Debian Codegraph @

media_softlet/linux/common/os/xe/include/xe_drm.h

raw
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 /*
16 * Please note that modifications to all structs defined here are
17 * subject to backwards-compatibility constraints.
18 * Sections in this file are organized as follows:
19 * 1. IOCTL definition
20 * 2. Extension definition and helper structs
21 * 3. IOCTL's Query structs in the order of the Query's entries.
22 * 4. The rest of IOCTL structs in the order of IOCTL declaration.
23 */
24
25 /**
26 * DOC: Xe Device Block Diagram
27 *
28 * The diagram below represents a high-level simplification of a discrete
29 * GPU supported by the Xe driver. It shows some device components which
30 * are necessary to understand this API, as well as how their relations
31 * to each other. This diagram does not represent real hardware::
32 *
33 * ┌──────────────────────────────────────────────────────────────────┐
34 * │ ┌──────────────────────────────────────────────────┐ ┌─────────┐ │
35 * │ │ ┌───────────────────────┐ ┌─────┐ │ │ ┌─────┐ │ │
36 * │ │ │ VRAM0 ├───┤ ... │ │ │ │VRAM1│ │ │
37 * │ │ └───────────┬───────────┘ └─GT1─┘ │ │ └──┬──┘ │ │
38 * │ │ ┌──────────────────┴───────────────────────────┐ │ │ ┌──┴──┐ │ │
39 * │ │ │ ┌─────────────────────┐ ┌─────────────────┐ │ │ │ │ │ │ │
40 * │ │ │ │ ┌──┐ ┌──┐ ┌──┐ ┌──┐ │ │ ┌─────┐ ┌─────┐ │ │ │ │ │ │ │ │
41 * │ │ │ │ │EU│ │EU│ │EU│ │EU│ │ │ │RCS0 │ │BCS0 │ │ │ │ │ │ │ │ │
42 * │ │ │ │ └──┘ └──┘ └──┘ └──┘ │ │ └─────┘ └─────┘ │ │ │ │ │ │ │ │
43 * │ │ │ │ ┌──┐ ┌──┐ ┌──┐ ┌──┐ │ │ ┌─────┐ ┌─────┐ │ │ │ │ │ │ │ │
44 * │ │ │ │ │EU│ │EU│ │EU│ │EU│ │ │ │VCS0 │ │VCS1 │ │ │ │ │ │ │ │ │
45 * │ │ │ │ └──┘ └──┘ └──┘ └──┘ │ │ └─────┘ └─────┘ │ │ │ │ │ │ │ │
46 * │ │ │ │ ┌──┐ ┌──┐ ┌──┐ ┌──┐ │ │ ┌─────┐ ┌─────┐ │ │ │ │ │ │ │ │
47 * │ │ │ │ │EU│ │EU│ │EU│ │EU│ │ │ │VECS0│ │VECS1│ │ │ │ │ │ ... │ │ │
48 * │ │ │ │ └──┘ └──┘ └──┘ └──┘ │ │ └─────┘ └─────┘ │ │ │ │ │ │ │ │
49 * │ │ │ │ ┌──┐ ┌──┐ ┌──┐ ┌──┐ │ │ ┌─────┐ ┌─────┐ │ │ │ │ │ │ │ │
50 * │ │ │ │ │EU│ │EU│ │EU│ │EU│ │ │ │CCS0 │ │CCS1 │ │ │ │ │ │ │ │ │
51 * │ │ │ │ └──┘ └──┘ └──┘ └──┘ │ │ └─────┘ └─────┘ │ │ │ │ │ │ │ │
52 * │ │ │ └─────────DSS─────────┘ │ ┌─────┐ ┌─────┐ │ │ │ │ │ │ │ │
53 * │ │ │ │ │CCS2 │ │CCS3 │ │ │ │ │ │ │ │ │
54 * │ │ │ ┌─────┐ ┌─────┐ ┌─────┐ │ └─────┘ └─────┘ │ │ │ │ │ │ │ │
55 * │ │ │ │ ... │ │ ... │ │ ... │ │ │ │ │ │ │ │ │ │
56 * │ │ │ └─DSS─┘ └─DSS─┘ └─DSS─┘ └─────Engines─────┘ │ │ │ │ │ │ │
57 * │ │ └───────────────────────────GT0────────────────┘ │ │ └─GT2─┘ │ │
58 * │ └────────────────────────────Tile0─────────────────┘ └─ Tile1──┘ │
59 * └─────────────────────────────Device0───────┬──────────────────────┘
60 * │
61 * ───────────────────────┴────────── PCI bus
62 */
63
64 /**
65 * DOC: Xe uAPI Overview
66 *
67 * This section aims to describe the Xe's IOCTL entries, its structs, and other
68 * Xe related uAPI such as uevents and PMU (Platform Monitoring Unit) related
69 * entries and usage.
70 *
71 * List of supported IOCTLs:
72 * - &DRM_IOCTL_XE_DEVICE_QUERY
73 * - &DRM_IOCTL_XE_GEM_CREATE
74 * - &DRM_IOCTL_XE_GEM_MMAP_OFFSET
75 * - &DRM_IOCTL_XE_VM_CREATE
76 * - &DRM_IOCTL_XE_VM_DESTROY
77 * - &DRM_IOCTL_XE_VM_BIND
78 * - &DRM_IOCTL_XE_EXEC_QUEUE_CREATE
79 * - &DRM_IOCTL_XE_EXEC_QUEUE_DESTROY
80 * - &DRM_IOCTL_XE_EXEC_QUEUE_GET_PROPERTY
81 * - &DRM_IOCTL_XE_EXEC
82 * - &DRM_IOCTL_XE_WAIT_USER_FENCE
83 * - &DRM_IOCTL_XE_OBSERVATION
84 */
85
86 /*
87 * xe specific ioctls.
88 *
89 * The device specific ioctl range is [DRM_COMMAND_BASE, DRM_COMMAND_END) ie
90 * [0x40, 0xa0) (a0 is excluded). The numbers below are defined as offset
91 * against DRM_COMMAND_BASE and should be between [0x0, 0x60).
92 */
93 #define DRM_XE_DEVICE_QUERY 0x00
94 #define DRM_XE_GEM_CREATE 0x01
95 #define DRM_XE_GEM_MMAP_OFFSET 0x02
96 #define DRM_XE_VM_CREATE 0x03
97 #define DRM_XE_VM_DESTROY 0x04
98 #define DRM_XE_VM_BIND 0x05
99 #define DRM_XE_EXEC_QUEUE_CREATE 0x06
100 #define DRM_XE_EXEC_QUEUE_DESTROY 0x07
101 #define DRM_XE_EXEC_QUEUE_GET_PROPERTY 0x08
102 #define DRM_XE_EXEC 0x09
103 #define DRM_XE_WAIT_USER_FENCE 0x0a
104 #define DRM_XE_OBSERVATION 0x0b
105
106 /* Must be kept compact -- no holes */
107
108 #define DRM_IOCTL_XE_DEVICE_QUERY DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_DEVICE_QUERY, struct drm_xe_device_query)
109 #define DRM_IOCTL_XE_GEM_CREATE DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_GEM_CREATE, struct drm_xe_gem_create)
110 #define DRM_IOCTL_XE_GEM_MMAP_OFFSET DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_GEM_MMAP_OFFSET, struct drm_xe_gem_mmap_offset)
111 #define DRM_IOCTL_XE_VM_CREATE DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_VM_CREATE, struct drm_xe_vm_create)
112 #define DRM_IOCTL_XE_VM_DESTROY DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_DESTROY, struct drm_xe_vm_destroy)
113 #define DRM_IOCTL_XE_VM_BIND DRM_IOW(DRM_COMMAND_BASE + DRM_XE_VM_BIND, struct drm_xe_vm_bind)
114 #define DRM_IOCTL_XE_EXEC_QUEUE_CREATE DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_CREATE, struct drm_xe_exec_queue_create)
115 #define DRM_IOCTL_XE_EXEC_QUEUE_DESTROY DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC_QUEUE_DESTROY, struct drm_xe_exec_queue_destroy)
116 #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)
117 #define DRM_IOCTL_XE_EXEC DRM_IOW(DRM_COMMAND_BASE + DRM_XE_EXEC, struct drm_xe_exec)
118 #define DRM_IOCTL_XE_WAIT_USER_FENCE DRM_IOWR(DRM_COMMAND_BASE + DRM_XE_WAIT_USER_FENCE, struct drm_xe_wait_user_fence)
119 #define DRM_IOCTL_XE_OBSERVATION DRM_IOW(DRM_COMMAND_BASE + DRM_XE_OBSERVATION, struct drm_xe_observation_param)
120
121 /**
122 * DOC: Xe IOCTL Extensions
123 *
124 * Before detailing the IOCTLs and its structs, it is important to highlight
125 * that every IOCTL in Xe is extensible.
126 *
127 * Many interfaces need to grow over time. In most cases we can simply
128 * extend the struct and have userspace pass in more data. Another option,
129 * as demonstrated by Vulkan's approach to providing extensions for forward
130 * and backward compatibility, is to use a list of optional structs to
131 * provide those extra details.
132 *
133 * The key advantage to using an extension chain is that it allows us to
134 * redefine the interface more easily than an ever growing struct of
135 * increasing complexity, and for large parts of that interface to be
136 * entirely optional. The downside is more pointer chasing; chasing across
137 * the __user boundary with pointers encapsulated inside u64.
138 *
139 * Example chaining:
140 *
141 * .. code-block:: C
142 *
143 * struct drm_xe_user_extension ext3 {
144 * .next_extension = 0, // end
145 * .name = ...,
146 * };
147 * struct drm_xe_user_extension ext2 {
148 * .next_extension = (uintptr_t)&ext3,
149 * .name = ...,
150 * };
151 * struct drm_xe_user_extension ext1 {
152 * .next_extension = (uintptr_t)&ext2,
153 * .name = ...,
154 * };
155 *
156 * Typically the struct drm_xe_user_extension would be embedded in some uAPI
157 * struct, and in this case we would feed it the head of the chain(i.e ext1),
158 * which would then apply all of the above extensions.
159 */
160
161 /**
162 * struct drm_xe_user_extension - Base class for defining a chain of extensions
163 */
164 struct drm_xe_user_extension {
165 /**
166 * @next_extension:
167 *
168 * Pointer to the next struct drm_xe_user_extension, or zero if the end.
