xref: /linux/include/uapi/drm/xe_drm.h (revision 8cdcef1c2f82d207aa8b2a02298fbc17191c6261)
1 /* SPDX-License-Identifier: MIT */
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #ifndef _UAPI_XE_DRM_H_
7 #define _UAPI_XE_DRM_H_
8 
9 #include "drm.h"
10 
11 #if defined(__cplusplus)
12 extern "C" {
13 #endif
14 
15 /* Please note that modifications to all structs defined here are
16  * subject to backwards-compatibility constraints.
17  */
18 
19 /**
20  * DOC: uevent generated by xe on it's pci node.
21  *
22  * DRM_XE_RESET_FAILED_UEVENT - Event is generated when attempt to reset gt
23  * fails. The value supplied with the event is always "NEEDS_RESET".
24  * Additional information supplied is tile id and gt id of the gt unit for
25  * which reset has failed.
26  */
27 #define DRM_XE_RESET_FAILED_UEVENT "DEVICE_STATUS"
28 
29 /**
30  * struct xe_user_extension - Base class for defining a chain of extensions
31  *
32  * Many interfaces need to grow over time. In most cases we can simply
33  * extend the struct and have userspace pass in more data. Another option,
34  * as demonstrated by Vulkan's approach to providing extensions for forward
35  * and backward compatibility, is to use a list of optional structs to
36  * provide those extra details.
37  *
38  * The key advantage to using an extension chain is that it allows us to
39  * redefine the interface more easily than an ever growing struct of
40  * increasing complexity, and for large parts of that interface to be
41  * entirely optional. The downside is more pointer chasing; chasing across
42  * the __user boundary with pointers encapsulated inside u64.
43  *
44  * Example chaining:
45  *
46  * .. code-block:: C
47  *
48  *	struct xe_user_extension ext3 {
49  *		.next_extension = 0, // end
50  *		.name = ...,
51  *	};
52  *	struct xe_user_extension ext2 {
53  *		.next_extension = (uintptr_t)&ext3,
54  *		.name = ...,
55  *	};
56  *	struct xe_user_extension ext1 {
57  *		.next_extension = (uintptr_t)&ext2,
58  *		.name = ...,
59  *	};
60  *
61  * Typically the struct xe_user_extension would be embedded in some uAPI
62  * struct, and in this case we would feed it the head of the chain(i.e ext1),
63  * which would then apply all of the above extensions.
64  *
65  */
66 struct xe_user_extension {
67 	/**
68 	 * @next_extension:
69 	 *
70 	 * Pointer to the next struct xe_user_extension, or zero if the end.
71 	 */
72 	__u64 next_extension;
73 
74 	/**
75 	 * @name: Name of the extension.
76 	 *
77 	 * Note that the name here is just some integer.
78 	 *
79 	 * Also note that the name space for this is not global for the whole
80 	 * driver, but rather its scope/meaning is limited to the specific piece
81 	 * of uAPI which has embedded the struct xe_user_extension.
82 	 */
83 	__u32 name;
84 
85 	/**
86 	 * @pad: MBZ
87 	 *
88 	 * All undefined bits must be zero.
89 	 */
90 	__u32 pad;
91 };
92 
93 /*
94  * xe specific ioctls.
95  *
96  * The device specific ioctl range is [DRM_COMMAND_BASE, DRM_COMMAND_END) ie
97  * [0x40, 0xa0) (a0 is excluded). The numbers below are defined as offset
98  * against DRM_COMMAND_BASE and should be between [0x0, 0x60).
99  */
100 #define DRM_XE_DEVICE_QUERY		0x00
101 #define DRM_XE_GEM_CREATE		0x01
102 #define DRM_XE_GEM_MMAP_OFFSET		0x02
103 #define DRM_XE_VM_CREATE		0x03
104 #define DRM_XE_VM_DESTROY		0x04
105 #define DRM_XE_VM_BIND			0x05
106 #define DRM_XE_EXEC			0x06
107 #define DRM_XE_EXEC_QUEUE_CREATE	0x07
108 #define DRM_XE_EXEC_QUEUE_DESTROY	0x08
109 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY	0x09
110 #define DRM_XE_EXEC_QUEUE_GET_PROPERTY	0x0a
111 #define DRM_XE_WAIT_USER_FENCE		0x0b
112 /* Must be kept compact -- no holes */
113 
114 #define DRM_IOCTL_XE_DEVICE_QUERY		DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_DEVICE_QUERY, struct drm_xe_device_query)
115 #define DRM_IOCTL_XE_GEM_CREATE			DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_GEM_CREATE, struct drm_xe_gem_create)
116 #define DRM_IOCTL_XE_GEM_MMAP_OFFSET		DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_GEM_MMAP_OFFSET, struct drm_xe_gem_mmap_offset)
117 #define DRM_IOCTL_XE_VM_CREATE			DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_VM_CREATE, struct drm_xe_vm_create)
118 #define DRM_IOCTL_XE_VM_DESTROY			DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_DESTROY, struct drm_xe_vm_destroy)
119 #define DRM_IOCTL_XE_VM_BIND			DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_BIND, struct drm_xe_vm_bind)
120 #define DRM_IOCTL_XE_EXEC			DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC, struct drm_xe_exec)
121 #define DRM_IOCTL_XE_EXEC_QUEUE_CREATE		DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_CREATE, struct drm_xe_exec_queue_create)
122 #define DRM_IOCTL_XE_EXEC_QUEUE_DESTROY		DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_DESTROY, struct drm_xe_exec_queue_destroy)
123 #define DRM_IOCTL_XE_EXEC_QUEUE_SET_PROPERTY	DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_SET_PROPERTY, struct drm_xe_exec_queue_set_property)
124 #define DRM_IOCTL_XE_EXEC_QUEUE_GET_PROPERTY	DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_GET_PROPERTY, struct drm_xe_exec_queue_get_property)
125 #define DRM_IOCTL_XE_WAIT_USER_FENCE		DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_WAIT_USER_FENCE, struct drm_xe_wait_user_fence)
126 
127 /** struct drm_xe_engine_class_instance - instance of an engine class */
128 struct drm_xe_engine_class_instance {
129 #define DRM_XE_ENGINE_CLASS_RENDER		0
130 #define DRM_XE_ENGINE_CLASS_COPY		1
131 #define DRM_XE_ENGINE_CLASS_VIDEO_DECODE	2
132 #define DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE	3
133 #define DRM_XE_ENGINE_CLASS_COMPUTE		4
134 	/*
135 	 * Kernel only classes (not actual hardware engine class). Used for
136 	 * creating ordered queues of VM bind operations.
