<html><head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
  </head>
  <body>
    <p><br>
    </p>
    <div class="moz-cite-prefix">On 2021-05-12 11:54 a.m., Felix
      Kuehling wrote:<br>
    </div>
    <blockquote type="cite" cite="mid:3fa73d87-bd13-4f60-b737-a5a7e0293a1d@amd.com">
      <pre class="moz-quote-pre" wrap="">Am 2021-05-12 um 8:38 a.m. schrieb Christian König:
</pre>
      <blockquote type="cite">
        <pre class="moz-quote-pre" wrap="">

Am 12.05.21 um 14:34 schrieb Philip Yang:
</pre>
        <blockquote type="cite">
          <pre class="moz-quote-pre" wrap="">Mapping huge page, 2MB aligned address with 2MB size, uses PDE0 as PTE.
If previously valid PDE0, PDE0.V=1 and PDE0.P=0 turns into PTE, this
requires TLB flush, otherwise page table walker will not read updated
PDE0.

Change page table update mapping to return free_table flag to indicate
the previously valid PDE may have turned into a PTE if page table is
freed.

Signed-off-by: Philip Yang <a class="moz-txt-link-rfc2396E" href="mailto:Philip.Yang@amd.com"><Philip.Yang@amd.com></a>
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 17 +++++++++++------
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h |  2 +-
  drivers/gpu/drm/amd/amdkfd/kfd_svm.c   | 12 ++++++++++--
  3 files changed, 22 insertions(+), 9 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index 3dcdcc9ff522..27418f9407f1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -1457,7 +1457,7 @@ static void amdgpu_vm_fragment(struct
amdgpu_vm_update_params *params,
   */
  static int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params
*params,
                   uint64_t start, uint64_t end,
-                 uint64_t dst, uint64_t flags)
+                 uint64_t dst, uint64_t flags, bool *free_table)
</pre>
        </blockquote>
        <pre class="moz-quote-pre" wrap="">
Please put that flag into the params structure instead.

Christian.
</pre>
      </blockquote>
      <pre class="moz-quote-pre" wrap="">
I agree. Christian, I think we also need to keep track of this for
graphics command submissions. amdgpu_cs_vm_handling needs get this flag
from amdgpu_vm_bo_update, and amdgpu_cs_submit needs to update
job->vm_needs_flushed based on this.
</pre>
    </blockquote>
    <p>amdgpu_vm_bo_update to pass NULL to ignore the free_table return
      flag.</p>
    <p>This new flag only used by SVM APIs. Old KFD and graphics command
      submission path is not changed.</p>
    <p>Thanks,</p>
    <p>Philip<br>
    </p>
    <blockquote type="cite" cite="mid:3fa73d87-bd13-4f60-b737-a5a7e0293a1d@amd.com">
      <pre class="moz-quote-pre" wrap="">
Regards,
  Felix


