On 27/10/2026 22:49, Matthew Brost wrote:
Servicing a page fault validates the VMA's backing store and binds it,
making the BO part of the VM's active working set, but unlike exec it
never moves anything on the TTM LRU. In fault mode, where a VM may run
for a long time without a single exec, the BOs the GPU actually touches
therefore drift towards the LRU head and are picked by eviction and by
the shrinker ahead of genuinely cold BOs.
Bump the BO's LRU on a successful page fault.
Fixes: dd08ebf6c352 ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Brost <[email protected]>
---
drivers/gpu/drm/xe/xe_pagefault.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/drivers/gpu/drm/xe/xe_pagefault.c
b/drivers/gpu/drm/xe/xe_pagefault.c
index 5247321c805e..6828f81df50e 100644
--- a/drivers/gpu/drm/xe/xe_pagefault.c
+++ b/drivers/gpu/drm/xe/xe_pagefault.c
@@ -201,6 +201,8 @@ static int xe_pagefault_handle_vma(struct xe_gt *gt, struct
xe_vma *vma,
xe_validation_ctx_init(&ctx, &vm->xe->val, &exec,
(struct xe_val_flags) {});
drm_exec_until_all_locked(&exec) {
+ struct xe_bo *bo = xe_vma_bo(vma);
+
err = xe_pagefault_begin(&exec, vma, tile->mem.vram,
needs_vram == 1);
drm_exec_retry_on_contention(&exec);
@@ -221,6 +223,11 @@ static int xe_pagefault_handle_vma(struct xe_gt *gt,
struct xe_vma *vma,
xe_pagefault_set_error(pf,
XE_PAGEFAULT_ERROR_VMA_REBIND);
break;
}
+
+ if (bo) {
+ scoped_guard(spinlock, &vm->xe->ttm.lru_lock)
+ ttm_bo_move_to_lru_tail(&bo->ttm);
Would it be cleaner to just use ttm_bo_move_to_lru_tail_unlocked()
without the scoped guard stuff?
Reviewed-by: Matthew Auld <[email protected]>
+ }
}
xe_validation_ctx_fini(&ctx);
} while (err == -EAGAIN);