137 	 */
138 #define DRM_XE_ENGINE_CLASS_VM_BIND_ASYNC	5
139 #define DRM_XE_ENGINE_CLASS_VM_BIND_SYNC	6
140 	__u16 engine_class;
141 
142 	__u16 engine_instance;
143 	__u16 gt_id;
144 	/** @pad: MBZ */
145 	__u16 pad;
146 };
147 
148 /**
149  * enum drm_xe_memory_class - Supported memory classes.
150  */
151 enum drm_xe_memory_class {
152 	/** @DRM_XE_MEM_REGION_CLASS_SYSMEM: Represents system memory. */
153 	DRM_XE_MEM_REGION_CLASS_SYSMEM = 0,
154 	/**
155 	 * @DRM_XE_MEM_REGION_CLASS_VRAM: On discrete platforms, this
156 	 * represents the memory that is local to the device, which we
157 	 * call VRAM. Not valid on integrated platforms.
158 	 */
159 	DRM_XE_MEM_REGION_CLASS_VRAM
160 };
161 
162 /**
163  * struct drm_xe_query_mem_region - Describes some region as known to
164  * the driver.
165  */
166 struct drm_xe_query_mem_region {
167 	/**
168 	 * @mem_class: The memory class describing this region.
169 	 *
170 	 * See enum drm_xe_memory_class for supported values.
171 	 */
172 	__u16 mem_class;
173 	/**
174 	 * @instance: The instance for this region.
175 	 *
176 	 * The @mem_class and @instance taken together will always give
177 	 * a unique pair.
178 	 */
179 	__u16 instance;
180 	/** @pad: MBZ */
181 	__u32 pad;
182 	/**
183 	 * @min_page_size: Min page-size in bytes for this region.
184 	 *
185 	 * When the kernel allocates memory for this region, the
186 	 * underlying pages will be at least @min_page_size in size.
187 	 *
188 	 * Important note: When userspace allocates a GTT address which
189 	 * can point to memory allocated from this region, it must also
190 	 * respect this minimum alignment. This is enforced by the
191 	 * kernel.
192 	 */
193 	__u32 min_page_size;
194 	/**
195 	 * @total_size: The usable size in bytes for this region.
196 	 */
197 	__u64 total_size;
198 	/**
199 	 * @used: Estimate of the memory used in bytes for this region.
200 	 *
201 	 * Requires CAP_PERFMON or CAP_SYS_ADMIN to get reliable
202 	 * accounting.  Without this the value here will always equal
203 	 * zero.
204 	 */
205 	__u64 used;
206 	/**
207 	 * @cpu_visible_size: How much of this region can be CPU
208 	 * accessed, in bytes.
209 	 *
210 	 * This will always be <= @total_size, and the remainder (if
211 	 * any) will not be CPU accessible. If the CPU accessible part
212 	 * is smaller than @total_size then this is referred to as a
213 	 * small BAR system.
214 	 *
215 	 * On systems without small BAR (full BAR), the probed_size will
216 	 * always equal the @total_size, since all of it will be CPU
217 	 * accessible.
218 	 *
219 	 * Note this is only tracked for DRM_XE_MEM_REGION_CLASS_VRAM
220 	 * regions (for other types the value here will always equal
221 	 * zero).
222 	 */
223 	__u64 cpu_visible_size;
224 	/**
225 	 * @cpu_visible_used: Estimate of CPU visible memory used, in
226 	 * bytes.
227 	 *
228 	 * Requires CAP_PERFMON or CAP_SYS_ADMIN to get reliable
229 	 * accounting. Without this the value here will always equal
230 	 * zero.  Note this is only currently tracked for
231 	 * DRM_XE_MEM_REGION_CLASS_VRAM regions (for other types the value
232 	 * here will always be zero).
233 	 */
234 	__u64 cpu_visible_used;
235 	/** @reserved: MBZ */
236 	__u64 reserved[6];
237 };
238 
239 /**
240  * struct drm_xe_query_engine_cycles - correlate CPU and GPU timestamps
241  *
242  * If a query is made with a struct drm_xe_device_query where .query is equal to
243  * DRM_XE_DEVICE_QUERY_ENGINE_CYCLES, then the reply uses struct drm_xe_query_engine_cycles
244  * in .data. struct drm_xe_query_engine_cycles is allocated by the user and
245  * .data points to this allocated structure.
246  *
247  * The query returns the engine cycles and the frequency that can
248  * be used to calculate the engine timestamp. In addition the
249  * query returns a set of cpu timestamps that indicate when the command
250  * streamer cycle count was captured.
