mm/memcontrol drm/ttm: charge ttm buffer backed by system memory
Subsystems (6)
cover mm/memcontrol drm/ttm: charge ttm buffer backed by system memory
Buffers created by GPU driver could be huge (often several MB and even hundred or thousand MB). Some GPU driver call drm_gem_get_pages() which uses shmem to allocate these buffers which will charge memcg already, while some GPU driver like amdgpu use TTM which just allocate these system memory backed buffers with alloc_pages() so won't charge memcg currently.
Not like pure kernel memory, GPU buffer need to be mapped to user space for user filling data and command then let GPU hardware consume these buffers. So it is not proper to use memcg kmem by adding __GFP_ACCOUNT to alloc_pages gfp flags.
Another reason is back memory of GPU buffer may be allocated latter after the buffer object is created, and even in other processes. So we need to record the memcg when buffer object creation, then charge it latter when needed.
TTM will use a page pool acting as a cache for write-combine/no-cache pages. So adding new GFP flags for alloc_pages also does not work.
Qiang Yu (3):
mm: memcontrol: add mem_cgroup_(un)charge_drvmem
mm: memcontrol: record driver memory statistics drm/ttm: support memcg for ttm_tt
drivers/gpu/drm/ttm/ttm_bo.c | 10 +++++ drivers/gpu/drm/ttm/ttm_page_alloc.c | 18 ++++++++- drivers/gpu/drm/ttm/ttm_tt.c | 3 ++ include/drm/ttm/ttm_bo_api.h | 5 +++ include/drm/ttm/ttm_tt.h | 4 ++ include/linux/memcontrol.h | 22 +++++++++++ mm/memcontrol.c | 58 ++++++++++++++++++++++++++++ 7 files changed, 119 insertions(+), 1 deletion(-)
Signature · 1 line
2.17.1
RFC PATCH 1/3 mm: memcontrol: add mem_cgroup_(un)charge_drvmem
This is for driver which will allocate memory for both user application and kernel device driver usage. For example, GPU driver will allocate some GFP_USER pages and mapped to user to fill commands and data like texture and vertex, then let GPU command processor "eat" these memory. These buffers can be huge (offen several MB and may get to hundred or even thousand MB).
--- include/linux/memcontrol.h | 21 ++++++++++++++++ mm/memcontrol.c | 49 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 70 insertions(+)
include/linux/memcontrol.h +21 -0
@@ -1363,6 +1363,27 @@ static inline void memcg_set_shrinker_bit(struct mem_cgroup *memcg, }#endif +#ifdef CONFIG_MEMCG +struct mem_cgroup *mem_cgroup_driver_get_from_current (void ); +int mem_cgroup_charge_drvmem (struct mem_cgroup *memcg,gfp_t gfp, +unsigned long nr_pages); +void mem_cgroup_uncharge_drvmem (struct mem_cgroup *memcg,unsigned long nr_pages); +#else +static inline struct mem_cgroup *mem_cgroup_get_from_current (void ) +{ +return NULL ; +} + +static inline int mem_cgroup_charge_drvmem (struct mem_cgroup *memcg,gfp_t gfp, +unsigned long nr_pages) +{ +return 0 ; +} + +static inline void mem_cgroup_uncharge_drvmem (struct mem_cgroup *memcg, +unsigned long nr_pages) { } +#endif +struct kmem_cache *memcg_kmem_get_cache (struct kmem_cache *cachep);void memcg_kmem_put_cache (struct kmem_cache *cachep);
mm/memcontrol.c +49 -0
55 lines not on this page · raw message
Signature · 1 line
2.17.1
