drm/i915: Fix a VMA UAF for multi-gt platform

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.
v3: Add more comment(Andi)
    Use the new gt var when possible(Andrzej)

Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Chris Wilson <chris.p.wilson@intel.com>
Cc: Andi Shyti <andi.shyti@linux.intel.com>
Cc: Andrzej Hajda <andrzej.hajda@intel.com>
Cc: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@intel.com>
Signed-off-by: Nirmoy Das <nirmoy.das@intel.com>
Tested-by: Andi Shyti <andi.shyti@linux.intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Reviewed-by: Andrzej Hajda <andrzej.hajda@intel.com>
Tested-by: Sushma Venkatesh Reddy <sushma.venkatesh.reddy@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230608110103.777594-1-andi.shyti@linux.intel.com

1 files changed, 19 insertions, 2 deletions

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
index 5fb459ea4294..1de9de1e4782 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c
@@ -2692,6 +2692,7 @@ static int
 eb_select_engine(struct i915_execbuffer *eb)
 {
 	struct intel_context *ce, *child;
+	struct intel_gt *gt;
 	unsigned int idx;
 	int err;
 
@@ -2715,10 +2716,17 @@ 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);
+	intel_gt_pm_get(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);
@@ -2757,7 +2765,10 @@ eb_select_engine(struct i915_execbuffer *eb)
 	return err;
 
 err:
-	intel_gt_pm_put(ce->engine->gt);
+	if (gt->info.id)
+		intel_gt_pm_put(to_gt(gt->i915));
+
+	intel_gt_pm_put(gt);
 	for_each_child(ce, child)
 		intel_context_put(child);
 	intel_context_put(ce);
@@ -2770,6 +2781,12 @@ eb_put_engine(struct i915_execbuffer *eb)
 	struct intel_context *child;
 
 	i915_vm_put(eb->context->vm);
+	/*
+	 * This works in conjunction with eb_select_engine() to prevent
+	 * i915_vma_parked() from interfering while execbuf validates vmas.
+	 */
+	if (eb->gt->info.id)
+		intel_gt_pm_put(to_gt(eb->gt->i915));
 	intel_gt_pm_put(eb->gt);
 	for_each_child(eb->context, child)
 		intel_context_put(child);