169 */
170 __u64 next_extension;
171
172 /**
173 * @name: Name of the extension.
174 *
175 * Note that the name here is just some integer.
176 *
177 * Also note that the name space for this is not global for the whole
178 * driver, but rather its scope/meaning is limited to the specific piece
179 * of uAPI which has embedded the struct drm_xe_user_extension.
180 */
181 __u32 name;
182
183 /**
184 * @pad: MBZ
185 *
186 * All undefined bits must be zero.
187 */
188 __u32 pad;
189 };
190
191 /**
192 * struct drm_xe_ext_set_property - Generic set property extension
193 *
194 * A generic struct that allows any of the Xe's IOCTL to be extended
195 * with a set_property operation.
196 */
197 struct drm_xe_ext_set_property {
198 /** @base: base user extension */
199 struct drm_xe_user_extension base;
200
201 /** @property: property to set */
202 __u32 property;
203
204 /** @pad: MBZ */
205 __u32 pad;
206
207 /** @value: property value */
208 __u64 value;
209
210 /** @reserved: Reserved */
211 __u64 reserved[2];
212 };
213
214 /**
215 * struct drm_xe_engine_class_instance - instance of an engine class
216 *
217 * It is returned as part of the @drm_xe_engine, but it also is used as
218 * the input of engine selection for both @drm_xe_exec_queue_create and
219 * @drm_xe_query_engine_cycles
220 *
221 * The @engine_class can be:
222 * - %DRM_XE_ENGINE_CLASS_RENDER
223 * - %DRM_XE_ENGINE_CLASS_COPY
224 * - %DRM_XE_ENGINE_CLASS_VIDEO_DECODE
225 * - %DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE
226 * - %DRM_XE_ENGINE_CLASS_COMPUTE
227 * - %DRM_XE_ENGINE_CLASS_VM_BIND - Kernel only classes (not actual
228 * hardware engine class). Used for creating ordered queues of VM
229 * bind operations.
230 */
231 struct drm_xe_engine_class_instance {
232 #define DRM_XE_ENGINE_CLASS_RENDER 0
233 #define DRM_XE_ENGINE_CLASS_COPY 1
234 #define DRM_XE_ENGINE_CLASS_VIDEO_DECODE 2
235 #define DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE 3
236 #define DRM_XE_ENGINE_CLASS_COMPUTE 4
237 #define DRM_XE_ENGINE_CLASS_VM_BIND 5
238 /** @engine_class: engine class id */
239 __u16 engine_class;
240 /** @engine_instance: engine instance id */
241 __u16 engine_instance;
242 /** @gt_id: Unique ID of this GT within the PCI Device */
243 __u16 gt_id;
244 /** @pad: MBZ */
245 __u16 pad;
246 };
247
248 /**
249 * struct drm_xe_engine - describe hardware engine
250 */
251 struct drm_xe_engine {
252 /** @instance: The @drm_xe_engine_class_instance */
253 struct drm_xe_engine_class_instance instance;
254
255 /** @reserved: Reserved */
256 __u64 reserved[3];
257 };
258
259 /**
260 * struct drm_xe_query_engines - describe engines
261 *
262 * If a query is made with a struct @drm_xe_device_query where .query
263 * is equal to %DRM_XE_DEVICE_QUERY_ENGINES, then the reply uses an array of
264 * struct @drm_xe_query_engines in .data.
265 */
266 struct drm_xe_query_engines {
267 /** @num_engines: number of engines returned in @engines */
268 __u32 num_engines;
269 /** @pad: MBZ */
270 __u32 pad;
271 /** @engines: The returned engines for this device */
272 struct drm_xe_engine engines[];
273 };
274
275 /**
276 * enum drm_xe_memory_class - Supported memory classes.
277 */
278 enum drm_xe_memory_class {
279 /** @DRM_XE_MEM_REGION_CLASS_SYSMEM: Represents system memory. */
280 DRM_XE_MEM_REGION_CLASS_SYSMEM = 0,
281 /**
282 * @DRM_XE_MEM_REGION_CLASS_VRAM: On discrete platforms, this
283 * represents the memory that is local to the device, which we
284 * call VRAM. Not valid on integrated platforms.
285 */
286 DRM_XE_MEM_REGION_CLASS_VRAM
287 };
288
289 /**
290 * struct drm_xe_mem_region - Describes some region as known to
291 * the driver.
292 */
293 struct drm_xe_mem_region {
294 /**
295 * @mem_class: The memory class describing this region.
296 *
297 * See enum drm_xe_memory_class for supported values.
298 */
299 __u16 mem_class;
300 /**
301 * @instance: The unique ID for this region, which serves as the
302 * index in the placement bitmask used as argument for
303 * &DRM_IOCTL_XE_GEM_CREATE
304 */
305 __u16 instance;
306 /**
307 * @min_page_size: Min page-size in bytes for this region.
308 *
309 * When the kernel allocates memory for this region, the
310 * underlying pages will be at least @min_page_size in size.
311 * Buffer objects with an allowable placement in this region must be
312 * created with a size aligned to this value.
313 * GPU virtual address mappings of (parts of) buffer objects that
314 * may be placed in this region must also have their GPU virtual
315 * address and range aligned to this value.
316 * Affected IOCTLS will return %-EINVAL if alignment restrictions are
317 * not met.
318 */
319 __u32 min_page_size;
320 /**
321 * @total_size: The usable size in bytes for this region.
322 */
323 __u64 total_size;
324 /**
325 * @used: Estimate of the memory used in bytes for this region.
326 *
327 * Requires CAP_PERFMON or CAP_SYS_ADMIN to get reliable
328 * accounting. Without this the value here will always equal
329 * zero.
330 */
331 __u64 used;
332 /**
333 * @cpu_visible_size: How much of this region can be CPU
334 * accessed, in bytes.
335 *
336 * This will always be <= @total_size, and the remainder (if
337 * any) will not be CPU accessible. If the CPU accessible part
338 * is smaller than @total_size then this is referred to as a
339 * small BAR system.
340 *
341 * On systems without small BAR (full BAR), the probed_size will
342 * always equal the @total_size, since all of it will be CPU
343 * accessible.
344 *
345 * Note this is only tracked for DRM_XE_MEM_REGION_CLASS_VRAM
346 * regions (for other types the value here will always equal
347 * zero).
348 */
349 __u64 cpu_visible_size;
350 /**
351 * @cpu_visible_used: Estimate of CPU visible memory used, in
352 * bytes.
353 *
354 * Requires CAP_PERFMON or CAP_SYS_ADMIN to get reliable
355 * accounting. Without this the value here will always equal
356 * zero. Note this is only currently tracked for
357 * DRM_XE_MEM_REGION_CLASS_VRAM regions (for other types the value
358 * here will always be zero).
359 */
360 __u64 cpu_visible_used;
361 /** @reserved: Reserved */
362 __u64 reserved[6];
363 };
364
365 /**
366 * struct drm_xe_query_mem_regions - describe memory regions
367 *
368 * If a query is made with a struct drm_xe_device_query where .query
369 * is equal to DRM_XE_DEVICE_QUERY_MEM_REGIONS, then the reply uses
370 * struct drm_xe_query_mem_regions in .data.
371 */
372 struct drm_xe_query_mem_regions {
373 /** @num_mem_regions: number of memory regions returned in @mem_regions */
374 __u32 num_mem_regions;
375 /** @pad: MBZ */
376 __u32 pad;
377 /** @mem_regions: The returned memory regions for this device */
378 struct drm_xe_mem_region mem_regions[];
379 };
380
381 /**
382 * struct drm_xe_query_config - describe the device configuration
383 *
384 * If a query is made with a struct drm_xe_device_query where .query
385 * is equal to DRM_XE_DEVICE_QUERY_CONFIG, then the reply uses
386 * struct drm_xe_query_config in .data.
387 *
388 * The index in @info can be:
389 * - %DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID - Device ID (lower 16 bits)
390 * and the device revision (next 8 bits)
391 * - %DRM_XE_QUERY_CONFIG_FLAGS - Flags describing the device
392 * configuration, see list below
393 *
394 * - %DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM - Flag is set if the device
395 * has usable VRAM
396 * - %DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT - Minimal memory alignment
397 * required by this device, typically SZ_4K or SZ_64K
398 * - %DRM_XE_QUERY_CONFIG_VA_BITS - Maximum bits of a virtual address
399 * - %DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY - Value of the highest
400 * available exec queue priority
401 */
402 struct drm_xe_query_config {
403 /** @num_params: number of parameters returned in info */
404 __u32 num_params;
405
406 /** @pad: MBZ */
407 __u32 pad;
408
409 #define DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID 0
410 #define DRM_XE_QUERY_CONFIG_FLAGS 1
411 #define DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM (1 << 0)
412 #define DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT 2
413 #define DRM_XE_QUERY_CONFIG_VA_BITS 3
414 #define DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY 4
415 /** @info: array of elements containing the config info */
416 __u64 info[];
417 };
418
419 /**
420 * struct drm_xe_gt - describe an individual GT.
421 *
422 * To be used with drm_xe_query_gt_list, which will return a list with all the
423 * existing GT individual descriptions.
424 * Graphics Technology (GT) is a subset of a GPU/tile that is responsible for
425 * implementing graphics and/or media operations.
426 *
427 * The index in @type can be:
428 * - %DRM_XE_QUERY_GT_TYPE_MAIN
429 * - %DRM_XE_QUERY_GT_TYPE_MEDIA
430 */
431 struct drm_xe_gt {
432 #define DRM_XE_QUERY_GT_TYPE_MAIN 0
433 #define DRM_XE_QUERY_GT_TYPE_MEDIA 1
434 /** @type: GT type: Main or Media */
435 __u16 type;
436 /** @tile_id: Tile ID where this GT lives (Information only) */
437 __u16 tile_id;
438 /** @gt_id: Unique ID of this GT within the PCI Device */
439 __u16 gt_id;
440 /** @pad: MBZ */
441 __u16 pad[3];
442 /** @reference_clock: A clock frequency for timestamp */
443 __u32 reference_clock;
444 /**
445 * @near_mem_regions: Bit mask of instances from
446 * drm_xe_query_mem_regions that are nearest to the current engines
447 * of this GT.
448 * Each index in this mask refers directly to the struct
449 * drm_xe_query_mem_regions' instance, no assumptions should
450 * be made about order. The type of each region is described
451 * by struct drm_xe_query_mem_regions' mem_class.
