Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
drm/xe/madvise: Update migration policy based on preferred location
When the user sets the valid devmem_fd as a preferred location, GPU fault
will trigger migration to tile of device associated with devmem_fd.
If the user sets an invalid devmem_fd the preferred location is current
placement(smem) only.
v2(Matthew Brost)
- Default should be faulting tile
- remove devmem_fd used as region
v3 (Matthew Brost)
- Add migration_policy
- Fix return condition
- fix migrate condition
v4
-Rebase
v5
- Add check for userptr and bo based vmas
Cc: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
Link: https://lore.kernel.org/r/20250821173104.3030148-11-himal.prasad.ghimiray@intel.com
Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
+76
-2
+44
-1
drivers/gpu/drm/xe/xe_svm.c
+44
-1
drivers/gpu/drm/xe/xe_svm.c
···
806
806
};
807
807
struct xe_svm_range *range;
808
808
struct dma_fence *fence;
809
+
struct drm_pagemap *dpagemap;
809
810
struct xe_tile *tile = gt_to_tile(gt);
810
811
int migrate_try_count = ctx.devmem_only ? 3 : 1;
811
812
ktime_t end = 0;
···
836
835
837
836
range_debug(range, "PAGE FAULT");
838
837
838
+
dpagemap = xe_vma_resolve_pagemap(vma, tile);
839
839
if (--migrate_try_count >= 0 &&
840
-
xe_svm_range_needs_migrate_to_vram(range, vma, IS_DGFX(vm->xe))) {
840
+
xe_svm_range_needs_migrate_to_vram(range, vma, !!dpagemap || ctx.devmem_only)) {
841
+
/* TODO : For multi-device dpagemap will be used to find the
842
+
* remote tile and remote device. Will need to modify
843
+
* xe_svm_alloc_vram to use dpagemap for future multi-device
844
+
* support.
845
+
*/
841
846
err = xe_svm_alloc_vram(tile, range, &ctx);
842
847
ctx.timeslice_ms <<= 1; /* Double timeslice if we have to retry */
843
848
if (err) {
···
1108
1101
}
1109
1102
1110
1103
/**
1104
+
* xe_vma_resolve_pagemap - Resolve the appropriate DRM pagemap for a VMA
1105
+
* @vma: Pointer to the xe_vma structure containing memory attributes
1106
+
* @tile: Pointer to the xe_tile structure used as fallback for VRAM mapping
1107
+
*
1108
+
* This function determines the correct DRM pagemap to use for a given VMA.
1109
+
* It first checks if a valid devmem_fd is provided in the VMA's preferred
1110
+
* location. If the devmem_fd is negative, it returns NULL, indicating no
1111
+
* pagemap is available and smem to be used as preferred location.
1112
+
* If the devmem_fd is equal to the default faulting
1113
+
* GT identifier, it returns the VRAM pagemap associated with the tile.
1114
+
*
1115
+
* Future support for multi-device configurations may use drm_pagemap_from_fd()
1116
+
* to resolve pagemaps from arbitrary file descriptors.
1117
+
*
1118
+
* Return: A pointer to the resolved drm_pagemap, or NULL if none is applicable.
1119
+
*/
1120
+
struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile)
1121
+
{
1122
+
s32 fd = (s32)vma->attr.preferred_loc.devmem_fd;
1123
+
1124
+
if (fd == DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM)
1125
+
return NULL;
1126
+
1127
+
if (fd == DRM_XE_PREFERRED_LOC_DEFAULT_DEVICE)
1128
+
return IS_DGFX(tile_to_xe(tile)) ? tile_local_pagemap(tile) : NULL;
1129
+
1130
+
/* TODO: Support multi-device with drm_pagemap_from_fd(fd) */
1131
+
return NULL;
1132
+
}
1133
+
1134
+
/**
1111
1135
* xe_svm_alloc_vram()- Allocate device memory pages for range,
1112
1136
* migrating existing data.
1113
1137
* @tile: tile to allocate vram from
···
1249
1211
int xe_devm_add(struct xe_tile *tile, struct xe_vram_region *vr)
1250
1212
{
1251
1213
return 0;
1214
+
}
1215
+
1216
+
struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile)
1217
+
{
1218
+
return NULL;
1252
1219
}
1253
1220
#endif
1254
1221
+8
drivers/gpu/drm/xe/xe_svm.h
+8
drivers/gpu/drm/xe/xe_svm.h
···
94
94
95
95
u8 xe_svm_ranges_zap_ptes_in_range(struct xe_vm *vm, u64 start, u64 end);
96
96
97
+
struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile);
98
+
97
99
/**
98
100
* xe_svm_range_has_dma_mapping() - SVM range has DMA mapping
99
101
* @range: SVM range
···
318
316
u8 xe_svm_ranges_zap_ptes_in_range(struct xe_vm *vm, u64 start, u64 end)
319
317
{
320
318
return 0;
319
+
}
320
+
321
+
static inline
322
+
struct drm_pagemap *xe_vma_resolve_pagemap(struct xe_vma *vma, struct xe_tile *tile)
323
+
{
324
+
return NULL;
321
325
}
322
326
323
327
#define xe_svm_assert_in_notifier(...) do {} while (0)
+24
-1
drivers/gpu/drm/xe/xe_vm_madvise.c
+24
-1
drivers/gpu/drm/xe/xe_vm_madvise.c
···
78
78
struct xe_vma **vmas, int num_vmas,
79
79
struct drm_xe_madvise *op)
80
80
{
81
-
/* Implementation pending */
81
+
int i;
82
+
83
+
xe_assert(vm->xe, op->type == DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC);
84
+
85
+
for (i = 0; i < num_vmas; i++) {
86
+
/*TODO: Extend attributes to bo based vmas */
87
+
if (!xe_vma_is_cpu_addr_mirror(vmas[i]))
88
+
continue;
89
+
90
+
vmas[i]->attr.preferred_loc.devmem_fd = op->preferred_mem_loc.devmem_fd;
91
+
92
+
/* Till multi-device support is not added migration_policy
93
+
* is of no use and can be ignored.
94
+
*/
95
+
vmas[i]->attr.preferred_loc.migration_policy =
96
+
op->preferred_mem_loc.migration_policy;
97
+
}
82
98
}
83
99
84
100
static void madvise_atomic(struct xe_device *xe, struct xe_vm *vm,
···
200
184
201
185
switch (args->type) {
202
186
case DRM_XE_MEM_RANGE_ATTR_PREFERRED_LOC:
187
+
{
188
+
s32 fd = (s32)args->preferred_mem_loc.devmem_fd;
189
+
190
+
if (XE_IOCTL_DBG(xe, fd < DRM_XE_PREFERRED_LOC_DEFAULT_SYSTEM))
191
+
return false;
192
+
203
193
if (XE_IOCTL_DBG(xe, args->preferred_mem_loc.migration_policy >
204
194
DRM_XE_MIGRATE_ONLY_SYSTEM_PAGES))
205
195
return false;
···
216
194
if (XE_IOCTL_DBG(xe, args->atomic.reserved))
217
195
return false;
218
196
break;
197
+
}
219
198
case DRM_XE_MEM_RANGE_ATTR_ATOMIC:
220
199
if (XE_IOCTL_DBG(xe, args->atomic.val > DRM_XE_ATOMIC_CPU))
221
200
return false;