Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
drm/amdgpu: expose more memory stats in fdinfo
This will be used for performance investigations.
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
authored by
Marek Olšák
and committed by
Alex Deucher
d6530c33
28f7e897
+76
-28
+19
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c
+19
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_fdinfo.c
···
60
60
struct amdgpu_fpriv *fpriv = file->driver_priv;
61
61
struct amdgpu_vm *vm = &fpriv->vm;
62
62
63
-
uint64_t vram_mem = 0, gtt_mem = 0, cpu_mem = 0;
63
+
struct amdgpu_mem_stats stats;
64
64
ktime_t usage[AMDGPU_HW_IP_NUM];
65
65
uint32_t bus, dev, fn, domain;
66
66
unsigned int hw_ip;
67
67
int ret;
68
68
69
+
memset(&stats, 0, sizeof(stats));
69
70
bus = adev->pdev->bus->number;
70
71
domain = pci_domain_nr(adev->pdev->bus);
71
72
dev = PCI_SLOT(adev->pdev->devfn);
···
76
75
if (ret)
77
76
return;
78
77
79
-
amdgpu_vm_get_memory(vm, &vram_mem, >t_mem, &cpu_mem);
78
+
amdgpu_vm_get_memory(vm, &stats);
80
79
amdgpu_bo_unreserve(vm->root.bo);
81
80
82
81
amdgpu_ctx_mgr_usage(&fpriv->ctx_mgr, usage);
···
91
90
seq_printf(m, "drm-driver:\t%s\n", file->minor->dev->driver->name);
92
91
seq_printf(m, "drm-pdev:\t%04x:%02x:%02x.%d\n", domain, bus, dev, fn);
93
92
seq_printf(m, "drm-client-id:\t%Lu\n", vm->immediate.fence_context);
94
-
seq_printf(m, "drm-memory-vram:\t%llu KiB\n", vram_mem/1024UL);
95
-
seq_printf(m, "drm-memory-gtt: \t%llu KiB\n", gtt_mem/1024UL);
96
-
seq_printf(m, "drm-memory-cpu: \t%llu KiB\n", cpu_mem/1024UL);
93
+
seq_printf(m, "drm-memory-vram:\t%llu KiB\n", stats.vram/1024UL);
94
+
seq_printf(m, "drm-memory-gtt: \t%llu KiB\n", stats.gtt/1024UL);
95
+
seq_printf(m, "drm-memory-cpu: \t%llu KiB\n", stats.cpu/1024UL);
96
+
seq_printf(m, "amd-memory-visible-vram:\t%llu KiB\n",
97
+
stats.visible_vram/1024UL);
98
+
seq_printf(m, "amd-evicted-vram:\t%llu KiB\n",
99
+
stats.evicted_vram/1024UL);
100
+
seq_printf(m, "amd-evicted-visible-vram:\t%llu KiB\n",
101
+
stats.evicted_visible_vram/1024UL);
102
+
seq_printf(m, "amd-requested-vram:\t%llu KiB\n",
103
+
stats.requested_vram/1024UL);
104
+
seq_printf(m, "amd-requested-visible-vram:\t%llu KiB\n",
105
+
stats.requested_visible_vram/1024UL);
106
+
seq_printf(m, "amd-requested-gtt:\t%llu KiB\n",
107
+
stats.requested_gtt/1024UL);
108
+
97
109
for (hw_ip = 0; hw_ip < AMDGPU_HW_IP_NUM; ++hw_ip) {
98
110
if (!usage[hw_ip])
99
111
continue;
+22
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+22
-5
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
···
1265
1265
trace_amdgpu_bo_move(abo, new_mem->mem_type, old_mem->mem_type);
1266
1266
}
1267
1267
1268
-
void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
1269
-
uint64_t *gtt_mem, uint64_t *cpu_mem)
1268
+
void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
1269
+
struct amdgpu_mem_stats *stats)
1270
1270
{
1271
1271
unsigned int domain;
1272
+
uint64_t size = amdgpu_bo_size(bo);
1272
1273
1273
1274
domain = amdgpu_mem_type_to_domain(bo->tbo.resource->mem_type);
1274
1275
switch (domain) {
1275
1276
case AMDGPU_GEM_DOMAIN_VRAM:
1276
-
*vram_mem += amdgpu_bo_size(bo);
1277
+
stats->vram += size;
1278
+
if (amdgpu_bo_in_cpu_visible_vram(bo))
1279
+
stats->visible_vram += size;
1277
1280
break;
1278
1281
case AMDGPU_GEM_DOMAIN_GTT:
1279
-
*gtt_mem += amdgpu_bo_size(bo);
1282
+
stats->gtt += size;
1280
1283
break;
1281
1284
case AMDGPU_GEM_DOMAIN_CPU:
1282
1285
default:
1283
-
*cpu_mem += amdgpu_bo_size(bo);
1286
+
stats->cpu += size;
1284
1287
break;
1288
+
}
1289
+
1290
+
if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_VRAM) {
1291
+
stats->requested_vram += size;
1292
+
if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1293
+
stats->requested_visible_vram += size;
1294
+
1295
+
if (domain != AMDGPU_GEM_DOMAIN_VRAM) {
1296
+
stats->evicted_vram += size;
1297
+
if (bo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED)
1298
+
stats->evicted_visible_vram += size;
1299
+
}
1300
+
} else if (bo->preferred_domains & AMDGPU_GEM_DOMAIN_GTT) {
1301
+
stats->requested_gtt += size;
1285
1302
}
1286
1303
}
1287
1304
+23
-2
drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
+23
-2
drivers/gpu/drm/amd/amdgpu/amdgpu_object.h
···
126
126
struct amdgpu_vm_bo_base entries[];
127
127
};
128
128
129
+
struct amdgpu_mem_stats {
130
+
