| Message ID | 20220510235653.933868-1-tjmercier@google.com |
|---|---|
| Headers | show |
| Series | Proposal for a GPU cgroup controller | expand |
On Thu, May 12, 2022 at 6:10 AM Nicolas Dufresne <nicolas@ndufresne.ca> wrote: > > Le mercredi 11 mai 2022 à 13:31 -0700, T.J. Mercier a écrit : > > On Wed, May 11, 2022 at 6:21 AM Nicolas Dufresne <nicolas@ndufresne.ca> wrote: > > > > > > Hi, > > > > > > Le mardi 10 mai 2022 à 23:56 +0000, T.J. Mercier a écrit : > > > > This patch series revisits the proposal for a GPU cgroup controller to > > > > track and limit memory allocations by various device/allocator > > > > subsystems. The patch series also contains a simple prototype to > > > > illustrate how Android intends to implement DMA-BUF allocator > > > > attribution using the GPU cgroup controller. The prototype does not > > > > include resource limit enforcements. > > > > > > I'm sorry, since I'm not in-depth technically involve. But from reading the > > > topic I don't understand the bound this creates between DMABuf Heaps and GPU. Is > > > this an attempt to really track the DMABuf allocated by userland, or just > > > something for GPU ? What about V4L2 devices ? Any way this can be clarified, > > > specially what would other subsystem needs to have cgroup DMABuf allocation > > > controller support ? > > > > > Hi Nicolas, > > > > The link between dmabufs, dmabuf heaps, and "GPU memory" is maybe > > somewhat of an Androidism. However this change aims to be usable for > > tracking all GPU related allocations. It's just that this initial > > series only adds support for tracking dmabufs allocated from dmabuf > > heaps. > > > > In Android most graphics buffers are dma buffers allocated from a > > dmabuf heap, so that is why these dmabuf heap allocations are being > > tracked under the GPU cgroup. Other dmabuf exporters like V4L2 might > > also want to track their buffers, but would probably want to do so > > under a bucket name of something like "v4l2". Same goes for GEM > > dmabufs. The naming scheme for this is still yet to be decided. It > > would be cool to be able to attribute memory at the driver level, or > > even different types of memory at the driver level, but I imagine > > there is a point of diminishing returns for fine-grained > > naming/bucketing. > > > > So far, I haven't tried to create a strict definition of what is and > > is not "GPU memory" for the purpose of this accounting, so I don't > > think we should be restricted to tracking just dmabufs. I don't see > > why this couldn't be anything a driver wants to consider as GPU memory > > as long as it is named/bucketed appropriately, such as both on-package > > graphics card memory use and CPU memory dedicated for graphics use > > like for host/device transfers. > > > > Is that helpful? > > I'm actually happy I've asked this question, wasn't silly after all. I think the > problem here is a naming issue. What you really are monitor is "video memory", > which consist of a memory segment allocated to store data used to render images > (its not always images of course, GPU an VPU have specialized buffers for their > purpose). > > Whether this should be split between what is used specifically by the GPU > drivers, the display drivers, the VPU (CODEC and pre/post-processor) or camera > drivers is something that should be discussed. But in the current approach, you > really meant Video memory as a superset of the above. Personally, I think > generically (to de-Andronized your work), en-globing all video memory is > sufficient. What I fail to understand is how you will manage to distinguished > DMABuf Heap allocation (which are used outside of Android btw), from Video > allocation or other type of usage. I'm sure non-video usage will exist in the > future (think of machine learning, compute, other high bandwidth streaming > thingy ...) > Ok thank you for pointing out the naming issue. The naming is a consequence of the initial use case, but I guess it's too specific. What I want out of this change is that android can track dmabufs that come out of heaps, and drm can track gpu memory. But other drivers could track different resources under different names. Imagine this were called a buffer cgroup controller instead of a GPU cgroup controller. Then the use component ("video memory") isn't tied up with the name of the controller, but it's up to the name of the bucket the resource is tracked under. I think this meets the needs of the two use cases I'm aware of now, while leaving the door open to other future needs. Really the controller is just enabling abstract named buckets for tracking and eventually limiting a type of resource. P.S. I will be unavailable starting tomorrow, but I'll be back on Monday. > > > > Best, > > T.J. > > > > > > > > > > Changelog: > > > > v7: > > > > Hide gpucg and gpucg_bucket struct definitions per Michal Koutný. > > > > This means gpucg_register_bucket now returns an internally allocated > > > > struct gpucg_bucket. > > > > > > > > Move all public function documentation to the cgroup_gpu.h header. > > > > > > > > Remove comment in documentation about duplicate name rejection which > > > > is not relevant to cgroups users per Michal Koutný. > > > > > > > > v6: > > > > Move documentation into cgroup-v2.rst per Tejun Heo. > > > > > > > > Rename BINDER_FD{A}_FLAG_SENDER_NO_NEED -> > > > > BINDER_FD{A}_FLAG_XFER_CHARGE per Carlos Llamas. > > > > > > > > Return error on transfer failure per Carlos Llamas. > > > > > > > > v5: > > > > Rebase on top of v5.18-rc3 > > > > > > > > Drop the global GPU cgroup "total" (sum of all device totals) portion > > > > of the design since there is no currently known use for this per > > > > Tejun Heo. > > > > > > > > Fix commit message which still contained the old name for > > > > dma_buf_transfer_charge per Michal Koutný. > > > > > > > > Remove all GPU cgroup code except what's necessary to support charge transfer > > > > from dma_buf. Previously charging was done in export, but for non-Android > > > > graphics use-cases this is not ideal since there may be a delay between > > > > allocation and export, during which time there is no accounting. > > > > > > > > Merge dmabuf: Use the GPU cgroup charge/uncharge APIs patch into > > > > dmabuf: heaps: export system_heap buffers with GPU cgroup charging as a > > > > result of above. > > > > > > > > Put the charge and uncharge code in the same file (system_heap_allocate, > > > > system_heap_dma_buf_release) instead of splitting them between the heap and > > > > the dma_buf_release. This avoids asymmetric management of the gpucg charges. > > > > > > > > Modify the dma_buf_transfer_charge API to accept a task_struct instead > > > > of a gpucg. This avoids requiring the caller to manage the refcount > > > > of the gpucg upon failure and confusing ownership transfer logic. > > > > > > > > Support all strings for gpucg_register_bucket instead of just string > > > > literals. > > > > > > > > Enforce globally unique gpucg_bucket names. > > > > > > > > Constrain gpucg_bucket name lengths to 64 bytes. > > > > > > > > Append "-heap" to gpucg_bucket names from dmabuf-heaps. > > > > > > > > Drop patch 7 from the series, which changed the types of > > > > binder_transaction_data's sender_pid and sender_euid fields. This was > > > > done in another commit here: > > > > https://lore.kernel.org/all/20220210021129.3386083-4-masahiroy@kernel.org/ > > > > > > > > Rename: > > > > gpucg_try_charge -> gpucg_charge > > > > find_cg_rpool_locked -> cg_rpool_find_locked > > > > init_cg_rpool -> cg_rpool_init > > > > get_cg_rpool_locked -> cg_rpool_get_locked > > > > "gpu cgroup controller" -> "GPU controller" > > > > gpucg_device -> gpucg_bucket > > > > usage -> size > > > > > > > > Tests: > > > > Support both binder_fd_array_object and binder_fd_object. This is > > > > necessary because new versions of Android will use binder_fd_object > > > > instead of binder_fd_array_object, and we need to support both. > > > > > > > > Tests for both binder_fd_array_object and binder_fd_object. > > > > > > > > For binder_utils return error codes instead of > > > > struct binder{fs}_ctx. > > > > > > > > Use ifdef __ANDROID__ to choose platform-dependent temp path instead > > > > of a runtime fallback. > > > > > > > > Ensure binderfs_mntpt ends with a trailing '/' character instead of > > > > prepending it where used. > > > > > > > > v4: > > > > Skip test if not run as root per Shuah Khan > > > > > > > > Add better test logging for abnormal child termination per Shuah Khan > > > > > > > > Adjust ordering of charge/uncharge during transfer to avoid potentially > > > > hitting cgroup limit per Michal Koutný > > > > > > > > Adjust gpucg_try_charge critical section for charge transfer functionality > > > > > > > > Fix uninitialized return code error for dmabuf_try_charge error case > > > > > > > > v3: > > > > Remove Upstreaming Plan from gpu-cgroup.rst per John Stultz > > > > > > > > Use more common dual author commit message format per John Stultz > > > > > > > > Remove android from binder changes title per Todd Kjos > > > > > > > > Add a kselftest for this new behavior per Greg Kroah-Hartman > > > > > > > > Include details on behavior for all combinations of kernel/userspace > > > > versions in changelog (thanks Suren Baghdasaryan) per Greg Kroah-Hartman. > > > > > > > > Fix pid and uid types in binder UAPI header > > > > > > > > v2: > > > > See the previous revision of this change submitted by Hridya Valsaraju > > > > at: https://lore.kernel.org/all/20220115010622.3185921-1-hridya@google.com/ > > > > > > > > Move dma-buf cgroup charge transfer from a dma_buf_op defined by every > > > > heap to a single dma-buf function for all heaps per Daniel Vetter and > > > > Christian König. Pointers to struct gpucg and struct gpucg_device > > > > tracking the current associations were added to the dma_buf struct to > > > > achieve this. > > > > > > > > Fix incorrect Kconfig help section indentation per Randy Dunlap. > > > > > > > > History of the GPU cgroup controller > > > > ==================================== > > > > The GPU/DRM cgroup controller came into being when a consensus[1] > > > > was reached that the resources it tracked were unsuitable to be integrated > > > > into memcg. Originally, the proposed controller was specific to the DRM > > > > subsystem and was intended to track GEM buffers and GPU-specific > > > > resources[2]. In order to help establish a unified memory accounting model > > > > for all GPU and all related subsystems, Daniel Vetter put forth a > > > > suggestion to move it out of the DRM subsystem so that it can be used by > > > > other DMA-BUF exporters as well[3]. This RFC proposes an interface that > > > > does the same. > > > > > > > > [1]: https://patchwork.kernel.org/project/dri-devel/cover/20190501140438.9506-1-brian.welty@intel.com/#22624705 > > > > [2]: https://lore.kernel.org/amd-gfx/20210126214626.16260-1-brian.welty@intel.com/ > > > > [3]: https://lore.kernel.org/amd-gfx/YCVOl8%2F87bqRSQei@phenom.ffwll.local/ > > > > > > > > Hridya Valsaraju (3): > > > > gpu: rfc: Proposal for a GPU cgroup controller > > > > cgroup: gpu: Add a cgroup controller for allocator attribution of GPU > > > > memory > > > > binder: Add flags to relinquish ownership of fds > > > > > > > > T.J. Mercier (3): > > > > dmabuf: heaps: export system_heap buffers with GPU cgroup charging > > > > dmabuf: Add gpu cgroup charge transfer function > > > > selftests: Add binder cgroup gpu memory transfer tests > > > > > > > > Documentation/admin-guide/cgroup-v2.rst | 23 + > > > > drivers/android/binder.c | 31 +- > > > > drivers/dma-buf/dma-buf.c | 80 ++- > > > > drivers/dma-buf/dma-heap.c | 38 ++ > > > > drivers/dma-buf/heaps/system_heap.c | 28 +- > > > > include/linux/cgroup_gpu.h | 146 +++++ > > > > include/linux/cgroup_subsys.h | 4 + > > > > include/linux/dma-buf.h | 49 +- > > > > include/linux/dma-heap.h | 15 + > > > > include/uapi/linux/android/binder.h | 23 +- > > > > init/Kconfig | 7 + > > > > kernel/cgroup/Makefile | 1 + > > > > kernel/cgroup/gpu.c | 390 +++++++++++++ > > > > .../selftests/drivers/android/binder/Makefile | 8 + > > > > .../drivers/android/binder/binder_util.c | 250 +++++++++ > > > > .../drivers/android/binder/binder_util.h | 32 ++ > > > > .../selftests/drivers/android/binder/config | 4 + > > > > .../binder/test_dmabuf_cgroup_transfer.c | 526 ++++++++++++++++++ > > > > 18 files changed, 1632 insertions(+), 23 deletions(-) > > > > create mode 100644 include/linux/cgroup_gpu.h > > > > create mode 100644 kernel/cgroup/gpu.c > > > > create mode 100644 tools/testing/selftests/drivers/android/binder/Makefile > > > > create mode 100644 tools/testing/selftests/drivers/android/binder/binder_util.c > > > > create mode 100644 tools/testing/selftests/drivers/android/binder/binder_util.h > > > > create mode 100644 tools/testing/selftests/drivers/android/binder/config > > > > create mode 100644 tools/testing/selftests/drivers/android/binder/test_dmabuf_cgroup_transfer.c > > > > > > > >
Hello, On Thu, May 12, 2022 at 08:43:52PM -0700, T.J. Mercier wrote: > > I'm actually happy I've asked this question, wasn't silly after all. I think the > > problem here is a naming issue. What you really are monitor is "video memory", > > which consist of a memory segment allocated to store data used to render images > > (its not always images of course, GPU an VPU have specialized buffers for their > > purpose). > > > > Whether this should be split between what is used specifically by the GPU > > drivers, the display drivers, the VPU (CODEC and pre/post-processor) or camera > > drivers is something that should be discussed. But in the current approach, you > > really meant Video memory as a superset of the above. Personally, I think > > generically (to de-Andronized your work), en-globing all video memory is > > sufficient. What I fail to understand is how you will manage to distinguished > > DMABuf Heap allocation (which are used outside of Android btw), from Video > > allocation or other type of usage. I'm sure non-video usage will exist in the > > future (think of machine learning, compute, other high bandwidth streaming > > thingy ...) > > > Ok thank you for pointing out the naming issue. The naming is a > consequence of the initial use case, but I guess it's too specific. > What I want out of this change is that android can track dmabufs that > come out of heaps, and drm can track gpu memory. But other drivers > could track different resources under different names. Imagine this > were called a buffer cgroup controller instead of a GPU cgroup > controller. Then the use component ("video memory") isn't tied up with > the name of the controller, but it's up to the name of the bucket the > resource is tracked under. I think this meets the needs of the two use > cases I'm aware of now, while leaving the door open to other future > needs. Really the controller is just enabling abstract named buckets > for tracking and eventually limiting a type of resource. So, there hasn't been whole lot of discussion w/ other GPU folks and what comes up still seems to indicate that we're still long way away from having a meaningful gpu controller. For your use case, would it make sense to just add dmabuf as a key to the misc controller? I'm not sure it makes sense to push "gpu controller" forward if there's no conceptual consensus around what resources are. Thanks.
