Re: [PATCH V4 14/19] accel/amdxdna: Move HMM invalidate wait into common GEM code

Thu, 15 Oct 2026 12:21:59 -0700 


On 10/13/26 11:45, Lizhi Hou wrote:


On 10/9/26 15:53, David Zhang wrote:

When a user mapping of a BO is invalidated, the driver must wait for
the BO's reservation fences before the pages are released. AIE2 does
this in aie2_hmm_invalidate() through the .hmm_invalidate callback,
and AIE4 needs the same wait.

Do the wait directly in amdxdna_hmm_invalidate() in amdxdna_gem.c and
remove the .hmm_invalidate callback and aie2_hmm_invalidate(). Devices
that do not attach fences to user BOs return from the wait at once.

Make the wait uninterruptible. With an interruptible wait, a pending
signal, such as SIGKILL during process exit, ends the wait early and
the pages can be freed while the NPU may still access them.



This is a bug fix. Please add "Fixes" tag and change the title to 
reflect the fix.



It is actually better to separate the bug fix and the movement to common 
code.


I would take the fix:

https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/dfa1a332824421f1713388a86331a4c4d84f680f

Then moving to common code on top.

Thanks,

Lizhi



Lizhi



Co-developed-by: Hayden Laccabue <[email protected]>
Signed-off-by: Hayden Laccabue <[email protected]>
Signed-off-by: David Zhang <[email protected]>
---
  drivers/accel/amdxdna/aie2_ctx.c        | 15 ---------------
  drivers/accel/amdxdna/aie2_pci.c        |  1 -
  drivers/accel/amdxdna/aie2_pci.h        |  1 -
  drivers/accel/amdxdna/amdxdna_gem.c     | 14 ++++++++++++--
  drivers/accel/amdxdna/amdxdna_pci_drv.h |  1 -
  5 files changed, 12 insertions(+), 20 deletions(-)


diff --git a/drivers/accel/amdxdna/aie2_ctx.c 
b/drivers/accel/amdxdna/aie2_ctx.c

index d927c8c9d557..8c1b29964f8f 100644
--- a/drivers/accel/amdxdna/aie2_ctx.c
+++ b/drivers/accel/amdxdna/aie2_ctx.c

@@ -1277,21 +1277,6 @@ int aie2_cmd_submit(struct amdxdna_hwctx 
*hwctx, struct amdxdna_sched_job *job,

      return ret;
  }
  -void aie2_hmm_invalidate(struct amdxdna_gem_obj *abo,
-             unsigned long cur_seq)
-{
-    struct amdxdna_dev *xdna = to_xdna_dev(to_gobj(abo)->dev);
-    struct drm_gem_object *gobj = to_gobj(abo);
-    long ret;
-
-    ret = dma_resv_wait_timeout(gobj->resv, DMA_RESV_USAGE_BOOKKEEP,
-                    true, MAX_SCHEDULE_TIMEOUT);
-    if (!ret)
-        XDNA_ERR(xdna, "Failed to wait for bo, ret %ld", ret);
-    else if (ret == -ERESTARTSYS)
-        XDNA_DBG(xdna, "Wait for bo interrupted by signal");
-}
-
  int aie2_hwctx_heap_expand(struct amdxdna_hwctx *hwctx,
                 struct amdxdna_gem_obj *heap)
  {

diff --git a/drivers/accel/amdxdna/aie2_pci.c 
b/drivers/accel/amdxdna/aie2_pci.c

index 1b029d33a24b..690ad035463c 100644
--- a/drivers/accel/amdxdna/aie2_pci.c
+++ b/drivers/accel/amdxdna/aie2_pci.c
@@ -1228,7 +1228,6 @@ const struct amdxdna_dev_ops aie2_ops = {
      .hwctx_config = aie2_hwctx_config,
      .hwctx_sync_debug_bo = aie2_hwctx_sync_debug_bo,
      .cmd_submit = aie2_cmd_submit,
-    .hmm_invalidate = aie2_hmm_invalidate,
      .get_array = aie2_get_array,
      .get_dev_revision = aie2_get_dev_rev,
      .hwctx_heap_expand = aie2_hwctx_heap_expand,

diff --git a/drivers/accel/amdxdna/aie2_pci.h 
b/drivers/accel/amdxdna/aie2_pci.h

index 709e7545c65f..13391275df31 100644
--- a/drivers/accel/amdxdna/aie2_pci.h
+++ b/drivers/accel/amdxdna/aie2_pci.h

