Skip to contentSkip to conversation

RFC: drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

PATCH Superseded Bommu Krishnaiah 2 participants 5 messages Dec 6, 2023

Subsystems (3)

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v2 cover RFC: drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

remove the num_engines/instances members from drm_xe_wait_user_fence structure and add a exec_queue_id member

Currently xe_wait_user_fence_ioctl is not checking exec_queue state and blocking until timeout, with this patch wakeup the blocking wait if exec_queue reset happen and returning proper error code

I am able to see exec_queue reset was happened and xe_wait_user_fence_ioctl returned EIO test output

root@DUT7075PVC:/home/gta# LD_LIBRARY_PATH=/home/gta/ ./xe_waitfence --r invalid-exec_queue-wait
IGT-Version: 1.28-g3c0162fc4 (x86_64) (Linux: 6.6.0-rc3-xe x86_64)
Opened device: /dev/dri/card0
Starting subtest: invalid-exec_queue-wait
Subtest invalid-exec_queue-wait: SUCCESS (0.993s)

test dmesg

[  807.680378] [IGT] xe_waitfence: executing
[  807.699796] [drm:drm_stub_open [drm]]
[  807.704536] xe 0000:51:00.0: [drm:drm_open_helper [drm]] comm="xe_waitfence", pid=2952, minor=0
[  807.715155] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, DRM_IOCTL_VERSION
[  807.727328] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, DRM_IOCTL_VERSION
[  807.739580] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.751518] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.763550] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.775525] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.787556] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.799494] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.811531] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.823476] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_DEVICE_QUERY
[  807.835577] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, DRM_IOCTL_VERSION
[  807.847921] [IGT] xe_waitfence: starting subtest invalid-exec_queue-wait
[  807.855528] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_VM_CREATE
[  807.891346] xe 0000:51:00.0: [drm:xe_reg_sr_apply_mmio [xe]] GT0: Applying GT save-restore MMIOs
[  807.901602] xe 0000:51:00.0: [drm:xe_reg_sr_apply_mmio [xe]] GT0: REG[0x9424] = 0x7ffffffc
-----------------
-----------------
[  808.560967] xe REG[0x4500-0x45ff]: deny rw access
[  808.566292] xe REG[0x1e3a8-0x1e3af]: allow read access
[  808.572161] xe 0000:51:00.0: [drm:xe_reg_sr_apply_mmio [xe]] GT0: Applying ccs3 save-restore MMIOs
[  808.582462] xe 0000:51:00.0: [drm:xe_reg_sr_apply_mmio [xe]] GT0: REG[0x260c4] = 0x3f7e0104
[  808.592096] xe 0000:51:00.0: [drm:xe_reg_sr_apply_whitelist [xe]] Whitelisting ccs3 registers
[  808.601962] xe REG[0x4400-0x45ff]: deny rw access
[  808.607281] xe REG[0x4500-0x45ff]: deny rw access
[  808.612608] xe REG[0x263a8-0x263af]: allow read access
[  808.618477] xe 0000:51:00.0: [drm] GT0: resumed
[  808.626283] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_EXEC_QUEUE_CREATE
[  808.638765] krishna xe_exec_queue_create_ioctl
[  808.645592] krishna args->exec_queue_id = 1
[  808.650328] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_GEM_CREATE
[  808.662621] xe 0000:51:00.0: [drm:xe_migrate_clear [xe]] Pass 0, size: 262144
[  808.672889] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_GEM_MMAP_OFFSET
[  808.685733] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_EXEC
[  808.696900] krishna args->exec_queue_id = 1
[  808.702700] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_WAIT_USER_FENCE
[  808.704620] xe 0000:51:00.0: [drm:pf_queue_work_func [xe]]
                ASID: 1048575
                VFID: 0
                PDATA: 0x00a3
                Faulted Address: 0x00000000001a0000
                FaultType: 0
                AccessType: 0
                FaultLevel: 4
                EngineClass: 3
                EngineInstance: 0
[  808.750685] xe 0000:51:00.0: [drm:pf_queue_work_func [xe]] Fault response: Unsuccessful -22
[  808.760519] xe 0000:51:00.0: [drm:xe_guc_exec_queue_memory_cat_error_handler [xe]] Engine memory cat error: guc_id=2
[  808.773237] xe 0000:51:00.0: [drm] exec gueue reset detected
[  808.773965] xe 0000:51:00.0: [drm] Timedout job: seqno=4294967169, guc_id=2, flags=0x8
[  808.779632] xe 0000:51:00.0: [drm:xe_wait_user_fence_ioctl [xe]] Ioctl argument check failed at drivers/gpu/drm/xe/xe_wait_user_fence.c:174: err < 0
[  808.789655] xe 0000:51:00.0: [drm] Xe device coredump has been created
[  808.803796] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence", pid=2952, ret=-5
[  808.811133] xe 0000:51:00.0: [drm] Check your /sys/class/drm/card0/device/devcoredump/data
[  808.811220] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, XE_EXEC_QUEUE_DESTROY
[  808.823605] xe 0000:51:00.0: [drm] Engine reset: guc_id=2
[  808.829862] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, DRM_IOCTL_GEM_CLOSE
[  808.843312] xe 0000:51:00.0: [drm:guc_exec_queue_timedout_job [xe]] Timedout signaled job: seqno=4294967169, guc_id=2, flags=0x9
[  808.848707] [IGT] xe_waitfence: finished subtest invalid-exec_queue-wait, SUCCESS
[  808.882255] xe 0000:51:00.0: [drm:drm_ioctl [drm]] comm="xe_waitfence" pid=2952, dev=0xe200, auth=1, DRM_IOCTL_VERSION
[  808.894404] xe 0000:51:00.0: [drm:drm_file_free.part.0 [drm]] comm="xe_waitfence", pid=2952, dev=0xe200, open_count=1
[  808.907374] xe 0000:51:00.0: [drm:drm_lastclose [drm]]
[  808.913594] xe 0000:51:00.0: [drm:drm_lastclose [drm]] driver lastclose completed
[  808.922416] [IGT] xe_waitfence: exiting, ret=0