</pre>
      <blockquote type="cite">
        <pre class="moz-quote-pre" wrap="">
</pre>
        <blockquote type="cite">
          <pre class="moz-quote-pre" wrap="">  {
      struct amdgpu_device *adev = params->adev;
      struct amdgpu_vm_pt_cursor cursor;
@@ -1583,6 +1583,8 @@ static int amdgpu_vm_update_ptes(struct
amdgpu_vm_update_params *params,
              while (cursor.pfn < frag_start) {
                  amdgpu_vm_free_pts(adev, params->vm, &cursor);
                  amdgpu_vm_pt_next(adev, &cursor);
+                if (free_table)
+                    *free_table = true;
              }
            } else if (frag >= shift) {
@@ -1610,6 +1612,7 @@ static int amdgpu_vm_update_ptes(struct
amdgpu_vm_update_params *params,
   * @nodes: array of drm_mm_nodes with the MC addresses
   * @pages_addr: DMA addresses to use for mapping
   * @fence: optional resulting fence
+ * @free_table: return true if page table is freed
   *
   * Fill in the page table entries between @start and @last.
   *
@@ -1624,7 +1627,8 @@ int amdgpu_vm_bo_update_mapping(struct
amdgpu_device *adev,
                  uint64_t flags, uint64_t offset,
                  struct drm_mm_node *nodes,
                  dma_addr_t *pages_addr,
-                struct dma_fence **fence)
+                struct dma_fence **fence,
+                bool *free_table)
  {
      struct amdgpu_vm_update_params params;
      enum amdgpu_sync_mode sync_mode;
@@ -1721,7 +1725,8 @@ int amdgpu_vm_bo_update_mapping(struct
amdgpu_device *adev,
          }
            tmp = start + num_entries;
-        r = amdgpu_vm_update_ptes(&params, start, tmp, addr, flags);
+        r = amdgpu_vm_update_ptes(&params, start, tmp, addr, flags,
+                      free_table);
          if (r)
              goto error_unlock;
  @@ -1879,7 +1884,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device
*adev, struct amdgpu_bo_va *bo_va,
                          resv, mapping->start,
                          mapping->last, update_flags,
                          mapping->offset, nodes,
-                        pages_addr, last_update);
+                        pages_addr, last_update, NULL);
          if (r)
              return r;
      }
@@ -2090,7 +2095,7 @@ int amdgpu_vm_clear_freed(struct amdgpu_device
*adev,
          r = amdgpu_vm_bo_update_mapping(adev, adev, vm, false, false,
                          resv, mapping->start,
                          mapping->last, init_pte_value,
-                        0, NULL, NULL, &f);
+                        0, NULL, NULL, &f, NULL);
          amdgpu_vm_free_mapping(adev, vm, mapping, f);
          if (r) {
              dma_fence_put(f);
@@ -3428,7 +3433,7 @@ bool amdgpu_vm_handle_fault(struct
amdgpu_device *adev, u32 pasid,
      }
        r = amdgpu_vm_bo_update_mapping(adev, adev, vm, true, false,
NULL, addr,
-                    addr, flags, value, NULL, NULL,
+                    addr, flags, value, NULL, NULL, NULL,
                      NULL);
      if (r)
          goto error_unlock;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
index ea60ec122b51..f61c20b70b79 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h
@@ -404,7 +404,7 @@ int amdgpu_vm_bo_update_mapping(struct
amdgpu_device *adev,
                  uint64_t flags, uint64_t offset,
                  struct drm_mm_node *nodes,
                  dma_addr_t *pages_addr,
-                struct dma_fence **fence);
+                struct dma_fence **fence, bool *free_table);
  int amdgpu_vm_bo_update(struct amdgpu_device *adev,
              struct amdgpu_bo_va *bo_va,
              bool clear);
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
index b665e9ff77e3..4f28052d44bf 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c
@@ -1084,7 +1084,7 @@ svm_range_unmap_from_gpu(struct amdgpu_device
*adev, struct amdgpu_vm *vm,
        return amdgpu_vm_bo_update_mapping(adev, adev, vm, false,
true, NULL,
                         start, last, init_pte_value, 0,
-                       NULL, NULL, fence);
+                       NULL, NULL, fence, NULL);
  }
    static int
@@ -1137,12 +1137,15 @@ svm_range_map_to_gpu(struct amdgpu_device
*adev, struct amdgpu_vm *vm,
               struct amdgpu_device *bo_adev, struct dma_fence **fence)
  {
      struct amdgpu_bo_va bo_va;
+    bool free_table = false;
+    struct kfd_process *p;
      uint64_t pte_flags;
      int r = 0;
        pr_debug("svms 0x%p [0x%lx 0x%lx]\n", prange->svms,
prange->start,
           prange->last);
  +    p = container_of(prange->svms, struct kfd_process, svms);
      if (prange->svm_bo && prange->ttm_res) {
          bo_va.is_xgmi = amdgpu_xgmi_same_hive(adev, bo_adev);
          prange->mapping.bo_va = &bo_va;
@@ -1159,7 +1162,8 @@ svm_range_map_to_gpu(struct amdgpu_device
*adev, struct amdgpu_vm *vm,
                      prange->mapping.offset,
                      prange->ttm_res ?
                          prange->ttm_res->mm_node : NULL,
-                    dma_addr, &vm->last_update);
+                    dma_addr, &vm->last_update,
+                    &free_table);
      if (r) {
          pr_debug("failed %d to map to gpu 0x%lx\n", r, prange->start);
          goto out;
@@ -1175,6 +1179,10 @@ svm_range_map_to_gpu(struct amdgpu_device
*adev, struct amdgpu_vm *vm,
      if (fence)
          *fence = dma_fence_get(vm->last_update);
  +    if (free_table)
+        amdgpu_amdkfd_flush_gpu_tlb_pasid((struct kgd_dev *)adev,
+                          p->pasid);
+
  out:
      prange->mapping.bo_va = NULL;
      return r;
</pre>
        </blockquote>
        <pre class="moz-quote-pre" wrap="">
_______________________________________________
amd-gfx mailing list
<a class="moz-txt-link-abbreviated" href="mailto:amd-gfx@lists.freedesktop.org">amd-gfx@lists.freedesktop.org</a>
<a class="moz-txt-link-freetext" href="https://lists.freedesktop.org/mailman/listinfo/amd-gfx">https://lists.freedesktop.org/mailman/listinfo/amd-gfx</a>
</pre>
      </blockquote>
    </blockquote>
  </body>
</html>