On Wed, Jan 20, 2021 at 10:52:11AM -0800, Yiwei Zhang wrote:
On Wed, Jan 20, 2021 at 5:33 AM Gerd Hoffmann kraxel@redhat.com wrote:
Hi,
> + select TRACE_GPU_MEM
> +#ifdef CONFIG_TRACE_GPU_MEM
That doesn't make sense btw.
Do you recommend we just select it or leave it an option?
The patch selects it (which makes sense given the small size). The #ifdef is pointless then ...
> +#ifdef CONFIG_TRACE_GPU_MEM > +static inline void virtio_gpu_trace_total_mem(struct virtio_gpu_device *vgdev, > + s64 delta) > +{ > + u64 total_mem = atomic64_add_return(delta, &vgdev->total_mem); > + > + trace_gpu_mem_total(0, 0, total_mem);
Hmm, so no per process tracking (pid arg hard-coded to zero)? Any plans for that? The cgroups patches mentioned by Daniel should address that btw.
Android GPU vendors do report the totals for each process as well. For Cuttlefish virtual platform, we haven't yet required that, and want to get the global total in place first.
That means no plans yet?
Android relies on this tracepoint + eBPF to make the GPU memory totals available at runtime on production devices, which has been enforced already. Not only game developers can have a reliable kernel total GPU memory to look at, but also Android leverages this to take GPU memory usage out from the system lost ram.
Sounds like you define "gpu memory" as "system memory used to store gpu data". Is that correct? What about device memory?
The total definition does include all device memory being used as well for numa devices.(If my understanding of your question is correct.)
device memory == gpu-owned memory, typically exposed to as pci memory bar.
qemu stdvga for example stores gem objects in device memory (unless it runs out of vram, then ttm allocates from / moves into main memory).
I'm not sure whether the other DRM drivers would like to integrate this tracepoint(maybe upstream drivers will move away from debugfs later as well?), but at least we hope virtio-gpu can take this.
Well, it is basically the same for all drivers using the gem shmem helpers. So I see little reason why we should do that at virtio-gpu level.
This can be a starting point. Another reason would be I'm fearing that this tracepoint approach might be more difficult to get upstreamed at drm layer level, since later we may want to get to per-process total tracking, which would be making more sense at device driver level.
Tracking in __drm_gem_shmem_create + drm_gem_shmem_free_object should give you pretty much the same results, with the major difference being that it works for all shmem-based drivers.
Of course just moving the trace points doesn't solve the other issues discussed.
Android GPU vendors have integrated this tracepoint to track gpu memory usage total(mapped into the gpu address space), which consists of below: (1) directly allocated via physical page allocator (2) imported external memory backed by dma-bufs (3) allocated exportable memory backed by dma-bufs
Hmm, the tracepoint doesn't track which of the three groups the memory belongs to. Which I think is important, specifically group (2) because that might already be accounted for by the exporting driver ...
The tracepoint only cares about a total number, but I'm not against the idea to extend the tracepoint with categorization. However, I believe the dma-bufs core can track which dma-buf gets attached/mapped to some devices. So that those overlap between dma-buf heaps and the gpu memory total we are tracking here can be canceled out.
Yep, maybe. Which is *exactly* why Daniel keeps asking for the big picture and how this integrates/interacts with the dma-buf accounting which seems to be in the works too.
Note that dma-bufs are not only used for cross-device sharing. They are also used to pass handles from one application to another (application to wayland compositor or x server for example). Which doesn't matter much for the totals, but for per-process accounting you need a plan how to account these shared buffers.
I suspect once you figured that you'll notice that this little hack is rather incomplete.
Despite the dma-buf side effort, we still wish to have this tracepoint integrated in virtio-gpu just for a global total at this moment.
I don't feel like merging patches with obvious shortcomings which have a high chance to end up as technical dept.
The question how this interacts with dma-buf accounting must be clarified.
I'd also suggest to join forces with the cgroups people. The problem space has alot of overlap. Even if we end up with multiple ways to export the accounting data the spots you have to hook into to actually do the accounting should be largely identical.
take care, Gerd
From: Yiwei Zhang zzyiwei@google.com
On the success of virtio_gpu_object_create, add size of newly allocated bo to the tracked total_mem. In drm_gem_object_funcs.free, after the gem bo lost its last refcount, subtract the bo size from the tracked total_mem if the original underlying memory allocation is successful.