Bommu Krishnaiah (2):
drm/xe/uapi: add exec_queue_id member to drm_xe_wait_user_fence
structure
drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

 drivers/gpu/drm/xe/xe_exec_queue_types.h |  2 +
 drivers/gpu/drm/xe/xe_execlist.c         |  7 ++
 drivers/gpu/drm/xe/xe_guc_submit.c       | 10 +++
 drivers/gpu/drm/xe/xe_wait_user_fence.c  | 81 +++++++-----------------
 include/uapi/drm/xe_drm.h                | 16 +----
 5 files changed, 44 insertions(+), 72 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 1/2 drm/xe/uapi: add exec_queue_id member to drm_xe_wait_user_fence structure

remove the num_engines/instances members from drm_xe_wait_user_fence structure and add a exec_queue_id member

Right now this is only checking if the engine list is sane and nothing else. In the end every operation with this IOCTL is a soft check. So, let's formalize that and only use this IOCTL to wait on the fence.

exec_queue_id member will help to user space to get proper error code from kernel while in exec_queue reset

---
 drivers/gpu/drm/xe/xe_wait_user_fence.c | 56 +------------------------
 include/uapi/drm/xe_drm.h               | 16 ++-----
 2 files changed, 4 insertions(+), 68 deletions(-)
drivers/gpu/drm/xe/xe_wait_user_fence.c +1 -55
index 4d5c2555ce41..447e4cfd204c 100644
@@ -58,29 +58,7 @@ static const enum xe_engine_class user_to_xe_engine_class[] = {
 	[DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
 };
 