RFC PATCH 2/3 mm: memcontrol: record driver memory statistics
In reply to Qiang Yu--- include/linux/memcontrol.h | 1 + mm/memcontrol.c | 9 +++++++++ 2 files changed, 10 insertions(+)
include/linux/memcontrol.h +1 -0
@@ -36,6 +36,7 @@ enum memcg_stat_item {MEMCG_SOCK ,/* XXX: why are these zone and not node counters? */ MEMCG_KERNEL_STACK_KB , +MEMCG_DRV , MEMCG_NR_STAT, };
mm/memcontrol.c +9 -0
@@ -1413,6 +1413,9 @@ static char *memory_stat_format(struct mem_cgroup *memcg)seq_buf_printf (&s,"sock %llu\n" , (u64 )memcg_page_state (memcg,MEMCG_SOCK ) *PAGE_SIZE ); +seq_buf_printf (&s,"driver %llu\n" , + (u64 )memcg_page_state (memcg,MEMCG_DRV ) * +PAGE_SIZE ); seq_buf_printf(&s, "shmem %llu\n", (u64)memcg_page_state(memcg, NR_SHMEM) *
@@ -6947,6 +6950,9 @@ EXPORT_SYMBOL(mem_cgroup_driver_get_from_current);int mem_cgroup_charge_drvmem (struct mem_cgroup *memcg,gfp_t gfp,unsigned long nr_pages) { +if (cgroup_subsys_on_dfl (memory_cgrp_subsys)) +mod_memcg_state (memcg,MEMCG_DRV , nr_pages); +return try_charge (memcg, gfp, nr_pages); }EXPORT_SYMBOL (mem_cgroup_charge_drvmem);
@@ -6958,6 +6964,9 @@ EXPORT_SYMBOL(mem_cgroup_charge_drvmem);*/ void mem_cgroup_uncharge_drvmem (struct mem_cgroup *memcg,unsigned long nr_pages) { +if (cgroup_subsys_on_dfl (memory_cgrp_subsys)) +mod_memcg_state (memcg,MEMCG_DRV , -nr_pages); +refill_stock (memcg, nr_pages); }EXPORT_SYMBOL (mem_cgroup_uncharge_drvmem);
Signature · 1 line
2.17.1
RFC PATCH 3/3 drm/ttm: support memcg for ttm_tt
In reply to Qiang YuCharge TTM allocated system memory to memory cgroup which will limit the memory usage of a group of processes.
The memory is always charged to the control group of task which create this buffer object and when it's created. For example, when a buffer is created by process A and exported to process B, then process B populate this buffer, the memory is still charged to process A's memcg; if a buffer is created by process A when in memcg B, then A is moved to memcg C and populate this buffer, it will charge memcg B.
--- drivers/gpu/drm/ttm/ttm_bo.c | 10 ++++++++++ drivers/gpu/drm/ttm/ttm_page_alloc.c | 18 +++++++++++++++++- drivers/gpu/drm/ttm/ttm_tt.c | 3 +++ include/drm/ttm/ttm_bo_api.h | 5 +++++ include/drm/ttm/ttm_tt.h | 4 ++++ 5 files changed, 39 insertions(+), 1 deletion(-)
drivers/gpu/drm/ttm/ttm_bo.c +10 -0
@@ -42,6 +42,7 @@#include <linux/module.h> #include <linux/atomic.h> #include <linux/dma-resv.h> +#include <linux/memcontrol.h> static void ttm_bo_global_kobj_release (struct kobject *kobj);
@@ -162,6 +163,10 @@ static void ttm_bo_release_list(struct kref *list_kref)if (!ttm_bo_uses_embedded_gem_object (bo))dma_resv_fini (&bo->base._resv);mutex_destroy (&bo->wu_mutex); +#ifdef CONFIG_MEMCG +if (bo->memcg) +css_put (&bo->memcg->css); +#endif bo->destroy (bo);ttm_mem_global_free (&ttm_mem_glob, acc_size); }
@@ -1330,6 +1335,11 @@ int ttm_bo_init_reserved(struct ttm_bo_device *bdev, }atomic_inc (&ttm_bo_glob.bo_count); +#ifdef CONFIG_MEMCG +if (bo->type == ttm_bo_type_device) + bo->memcg =mem_cgroup_driver_get_from_current (); +#endif +/* * For ttm_bo_type_device buffers, allocate * address space from the device.