On Fri, May 13, 2022 at 9:13 AM Tejun Heo <tj@kernel.org> wrote: > > Hello, > > On Thu, May 12, 2022 at 08:43:52PM -0700, T.J. Mercier wrote: > > > I'm actually happy I've asked this question, wasn't silly after all. I think the > > > problem here is a naming issue. What you really are monitor is "video memory", > > > which consist of a memory segment allocated to store data used to render images > > > (its not always images of course, GPU an VPU have specialized buffers for their > > > purpose). > > > > > > Whether this should be split between what is used specifically by the GPU > > > drivers, the display drivers, the VPU (CODEC and pre/post-processor) or camera > > > drivers is something that should be discussed. But in the current approach, you > > > really meant Video memory as a superset of the above. Personally, I think > > > generically (to de-Andronized your work), en-globing all video memory is > > > sufficient. What I fail to understand is how you will manage to distinguished > > > DMABuf Heap allocation (which are used outside of Android btw), from Video > > > allocation or other type of usage. I'm sure non-video usage will exist in the > > > future (think of machine learning, compute, other high bandwidth streaming > > > thingy ...) > > > > > Ok thank you for pointing out the naming issue. The naming is a > > consequence of the initial use case, but I guess it's too specific. > > What I want out of this change is that android can track dmabufs that > > come out of heaps, and drm can track gpu memory. But other drivers > > could track different resources under different names. Imagine this > > were called a buffer cgroup controller instead of a GPU cgroup > > controller. Then the use component ("video memory") isn't tied up with > > the name of the controller, but it's up to the name of the bucket the > > resource is tracked under. I think this meets the needs of the two use > > cases I'm aware of now, while leaving the door open to other future > > needs. Really the controller is just enabling abstract named buckets > > for tracking and eventually limiting a type of resource. > > So, there hasn't been whole lot of discussion w/ other GPU folks and what > comes up still seems to indicate that we're still long way away from having > a meaningful gpu controller. > Yes, and I would still be happy to collaborate. > For your use case, would it make sense to just > add dmabuf as a key to the misc controller? > Thanks for your suggestion. This almost works. "dmabuf" as a key could work, but I'd actually like to account for each heap. Since heaps can be dynamically added, I can't accommodate every potential heap name by hardcoding registrations in the misc controller. > I'm not sure it makes sense to > push "gpu controller" forward if there's no conceptual consensus around what > resources are. > > Thanks. > > -- > tejun
From: Enric Balletbo i Serra <eballetbo@kernel.org> On Tue, 10 May 2022 23:56:45 +0000, T.J. Mercier wrote: > From: Hridya Valsaraju <hridya@google.com> > Hi T.J. Mercier, Many thanks for this effort. It caught my attention because we might have a use case where this feature can be useful for us. Hence I'd like to jump and be part of the discussion, I'd really appreciate if you can cc'me for next versions. While reading the full patchset I was a bit confused about the status of this proposal. In fact, the rfc in the subject combined with the number of iterations (already seven) confused me. So I'm wondering if this is a RFC or a 'real' proposal already that you want to land. If this is still a RFC I'd remove the 'rfc: Proposal' and use the more canonical way that is put RFC in the []. I.e [PATCH RFC v7] cgroup: Add a GPU cgroup controller. If it is not, I'd just remove the RFC and make the subject in the cgroup subsystem instead of the gpu. I.E [PATCH v7] cgroup: Add a GPU cgroup I don't want to nitpick but IMO that helps new people to join to the history of the patchset. > This patch adds a proposal for a new GPU cgroup controller for > accounting/limiting GPU and GPU-related memory allocations. As far as I can see the only thing that is adding here is the accounting, so I'd remove any reference to limiting and just explain what the patch really introduces, not the future, otherwise is confusing an you expect more than the patch really does. It is important maintain the commit message sync with what the patch really does. > The proposed controller is based on the DRM cgroup controller[1] and > follows the design of the RDMA cgroup controller. > > The new cgroup controller would: > * Allow setting per-device limits on the total size of buffers > allocated by device within a cgroup. > * Expose a per-device/allocator breakdown of the buffers charged to a > cgroup. > > The prototype in the following patches is only for memory accounting > using the GPU cgroup controller and does not implement limit setting. > > [1]: https://lore.kernel.org/amd-gfx/20210126214626.16260-1-brian.welty@intel.com/ > I think this is material for the cover more than the commit message. When I read this I was expecting all this in this patch. > Signed-off-by: Hridya Valsaraju <hridya@google.com> > Signed-off-by: T.J. Mercier <tjmercier@google.com> > --- > v7 changes > Remove comment about duplicate name rejection which is not relevant to > cgroups users per Michal Koutný. > > v6 changes > Move documentation into cgroup-v2.rst per Tejun Heo. > > v5 changes > Drop the global GPU cgroup "total" (sum of all device totals) portion > of the design since there is no currently known use for this per > Tejun Heo. > > Update for renamed functions/variables. > > v3 changes > Remove Upstreaming Plan from gpu-cgroup.rst per John Stultz. > > Use more common dual author commit message format per John Stultz. > --- > Documentation/admin-guide/cgroup-v2.rst | 23 +++++++++++++++++++++++ > 1 file changed, 23 insertions(+) > > diff --git a/Documentation/admin-guide/cgroup-v2.rst b/Documentation/admin-guide/cgroup-v2.rst > index 69d7a6983f78..2e1d26e327c7 100644 > --- a/Documentation/admin-guide/cgroup-v2.rst > +++ b/Documentation/admin-guide/cgroup-v2.rst > @@ -2352,6 +2352,29 @@ first, and stays charged to that cgroup until that resource is freed. Migrating > a process to a different cgroup does not move the charge to the destination > cgroup where the process has moved. > > + > +GPU > +--- > + > +The GPU controller accounts for device and system memory allocated by the GPU > +and related subsystems for graphics use. Resource limits are not currently > +supported. > + > +GPU Interface Files > +~~~~~~~~~~~~~~~~~~~~ > + > + gpu.memory.current > + A read-only file containing memory allocations in flat-keyed format. The key > + is a string representing the device name. The value is the size of the memory > + charged to the device in bytes. The device names are globally unique.:: > + > + $ cat /sys/kernel/fs/cgroup1/gpu.memory.current I think this is outdated, you are using cgroup v2, right? > + dev1 4194304 > + dev2 104857600 > + When I applied the full series I was expecting see the memory allocated by the gpu devices or users of the gpu in this file but, after some experiments, what I saw is the memory allocated via any process that uses the dma-buf heap API (not necessary gpu users). For example, if you create a small program that allocates some memory via the dma-buf heap API and then you cat the gpu.memory.current file, you see that the memory accounted is not related to the gpu. This is really confusing, looks to me that the patches evolved to account memory that is not really related to the GPU but allocated vi the dma-buf heap API. IMO the name of the file should be according to what really does to avoid confusions. So, is this patchset meant to be GPU specific? If the answer is yes that's good but that's not what I experienced. I'm missing something? If the answer is that evolved to track dma-buf heap allocations I think all the patches need some rework to adapt the wording as right now, the gpu wording seems confusing to me. > + The device name string is set by a device driver when it registers with the > + GPU cgroup controller to participate in resource accounting. > + > Others > ------ > > Thanks, Enric
Hello, On Tue, May 17, 2022 at 04:30:29PM -0700, T.J. Mercier wrote: > Thanks for your suggestion. This almost works. "dmabuf" as a key could > work, but I'd actually like to account for each heap. Since heaps can > be dynamically added, I can't accommodate every potential heap name by > hardcoding registrations in the misc controller. On its own, that's a pretty weak reason to be adding a separate gpu controller especially given that it doesn't really seem to be one with proper abstractions for gpu resources. We don't want to keep adding random keys to misc controller but can definitely add limited flexibility. What kind of keys do you need? Thanks.
