this approach is there is no need have a charge transfer
> >>>>>> functionality and the charge/uncharge being closely tied to the actual
> >>>>>> memory allocation and free.
> >>>>>>
> >>>>>> Personally I would prefer the second approach but I don't want to just
> >>>>>> block this work if the dmabuf folks are ok with the cons mentioned of
> >>>>>> the first approach.
> >>>>>
> >>>>> I am not familiar with dmabuf internals to judge complexity on their end
> >>>>> but I fully agree that charge-when-used is much more easier to reason
> >>>>> about and it should have less subtle surprises.
> >>>>
> >>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
> >>>> am missing something, but in principle yes, I agree that the 2nd option
> >>>> (charge the user, not exporter) should be preferred. Thing being that at
> >>>> export time there may not be any backing store allocated, plus if the
> >>>> series is restricting the charge transfer to just Android clients then
> >>>> it seems it has the potential to miss many other use cases. At least
> >>>> needs to outline a description on how the feature will be useful outside
> >>>> Android.
> >>>>
> >>> There is no restriction like that. It's available to anybody who wants
> >>> to call dma_buf_charge_transfer if they actually have a need for that,
> >>> which I don't really expect to be common since most users/owners of
> >>> the buffers will be the ones causing the export in the first place.
> >>> It's just not like that on Android with the extra allocator process in
> >>> the middle most of the time.
> >>
> >> Yeah I used the wrong term "restrict", apologies. What I meant was, if
> >> the idea was to allow spotting memory leaks, with the charge transfer
> >> being optional and in the series only wired up for Android Binder, then
> >> it obviously only fully works for that one case. So a step back..
> >>
> > Oh, spotting kernel memory leaks is a side-benefit of accounting
> > kernel-only buffers in the root cgroup. The primary goal is to
> > attribute buffers to applications that originated them (via
> > per-application cgroups) simply for accounting purposes. Buffers are
> > using memory on the system, and we want to know who created them and
> > how much memory is used. That information is/will no longer available
> > with the recent deprecation of the dmabuf sysfs statistics.
> >
> >> .. For instance, it is not feasible to transfer the charge when dmabuf
> >> is attached, or imported? That would attribute the usage to the
> >> user/importer so give better visibility on who is actually causing the
> >> memory leak.
> >>
> > Instead of accounting at export, we could account at attach. That just
> > turns out not to be very useful when the majority of our
> > heap-allocated buffers don't have attachments at any particular point
> > in time. :\ But again it's less about leaks and more about knowing
> > which buffers exist in the first place.
> >
> >> Further more, if above is feasible, then could it also be implemented in
> >> the common layer so it would automatically cover all drivers?
> >>
> > Which common layer code specifically? The dmabuf interface appears to
> > be the most central/common place to me.
>
> Yes, I meant dma_buf_attach / detach. More below.
> >>>> Also stepping back for a moment - is a new memory category really
> >>>> needed, versus perhaps attempting to charge the actual backing store
> >>>> memory to the correct client? (There might have been many past
> >>>> discussions on this so it's okay to point me towards something in the
> >>>> archives.)
> >>>>
> >>> Well the dmabuf counter for the stat file is really just a subcategory
> >>> of memory that is charged. Its existence is not related to getting the
> >>> charge attributed to the right process/cgroup. We do want to know how
> >>> much of the memory attributed to a process is for dmabufs, which is
> >>> the main point of this series.
> >>
> >> Then I am probably missing something because the statement how proposal
> >> is not intended to charge to the right process, but wants to know how
> >> much dmabuf "size" is attributed to a process, confuses me due a seeming
> >> contradiction. And the fact it would not be externally observable how
> >> much of the stats is accurate and how much is not (without knowing the
> >> implementation detail of which drivers implement charge transfer and
> >> when). Maybe I completely misunderstood the use case.
> >>
> > Hmm, did I clear this up above or no? The current proposal is for the
> > process causing the export of a buffer to be charged for it,
> > regardless of whatever happens afterwards. (Unless that process is
> > like gralloc on Android, in which case the charge is transferred from
> > gralloc to whoever called gralloc to allocate the buffer on their
> > behalf.)
>
> Main problem for me is that charging at export time has no relation to memory used. But I am not familiar with the memcg counters to know if any other counter sets that same precedent. If all other are about real memory use then IMO this does not fit that well. I mean specifically this:
>
> +         dmabuf (npn)
> +               Amount of memory used for exported DMA buffers allocated by the cgroup.
> +               Stays with the allocating cgroup regardless of how the buffer is shared.
> +
>
> I think that "Amount of memory used for exported..." is not correct. As implemented it is more akin the virtual address space size in the cpu space - it can have no relation to the actual usage since backing store is not allocated until the attachment is made.
>
> Then also this:
>
> @@ -446,6 +447,8 @@ struct dma_buf {
>                 struct dma_buf *dmabuf;
>         } *sysfs_entry;
>   #endif
> +       /* The cgroup to which this buffer is currently attributed */
> +       struct mem_cgroup *memcg;
>   };
>
> Does not conceptually fit in my mind. Dmabufs are not associated with one cgroup at a time.
>
It's true that a dmabuf could be shared among processes in different
cgroups, but this refers to the one that's charged for it. Similar to
how the shmem pages that back memfds which can be similarly shared get
charged to the first cgroup that touches each page, here it's the
entire buffer instead of each individual page. Maybe it'd be possible
to charge whoever attaches / maps first, but I have to point out
there'd be a gap between then and export where we'd have no accounting
of the memory for cases where pages actually do get allocated during
export (like in the system_heap).

