On 6/6/2023 10:56 PM, Andi Shyti wrote:
Hi Nirmoy,
On Tue, Jun 06, 2023 at 10:27:55PM +0200, Nirmoy Das wrote:
Ensure correct handling of closed VMAs on multi-gt platforms to prevent
Use-After-Free. Currently, when GT0 goes idle, closed VMAs that are
exclusively added to GT0's closed_vma link (gt->closed_vma) and
subsequently freed by i915_vma_parked(), which assumes the entire GPU is
idle. However, on platforms with multiple GTs, such as MTL, GT1 may
remain active while GT0 is idle. This causes GT0 to mistakenly consider
the closed VMAs in its closed_vma list as unnecessary, potentially
leading to Use-After-Free issues if a job for GT1 attempts to access a
freed VMA.
Although we do take a wakeref for GT0 but it happens later, after
evaluating VMAs. To mitigate this, it is necessary to hold a GT0 wakeref
early.
v2: Use gt id to detect multi-tile(Andi)
Fix the incorrect error path.
Cc: Joonas Lahtinen <joonas.lahtinen@xxxxxxxxxxxxxxx>
Cc: Rodrigo Vivi <rodrigo.vivi@xxxxxxxxx>
Cc: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxxxxxxx>
Cc: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
Cc: Chris Wilson <chris.p.wilson@xxxxxxxxx>
Cc: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>
Cc: Andrzej Hajda <andrzej.hajda@xxxxxxxxx>
Cc: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@xxxxxxxxx>
Tested-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>
Signed-off-by: Nirmoy Das <nirmoy.das@xxxxxxxxx>
I wonder if we need any Fixes tag here, maybe this:
Fixes: d93939730347 ("drm/i915: Remove the vma refcount")
No, vma refcount is not enough unfortunately. Issue is i915_vma_parked()
expects only one GT.
---
drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index 3aeede6aee4d..c2a67435acfa 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -2683,6 +2683,7 @@ static int
eb_select_engine(struct i915_execbuffer *eb)
{
struct intel_context *ce, *child;
+ struct intel_gt *gt;
unsigned int idx;
int err;
@@ -2706,10 +2707,16 @@ eb_select_engine(struct i915_execbuffer *eb)
}
}
eb->num_batches = ce->parallel.number_children + 1;
+ gt = ce->engine->gt;
for_each_child(ce, child)
intel_context_get(child);
intel_gt_pm_get(ce->engine->gt);
+ /* Keep GT0 active on MTL so that i915_vma_parked() doesn't
+ * free VMAs while execbuf ioctl is validating VMAs.
+ */
+ if (gt->info.id)
+ intel_gt_pm_get(to_gt(gt->i915));
if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
err = intel_context_alloc_state(ce);
@@ -2748,6 +2755,9 @@ eb_select_engine(struct i915_execbuffer *eb)
return err;
err:
+ if (gt->info.id)
+ intel_gt_pm_put(to_gt(gt->i915));
+
intel_gt_pm_put(ce->engine->gt);
for_each_child(ce, child)
intel_context_put(child);
@@ -2761,6 +2771,8 @@ eb_put_engine(struct i915_execbuffer *eb)
struct intel_context *child;
i915_vm_put(eb->context->vm);
+ if (eb->gt->info.id)
+ intel_gt_pm_put(to_gt(eb->gt->i915));
intel_gt_pm_put(eb->gt);
I would add a comment up here, just not to scare people when they
see this.
I can add a comment pairing comment mentioning
eb_select_engine().
Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx>
Thanks,
Nirmoy
Andi
for_each_child(eb->context, child)
intel_context_put(child);
--
2.39.0