On Fri, May 20, 2022 at 12:47 AM Tejun Heo <tj@kernel.org> wrote: > > Hello, > > On Tue, May 17, 2022 at 04:30:29PM -0700, T.J. Mercier wrote: > > Thanks for your suggestion. This almost works. "dmabuf" as a key could > > work, but I'd actually like to account for each heap. Since heaps can > > be dynamically added, I can't accommodate every potential heap name by > > hardcoding registrations in the misc controller. > > On its own, that's a pretty weak reason to be adding a separate gpu > controller especially given that it doesn't really seem to be one with > proper abstractions for gpu resources. We don't want to keep adding random > keys to misc controller but can definitely add limited flexibility. What > kind of keys do you need? > Well the dmabuf-from-heaps component of this is the initial use case. I was envisioning we'd have additional keys as discussed here: https://lore.kernel.org/lkml/20220328035951.1817417-1-tjmercier@google.com/T/#m82e5fe9d8674bb60160701e52dae4356fea2ddfa So we'd end up with a well-defined core set of keys like "system", and then drivers would be free to use their own keys for their own unique purposes which could be complementary or orthogonal to the core set. Yesterday I was talking with someone who is interested in limiting gpu cores and bus IDs in addition to gpu memory. How to define core keys is the part where it looks like there's trouble. For my use case it would be sufficient to have current and maximum values for an arbitrary number of keys - one per heap. So the only part missing from the misc controller (for my use case) is the ability to register a new key at runtime as heaps are added. Instead of keeping track of resources with enum misc_res_type, requesting a resource handle/ID from the misc controller at runtime is what I think would be required instead. > Thanks. > > -- > tejun
On Thu, May 19, 2022 at 3:53 AM <eballetbo@kernel.org> wrote: > > From: Enric Balletbo i Serra <eballetbo@kernel.org> > > On Tue, 10 May 2022 23:56:46 +0000, T.J. Mercier wrote > > From: Hridya Valsaraju <hridya@google.com> > > > > The cgroup controller provides accounting for GPU and GPU-related > > memory allocations. The memory being accounted can be device memory or > > memory allocated from pools dedicated to serve GPU-related tasks. > > > > This patch adds APIs to: > > -allow a device to register for memory accounting using the GPU cgroup > > controller. > > -charge and uncharge allocated memory to a cgroup. > > > > When the cgroup controller is enabled, it would expose information about > > the memory allocated by each device(registered for GPU cgroup memory > > accounting) for each cgroup. > > > > The API/UAPI can be extended to set per-device/total allocation limits > > in the future. > > > > The cgroup controller has been named following the discussion in [1]. > > > > [1]: https://lore.kernel.org/amd-gfx/YCJp%2F%2FkMC7YjVMXv@phenom.ffwll.local/ > > > > Signed-off-by: Hridya Valsaraju <hridya@google.com> > > Signed-off-by: T.J. Mercier <tjmercier@google.com> > > --- > > v7 changes > > Hide gpucg and gpucg_bucket struct definitions per Michal Koutný. > > This means gpucg_register_bucket now returns an internally allocated > > struct gpucg_bucket. > > > > Move all public function documentation to the cgroup_gpu.h header. > > > > v5 changes > > Support all strings for gpucg_register_device instead of just string > > literals. > > > > Enforce globally unique gpucg_bucket names. > > > > Constrain gpucg_bucket name lengths to 64 bytes. > > > > Obtain just a single css refcount instead of nr_pages for each > > charge. > > > > Rename: > > gpucg_try_charge -> gpucg_charge > > find_cg_rpool_locked -> cg_rpool_find_locked > > init_cg_rpool -> cg_rpool_init > > get_cg_rpool_locked -> cg_rpool_get_locked > > "gpu cgroup controller" -> "GPU controller" > > gpucg_device -> gpucg_bucket > > usage -> size > > > > v4 changes > > Adjust gpucg_try_charge critical section for future charge transfer > > functionality. > > > > v3 changes > > Use more common dual author commit message format per John Stultz. > > > > v2 changes > > Fix incorrect Kconfig help section indentation per Randy Dunlap. > > --- > > include/linux/cgroup_gpu.h | 122 ++++++++++++ > > include/linux/cgroup_subsys.h | 4 + > > init/Kconfig | 7 + > > kernel/cgroup/Makefile | 1 + > > kernel/cgroup/gpu.c | 339 ++++++++++++++++++++++++++++++++++ > > 5 files changed, 473 insertions(+) > > create mode 100644 include/linux/cgroup_gpu.h > > create mode 100644 kernel/cgroup/gpu.c > > > > diff --git a/include/linux/cgroup_gpu.h b/include/linux/cgroup_gpu.h > > new file mode 100644 > > index 000000000000..cb228a16aa1f > > --- /dev/null > > +++ b/include/linux/cgroup_gpu.h > > @@ -0,0 +1,122 @@ > > +/* SPDX-License-Identifier: MIT > > + * Copyright 2019 Advanced Micro Devices, Inc. > > + * Copyright (C) 2022 Google LLC. > > + */ > > +#ifndef _CGROUP_GPU_H > > +#define _CGROUP_GPU_H > > + > > +#include <linux/cgroup.h> > > + > > +#define GPUCG_BUCKET_NAME_MAX_LEN 64 > > + > > +struct gpucg; > > +struct gpucg_bucket; > > + > > +#ifdef CONFIG_CGROUP_GPU > > + > > +/** > > + * css_to_gpucg - get the corresponding gpucg ref from a cgroup_subsys_state > > + * @css: the target cgroup_subsys_state > > + * > > + * Returns: gpu cgroup that contains the @css > > + */ > > +struct gpucg *css_to_gpucg(struct cgroup_subsys_state *css); > > + > > +/** > > + * gpucg_get - get the gpucg reference that a task belongs to > > + * @task: the target task > > + * > > + * This increases the reference count of the css that the @task belongs to. > > + * > > + * Returns: reference to the gpu cgroup the task belongs to. > > + */ > > +struct gpucg *gpucg_get(struct task_struct *task); > > + > > +/** > > + * gpucg_put - put a gpucg reference > > + * @gpucg: the target gpucg > > + * > > + * Put a reference obtained via gpucg_get > > + */ > > +void gpucg_put(struct gpucg *gpucg); > > + > > +/** > > + * gpucg_parent - find the parent of a gpu cgroup > > + * @cg: the target gpucg > > + * > > + * This does not increase the reference count of the parent cgroup > > + * > > + * Returns: parent gpu cgroup of @cg > > + */ > > +struct gpucg *gpucg_parent(struct gpucg *cg); > > + > > +/** > > + * gpucg_charge - charge memory to the specified gpucg and gpucg_bucket. > > + * Caller must hold a reference to @gpucg obtained through gpucg_get(). The size of the memory is > > + * rounded up to be a multiple of the page size. > > + * > > + * @gpucg: The gpu cgroup to charge the memory to. > > + * @bucket: The bucket to charge the memory to. > > + * @size: The size of memory to charge in bytes. > > + * This size will be rounded up to the nearest page size. > > + * > > + * Return: returns 0 if the charging is successful and otherwise returns an error code. > > + */ > > +int gpucg_charge(struct gpucg *gpucg, struct gpucg_bucket *bucket, u64 size); > > + > > +/** > > + * gpucg_uncharge - uncharge memory from the specified gpucg and gpucg_bucket. > > + * The caller must hold a reference to @gpucg obtained through gpucg_get(). > > + * > > + * @gpucg: The gpu cgroup to