> So if you would place tracking into dma_buf_attach/detach you would be able to charge to correct cgroup regardless of a driver and since by contract at this stage there is backing store, the reflected memory usage counter would be truthful.
>
> But then you state a problem, that majority of the time there are no attachments in your setup, and you also say the proposal is not so much about leaks but more about knowing what is exported.
>
> In this case you could additionally track that via dma_buf_getfile / dma_buf_file_release as a separate category like dmabuf-exported? But again, I personally don't know if such "may not really be using memory" counters fit in memcg.
>
> (Hm you'd probably still need dmabuf->export_memcg to store who was the original caller of dma_buf_getfile, in case last reference is dropped from a different process/context. Even dmabuf->attach_memcg for attach/detach to work correctly for the same reason.)
>
> Regards,
>
> Tvrtko
  
T.J. Mercier Feb. 2, 2023, 11:43 p.m. UTC | #14
On Wed, Feb 1, 2023 at 6:52 AM Tvrtko Ursulin
<tvrtko.ursulin@linux.intel.com> wrote:
>
>
> On 01/02/2023 14:23, Tvrtko Ursulin wrote:
> >
> > On 01/02/2023 01:49, T.J. Mercier wrote:
> >> On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
> >> <tvrtko.ursulin@linux.intel.com> wrote:
> >>>
> >>>
> >>> On 25/01/2023 20:04, T.J. Mercier wrote:
> >>>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
> >>>> <tvrtko.ursulin@linux.intel.com> wrote:
> >>>>>
> >>>>>
> >>>>> Hi,
> >>>>>
> >>>>> On 25/01/2023 11:52, Michal Hocko wrote:
> >>>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
> >>>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
> >>>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
> >>>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
> >>>>>>>>> buffer to the allocating cgroup until all buffer references are
> >>>>>>>>> released.
> >>>>>>>>
> >>>>>>>> Is there any reason why this memory cannot be charged during the
> >>>>>>>> allocation (__GFP_ACCOUNT used)?
> >>>>>>>> Also you do charge and account the memory but underlying pages
> >>>>>>>> do not
> >>>>>>>> know about their memcg (this is normally done with commit_charge
> >>>>>>>> for
> >>>>>>>> user mapped pages). This would become a problem if the memory is
> >>>>>>>> migrated for example.
> >>>>>>>
> >>>>>>> I don't think this is movable memory.
> >>>>>>>
> >>>>>>>> This also means that you have to maintain memcg
> >>>>>>>> reference outside of the memcg proper which is not really nice
> >>>>>>>> either.
> >>>>>>>> This mimicks tcp kmem limit implementation which I really have
> >>>>>>>> to say I
> >>>>>>>> am not a great fan of and this pattern shouldn't be coppied.
> >>>>>>>>
> >>>>>>>
> >>>>>>> I think we should keep the discussion on technical merits instead of
> >>>>>>> personal perference. To me using skmem like interface is totally
> >>>>>>> fine
> >>>>>>> but the pros/cons need to be very explicit and the clear reasons to
> >>>>>>> select that option should be included.
> >>>>>>
> >>>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
> >>>>>> accounting but the overall code maintenance cost is higher because
> >>>>>> of how tcp take on accounting differs from anything else in the memcg
> >>>>>> proper. I would prefer to not grow another example like that.
> >>>>>>
> >>>>>>> To me there are two options:
> >>>>>>>
> >>>>>>> 1. Using skmem like interface as this patch series:
> >>>>>>>
> >>>>>>> The main pros of this option is that it is very simple. Let me
> >>>>>>> list down
> >>>>>>> the cons of this approach:
> >>>>>>>
> >>>>>>> a. There is time window between the actual memory allocation/free
> >>>>>>> and
> >>>>>>> the charge and uncharge and [un]charge happen when the whole
> >>>>>>> memory is
> >>>>>>> allocated or freed. I think for the charge path that might not be
> >>>>>>> a big
> >>>>>>> issue but on the uncharge, this can cause issues. The application
> >>>>>>> and
> >>>>>>> the potential shrinkers have freed some of this dmabuf memory but
> >>>>>>> until
> >>>>>>> the whole dmabuf is freed, the memcg uncharge will not happen.
> >>>>>>> This can
> >>>>>>> consequences on reclaim and oom behavior of the application.
> >>>>>>>
> >>>>>>> b. Due to the usage model i.e. a central daemon allocating the
> >>>>>>> dmabuf
> >>>>>>> memory upfront, there is a requirement to have a memcg charge
> >>>>>>> transfer
> >>>>>>> functionality to transfer the charge from the central daemon to the
> >>>>>>> client applications. This does introduce complexity and avenues
> >>>>>>> of weird
> >>>>>>> reclaim and oom behavior.
> >>>>>>>
> >>>>>>>
> >>>>>>> 2. Allocate and charge the memory on page fault by actual user
> >>>>>>>
> >>>>>>> In this approach, the memory is not allocated upfront by the central
> >>>>>>> daemon but rather on the page fault by the client application and
> >>>>>>> the
> >>>>>>> memcg charge happen at the same time.
> >>>>>>>
> >>>>>>> The only cons I can think of is this approach is more involved
> >>>>>>> and may
> >>>>>>> need some clever tricks to track the page on the free patch i.e.
> >>>>>>> we to
> >>>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
> >>>>>>>
> >>>>>>> The pros of this approach is there is no need have a charge transfer
> >>>>>>> functionality and the charge/uncharge being closely tied to the
> >>>>>>> actual
> >>>>>>> memory allocation and free.
> >>>>>>>
> >>>>>>> Personally I would prefer the second approach but I don't want to
> >>>>>>> just
> >>>>>>> block this work if the dmabuf folks are ok with the cons
> >>>>>>> mentioned of
> >>>>>>> the first approach.
> >>>>>>
> >>>>>> I am not familiar with dmabuf internals to judge complexity on
> >>>>>> their end
> >>>>>> but I fully agree that charge-when-used is much more easier to reason
> >>>>>> about and it should have less subtle surprises.
> >>>>>
> >>>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so
> >>>>> maybe I
> >>>>> am missing something, but in principle yes, I agree that the 2nd
> >>>>> option
> >>>>> (charge the user, not exporter) should be preferred. Thing being
> >>>>> that at
> >>>>> export time there may not be any backing store allocated, plus if the
> >>>>> series is restricting the charge transfer to just Android clients then
> >>>>> it seems it has the potential to miss many other use cases. At least
> >>>>> needs to outline a description on how the feature will be useful
> >>>>> outside
> >>>>> Android.