-static int check_hw_engines(struct xe_device *xe,
-			    struct drm_xe_engine_class_instance *eci,
-			    int num_engines)
-{
-	int i;
-
-	for (i = 0; i < num_engines; ++i) {
-		enum xe_engine_class user_class =
-			user_to_xe_engine_class[eci[i].engine_class];
-
-		if (eci[i].gt_id >= xe->info.tile_count)
-			return -EINVAL;
-
-		if (!xe_gt_hw_engine(xe_device_get_gt(xe, eci[i].gt_id),
-				     user_class, eci[i].engine_instance, true))
-			return -EINVAL;
-	}
-
-	return 0;
-}
-
-#define VALID_FLAGS	(DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP | \
-			 DRM_XE_UFENCE_WAIT_FLAG_ABSTIME)
+#define VALID_FLAGS	DRM_XE_UFENCE_WAIT_FLAG_ABSTIME
 #define MAX_OP		DRM_XE_UFENCE_WAIT_OP_LTE
 
 static long to_jiffies_timeout(struct xe_device *xe,
@@ -132,12 +110,8 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 	struct xe_device *xe = to_xe_device(dev);
 	DEFINE_WAIT_FUNC(w_wait, woken_wake_function);
 	struct drm_xe_wait_user_fence *args = data;
-	struct drm_xe_engine_class_instance eci[XE_HW_ENGINE_MAX_INSTANCE];
-	struct drm_xe_engine_class_instance __user *user_eci =
-		u64_to_user_ptr(args->instances);
 	u64 addr = args->addr;
 	int err;
-	bool no_engines = args->flags & DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP;
 	long timeout;
 	ktime_t start;
 
@@ -151,41 +125,13 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 	if (XE_IOCTL_DBG(xe, args->op > MAX_OP))
 		return -EINVAL;
 
-	if (XE_IOCTL_DBG(xe, no_engines &&
-			 (args->num_engines || args->instances)))
-		return -EINVAL;
-
-	if (XE_IOCTL_DBG(xe, !no_engines && !args->num_engines))
-		return -EINVAL;
-
 	if (XE_IOCTL_DBG(xe, addr & 0x7))
 		return -EINVAL;
 
-	if (XE_IOCTL_DBG(xe, args->num_engines > XE_HW_ENGINE_MAX_INSTANCE))
-		return -EINVAL;
-
-	if (!no_engines) {
-		err = copy_from_user(eci, user_eci,
-				     sizeof(struct drm_xe_engine_class_instance) *
-			     args->num_engines);
-		if (XE_IOCTL_DBG(xe, err))
-			return -EFAULT;
-
-		if (XE_IOCTL_DBG(xe, check_hw_engines(xe, eci,
-						      args->num_engines)))
-			return -EINVAL;
-	}
-
 	timeout = to_jiffies_timeout(xe, args);
 
 	start = ktime_get();
 
-	/*
-	 * FIXME: Very simple implementation at the moment, single wait queue
-	 * for everything. Could be optimized to have a wait queue for every
-	 * hardware engine. Open coding as 'do_compare' can sleep which doesn't
-	 * work with the wait_event_* macros.
-	 */
 	add_wait_queue(&xe->ufence_wq, &w_wait);
 	for (;;) {
 		err = do_compare(addr, args->value, args->mask, args->op);
include/uapi/drm/xe_drm.h +3 -13
index eb03a49c17a1..d922641e11e8 100644
@@ -1030,8 +1030,7 @@ struct drm_xe_wait_user_fence {
 	/** @op: wait operation (type of comparison) */
 	__u16 op;
 
-#define DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP	(1 << 0)	/* e.g. Wait on VM bind */
-#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME	(1 << 1)
+#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME	(1 << 0)
 	/** @flags: wait flags */
 	__u16 flags;
 
@@ -1064,17 +1063,8 @@ struct drm_xe_wait_user_fence {
 	 */
 	__s64 timeout;
 
-	/**
-	 * @num_engines: number of engine instances to wait on, must be zero
-	 * when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
-	 */
-	__u64 num_engines;
-
-	/**
-	 * @instances: user pointer to array of drm_xe_engine_class_instance to
-	 * wait on, must be NULL when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
-	 */
-	__u64 instances;
+	/** @exec_queue_id: exec_queue_id returned from xe_exec_queue_create_ioctl */
+	__u32 exec_queue_id;
 
 	/** @reserved: Reserved */
 	__u64 reserved[2];
Signature · 1 line
2.25.1
#2 lore raw

Re: [Intel-xe] [PATCH v2 1/2] drm/xe/uapi: add exec_queue_id member to drm_xe_wait_user_fence structure