uncharge the memory from. > > + * @bucket: The bucket to uncharge the memory from. > > + * @size: The size of memory to uncharge in bytes. > > + * This size will be rounded up to the nearest page size. > > + */ > > +void gpucg_uncharge(struct gpucg *gpucg, struct gpucg_bucket *bucket, u64 size); > > + > > +/** > > + * gpucg_register_bucket - Registers a bucket for memory accounting using the GPU cgroup controller. > > + * > > + * @name: Pointer to a null-terminated string to denote the name of the bucket. This name should be > > + * globally unique, and should not exceed @GPUCG_BUCKET_NAME_MAX_LEN bytes. > > + * > > + * @bucket must remain valid. @name will be copied. > > + * > > + * Returns a pointer to a newly allocated bucket on success, or an errno code otherwise. As buckets > > + * cannot be unregistered, this can never be freed. > > + */ > > +struct gpucg_bucket *gpucg_register_bucket(const char *name); > > +#else /* CONFIG_CGROUP_GPU */ > > + > > +static inline struct gpucg *css_to_gpucg(struct cgroup_subsys_state *css) > > +{ > > + return NULL; > > +} > > + > > +static inline struct gpucg *gpucg_get(struct task_struct *task) > > +{ > > + return NULL; > > +} > > + > > +static inline void gpucg_put(struct gpucg *gpucg) {} > > + > > +static inline struct gpucg *gpucg_parent(struct gpucg *cg) > > +{ > > + return NULL; > > +} > > + > > +static inline int gpucg_charge(struct gpucg *gpucg, > > + struct gpucg_bucket *bucket, > > + u64 size) > > +{ > > + return 0; > > +} > > + > > +static inline void gpucg_uncharge(struct gpucg *gpucg, > > + struct gpucg_bucket *bucket, > > + u64 size) {} > > + > > +static inline struct gpucg_bucket *gpucg_register_bucket(const char *name) {} > > I think this needs to return NULL, otherwise you'll get a compiler error when > CONFIG_CGROUP_GPU is not set. > > I found other build errors when CONFIG_CGROUP_GPU is not set, please fix them in > the next versioon. > > Thanks, > Enric > Thanks. I have been building each patch with allnoconfig and allyesconfig before posting, but clearly this was not sufficient. I'll fix this up. > > +#endif /* CONFIG_CGROUP_GPU */ > > +#endif /* _CGROUP_GPU_H */ > > diff --git a/include/linux/cgroup_subsys.h b/include/linux/cgroup_subsys.h > > index 445235487230..46a2a7b93c41 100644 > > --- a/include/linux/cgroup_subsys.h > > +++ b/include/linux/cgroup_subsys.h > > @@ -65,6 +65,10 @@ SUBSYS(rdma) > > SUBSYS(misc) > > #endif > > > > +#if IS_ENABLED(CONFIG_CGROUP_GPU) > > +SUBSYS(gpu) > > +#endif > > + > > /* > > * The following subsystems are not supported on the default hierarchy. > > */ > > diff --git a/init/Kconfig b/init/Kconfig > > index ddcbefe535e9..2e00a190e170 100644 > > --- a/init/Kconfig > > +++ b/init/Kconfig > > @@ -984,6 +984,13 @@ config BLK_CGROUP > > > > See Documentation/admin-guide/cgroup-v1/blkio-controller.rst for more information. > > > > +config CGROUP_GPU > > + bool "GPU controller (EXPERIMENTAL)" > > + select PAGE_COUNTER > > + help > > + Provides accounting and limit setting for memory allocations by the GPU and > > + GPU-related subsystems. > > + > > config CGROUP_WRITEBACK > > bool > > depends on MEMCG && BLK_CGROUP > > diff --git a/kernel/cgroup/Makefile b/kernel/cgroup/Makefile > > index 12f8457ad1f9..be95a5a532fc 100644 > > --- a/kernel/cgroup/Makefile > > +++ b/kernel/cgroup/Makefile > > @@ -7,3 +7,4 @@ obj-$(CONFIG_CGROUP_RDMA) += rdma.o > > obj-$(CONFIG_CPUSETS) += cpuset.o > > obj-$(CONFIG_CGROUP_MISC) += misc.o > > obj-$(CONFIG_CGROUP_DEBUG) += debug.o > > +obj-$(CONFIG_CGROUP_GPU) += gpu.o > > diff --git a/kernel/cgroup/gpu.c b/kernel/cgroup/gpu.c > > new file mode 100644 > > index 000000000000..ad16ea15d427 > > --- /dev/null > > +++ b/kernel/cgroup/gpu.c > > @@ -0,0 +1,339 @@ > > +// SPDX-License-Identifier: MIT > > +// Copyright 2019 Advanced Micro Devices, Inc. > > +// Copyright (C) 2022 Google LLC. > > + > > +#include <linux/cgroup.h> > > +#include <linux/cgroup_gpu.h> > > +#include <linux/err.h> > > +#include <linux/gfp.h> > > +#include <linux/list.h> > > +#include <linux/mm.h> > > +#include <linux/page_counter.h> > > +#include <linux/seq_file.h> > > +#include <linux/slab.h> > > +#include <linux/string.h> > > + > > +static struct gpucg *root_gpucg __read_mostly; > > + > > +/* > > + * Protects list of resource pools maintained on per cgroup basis and list > > + * of buckets registered for memory accounting using the GPU cgroup controller. > > + */ > > +static DEFINE_MUTEX(gpucg_mutex); > > +static LIST_HEAD(gpucg_buckets); > > + > > +/* The GPU cgroup controller data structure */ > > +struct gpucg { > > + struct cgroup_subsys_state css; > > + > > + /* list of all resource pools that belong to this cgroup */ > > + struct list_head rpools; > > +}; > > + > > +/* A named entity representing bucket of tracked memory. */ > > +struct gpucg_bucket { > > + /* list of various resource pools in various cgroups that the bucket is part of */ > > + struct list_head rpools; > > + > > + /* list of all buckets registered for GPU cgroup accounting */ > > + struct list_head bucket_node; > > + > > + /* string to be used as identifier for accounting and limit setting */ > > + const char *name; > > +}; > > + > > +struct gpucg_resource_pool { > > + /* The bucket whose resource usage is tracked by this resource pool */ > > + struct gpucg_bucket *bucket; > > + > > + /* list of all resource pools for the cgroup */ > > + struct list_head cg_node; > > + > > + /* list maintained by the gpucg_bucket to keep track of its resource pools */ > > + struct list_head bucket_node; > > + > > + /* tracks memory usage of the resource pool */ > > + struct page_counter total; > > +}; > > + > > +static void free_cg_rpool_locked(struct gpucg_resource_pool *rpool) > > +{ > > + lockdep_assert_held(&gpucg_mutex); > > + > > + list_del(&rpool->cg_node); > > + list_del(&rpool->bucket_node); > > + kfree(rpool); > > +} > > + > > +static void gpucg_css_free(struct cgroup_subsys_state *css) > > +{ > > + struct gpucg_resource_pool *rpool, *tmp; > > + struct gpucg *gpucg = css_to_gpucg(css); > > + > > + // delete all resource pools > > + mutex_lock(&gpucg_mutex); > > + list_for_each_entry_safe(rpool, tmp, &gpucg->rpools, cg_node) > > + free_cg_rpool_locked(rpool); > > + mutex_unlock(&gpucg_mutex); > > + > > + kfree(gpucg); > > +} > > + > > +static struct cgroup_subsys_state * > > +gpucg_css_alloc(struct cgroup_subsys_state *parent_css) > > +{ > > + struct gpucg *gpucg, *parent; > > + > > + gpucg = kzalloc(sizeof(struct gpucg), GFP_KERNEL); > > + if (!gpucg) > > + return ERR_PTR(-ENOMEM); > > + > > + parent = css_to_gpucg(parent_css); > > + if (!parent) > > + root_gpucg = gpucg; > > + > > + INIT_LIST_HEAD(&gpucg->rpools); > > + > > + return &gpucg->css; > > +} > > + > > +static struct gpucg_resource_pool *cg_rpool_find_locked( > > + struct gpucg *cg, > > + struct gpucg_bucket *bucket) > > +{ > > + struct gpucg_resource_pool *rpool; > > + > > + lockdep_assert_held(&gpucg_mutex); > > + > > + list_for_each_entry(rpool, &cg->rpools, cg_node) > > + if (rpool->bucket == bucket) > > + return rpool; > > + > > + return NULL; > > +} > > + > > +static