It's more accurate to do this in device driver layer to best match when the underlying resource gets allocated and destroyed during tracing.
Signed-off-by: Yiwei Zhang zzyiwei@android.com --- drivers/gpu/drm/virtio/Kconfig | 1 + drivers/gpu/drm/virtio/virtgpu_drv.h | 2 ++ drivers/gpu/drm/virtio/virtgpu_object.c | 11 +++++++++++ 3 files changed, 14 insertions(+)
diff --git a/drivers/gpu/drm/virtio/Kconfig b/drivers/gpu/drm/virtio/Kconfig index b925b8b1da16..e103b7e883b1 100644 --- a/drivers/gpu/drm/virtio/Kconfig +++ b/drivers/gpu/drm/virtio/Kconfig @@ -5,6 +5,7 @@ config DRM_VIRTIO_GPU select DRM_KMS_HELPER select DRM_GEM_SHMEM_HELPER select VIRTIO_DMA_SHARED_BUFFER + select TRACE_GPU_MEM help This is the virtual GPU driver for virtio. It can be used with QEMU based VMMs (like KVM or Xen). diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h index 6a232553c99b..7ab63ce9c6a9 100644 --- a/drivers/gpu/drm/virtio/virtgpu_drv.h +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h @@ -249,6 +249,8 @@ struct virtio_gpu_device { spinlock_t resource_export_lock; /* protects map state and host_visible_mm */ spinlock_t host_visible_lock; + /* total memory backing gem bos */ + atomic64_t total_mem; };
struct virtio_gpu_fpriv { diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c index d69a5b6da553..e2251fc41509 100644 --- a/drivers/gpu/drm/virtio/virtgpu_object.c +++ b/drivers/gpu/drm/virtio/virtgpu_object.c @@ -25,12 +25,21 @@
#include <linux/dma-mapping.h> #include <linux/moduleparam.h> +#include <trace/events/gpu_mem.h>
#include "virtgpu_drv.h"
static int virtio_gpu_virglrenderer_workaround = 1; module_param_named(virglhack, virtio_gpu_virglrenderer_workaround, int, 0400);
+static inline void virtio_gpu_trace_total_mem(struct virtio_gpu_device *vgdev, + s64 delta) +{ + u64 total_mem = atomic64_add_return(delta, &vgdev->total_mem); + + trace_gpu_mem_total(vgdev->ddev->primary->index, 0, total_mem); +} + int virtio_gpu_resource_id_get(struct virtio_gpu_device *vgdev, uint32_t *resid) { if (virtio_gpu_virglrenderer_workaround) { @@ -104,6 +113,7 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) struct virtio_gpu_device *vgdev = bo->base.base.dev->dev_private;
if (bo->created) { + virtio_gpu_trace_total_mem(vgdev, -(obj->size)); virtio_gpu_cmd_unref_resource(vgdev, bo); virtio_gpu_notify(vgdev); /* completion handler calls virtio_gpu_cleanup_object() */ @@ -265,6 +275,7 @@ int virtio_gpu_object_create(struct virtio_gpu_device *vgdev, virtio_gpu_object_attach(vgdev, bo, ents, nents); }
+ virtio_gpu_trace_total_mem(vgdev, shmem_obj->base.size); *bo_ptr = bo; return 0;
On the success of virtio_gpu_object_create, add size of newly allocated bo to the tracked total_mem. In drm_gem_object_funcs.free, after the gem bo loses its last refcount, subtract the bo size from the tracked total_mem if the original underlying memory allocation is successful.
It's more accurate to do this in device driver layer to best match when the underlying resource gets allocated and destroyed during tracing.