> >>>>>
> >>>> There is no restriction like that. It's available to anybody who wants
> >>>> to call dma_buf_charge_transfer if they actually have a need for that,
> >>>> which I don't really expect to be common since most users/owners of
> >>>> the buffers will be the ones causing the export in the first place.
> >>>> It's just not like that on Android with the extra allocator process in
> >>>> the middle most of the time.
> >>>
> >>> Yeah I used the wrong term "restrict", apologies. What I meant was, if
> >>> the idea was to allow spotting memory leaks, with the charge transfer
> >>> being optional and in the series only wired up for Android Binder, then
> >>> it obviously only fully works for that one case. So a step back..
> >>>
> >> Oh, spotting kernel memory leaks is a side-benefit of accounting
> >> kernel-only buffers in the root cgroup. The primary goal is to
> >> attribute buffers to applications that originated them (via
> >> per-application cgroups) simply for accounting purposes. Buffers are
> >> using memory on the system, and we want to know who created them and
> >> how much memory is used. That information is/will no longer available
> >> with the recent deprecation of the dmabuf sysfs statistics.
> >>
> >>> .. For instance, it is not feasible to transfer the charge when dmabuf
> >>> is attached, or imported? That would attribute the usage to the
> >>> user/importer so give better visibility on who is actually causing the
> >>> memory leak.
> >>>
> >> Instead of accounting at export, we could account at attach. That just
> >> turns out not to be very useful when the majority of our
> >> heap-allocated buffers don't have attachments at any particular point
> >> in time. :\ But again it's less about leaks and more about knowing
> >> which buffers exist in the first place.
> >>
> >>> Further more, if above is feasible, then could it also be implemented in
> >>> the common layer so it would automatically cover all drivers?
> >>>
> >> Which common layer code specifically? The dmabuf interface appears to
> >> be the most central/common place to me.
> >
> > Yes, I meant dma_buf_attach / detach. More below.
> >>>>> Also stepping back for a moment - is a new memory category really
> >>>>> needed, versus perhaps attempting to charge the actual backing store
> >>>>> memory to the correct client? (There might have been many past
> >>>>> discussions on this so it's okay to point me towards something in the
> >>>>> archives.)
> >>>>>
> >>>> Well the dmabuf counter for the stat file is really just a subcategory
> >>>> of memory that is charged. Its existence is not related to getting the
> >>>> charge attributed to the right process/cgroup. We do want to know how
> >>>> much of the memory attributed to a process is for dmabufs, which is
> >>>> the main point of this series.
> >>>
> >>> Then I am probably missing something because the statement how proposal
> >>> is not intended to charge to the right process, but wants to know how
> >>> much dmabuf "size" is attributed to a process, confuses me due a seeming
> >>> contradiction. And the fact it would not be externally observable how
> >>> much of the stats is accurate and how much is not (without knowing the
> >>> implementation detail of which drivers implement charge transfer and
> >>> when). Maybe I completely misunderstood the use case.
> >>>
> >> Hmm, did I clear this up above or no? The current proposal is for the
> >> process causing the export of a buffer to be charged for it,
> >> regardless of whatever happens afterwards. (Unless that process is
> >> like gralloc on Android, in which case the charge is transferred from
> >> gralloc to whoever called gralloc to allocate the buffer on their
> >> behalf.)
> >
> > Main problem for me is that charging at export time has no relation to
> > memory used. But I am not familiar with the memcg counters to know if
> > any other counter sets that same precedent. If all other are about real
> > memory use then IMO this does not fit that well. I mean specifically this:
> >
> > +      dmabuf (npn)
> > +        Amount of memory used for exported DMA buffers allocated by the
> > cgroup.
> > +        Stays with the allocating cgroup regardless of how the buffer
> > is shared.
> > +
> >
> > I think that "Amount of memory used for exported..." is not correct. As
> > implemented it is more akin the virtual address space size in the cpu
> > space - it can have no relation to the actual usage since backing store
> > is not allocated until the attachment is made.
> >
> > Then also this:
> >
> > @@ -446,6 +447,8 @@ struct dma_buf {
> >           struct dma_buf *dmabuf;
> >       } *sysfs_entry;
> >   #endif
> > +    /* The cgroup to which this buffer is currently attributed */
> > +    struct mem_cgroup *memcg;
> >   };
> >
> > Does not conceptually fit in my mind. Dmabufs are not associated with
> > one cgroup at a time.
> >
> > So if you would place tracking into dma_buf_attach/detach you would be
> > able to charge to correct cgroup regardless of a driver and since by
> > contract at this stage there is backing store, the reflected memory
> > usage counter would be truthful.
> >
> > But then you state a problem, that majority of the time there are no
> > attachments in your setup, and you also say the proposal is not so much
> > about leaks but more about knowing what is exported.
> >
> > In this case you could additionally track that via dma_buf_getfile /
> > dma_buf_file_release as a separate category like dmabuf-exported? But
> > again, I personally don't know if such "may not really be using memory"
> > counters fit in memcg.
> >
> > (Hm you'd probably still need dmabuf->export_memcg to store who was the
> > original caller of dma_buf_getfile, in case last reference is dropped
> > from a different process/context. Even dmabuf->attach_memcg for
> > attach/detach to work correctly for the same reason.)
>
> Or to work around the "may not really be using memory" problem with the
> exported tracking, perhaps you could record dmabuf->export_memcg at
> dma_buf_export time, but only charge against it at dma_buf_getfile time.
> Assuming it is possible to keep references to those memcg's over the
> dmabuf lifetime without any issues.
>
I don't follow here. dma_buf_export calls dma_buf_getfile. Did you
mean dma_buf_attach / dma_buf_mmap instead of dma_buf_getfile? If so
that's an interesting idea, but want to make sure I'm tracking
correctly.