452 */
453 __u64 near_mem_regions;
454 /**
455 * @far_mem_regions: Bit mask of instances from
456 * drm_xe_query_mem_regions that are far from the engines of this GT.
457 * In general, they have extra indirections when compared to the
458 * @near_mem_regions. For a discrete device this could mean system
459 * memory and memory living in a different tile.
460 * Each index in this mask refers directly to the struct
461 * drm_xe_query_mem_regions' instance, no assumptions should
462 * be made about order. The type of each region is described
463 * by struct drm_xe_query_mem_regions' mem_class.
464 */
465 __u64 far_mem_regions;
466 /** @ip_ver_major: Graphics/media IP major version on GMD_ID platforms */
467 __u16 ip_ver_major;
468 /** @ip_ver_minor: Graphics/media IP minor version on GMD_ID platforms */
469 __u16 ip_ver_minor;
470 /** @ip_ver_rev: Graphics/media IP revision version on GMD_ID platforms */
471 __u16 ip_ver_rev;
472 /** @pad2: MBZ */
473 __u16 pad2;
474 /** @reserved: Reserved */
475 __u64 reserved[7];
476 };
477
478 /**
479 * struct drm_xe_query_gt_list - A list with GT description items.
480 *
481 * If a query is made with a struct drm_xe_device_query where .query
482 * is equal to DRM_XE_DEVICE_QUERY_GT_LIST, then the reply uses struct
483 * drm_xe_query_gt_list in .data.
484 */
485 struct drm_xe_query_gt_list {
486 /** @num_gt: number of GT items returned in gt_list */
487 __u32 num_gt;
488 /** @pad: MBZ */
489 __u32 pad;
490 /** @gt_list: The GT list returned for this device */
491 struct drm_xe_gt gt_list[];
492 };
493
494 /**
495 * struct drm_xe_query_topology_mask - describe the topology mask of a GT
496 *
497 * This is the hardware topology which reflects the internal physical
498 * structure of the GPU.
499 *
500 * If a query is made with a struct drm_xe_device_query where .query
501 * is equal to DRM_XE_DEVICE_QUERY_GT_TOPOLOGY, then the reply uses
502 * struct drm_xe_query_topology_mask in .data.
503 *
504 * The @type can be:
505 * - %DRM_XE_TOPO_DSS_GEOMETRY - To query the mask of Dual Sub Slices
506 * (DSS) available for geometry operations. For example a query response
507 * containing the following in mask:
508 * ``DSS_GEOMETRY ff ff ff ff 00 00 00 00``
509 * means 32 DSS are available for geometry.
510 * - %DRM_XE_TOPO_DSS_COMPUTE - To query the mask of Dual Sub Slices
511 * (DSS) available for compute operations. For example a query response
512 * containing the following in mask:
513 * ``DSS_COMPUTE ff ff ff ff 00 00 00 00``
514 * means 32 DSS are available for compute.
515 * - %DRM_XE_TOPO_L3_BANK - To query the mask of enabled L3 banks. This type
516 * may be omitted if the driver is unable to query the mask from the
517 * hardware.
518 * - %DRM_XE_TOPO_EU_PER_DSS - To query the mask of Execution Units (EU)
519 * available per Dual Sub Slices (DSS). For example a query response
520 * containing the following in mask:
521 * ``EU_PER_DSS ff ff 00 00 00 00 00 00``
522 * means each DSS has 16 SIMD8 EUs. This type may be omitted if device
523 * doesn't have SIMD8 EUs.
524 * - %DRM_XE_TOPO_SIMD16_EU_PER_DSS - To query the mask of SIMD16 Execution
525 * Units (EU) available per Dual Sub Slices (DSS). For example a query
526 * response containing the following in mask:
527 * ``SIMD16_EU_PER_DSS ff ff 00 00 00 00 00 00``
528 * means each DSS has 16 SIMD16 EUs. This type may be omitted if device
529 * doesn't have SIMD16 EUs.
530 */
531 struct drm_xe_query_topology_mask {
532 /** @gt_id: GT ID the mask is associated with */
533 __u16 gt_id;
534
535 #define DRM_XE_TOPO_DSS_GEOMETRY 1
536 #define DRM_XE_TOPO_DSS_COMPUTE 2
537 #define DRM_XE_TOPO_L3_BANK 3
538 #define DRM_XE_TOPO_EU_PER_DSS 4
539 #define DRM_XE_TOPO_SIMD16_EU_PER_DSS 5
540 /** @type: type of mask */
541 __u16 type;
542
543 /** @num_bytes: number of bytes in requested mask */
544 __u32 num_bytes;
545
546 /** @mask: little-endian mask of @num_bytes */
547 __u8 mask[];
548 };
549
550 /**
551 * struct drm_xe_query_engine_cycles - correlate CPU and GPU timestamps
552 *
553 * If a query is made with a struct drm_xe_device_query where .query is equal to
554 * DRM_XE_DEVICE_QUERY_ENGINE_CYCLES, then the reply uses struct drm_xe_query_engine_cycles
555 * in .data. struct drm_xe_query_engine_cycles is allocated by the user and
556 * .data points to this allocated structure.
557 *
558 * The query returns the engine cycles, which along with GT's @reference_clock,
559 * can be used to calculate the engine timestamp. In addition the
560 * query returns a set of cpu timestamps that indicate when the command
561 * streamer cycle count was captured.
562 */
563 struct drm_xe_query_engine_cycles {
564 /**
565 * @eci: This is input by the user and is the engine for which command
566 * streamer cycles is queried.
567 */
568 struct drm_xe_engine_class_instance eci;
569
570 /**
571 * @clockid: This is input by the user and is the reference clock id for
572 * CPU timestamp. For definition, see clock_gettime(2) and
573 * perf_event_open(2). Supported clock ids are CLOCK_MONOTONIC,
574 * CLOCK_MONOTONIC_RAW, CLOCK_REALTIME, CLOCK_BOOTTIME, CLOCK_TAI.
575 */
576 __s32 clockid;
577
578 /** @width: Width of the engine cycle counter in bits. */
579 __u32 width;
580
581 /**
582 * @engine_cycles: Engine cycles as read from its register
583 * at 0x358 offset.
584 */
585 __u64 engine_cycles;
586
587 /**
588 * @cpu_timestamp: CPU timestamp in ns. The timestamp is captured before
589 * reading the engine_cycles register using the reference clockid set by the
590 * user.
591 */
592 __u64 cpu_timestamp;
593
594 /**
595 * @cpu_delta: Time delta in ns captured around reading the lower dword
596 * of the engine_cycles register.
597 */
598 __u64 cpu_delta;
599 };
600
601 /**
602 * struct drm_xe_query_uc_fw_version - query a micro-controller firmware version
603 *
604 * Given a uc_type this will return the branch, major, minor and patch version
605 * of the micro-controller firmware.
606 */
607 struct drm_xe_query_uc_fw_version {
608 /** @uc_type: The micro-controller type to query firmware version */
609 #define XE_QUERY_UC_TYPE_GUC_SUBMISSION 0
610 #define XE_QUERY_UC_TYPE_HUC 1
611 __u16 uc_type;
612
613 /** @pad: MBZ */
614 __u16 pad;
615
616 /** @branch_ver: branch uc fw version */
617 __u32 branch_ver;
618 /** @major_ver: major uc fw version */
619 __u32 major_ver;
620 /** @minor_ver: minor uc fw version */
621 __u32 minor_ver;
622 /** @patch_ver: patch uc fw version */
623 __u32 patch_ver;
624
625 /** @pad2: MBZ */
626 __u32 pad2;
627
628 /** @reserved: Reserved */
629 __u64 reserved;
630 };
631
632 /**
633 * struct drm_xe_query_pxp_status - query if PXP is ready
634 *
635 * If PXP is enabled and no fatal error has occurred, the status will be set to
636 * one of the following values:
637 * 0: PXP init still in progress
638 * 1: PXP init complete
639 *
640 * If PXP is not enabled or something has gone wrong, the query will be failed
641 * with one of the following error codes:
642 * -ENODEV: PXP not supported or disabled;
643 * -EIO: fatal error occurred during init, so PXP will never be enabled;
644 * -EINVAL: incorrect value provided as part of the query;
645 * -EFAULT: error copying the memory between kernel and userspace.
646 *
647 * The status can only be 0 in the first few seconds after driver load. If
648 * everything works as expected, the status will transition to init complete in
649 * less than 1 second, while in case of errors the driver might take longer to
650 * start returning an error code, but it should still take less than 10 seconds.
651 *
652 * The supported session type bitmask is based on the values in
653 * enum drm_xe_pxp_session_type. TYPE_NONE is always supported and therefore
654 * is not reported in the bitmask.
655 *
656 */
657 struct drm_xe_query_pxp_status {
658 /** @status: current PXP status */
659 __u32 status;
660
661 /** @supported_session_types: bitmask of supported PXP session types */
662 __u32 supported_session_types;
663 };
664
665 /**
666 * struct drm_xe_device_query - Input of &DRM_IOCTL_XE_DEVICE_QUERY - main
667 * structure to query device information
668 *
669 * The user selects the type of data to query among DRM_XE_DEVICE_QUERY_*
670 * and sets the value in the query member. This determines the type of
671 * the structure provided by the driver in data, among struct drm_xe_query_*.
672 *
673 * The @query can be:
674 * - %DRM_XE_DEVICE_QUERY_ENGINES
675 * - %DRM_XE_DEVICE_QUERY_MEM_REGIONS
676 * - %DRM_XE_DEVICE_QUERY_CONFIG
677 * - %DRM_XE_DEVICE_QUERY_GT_LIST
678 * - %DRM_XE_DEVICE_QUERY_HWCONFIG - Query type to retrieve the hardware
679 * configuration of the device such as information on slices, memory,
680 * caches, and so on. It is provided as a table of key / value
681 * attributes.
682 * - %DRM_XE_DEVICE_QUERY_GT_TOPOLOGY
683 * - %DRM_XE_DEVICE_QUERY_ENGINE_CYCLES
684 * - %DRM_XE_DEVICE_QUERY_PXP_STATUS
685 *
686 * If size is set to 0, the driver fills it with the required size for
687 * the requested type of data to query. If size is equal to the required
688 * size, the queried information is copied into data. If size is set to
689 * a value different from 0 and different from the required size, the
690 * IOCTL call returns -EINVAL.
