drm/xe: Enforce correct user fence signaling order using

Prevent application hangs caused by out-of-order fence signaling when
user fences are attached. Use drm_syncobj (via dma-fence-chain) to
guarantee that each user fence signals in order, regardless of the
signaling order of the attached fences. Ensure user fence writebacks to
user space occur in the correct sequence.

v7:
 - Skip drm_syncbj create of error (CI)

Fixes: dd08ebf ("drm/xe: Introduce a new DRM driver for Intel GPUs")
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Reviewed-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Link: https://patch.msgid.link/20251031234050.3043507-2-matthew.brost@intel.com

Author: mbrost05

drivers/gpu/drm/xe/xe_exec.c

@@ -173,7 +173,8 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 
 	for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {
 		err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs],
-					  &syncs_user[num_syncs], SYNC_PARSE_FLAG_EXEC |
+					  &syncs_user[num_syncs], NULL, 0,
+					  SYNC_PARSE_FLAG_EXEC |
 					  (xe_vm_in_lr_mode(vm) ?
 					   SYNC_PARSE_FLAG_LR_MODE : 0));
 		if (err)

drivers/gpu/drm/xe/xe_exec_queue.c

@@ -10,6 +10,7 @@
 #include <drm/drm_device.h>
 #include <drm/drm_drv.h>
 #include <drm/drm_file.h>
+#include <drm/drm_syncobj.h>
 #include <uapi/drm/xe_drm.h>
 
 #include "xe_dep_scheduler.h"
@@ -368,6 +369,16 @@ struct xe_exec_queue *xe_exec_queue_create_bind(struct xe_device *xe,
 	}
 	xe_vm_put(migrate_vm);
 
+	if (!IS_ERR(q)) {
+		int err = drm_syncobj_create(&q->ufence_syncobj,
+					     DRM_SYNCOBJ_CREATE_SIGNALED,
+					     NULL);
+		if (err) {
+			xe_exec_queue_put(q);
+			return ERR_PTR(err);
+		}
+	}
+
 	return q;
 }
 ALLOW_ERROR_INJECTION(xe_exec_queue_create_bind, ERRNO);
@@ -379,6 +390,9 @@ void xe_exec_queue_destroy(struct kref *ref)
 
 	xe_assert(gt_to_xe(q->gt), atomic_read(&q->job_cnt) == 0);
 
+	if (q->ufence_syncobj)
+		drm_syncobj_put(q->ufence_syncobj);
+
 	if (xe_exec_queue_uses_pxp(q))
 		xe_pxp_exec_queue_remove(gt_to_xe(q->gt)->pxp, q);
 

drivers/gpu/drm/xe/xe_exec_queue_types.h

@@ -15,6 +15,7 @@
 #include "xe_hw_fence_types.h"
 #include "xe_lrc_types.h"
 
+struct drm_syncobj;
 struct xe_execlist_exec_queue;
 struct xe_gt;
 struct xe_guc_exec_queue;
@@ -155,6 +156,12 @@ struct xe_exec_queue {
 		struct list_head link;
 	} pxp;
 
+	/** @ufence_syncobj: User fence syncobj */
+	struct drm_syncobj *ufence_syncobj;
+
+	/** @ufence_timeline_value: User fence timeline value */
+	u64 ufence_timeline_value;
+
 	/** @ops: submission backend exec queue operations */
 	const struct xe_exec_queue_ops *ops;
 

drivers/gpu/drm/xe/xe_oa.c

@@ -10,6 +10,7 @@
 
 #include <drm/drm_drv.h>
 #include <drm/drm_managed.h>
+#include <drm/drm_syncobj.h>
 #include <uapi/drm/xe_drm.h>
 
 #include <generated/xe_wa_oob.h>
@@ -1390,7 +1391,9 @@ static int xe_oa_user_extensions(struct xe_oa *oa, enum xe_oa_user_extn_from fro
 	return 0;
 }
 
-static int xe_oa_parse_syncs(struct xe_oa *oa, struct xe_oa_open_param *param)
+static int xe_oa_parse_syncs(struct xe_oa *oa,
+			     struct xe_oa_stream *stream,
+			     struct xe_oa_open_param *param)
 {
 	int ret, num_syncs, num_ufence = 0;
 
@@ -1410,7 +1413,9 @@ static int xe_oa_parse_syncs(struct xe_oa *oa, struct xe_oa_open_param *param)
 
 	for (num_syncs = 0; num_syncs < param->num_syncs; num_syncs++) {
 		ret = xe_sync_entry_parse(oa->xe, param->xef, &param->syncs[num_syncs],
-					  &param->syncs_user[num_syncs], 0);
+					  &param->syncs_user[num_syncs],
+					  stream->ufence_syncobj,
+					  ++stream->ufence_timeline_value, 0);
 		if (ret)
 			goto err_syncs;
 