drivers/gpu/drm/ttm/ttm_page_alloc.c +17 -1
@@ -42,7 +42,7 @@#include <linux/seq_file.h> /* for seq_printf */ #include <linux/slab.h> #include <linux/dma-mapping.h> - +#include <linux/memcontrol.h> #include <linux/atomic.h> #include <drm/ttm/ttm_bo_driver.h>
@@ -1045,6 +1045,11 @@ ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)ttm_put_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state); ttm->state = tt_unpopulated; + +#ifdef CONFIG_MEMCG +if (ttm->memcg) +mem_cgroup_uncharge_drvmem (ttm->memcg, ttm->num_pages); +#endif }int ttm_pool_populate (struct ttm_tt *ttm,struct ttm_operation_ctx *ctx)
@@ -1059,6 +1064,17 @@ int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)if (ttm_check_under_lowerlimit (mem_glob, ttm->num_pages, ctx))return -ENOMEM ; +#ifdef CONFIG_MEMCG +if (ttm->memcg) { +gfp_t gfp_flags =GFP_USER ; +if (ttm->page_flags &TTM_PAGE_FLAG_NO_RETRY ) + gfp_flags |= __GFP_RETRY_MAYFAIL; + ret =mem_cgroup_charge_drvmem (ttm->memcg, gfp_flags, ttm->num_pages); +if (ret) +return ret; + } +#endif + ret =ttm_get_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state);if (unlikely (ret !=0 )) {
drivers/gpu/drm/ttm/ttm_tt.c +3 -0
@@ -233,6 +233,9 @@ void ttm_tt_init_fields(struct ttm_tt *ttm, struct ttm_buffer_object *bo, ttm->state = tt_unpopulated; ttm->swap_storage =NULL ; ttm->sg = bo->sg; +#ifdef CONFIG_MEMCG + ttm->memcg = bo->memcg; +#endif } int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,
include/drm/ttm/ttm_bo_api.h +5 -0
@@ -54,6 +54,8 @@ struct ttm_place;struct ttm_lru_bulk_move ; +struct mem_cgroup ; +/** * struct ttm_bus_placement *
@@ -180,6 +182,9 @@ struct ttm_buffer_object {void (*destroy) (struct ttm_buffer_object *);unsigned long num_pages;size_t acc_size; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif /** * Members not needing protection.
include/drm/ttm/ttm_tt.h +4 -0
@@ -33,6 +33,7 @@ struct ttm_tt;struct ttm_mem_reg ;struct ttm_buffer_object ;struct ttm_operation_ctx ; +struct mem_cgroup ;#define TTM_PAGE_FLAG_WRITE (1 << 3)#define TTM_PAGE_FLAG_SWAPPED (1 << 4)
@@ -116,6 +117,9 @@ struct ttm_tt { tt_unbound, tt_unpopulated, } state; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif };/**
Signature · 1 line
2.17.1
Re: [PATCH RFC 3/3] drm/ttm: support memcg for ttm_tt
Am 13.01.20 um 16:35 schrieb Qiang Yu:
Quoted 2 lines
Charge TTM allocated system memory to memory cgroup which will limit the memory usage of a group of processes.
NAK to the whole approach. This belongs into the GEM or driver layer, but not into TTM.
Quoted 7 lines
The memory is always charged to the control group of task which create this buffer object and when it's created. For example, when a buffer is created by process A and exported to process B, then process B populate this buffer, the memory is still charged to process A's memcg; if a buffer is created by process A when in memcg B, then A is moved to memcg C and populate this buffer, it will charge memcg B.
This is actually the most common use case for graphics application where the X server allocates most of the backing store.
So we need a better handling than just accounting the memory to whoever allocated it first.
Regards,
Christian.
Quoted 147 lines
Signed-off-by Qiang Yu <[email protected]>--- drivers/gpu/drm/ttm/ttm_bo.c | 10 ++++++++++ drivers/gpu/drm/ttm/ttm_page_alloc.c | 18 +++++++++++++++++- drivers/gpu/drm/ttm/ttm_tt.c | 3 +++ include/drm/ttm/ttm_bo_api.h | 5 +++++ include/drm/ttm/ttm_tt.h | 4 ++++ 5 files changed, 39 insertions(+), 1 deletion(-)drivers/gpu/drm/ttm/ttm_bo.c +10
-0@@ -42,6 +42,7 @@#include <linux/module.h> #include <linux/atomic.h> #include <linux/dma-resv.h> +#include <linux/memcontrol.h> static void ttm_bo_global_kobj_release (struct kobject *kobj);@@ -162,6 +163,10 @@ static void ttm_bo_release_list(struct kref *list_kref)if (!ttm_bo_uses_embedded_gem_object (bo))dma_resv_fini (&bo->base._resv);mutex_destroy (&bo->wu_mutex); +#ifdef CONFIG_MEMCG +if (bo->memcg) +css_put (&bo->memcg->css); +#endif bo->destroy (bo);ttm_mem_global_free (&ttm_mem_glob, acc_size); }@@ -1330,6 +1335,11 @@ int ttm_bo_init_reserved(struct ttm_bo_device *bdev, }atomic_inc (&ttm_bo_glob.bo_count); +#ifdef CONFIG_MEMCG +if (bo->type == ttm_bo_type_device) + bo->memcg =mem_cgroup_driver_get_from_current (); +#endif +/* * For ttm_bo_type_device buffers, allocate * address space from the device. drivers/gpu/drm/ttm/ttm_page_alloc.c +17