691 *
692 * For example the following code snippet allows retrieving and printing
693 * information about the device engines with DRM_XE_DEVICE_QUERY_ENGINES:
694 *
695 * .. code-block:: C
696 *
697 * struct drm_xe_query_engines *engines;
698 * struct drm_xe_device_query query = {
699 * .extensions = 0,
700 * .query = DRM_XE_DEVICE_QUERY_ENGINES,
701 * .size = 0,
702 * .data = 0,
703 * };
704 * ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query);
705 * engines = malloc(query.size);
706 * query.data = (uintptr_t)engines;
707 * ioctl(fd, DRM_IOCTL_XE_DEVICE_QUERY, &query);
708 * for (int i = 0; i < engines->num_engines; i++) {
709 * printf("Engine %d: %s\n", i,
710 * engines->engines[i].instance.engine_class ==
711 * DRM_XE_ENGINE_CLASS_RENDER ? "RENDER":
712 * engines->engines[i].instance.engine_class ==
713 * DRM_XE_ENGINE_CLASS_COPY ? "COPY":
714 * engines->engines[i].instance.engine_class ==
715 * DRM_XE_ENGINE_CLASS_VIDEO_DECODE ? "VIDEO_DECODE":
716 * engines->engines[i].instance.engine_class ==
717 * DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE ? "VIDEO_ENHANCE":
718 * engines->engines[i].instance.engine_class ==
719 * DRM_XE_ENGINE_CLASS_COMPUTE ? "COMPUTE":
720 * "UNKNOWN");
721 * }
722 * free(engines);
723 */
724 struct drm_xe_device_query {
725 /** @extensions: Pointer to the first extension struct, if any */
726 __u64 extensions;
727
728 #define DRM_XE_DEVICE_QUERY_ENGINES 0
729 #define DRM_XE_DEVICE_QUERY_MEM_REGIONS 1
730 #define DRM_XE_DEVICE_QUERY_CONFIG 2
731 #define DRM_XE_DEVICE_QUERY_GT_LIST 3
732 #define DRM_XE_DEVICE_QUERY_HWCONFIG 4
733 #define DRM_XE_DEVICE_QUERY_GT_TOPOLOGY 5
734 #define DRM_XE_DEVICE_QUERY_ENGINE_CYCLES 6
735 #define DRM_XE_DEVICE_QUERY_UC_FW_VERSION 7
736 #define DRM_XE_DEVICE_QUERY_OA_UNITS 8
737 #define DRM_XE_DEVICE_QUERY_PXP_STATUS 9
738 /** @query: The type of data to query */
739 __u32 query;
740
741 /** @size: Size of the queried data */
742 __u32 size;
743
744 /** @data: Queried data is placed here */
745 __u64 data;
746
747 /** @reserved: Reserved */
748 __u64 reserved[2];
749 };
750
751 /**
752 * struct drm_xe_gem_create - Input of &DRM_IOCTL_XE_GEM_CREATE - A structure for
753 * gem creation
754 *
755 * The @flags can be:
756 * - %DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING
757 * - %DRM_XE_GEM_CREATE_FLAG_SCANOUT
758 * - %DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM - When using VRAM as a
759 * possible placement, ensure that the corresponding VRAM allocation
760 * will always use the CPU accessible part of VRAM. This is important
761 * for small-bar systems (on full-bar systems this gets turned into a
762 * noop).
763 * Note1: System memory can be used as an extra placement if the kernel
764 * should spill the allocation to system memory, if space can't be made
765 * available in the CPU accessible part of VRAM (giving the same
766 * behaviour as the i915 interface, see
767 * I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS).
768 * Note2: For clear-color CCS surfaces the kernel needs to read the
769 * clear-color value stored in the buffer, and on discrete platforms we
770 * need to use VRAM for display surfaces, therefore the kernel requires
771 * setting this flag for such objects, otherwise an error is thrown on
772 * small-bar systems.
773 *
774 * @cpu_caching supports the following values:
775 * - %DRM_XE_GEM_CPU_CACHING_WB - Allocate the pages with write-back
776 * caching. On iGPU this can't be used for scanout surfaces. Currently
777 * not allowed for objects placed in VRAM.
778 * - %DRM_XE_GEM_CPU_CACHING_WC - Allocate the pages as write-combined. This
779 * is uncached. Scanout surfaces should likely use this. All objects
780 * that can be placed in VRAM must use this.
781 *
782 * This ioctl supports setting the following properties via the
783 * %DRM_XE_GEM_CREATE_EXTENSION_SET_PROPERTY extension, which uses the
784 * generic @drm_xe_ext_set_property struct:
785 *
786 * - %DRM_XE_GEM_CREATE_SET_PROPERTY_PXP_TYPE - set the type of PXP session
787 * this object will be used with. Valid values are listed in enum
788 * drm_xe_pxp_session_type. %DRM_XE_PXP_TYPE_NONE is the default behavior, so
789 * there is no need to explicitly set that. Objects used with session of type
790 * %DRM_XE_PXP_TYPE_HWDRM will be marked as invalid if a PXP invalidation
791 * event occurs after their creation. Attempting to flip an invalid object
792 * will cause a black frame to be displayed instead. Submissions with invalid
793 * objects mapped in the VM will be rejected.
794 */
795 struct drm_xe_gem_create {
796 #define DRM_XE_GEM_CREATE_EXTENSION_SET_PROPERTY 0
797 #define DRM_XE_GEM_CREATE_SET_PROPERTY_PXP_TYPE 0
798 /** @extensions: Pointer to the first extension struct, if any */
799 __u64 extensions;
800
801 /**
802 * @size: Size of the object to be created, must match region
803 * (system or vram) minimum alignment (&min_page_size).
804 */
805 __u64 size;
806
807 /**
808 * @placement: A mask of memory instances of where BO can be placed.
809 * Each index in this mask refers directly to the struct
810 * drm_xe_query_mem_regions' instance, no assumptions should
811 * be made about order. The type of each region is described
812 * by struct drm_xe_query_mem_regions' mem_class.
813 */
814 __u32 placement;
815
816 #define DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING (1 << 0)
817 #define DRM_XE_GEM_CREATE_FLAG_SCANOUT (1 << 1)
818 #define DRM_XE_GEM_CREATE_FLAG_NEEDS_VISIBLE_VRAM (1 << 2)
819 /**
820 * @flags: Flags, currently a mask of memory instances of where BO can
821 * be placed
822 */
823 __u32 flags;
824
825 /**
826 * @vm_id: Attached VM, if any
827 *
828 * If a VM is specified, this BO must:
829 *
830 * 1. Only ever be bound to that VM.
831 * 2. Cannot be exported as a PRIME fd.
832 */
833 __u32 vm_id;
834
835 /**
836 * @handle: Returned handle for the object.
837 *
838 * Object handles are nonzero.
839 */
840 __u32 handle;
841
842 #define DRM_XE_GEM_CPU_CACHING_WB 1
843 #define DRM_XE_GEM_CPU_CACHING_WC 2
844 /**
845 * @cpu_caching: The CPU caching mode to select for this object. If
846 * mmaping the object the mode selected here will also be used. The
847 * exception is when mapping system memory (including data evicted
848 * to system) on discrete GPUs. The caching mode selected will
849 * then be overridden to DRM_XE_GEM_CPU_CACHING_WB, and coherency
850 * between GPU- and CPU is guaranteed. The caching mode of
851 * existing CPU-mappings will be updated transparently to
852 * user-space clients.
853 */
854 __u16 cpu_caching;
855 /** @pad: MBZ */
856 __u16 pad[3];
857
858 /** @reserved: Reserved */
859 __u64 reserved[2];
860 };
861
862 /**
863 * struct drm_xe_gem_mmap_offset - Input of &DRM_IOCTL_XE_GEM_MMAP_OFFSET
864 *
865 * The @flags can be:
866 * - %DRM_XE_MMAP_OFFSET_FLAG_PCI_BARRIER - For user to query special offset
867 * for use in mmap ioctl. Writing to the returned mmap address will generate a
868 * PCI memory barrier with low overhead (avoiding IOCTL call as well as writing
869 * to VRAM which would also add overhead), acting like an MI_MEM_FENCE
870 * instruction.
871 *
872 * Note: The mmap size can be at most 4K, due to HW limitations. As a result
873 * this interface is only supported on CPU architectures that support 4K page
874 * size. The mmap_offset ioctl will detect this and gracefully return an
875 * error, where userspace is expected to have a different fallback method for
876 * triggering a barrier.
877 *
878 * Roughly the usage would be as follows:
879 *
880 * .. code-block:: C
881 *
882 * struct drm_xe_gem_mmap_offset mmo = {
883 * .handle = 0, // must be set to 0
884 * .flags = DRM_XE_MMAP_OFFSET_FLAG_PCI_BARRIER,
885 * };
886 *
887 * err = ioctl(fd, DRM_IOCTL_XE_GEM_MMAP_OFFSET, &mmo);
888 * map = mmap(NULL, size, PROT_WRITE, MAP_SHARED, fd, mmo.offset);
889 * map[i] = 0xdeadbeaf; // issue barrier
890 */
891 struct drm_xe_gem_mmap_offset {
892 /** @extensions: Pointer to the first extension struct, if any */
893 __u64 extensions;
894
895 /** @handle: Handle for the object being mapped. */
896 __u32 handle;
897
898 #define DRM_XE_MMAP_OFFSET_FLAG_PCI_BARRIER (1 << 0)
899 /** @flags: Flags */
900 __u32 flags;
901
902 /** @offset: The fake offset to use for subsequent mmap call */
903 __u64 offset;
904
905 /** @reserved: Reserved */
906 __u64 reserved[2];
907 };
908
909 /**
910 * struct drm_xe_vm_create - Input of &DRM_IOCTL_XE_VM_CREATE
911 *
912 * The @flags can be:
913 * - %DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE
914 * - %DRM_XE_VM_CREATE_FLAG_LR_MODE - An LR, or Long Running VM accepts
915 * exec submissions to its exec_queues that don't have an upper time
916 * limit on the job execution time. But exec submissions to these
917 * don't allow any of the flags DRM_XE_SYNC_FLAG_SYNCOBJ,
918 * DRM_XE_SYNC_FLAG_TIMELINE_SYNCOBJ, DRM_XE_SYNC_FLAG_DMA_BUF,
919 * used as out-syncobjs, that is, together with DRM_XE_SYNC_FLAG_SIGNAL.
920 * LR VMs can be created in recoverable page-fault mode using
921 * DRM_XE_VM_CREATE_FLAG_FAULT_MODE, if the device supports it.