struct gpucg_resource_pool *cg_rpool_init(struct gpucg *cg, > > + struct gpucg_bucket *bucket) > > +{ > > + struct gpucg_resource_pool *rpool = kzalloc(sizeof(*rpool), > > + GFP_KERNEL); > > + if (!rpool) > > + return ERR_PTR(-ENOMEM); > > + > > + rpool->bucket = bucket; > > + > > + page_counter_init(&rpool->total, NULL); > > + INIT_LIST_HEAD(&rpool->cg_node); > > + INIT_LIST_HEAD(&rpool->bucket_node); > > + list_add_tail(&rpool->cg_node, &cg->rpools); > > + list_add_tail(&rpool->bucket_node, &bucket->rpools); > > + > > + return rpool; > > +} > > + > > +/** > > + * get_cg_rpool_locked - find the resource pool for the specified bucket and > > + * specified cgroup. If the resource pool does not exist for the cg, it is > > + * created in a hierarchical manner in the cgroup and its ancestor cgroups who > > + * do not already have a resource pool entry for the bucket. > > + * > > + * @cg: The cgroup to find the resource pool for. > > + * @bucket: The bucket associated with the returned resource pool. > > + * > > + * Return: return resource pool entry corresponding to the specified bucket in > > + * the specified cgroup (hierarchically creating them if not existing already). > > + * > > + */ > > +static struct gpucg_resource_pool * > > +cg_rpool_get_locked(struct gpucg *cg, struct gpucg_bucket *bucket) > > +{ > > + struct gpucg *parent_cg, *p, *stop_cg; > > + struct gpucg_resource_pool *rpool, *tmp_rpool; > > + struct gpucg_resource_pool *parent_rpool = NULL, *leaf_rpool = NULL; > > + > > + rpool = cg_rpool_find_locked(cg, bucket); > > + if (rpool) > > + return rpool; > > + > > + stop_cg = cg; > > + do { > > + rpool = cg_rpool_init(stop_cg, bucket); > > + if (IS_ERR(rpool)) > > + goto err; > > + > > + if (!leaf_rpool) > > + leaf_rpool = rpool; > > + > > + stop_cg = gpucg_parent(stop_cg); > > + if (!stop_cg) > > + break; > > + > > + rpool = cg_rpool_find_locked(stop_cg, bucket); > > + } while (!rpool); > > + > > + /* > > + * Re-initialize page counters of all rpools created in this invocation > > + * to enable hierarchical charging. > > + * stop_cg is the first ancestor cg who already had a resource pool for > > + * the bucket. It can also be NULL if no ancestors had a pre-existing > > + * resource pool for the bucket before this invocation. > > + */ > > + rpool = leaf_rpool; > > + for (p = cg; p != stop_cg; p = parent_cg) { > > + parent_cg = gpucg_parent(p); > > + if (!parent_cg) > > + break; > > + parent_rpool = cg_rpool_find_locked(parent_cg, bucket); > > + page_counter_init(&rpool->total, &parent_rpool->total); > > + > > + rpool = parent_rpool; > > + } > > + > > + return leaf_rpool; > > +err: > > + for (p = cg; p != stop_cg; p = gpucg_parent(p)) { > > + tmp_rpool = cg_rpool_find_locked(p, bucket); > > + free_cg_rpool_locked(tmp_rpool); > > + } > > + return rpool; > > +} > > + > > +struct gpucg *css_to_gpucg(struct cgroup_subsys_state *css) > > +{ > > + return css ? container_of(css, struct gpucg, css) : NULL; > > +} > > + > > +struct gpucg *gpucg_get(struct task_struct *task) > > +{ > > + if (!cgroup_subsys_enabled(gpu_cgrp_subsys)) > > + return NULL; > > + return css_to_gpucg(task_get_css(task, gpu_cgrp_id)); > > +} > > + > > +void gpucg_put(struct gpucg *gpucg) > > +{ > > + if (gpucg) > > + css_put(&gpucg->css); > > +} > > + > > +struct gpucg *gpucg_parent(struct gpucg *cg) > > +{ > > + return css_to_gpucg(cg->css.parent); > > +} > > + > > +int gpucg_charge(struct gpucg *gpucg, struct gpucg_bucket *bucket, u64 size) > > +{ > > + struct page_counter *counter; > > + u64 nr_pages; > > + struct gpucg_resource_pool *rp; > > + int ret = 0; > > + > > + nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; > > + > > + mutex_lock(&gpucg_mutex); > > + rp = cg_rpool_get_locked(gpucg, bucket); > > + /* > > + * Continue to hold gpucg_mutex because we use it to block charges while transfers are in > > + * progress to avoid potentially exceeding a limit. > > + */ > > + if (IS_ERR(rp)) { > > + mutex_unlock(&gpucg_mutex); > > + return PTR_ERR(rp); > > + } > > + > > + if (page_counter_try_charge(&rp->total, nr_pages, &counter)) > > + css_get(&gpucg->css); > > + else > > + ret = -ENOMEM; > > + mutex_unlock(&gpucg_mutex); > > + > > + return ret; > > +} > > + > > +void gpucg_uncharge(struct gpucg *gpucg, struct gpucg_bucket *bucket, u64 size) > > +{ > > + u64 nr_pages; > > + struct gpucg_resource_pool *rp; > > + > > + mutex_lock(&gpucg_mutex); > > + rp = cg_rpool_find_locked(gpucg, bucket); > > + /* > > + * gpucg_mutex can be unlocked here, rp will stay valid until gpucg is freed and there are > > + * active refs on gpucg. Uncharges are fine while transfers are in progress since there is > > + * no potential to exceed a limit while uncharging and transferring. > > + */ > > + mutex_unlock(&gpucg_mutex); > > + > > + if (unlikely(!rp)) { > > + pr_err("Resource pool not found, incorrect charge/uncharge ordering?\n"); > > + return; > > + } > > + > > + nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; > > + page_counter_uncharge(&rp->total, nr_pages); > > + css_put(&gpucg->css); > > +} > > + > > +struct gpucg_bucket *gpucg_register_bucket(const char *name) > > +{ > > + struct gpucg_bucket *bucket, *b; > > + > > + if (!name) > > + return ERR_PTR(-EINVAL); > > + > > + if (strlen(name) >= GPUCG_BUCKET_NAME_MAX_LEN) > > + return ERR_PTR(-ENAMETOOLONG); > > + > > + bucket = kzalloc(sizeof(struct gpucg_bucket), GFP_KERNEL); > > + if (!bucket) > > + return ERR_PTR(-ENOMEM); > > + > > + INIT_LIST_HEAD(&bucket->bucket_node); > > + INIT_LIST_HEAD(&bucket->rpools); > > + bucket->name = kstrdup_const(name, GFP_KERNEL); > > + > > + mutex_lock(&gpucg_mutex); > > + list_for_each_entry(b, &gpucg_buckets, bucket_node) { > > + if (strncmp(b->name, bucket->name, GPUCG_BUCKET_NAME_MAX_LEN) == 0) { > > + mutex_unlock(&gpucg_mutex); > > + kfree_const(bucket->name); > > + kfree(bucket); > > + return ERR_PTR(-EEXIST); > > + } > > + } > > + list_add_tail(&bucket->bucket_node, &gpucg_buckets); > > + mutex_unlock(&gpucg_mutex); > > + > > + return bucket; > > +} > > + > > +static int gpucg_resource_show(struct seq_file *sf, void *v) > > +{ > > + struct gpucg_resource_pool *rpool; > > + struct gpucg *cg = css_to_gpucg(seq_css(sf)); > > + > > + mutex_lock(&gpucg_mutex); > > + list_for_each_entry(rpool, &cg->rpools, cg_node) { > > + seq_printf(sf, "%s %lu\n", rpool->bucket->name, > > + page_counter_read(&rpool->total) * PAGE_SIZE); > > + } > > + mutex_unlock(&gpucg_mutex); > > + > > + return 0; > > +} > > + > > +struct cftype files[] = { > > + { > > + .name = "memory.current", > > + .seq_show = gpucg_resource_show, > > + }, > > + { } /* terminate */ > > +}; > > + > > +struct cgroup_subsys gpu_cgrp_subsys = { > > + .css_alloc = gpucg_css_alloc, > > + .css_free = gpucg_css_free, > > + .early_init = false, > > + .legacy_cftypes = files, > > + .dfl_cftypes = files, > > +}; > > > > -- > > 2.36.0.512.ge40c2bad7a-goog > > > > >