251  */
252 struct drm_xe_query_engine_cycles {
253 	/**
254 	 * @eci: This is input by the user and is the engine for which command
255 	 * streamer cycles is queried.
256 	 */
257 	struct drm_xe_engine_class_instance eci;
258 
259 	/**
260 	 * @clockid: This is input by the user and is the reference clock id for
261 	 * CPU timestamp. For definition, see clock_gettime(2) and
262 	 * perf_event_open(2). Supported clock ids are CLOCK_MONOTONIC,
263 	 * CLOCK_MONOTONIC_RAW, CLOCK_REALTIME, CLOCK_BOOTTIME, CLOCK_TAI.
264 	 */
265 	__s32 clockid;
266 
267 	/** @width: Width of the engine cycle counter in bits. */
268 	__u32 width;
269 
270 	/**
271 	 * @engine_cycles: Engine cycles as read from its register
272 	 * at 0x358 offset.
273 	 */
274 	__u64 engine_cycles;
275 
276 	/** @engine_frequency: Frequency of the engine cycles in Hz. */
277 	__u64 engine_frequency;
278 
279 	/**
280 	 * @cpu_timestamp: CPU timestamp in ns. The timestamp is captured before
281 	 * reading the engine_cycles register using the reference clockid set by the
282 	 * user.
283 	 */
284 	__u64 cpu_timestamp;
285 
286 	/**
287 	 * @cpu_delta: Time delta in ns captured around reading the lower dword
288 	 * of the engine_cycles register.
289 	 */
290 	__u64 cpu_delta;
291 };
292 
293 /**
294  * struct drm_xe_query_mem_regions - describe memory regions
295  *
296  * If a query is made with a struct drm_xe_device_query where .query
297  * is equal to DRM_XE_DEVICE_QUERY_MEM_REGIONS, then the reply uses
298  * struct drm_xe_query_mem_regions in .data.
299  */
300 struct drm_xe_query_mem_regions {
301 	/** @num_regions: number of memory regions returned in @regions */
302 	__u32 num_regions;
303 	/** @pad: MBZ */
304 	__u32 pad;
305 	/** @regions: The returned regions for this device */
306 	struct drm_xe_query_mem_region regions[];
307 };
308 
309 /**
310  * struct drm_xe_query_config - describe the device configuration
311  *
312  * If a query is made with a struct drm_xe_device_query where .query
313  * is equal to DRM_XE_DEVICE_QUERY_CONFIG, then the reply uses
314  * struct drm_xe_query_config in .data.
315  *
316  */
317 struct drm_xe_query_config {
318 	/** @num_params: number of parameters returned in info */
319 	__u32 num_params;
320 
321 	/** @pad: MBZ */
322 	__u32 pad;
323 
324 #define DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID	0
325 #define DRM_XE_QUERY_CONFIG_FLAGS			1
326 	#define DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM	(1 << 0)
327 #define DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT		2
328 #define DRM_XE_QUERY_CONFIG_VA_BITS			3
329 #define DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY	4
330 	/** @info: array of elements containing the config info */
331 	__u64 info[];
332 };
333 
334 /**
335  * struct drm_xe_query_gt - describe an individual GT.
336  *
337  * To be used with drm_xe_query_gt_list, which will return a list with all the
338  * existing GT individual descriptions.
339  * Graphics Technology (GT) is a subset of a GPU/tile that is responsible for
340  * implementing graphics and/or media operations.
341  */
342 struct drm_xe_query_gt {
343 #define DRM_XE_QUERY_GT_TYPE_MAIN		0
344 #define DRM_XE_QUERY_GT_TYPE_MEDIA		1
345 	/** @type: GT type: Main or Media */
346 	__u16 type;
347 	/** @gt_id: Unique ID of this GT within the PCI Device */
348 	__u16 gt_id;
349 	/** @clock_freq: A clock frequency for timestamp */
350 	__u32 clock_freq;
351 	/**
352 	 * @near_mem_regions: Bit mask of instances from
353 	 * drm_xe_query_mem_regions that are nearest to the current engines
354 	 * of this GT.
355 	 */
356 	__u64 near_mem_regions;
357 	/**
358 	 * @far_mem_regions: Bit mask of instances from
359 	 * drm_xe_query_mem_regions that are far from the engines of this GT.
360 	 * In general, they have extra indirections when compared to the
361 	 * @near_mem_regions. For a discrete device this could mean system
362 	 * memory and memory living in a different tile.
363 	 */
364 	__u64 far_mem_regions;
365 	/** @reserved: Reserved */
366 	__u64 reserved[8];
367 };
368 
369 /**
370  * struct drm_xe_query_gt_list - A list with GT description items.
371  *
372  * If a query is made with a struct drm_xe_device_query where .query
373  * is equal to DRM_XE_DEVICE_QUERY_GT_LIST, then the reply uses struct
374  * drm_xe_query_gt_list in .data.
375  */
376 struct drm_xe_query_gt_list {
377 	/** @num_gt: number of GT items returned in gt_list */
378 	__u32 num_gt;
379 	/** @pad: MBZ */
380 	__u32 pad;
381 	/** @gt_list: The GT list returned for this device */
382 	struct drm_xe_query_gt gt_list[];
383 };
384 
385 /**
386  * struct drm_xe_query_topology_mask - describe the topology mask of a GT
387  *
388  * This is the hardware topology which reflects the internal physical
389  * structure of the GPU.
390  *
391  * If a query is made with a struct drm_xe_device_query where .query
392  * is equal to DRM_XE_DEVICE_QUERY_GT_TOPOLOGY, then the reply uses
393  * struct drm_xe_query_topology_mask in .data.