922 * If that flag is omitted, the UMD can not rely on the slightly
923 * different per-VM overcommit semantics that are enabled by
924 * DRM_XE_VM_CREATE_FLAG_FAULT_MODE (see below), but KMD may
925 * still enable recoverable pagefaults if supported by the device.
926 * - %DRM_XE_VM_CREATE_FLAG_FAULT_MODE - Requires also
927 * DRM_XE_VM_CREATE_FLAG_LR_MODE. It allows memory to be allocated on
928 * demand when accessed, and also allows per-VM overcommit of memory.
929 * The xe driver internally uses recoverable pagefaults to implement
930 * this.
931 */
932 struct drm_xe_vm_create {
933 /** @extensions: Pointer to the first extension struct, if any */
934 __u64 extensions;
935
936 #define DRM_XE_VM_CREATE_FLAG_SCRATCH_PAGE (1 << 0)
937 #define DRM_XE_VM_CREATE_FLAG_LR_MODE (1 << 1)
938 #define DRM_XE_VM_CREATE_FLAG_FAULT_MODE (1 << 2)
939 /** @flags: Flags */
940 __u32 flags;
941
942 /** @vm_id: Returned VM ID */
943 __u32 vm_id;
944
945 /** @reserved: Reserved */
946 __u64 reserved[2];
947 };
948
949 /**
950 * struct drm_xe_vm_destroy - Input of &DRM_IOCTL_XE_VM_DESTROY
951 */
952 struct drm_xe_vm_destroy {
953 /** @vm_id: VM ID */
954 __u32 vm_id;
955
956 /** @pad: MBZ */
957 __u32 pad;
958
959 /** @reserved: Reserved */
960 __u64 reserved[2];
961 };
962
963 /**
964 * struct drm_xe_vm_bind_op - run bind operations
965 *
966 * The @op can be:
967 * - %DRM_XE_VM_BIND_OP_MAP
968 * - %DRM_XE_VM_BIND_OP_UNMAP
969 * - %DRM_XE_VM_BIND_OP_MAP_USERPTR
970 * - %DRM_XE_VM_BIND_OP_UNMAP_ALL
971 * - %DRM_XE_VM_BIND_OP_PREFETCH
972 *
973 * and the @flags can be:
974 * - %DRM_XE_VM_BIND_FLAG_READONLY - Setup the page tables as read-only
975 * to ensure write protection
976 * - %DRM_XE_VM_BIND_FLAG_IMMEDIATE - On a faulting VM, do the
977 * MAP operation immediately rather than deferring the MAP to the page
978 * fault handler. This is implied on a non-faulting VM as there is no
979 * fault handler to defer to.
980 * - %DRM_XE_VM_BIND_FLAG_NULL - When the NULL flag is set, the page
981 * tables are setup with a special bit which indicates writes are
982 * dropped and all reads return zero. In the future, the NULL flags
983 * will only be valid for DRM_XE_VM_BIND_OP_MAP operations, the BO
984 * handle MBZ, and the BO offset MBZ. This flag is intended to
985 * implement VK sparse bindings.
986 * - %DRM_XE_VM_BIND_FLAG_CHECK_PXP - If the object is encrypted via PXP,
987 * reject the binding if the encryption key is no longer valid. This
988 * flag has no effect on BOs that are not marked as using PXP.
989 */
990 struct drm_xe_vm_bind_op {
991 /** @extensions: Pointer to the first extension struct, if any */
992 __u64 extensions;
993
994 /**
995 * @obj: GEM object to operate on, MBZ for MAP_USERPTR, MBZ for UNMAP
996 */
997 __u32 obj;
998
999 /**
1000 * @pat_index: The platform defined @pat_index to use for this mapping.
1001 * The index basically maps to some predefined memory attributes,
1002 * including things like caching, coherency, compression etc. The exact
1003 * meaning of the pat_index is platform specific and defined in the
1004 * Bspec and PRMs. When the KMD sets up the binding the index here is
1005 * encoded into the ppGTT PTE.
1006 *
1007 * For coherency the @pat_index needs to be at least 1way coherent when
1008 * drm_xe_gem_create.cpu_caching is DRM_XE_GEM_CPU_CACHING_WB. The KMD
1009 * will extract the coherency mode from the @pat_index and reject if
1010 * there is a mismatch (see note below for pre-MTL platforms).
1011 *
1012 * Note: On pre-MTL platforms there is only a caching mode and no
1013 * explicit coherency mode, but on such hardware there is always a
1014 * shared-LLC (or is dgpu) so all GT memory accesses are coherent with
1015 * CPU caches even with the caching mode set as uncached. It's only the
1016 * display engine that is incoherent (on dgpu it must be in VRAM which
1017 * is always mapped as WC on the CPU). However to keep the uapi somewhat
1018 * consistent with newer platforms the KMD groups the different cache
1019 * levels into the following coherency buckets on all pre-MTL platforms:
1020 *
1021 * ppGTT UC -> COH_NONE
1022 * ppGTT WC -> COH_NONE
1023 * ppGTT WT -> COH_NONE
1024 * ppGTT WB -> COH_AT_LEAST_1WAY
1025 *
1026 * In practice UC/WC/WT should only ever used for scanout surfaces on
1027 * such platforms (or perhaps in general for dma-buf if shared with
1028 * another device) since it is only the display engine that is actually
1029 * incoherent. Everything else should typically use WB given that we
1030 * have a shared-LLC. On MTL+ this completely changes and the HW
1031 * defines the coherency mode as part of the @pat_index, where
1032 * incoherent GT access is possible.
1033 *
1034 * Note: For userptr and externally imported dma-buf the kernel expects
1035 * either 1WAY or 2WAY for the @pat_index.
1036 *
1037 * For DRM_XE_VM_BIND_FLAG_NULL bindings there are no KMD restrictions
1038 * on the @pat_index. For such mappings there is no actual memory being
1039 * mapped (the address in the PTE is invalid), so the various PAT memory
1040 * attributes likely do not apply. Simply leaving as zero is one
1041 * option (still a valid pat_index).
1042 */
1043 __u16 pat_index;
1044
1045 /** @pad: MBZ */
1046 __u16 pad;
1047
1048 union {
1049 /**
1050 * @obj_offset: Offset into the object, MBZ for CLEAR_RANGE,
1051 * ignored for unbind
1052 */
1053 __u64 obj_offset;
1054
1055 /** @userptr: user pointer to bind on */
1056 __u64 userptr;
1057 };
1058
1059 /**
1060 * @range: Number of bytes from the object to bind to addr, MBZ for UNMAP_ALL
1061 */
1062 __u64 range;
1063
1064 /** @addr: Address to operate on, MBZ for UNMAP_ALL */
1065 __u64 addr;
1066
1067 #define DRM_XE_VM_BIND_OP_MAP 0x0
1068 #define DRM_XE_VM_BIND_OP_UNMAP 0x1
1069 #define DRM_XE_VM_BIND_OP_MAP_USERPTR 0x2
1070 #define DRM_XE_VM_BIND_OP_UNMAP_ALL 0x3
1071 #define DRM_XE_VM_BIND_OP_PREFETCH 0x4
1072 /** @op: Bind operation to perform */
1073 __u32 op;
1074
1075 #define DRM_XE_VM_BIND_FLAG_READONLY (1 << 0)
1076 #define DRM_XE_VM_BIND_FLAG_IMMEDIATE (1 << 1)
1077 #define DRM_XE_VM_BIND_FLAG_NULL (1 << 2)
1078 #define DRM_XE_VM_BIND_FLAG_DUMPABLE (1 << 3)
1079 #define DRM_XE_VM_BIND_FLAG_CHECK_PXP (1 << 4)
1080 /** @flags: Bind flags */
1081 __u32 flags;
1082
1083 /**
1084 * @prefetch_mem_region_instance: Memory region to prefetch VMA to.
1085 * It is a region instance, not a mask.
1086 * To be used only with %DRM_XE_VM_BIND_OP_PREFETCH operation.
1087 */
1088 __u32 prefetch_mem_region_instance;
1089
1090 /** @pad2: MBZ */
1091 __u32 pad2;
1092
1093 /** @reserved: Reserved */
1094 __u64 reserved[3];
1095 };
1096
1097 /**
1098 * struct drm_xe_vm_bind - Input of &DRM_IOCTL_XE_VM_BIND
1099 *
1100 * Below is an example of a minimal use of @drm_xe_vm_bind to
1101 * asynchronously bind the buffer `data` at address `BIND_ADDRESS` to
1102 * illustrate `userptr`. It can be synchronized by using the example
1103 * provided for @drm_xe_sync.
1104 *
1105 * .. code-block:: C
1106 *
1107 * data = aligned_alloc(ALIGNMENT, BO_SIZE);
1108 * struct drm_xe_vm_bind bind = {
1109 * .vm_id = vm,
1110 * .num_binds = 1,
1111 * .bind.obj = 0,
1112 * .bind.obj_offset = to_user_pointer(data),
1113 * .bind.range = BO_SIZE,
1114 * .bind.addr = BIND_ADDRESS,
1115 * .bind.op = DRM_XE_VM_BIND_OP_MAP_USERPTR,
1116 * .bind.flags = 0,
1117 * .num_syncs = 1,
1118 * .syncs = &sync,
1119 * .exec_queue_id = 0,
1120 * };
1121 * ioctl(fd, DRM_IOCTL_XE_VM_BIND, &bind);
1122 *
1123 */
1124 struct drm_xe_vm_bind {
1125 /** @extensions: Pointer to the first extension struct, if any */
1126 __u64 extensions;
1127
1128 /** @vm_id: The ID of the VM to bind to */
1129 __u32 vm_id;
1130
1131 /**
1132 * @exec_queue_id: exec_queue_id, must be of class DRM_XE_ENGINE_CLASS_VM_BIND
1133 * and exec queue must have same vm_id. If zero, the default VM bind engine
1134 * is used.