@@ -272,7 +272,6 @@ int aie2_hwctx_sync_debug_bo(struct amdxdna_hwctx 
*hwctx, u32 debug_bo_hdl);

  void aie2_hwctx_suspend(struct amdxdna_client *client);
  int aie2_hwctx_resume(struct amdxdna_client *client);

  int aie2_cmd_submit(struct amdxdna_hwctx *hwctx, struct 
amdxdna_sched_job *job, u64 *seq);
-void aie2_hmm_invalidate(struct amdxdna_gem_obj *abo, unsigned long 
cur_seq);
  int aie2_hwctx_heap_expand(struct amdxdna_hwctx *hwctx, struct 
amdxdna_gem_obj *heap);

    #endif /* _AIE2_PCI_H_ */

diff --git a/drivers/accel/amdxdna/amdxdna_gem.c 
b/drivers/accel/amdxdna/amdxdna_gem.c

index f4832337ec31..e1b72cef51bd 100644
--- a/drivers/accel/amdxdna/amdxdna_gem.c
+++ b/drivers/accel/amdxdna/amdxdna_gem.c
@@ -12,6 +12,7 @@
  #include <drm/gpu_scheduler.h>
  #include <linux/dma-buf.h>
  #include <linux/dma-direct.h>
+#include <linux/dma-resv.h>
  #include <linux/iosys-map.h>
  #include <linux/pagemap.h>
  #include <linux/swap.h>

@@ -232,6 +233,7 @@ static bool amdxdna_hmm_invalidate(struct 
mmu_interval_notifier *mni,
      struct amdxdna_umap *mapp = container_of(mni, struct 
amdxdna_umap, notifier);

      struct amdxdna_gem_obj *abo = mapp->abo;
      struct amdxdna_dev *xdna;
+    long ret;
        if (!mmu_notifier_range_blockable(range))
          return false;

@@ -249,8 +251,16 @@ static bool amdxdna_hmm_invalidate(struct 
mmu_interval_notifier *mni,

      mmu_interval_set_seq(&mapp->notifier, cur_seq);
      up_write(&xdna->notifier_lock);
  -    if (xdna->dev_info->ops->hmm_invalidate)
-        xdna->dev_info->ops->hmm_invalidate(abo, cur_seq);
+    /*
+     * Wait uninterruptibly. Returning early on a signal would let the
+     * pages be freed while the NPU may still access them. This relies
+     * on job fences signaling in finite time; job timeout detection
+     * and recovery (TDR) for AIE4 will be added in a separate change.
+     */

+    ret = dma_resv_wait_timeout(to_gobj(abo)->resv, 
DMA_RESV_USAGE_BOOKKEEP,

+                    false, MAX_SCHEDULE_TIMEOUT);
+    if (!ret)
+        XDNA_ERR(xdna, "Failed to wait for bo, ret %ld", ret);
        if (range->event == MMU_NOTIFY_UNMAP) {
          down_write(&xdna->notifier_lock);

diff --git a/drivers/accel/amdxdna/amdxdna_pci_drv.h 
b/drivers/accel/amdxdna/amdxdna_pci_drv.h

index 11f46ec738d7..2a1d6b33363c 100644
--- a/drivers/accel/amdxdna/amdxdna_pci_drv.h
+++ b/drivers/accel/amdxdna/amdxdna_pci_drv.h
@@ -63,7 +63,6 @@ struct amdxdna_dev_ops {

      int (*hwctx_config)(struct amdxdna_hwctx *hwctx, u32 type, u64 
value, void *buf, u32 size);
      int (*hwctx_sync_debug_bo)(struct amdxdna_hwctx *hwctx, u32 
debug_bo_hdl);
      int (*hwctx_heap_expand)(struct amdxdna_hwctx *hwctx, struct 
amdxdna_gem_obj *heap);
-    void (*hmm_invalidate)(struct amdxdna_gem_obj *abo, unsigned 
long cur_seq);
      int (*cmd_submit)(struct amdxdna_hwctx *hwctx, struct 
amdxdna_sched_job *job, u64 *seq);
      int (*cmd_wait)(struct amdxdna_hwctx *hwctx, u64 seq, u32 
timeout);
      int (*get_aie_info)(struct amdxdna_client *client, struct 
amdxdna_drm_get_info *args);























					Reply via email to