> That way we could have dmabuf-exported and dmabuf-imported memcg
> categories which would better correlate with real memory usage. I say
> better, because I don't think it would still be perfect since individual
> drivers are allowed to hold onto the backing store post detach and that
> is invisible to dmabuf API. But that probably is a different problem.
>
Oh, that sounds... broken.

> Regards,
>
> Tvrtko
  
Tvrtko Ursulin Feb. 3, 2023, 9:27 a.m. UTC | #15
On 02/02/2023 23:43, T.J. Mercier wrote:
> On Wed, Feb 1, 2023 at 6:52 AM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>>
>>
>> On 01/02/2023 14:23, Tvrtko Ursulin wrote:
>>>
>>> On 01/02/2023 01:49, T.J. Mercier wrote:
>>>> On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
>>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>>
>>>>>
>>>>> On 25/01/2023 20:04, T.J. Mercier wrote:
>>>>>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
>>>>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>>>>
>>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> On 25/01/2023 11:52, Michal Hocko wrote:
>>>>>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>>>>>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>>>>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>>>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>>>>>>>>> buffer to the allocating cgroup until all buffer references are
>>>>>>>>>>> released.
>>>>>>>>>>
>>>>>>>>>> Is there any reason why this memory cannot be charged during the
>>>>>>>>>> allocation (__GFP_ACCOUNT used)?
>>>>>>>>>> Also you do charge and account the memory but underlying pages
>>>>>>>>>> do not
>>>>>>>>>> know about their memcg (this is normally done with commit_charge
>>>>>>>>>> for
>>>>>>>>>> user mapped pages). This would become a problem if the memory is
>>>>>>>>>> migrated for example.
>>>>>>>>>
>>>>>>>>> I don't think this is movable memory.
>>>>>>>>>
>>>>>>>>>> This also means that you have to maintain memcg
>>>>>>>>>> reference outside of the memcg proper which is not really nice
>>>>>>>>>> either.
>>>>>>>>>> This mimicks tcp kmem limit implementation which I really have
>>>>>>>>>> to say I
>>>>>>>>>> am not a great fan of and this pattern shouldn't be coppied.
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>> I think we should keep the discussion on technical merits instead of
>>>>>>>>> personal perference. To me using skmem like interface is totally
>>>>>>>>> fine
>>>>>>>>> but the pros/cons need to be very explicit and the clear reasons to
>>>>>>>>> select that option should be included.
>>>>>>>>
>>>>>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
>>>>>>>> accounting but the overall code maintenance cost is higher because
>>>>>>>> of how tcp take on accounting differs from anything else in the memcg
>>>>>>>> proper. I would prefer to not grow another example like that.
>>>>>>>>
>>>>>>>>> To me there are two options:
>>>>>>>>>
>>>>>>>>> 1. Using skmem like interface as this patch series:
>>>>>>>>>
>>>>>>>>> The main pros of this option is that it is very simple. Let me
>>>>>>>>> list down
>>>>>>>>> the cons of this approach:
>>>>>>>>>
>>>>>>>>> a. There is time window between the actual memory allocation/free
>>>>>>>>> and
>>>>>>>>> the charge and uncharge and [un]charge happen when the whole
>>>>>>>>> memory is
>>>>>>>>> allocated or freed. I think for the charge path that might not be
>>>>>>>>> a big
>>>>>>>>> issue but on the uncharge, this can cause issues. The application
>>>>>>>>> and
>>>>>>>>> the potential shrinkers have freed some of this dmabuf memory but
>>>>>>>>> until
>>>>>>>>> the whole dmabuf is freed, the memcg uncharge will not happen.
>>>>>>>>> This can
>>>>>>>>> consequences on reclaim and oom behavior of the application.
>>>>>>>>>
>>>>>>>>> b. Due to the usage model i.e. a central daemon allocating the
>>>>>>>>> dmabuf
>>>>>>>>> memory upfront, there is a requirement to have a memcg charge
>>>>>>>>> transfer
>>>>>>>>> functionality to transfer the charge from the central daemon to the
>>>>>>>>> client applications. This does introduce complexity and avenues
>>>>>>>>> of weird
>>>>>>>>> reclaim and oom behavior.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> 2. Allocate and charge the memory on page fault by actual user
>>>>>>>>>
>>>>>>>>> In this approach, the memory is not allocated upfront by the central
>>>>>>>>> daemon but rather on the page fault by the client application and
>>>>>>>>> the
>>>>>>>>> memcg charge happen at the same time.
>>>>>>>>>
>>>>>>>>> The only cons I can think of is this approach is more involved
>>>>>>>>> and may
>>>>>>>>> need some clever tricks to track the page on the free patch i.e.
>>>>>>>>> we to
>>>>>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>>>>>>>>
>>>>>>>>> The pros of this approach is there is no need have a charge transfer
>>>>>>>>> functionality and the charge/uncharge being closely tied to the
>>>>>>>>> actual
>>>>>>>>> memory allocation and free.
>>>>>>>>>
>>>>>>>>> Personally I would prefer the second approach but I don't want to
>>>>>>>>> just
>>>>>>>>> block this work if the dmabuf folks are ok with the cons
>>>>>>>>> mentioned of
>>>>>>>>> the first approach.
>>>>>>>>
>>>>>>>> I am not familiar with dmabuf internals to judge complexity on
>>>>>>>> their end
>>>>>>>> but I fully agree that charge-when-used is much more easier to reason
>>>>>>>> about and it should have less subtle surprises.
>>>>>>>
>>>>>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so
>>>>>>> maybe I
>>>>>>> am missing something, but in principle yes, I agree that the 2nd
>>>>>>> option
>>>>>>> (charge the user, not exporter) should be preferred. Thing being
>>>>>>> that at
>>>>>>> export time there may not be any backing store allocated, plus if the
>>>>>>> series is restricting the charge transfer to just Android clients then
>>>>>>> it seems it has the potential to miss many other use cases. At least
>>>>>>> needs to outline a description on how the feature will be useful
>>>>>>> outside
>>>>>>> Android.
>>>>>>>
>>>>>> There is no restriction like that. It's available to anybody who wants
>>>>>> to call dma_buf_charge_transfer if they actually have a need for that,
>>>>>> which I don't really expect to be common since most users/owners of