1135 */
1136 __u32 exec_queue_id;
1137
1138 /** @pad: MBZ */
1139 __u32 pad;
1140
1141 /** @num_binds: number of binds in this IOCTL */
1142 __u32 num_binds;
1143
1144 union {
1145 /** @bind: used if num_binds == 1 */
1146 struct drm_xe_vm_bind_op bind;
1147
1148 /**
1149 * @vector_of_binds: userptr to array of struct
1150 * drm_xe_vm_bind_op if num_binds > 1
1151 */
1152 __u64 vector_of_binds;
1153 };
1154
1155 /** @pad2: MBZ */
1156 __u32 pad2;
1157
1158 /** @num_syncs: amount of syncs to wait on */
1159 __u32 num_syncs;
1160
1161 /** @syncs: pointer to struct drm_xe_sync array */
1162 __u64 syncs;
1163
1164 /** @reserved: Reserved */
1165 __u64 reserved[2];
1166 };
1167
1168 /**
1169 * struct drm_xe_exec_queue_create - Input of &DRM_IOCTL_XE_EXEC_QUEUE_CREATE
1170 *
1171 * This ioctl supports setting the following properties via the
1172 * %DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY extension, which uses the
1173 * generic @drm_xe_ext_set_property struct:
1174 *
1175 * - %DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY - set the queue priority.
1176 * CAP_SYS_NICE is required to set a value above normal.
1177 * - %DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE - set the queue timeslice
1178 * duration in microseconds.
1179 * - %DRM_XE_EXEC_QUEUE_SET_PROPERTY_PXP_TYPE - set the type of PXP session
1180 * this queue will be used with. Valid values are listed in enum
1181 * drm_xe_pxp_session_type. %DRM_XE_PXP_TYPE_NONE is the default behavior, so
1182 * there is no need to explicitly set that. When a queue of type
1183 * %DRM_XE_PXP_TYPE_HWDRM is created, the PXP default HWDRM session
1184 * (%XE_PXP_HWDRM_DEFAULT_SESSION) will be started, if isn't already running.
1185 * Given that going into a power-saving state kills PXP HWDRM sessions,
1186 * runtime PM will be blocked while queues of this type are alive.
1187 * All PXP queues will be killed if a PXP invalidation event occurs.
1188 *
1189 * The example below shows how to use @drm_xe_exec_queue_create to create
1190 * a simple exec_queue (no parallel submission) of class
1191 * &DRM_XE_ENGINE_CLASS_RENDER.
1192 *
1193 * .. code-block:: C
1194 *
1195 * struct drm_xe_engine_class_instance instance = {
1196 * .engine_class = DRM_XE_ENGINE_CLASS_RENDER,
1197 * };
1198 * struct drm_xe_exec_queue_create exec_queue_create = {
1199 * .extensions = 0,
1200 * .vm_id = vm,
1201 * .num_bb_per_exec = 1,
1202 * .num_eng_per_bb = 1,
1203 * .instances = to_user_pointer(&instance),
1204 * };
1205 * ioctl(fd, DRM_IOCTL_XE_EXEC_QUEUE_CREATE, &exec_queue_create);
1206 *
1207 */
1208 struct drm_xe_exec_queue_create {
1209 #define DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY 0
1210 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY 0
1211 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE 1
1212 #define DRM_XE_EXEC_QUEUE_SET_PROPERTY_PXP_TYPE 2
1213 /** @extensions: Pointer to the first extension struct, if any */
1214 __u64 extensions;
1215
1216 /** @width: submission width (number BB per exec) for this exec queue */
1217 __u16 width;
1218
1219 /** @num_placements: number of valid placements for this exec queue */
1220 __u16 num_placements;
1221
1222 /** @vm_id: VM to use for this exec queue */
1223 __u32 vm_id;
1224
1225 /** @flags: MBZ */
1226 __u32 flags;
1227
1228 /** @exec_queue_id: Returned exec queue ID */
1229 __u32 exec_queue_id;
1230
1231 /**
1232 * @instances: user pointer to a 2-d array of struct
1233 * drm_xe_engine_class_instance
1234 *
1235 * length = width (i) * num_placements (j)
1236 * index = j + i * width
1237 */
1238 __u64 instances;
1239
1240 /** @reserved: Reserved */
1241 __u64 reserved[2];
1242 };
1243
1244 /**
1245 * struct drm_xe_exec_queue_destroy - Input of &DRM_IOCTL_XE_EXEC_QUEUE_DESTROY
1246 */
1247 struct drm_xe_exec_queue_destroy {
1248 /** @exec_queue_id: Exec queue ID */
1249 __u32 exec_queue_id;
1250
1251 /** @pad: MBZ */
1252 __u32 pad;
1253
1254 /** @reserved: Reserved */
1255 __u64 reserved[2];
1256 };
1257
1258 /**
1259 * struct drm_xe_exec_queue_get_property - Input of &DRM_IOCTL_XE_EXEC_QUEUE_GET_PROPERTY
1260 *
1261 * The @property can be:
1262 * - %DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN
1263 */
1264 struct drm_xe_exec_queue_get_property {
1265 /** @extensions: Pointer to the first extension struct, if any */
1266 __u64 extensions;
1267
1268 /** @exec_queue_id: Exec queue ID */
1269 __u32 exec_queue_id;
1270
1271 #define DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN 0
1272 /** @property: property to get */
1273 __u32 property;
1274
1275 /** @value: property value */
1276 __u64 value;
1277
1278 /** @reserved: Reserved */
1279 __u64 reserved[2];
1280 };
1281
1282 /**
1283 * struct drm_xe_sync - sync object
1284 *
1285 * The @type can be:
1286 * - %DRM_XE_SYNC_TYPE_SYNCOBJ
1287 * - %DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ
1288 * - %DRM_XE_SYNC_TYPE_USER_FENCE
1289 *
1290 * and the @flags can be:
1291 * - %DRM_XE_SYNC_FLAG_SIGNAL
1292 *
1293 * A minimal use of @drm_xe_sync looks like this:
1294 *
1295 * .. code-block:: C
1296 *
1297 * struct drm_xe_sync sync = {
1298 * .flags = DRM_XE_SYNC_FLAG_SIGNAL,
1299 * .type = DRM_XE_SYNC_TYPE_SYNCOBJ,
1300 * };
1301 * struct drm_syncobj_create syncobj_create = { 0 };
1302 * ioctl(fd, DRM_IOCTL_SYNCOBJ_CREATE, &syncobj_create);
1303 * sync.handle = syncobj_create.handle;
1304 * ...
1305 * use of &sync in drm_xe_exec or drm_xe_vm_bind
1306 * ...
1307 * struct drm_syncobj_wait wait = {
1308 * .handles = &sync.handle,
1309 * .timeout_nsec = INT64_MAX,
1310 * .count_handles = 1,
1311 * .flags = 0,
1312 * .first_signaled = 0,
1313 * .pad = 0,
1314 * };
1315 * ioctl(fd, DRM_IOCTL_SYNCOBJ_WAIT, &wait);
1316 */
1317 struct drm_xe_sync {
1318 /** @extensions: Pointer to the first extension struct, if any */
1319 __u64 extensions;
1320
1321 #define DRM_XE_SYNC_TYPE_SYNCOBJ 0x0
1322 #define DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ 0x1
1323 #define DRM_XE_SYNC_TYPE_USER_FENCE 0x2
1324 /** @type: Type of the this sync object */
1325 __u32 type;
1326
1327 #define DRM_XE_SYNC_FLAG_SIGNAL (1 << 0)
1328 /** @flags: Sync Flags */
1329 __u32 flags;
1330
1331 union {
1332 /** @handle: Handle for the object */
1333 __u32 handle;
1334
1335 /**
1336 * @addr: Address of user fence. When sync is passed in via exec
1337 * IOCTL this is a GPU address in the VM. When sync passed in via
1338 * VM bind IOCTL this is a user pointer. In either case, it is
1339 * the users responsibility that this address is present and
1340 * mapped when the user fence is signalled. Must be qword
1341 * aligned.
1342 */
1343 __u64 addr;
1344 };
1345
1346 /**
1347 * @timeline_value: Input for the timeline sync object. Needs to be
1348 * different than 0 when used with %DRM_XE_SYNC_FLAG_TIMELINE_SYNCOBJ.
1349 */
1350 __u64 timeline_value;
1351
1352 /** @reserved: Reserved */
1353 __u64 reserved[2];
1354 };
1355
1356 /**
1357 * struct drm_xe_exec - Input of &DRM_IOCTL_XE_EXEC
1358 *
1359 * This is an example to use @drm_xe_exec for execution of the object
1360 * at BIND_ADDRESS (see example in @drm_xe_vm_bind) by an exec_queue
1361 * (see example in @drm_xe_exec_queue_create). It can be synchronized
1362 * by using the example provided for @drm_xe_sync.
1363 *
1364 * .. code-block:: C
1365 *
1366 * struct drm_xe_exec exec = {
1367 * .exec_queue_id = exec_queue,
1368 * .syncs = &sync,
1369 * .num_syncs = 1,
1370 * .address = BIND_ADDRESS,
1371 * .num_batch_buffer = 1,
1372 * };
1373 * ioctl(fd, DRM_IOCTL_XE_EXEC, &exec);
1374 *
1375 */
1376 struct drm_xe_exec {
1377 /** @extensions: Pointer to the first extension struct, if any */
1378 __u64 extensions;
1379
1380 /** @exec_queue_id: Exec queue ID for the batch buffer */
1381 __u32 exec_queue_id;
1382
1383 /** @num_syncs: Amount of struct drm_xe_sync in array. */
1384 __u32 num_syncs;
1385
1386 /** @syncs: Pointer to struct drm_xe_sync array. */
1387 __u64 syncs;
1388
1389 /**
1390 * @address: address of batch buffer if num_batch_buffer == 1 or an
1391 * array of batch buffer addresses
1392 */
1393 __u64 address;
1394
1395 /**
1396 * @num_batch_buffer: number of batch buffer in this exec, must match
1397 * the width of the engine
1398 */
1399 __u16 num_batch_buffer;
1400
1401 /** @pad: MBZ */
1402 __u16 pad[3];
1403
1404 /** @reserved: Reserved */
1405 __u64 reserved[2];
1406 };
1407
1408 /**
1409 * struct drm_xe_wait_user_fence - Input of &DRM_IOCTL_XE_WAIT_USER_FENCE
1410 *
1411 * Wait on user fence, XE will wake-up on every HW engine interrupt in the
1412 * instances list and check if user fence is complete::
1413 *
1414 * (*addr & MASK) OP (VALUE & MASK)
1415 *
1416 * Returns to user on user fence completion or timeout.
