drm/locking: add backtrace for locking contended locks without backoff

If drm_modeset_lock() returns -EDEADLK, the caller is supposed to drop
all currently held locks using drm_modeset_backoff(). Failing to do so
will result in warnings and backtraces on the paths trying to lock a
contended lock. Add support for optionally printing the backtrace on the
path that hit the deadlock and didn't gracefully handle the situation.

For example, the patch [1] inadvertently dropped the return value check
and error return on replacing calc_watermark_data() with
intel_compute_global_watermarks(). The backtraces on the subsequent
locking paths hitting WARN_ON(ctx->contended) were unhelpful, but adding
the backtrace to the deadlock path produced this helpful printout:

<7> [98.002465] drm_modeset_lock attempting to lock a contended lock without backoff:
   drm_modeset_lock+0x107/0x130
   drm_atomic_get_plane_state+0x76/0x150
   skl_compute_wm+0x251d/0x2b20 [i915]
   intel_atomic_check+0x1942/0x29e0 [i915]
   drm_atomic_check_only+0x554/0x910
   drm_atomic_nonblocking_commit+0xe/0x50
   drm_mode_atomic_ioctl+0x8c2/0xab0
   drm_ioctl_kernel+0xac/0x140

Add new CONFIG_DRM_DEBUG_MODESET_LOCK to enable modeset lock debugging
with stack depot and trace.

[1] https://lore.kernel.org/r/20210924114741.15940-4-jani.nikula@intel.com

v2:
- default y if DEBUG_WW_MUTEX_SLOWPATH (Daniel)
- depends on DEBUG_KERNEL

Cc: Daniel Vetter <daniel@ffwll.ch>
Cc: Dave Airlie <airlied@gmail.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211001091444.8177-1-jani.nikula@intel.com

1 files changed, 47 insertions, 2 deletions

diff --git a/drivers/gpu/drm/drm_modeset_lock.c b/drivers/gpu/drm/drm_modeset_lock.c
index fcfe1a03c4a1..62f6fe482d25 100644
--- a/drivers/gpu/drm/drm_modeset_lock.c
+++ b/drivers/gpu/drm/drm_modeset_lock.c
@@ -25,6 +25,7 @@
 #include <drm/drm_crtc.h>
 #include <drm/drm_device.h>
 #include <drm/drm_modeset_lock.h>
+#include <drm/drm_print.h>
 
 /**
  * DOC: kms locking
@@ -77,6 +78,45 @@
 
 static DEFINE_WW_CLASS(crtc_ww_class);
 
+#if IS_ENABLED(CONFIG_DRM_DEBUG_MODESET_LOCK)
+static noinline depot_stack_handle_t __stack_depot_save(void)
+{
+	unsigned long entries[8];
+	unsigned int n;
+
+	n = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
+
+	return stack_depot_save(entries, n, GFP_NOWAIT | __GFP_NOWARN);
+}
+
+static void __stack_depot_print(depot_stack_handle_t stack_depot)
+{
+	struct drm_printer p = drm_debug_printer("drm_modeset_lock");
+	unsigned long *entries;
+	unsigned int nr_entries;
+	char *buf;
+
+	buf = kmalloc(PAGE_SIZE, GFP_NOWAIT | __GFP_NOWARN);
+	if (!buf)
+		return;
+
+	nr_entries = stack_depot_fetch(stack_depot, &entries);
+	stack_trace_snprint(buf, PAGE_SIZE, entries, nr_entries, 2);
+
+	drm_printf(&p, "attempting to lock a contended lock without backoff:\n%s", buf);
+
+	kfree(buf);
+}
+#else /* CONFIG_DRM_DEBUG_MODESET_LOCK */
+static depot_stack_handle_t __stack_depot_save(void)
+{
+	return 0;
+}
+static void __stack_depot_print(depot_stack_handle_t stack_depot)
+{
+}
+#endif /* CONFIG_DRM_DEBUG_MODESET_LOCK */
+
 /**
  * drm_modeset_lock_all - take all modeset locks
  * @dev: DRM device
@@ -225,7 +265,9 @@ EXPORT_SYMBOL(drm_modeset_acquire_fini);
  */
 void drm_modeset_drop_locks(struct drm_modeset_acquire_ctx *ctx)
 {
-	WARN_ON(ctx->contended);
+	if (WARN_ON(ctx->contended))
+		__stack_depot_print(ctx->stack_depot);
+
 	while (!list_empty(&ctx->locked)) {
 		struct drm_modeset_lock *lock;
 
@@ -243,7 +285,8 @@ static inline int modeset_lock(struct drm_modeset_lock *lock,
 {
 	int ret;
 
-	WARN_ON(ctx->contended);
+	if (WARN_ON(ctx->contended))
+		__stack_depot_print(ctx->stack_depot);
 
 	if (ctx->trylock_only) {
 		lockdep_assert_held(&ctx->ww_ctx);
@@ -274,6 +317,7 @@ static inline int modeset_lock(struct drm_modeset_lock *lock,
 		ret = 0;
 	} else if (ret == -EDEADLK) {
 		ctx->contended = lock;
+		ctx->stack_depot = __stack_depot_save();
 	}
 
 	return ret;
@@ -296,6 +340,7 @@ int drm_modeset_backoff(struct drm_modeset_acquire_ctx *ctx)
 	struct drm_modeset_lock *contended = ctx->contended;
 
 	ctx->contended = NULL;
+	ctx->stack_depot = 0;
 
 	if (WARN_ON(!contended))
 		return 0;