-1@@ -42,7 +42,7 @@#include <linux/seq_file.h> /* for seq_printf */ #include <linux/slab.h> #include <linux/dma-mapping.h> - +#include <linux/memcontrol.h> #include <linux/atomic.h> #include <drm/ttm/ttm_bo_driver.h> @@ -1045,6 +1045,11 @@ ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)ttm_put_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state); ttm->state = tt_unpopulated; + +#ifdef CONFIG_MEMCG +if (ttm->memcg) +mem_cgroup_uncharge_drvmem (ttm->memcg, ttm->num_pages); +#endif }int ttm_pool_populate (struct ttm_tt *ttm,struct ttm_operation_ctx *ctx)@@ -1059,6 +1064,17 @@ int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)if (ttm_check_under_lowerlimit (mem_glob, ttm->num_pages, ctx))return -ENOMEM ; +#ifdef CONFIG_MEMCG +if (ttm->memcg) { +gfp_t gfp_flags =GFP_USER ; +if (ttm->page_flags &TTM_PAGE_FLAG_NO_RETRY ) + gfp_flags |= __GFP_RETRY_MAYFAIL; + ret =mem_cgroup_charge_drvmem (ttm->memcg, gfp_flags, ttm->num_pages); +if (ret) +return ret; + } +#endif + ret =ttm_get_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state);if (unlikely (ret !=0 )) {drivers/gpu/drm/ttm/ttm_tt.c +3
-0@@ -233,6 +233,9 @@ void ttm_tt_init_fields(struct ttm_tt *ttm, struct ttm_buffer_object *bo, ttm->state = tt_unpopulated; ttm->swap_storage =NULL ; ttm->sg = bo->sg; +#ifdef CONFIG_MEMCG + ttm->memcg = bo->memcg; +#endif } int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,include/drm/ttm/ttm_bo_api.h +5
-0@@ -54,6 +54,8 @@ struct ttm_place;struct ttm_lru_bulk_move ; +struct mem_cgroup ; +/** * struct ttm_bus_placement * @@ -180,6 +182,9 @@ struct ttm_buffer_object {void (*destroy) (struct ttm_buffer_object *);unsigned long num_pages;size_t acc_size; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif /** * Members not needing protection. include/drm/ttm/ttm_tt.h +4
-0@@ -33,6 +33,7 @@ struct ttm_tt;struct ttm_mem_reg ;struct ttm_buffer_object ;struct ttm_operation_ctx ; +struct mem_cgroup ;#define TTM_PAGE_FLAG_WRITE (1 << 3)#define TTM_PAGE_FLAG_SWAPPED (1 << 4)@@ -116,6 +117,9 @@ struct ttm_tt { tt_unbound, tt_unpopulated, } state; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif };/**
Re: [PATCH RFC 3/3] drm/ttm: support memcg for ttm_tt
Quoted 8 lines · Christian König
Am 13.01.20 um 16:35 schrieb Qiang Yu:
Quoted 2 lines
Charge TTM allocated system memory to memory cgroup which will limit the memory usage of a group of processes.
NAK to the whole approach. This belongs into the GEM or driver layer, but not into TTM.
Sorry for responding late.
GEM layer seems not a proper place to handle this as:
1. it is not aware of the back storage (system mem or device mem) unless we add this information up to GEM which I think is not appropriate
2. system memory allocated by GEM with drm_gem_get_pages() is already charged to memcg, it's only the ttm system memory not charged to memcg
Implement in driver like amdgpu is an option. But seems the problem is inside TTM which does not charge pages allocated by itself to memcg, won't it be better to solve it in TTM so that all drivers using it can benefit? Or you just think we should not rely on memcg for GPU system memory limitation?
Quoted 8 lines
Quoted 7 lines
The memory is always charged to the control group of task which create this buffer object and when it's created. For example, when a buffer is created by process A and exported to process B, then process B populate this buffer, the memory is still charged to process A's memcg; if a buffer is created by process A when in memcg B, then A is moved to memcg C and populate this buffer, it will charge memcg B.