394  */
395 struct drm_xe_query_topology_mask {
396 	/** @gt_id: GT ID the mask is associated with */
397 	__u16 gt_id;
398 
399 	/*
400 	 * To query the mask of Dual Sub Slices (DSS) available for geometry
401 	 * operations. For example a query response containing the following
402 	 * in mask:
403 	 *   DSS_GEOMETRY    ff ff ff ff 00 00 00 00
404 	 * means 32 DSS are available for geometry.
405 	 */
406 #define DRM_XE_TOPO_DSS_GEOMETRY	(1 << 0)
407 	/*
408 	 * To query the mask of Dual Sub Slices (DSS) available for compute
409 	 * operations. For example a query response containing the following
410 	 * in mask:
411 	 *   DSS_COMPUTE    ff ff ff ff 00 00 00 00
412 	 * means 32 DSS are available for compute.
413 	 */
414 #define DRM_XE_TOPO_DSS_COMPUTE		(1 << 1)
415 	/*
416 	 * To query the mask of Execution Units (EU) available per Dual Sub
417 	 * Slices (DSS). For example a query response containing the following
418 	 * in mask:
419 	 *   EU_PER_DSS    ff ff 00 00 00 00 00 00
420 	 * means each DSS has 16 EU.
421 	 */
422 #define DRM_XE_TOPO_EU_PER_DSS		(1 << 2)
423 	/** @type: type of mask */
424 	__u16 type;
425 
426 	/** @num_bytes: number of bytes in requested mask */
427 	__u32 num_bytes;
428 
429 	/** @mask: little-endian mask of @num_bytes */
430 	__u8 mask[];
431 };
432 
433 /**
434  * struct drm_xe_device_query - main structure to query device information
435  *
436  * If size is set to 0, the driver fills it with the required size for the
437  * requested type of data to query. If size is equal to the required size,
438  * the queried information is copied into data.
439  *
440  * For example the following code snippet allows retrieving and printing
441  * information about the device engines with DRM_XE_DEVICE_QUERY_ENGINES:
442  *
443  * .. code-block:: C
444  *
445  *	struct drm_xe_engine_class_instance *hwe;
446  *	struct drm_xe_device_query query = {
447  *		.extensions = 0,
448  *		.query = DRM_XE_DEVICE_QUERY_ENGINES,
449  *		.size = 0,
450  *		.data = 0,
451  *	};
452  *	ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query);
453  *	hwe = malloc(query.size);
454  *	query.data = (uintptr_t)hwe;
455  *	ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query);
456  *	int num_engines = query.size / sizeof(*hwe);
457  *	for (int i = 0; i < num_engines; i++) {
458  *		printf("Engine %d: %s\n", i,
459  *			hwe[i].engine_class == DRM_XE_ENGINE_CLASS_RENDER ? "RENDER":
460  *			hwe[i].engine_class == DRM_XE_ENGINE_CLASS_COPY ? "COPY":
461  *			hwe[i].engine_class == DRM_XE_ENGINE_CLASS_VIDEO_DECODE ? "VIDEO_DECODE":
462  *			hwe[i].engine_class == DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE ? "VIDEO_ENHANCE":
463  *			hwe[i].engine_class == DRM_XE_ENGINE_CLASS_COMPUTE ? "COMPUTE":
464  *			"UNKNOWN");
465  *	}
466  *	free(hwe);
467  */
468 struct drm_xe_device_query {
469 	/** @extensions: Pointer to the first extension struct, if any */
470 	__u64 extensions;
471 
472 #define DRM_XE_DEVICE_QUERY_ENGINES		0
473 #define DRM_XE_DEVICE_QUERY_MEM_REGIONS		1
474 #define DRM_XE_DEVICE_QUERY_CONFIG		2
475 #define DRM_XE_DEVICE_QUERY_GT_LIST		3
476 #define DRM_XE_DEVICE_QUERY_HWCONFIG		4
477 #define DRM_XE_DEVICE_QUERY_GT_TOPOLOGY		5
478 #define DRM_XE_DEVICE_QUERY_ENGINE_CYCLES	6
479 	/** @query: The type of data to query */
480 	__u32 query;
481 
482 	/** @size: Size of the queried data */
483 	__u32 size;
484 
485 	/** @data: Queried data is placed here */
486 	__u64 data;
487 
488 	/** @reserved: Reserved */
489 	__u64 reserved[2];
490 };
491 
492 struct drm_xe_gem_create {
493 	/** @extensions: Pointer to the first extension struct, if any */
494 	__u64 extensions;
495 
496 	/**
497 	 * @size: Requested size for the object
498 	 *
499 	 * The (page-aligned) allocated size for the object will be returned.
500 	 */
501 	__u64 size;
502 
503 #define DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING		(0x1 << 24)
504 #define DRM_XE_GEM_CREATE_FLAG_SCANOUT			(0x1 << 25)
505 /*
506  * When using VRAM as a possible placement, ensure that the corresponding VRAM
507  * allocation will always use the CPU accessible part of VRAM. This is important
508  * for small-bar systems (on full-bar systems this gets turned into a noop).
509  *
510  * Note: System memory can be used as an extra placement if the kernel should
511  * spill the allocation to system memory, if space can't be made available in
512  * the CPU accessible part of VRAM (giving the same behaviour as the i915
513  * interface, see I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS).
514  *
515  * Note: For clear-color CCS surfaces the kernel needs to read the clear-color
516  * value stored in the buffer, and on discrete platforms we need to use VRAM for
517  * display surfaces, therefore the kernel requires setting this flag for such
518  * objects, otherwise an error is thrown on small-bar systems.