Signed-off-by: Yiwei Zhang zzyiwei@android.com --- drivers/gpu/drm/virtio/Kconfig | 1 + drivers/gpu/drm/virtio/virtgpu_drv.h | 2 ++ drivers/gpu/drm/virtio/virtgpu_object.c | 11 +++++++++++ 3 files changed, 14 insertions(+)
diff --git a/drivers/gpu/drm/virtio/Kconfig b/drivers/gpu/drm/virtio/Kconfig index b925b8b1da16..e103b7e883b1 100644 --- a/drivers/gpu/drm/virtio/Kconfig +++ b/drivers/gpu/drm/virtio/Kconfig @@ -5,6 +5,7 @@ config DRM_VIRTIO_GPU select DRM_KMS_HELPER select DRM_GEM_SHMEM_HELPER select VIRTIO_DMA_SHARED_BUFFER + select TRACE_GPU_MEM help This is the virtual GPU driver for virtio. It can be used with QEMU based VMMs (like KVM or Xen). diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h index 6a232553c99b..c5622f9b591f 100644 --- a/drivers/gpu/drm/virtio/virtgpu_drv.h +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h @@ -249,6 +249,8 @@ struct virtio_gpu_device { spinlock_t resource_export_lock; /* protects map state and host_visible_mm */ spinlock_t host_visible_lock; + + atomic64_t total_mem; };
struct virtio_gpu_fpriv { diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c index d69a5b6da553..e2251fc41509 100644 --- a/drivers/gpu/drm/virtio/virtgpu_object.c +++ b/drivers/gpu/drm/virtio/virtgpu_object.c @@ -25,12 +25,21 @@
#include <linux/dma-mapping.h> #include <linux/moduleparam.h> +#include <trace/events/gpu_mem.h>
#include "virtgpu_drv.h"
static int virtio_gpu_virglrenderer_workaround = 1; module_param_named(virglhack, virtio_gpu_virglrenderer_workaround, int, 0400);
+static inline void virtio_gpu_trace_total_mem(struct virtio_gpu_device *vgdev, + s64 delta) +{ + u64 total_mem = atomic64_add_return(delta, &vgdev->total_mem); + + trace_gpu_mem_total(vgdev->ddev->primary->index, 0, total_mem); +} + int virtio_gpu_resource_id_get(struct virtio_gpu_device *vgdev, uint32_t *resid) { if (virtio_gpu_virglrenderer_workaround) { @@ -104,6 +113,7 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) struct virtio_gpu_device *vgdev = bo->base.base.dev->dev_private;
if (bo->created) { + virtio_gpu_trace_total_mem(vgdev, -(obj->size)); virtio_gpu_cmd_unref_resource(vgdev, bo); virtio_gpu_notify(vgdev); /* completion handler calls virtio_gpu_cleanup_object() */ @@ -265,6 +275,7 @@ int virtio_gpu_object_create(struct virtio_gpu_device *vgdev, virtio_gpu_object_attach(vgdev, bo, ents, nents); }
+ virtio_gpu_trace_total_mem(vgdev, shmem_obj->base.size); *bo_ptr = bo; return 0;
On Thu, Jan 21, 2021 at 9:40 PM Yiwei Zhang zzyiwei@android.com wrote:
On the success of virtio_gpu_object_create, add size of newly allocated bo to the tracked total_mem. In drm_gem_object_funcs.free, after the gem bo loses its last refcount, subtract the bo size from the tracked total_mem if the original underlying memory allocation is successful.
It's more accurate to do this in device driver layer to best match when the underlying resource gets allocated and destroyed during tracing.