@@ -1540,7 +1545,7 @@ static long xe_oa_config_locked(struct xe_oa_stream *stream, u64 arg)
 		return -ENODEV;
 
 	param.xef = stream->xef;
-	err = xe_oa_parse_syncs(stream->oa, &param);
+	err = xe_oa_parse_syncs(stream->oa, stream, &param);
 	if (err)
 		goto err_config_put;
 
@@ -1636,6 +1641,7 @@ static void xe_oa_destroy_locked(struct xe_oa_stream *stream)
 	if (stream->exec_q)
 		xe_exec_queue_put(stream->exec_q);
 
+	drm_syncobj_put(stream->ufence_syncobj);
 	kfree(stream);
 }
 
@@ -1827,6 +1833,7 @@ static int xe_oa_stream_open_ioctl_locked(struct xe_oa *oa,
 					  struct xe_oa_open_param *param)
 {
 	struct xe_oa_stream *stream;
+	struct drm_syncobj *ufence_syncobj;
 	int stream_fd;
 	int ret;
 
@@ -1837,17 +1844,31 @@ static int xe_oa_stream_open_ioctl_locked(struct xe_oa *oa,
 		goto exit;
 	}
 
+	ret = drm_syncobj_create(&ufence_syncobj, DRM_SYNCOBJ_CREATE_SIGNALED,
+				 NULL);
+	if (ret)
+		goto exit;
+
 	stream = kzalloc(sizeof(*stream), GFP_KERNEL);
 	if (!stream) {
 		ret = -ENOMEM;
-		goto exit;
+		goto err_syncobj;
 	}
-
+	stream->ufence_syncobj = ufence_syncobj;
 	stream->oa = oa;
-	ret = xe_oa_stream_init(stream, param);
+
+	ret = xe_oa_parse_syncs(oa, stream, param);
 	if (ret)
 		goto err_free;
 
+	ret = xe_oa_stream_init(stream, param);
+	if (ret) {
+		while (param->num_syncs--)
+			xe_sync_entry_cleanup(&param->syncs[param->num_syncs]);
+		kfree(param->syncs);
+		goto err_free;
+	}
+
 	if (!param->disabled) {
 		ret = xe_oa_enable_locked(stream);
 		if (ret)
@@ -1871,6 +1892,8 @@ static int xe_oa_stream_open_ioctl_locked(struct xe_oa *oa,
 	xe_oa_stream_destroy(stream);
 err_free:
 	kfree(stream);
+err_syncobj:
+	drm_syncobj_put(ufence_syncobj);
 exit:
 	return ret;
 }
@@ -2084,22 +2107,14 @@ int xe_oa_stream_open_ioctl(struct drm_device *dev, u64 data, struct drm_file *f
 		goto err_exec_q;
 	}
 
-	ret = xe_oa_parse_syncs(oa, &param);
-	if (ret)
-		goto err_exec_q;
-
 	mutex_lock(&param.hwe->gt->oa.gt_lock);
 	ret = xe_oa_stream_open_ioctl_locked(oa, &param);
 	mutex_unlock(&param.hwe->gt->oa.gt_lock);
 	if (ret < 0)
-		goto err_sync_cleanup;
+		goto err_exec_q;
 
 	return ret;
 
-err_sync_cleanup:
-	while (param.num_syncs--)
-		xe_sync_entry_cleanup(&param.syncs[param.num_syncs]);
-	kfree(param.syncs);
 err_exec_q:
 	if (param.exec_q)
 		xe_exec_queue_put(param.exec_q);

drivers/gpu/drm/xe/xe_oa_types.h

@@ -15,6 +15,8 @@
 #include "regs/xe_reg_defs.h"
 #include "xe_hw_engine_types.h"
 
+struct drm_syncobj;
+
 #define DEFAULT_XE_OA_BUFFER_SIZE SZ_16M
 
 enum xe_oa_report_header {
@@ -248,6 +250,12 @@ struct xe_oa_stream {
 	/** @xef: xe_file with which the stream was opened */
 	struct xe_file *xef;
 