519  */
520 #define DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM	(0x1 << 26)
521 	/**
522 	 * @flags: Flags, currently a mask of memory instances of where BO can
523 	 * be placed
524 	 */
525 	__u32 flags;
526 
527 	/**
528 	 * @vm_id: Attached VM, if any
529 	 *
530 	 * If a VM is specified, this BO must:
531 	 *
532 	 *  1. Only ever be bound to that VM.
533 	 *  2. Cannot be exported as a PRIME fd.
534 	 */
535 	__u32 vm_id;
536 
537 	/**
538 	 * @handle: Returned handle for the object.
539 	 *
540 	 * Object handles are nonzero.
541 	 */
542 	__u32 handle;
543 
544 	/**
545 	 * @cpu_caching: The CPU caching mode to select for this object. If
546 	 * mmaping the object the mode selected here will also be used.
547 	 *
548 	 * Supported values:
549 	 *
550 	 * DRM_XE_GEM_CPU_CACHING_WB: Allocate the pages with write-back
551 	 * caching.  On iGPU this can't be used for scanout surfaces. Currently
552 	 * not allowed for objects placed in VRAM.
553 	 *
554 	 * DRM_XE_GEM_CPU_CACHING_WC: Allocate the pages as write-combined. This
555 	 * is uncached. Scanout surfaces should likely use this. All objects
556 	 * that can be placed in VRAM must use this.
557 	 */
558 #define DRM_XE_GEM_CPU_CACHING_WB                      1
559 #define DRM_XE_GEM_CPU_CACHING_WC                      2
560 	__u16 cpu_caching;
561 	/** @pad: MBZ */
562 	__u16 pad;
563 
564 	/** @reserved: Reserved */
565 	__u64 reserved[2];
566 };
567 
568 struct drm_xe_gem_mmap_offset {
569 	/** @extensions: Pointer to the first extension struct, if any */
570 	__u64 extensions;
571 
572 	/** @handle: Handle for the object being mapped. */
573 	__u32 handle;
574 
575 	/** @flags: Must be zero */
576 	__u32 flags;
577 
578 	/** @offset: The fake offset to use for subsequent mmap call */
579 	__u64 offset;
580 
581 	/** @reserved: Reserved */
582 	__u64 reserved[2];
583 };
584 
585 /** struct drm_xe_ext_set_property - XE set property extension */
586 struct drm_xe_ext_set_property {
587 	/** @base: base user extension */
588 	struct xe_user_extension base;
589 
590 	/** @property: property to set */
591 	__u32 property;
592 
593 	/** @pad: MBZ */
594 	__u32 pad;
595 
596 	/** @value: property value */
597 	__u64 value;
598 
599 	/** @reserved: Reserved */
600 	__u64 reserved[2];
601 };
602 
603 struct drm_xe_vm_create {
604 #define DRM_XE_VM_EXTENSION_SET_PROPERTY	0
605 	/** @extensions: Pointer to the first extension struct, if any */
606 	__u64 extensions;
607 
608 #define DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE	(1 << 0)
609 #define DRM_XE_VM_CREATE_FLAG_COMPUTE_MODE	(1 << 1)
610 #define DRM_XE_VM_CREATE_FLAG_ASYNC_DEFAULT	(1 << 2)
611 #define DRM_XE_VM_CREATE_FLAG_FAULT_MODE	(1 << 3)
612 	/** @flags: Flags */
613 	__u32 flags;
614 
615 	/** @vm_id: Returned VM ID */
616 	__u32 vm_id;
617 
618 	/** @reserved: Reserved */
619 	__u64 reserved[2];
620 };
621 
622 struct drm_xe_vm_destroy {
623 	/** @vm_id: VM ID */
624 	__u32 vm_id;
625 
626 	/** @pad: MBZ */
627 	__u32 pad;
628 
629 	/** @reserved: Reserved */
630 	__u64 reserved[2];
631 };
632 
633 struct drm_xe_vm_bind_op {
634 	/**
635 	 * @obj: GEM object to operate on, MBZ for MAP_USERPTR, MBZ for UNMAP
636 	 */
637 	__u32 obj;
638 
639 	/**
640 	 * @pat_index: The platform defined @pat_index to use for this mapping.
641 	 * The index basically maps to some predefined memory attributes,
642 	 * including things like caching, coherency, compression etc.  The exact
643 	 * meaning of the pat_index is platform specific and defined in the
644 	 * Bspec and PRMs.  When the KMD sets up the binding the index here is
645 	 * encoded into the ppGTT PTE.
646 	 *
647 	 * For coherency the @pat_index needs to be at least 1way coherent when
648 	 * drm_xe_gem_create.cpu_caching is DRM_XE_GEM_CPU_CACHING_WB. The KMD
649 	 * will extract the coherency mode from the @pat_index and reject if
650 	 * there is a mismatch (see note below for pre-MTL platforms).
651 	 *
652 	 * Note: On pre-MTL platforms there is only a caching mode and no
653 	 * explicit coherency mode, but on such hardware there is always a
654 	 * shared-LLC (or is dgpu) so all GT memory accesses are coherent with
655 	 * CPU caches even with the caching mode set as uncached.  It's only the
656 	 * display engine that is incoherent (on dgpu it must be in VRAM which
657 	 * is always mapped as WC on the CPU). However to keep the uapi somewhat
658 	 * consistent with newer platforms the KMD groups the different cache
659 	 * levels into the following coherency buckets on all pre-MTL platforms:
660 	 *
661 	 *	ppGTT UC -> COH_NONE
662 	 *	ppGTT WC -> COH_NONE
663 	 *	ppGTT WT -> COH_NONE
664 	 *	ppGTT WB -> COH_AT_LEAST_1WAY
665 	 *
666 	 * In practice UC/WC/WT should only ever used for scanout surfaces on
667 	 * such platforms (or perhaps in general for dma-buf if shared with
668 	 * another device) since it is only the display engine that is actually
669 	 * incoherent.  Everything else should typically use WB given that we
670 	 * have a shared-LLC.  On MTL+ this completely changes and the HW
671 	 * defines the coherency mode as part of the @pat_index, where
672 	 * incoherent GT access is possible.