1417 *
1418 * The @op can be:
1419 * - %DRM_XE_UFENCE_WAIT_OP_EQ
1420 * - %DRM_XE_UFENCE_WAIT_OP_NEQ
1421 * - %DRM_XE_UFENCE_WAIT_OP_GT
1422 * - %DRM_XE_UFENCE_WAIT_OP_GTE
1423 * - %DRM_XE_UFENCE_WAIT_OP_LT
1424 * - %DRM_XE_UFENCE_WAIT_OP_LTE
1425 *
1426 * and the @flags can be:
1427 * - %DRM_XE_UFENCE_WAIT_FLAG_ABSTIME
1428 * - %DRM_XE_UFENCE_WAIT_FLAG_SOFT_OP
1429 *
1430 * The @mask values can be for example:
1431 * - 0xffu for u8
1432 * - 0xffffu for u16
1433 * - 0xffffffffu for u32
1434 * - 0xffffffffffffffffu for u64
1435 */
1436 struct drm_xe_wait_user_fence {
1437 /** @extensions: Pointer to the first extension struct, if any */
1438 __u64 extensions;
1439
1440 /**
1441 * @addr: user pointer address to wait on, must qword aligned
1442 */
1443 __u64 addr;
1444
1445 #define DRM_XE_UFENCE_WAIT_OP_EQ 0x0
1446 #define DRM_XE_UFENCE_WAIT_OP_NEQ 0x1
1447 #define DRM_XE_UFENCE_WAIT_OP_GT 0x2
1448 #define DRM_XE_UFENCE_WAIT_OP_GTE 0x3
1449 #define DRM_XE_UFENCE_WAIT_OP_LT 0x4
1450 #define DRM_XE_UFENCE_WAIT_OP_LTE 0x5
1451 /** @op: wait operation (type of comparison) */
1452 __u16 op;
1453
1454 #define DRM_XE_UFENCE_WAIT_FLAG_ABSTIME (1 << 0)
1455 /** @flags: wait flags */
1456 __u16 flags;
1457
1458 /** @pad: MBZ */
1459 __u32 pad;
1460
1461 /** @value: compare value */
1462 __u64 value;
1463
1464 /** @mask: comparison mask */
1465 __u64 mask;
1466
1467 /**
1468 * @timeout: how long to wait before bailing, value in nanoseconds.
1469 * Without DRM_XE_UFENCE_WAIT_FLAG_ABSTIME flag set (relative timeout)
1470 * it contains timeout expressed in nanoseconds to wait (fence will
1471 * expire at now() + timeout).
1472 * When DRM_XE_UFENCE_WAIT_FLAG_ABSTIME flat is set (absolute timeout) wait
1473 * will end at timeout (uses system MONOTONIC_CLOCK).
1474 * Passing negative timeout leads to neverending wait.
1475 *
1476 * On relative timeout this value is updated with timeout left
1477 * (for restarting the call in case of signal delivery).
1478 * On absolute timeout this value stays intact (restarted call still
1479 * expire at the same point of time).
1480 */
1481 __s64 timeout;
1482
1483 /** @exec_queue_id: exec_queue_id returned from xe_exec_queue_create_ioctl */
1484 __u32 exec_queue_id;
1485
1486 /** @pad2: MBZ */
1487 __u32 pad2;
1488
1489 /** @reserved: Reserved */
1490 __u64 reserved[2];
1491 };
1492
1493 /**
1494 * enum drm_xe_observation_type - Observation stream types
1495 */
1496 enum drm_xe_observation_type {
1497 /** @DRM_XE_OBSERVATION_TYPE_OA: OA observation stream type */
1498 DRM_XE_OBSERVATION_TYPE_OA,
1499 };
1500
1501 /**
1502 * enum drm_xe_observation_op - Observation stream ops
1503 */
1504 enum drm_xe_observation_op {
1505 /** @DRM_XE_OBSERVATION_OP_STREAM_OPEN: Open an observation stream */
1506 DRM_XE_OBSERVATION_OP_STREAM_OPEN,
1507
1508 /** @DRM_XE_OBSERVATION_OP_ADD_CONFIG: Add observation stream config */
1509 DRM_XE_OBSERVATION_OP_ADD_CONFIG,
1510
1511 /** @DRM_XE_OBSERVATION_OP_REMOVE_CONFIG: Remove observation stream config */
1512 DRM_XE_OBSERVATION_OP_REMOVE_CONFIG,
1513 };
1514
1515 /**
1516 * struct drm_xe_observation_param - Input of &DRM_XE_OBSERVATION
1517 *
1518 * The observation layer enables multiplexing observation streams of
1519 * multiple types. The actual params for a particular stream operation are
1520 * supplied via the @param pointer (use __copy_from_user to get these
1521 * params).
1522 */
1523 struct drm_xe_observation_param {
1524 /** @extensions: Pointer to the first extension struct, if any */
1525 __u64 extensions;
1526 /** @observation_type: observation stream type, of enum @drm_xe_observation_type */
1527 __u64 observation_type;
1528 /** @observation_op: observation stream op, of enum @drm_xe_observation_op */
1529 __u64 observation_op;
1530 /** @param: Pointer to actual stream params */
1531 __u64 param;
1532 };
1533
1534 /**
1535 * enum drm_xe_observation_ioctls - Observation stream fd ioctl's
1536 *
1537 * Information exchanged between userspace and kernel for observation fd
1538 * ioctl's is stream type specific
1539 */
1540 enum drm_xe_observation_ioctls {
1541 /** @DRM_XE_OBSERVATION_IOCTL_ENABLE: Enable data capture for an observation stream */
1542 DRM_XE_OBSERVATION_IOCTL_ENABLE = _IO('i', 0x0),
1543
1544 /** @DRM_XE_OBSERVATION_IOCTL_DISABLE: Disable data capture for a observation stream */
1545 DRM_XE_OBSERVATION_IOCTL_DISABLE = _IO('i', 0x1),
1546
1547 /** @DRM_XE_OBSERVATION_IOCTL_CONFIG: Change observation stream configuration */
1548 DRM_XE_OBSERVATION_IOCTL_CONFIG = _IO('i', 0x2),
1549
1550 /** @DRM_XE_OBSERVATION_IOCTL_STATUS: Return observation stream status */
1551 DRM_XE_OBSERVATION_IOCTL_STATUS = _IO('i', 0x3),
1552
1553 /** @DRM_XE_OBSERVATION_IOCTL_INFO: Return observation stream info */
1554 DRM_XE_OBSERVATION_IOCTL_INFO = _IO('i', 0x4),
1555 };
1556
1557 /**
1558 * enum drm_xe_oa_unit_type - OA unit types
1559 */
1560 enum drm_xe_oa_unit_type {
1561 /**
1562 * @DRM_XE_OA_UNIT_TYPE_OAG: OAG OA unit. OAR/OAC are considered
1563 * sub-types of OAG. For OAR/OAC, use OAG.
1564 */
1565 DRM_XE_OA_UNIT_TYPE_OAG,
1566
1567 /** @DRM_XE_OA_UNIT_TYPE_OAM: OAM OA unit */
1568 DRM_XE_OA_UNIT_TYPE_OAM,
1569 };
1570
1571 /**
1572 * struct drm_xe_oa_unit - describe OA unit
1573 */
1574 struct drm_xe_oa_unit {
1575 /** @extensions: Pointer to the first extension struct, if any */
1576 __u64 extensions;
1577
1578 /** @oa_unit_id: OA unit ID */
1579 __u32 oa_unit_id;
1580
1581 /** @oa_unit_type: OA unit type of @drm_xe_oa_unit_type */
1582 __u32 oa_unit_type;
1583
1584 /** @capabilities: OA capabilities bit-mask */
1585 __u64 capabilities;
1586 #define DRM_XE_OA_CAPS_BASE (1 << 0)
1587 #define DRM_XE_OA_CAPS_SYNCS (1 << 1)
1588 #define DRM_XE_OA_CAPS_OA_BUFFER_SIZE (1 << 2)
1589 #define DRM_XE_OA_CAPS_WAIT_NUM_REPORTS (1 << 3)
1590
1591 /** @oa_timestamp_freq: OA timestamp freq */
1592 __u64 oa_timestamp_freq;
1593
1594 /** @reserved: MBZ */
1595 __u64 reserved[4];
1596
1597 /** @num_engines: number of engines in @eci array */
1598 __u64 num_engines;
1599
1600 /** @eci: engines attached to this OA unit */
1601 struct drm_xe_engine_class_instance eci[];
1602 };
1603
1604 /**
1605 * struct drm_xe_query_oa_units - describe OA units
1606 *
1607 * If a query is made with a struct drm_xe_device_query where .query
1608 * is equal to DRM_XE_DEVICE_QUERY_OA_UNITS, then the reply uses struct
1609 * drm_xe_query_oa_units in .data.
1610 *
1611 * OA unit properties for all OA units can be accessed using a code block
1612 * such as the one below:
1613 *
1614 * .. code-block:: C
1615 *
1616 * struct drm_xe_query_oa_units *qoa;
1617 * struct drm_xe_oa_unit *oau;
1618 * u8 *poau;
1619 *
1620 * // malloc qoa and issue DRM_XE_DEVICE_QUERY_OA_UNITS. Then:
1621 * poau = (u8 *)&qoa->oa_units[0];
1622 * for (int i = 0; i < qoa->num_oa_units; i++) {
1623 * oau = (struct drm_xe_oa_unit *)poau;
1624 * // Access 'struct drm_xe_oa_unit' fields here
1625 * poau += sizeof(*oau) + oau->num_engines * sizeof(oau->eci[0]);
1626 * }
1627 */
1628 struct drm_xe_query_oa_units {
1629 /** @extensions: Pointer to the first extension struct, if any */
1630 __u64 extensions;
1631 /** @num_oa_units: number of OA units returned in oau[] */
1632 __u32 num_oa_units;
1633 /** @pad: MBZ */
1634 __u32 pad;
1635 /**
1636 * @oa_units: struct @drm_xe_oa_unit array returned for this device.