See one question below. With that addressed, patch is Reviewed-by: Oak Zeng <[email protected]>

Quoted 122 lines

-----Original Message-----
From: Intel-xe <[email protected]> On Behalf Of Bommu Krishnaiah
Sent: Tuesday, December 5, 2023 9:45 PM
To: [email protected]
Cc: Dugast, Francois <[email protected]>; Vivi, Rodrigo <[email protected]>
Subject: [Intel-xe] [PATCH v2 1/2] drm/xe/uapi: add exec_queue_id member to drm_xe_wait_user_fence structure

remove the num_engines/instances members from drm_xe_wait_user_fence structure and add a exec_queue_id member

Right now this is only checking if the engine list is sane and nothing else. In the end every operation with this IOCTL is a soft check. So, let's formalize that and only use this IOCTL to wait on the fence.

exec_queue_id member will help to user space to get proper error code from kernel while in exec_queue reset

---
 drivers/gpu/drm/xe/xe_wait_user_fence.c | 56 +------------------------
 include/uapi/drm/xe_drm.h               | 16 ++-----
 2 files changed, 4 insertions(+), 68 deletions(-)
drivers/gpu/drm/xe/xe_wait_user_fence.c +1 -23
index 4d5c2555ce41..447e4cfd204c 100644
@@ -58,29 +58,7 @@ static const enum xe_engine_class
 user_to_xe_engine_class[] = {
 	[DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
 };
 
-static int check_hw_engines(struct xe_device *xe,
-			    struct drm_xe_engine_class_instance *eci,
-			    int num_engines)
-{
-	int i;
-
-	for (i = 0; i < num_engines; ++i) {
-		enum xe_engine_class user_class =
-			user_to_xe_engine_class[eci[i].engine_class];
-
-		if (eci[i].gt_id >= xe->info.tile_count)
-			return -EINVAL;
-
-		if (!xe_gt_hw_engine(xe_device_get_gt(xe, eci[i].gt_id),
-				     user_class, eci[i].engine_instance, true))
-			return -EINVAL;
-	}
-
-	return 0;
-}
-
-#define VALID_FLAGS	(DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP | \
-			 DRM_XE_UFENCE_WAIT_FLAG_ABSTIME)
+#define VALID_FLAGS	DRM_XE_UFENCE_WAIT_FLAG_ABSTIME
 #define MAX_OP		DRM_XE_UFENCE_WAIT_OP_LTE
 

static long to_jiffies_timeout(struct xe_device *xe,

drivers/gpu/drm/xe/xe_wait_user_fence.c +0 -4
@@ -132,12 +110,8 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,
 	struct xe_device *xe = to_xe_device(dev);
 	DEFINE_WAIT_FUNC(w_wait, woken_wake_function);
 	struct drm_xe_wait_user_fence *args = data;
-	struct drm_xe_engine_class_instance
 eci[XE_HW_ENGINE_MAX_INSTANCE];
-	struct drm_xe_engine_class_instance __user *user_eci =
-		u64_to_user_ptr(args->instances);
 	u64 addr = args->addr;
 	int err;
-	bool no_engines = args->flags &
 DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP;
drivers/gpu/drm/xe/xe_wait_user_fence.c +0 -26
 	long timeout;
 	ktime_t start;
@@ -151,41 +125,13 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,
 	if (XE_IOCTL_DBG(xe, args->op > MAX_OP))
 		return -EINVAL;
 
-	if (XE_IOCTL_DBG(xe, no_engines &&
-			 (args->num_engines || args->instances)))
-		return -EINVAL;
-
-	if (XE_IOCTL_DBG(xe, !no_engines && !args->num_engines))
-		return -EINVAL;
-
 	if (XE_IOCTL_DBG(xe, addr & 0x7))
 		return -EINVAL;
 
-	if (XE_IOCTL_DBG(xe, args->num_engines >
 XE_HW_ENGINE_MAX_INSTANCE))
-		return -EINVAL;
-
-	if (!no_engines) {
-		err = copy_from_user(eci, user_eci,
-				     sizeof(struct drm_xe_engine_class_instance) *
-			     args->num_engines);
-		if (XE_IOCTL_DBG(xe, err))
-			return -EFAULT;
-
-		if (XE_IOCTL_DBG(xe, check_hw_engines(xe, eci,
-						      args->num_engines)))
-			return -EINVAL;
-	}
-
 	timeout = to_jiffies_timeout(xe, args);
 
 	start = ktime_get();
 
