Re: [PATCH v5 4/4] drm/i915: add guard page to ggtt->error_capture

From: Tvrtko Ursulin
Date: Thu Mar 09 2023 - 05:18:34 EST


On 09/03/2023 09:59, Andrzej Hajda wrote:


On 09.03.2023 10:43, Tvrtko Ursulin wrote:

On 09/03/2023 09:34, Andrzej Hajda wrote:


On 09.03.2023 10:08, Tvrtko Ursulin wrote:

On 08/03/2023 15:39, Andrzej Hajda wrote:
Write-combining memory allows speculative reads by CPU.
ggtt->error_capture is WC mapped to CPU, so CPU/MMU can try
to prefetch memory beyond the error_capture, ie it tries
to read memory pointed by next PTE in GGTT.
If this PTE points to invalid address DMAR errors will occur.
This behaviour was observed on ADL and RPL platforms.
To avoid it, guard scratch page should be added after error_capture.
The patch fixes the most annoying issue with error capture but
since WC reads are used also in other places there is a risk similar
problem can affect them as well.

v2:
   - modified commit message (I hope the diagnosis is correct),
   - added bug checks to ensure scratch is initialized on gen3 platforms.
     CI produces strange stacktrace for it suggesting scratch[0] is NULL,
     to be removed after resolving the issue with gen3 platforms.
v3:
   - removed bug checks, replaced with gen check.
v4:
   - change code for scratch page insertion to support all platforms,
   - add info in commit message there could be more similar issues
v5:
   - check for nop_clear_range instead of gen8 (Tvrtko),
   - re-insert scratch pages on resume (Tvrtko)

Signed-off-by: Andrzej Hajda <andrzej.hajda@xxxxxxxxx>
Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>
---
  drivers/gpu/drm/i915/gt/intel_ggtt.c | 35 +++++++++++++++++++++++++++++++----
  1 file changed, 31 insertions(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/i915/gt/intel_ggtt.c b/drivers/gpu/drm/i915/gt/intel_ggtt.c
index b925da42c7cfc4..8fb700fde85c8f 100644
--- a/drivers/gpu/drm/i915/gt/intel_ggtt.c
+++ b/drivers/gpu/drm/i915/gt/intel_ggtt.c
@@ -502,6 +502,21 @@ static void cleanup_init_ggtt(struct i915_ggtt *ggtt)
      mutex_destroy(&ggtt->error_mutex);
  }
  +static void
+ggtt_insert_scratch_pages(struct i915_ggtt *ggtt, u64 offset, u64 length)
+{
+    struct i915_address_space *vm = &ggtt->vm;
+
+    if (vm->clear_range != nop_clear_range)

Hm I thought usually we would add a prefix for exported stuff, like in this case i915_vm_nop_clear_range, however I see intel_gtt.h exports a bunch of stuff with no prefixes already so I guess you could continue like that by inertia. The conundrum also could have been avoided if you left it static (leaving out dpt and mock_gtt patches) but no strong opinion from me.

+        return vm->clear_range(vm, offset, length);
+
+    while (length > 0) {
+        vm->insert_page(vm, px_dma(vm->scratch[0]), offset, I915_CACHE_NONE, 0);
+        offset += I915_GTT_PAGE_SIZE;
+        length -= I915_GTT_PAGE_SIZE;
+    }
+}
+
  static int init_ggtt(struct i915_ggtt *ggtt)
  {
      /*
@@ -550,8 +565,12 @@ static int init_ggtt(struct i915_ggtt *ggtt)
           * paths, and we trust that 0 will remain reserved. However,
           * the only likely reason for failure to insert is a driver
           * bug, which we expect to cause other failures...
+         *
+         * Since CPU can perform speculative reads on error capture
+         * (write-combining allows it) add scratch page after error
+         * capture to avoid DMAR errors.
           */
-        ggtt->error_capture.size = I915_GTT_PAGE_SIZE;
+        ggtt->error_capture.size = 2 * I915_GTT_PAGE_SIZE;
          ggtt->error_capture.color = I915_COLOR_UNEVICTABLE;
          if (drm_mm_reserve_node(&ggtt->vm.mm, &ggtt->error_capture))
drm_mm_insert_node_in_range(&ggtt->vm.mm,
@@ -561,11 +580,15 @@ static int init_ggtt(struct i915_ggtt *ggtt)
                              0, ggtt->mappable_end,
                              DRM_MM_INSERT_LOW);
      }
-    if (drm_mm_node_allocated(&ggtt->error_capture))
+    if (drm_mm_node_allocated(&ggtt->error_capture)) {
+        u64 start = ggtt->error_capture.start;
+        u64 size = ggtt->error_capture.size;
+
+        ggtt_insert_scratch_pages(ggtt, start, size);
          drm_dbg(&ggtt->vm.i915->drm,
              "Reserved GGTT:[%llx, %llx] for use by error capture\n",
-            ggtt->error_capture.start,
-            ggtt->error_capture.start + ggtt->error_capture.size);
+            start, start + size);
+    }
        /*
       * The upper portion of the GuC address space has a sizeable hole
@@ -1256,6 +1279,10 @@ void i915_ggtt_resume(struct i915_ggtt *ggtt)
        flush = i915_ggtt_resume_vm(&ggtt->vm);
  +    if (drm_mm_node_allocated(&ggtt->error_capture))
+        ggtt_insert_scratch_pages(ggtt, ggtt->error_capture.start,
+                      ggtt->error_capture.size);

Maybe it belongs in i915_ggtt_resume_vm since that one deals with PTEs? Looks like it to me, but ack either way.

i915_ggtt_resume_vm is called for ggtt and dpt. Of course I could add conditionals there checking if it is ggtt, but in such situation i915_ggtt_resume seems more natural candidate.

"if (drm_mm_node_allocated(&ggtt->error_capture))" check would handle that automatically, no? i915_ggtt_resume has nothing about PTEs at the moment..

Yes but since i915_ggtt_resume_vm  has vm as an argument (ie it operates on generic vm), there will be needed downcasting somewhere:
if (vm->is_ggtt) {
    struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
    if (drm_mm_node_allocated(&ggtt->error_capture))
        ...
}

In i915_ggtt_resume we have it for free, but moreover i915_ggtt_resume_vm (despite its name) seems to handle common stuff of ggtt and dpt, and i915_ggtt_resume looks as specific for ggtt, similarly intel_dpt_resume is specific for dpt.
If it does not convince you, I will update patch with above code.

Right, I see your point - I was mislead by the name i915_ggtt_resume_vm thinking it signifies it working on i915_ggtt. It's all good then.

Regards,

Tvrtko