>>>>>> the buffers will be the ones causing the export in the first place.
>>>>>> It's just not like that on Android with the extra allocator process in
>>>>>> the middle most of the time.
>>>>>
>>>>> Yeah I used the wrong term "restrict", apologies. What I meant was, if
>>>>> the idea was to allow spotting memory leaks, with the charge transfer
>>>>> being optional and in the series only wired up for Android Binder, then
>>>>> it obviously only fully works for that one case. So a step back..
>>>>>
>>>> Oh, spotting kernel memory leaks is a side-benefit of accounting
>>>> kernel-only buffers in the root cgroup. The primary goal is to
>>>> attribute buffers to applications that originated them (via
>>>> per-application cgroups) simply for accounting purposes. Buffers are
>>>> using memory on the system, and we want to know who created them and
>>>> how much memory is used. That information is/will no longer available
>>>> with the recent deprecation of the dmabuf sysfs statistics.
>>>>
>>>>> .. For instance, it is not feasible to transfer the charge when dmabuf
>>>>> is attached, or imported? That would attribute the usage to the
>>>>> user/importer so give better visibility on who is actually causing the
>>>>> memory leak.
>>>>>
>>>> Instead of accounting at export, we could account at attach. That just
>>>> turns out not to be very useful when the majority of our
>>>> heap-allocated buffers don't have attachments at any particular point
>>>> in time. :\ But again it's less about leaks and more about knowing
>>>> which buffers exist in the first place.
>>>>
>>>>> Further more, if above is feasible, then could it also be implemented in
>>>>> the common layer so it would automatically cover all drivers?
>>>>>
>>>> Which common layer code specifically? The dmabuf interface appears to
>>>> be the most central/common place to me.
>>>
>>> Yes, I meant dma_buf_attach / detach. More below.
>>>>>>> Also stepping back for a moment - is a new memory category really
>>>>>>> needed, versus perhaps attempting to charge the actual backing store
>>>>>>> memory to the correct client? (There might have been many past
>>>>>>> discussions on this so it's okay to point me towards something in the
>>>>>>> archives.)
>>>>>>>
>>>>>> Well the dmabuf counter for the stat file is really just a subcategory
>>>>>> of memory that is charged. Its existence is not related to getting the
>>>>>> charge attributed to the right process/cgroup. We do want to know how
>>>>>> much of the memory attributed to a process is for dmabufs, which is
>>>>>> the main point of this series.
>>>>>
>>>>> Then I am probably missing something because the statement how proposal
>>>>> is not intended to charge to the right process, but wants to know how
>>>>> much dmabuf "size" is attributed to a process, confuses me due a seeming
>>>>> contradiction. And the fact it would not be externally observable how
>>>>> much of the stats is accurate and how much is not (without knowing the
>>>>> implementation detail of which drivers implement charge transfer and
>>>>> when). Maybe I completely misunderstood the use case.
>>>>>
>>>> Hmm, did I clear this up above or no? The current proposal is for the
>>>> process causing the export of a buffer to be charged for it,
>>>> regardless of whatever happens afterwards. (Unless that process is
>>>> like gralloc on Android, in which case the charge is transferred from
>>>> gralloc to whoever called gralloc to allocate the buffer on their
>>>> behalf.)
>>>
>>> Main problem for me is that charging at export time has no relation to
>>> memory used. But I am not familiar with the memcg counters to know if
>>> any other counter sets that same precedent. If all other are about real
>>> memory use then IMO this does not fit that well. I mean specifically this:
>>>
>>> +      dmabuf (npn)
>>> +        Amount of memory used for exported DMA buffers allocated by the
>>> cgroup.
>>> +        Stays with the allocating cgroup regardless of how the buffer
>>> is shared.
>>> +
>>>
>>> I think that "Amount of memory used for exported..." is not correct. As
>>> implemented it is more akin the virtual address space size in the cpu
>>> space - it can have no relation to the actual usage since backing store
>>> is not allocated until the attachment is made.
>>>
>>> Then also this:
>>>
>>> @@ -446,6 +447,8 @@ struct dma_buf {
>>>            struct dma_buf *dmabuf;
>>>        } *sysfs_entry;
>>>    #endif
>>> +    /* The cgroup to which this buffer is currently attributed */
>>> +    struct mem_cgroup *memcg;
>>>    };
>>>
>>> Does not conceptually fit in my mind. Dmabufs are not associated with
>>> one cgroup at a time.
>>>
>>> So if you would place tracking into dma_buf_attach/detach you would be
>>> able to charge to correct cgroup regardless of a driver and since by
>>> contract at this stage there is backing store, the reflected memory
>>> usage counter would be truthful.
>>>
>>> But then you state a problem, that majority of the time there are no
>>> attachments in your setup, and you also say the proposal is not so much
>>> about leaks but more about knowing what is exported.
>>>
>>> In this case you could additionally track that via dma_buf_getfile /
>>> dma_buf_file_release as a separate category like dmabuf-exported? But
>>> again, I personally don't know if such "may not really be using memory"
>>> counters fit in memcg.
>>>
>>> (Hm you'd probably still need dmabuf->export_memcg to store who was the
>>> original caller of dma_buf_getfile, in case last reference is dropped
>>> from a different process/context. Even dmabuf->attach_memcg for
>>> attach/detach to work correctly for the same reason.)
>>
>> Or to work around the "may not really be using memory" problem with the
>> exported tracking, perhaps you could record dmabuf->export_memcg at
>> dma_buf_export time, but only charge against it at dma_buf_getfile time.
>> Assuming it is possible to keep references to those memcg's over the
>> dmabuf lifetime without any issues.
>>
> I don't follow here. dma_buf_export calls dma_buf_getfile. Did you
> mean dma_buf_attach / dma_buf_mmap instead of dma_buf_getfile? If so
> that's an interesting idea, but want to make sure I'm tracking
> correctly.