Quoted 6 lines
This is actually the most common use case for graphics application where the X server allocates most of the backing store.
So we need a better handling than just accounting the memory to whoever allocated it first.
You mean the application based on DRI2 and X11 protocol draw? I think this is still reasonable to charge xserver for the memory, because xserver allocate the buffer and share to application which is its design and implementation nature. With DRI3, the buffer is allocated by application, also suitable for this
approach.
Regards,
Qiang
Quoted 155 lines
Regards,
Christian.Quoted 138 lines
Signed-off-by Qiang Yu <[email protected]>--- drivers/gpu/drm/ttm/ttm_bo.c | 10 ++++++++++ drivers/gpu/drm/ttm/ttm_page_alloc.c | 18 +++++++++++++++++- drivers/gpu/drm/ttm/ttm_tt.c | 3 +++ include/drm/ttm/ttm_bo_api.h | 5 +++++ include/drm/ttm/ttm_tt.h | 4 ++++ 5 files changed, 39 insertions(+), 1 deletion(-)drivers/gpu/drm/ttm/ttm_bo.c +10
-0@@ -42,6 +42,7 @@#include <linux/module.h> #include <linux/atomic.h> #include <linux/dma-resv.h> +#include <linux/memcontrol.h> static void ttm_bo_global_kobj_release (struct kobject *kobj);@@ -162,6 +163,10 @@ static void ttm_bo_release_list(struct kref *list_kref)if (!ttm_bo_uses_embedded_gem_object (bo))dma_resv_fini (&bo->base._resv);mutex_destroy (&bo->wu_mutex); +#ifdef CONFIG_MEMCG +if (bo->memcg) +css_put (&bo->memcg->css); +#endif bo->destroy (bo);ttm_mem_global_free (&ttm_mem_glob, acc_size); }@@ -1330,6 +1335,11 @@ int ttm_bo_init_reserved(struct ttm_bo_device *bdev, }atomic_inc (&ttm_bo_glob.bo_count); +#ifdef CONFIG_MEMCG +if (bo->type == ttm_bo_type_device) + bo->memcg =mem_cgroup_driver_get_from_current (); +#endif +/* * For ttm_bo_type_device buffers, allocate * address space from the device. drivers/gpu/drm/ttm/ttm_page_alloc.c +17
-1@@ -42,7 +42,7 @@#include <linux/seq_file.h> /* for seq_printf */ #include <linux/slab.h> #include <linux/dma-mapping.h> - +#include <linux/memcontrol.h> #include <linux/atomic.h> #include <drm/ttm/ttm_bo_driver.h> @@ -1045,6 +1045,11 @@ ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)ttm_put_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state); ttm->state = tt_unpopulated; + +#ifdef CONFIG_MEMCG +if (ttm->memcg) +mem_cgroup_uncharge_drvmem (ttm->memcg, ttm->num_pages); +#endif }int ttm_pool_populate (struct ttm_tt *ttm,struct ttm_operation_ctx *ctx)@@ -1059,6 +1064,17 @@ int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)if (ttm_check_under_lowerlimit (mem_glob, ttm->num_pages, ctx))return -ENOMEM ; +#ifdef CONFIG_MEMCG +if (ttm->memcg) { +gfp_t gfp_flags =GFP_USER ; +if (ttm->page_flags &TTM_PAGE_FLAG_NO_RETRY ) + gfp_flags |= __GFP_RETRY_MAYFAIL; + ret =mem_cgroup_charge_drvmem (ttm->memcg, gfp_flags, ttm->num_pages); +if (ret) +return ret; + } +#endif + ret =ttm_get_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state);if (unlikely (ret !=0 )) {drivers/gpu/drm/ttm/ttm_tt.c +3
-0@@ -233,6 +233,9 @@ void ttm_tt_init_fields(struct ttm_tt *ttm, struct ttm_buffer_object *bo, ttm->state = tt_unpopulated; ttm->swap_storage =NULL ; ttm->sg = bo->sg; +#ifdef CONFIG_MEMCG + ttm->memcg = bo->memcg; +#endif } int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,include/drm/ttm/ttm_bo_api.h +5
-0@@ -54,6 +54,8 @@ struct ttm_place;struct ttm_lru_bulk_move ; +struct mem_cgroup ; +/** * struct ttm_bus_placement * @@ -180,6 +182,9 @@ struct ttm_buffer_object {void (*destroy) (struct ttm_buffer_object *);unsigned long num_pages;size_t acc_size; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif /** * Members not needing protection. include/drm/ttm/ttm_tt.h +1
-0@@ -33,6 +33,7 @@ struct ttm_tt;struct ttm_mem_reg ;struct ttm_buffer_object ;struct ttm_operation_ctx ; +struct mem_cgroup ;#define TTM_PAGE_FLAG_WRITE (1 << 3)#define TTM_PAGE_FLAG_SWAPPED (1 << 4)include/drm/ttm/ttm_tt.h struct ttm_tt { · line 125
Quoted 9 lines
include/drm/ttm/ttm_tt.h +3
-0@@ -116,6 +117,9 @@ struct ttm_tt { tt_unbound, tt_unpopulated, } state; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif };/** _______________________________________________
dri-devel mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
Re: [PATCH RFC 3/3] drm/ttm: support memcg for ttm_tt
Am 19.01.20 um 03:47 schrieb Qiang Yu:
Quoted 8 lines
Quoted 6 lines · Christian König
Am 13.01.20 um 16:35 schrieb Qiang Yu:
Quoted 2 lines
Charge TTM allocated system memory to memory cgroup which will limit the memory usage of a group of processes.