673 	 *
674 	 * Note: For userptr and externally imported dma-buf the kernel expects
675 	 * either 1WAY or 2WAY for the @pat_index.
676 	 *
677 	 * For DRM_XE_VM_BIND_FLAG_NULL bindings there are no KMD restrictions
678 	 * on the @pat_index. For such mappings there is no actual memory being
679 	 * mapped (the address in the PTE is invalid), so the various PAT memory
680 	 * attributes likely do not apply.  Simply leaving as zero is one
681 	 * option (still a valid pat_index).
682 	 */
683 	__u16 pat_index;
684 
685 	/** @pad: MBZ */
686 	__u16 pad;
687 
688 	union {
689 		/**
690 		 * @obj_offset: Offset into the object, MBZ for CLEAR_RANGE,
691 		 * ignored for unbind
692 		 */
693 		__u64 obj_offset;
694 
695 		/** @userptr: user pointer to bind on */
696 		__u64 userptr;
697 	};
698 
699 	/**
700 	 * @range: Number of bytes from the object to bind to addr, MBZ for UNMAP_ALL
701 	 */
702 	__u64 range;
703 
704 	/** @addr: Address to operate on, MBZ for UNMAP_ALL */
705 	__u64 addr;
706 
707 	/**
708 	 * @tile_mask: Mask for which tiles to create binds for, 0 == All tiles,
709 	 * only applies to creating new VMAs
710 	 */
711 	__u64 tile_mask;
712 
713 #define DRM_XE_VM_BIND_OP_MAP		0x0
714 #define DRM_XE_VM_BIND_OP_UNMAP		0x1
715 #define DRM_XE_VM_BIND_OP_MAP_USERPTR	0x2
716 #define DRM_XE_VM_BIND_OP_UNMAP_ALL	0x3
717 #define DRM_XE_VM_BIND_OP_PREFETCH	0x4
718 	/** @op: Bind operation to perform */
719 	__u32 op;
720 
721 #define DRM_XE_VM_BIND_FLAG_READONLY	(1 << 0)
722 #define DRM_XE_VM_BIND_FLAG_ASYNC	(1 << 1)
723 	/*
724 	 * Valid on a faulting VM only, do the MAP operation immediately rather
725 	 * than deferring the MAP to the page fault handler.
726 	 */
727 #define DRM_XE_VM_BIND_FLAG_IMMEDIATE	(1 << 2)
728 	/*
729 	 * When the NULL flag is set, the page tables are setup with a special
730 	 * bit which indicates writes are dropped and all reads return zero.  In
731 	 * the future, the NULL flags will only be valid for DRM_XE_VM_BIND_OP_MAP
732 	 * operations, the BO handle MBZ, and the BO offset MBZ. This flag is
733 	 * intended to implement VK sparse bindings.
734 	 */
735 #define DRM_XE_VM_BIND_FLAG_NULL	(1 << 3)
736 	/** @flags: Bind flags */
737 	__u32 flags;
738 
739 	/**
740 	 * @prefetch_mem_region_instance: Memory region to prefetch VMA to.
741 	 * It is a region instance, not a mask.
742 	 * To be used only with %DRM_XE_VM_BIND_OP_PREFETCH operation.
743 	 */
744 	__u32 prefetch_mem_region_instance;
745 
746 	/** @reserved: Reserved */
747 	__u64 reserved[2];
748 };
749 
750 struct drm_xe_vm_bind {
751 	/** @extensions: Pointer to the first extension struct, if any */
752 	__u64 extensions;
753 
754 	/** @vm_id: The ID of the VM to bind to */
755 	__u32 vm_id;
756 
757 	/**
758 	 * @exec_queue_id: exec_queue_id, must be of class DRM_XE_ENGINE_CLASS_VM_BIND
759 	 * and exec queue must have same vm_id. If zero, the default VM bind engine
760 	 * is used.
