[PATCH v3] drm/xe: Guard page-fault worker with runtime PM check

[Varun Gupta <varun.gupta@intel.com>]

During VM teardown, the VM's runtime PM reference is dropped
asynchronously, allowing the device to autosuspend while stale page
faults belonging to the now-dead VM are still queued. When the
page-fault worker later tries to ack one of these, it calls into
guc_ct_send_locked() on an already-suspended device, tripping:

  Assertion `!xe_pm_runtime_suspended(xe)` failed!
  WARNING at xe_device.c:1267 xe_device_assert_mem_access+0x11c/0x140 [xe]

A live VM/exec queue always holds a PM reference while it has
outstanding work, so if the device is suspended at ack time, the
owning context is already gone and the fault is stale.

Take a runtime PM reference across the entire pagefault
queue worker to safely deliver acks for torn-down VMs.

v3:
 - Move PM ref to the generic xe_pagefault_queue_work using
   guard(xe_pm_runtime)(xe) instead of tracking it in the GuC
   backend(Matt Brost).

v2:
 - Hold PM ref across the entire batch (begin/end) instead of per-ack.
   This prevents the device from autosuspending mid-batch, which would
   leave write_only acks written but the end flush skipped, and skip
   counter++, desyncing the cadence check.(Himal)
 - Add a comment explaining stale faults.(Himal)

Fixes: f289f7807119 ("drm/xe: Add xe_guc_pagefault layer")
Reported-by: Szymon Markiewicz <szymon.markiewicz@intel.com>
Signed-off-by: Varun Gupta <varun.gupta@intel.com>
Reviewed-by: Matthew Brost <matthew.brost@intel.com>
---
 drivers/gpu/drm/xe/xe_pagefault.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/drivers/gpu/drm/xe/xe_pagefault.c b/drivers/gpu/drm/xe/xe_pagefault.c
index 2e415995f067..3e8e8f7f5f40 100644
--- a/drivers/gpu/drm/xe/xe_pagefault.c
+++ b/drivers/gpu/drm/xe/xe_pagefault.c
@@ -17,6 +17,7 @@
 #include "xe_log.h"
 #include "xe_pagefault.h"
 #include "xe_pagefault_types.h"
+#include "xe_pm.h"
 #include "xe_svm.h"
 #include "xe_trace_bo.h"
 #include "xe_vm.h"
@@ -604,6 +605,13 @@ static void xe_pagefault_queue_work(struct work_struct *w)
 	u64 cache_start = XE_PAGEFAULT_CACHE_START_INVALID, cache_end = 0;
 	u32 cache_asid = 0;
 
+	/*
+	 * A live VM holds a PM reference, but a torn-down VM does not.
+	 * Guard the entire worker loop to safely drain stale faults and
+	 * prevent autosuspends from desyncing batched CT flushes.
+	 */
+	guard(xe_pm_runtime)(xe);
+
 #define USM_QUEUE_MAX_RUNTIME_MS      20
 	threshold = jiffies + msecs_to_jiffies(USM_QUEUE_MAX_RUNTIME_MS);
 
-- 
2.43.0