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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 




On 02/02/2023 23:43, T.J. Mercier wrote:
On Wed, Feb 1, 2023 at 6:23 AM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:


On 01/02/2023 01:49, T.J. Mercier wrote:
On Tue, Jan 31, 2023 at 6:01 AM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> wrote:


On 25/01/2023 20:04, T.J. Mercier wrote:
On Wed, Jan 25, 2023 at 9:31 AM Tvrtko Ursulin
<tvrtko.ursulin@xxxxxxxxxxxxxxx> 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



[Index of Archives]     [Linux Input]     [Video for Linux]     [Gstreamer Embedded]     [Mplayer Users]     [Linux USB Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [Yosemite Backpacking]

  Powered by Linux