1637 * Written below as a u64 array to avoid problems with nested flexible
1638 * arrays with some compilers
1639 */
1640 __u64 oa_units[];
1641 };
1642
1643 /**
1644 * enum drm_xe_oa_format_type - OA format types as specified in PRM/Bspec
1645 * 52198/60942
1646 */
1647 enum drm_xe_oa_format_type {
1648 /** @DRM_XE_OA_FMT_TYPE_OAG: OAG report format */
1649 DRM_XE_OA_FMT_TYPE_OAG,
1650 /** @DRM_XE_OA_FMT_TYPE_OAR: OAR report format */
1651 DRM_XE_OA_FMT_TYPE_OAR,
1652 /** @DRM_XE_OA_FMT_TYPE_OAM: OAM report format */
1653 DRM_XE_OA_FMT_TYPE_OAM,
1654 /** @DRM_XE_OA_FMT_TYPE_OAC: OAC report format */
1655 DRM_XE_OA_FMT_TYPE_OAC,
1656 /** @DRM_XE_OA_FMT_TYPE_OAM_MPEC: OAM SAMEDIA or OAM MPEC report format */
1657 DRM_XE_OA_FMT_TYPE_OAM_MPEC,
1658 /** @DRM_XE_OA_FMT_TYPE_PEC: PEC report format */
1659 DRM_XE_OA_FMT_TYPE_PEC,
1660 };
1661
1662 /**
1663 * enum drm_xe_oa_property_id - OA stream property id's
1664 *
1665 * Stream params are specified as a chain of @drm_xe_ext_set_property
1666 * struct's, with @property values from enum @drm_xe_oa_property_id and
1667 * @drm_xe_user_extension base.name set to @DRM_XE_OA_EXTENSION_SET_PROPERTY.
1668 * @param field in struct @drm_xe_observation_param points to the first
1669 * @drm_xe_ext_set_property struct.
1670 *
1671 * Exactly the same mechanism is also used for stream reconfiguration using the
1672 * @DRM_XE_OBSERVATION_IOCTL_CONFIG observation stream fd ioctl, though only a
1673 * subset of properties below can be specified for stream reconfiguration.
1674 */
1675 enum drm_xe_oa_property_id {
1676 #define DRM_XE_OA_EXTENSION_SET_PROPERTY 0
1677 /**
1678 * @DRM_XE_OA_PROPERTY_OA_UNIT_ID: ID of the OA unit on which to open
1679 * the OA stream, see @oa_unit_id in 'struct
1680 * drm_xe_query_oa_units'. Defaults to 0 if not provided.
1681 */
1682 DRM_XE_OA_PROPERTY_OA_UNIT_ID = 1,
1683
1684 /**
1685 * @DRM_XE_OA_PROPERTY_SAMPLE_OA: A value of 1 requests inclusion of raw
1686 * OA unit reports or stream samples in a global buffer attached to an
1687 * OA unit.
1688 */
1689 DRM_XE_OA_PROPERTY_SAMPLE_OA,
1690
1691 /**
1692 * @DRM_XE_OA_PROPERTY_OA_METRIC_SET: OA metrics defining contents of OA
1693 * reports, previously added via @DRM_XE_OBSERVATION_OP_ADD_CONFIG.
1694 */
1695 DRM_XE_OA_PROPERTY_OA_METRIC_SET,
1696
1697 /** @DRM_XE_OA_PROPERTY_OA_FORMAT: OA counter report format */
1698 DRM_XE_OA_PROPERTY_OA_FORMAT,
1699 /*
1700 * OA_FORMAT's are specified the same way as in PRM/Bspec 52198/60942,
1701 * in terms of the following quantities: a. enum @drm_xe_oa_format_type
1702 * b. Counter select c. Counter size and d. BC report. Also refer to the
1703 * oa_formats array in drivers/gpu/drm/xe/xe_oa.c.
1704 */
1705 #define DRM_XE_OA_FORMAT_MASK_FMT_TYPE (0xffu << 0)
1706 #define DRM_XE_OA_FORMAT_MASK_COUNTER_SEL (0xffu << 8)
1707 #define DRM_XE_OA_FORMAT_MASK_COUNTER_SIZE (0xffu << 16)
1708 #define DRM_XE_OA_FORMAT_MASK_BC_REPORT (0xffu << 24)
1709
1710 /**
1711 * @DRM_XE_OA_PROPERTY_OA_PERIOD_EXPONENT: Requests periodic OA unit
1712 * sampling with sampling frequency proportional to 2^(period_exponent + 1)
1713 */
1714 DRM_XE_OA_PROPERTY_OA_PERIOD_EXPONENT,
1715
1716 /**
1717 * @DRM_XE_OA_PROPERTY_OA_DISABLED: A value of 1 will open the OA
1718 * stream in a DISABLED state (see @DRM_XE_OBSERVATION_IOCTL_ENABLE).
1719 */
1720 DRM_XE_OA_PROPERTY_OA_DISABLED,
1721
1722 /**
1723 * @DRM_XE_OA_PROPERTY_EXEC_QUEUE_ID: Open the stream for a specific
1724 * @exec_queue_id. OA queries can be executed on this exec queue.
1725 */
1726 DRM_XE_OA_PROPERTY_EXEC_QUEUE_ID,
1727
1728 /**
1729 * @DRM_XE_OA_PROPERTY_OA_ENGINE_INSTANCE: Optional engine instance to
1730 * pass along with @DRM_XE_OA_PROPERTY_EXEC_QUEUE_ID or will default to 0.
1731 */
1732 DRM_XE_OA_PROPERTY_OA_ENGINE_INSTANCE,
1733
1734 /**
1735 * @DRM_XE_OA_PROPERTY_NO_PREEMPT: Allow preemption and timeslicing
1736 * to be disabled for the stream exec queue.
1737 */
1738 DRM_XE_OA_PROPERTY_NO_PREEMPT,
1739
1740 /**
1741 * @DRM_XE_OA_PROPERTY_NUM_SYNCS: Number of syncs in the sync array
1742 * specified in @DRM_XE_OA_PROPERTY_SYNCS
1743 */
1744 DRM_XE_OA_PROPERTY_NUM_SYNCS,
1745
1746 /**
1747 * @DRM_XE_OA_PROPERTY_SYNCS: Pointer to struct @drm_xe_sync array
1748 * with array size specified via @DRM_XE_OA_PROPERTY_NUM_SYNCS. OA
1749 * configuration will wait till input fences signal. Output fences
1750 * will signal after the new OA configuration takes effect. For
1751 * @DRM_XE_SYNC_TYPE_USER_FENCE, @addr is a user pointer, similar
1752 * to the VM bind case.
1753 */
1754 DRM_XE_OA_PROPERTY_SYNCS,
1755
1756 /**
1757 * @DRM_XE_OA_PROPERTY_OA_BUFFER_SIZE: Size of OA buffer to be
1758 * allocated by the driver in bytes. Supported sizes are powers of
1759 * 2 from 128 KiB to 128 MiB. When not specified, a 16 MiB OA
1760 * buffer is allocated by default.
1761 */
1762 DRM_XE_OA_PROPERTY_OA_BUFFER_SIZE,
1763
1764 /**
1765 * @DRM_XE_OA_PROPERTY_WAIT_NUM_REPORTS: Number of reports to wait
1766 * for before unblocking poll or read
1767 */
1768 DRM_XE_OA_PROPERTY_WAIT_NUM_REPORTS,
1769 };
1770
1771 /**
1772 * struct drm_xe_oa_config - OA metric configuration
1773 *
1774 * Multiple OA configs can be added using @DRM_XE_OBSERVATION_OP_ADD_CONFIG. A
1775 * particular config can be specified when opening an OA stream using
1776 * @DRM_XE_OA_PROPERTY_OA_METRIC_SET property.
1777 */
1778 struct drm_xe_oa_config {
1779 /** @extensions: Pointer to the first extension struct, if any */
1780 __u64 extensions;
1781
1782 /** @uuid: String formatted like "%\08x-%\04x-%\04x-%\04x-%\012x" */
1783 char uuid[36];
1784
1785 /** @n_regs: Number of regs in @regs_ptr */
1786 __u32 n_regs;
1787
1788 /**
1789 * @regs_ptr: Pointer to (register address, value) pairs for OA config
1790 * registers. Expected length of buffer is: (2 * sizeof(u32) * @n_regs).
1791 */
1792 __u64 regs_ptr;
1793 };
1794
1795 /**
1796 * struct drm_xe_oa_stream_status - OA stream status returned from
1797 * @DRM_XE_OBSERVATION_IOCTL_STATUS observation stream fd ioctl. Userspace can
1798 * call the ioctl to query stream status in response to EIO errno from
1799 * observation fd read().
1800 */
1801 struct drm_xe_oa_stream_status {
1802 /** @extensions: Pointer to the first extension struct, if any */
1803 __u64 extensions;
1804
1805 /** @oa_status: OA stream status (see Bspec 46717/61226) */
1806 __u64 oa_status;
1807 #define DRM_XE_OASTATUS_MMIO_TRG_Q_FULL (1 << 3)
1808 #define DRM_XE_OASTATUS_COUNTER_OVERFLOW (1 << 2)
1809 #define DRM_XE_OASTATUS_BUFFER_OVERFLOW (1 << 1)
1810 #define DRM_XE_OASTATUS_REPORT_LOST (1 << 0)
1811
1812 /** @reserved: reserved for future use */
1813 __u64 reserved[3];
1814 };
1815
1816 /**
1817 * struct drm_xe_oa_stream_info - OA stream info returned from
1818 * @DRM_XE_OBSERVATION_IOCTL_INFO observation stream fd ioctl
1819 */
1820 struct drm_xe_oa_stream_info {
1821 /** @extensions: Pointer to the first extension struct, if any */
1822 __u64 extensions;
1823
1824 /** @oa_buf_size: OA buffer size */
1825 __u64 oa_buf_size;
1826
1827 /** @reserved: reserved for future use */
1828 __u64 reserved[3];
1829 };
1830
1831 /**
1832 * enum drm_xe_pxp_session_type - Supported PXP session types.
1833 *
1834 * We currently only support HWDRM sessions, which are used for protected
1835 * content that ends up being displayed, but the HW supports multiple types, so
1836 * we might extend support in the future.
1837 */
1838 enum drm_xe_pxp_session_type {
1839 /** @DRM_XE_PXP_TYPE_NONE: PXP not used */
1840 DRM_XE_PXP_TYPE_NONE = 0,
1841 /**
1842 * @DRM_XE_PXP_TYPE_HWDRM: HWDRM sessions are used for content that ends
1843 * up on the display.
1844 */
1845 DRM_XE_PXP_TYPE_HWDRM = 1,
1846 };
1847
1848 /* ID of the protected content session managed by Xe when PXP is active */
1849 #define DRM_XE_PXP_HWDRM_DEFAULT_SESSION 0xf
1850
1851 #if defined(__cplusplus)
1852 }
1853 #endif
1854
1855 #endif /* _UAPI_XE_DRM_H_ */