Yes sorry, I confused the two sides when typing.

Exported lifetime: dma_buf_getfile to dma_buf_file_release.
Imported lifetime: dma_buf_attach to dma_buf_detach.

Multiple attachments though, so if you want to track imported size the 
importer memcg would probably need to be stored in struct 
dma_buf_attachment.

And exported size would only need to be charged once on first importer 
attaching.

I am not familiar if cgroup migrations would automatically be handled or 
not if you permanently store memcg pointers in the respective dmabuf 
structures.

>> That way we could have dmabuf-exported and dmabuf-imported memcg
>> categories which would better correlate with real memory usage. I say
>> better, because I don't think it would still be perfect since individual
>> drivers are allowed to hold onto the backing store post detach and that
>> is invisible to dmabuf API. But that probably is a different problem.
>>
> Oh, that sounds... broken.

Not broken in general, but definitely an asterisk on the dmabuf charging 
semantics. Unless it is completely incompatible with anything to be 
tracked under memcg?

Regards,

Tvrtko
  
Tvrtko Ursulin Feb. 3, 2023, 9:46 a.m. UTC | #16
On 02/02/2023 23:43, T.J. Mercier wrote:
> On Wed, Feb 1, 2023 at 6:23 AM Tvrtko Ursulin
> <tvrtko.ursulin@linux.intel.com> wrote:
>>
>>
>> On 01/02/2023 01:49, T.J. Mercier wrote:
>>> On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>
>>>>
>>>> On 25/01/2023 20:04, T.J. Mercier wrote:
>>>>> On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
>>>>> <tvrtko.ursulin@linux.intel.com> wrote:
>>>>>>
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> On 25/01/2023 11:52, Michal Hocko wrote:
>>>>>>> On Tue 24-01-23 19:46:28, Shakeel Butt wrote:
>>>>>>>> On Tue, Jan 24, 2023 at 03:59:58PM +0100, Michal Hocko wrote:
>>>>>>>>> On Mon 23-01-23 19:17:23, T.J. Mercier wrote:
>>>>>>>>>> When a buffer is exported to userspace, use memcg to attribute the
>>>>>>>>>> buffer to the allocating cgroup until all buffer references are
>>>>>>>>>> released.
>>>>>>>>>
>>>>>>>>> Is there any reason why this memory cannot be charged during the
>>>>>>>>> allocation (__GFP_ACCOUNT used)?
>>>>>>>>> Also you do charge and account the memory but underlying pages do not
>>>>>>>>> know about their memcg (this is normally done with commit_charge for
>>>>>>>>> user mapped pages). This would become a problem if the memory is
>>>>>>>>> migrated for example.
>>>>>>>>
>>>>>>>> I don't think this is movable memory.
>>>>>>>>
>>>>>>>>> This also means that you have to maintain memcg
>>>>>>>>> reference outside of the memcg proper which is not really nice either.
>>>>>>>>> This mimicks tcp kmem limit implementation which I really have to say I
>>>>>>>>> am not a great fan of and this pattern shouldn't be coppied.
>>>>>>>>>
>>>>>>>>
>>>>>>>> I think we should keep the discussion on technical merits instead of
>>>>>>>> personal perference. To me using skmem like interface is totally fine
>>>>>>>> but the pros/cons need to be very explicit and the clear reasons to
>>>>>>>> select that option should be included.
>>>>>>>
>>>>>>> I do agree with that. I didn't want sound to be personal wrt tcp kmem
>>>>>>> accounting but the overall code maintenance cost is higher because
>>>>>>> of how tcp take on accounting differs from anything else in the memcg
>>>>>>> proper. I would prefer to not grow another example like that.
>>>>>>>
>>>>>>>> To me there are two options:
>>>>>>>>
>>>>>>>> 1. Using skmem like interface as this patch series:
>>>>>>>>
>>>>>>>> The main pros of this option is that it is very simple. Let me list down
>>>>>>>> the cons of this approach:
>>>>>>>>
>>>>>>>> a. There is time window between the actual memory allocation/free and
>>>>>>>> the charge and uncharge and [un]charge happen when the whole memory is
>>>>>>>> allocated or freed. I think for the charge path that might not be a big
>>>>>>>> issue but on the uncharge, this can cause issues. The application and
>>>>>>>> the potential shrinkers have freed some of this dmabuf memory but until
>>>>>>>> the whole dmabuf is freed, the memcg uncharge will not happen. This can
>>>>>>>> consequences on reclaim and oom behavior of the application.
>>>>>>>>
>>>>>>>> b. Due to the usage model i.e. a central daemon allocating the dmabuf
>>>>>>>> memory upfront, there is a requirement to have a memcg charge transfer
>>>>>>>> functionality to transfer the charge from the central daemon to the
>>>>>>>> client applications. This does introduce complexity and avenues of weird
>>>>>>>> reclaim and oom behavior.
>>>>>>>>
>>>>>>>>
>>>>>>>> 2. Allocate and charge the memory on page fault by actual user
>>>>>>>>
>>>>>>>> In this approach, the memory is not allocated upfront by the central
>>>>>>>> daemon but rather on the page fault by the client application and the
>>>>>>>> memcg charge happen at the same time.
>>>>>>>>
>>>>>>>> The only cons I can think of is this approach is more involved and may
>>>>>>>> need some clever tricks to track the page on the free patch i.e. we to
>>>>>>>> decrement the dmabuf memcg stat on free path. Maybe a page flag.
>>>>>>>>
>>>>>>>> The pros of this approach is there is no need have a charge transfer
>>>>>>>> functionality and the charge/uncharge being closely tied to the actual
>>>>>>>> memory allocation and free.
>>>>>>>>
>>>>>>>> Personally I would prefer the second approach but I don't want to just
>>>>>>>> block this work if the dmabuf folks are ok with the cons mentioned of
>>>>>>>> the first approach.
>>>>>>>
>>>>>>> I am not familiar with dmabuf internals to judge complexity on their end
>>>>>>> but I fully agree that charge-when-used is much more easier to reason
>>>>>>> about and it should have less subtle surprises.
>>>>>>
>>>>>> Disclaimer that I don't seem to see patches 3&4 on dri-devel so maybe I
>>>>>> am missing something, but in principle yes, I agree that the 2nd option
>>>>>> (charge the user, not exporter) should be preferred. Thing being that at
>>>>>> export time there may not be any backing store allocated, plus if the
>>>>>> series is restricting the charge transfer to just Android clients then
>>>>>> it seems it has the potential to miss many other use cases. At least
>>>>>> needs to outline a description on how the feature will be useful outside
>>>>>> Android.
>>>>>>
>>>>> There is no restriction like that. It's available to anybody who wants
>>>>> to call dma_buf_charge_transfer if they actually have a need for that,
>>>>> which I don't really expect to be common since most users/owners of
>>>>> the buffers will be the ones causing the export in the first place.
>>>>> It's just not like that on Android with the extra allocator process in
>>>>> the middle most of the time.
>>>>
>>>> Yeah I used the wrong term "restrict", apologies. What I meant was, if
>>>> the idea was to allow spotting memory leaks, with the charge transfer
>>>> being optional and in the series only wired up for Android Binder, then
>>>> it obviously only fully works for that one case. So a step back..
>>>>
>>> Oh, spotting kernel memory leaks is a side-benefit of accounting
>>> kernel-only buffers in the root cgroup. The primary goal is to
>>> attribute buffers to applications that originated them (via
>>> per-application cgroups) simply for accounting purposes. Buffers are
>>> using memory on the system, and we want to know who created them and
>>> how much memory is used. That information is/will no longer available
>>> with the recent deprecation of the dmabuf sysfs statistics.
>>>
>>>> .. For instance, it is not feasible to transfer the charge when dmabuf
>>>> is attached, or imported? That would attribute the usage to the
>>>> user/importer so give better visibility on who is actually causing the
>>>> memory leak.
>>>>
>>> Instead of accounting at export, we could account at attach. That just
>>> turns out not to be very useful when the majority of our
>>> heap-allocated buffers don't have attachments at any particular point
>>> in time. :\ But again it's less about leaks and more about knowing
>>> which buffers exist in the first place.
>>>
>>>> Further more, if above is feasible, then could it also be implemented in
>>>> the common layer so it would automatically cover all drivers?
>>>>
>>> Which common layer code specifically? The dmabuf interface appears to
>>> be the most central/common place to me.
>>
>> Yes, I meant dma_buf_attach / detach. More below.
>>>>>> Also stepping back for a moment - is a new memory category really
>>>>>> needed, versus perhaps attempting to charge the actual backing store
>>>>>> memory to the correct client? (There might have been many past
>>>>>> discussions on this so it's okay to point me towards something in the
>>>>>> archives.)
>>>>>>
>>>>> Well the dmabuf counter for the stat file is really just a subcategory
>>>>> of memory that is charged. Its existence is not related to getting the
>>>>> charge attributed to the right process/cgroup. We do want to know how
>>>>> much of the memory attributed to a process is for dmabufs, which is
>>>>> the main point of this series.
>>>>
>>>> Then I am probably missing something because the statement how proposal
>>>> is not intended to charge to the right process, but wants to know how
>>>> much dmabuf "size" is attributed to a process, confuses me due a seeming
>>>> contradiction. And the fact it would not be externally observable how
>>>> much of the stats is accurate and how much is not (without knowing the
>>>> implementation detail of which drivers implement charge transfer and
>>>> when). Maybe I completely misunderstood the use case.
>>>>
>>> Hmm, did I clear this up above or no? The current proposal is for the
>>> process causing the export of a buffer to be charged for it,
>>> regardless of whatever happens afterwards. (Unless that process is
>>> like gralloc on Android, in which case the charge is transferred from
>>> gralloc to whoever called gralloc to allocate the buffer on their
>>> behalf.)
>>
>> Main problem for me is that charging at export time has no relation to memory used. But I am not familiar with the memcg counters to know if any other counter sets that same precedent. If all other are about real memory use then IMO this does not fit that well. I mean specifically this:
>>
>> +         dmabuf (npn)
>> +               Amount of memory used for exported DMA buffers allocated by the cgroup.
>> +               Stays with the allocating cgroup regardless of how the buffer is shared.
>> +
>>
>> I think that "Amount of memory used for exported..." is not correct. As implemented it is more akin the virtual address space size in the cpu space - it can have no relation to the actual usage since backing store is not allocated until the attachment is made.
>>
>> Then also this:
>>
>> @@ -446,6 +447,8 @@ struct dma_buf {
>>                  struct dma_buf *dmabuf;
>>          } *sysfs_entry;
>>    #endif
>> +       /* The cgroup to which this buffer is currently attributed */
>> +       struct mem_cgroup *memcg;
>>    };
>>
>> Does not conceptually fit in my mind. Dmabufs are not associated with one cgroup at a time.
>>
> It's true that a dmabuf could be shared among processes in different
> cgroups, but this refers to the one that's charged for it. Similar to
> how the shmem pages that back memfds which can be similarly shared get
> charged to the first cgroup that touches each page, here it's the
> entire buffer instead of each individual page. Maybe it'd be possible
> to charge whoever attaches / maps first, but I have to point out
> there'd be a gap between then and export where we'd have no accounting
> of the memory for cases where pages actually do get allocated during
> export (like in the system_heap).