NAK to the whole approach. This belongs into the GEM or driver layer, but not into TTM.
Quoted 7 lines
Sorry for responding late.
GEM layer seems not a proper place to handle this as:
1. it is not aware of the back storage (system mem or device mem) unless we add this information up to GEM which I think is not appropriate
2. system memory allocated by GEM withdrm_gem_get_pages()is already charged to memcg, it's only the ttm system memory not charged to memcg
The key point is that we already discussed this on the mailing list and GEM was agreed on to be the right place for this.
That's the reason why the Intel developers already proposed a way to expose the buffer location in GEM.
Please sync up with Kenny who is leading the development efforts and with the Intel developers before warming up an old discussion again.
Adding that to TTM is an absolute no-go from my maintainers perspective.
Quoted 19 lines
Implement in driver like amdgpu is an option. But seems the problem is inside TTM which does not charge pages allocated by itself to memcg, won't it be better to solve it in TTM so that all drivers using it can benefit? Or you just think we should not rely on memcg for GPU system memory limitation?
Quoted 7 lines
Quoted 7 lines
The memory is always charged to the control group of task which create this buffer object and when it's created. For example, when a buffer is created by process A and exported to process B, then process B populate this buffer, the memory is still charged to process A's memcg; if a buffer is created by process A when in memcg B, then A is moved to memcg C and populate this buffer, it will charge memcg B.
Quoted 6 lines
This is actually the most common use case for graphics application where the X server allocates most of the backing store.
So we need a better handling than just accounting the memory to whoever allocated it first.
Quoted 6 lines
You mean the application based on DRI2 and X11 protocol draw? I think this is still reasonable to charge xserver for the memory, because xserver allocate the buffer and share to application which is its design and implementation nature. With DRI3, the buffer is allocated by application, also suitable for this
approach.
That is a way to simplistic.
Again we already discussed this and the agreed compromise is to charge the application which is using the memory and not who has allocated it.
So you need to add the charge on importing a buffer and not just when it is created.
Regards,
Christian.