On Fri, Jun 24, 2022 at 1:17 PM Daniel Vetter <daniel@ffwll.ch> wrote: > > On Wed, Jun 15, 2022 at 10:31:21AM -0700, T.J. Mercier wrote: > > On Fri, May 20, 2022 at 9:25 AM T.J. Mercier <tjmercier@google.com> wrote: > > > > > > On Fri, May 20, 2022 at 12:47 AM Tejun Heo <tj@kernel.org> wrote: > > > > > > > > Hello, > > > > > > > > On Tue, May 17, 2022 at 04:30:29PM -0700, T.J. Mercier wrote: > > > > > Thanks for your suggestion. This almost works. "dmabuf" as a key could > > > > > work, but I'd actually like to account for each heap. Since heaps can > > > > > be dynamically added, I can't accommodate every potential heap name by > > > > > hardcoding registrations in the misc controller. > > > > > > > > On its own, that's a pretty weak reason to be adding a separate gpu > > > > controller especially given that it doesn't really seem to be one with > > > > proper abstractions for gpu resources. We don't want to keep adding random > > > > keys to misc controller but can definitely add limited flexibility. What > > > > kind of keys do you need? > > > > > > > Well the dmabuf-from-heaps component of this is the initial use case. > > > I was envisioning we'd have additional keys as discussed here: > > > https://lore.kernel.org/lkml/20220328035951.1817417-1-tjmercier@google.com/T/#m82e5fe9d8674bb60160701e52dae4356fea2ddfa > > > So we'd end up with a well-defined core set of keys like "system", and > > > then drivers would be free to use their own keys for their own unique > > > purposes which could be complementary or orthogonal to the core set. > > > Yesterday I was talking with someone who is interested in limiting gpu > > > cores and bus IDs in addition to gpu memory. How to define core keys > > > is the part where it looks like there's trouble. > > > > > > For my use case it would be sufficient to have current and maximum > > > values for an arbitrary number of keys - one per heap. So the only > > > part missing from the misc controller (for my use case) is the ability > > > to register a new key at runtime as heaps are added. Instead of > > > keeping track of resources with enum misc_res_type, requesting a > > > resource handle/ID from the misc controller at runtime is what I think > > > would be required instead. > > > > > Quick update: I'm going to make an attempt to modify the misc > > controller to support a limited amount of dynamic resource > > registration/tracking in place of the new controller in this series. > > > > Thanks everyone for the feedback. > > Somehow I missed this entire chain here. > > I'm not a fan, because I'm kinda hoping we could finally unify gpu memory > account. Atm everyone just adds their one-off solution in a random corner: > - total tracking in misc cgroup controller > - dma-buf sysfs files (except apparently too slow so it'll get deleted > again) > - random other stuff on open device files os OOM killer can see it > > This doesn't look good. But I also think one could see it as "gpu memory" is the drm subsystem doing the same thing (in that it's artificially narrow to gpus). It seems we need something to account for buffers allocated by drivers, no matter which subsystem it was in (drm, v4l2, or networking or whatever). thanks -john
On Fri, Jun 24, 2022 at 1:36 PM Daniel Vetter <daniel@ffwll.ch> wrote: > > On Fri, Jun 24, 2022 at 01:32:45PM -0700, John Stultz wrote: > > On Fri, Jun 24, 2022 at 1:17 PM Daniel Vetter <daniel@ffwll.ch> wrote: > > > > > > On Wed, Jun 15, 2022 at 10:31:21AM -0700, T.J. Mercier wrote: > > > > On Fri, May 20, 2022 at 9:25 AM T.J. Mercier <tjmercier@google.com> wrote: > > > > > > > > > > On Fri, May 20, 2022 at 12:47 AM Tejun Heo <tj@kernel.org> wrote: > > > > > > > > > > > > Hello, > > > > > > > > > > > > On Tue, May 17, 2022 at 04:30:29PM -0700, T.J. Mercier wrote: > > > > > > > Thanks for your suggestion. This almost works. "dmabuf" as a key could > > > > > > > work, but I'd actually like to account for each heap. Since heaps can > > > > > > > be dynamically added, I can't accommodate every potential heap name by > > > > > > > hardcoding registrations in the misc controller. > > > > > > > > > > > > On its own, that's a pretty weak reason to be adding a separate gpu > > > > > > controller especially given that it doesn't really seem to be one with > > > > > > proper abstractions for gpu resources. We don't want to keep adding random > > > > > > keys to misc controller but can definitely add limited flexibility. What > > > > > > kind of keys do you need? > > > > > > > > > > > Well the dmabuf-from-heaps component of this is the initial use case. > > > > > I was envisioning we'd have additional keys as discussed here: > > > > > https://lore.kernel.org/lkml/20220328035951.1817417-1-tjmercier@google.com/T/#m82e5fe9d8674bb60160701e52dae4356fea2ddfa > > > > > So we'd end up with a well-defined core set of keys like "system", and > > > > > then drivers would be free to use their own keys for their own unique > > > > > purposes which could be complementary or orthogonal to the core set. > > > > > Yesterday I was talking with someone who is interested in limiting gpu > > > > > cores and bus IDs in addition to gpu memory. How to define core keys > > > > > is the part where it looks like there's trouble. > > > > > > > > > > For my use case it would be sufficient to have current and maximum > > > > > values for an arbitrary number of keys - one per heap. So the only > > > > > part missing from the misc controller (for my use case) is the ability > > > > > to register a new key at runtime as heaps are added. Instead of > > > > > keeping track of resources with enum misc_res_type, requesting a > > > > > resource handle/ID from the misc controller at runtime is what I think > > > > > would be required instead. > > > > > > > > > Quick update: I'm going to make an attempt to modify the misc > > > > controller to support a limited amount of dynamic resource > > > > registration/tracking in place of the new controller in this series. > > > > > > > > Thanks everyone for the feedback. > > > > > > Somehow I missed this entire chain here. > > > > > > I'm not a fan, because I'm kinda hoping we could finally unify gpu memory > > > account. Atm everyone just adds their one-off solution in a random corner: > > > - total tracking in misc cgroup controller > > > - dma-buf sysfs files (except apparently too slow so it'll get deleted > > > again) > > > - random other stuff on open device files os OOM killer can see it > > > > > > This doesn't look good. > > > > But I also think one could see it as "gpu memory" is the drm subsystem > > doing the same thing (in that it's artificially narrow to gpus). It > > seems we need something to account for buffers allocated by drivers, > > no matter which subsystem it was in (drm, v4l2, or networking or > > whatever). > > This is what the gpucg was. It wasn't called the dmabuf cg because we want > to account also memory of other types (e.g. drm gem buffer objects which > aren't exported), and I guess people didn't dare call it an xpu. > > But this was absolutely for a lot more than just "gpu drivers in drm". > Better names welcome. > -Daniel > -- > Daniel Vetter > Software Engineer, Intel Corporation > http://blog.ffwll.ch