-	/*
-	 * FIXME: Very simple implementation at the moment, single wait queue
-	 * for everything. Could be optimized to have a wait queue for every
-	 * hardware engine. Open coding as 'do_compare' can sleep which

doesn't

drivers/gpu/drm/xe/xe_wait_user_fence.c int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data, · line 188

Quoted 2 lines
drivers/gpu/drm/xe/xe_wait_user_fence.c +0 -2
-	 * work with the wait_event_* macros.
-	 */

Above comment suggested a optimization to use per engine based wait queue. You deleted this comment, but it is still per device based wait queue.

Personally I think it is okay to use per device base wait queue. So I think it Is fine to delete this comment though...

Oak

Quoted 42 lines
drivers/gpu/drm/xe/xe_wait_user_fence.c +0 -0
 	add_wait_queue(&xe->ufence_wq, &w_wait);
 	for (;;) {
 		err = do_compare(addr, args->value, args->mask, args->op);
include/uapi/drm/xe_drm.h +0 -1
index eb03a49c17a1..d922641e11e8 100644
@@ -1030,8 +1030,7 @@ struct drm_xe_wait_user_fence {
 	/** @op: wait operation (type of comparison) */
 	__u16 op;
 
-#define DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP	(1 << 0)	/* e.g. Wait on

VM bind */

include/uapi/drm/xe_drm.h +1 -10
-#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME	(1 << 1)
+#define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME	(1 << 0)
 	/** @flags: wait flags */
 	__u16 flags;
@@ -1064,17 +1063,8 @@ struct drm_xe_wait_user_fence {
 	 */
 	__s64 timeout;
 
-	/**
-	 * @num_engines: number of engine instances to wait on, must be zero
-	 * when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
-	 */
-	__u64 num_engines;
-
-	/**
-	 * @instances: user pointer to array of drm_xe_engine_class_instance to
-	 * wait on, must be NULL when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP

set

include/uapi/drm/xe_drm.h +1 -2
-	 */
-	__u64 instances;
+	/** @exec_queue_id: exec_queue_id returned from

xe_exec_queue_create_ioctl */

include/uapi/drm/xe_drm.h +1 -1
+	__u32 exec_queue_id;
 
 	/** @reserved: Reserved */
 	__u64 reserved[2];
--

2.25.1

#3 lore raw

PATCH v2 2/2 drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

In reply to Bommu Krishnaiah

Currently xe_wait_user_fence_ioctl is not checking exec_queue state and blocking until timeout, with this patch wakeup the blocking wait if exec_queue reset happen and returning proper error code

Cc Oak Zeng <[email protected]>
Cc Kempczynski Zbigniew <[email protected]>
Cc Matthew Brost <[email protected]>
---
 drivers/gpu/drm/xe/xe_exec_queue_types.h |  2 ++
 drivers/gpu/drm/xe/xe_execlist.c         |  7 ++++++
 drivers/gpu/drm/xe/xe_guc_submit.c       | 10 +++++++++
 drivers/gpu/drm/xe/xe_wait_user_fence.c  | 27 +++++++++++++++++++-----
 4 files changed, 41 insertions(+), 5 deletions(-)
drivers/gpu/drm/xe/xe_exec_queue_types.h +2 -0
index 5ba47a5cfdbd..84ccf7242247 100644
@@ -206,6 +206,8 @@ struct xe_exec_queue_ops {
 	 * signalled when this function is called.
 	 */
 	void (*resume)(struct xe_exec_queue *q);
+	/** @reset_status: check exec queue reset status */
+	bool (*reset_status)(struct xe_exec_queue *q);
 };
 
 #endif
drivers/gpu/drm/xe/xe_execlist.c +7 -0
index e8754adfc52a..dab41e99f3d3 100644
@@ -446,6 +446,12 @@ static void execlist_exec_queue_resume(struct xe_exec_queue *q)
 	/* NIY */
 }
 