Signed-off-by: Yiwei Zhang zzyiwei@android.com
drivers/gpu/drm/virtio/Kconfig | 1 + drivers/gpu/drm/virtio/virtgpu_drv.h | 2 ++ drivers/gpu/drm/virtio/virtgpu_object.c | 11 +++++++++++ 3 files changed, 14 insertions(+)
diff --git a/drivers/gpu/drm/virtio/Kconfig b/drivers/gpu/drm/virtio/Kconfig index b925b8b1da16..e103b7e883b1 100644 --- a/drivers/gpu/drm/virtio/Kconfig +++ b/drivers/gpu/drm/virtio/Kconfig @@ -5,6 +5,7 @@ config DRM_VIRTIO_GPU select DRM_KMS_HELPER select DRM_GEM_SHMEM_HELPER select VIRTIO_DMA_SHARED_BUFFER
select TRACE_GPU_MEM help This is the virtual GPU driver for virtio. It can be used with QEMU based VMMs (like KVM or Xen).diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h index 6a232553c99b..c5622f9b591f 100644 --- a/drivers/gpu/drm/virtio/virtgpu_drv.h +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h @@ -249,6 +249,8 @@ struct virtio_gpu_device { spinlock_t resource_export_lock; /* protects map state and host_visible_mm */ spinlock_t host_visible_lock;
atomic64_t total_mem;};
struct virtio_gpu_fpriv { diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c index d69a5b6da553..e2251fc41509 100644 --- a/drivers/gpu/drm/virtio/virtgpu_object.c +++ b/drivers/gpu/drm/virtio/virtgpu_object.c @@ -25,12 +25,21 @@
#include <linux/dma-mapping.h> #include <linux/moduleparam.h> +#include <trace/events/gpu_mem.h>
#include "virtgpu_drv.h"
static int virtio_gpu_virglrenderer_workaround = 1; module_param_named(virglhack, virtio_gpu_virglrenderer_workaround, int, 0400);
+static inline void virtio_gpu_trace_total_mem(struct virtio_gpu_device *vgdev,
s64 delta)+{
u64 total_mem = atomic64_add_return(delta, &vgdev->total_mem);trace_gpu_mem_total(vgdev->ddev->primary->index, 0, total_mem);+}
int virtio_gpu_resource_id_get(struct virtio_gpu_device *vgdev, uint32_t *resid) { if (virtio_gpu_virglrenderer_workaround) { @@ -104,6 +113,7 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) struct virtio_gpu_device *vgdev = bo->base.base.dev->dev_private;
if (bo->created) {
virtio_gpu_trace_total_mem(vgdev, -(obj->size)); virtio_gpu_cmd_unref_resource(vgdev, bo); virtio_gpu_notify(vgdev); /* completion handler calls virtio_gpu_cleanup_object() */@@ -265,6 +275,7 @@ int virtio_gpu_object_create(struct virtio_gpu_device *vgdev, virtio_gpu_object_attach(vgdev, bo, ents, nents); }
virtio_gpu_trace_total_mem(vgdev, shmem_obj->base.size); *bo_ptr = bo; return 0;-- 2.30.0.280.ga3ce27912f-goog
Re Gerd and Daniel:
I'm not sure why we want to couple this patch too much with the dma-bufs tracking. The tracepoint added here itself is pretty useful for tracking gem bo total usage in virtio gpu upon tracing. The original purpose for integrating this tracepoint in all Android gpu kernel drivers is to just track total gpu memory usage and serve the accurate data to game developers in a much easier way. It's something they can rely on for robust testing and regression monitoring.
The only overlap with the dma-buf side is when we export a bo via prime to a dma-buf. But still, the total here is already useful for this particular device. Using which approach to account for the overlap wouldn't block this small integration from my understanding.
Besides, there's no plan for adding per-process gem total tracking in virtio-gpu at this moment. This patch should be light enough to carry without worrying about tech debt I believe.
Many thanks! Yiwei
On Thu, Jan 21, 2021 at 11:58:22PM -0800, Yiwei Zhang wrote:
On Thu, Jan 21, 2021 at 9:40 PM Yiwei Zhang zzyiwei@android.com wrote:
On the success of virtio_gpu_object_create, add size of newly allocated bo to the tracked total_mem. In drm_gem_object_funcs.free, after the gem bo loses its last refcount, subtract the bo size from the tracked total_mem if the original underlying memory allocation is successful.
It's more accurate to do this in device driver layer to best match when the underlying resource gets allocated and destroyed during tracing.