+	/** @ufence_syncobj: User fence syncobj */
+	struct drm_syncobj *ufence_syncobj;
+
+	/** @ufence_timeline_value: User fence timeline value */
+	u64 ufence_timeline_value;
+
 	/** @last_fence: fence to use in stream destroy when needed */
 	struct dma_fence *last_fence;
 

drivers/gpu/drm/xe/xe_sync.c

@@ -113,6 +113,8 @@ static void user_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 int xe_sync_entry_parse(struct xe_device *xe, struct xe_file *xef,
 			struct xe_sync_entry *sync,
 			struct drm_xe_sync __user *sync_user,
+			struct drm_syncobj *ufence_syncobj,
+			u64 ufence_timeline_value,
 			unsigned int flags)
 {
 	struct drm_xe_sync sync_in;
@@ -192,10 +194,15 @@ int xe_sync_entry_parse(struct xe_device *xe, struct xe_file *xef,
 		if (exec) {
 			sync->addr = sync_in.addr;
 		} else {
+			sync->ufence_timeline_value = ufence_timeline_value;
 			sync->ufence = user_fence_create(xe, sync_in.addr,
 							 sync_in.timeline_value);
 			if (XE_IOCTL_DBG(xe, IS_ERR(sync->ufence)))
 				return PTR_ERR(sync->ufence);
+			sync->ufence_chain_fence = dma_fence_chain_alloc();
+			if (!sync->ufence_chain_fence)
+				return -ENOMEM;
+			sync->ufence_syncobj = ufence_syncobj;
 		}
 
 		break;
@@ -239,7 +246,12 @@ void xe_sync_entry_signal(struct xe_sync_entry *sync, struct dma_fence *fence)
 	} else if (sync->ufence) {
 		int err;
 
-		dma_fence_get(fence);
+		drm_syncobj_add_point(sync->ufence_syncobj,
+				      sync->ufence_chain_fence,
+				      fence, sync->ufence_timeline_value);
+		sync->ufence_chain_fence = NULL;
+
+		fence = drm_syncobj_fence_get(sync->ufence_syncobj);
 		user_fence_get(sync->ufence);
 		err = dma_fence_add_callback(fence, &sync->ufence->cb,
 					     user_fence_cb);
@@ -259,7 +271,8 @@ void xe_sync_entry_cleanup(struct xe_sync_entry *sync)
 		drm_syncobj_put(sync->syncobj);
 	dma_fence_put(sync->fence);
 	dma_fence_chain_free(sync->chain_fence);
-	if (sync->ufence)
+	dma_fence_chain_free(sync->ufence_chain_fence);
+	if (!IS_ERR_OR_NULL(sync->ufence))
 		user_fence_put(sync->ufence);
 }
 

drivers/gpu/drm/xe/xe_sync.h

@@ -8,6 +8,7 @@
 
 #include "xe_sync_types.h"
 
+struct drm_syncobj;
 struct xe_device;
 struct xe_exec_queue;
 struct xe_file;
@@ -21,6 +22,8 @@ struct xe_vm;
 int xe_sync_entry_parse(struct xe_device *xe, struct xe_file *xef,
 			struct xe_sync_entry *sync,
 			struct drm_xe_sync __user *sync_user,
+			struct drm_syncobj *ufence_syncobj,
+			u64 ufence_timeline_value,
 			unsigned int flags);
 int xe_sync_entry_add_deps(struct xe_sync_entry *sync,
 			   struct xe_sched_job *job);

drivers/gpu/drm/xe/xe_sync_types.h

@@ -18,9 +18,12 @@ struct xe_sync_entry {
 	struct drm_syncobj *syncobj;
 	struct dma_fence *fence;
 	struct dma_fence_chain *chain_fence;
+	struct dma_fence_chain *ufence_chain_fence;
+	struct drm_syncobj *ufence_syncobj;
 	struct xe_user_fence *ufence;
 	u64 addr;
 	u64 timeline_value;
+	u64 ufence_timeline_value;
 	u32 type;
 	u32 flags;
 };

drivers/gpu/drm/xe/xe_vm.c

@@ -3633,8 +3633,12 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
 
 	syncs_user = u64_to_user_ptr(args->syncs);
 	for (num_syncs = 0; num_syncs < args->num_syncs; num_syncs++) {
+		struct xe_exec_queue *__q = q ?: vm->q[0];
+
 		err = xe_sync_entry_parse(xe, xef, &syncs[num_syncs],
 					  &syncs_user[num_syncs],
+					  __q->ufence_syncobj,
+					  ++__q->ufence_timeline_value,
 					  (xe_vm_in_lr_mode(vm) ?
 					   SYNC_PARSE_FLAG_LR_MODE : 0) |
 					  (!args->num_binds ?