+static bool execlist_exec_queue_reset_status(struct xe_exec_queue *q)
+{
+	/* NIY */
+	return false;
+}
+
 static const struct xe_exec_queue_ops execlist_exec_queue_ops = {
 	.init = execlist_exec_queue_init,
 	.kill = execlist_exec_queue_kill,
@@ -457,6 +463,7 @@ static const struct xe_exec_queue_ops execlist_exec_queue_ops = {
 	.suspend = execlist_exec_queue_suspend,
 	.suspend_wait = execlist_exec_queue_suspend_wait,
 	.resume = execlist_exec_queue_resume,
+	.reset_status = execlist_exec_queue_reset_status,
 };
 
 int xe_execlist_init(struct xe_gt *gt)
drivers/gpu/drm/xe/xe_guc_submit.c +10 -0
index 46b132ee1d3a..a42c3574cbec 100644
@@ -854,6 +854,10 @@ static void simple_error_capture(struct xe_exec_queue *q)
 static void xe_guc_exec_queue_trigger_cleanup(struct xe_exec_queue *q)
 {
 	struct xe_guc *guc = exec_queue_to_guc(q);
+	struct xe_device *xe = guc_to_xe(guc);
+
+	/** to wakeup xe_wait_user_fence ioctl if exec queue is reset */
+	wake_up_all(&xe->ufence_wq);
 
 	if (xe_exec_queue_is_lr(q))
 		queue_work(guc_to_gt(guc)->ordered_wq, &q->guc->lr_tdr);
@@ -1394,6 +1398,11 @@ static void guc_exec_queue_resume(struct xe_exec_queue *q)
 	guc_exec_queue_add_msg(q, msg, RESUME);
 }
 
+static bool guc_exec_queue_reset_status(struct xe_exec_queue *q)
+{
+	return exec_queue_reset(q);
+}
+
 /*
  * All of these functions are an abstraction layer which other parts of XE can
  * use to trap into the GuC backend. All of these functions, aside from init,
@@ -1411,6 +1420,7 @@ static const struct xe_exec_queue_ops guc_exec_queue_ops = {
 	.suspend = guc_exec_queue_suspend,
 	.suspend_wait = guc_exec_queue_suspend_wait,
 	.resume = guc_exec_queue_resume,
+	.reset_status = guc_exec_queue_reset_status,
 };
 
 static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q)
drivers/gpu/drm/xe/xe_wait_user_fence.c +22 -5
index 447e4cfd204c..b3740c3d306c 100644
@@ -13,6 +13,7 @@
 #include "xe_device.h"
 #include "xe_gt.h"
 #include "xe_macros.h"
+#include "xe_exec_queue.h"
 
 static int do_compare(u64 addr, u64 value, u64 mask, u16 op)
 {
@@ -108,10 +109,12 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 			     struct drm_file *file)
 {
 	struct xe_device *xe = to_xe_device(dev);
+	struct xe_file *xef = to_xe_file(file);
 	DEFINE_WAIT_FUNC(w_wait, woken_wake_function);
 	struct drm_xe_wait_user_fence *args = data;
+	struct xe_exec_queue *q = NULL;
 	u64 addr = args->addr;
-	int err;
+	int err = 0;
 	long timeout;
 	ktime_t start;
 
@@ -128,6 +131,10 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 	if (XE_IOCTL_DBG(xe, addr & 0x7))
 		return -EINVAL;
 
+	q = xe_exec_queue_lookup(xef, args->exec_queue_id);
+	if (XE_IOCTL_DBG(xe, !q))
+		return -ENOENT;
+
 	timeout = to_jiffies_timeout(xe, args);
 
 	start = ktime_get();
@@ -143,6 +150,12 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 			break;
 		}
 
+		if (q->ops->reset_status(q)) {
+			drm_info(&xe->drm, "exec gueue reset detected\n");
+			err = -EIO;
+			break;
+		}
+
 		if (!timeout) {
 			err = -ETIME;
 			break;
@@ -159,9 +172,13 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev, void *data,
 	}
 
 	if (XE_IOCTL_DBG(xe, err < 0))
-		return err;
-	else if (XE_IOCTL_DBG(xe, !timeout))
-		return -ETIME;
+		goto err_exec_queue;
+	else if (XE_IOCTL_DBG(xe, !timeout)) {
+		err = -ETIME;
+		goto err_exec_queue;
+	}
 