761 	 */
762 	__u32 exec_queue_id;
763 
764 	/** @num_binds: number of binds in this IOCTL */
765 	__u32 num_binds;
766 
767 	/** @pad: MBZ */
768 	__u32 pad;
769 
770 	union {
771 		/** @bind: used if num_binds == 1 */
772 		struct drm_xe_vm_bind_op bind;
773 
774 		/**
775 		 * @vector_of_binds: userptr to array of struct
776 		 * drm_xe_vm_bind_op if num_binds > 1
777 		 */
778 		__u64 vector_of_binds;
779 	};
780 
781 	/** @num_syncs: amount of syncs to wait on */
782 	__u32 num_syncs;
783 
784 	/** @pad2: MBZ */
785 	__u32 pad2;
786 
787 	/** @syncs: pointer to struct drm_xe_sync array */
788 	__u64 syncs;
789 
790 	/** @reserved: Reserved */
791 	__u64 reserved[2];
792 };
793 
794 /* For use with DRM_XE_EXEC_QUEUE_SET_PROPERTY_ACC_GRANULARITY */
795 
796 /* Monitor 128KB contiguous region with 4K sub-granularity */
797 #define DRM_XE_ACC_GRANULARITY_128K 0
798 
799 /* Monitor 2MB contiguous region with 64KB sub-granularity */
800 #define DRM_XE_ACC_GRANULARITY_2M 1
801 
802 /* Monitor 16MB contiguous region with 512KB sub-granularity */
803 #define DRM_XE_ACC_GRANULARITY_16M 2
804 
805 /* Monitor 64MB contiguous region with 2M sub-granularity */
806 #define DRM_XE_ACC_GRANULARITY_64M 3
807 
808 /**
809  * struct drm_xe_exec_queue_set_property - exec queue set property
810  *
811  * Same namespace for extensions as drm_xe_exec_queue_create
812  */
813 struct drm_xe_exec_queue_set_property {
814 	/** @extensions: Pointer to the first extension struct, if any */
815 	__u64 extensions;
816 
817 	/** @exec_queue_id: Exec queue ID */
818 	__u32 exec_queue_id;
819 
820 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY			0
821 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE		1
822 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_PREEMPTION_TIMEOUT	2
823 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_PERSISTENCE		3
824 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_JOB_TIMEOUT		4
825 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_ACC_TRIGGER		5
826 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_ACC_NOTIFY		6
827 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_ACC_GRANULARITY		7
828 	/** @property: property to set */
829 	__u32 property;
830 
831 	/** @value: property value */
832 	__u64 value;
833 
834 	/** @reserved: Reserved */
835 	__u64 reserved[2];
836 };
837 
838 struct drm_xe_exec_queue_create {
839 #define DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY               0
840 	/** @extensions: Pointer to the first extension struct, if any */
841 	__u64 extensions;
842 
843 	/** @width: submission width (number BB per exec) for this exec queue */
844 	__u16 width;
845 
846 	/** @num_placements: number of valid placements for this exec queue */
847 	__u16 num_placements;
848 
849 	/** @vm_id: VM to use for this exec queue */
850 	__u32 vm_id;
851 
852 	/** @flags: MBZ */
853 	__u32 flags;
854 
855 	/** @exec_queue_id: Returned exec queue ID */
856 	__u32 exec_queue_id;
857 
858 	/**
859 	 * @instances: user pointer to a 2-d array of struct
860 	 * drm_xe_engine_class_instance
861 	 *
862 	 * length = width (i) * num_placements (j)
863 	 * index = j + i * width
864 	 */
865 	__u64 instances;
866 
867 	/** @reserved: Reserved */
868 	__u64 reserved[2];
869 };
870 
871 struct drm_xe_exec_queue_get_property {
872 	/** @extensions: Pointer to the first extension struct, if any */
873 	__u64 extensions;
874 
875 	/** @exec_queue_id: Exec queue ID */
876 	__u32 exec_queue_id;
877 
878 #define DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN	0
879 	/** @property: property to get */
880 	__u32 property;
881 
882 	/** @value: property value */
883 	__u64 value;
884 
885 	/** @reserved: Reserved */
886 	__u64 reserved[2];
887 };
888 
889 struct drm_xe_exec_queue_destroy {
890 	/** @exec_queue_id: Exec queue ID */
891 	__u32 exec_queue_id;
892 
893 	/** @pad: MBZ */
894 	__u32 pad;
895 
896 	/** @reserved: Reserved */
897 	__u64 reserved[2];
898 };
899 
900 struct drm_xe_sync {
901 	/** @extensions: Pointer to the first extension struct, if any */
902 	__u64 extensions;
903 
904 #define DRM_XE_SYNC_FLAG_SYNCOBJ		0x0
905 #define DRM_XE_SYNC_FLAG_TIMELINE_SYNCOBJ	0x1
906 #define DRM_XE_SYNC_FLAG_DMA_BUF		0x2
907 #define DRM_XE_SYNC_FLAG_USER_FENCE		0x3
908 #define DRM_XE_SYNC_FLAG_SIGNAL		0x10
909 	__u32 flags;
910 
911 	/** @pad: MBZ */
912 	__u32 pad;
913 
914 	union {
915 		__u32 handle;
916 
917 		/**
918 		 * @addr: Address of user fence. When sync passed in via exec
919 		 * IOCTL this a GPU address in the VM. When sync passed in via
920 		 * VM bind IOCTL this is a user pointer. In either case, it is
921 		 * the users responsibility that this address is present and
922 		 * mapped when the user fence is signalled. Must be qword
923 		 * aligned.
924 		 */
925 		__u64 addr;
926 	};
927 
928 	__u64 timeline_value;
929 
930 	/** @reserved: Reserved */
931 	__u64 reserved[2];
932 };
933 
934 struct drm_xe_exec {
935 	/** @extensions: Pointer to the first extension struct, if any */
936 	__u64 extensions;
937 
938 	/** @exec_queue_id: Exec queue ID for the batch buffer */
939 	__u32 exec_queue_id;
940 
941 	/** @num_syncs: Amount of struct drm_xe_sync in array. */
942 	__u32 num_syncs;
943 
944 	/** @syncs: Pointer to struct drm_xe_sync array. */
945 	__u64 syncs;
946 
947 	/**
948 	 * @address: address of batch buffer if num_batch_buffer == 1 or an
949 	 * array of batch buffer addresses
950 	 */
951 	__u64 address;
952 
953 	/**
954 	 * @num_batch_buffer: number of batch buffer in this exec, must match
955 	 * the width of the engine
956 	 */
957 	__u16 num_batch_buffer;
958 
959 	/** @pad: MBZ */
960 	__u16 pad[3];
961 
962 	/** @reserved: Reserved */
963 	__u64 reserved[2];
964 };
965 
966 /**
967  * struct drm_xe_wait_user_fence - wait user fence
968  *
969  * Wait on user fence, XE will wake-up on every HW engine interrupt in the
970  * instances list and check if user fence is complete::
971  *
972  *	(*addr & MASK) OP (VALUE & MASK)
973  *
974  * Returns to user on user fence completion or timeout.