Signed-off-by: Yiwei Zhang zzyiwei@android.com
drivers/gpu/drm/virtio/Kconfig | 1 + drivers/gpu/drm/virtio/virtgpu_drv.h | 2 ++ drivers/gpu/drm/virtio/virtgpu_object.c | 11 +++++++++++ 3 files changed, 14 insertions(+)
diff --git a/drivers/gpu/drm/virtio/Kconfig b/drivers/gpu/drm/virtio/Kconfig index b925b8b1da16..e103b7e883b1 100644 --- a/drivers/gpu/drm/virtio/Kconfig +++ b/drivers/gpu/drm/virtio/Kconfig @@ -5,6 +5,7 @@ config DRM_VIRTIO_GPU select DRM_KMS_HELPER select DRM_GEM_SHMEM_HELPER select VIRTIO_DMA_SHARED_BUFFER
select TRACE_GPU_MEM help This is the virtual GPU driver for virtio. It can be used with QEMU based VMMs (like KVM or Xen).diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h b/drivers/gpu/drm/virtio/virtgpu_drv.h index 6a232553c99b..c5622f9b591f 100644 --- a/drivers/gpu/drm/virtio/virtgpu_drv.h +++ b/drivers/gpu/drm/virtio/virtgpu_drv.h @@ -249,6 +249,8 @@ struct virtio_gpu_device { spinlock_t resource_export_lock; /* protects map state and host_visible_mm */ spinlock_t host_visible_lock;
atomic64_t total_mem;};
struct virtio_gpu_fpriv { diff --git a/drivers/gpu/drm/virtio/virtgpu_object.c b/drivers/gpu/drm/virtio/virtgpu_object.c index d69a5b6da553..e2251fc41509 100644 --- a/drivers/gpu/drm/virtio/virtgpu_object.c +++ b/drivers/gpu/drm/virtio/virtgpu_object.c @@ -25,12 +25,21 @@
#include <linux/dma-mapping.h> #include <linux/moduleparam.h> +#include <trace/events/gpu_mem.h>
#include "virtgpu_drv.h"
static int virtio_gpu_virglrenderer_workaround = 1; module_param_named(virglhack, virtio_gpu_virglrenderer_workaround, int, 0400);
+static inline void virtio_gpu_trace_total_mem(struct virtio_gpu_device *vgdev,
s64 delta)+{
u64 total_mem = atomic64_add_return(delta, &vgdev->total_mem);trace_gpu_mem_total(vgdev->ddev->primary->index, 0, total_mem);+}
int virtio_gpu_resource_id_get(struct virtio_gpu_device *vgdev, uint32_t *resid) { if (virtio_gpu_virglrenderer_workaround) { @@ -104,6 +113,7 @@ static void virtio_gpu_free_object(struct drm_gem_object *obj) struct virtio_gpu_device *vgdev = bo->base.base.dev->dev_private;
if (bo->created) {
virtio_gpu_trace_total_mem(vgdev, -(obj->size)); virtio_gpu_cmd_unref_resource(vgdev, bo); virtio_gpu_notify(vgdev); /* completion handler calls virtio_gpu_cleanup_object() */@@ -265,6 +275,7 @@ int virtio_gpu_object_create(struct virtio_gpu_device *vgdev, virtio_gpu_object_attach(vgdev, bo, ents, nents); }
virtio_gpu_trace_total_mem(vgdev, shmem_obj->base.size); *bo_ptr = bo; return 0;-- 2.30.0.280.ga3ce27912f-goog
Re Gerd and Daniel:
I'm not sure why we want to couple this patch too much with the dma-bufs tracking. The tracepoint added here itself is pretty useful for tracking gem bo total usage in virtio gpu upon tracing. The original purpose for integrating this tracepoint in all Android gpu kernel drivers is to just track total gpu memory usage and serve the accurate data to game developers in a much easier way. It's something they can rely on for robust testing and regression monitoring.
The only overlap with the dma-buf side is when we export a bo via prime to a dma-buf. But still, the total here is already useful for this particular device. Using which approach to account for the overlap wouldn't block this small integration from my understanding.
Besides, there's no plan for adding per-process gem total tracking in virtio-gpu at this moment. This patch should be light enough to carry without worrying about tech debt I believe.
The tracepoint is clearly more generic than just what you implement here, to support the full use cases on Android's closed stacks. And it is uapi.
Tech debt isn't measured in lines of code, but in how expensive it's going to be to fix up the mess in the future. uapi is expensive no matter how few lines are used to implement it.
So yeah this needs to be properly thought out, properly implemented (not just on the virtual demo stack but something that looks like actual production stack), with open drivers, proper alignment with other efforts like tracking memory with cgroups, and the interactions with dma-buf tracking resolved, and igt testcases (this is meant to be generic after all), and at least solid proposals for rolling this out across the drm drivers, and ...
In other words, new uapi needs to be done right. -Daniel
dri-devel@lists.freedesktop.org
-
Daniel Vetter -
Gerd Hoffmann -
Yiwei Zhang -
Yiwei Zhang