Quoted 157 lines
Regards,
QiangQuoted 153 lines
Regards,
Christian.Quoted 137 lines
Signed-off-by Qiang Yu <[email protected]>--- drivers/gpu/drm/ttm/ttm_bo.c | 10 ++++++++++ drivers/gpu/drm/ttm/ttm_page_alloc.c | 18 +++++++++++++++++- drivers/gpu/drm/ttm/ttm_tt.c | 3 +++ include/drm/ttm/ttm_bo_api.h | 5 +++++ include/drm/ttm/ttm_tt.h | 4 ++++ 5 files changed, 39 insertions(+), 1 deletion(-)drivers/gpu/drm/ttm/ttm_bo.c +10
-0@@ -42,6 +42,7 @@#include <linux/module.h> #include <linux/atomic.h> #include <linux/dma-resv.h> +#include <linux/memcontrol.h> static void ttm_bo_global_kobj_release (struct kobject *kobj);@@ -162,6 +163,10 @@ static void ttm_bo_release_list(struct kref *list_kref)if (!ttm_bo_uses_embedded_gem_object (bo))dma_resv_fini (&bo->base._resv);mutex_destroy (&bo->wu_mutex); +#ifdef CONFIG_MEMCG +if (bo->memcg) +css_put (&bo->memcg->css); +#endif bo->destroy (bo);ttm_mem_global_free (&ttm_mem_glob, acc_size); }@@ -1330,6 +1335,11 @@ int ttm_bo_init_reserved(struct ttm_bo_device *bdev, }atomic_inc (&ttm_bo_glob.bo_count); +#ifdef CONFIG_MEMCG +if (bo->type == ttm_bo_type_device) + bo->memcg =mem_cgroup_driver_get_from_current (); +#endif +/* * For ttm_bo_type_device buffers, allocate * address space from the device. drivers/gpu/drm/ttm/ttm_page_alloc.c +17
-1@@ -42,7 +42,7 @@#include <linux/seq_file.h> /* for seq_printf */ #include <linux/slab.h> #include <linux/dma-mapping.h> - +#include <linux/memcontrol.h> #include <linux/atomic.h> #include <drm/ttm/ttm_bo_driver.h> @@ -1045,6 +1045,11 @@ ttm_pool_unpopulate_helper(struct ttm_tt *ttm, unsigned mem_count_update)ttm_put_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state); ttm->state = tt_unpopulated; + +#ifdef CONFIG_MEMCG +if (ttm->memcg) +mem_cgroup_uncharge_drvmem (ttm->memcg, ttm->num_pages); +#endif }int ttm_pool_populate (struct ttm_tt *ttm,struct ttm_operation_ctx *ctx)@@ -1059,6 +1064,17 @@ int ttm_pool_populate(struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)if (ttm_check_under_lowerlimit (mem_glob, ttm->num_pages, ctx))return -ENOMEM ; +#ifdef CONFIG_MEMCG +if (ttm->memcg) { +gfp_t gfp_flags =GFP_USER ; +if (ttm->page_flags &TTM_PAGE_FLAG_NO_RETRY ) + gfp_flags |= __GFP_RETRY_MAYFAIL; + ret =mem_cgroup_charge_drvmem (ttm->memcg, gfp_flags, ttm->num_pages); +if (ret) +return ret; + } +#endif + ret =ttm_get_pages (ttm->pages, ttm->num_pages, ttm->page_flags, ttm->caching_state);if (unlikely (ret !=0 )) {drivers/gpu/drm/ttm/ttm_tt.c +3
-0@@ -233,6 +233,9 @@ void ttm_tt_init_fields(struct ttm_tt *ttm, struct ttm_buffer_object *bo, ttm->state = tt_unpopulated; ttm->swap_storage =NULL ; ttm->sg = bo->sg; +#ifdef CONFIG_MEMCG + ttm->memcg = bo->memcg; +#endif } int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,include/drm/ttm/ttm_bo_api.h +5
-0@@ -54,6 +54,8 @@ struct ttm_place;struct ttm_lru_bulk_move ; +struct mem_cgroup ; +/** * struct ttm_bus_placement * @@ -180,6 +182,9 @@ struct ttm_buffer_object {void (*destroy) (struct ttm_buffer_object *);unsigned long num_pages;size_t acc_size; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif /** * Members not needing protection. include/drm/ttm/ttm_tt.h +1
-0@@ -33,6 +33,7 @@ struct ttm_tt;struct ttm_mem_reg ;struct ttm_buffer_object ;struct ttm_operation_ctx ; +struct mem_cgroup ;#define TTM_PAGE_FLAG_WRITE (1 << 3)#define TTM_PAGE_FLAG_SWAPPED (1 << 4)include/drm/ttm/ttm_tt.h struct ttm_tt { · line 125
Quoted 9 lines
include/drm/ttm/ttm_tt.h +3
-0@@ -116,6 +117,9 @@ struct ttm_tt { tt_unbound, tt_unpopulated, } state; +#ifdef CONFIG_MEMCG +struct mem_cgroup *memcg; +#endif };/** _______________________________________________
dri-devel mailing list [email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fdri-devel&data=02%7C01%7Cchristian.koenig%40amd.com%7C5d3b70a43b80444c550808d79c89e968%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637149988466853762&sdata=ni3ku7nC%2FD5E8kivppfuuF7ZoiyfLQ8L3Y4j9IfHYUU%3D&reserved=0