-	return 0;
+err_exec_queue:
+	xe_exec_queue_put(q);
+	return err;
 }
Signature · 1 line
2.25.1
#4 lore raw

Re: [Intel-xe] [PATCH v2 2/2] drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

Quoted 165 lines

-----Original Message-----
From: Bommu, Krishnaiah <[email protected]>
Sent: Tuesday, December 5, 2023 9:45 PM
To: [email protected]
Cc: Bommu, Krishnaiah <[email protected]>; Zeng, Oak <[email protected]>; Kempczynski, Zbigniew <[email protected]>; Brost, Matthew <[email protected]>
Subject: [PATCH v2 2/2] drm/xe/uapi: Return correct error code for xe_wait_user_fence_ioctl

Currently xe_wait_user_fence_ioctl is not checking exec_queue state and blocking until timeout, with this patch wakeup the blocking wait if exec_queue reset happen and returning proper error code

Cc Oak Zeng <[email protected]>
Cc Kempczynski Zbigniew <[email protected]>
Cc Matthew Brost <[email protected]>
---
 drivers/gpu/drm/xe/xe_exec_queue_types.h |  2 ++
 drivers/gpu/drm/xe/xe_execlist.c         |  7 ++++++
 drivers/gpu/drm/xe/xe_guc_submit.c       | 10 +++++++++
 drivers/gpu/drm/xe/xe_wait_user_fence.c  | 27 +++++++++++++++++++-----
 4 files changed, 41 insertions(+), 5 deletions(-)
drivers/gpu/drm/xe/xe_exec_queue_types.h +2 -0
index 5ba47a5cfdbd..84ccf7242247 100644
@@ -206,6 +206,8 @@ struct xe_exec_queue_ops {
 	 * signalled when this function is called.
 	 */
 	void (*resume)(struct xe_exec_queue *q);
+	/** @reset_status: check exec queue reset status */
+	bool (*reset_status)(struct xe_exec_queue *q);
 };
 
 #endif
drivers/gpu/drm/xe/xe_execlist.c +6 -0
index e8754adfc52a..dab41e99f3d3 100644
@@ -446,6 +446,12 @@ static void execlist_exec_queue_resume(struct
 xe_exec_queue *q)
 	/* NIY */
 }
 
+static bool execlist_exec_queue_reset_status(struct xe_exec_queue *q)
+{
+	/* NIY */
+	return false;
+}
+
 static const struct xe_exec_queue_ops execlist_exec_queue_ops = {
 	.init = execlist_exec_queue_init,
drivers/gpu/drm/xe/xe_execlist.c +1 -0
 	.kill = execlist_exec_queue_kill,
@@ -457,6 +463,7 @@ static const struct xe_exec_queue_ops
 execlist_exec_queue_ops = {
 	.suspend = execlist_exec_queue_suspend,
 	.suspend_wait = execlist_exec_queue_suspend_wait,
 	.resume = execlist_exec_queue_resume,
+	.reset_status = execlist_exec_queue_reset_status,
 };
 

int xe_execlist_init(struct xe_gt *gt)

drivers/gpu/drm/xe/xe_guc_submit.c +4 -0
index 46b132ee1d3a..a42c3574cbec 100644
@@ -854,6 +854,10 @@ static void simple_error_capture(struct xe_exec_queue
 *q)
 static void xe_guc_exec_queue_trigger_cleanup(struct xe_exec_queue *q)
 {
 	struct xe_guc *guc = exec_queue_to_guc(q);
+	struct xe_device *xe = guc_to_xe(guc);
+
+	/** to wakeup xe_wait_user_fence ioctl if exec queue is reset */
+	wake_up_all(&xe->ufence_wq);
 
 	if (xe_exec_queue_is_lr(q))
drivers/gpu/drm/xe/xe_guc_submit.c +5 -0
 		queue_work(guc_to_gt(guc)->ordered_wq, &q->guc->lr_tdr);
@@ -1394,6 +1398,11 @@ static void guc_exec_queue_resume(struct
 xe_exec_queue *q)
 	guc_exec_queue_add_msg(q, msg, RESUME);
 }
 