Okay I wasn't familiar with heaps until now - indeed - allocating a dma 
buf from there is allocation and export in one, no delayed/lazy anything 
on neither edge. Therefore charge at exports works there.

One option - rename the proposed memcg category to be clear it is only 
for dma buf heaps?

But does it not create double accounting btw? Since there are both 
pages/cma allocations that would be tracked and the new dma buf category.

Another option was allow each "backend" to specify if export charge 
needs to happen on export or import to be more accurate? (Like a flag 
for dma_buf_export_info maybe.)

Regards,

Tvrtko

>> So if you would place tracking into dma_buf_attach/detach you would be able to charge to correct cgroup regardless of a driver and since by contract at this stage there is backing store, the reflected memory usage counter would be truthful.
>>
>> But then you state a problem, that majority of the time there are no attachments in your setup, and you also say the proposal is not so much about leaks but more about knowing what is exported.
>>
>> In this case you could additionally track that via dma_buf_getfile / dma_buf_file_release as a separate category like dmabuf-exported? But again, I personally don't know if such "may not really be using memory" counters fit in memcg.
>>
>> (Hm you'd probably still need dmabuf->export_memcg to store who was the original caller of dma_buf_getfile, in case last reference is dropped from a different process/context. Even dmabuf->attach_memcg for attach/detach to work correctly for the same reason.)
>>
>> Regards,
>>
>> Tvrtko
  

Patch

diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst
index c8ae7c897f14..538ae22bc514 100644
--- a/Documentation/admin-guide/cgroup-v2.rst
+++ b/Documentation/admin-guide/cgroup-v2.rst
@@ -1455,6 +1455,10 @@  PAGE_SIZE multiple when read back.
 		Amount of memory used for storing in-kernel data
 		structures.
 
+	  dmabuf (npn)
+		Amount of memory used for exported DMA buffers allocated by the cgroup.
+		Stays with the allocating cgroup regardless of how the buffer is shared.
+
 	  workingset_refault_anon
 		Number of refaults of previously evicted anonymous pages.
 
diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
index e6528767efc7..a6a8cb5cb32d 100644
--- a/drivers/dma-buf/dma-buf.c
+++ b/drivers/dma-buf/dma-buf.c
@@ -75,6 +75,9 @@  static void dma_buf_release(struct dentry *dentry)
 	 */
 	BUG_ON(dmabuf->cb_in.active || dmabuf->cb_out.active);
 
+	mem_cgroup_uncharge_dmabuf(dmabuf->memcg, PAGE_ALIGN(dmabuf->size) / PAGE_SIZE);
+	mem_cgroup_put(dmabuf->memcg);
+
 	dma_buf_stats_teardown(dmabuf);
 	dmabuf->ops->release(dmabuf);
 