/* current VRAM usage, includes visible VRAM */
131
+
uint64_t vram;
132
+
/* current visible VRAM usage */
133
+
uint64_t visible_vram;
134
+
/* current GTT usage */
135
+
uint64_t gtt;
136
+
/* current system memory usage */
137
+
uint64_t cpu;
138
+
/* sum of evicted buffers, includes visible VRAM */
139
+
uint64_t evicted_vram;
140
+
/* sum of evicted buffers due to CPU access */
141
+
uint64_t evicted_visible_vram;
142
+
/* how much userspace asked for, includes vis.VRAM */
143
+
uint64_t requested_vram;
144
+
/* how much userspace asked for */
145
+
uint64_t requested_visible_vram;
146
+
/* how much userspace asked for */
147
+
uint64_t requested_gtt;
148
+
};
149
+
129
150
static inline struct amdgpu_bo *ttm_to_amdgpu_bo(struct ttm_buffer_object *tbo)
130
151
{
131
152
return container_of(tbo, struct amdgpu_bo, tbo);
···
346
325
int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr);
347
326
u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo);
348
327
u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo);
349
-
void amdgpu_bo_get_memory(struct amdgpu_bo *bo, uint64_t *vram_mem,
350
-
uint64_t *gtt_mem, uint64_t *cpu_mem);
328
+
void amdgpu_bo_get_memory(struct amdgpu_bo *bo,
329
+
struct amdgpu_mem_stats *stats);
351
330
void amdgpu_bo_add_to_shadow_list(struct amdgpu_bo_vm *vmbo);
352
331
int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow,
353
332
struct dma_fence **fence);
+9
-14
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+9
-14
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
···
920
920
return r;
921
921
}
922
922
923
-
void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
924
-
uint64_t *gtt_mem, uint64_t *cpu_mem)
923
+
void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
924
+
struct amdgpu_mem_stats *stats)
925
925
{
926
926
struct amdgpu_bo_va *bo_va, *tmp;
927
927
···
929
929
list_for_each_entry_safe(bo_va, tmp, &vm->idle, base.vm_status) {
930
930
if (!bo_va->base.bo)
931
931
continue;
932
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
933
-
gtt_mem, cpu_mem);
932
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
934
933
}
935
934
list_for_each_entry_safe(bo_va, tmp, &vm->evicted, base.vm_status) {
936
935
if (!bo_va->base.bo)
937
936
continue;
938
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
939
-
gtt_mem, cpu_mem);
937
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
940
938
}
941
939
list_for_each_entry_safe(bo_va, tmp, &vm->relocated, base.vm_status) {
942
940
if (!bo_va->base.bo)
943
941
continue;
944
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
945
-
gtt_mem, cpu_mem);
942
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
946
943
}
947
944
list_for_each_entry_safe(bo_va, tmp, &vm->moved, base.vm_status) {
948
945
if (!bo_va->base.bo)
949
946
continue;
950
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
951
-
gtt_mem, cpu_mem);
947
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
952
948
}
953
949
list_for_each_entry_safe(bo_va, tmp, &vm->invalidated, base.vm_status) {
954
950
if (!bo_va->base.bo)
955
951
continue;
956
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
957
-
gtt_mem, cpu_mem);
952
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
958
953
}
959
954
list_for_each_entry_safe(bo_va, tmp, &vm->done, base.vm_status) {
960
955
if (!bo_va->base.bo)
961
956
continue;
962
-
amdgpu_bo_get_memory(bo_va->base.bo, vram_mem,
963
-
gtt_mem, cpu_mem);
957
+
amdgpu_bo_get_memory(bo_va->base.bo, stats);
964
958
}
965
959
spin_unlock(&vm->status_lock);
966
960
}
961
+
967
962
/**
968
963
* amdgpu_vm_bo_update - update all BO mappings in the vm page table
969
964
*
+3
-2
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+3
-2
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
···
40
40
struct amdgpu_job;
41
41
struct amdgpu_bo_list_entry;
42
42
struct amdgpu_bo_vm;
43
+
struct amdgpu_mem_stats;
43
44
44
45
/*
45
46
* GPUVM handling
···
458
457
459
458
void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev,
460
459
struct amdgpu_vm *vm);
461
-
void amdgpu_vm_get_memory(struct amdgpu_vm *vm, uint64_t *vram_mem,
462
-
uint64_t *gtt_mem, uint64_t *cpu_mem);
460
+
void amdgpu_vm_get_memory(struct amdgpu_vm *vm,
461
+
struct amdgpu_mem_stats *stats);
463
462
464
463
int amdgpu_vm_pt_clear(struct amdgpu_device *adev, struct amdgpu_vm *vm,
465
464
struct amdgpu_bo_vm *vmbo, bool immediate);