+static bool guc_exec_queue_reset_status(struct xe_exec_queue *q)
+{
+	return exec_queue_reset(q);
+}
+
 /*
  * All of these functions are an abstraction layer which other parts of XE can

* use to trap into the GuC backend. All of these functions, aside from init,

drivers/gpu/drm/xe/xe_guc_submit.c +1 -0
@@ -1411,6 +1420,7 @@ static const struct xe_exec_queue_ops
 guc_exec_queue_ops = {
 	.suspend = guc_exec_queue_suspend,
 	.suspend_wait = guc_exec_queue_suspend_wait,
 	.resume = guc_exec_queue_resume,
+	.reset_status = guc_exec_queue_reset_status,
 };
 

static void guc_exec_queue_stop(struct xe_guc *guc, struct xe_exec_queue *q)

drivers/gpu/drm/xe/xe_wait_user_fence.c +8 -1
index 447e4cfd204c..b3740c3d306c 100644
@@ -13,6 +13,7 @@
 #include "xe_device.h"
 #include "xe_gt.h"
 #include "xe_macros.h"
+#include "xe_exec_queue.h"
 
 static int do_compare(u64 addr, u64 value, u64 mask, u16 op)
 {
@@ -108,10 +109,12 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,
 			     struct drm_file *file)
 {
 	struct xe_device *xe = to_xe_device(dev);
+	struct xe_file *xef = to_xe_file(file);
 	DEFINE_WAIT_FUNC(w_wait, woken_wake_function);
 	struct drm_xe_wait_user_fence *args = data;
+	struct xe_exec_queue *q = NULL;
 	u64 addr = args->addr;
-	int err;
+	int err = 0;
 	long timeout;
 	ktime_t start;
@@ -128,6 +131,10 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,
 	if (XE_IOCTL_DBG(xe, addr & 0x7))
 		return -EINVAL;
 
+	q = xe_exec_queue_lookup(xef, args->exec_queue_id);
+	if (XE_IOCTL_DBG(xe, !q))
+		return -ENOENT;
+
 	timeout = to_jiffies_timeout(xe, args);
 
drivers/gpu/drm/xe/xe_wait_user_fence.c +6 -0
 	start = ktime_get();
@@ -143,6 +150,12 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,
 			break;
 		}
 
+		if (q->ops->reset_status(q)) {
+			drm_info(&xe->drm, "exec gueue reset detected\n");
+			err = -EIO;
+			break;
+		}
+
 		if (!timeout) {
 			err = -ETIME;
drivers/gpu/drm/xe/xe_wait_user_fence.c +0 -0
 			break;
@@ -159,9 +172,13 @@ int xe_wait_user_fence_ioctl(struct drm_device *dev,
 void *data,

drivers/gpu/drm/xe/xe_wait_user_fence.c int xe_wait_user_fence_ioctl(struct drm_device *dev, · line 179

Quoted 10 lines
drivers/gpu/drm/xe/xe_wait_user_fence.c +4 -3
 	}
 
 	if (XE_IOCTL_DBG(xe, err < 0))
-		return err;
-	else if (XE_IOCTL_DBG(xe, !timeout))
-		return -ETIME;
+		goto err_exec_queue;
+	else if (XE_IOCTL_DBG(xe, !timeout)) {
+		err = -ETIME;
+		goto err_exec_queue;

Those goto seems not necessary.

You can just write:

if (XE_IOCTL_DBG(xe, !timeout))

	err = -ETIME;

xe_exec_queue_put(q);
return err;

Otherwise patch looks good to me. Reviewed-by: Oak Zeng <[email protected]>

Oak

Quoted 9 lines
drivers/gpu/drm/xe/xe_wait_user_fence.c +4 -1
+	}
 
-	return 0;
+err_exec_queue:
+	xe_exec_queue_put(q);
+	return err;

}
--
2.25.1