@@ -673,6 +676,13 @@  struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
 	if (ret)
 		goto err_dmabuf;
 
+	dmabuf->memcg = get_mem_cgroup_from_mm(current->mm);
+	if (!mem_cgroup_charge_dmabuf(dmabuf->memcg, PAGE_ALIGN(dmabuf->size) / PAGE_SIZE,
+				      GFP_KERNEL)) {
+		ret = -ENOMEM;
+		goto err_memcg;
+	}
+
 	file->private_data = dmabuf;
 	file->f_path.dentry->d_fsdata = dmabuf;
 	dmabuf->file = file;
@@ -683,6 +693,9 @@  struct dma_buf *dma_buf_export(const struct dma_buf_export_info *exp_info)
 
 	return dmabuf;
 
+err_memcg:
+	mem_cgroup_put(dmabuf->memcg);
+	dma_buf_stats_teardown(dmabuf);
 err_dmabuf:
 	if (!resv)
 		dma_resv_fini(dmabuf->resv);
diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
index 6fa8d4e29719..1f0ffb8e4bf5 100644
--- a/include/linux/dma-buf.h
+++ b/include/linux/dma-buf.h
@@ -22,6 +22,7 @@ 
 #include <linux/fs.h>
 #include <linux/dma-fence.h>
 #include <linux/wait.h>
+#include <linux/memcontrol.h>
 
 struct device;
 struct dma_buf;
@@ -446,6 +447,8 @@  struct dma_buf {
 		struct dma_buf *dmabuf;
 	} *sysfs_entry;
 #endif
+	/* The cgroup to which this buffer is currently attributed */
+	struct mem_cgroup *memcg;
 };
 
 /**
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index d3c8203cab6c..c10b8565fdbf 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -37,6 +37,7 @@  enum memcg_stat_item {
 	MEMCG_KMEM,
 	MEMCG_ZSWAP_B,
 	MEMCG_ZSWAPPED,
+	MEMCG_DMABUF,
 	MEMCG_NR_STAT,
 };
 
@@ -673,6 +674,25 @@  static inline int mem_cgroup_charge(struct folio *folio, struct mm_struct *mm,
 
 int mem_cgroup_swapin_charge_folio(struct folio *folio, struct mm_struct *mm,
 				  gfp_t gfp, swp_entry_t entry);
+
+/**
+ * mem_cgroup_charge_dmabuf - Charge dma-buf memory to a cgroup and update stat counter
+ * @memcg: memcg to charge
+ * @nr_pages: number of pages to charge
+ * @gfp_mask: reclaim mode
+ *
+ * Charges @nr_pages to @memcg. Returns %true if the charge fit within
+ * @memcg's configured limit, %false if it doesn't.
+ */
+bool __mem_cgroup_charge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages, gfp_t gfp_mask);
+static inline bool mem_cgroup_charge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages,
+					    gfp_t gfp_mask)
+{
+	if (mem_cgroup_disabled())
+		return 0;
+	return __mem_cgroup_charge_dmabuf(memcg, nr_pages, gfp_mask);
+}
+
 void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry);
 
 void __mem_cgroup_uncharge(struct folio *folio);
@@ -690,6 +710,14 @@  static inline void mem_cgroup_uncharge(struct folio *folio)
 	__mem_cgroup_uncharge(folio);
 }
 
+void __mem_cgroup_uncharge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages);
+static inline void mem_cgroup_uncharge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages)
+{
+	if (mem_cgroup_disabled())
+		return;
+	__mem_cgroup_uncharge_dmabuf(memcg, nr_pages);
+}
+
 void __mem_cgroup_uncharge_list(struct list_head *page_list);
 static inline void mem_cgroup_uncharge_list(struct list_head *page_list)
 {
@@ -1242,6 +1270,12 @@  static inline int mem_cgroup_swapin_charge_folio(struct folio *folio,
 	return 0;
 }
 
+static inline bool mem_cgroup_charge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages,
+					    gfp_t gfp_mask)
+{
+	return true;
+}
+
 static inline void mem_cgroup_swapin_uncharge_swap(swp_entry_t entry)
 {
 }
@@ -1250,6 +1284,10 @@  static inline void mem_cgroup_uncharge(struct folio *folio)
 {
 }
 
+static inline void mem_cgroup_uncharge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages)
+{
+}
+
 static inline void mem_cgroup_uncharge_list(struct list_head *page_list)
 {
 }
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index ab457f0394ab..375d18370f4b 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1502,6 +1502,7 @@  static const struct memory_stat memory_stats[] = {
 	{ "unevictable",		NR_UNEVICTABLE			},
 	{ "slab_reclaimable",		NR_SLAB_RECLAIMABLE_B		},
 	{ "slab_unreclaimable",		NR_SLAB_UNRECLAIMABLE_B		},
+	{ "dmabuf",			MEMCG_DMABUF			},
 
 	/* The memory events */
 	{ "workingset_refault_anon",	WORKINGSET_REFAULT_ANON		},
@@ -4042,6 +4043,7 @@  static const unsigned int memcg1_stats[] = {
 	WORKINGSET_REFAULT_ANON,
 	WORKINGSET_REFAULT_FILE,
 	MEMCG_SWAP,
+	MEMCG_DMABUF,
 };
 
 static const char *const memcg1_stat_names[] = {
@@ -4057,6 +4059,7 @@  static const char *const memcg1_stat_names[] = {
 	"workingset_refault_anon",
 	"workingset_refault_file",
 	"swap",
+	"dmabuf",
 };
 
 /* Universal VM events cgroup1 shows, original sort order */
@@ -7299,6 +7302,22 @@  void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
 	refill_stock(memcg, nr_pages);
 }
 
+bool __mem_cgroup_charge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages, gfp_t gfp_mask)
+{
+	if (try_charge(memcg, gfp_mask, nr_pages) == 0) {
+		mod_memcg_state(memcg, MEMCG_DMABUF, nr_pages);
+		return true;
+	}
+
+	return false;
+}
+
+void __mem_cgroup_uncharge_dmabuf(struct mem_cgroup *memcg, unsigned int nr_pages)
+{
+	mod_memcg_state(memcg, MEMCG_DMABUF, -nr_pages);
+	refill_stock(memcg, nr_pages);
+}
+
 static int __init cgroup_memory(char *s)
 {
 	char *token;