This patch series revisits the proposal for a GPU cgroup controller to track and limit memory allocations by various device/allocator subsystems. The patch series also contains a simple prototype to illustrate how Android intends to implement DMA-BUF allocator attribution using the GPU cgroup controller. The prototype does not include resource limit enforcements. Changelog: v7: Hide gpucg and gpucg_bucket struct definitions per Michal Koutný. This means gpucg_register_bucket now returns an internally allocated struct gpucg_bucket. Move all public function documentation to the cgroup_gpu.h header. Remove comment in documentation about duplicate name rejection which is not relevant to cgroups users per Michal Koutný. v6: Move documentation into cgroup-v2.rst per Tejun Heo. Rename BINDER_FD{A}_FLAG_SENDER_NO_NEED -> BINDER_FD{A}_FLAG_XFER_CHARGE per Carlos Llamas. Return error on transfer failure per Carlos Llamas. v5: Rebase on top of v5.18-rc3 Drop the global GPU cgroup "total" (sum of all device totals) portion of the design since there is no currently known use for this per Tejun Heo. Fix commit message which still contained the old name for dma_buf_transfer_charge per Michal Koutný. Remove all GPU cgroup code except what's necessary to support charge transfer from dma_buf. Previously charging was done in export, but for non-Android graphics use-cases this is not ideal since there may be a delay between allocation and export, during which time there is no accounting. Merge dmabuf: Use the GPU cgroup charge/uncharge APIs patch into dmabuf: heaps: export system_heap buffers with GPU cgroup charging as a result of above. Put the charge and uncharge code in the same file (system_heap_allocate, system_heap_dma_buf_release) instead of splitting them between the heap and the dma_buf_release. This avoids asymmetric management of the gpucg charges. Modify the dma_buf_transfer_charge API to accept a task_struct instead of a gpucg. This avoids requiring the caller to manage the refcount of the gpucg upon failure and confusing ownership transfer logic. Support all strings for gpucg_register_bucket instead of just string literals. Enforce globally unique gpucg_bucket names. Constrain gpucg_bucket name lengths to 64 bytes. Append "-heap" to gpucg_bucket names from dmabuf-heaps. Drop patch 7 from the series, which changed the types of binder_transaction_data's sender_pid and sender_euid fields. This was done in another commit here: https://lore.kernel.org/all/20220210021129.3386083-4-masahiroy@kernel.org/ Rename: gpucg_try_charge -> gpucg_charge find_cg_rpool_locked -> cg_rpool_find_locked init_cg_rpool -> cg_rpool_init get_cg_rpool_locked -> cg_rpool_get_locked "gpu cgroup controller" -> "GPU controller" gpucg_device -> gpucg_bucket usage -> size Tests: Support both binder_fd_array_object and binder_fd_object. This is necessary because new versions of Android will use binder_fd_object instead of binder_fd_array_object, and we need to support both. Tests for both binder_fd_array_object and binder_fd_object. For binder_utils return error codes instead of struct binder{fs}_ctx. Use ifdef __ANDROID__ to choose platform-dependent temp path instead of a runtime fallback. Ensure binderfs_mntpt ends with a trailing '/' character instead of prepending it where used. v4: Skip test if not run as root per Shuah Khan Add better test logging for abnormal child termination per Shuah Khan Adjust ordering of charge/uncharge during transfer to avoid potentially hitting cgroup limit per Michal Koutný Adjust gpucg_try_charge critical section for charge transfer functionality Fix uninitialized return code error for dmabuf_try_charge error case v3: Remove Upstreaming Plan from gpu-cgroup.rst per John Stultz Use more common dual author commit message format per John Stultz Remove android from binder changes title per Todd Kjos Add a kselftest for this new behavior per Greg Kroah-Hartman Include details on behavior for all combinations of kernel/userspace versions in changelog (thanks Suren Baghdasaryan) per Greg Kroah-Hartman. Fix pid and uid types in binder UAPI header v2: See the previous revision of this change submitted by Hridya Valsaraju at: https://lore.kernel.org/all/20220115010622.3185921-1-hridya@google.com/ Move dma-buf cgroup charge transfer from a dma_buf_op defined by every heap to a single dma-buf function for all heaps per Daniel Vetter and Christian König. Pointers to struct gpucg and struct gpucg_device tracking the current associations were added to the dma_buf struct to achieve this. Fix incorrect Kconfig help section indentation per Randy Dunlap. History of the GPU cgroup controller ==================================== The GPU/DRM cgroup controller came into being when a consensus[1] was reached that the resources it tracked were unsuitable to be integrated into memcg. Originally, the proposed controller was specific to the DRM subsystem and was intended to track GEM buffers and GPU-specific resources[2]. In order to help establish a unified memory accounting model for all GPU and all related subsystems, Daniel Vetter put forth a suggestion to move it out of the DRM subsystem so that it can be used by other DMA-BUF exporters as well[3]. This RFC proposes an interface that does the same. [1]: https://patchwork.kernel.org/project/dri-devel/cover/20190501140438.9506-1-brian.welty@intel.com/#22624705 [2]: https://lore.kernel.org/amd-gfx/20210126214626.16260-1-brian.welty@intel.com/ [3]: https://lore.kernel.org/amd-gfx/YCVOl8%2F87bqRSQei@phenom.ffwll.local/ Hridya Valsaraju (3): gpu: rfc: Proposal for a GPU cgroup controller cgroup: gpu: Add a cgroup controller for allocator attribution of GPU memory binder: Add flags to relinquish ownership of fds T.J. Mercier (3): dmabuf: heaps: export system_heap buffers with GPU cgroup charging dmabuf: Add gpu cgroup charge transfer function selftests: Add binder cgroup gpu memory transfer tests Documentation/admin-guide/cgroup-v2.rst | 23 + drivers/android/binder.c | 31 +- drivers/dma-buf/dma-buf.c | 80 ++- drivers/dma-buf/dma-heap.c | 38 ++ drivers/dma-buf/heaps/system_heap.c | 28 +- include/linux/cgroup_gpu.h | 146 +++++ include/linux/cgroup_subsys.h | 4 + include/linux/dma-buf.h | 49 +- include/linux/dma-heap.h | 15 + include/uapi/linux/android/binder.h | 23 +- init/Kconfig | 7 + kernel/cgroup/Makefile | 1 + kernel/cgroup/gpu.c | 390 +++++++++++++ .../selftests/drivers/android/binder/Makefile | 8 + .../drivers/android/binder/binder_util.c | 250 +++++++++ .../drivers/android/binder/binder_util.h | 32 ++ .../selftests/drivers/android/binder/config | 4 + .../binder/test_dmabuf_cgroup_transfer.c | 526 ++++++++++++++++++ 18 files changed, 1632 insertions(+), 23 deletions(-) create mode 100644 include/linux/cgroup_gpu.h create mode 100644 kernel/cgroup/gpu.c create mode 100644 tools/testing/selftests/drivers/android/binder/Makefile create mode 100644 tools/testing/selftests/drivers/android/binder/binder_util.c create mode 100644 tools/testing/selftests/drivers/android/binder/binder_util.h create mode 100644 tools/testing/selftests/drivers/android/binder/config create mode 100644 tools/testing/selftests/drivers/android/binder/test_dmabuf_cgroup_transfer.c