975  */
976 struct drm_xe_wait_user_fence {
977 	/** @extensions: Pointer to the first extension struct, if any */
978 	__u64 extensions;
979 
980 	/**
981 	 * @addr: user pointer address to wait on, must qword aligned
982 	 */
983 	__u64 addr;
984 
985 #define DRM_XE_UFENCE_WAIT_OP_EQ	0x0
986 #define DRM_XE_UFENCE_WAIT_OP_NEQ	0x1
987 #define DRM_XE_UFENCE_WAIT_OP_GT	0x2
988 #define DRM_XE_UFENCE_WAIT_OP_GTE	0x3
989 #define DRM_XE_UFENCE_WAIT_OP_LT	0x4
990 #define DRM_XE_UFENCE_WAIT_OP_LTE	0x5
991 	/** @op: wait operation (type of comparison) */
992 	__u16 op;
993 
994 #define DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP	(1 << 0)	/* e.g. Wait on VM bind */
995 #define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME	(1 << 1)
996 	/** @flags: wait flags */
997 	__u16 flags;
998 
999 	/** @pad: MBZ */
1000 	__u32 pad;
1001 
1002 	/** @value: compare value */
1003 	__u64 value;
1004 
1005 #define DRM_XE_UFENCE_WAIT_MASK_U8	0xffu
1006 #define DRM_XE_UFENCE_WAIT_MASK_U16	0xffffu
1007 #define DRM_XE_UFENCE_WAIT_MASK_U32	0xffffffffu
1008 #define DRM_XE_UFENCE_WAIT_MASK_U64	0xffffffffffffffffu
1009 	/** @mask: comparison mask */
1010 	__u64 mask;
1011 
1012 	/**
1013 	 * @timeout: how long to wait before bailing, value in nanoseconds.
1014 	 * Without DRM_XE_UFENCE_WAIT_FLAG_ABSTIME flag set (relative timeout)
1015 	 * it contains timeout expressed in nanoseconds to wait (fence will
1016 	 * expire at now() + timeout).
1017 	 * When DRM_XE_UFENCE_WAIT_FLAG_ABSTIME flat is set (absolute timeout) wait
1018 	 * will end at timeout (uses system MONOTONIC_CLOCK).
1019 	 * Passing negative timeout leads to neverending wait.
1020 	 *
1021 	 * On relative timeout this value is updated with timeout left
1022 	 * (for restarting the call in case of signal delivery).
1023 	 * On absolute timeout this value stays intact (restarted call still
1024 	 * expire at the same point of time).
1025 	 */
1026 	__s64 timeout;
1027 
1028 	/**
1029 	 * @num_engines: number of engine instances to wait on, must be zero
1030 	 * when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
1031 	 */
1032 	__u64 num_engines;
1033 
1034 	/**
1035 	 * @instances: user pointer to array of drm_xe_engine_class_instance to
1036 	 * wait on, must be NULL when DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP set
1037 	 */
1038 	__u64 instances;
1039 
1040 	/** @reserved: Reserved */
1041 	__u64 reserved[2];
1042 };
1043 
1044 /**
1045  * DOC: XE PMU event config IDs
1046  *
1047  * Check 'man perf_event_open' to use the ID's DRM_XE_PMU_XXXX listed in xe_drm.h
1048  * in 'struct perf_event_attr' as part of perf_event_open syscall to read a
1049  * particular event.
1050  *
1051  * For example to open the DRMXE_PMU_RENDER_GROUP_BUSY(0):
1052  *
1053  * .. code-block:: C
1054  *
1055  *	struct perf_event_attr attr;
1056  *	long long count;
1057  *	int cpu = 0;
1058  *	int fd;
1059  *
1060  *	memset(&attr, 0, sizeof(struct perf_event_attr));
1061  *	attr.type = type; // eg: /sys/bus/event_source/devices/xe_0000_56_00.0/type
1062  *	attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED;
1063  *	attr.use_clockid = 1;
1064  *	attr.clockid = CLOCK_MONOTONIC;
1065  *	attr.config = DRM_XE_PMU_RENDER_GROUP_BUSY(0);
1066  *
1067  *	fd = syscall(__NR_perf_event_open, &attr, -1, cpu, -1, 0);
1068  */
1069 
1070 /*
1071  * Top bits of every counter are GT id.
1072  */
1073 #define __DRM_XE_PMU_GT_SHIFT (56)
1074 
1075 #define ___DRM_XE_PMU_OTHER(gt, x) \
1076 	(((__u64)(x)) | ((__u64)(gt) << __DRM_XE_PMU_GT_SHIFT))
1077 
1078 #define DRM_XE_PMU_RENDER_GROUP_BUSY(gt)	___DRM_XE_PMU_OTHER(gt, 0)
1079 #define DRM_XE_PMU_COPY_GROUP_BUSY(gt)		___DRM_XE_PMU_OTHER(gt, 1)
1080 #define DRM_XE_PMU_MEDIA_GROUP_BUSY(gt)		___DRM_XE_PMU_OTHER(gt, 2)
1081 #define DRM_XE_PMU_ANY_ENGINE_GROUP_BUSY(gt)	___DRM_XE_PMU_OTHER(gt, 3)
1082 
1083 #if defined(__cplusplus)
1084 }
1085 #endif
1086 
1087 #endif /* _UAPI_XE_DRM_H_ */
1088