xref: /linux/drivers/gpu/drm/xe/xe_query.c (revision 220994d61cebfc04f071d69049127657c7e8191b)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #include "xe_query.h"
7 
8 #include <linux/nospec.h>
9 #include <linux/sched/clock.h>
10 
11 #include <drm/ttm/ttm_placement.h>
12 #include <generated/xe_wa_oob.h>
13 #include <uapi/drm/xe_drm.h>
14 
15 #include "regs/xe_engine_regs.h"
16 #include "regs/xe_gt_regs.h"
17 #include "xe_bo.h"
18 #include "xe_device.h"
19 #include "xe_eu_stall.h"
20 #include "xe_exec_queue.h"
21 #include "xe_force_wake.h"
22 #include "xe_ggtt.h"
23 #include "xe_gt.h"
24 #include "xe_guc_hwconfig.h"
25 #include "xe_macros.h"
26 #include "xe_mmio.h"
27 #include "xe_oa.h"
28 #include "xe_pxp.h"
29 #include "xe_ttm_vram_mgr.h"
30 #include "xe_wa.h"
31 
32 static const u16 xe_to_user_engine_class[] = {
33 	[XE_ENGINE_CLASS_RENDER] = DRM_XE_ENGINE_CLASS_RENDER,
34 	[XE_ENGINE_CLASS_COPY] = DRM_XE_ENGINE_CLASS_COPY,
35 	[XE_ENGINE_CLASS_VIDEO_DECODE] = DRM_XE_ENGINE_CLASS_VIDEO_DECODE,
36 	[XE_ENGINE_CLASS_VIDEO_ENHANCE] = DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE,
37 	[XE_ENGINE_CLASS_COMPUTE] = DRM_XE_ENGINE_CLASS_COMPUTE,
38 };
39 
40 static const enum xe_engine_class user_to_xe_engine_class[] = {
41 	[DRM_XE_ENGINE_CLASS_RENDER] = XE_ENGINE_CLASS_RENDER,
42 	[DRM_XE_ENGINE_CLASS_COPY] = XE_ENGINE_CLASS_COPY,
43 	[DRM_XE_ENGINE_CLASS_VIDEO_DECODE] = XE_ENGINE_CLASS_VIDEO_DECODE,
44 	[DRM_XE_ENGINE_CLASS_VIDEO_ENHANCE] = XE_ENGINE_CLASS_VIDEO_ENHANCE,
45 	[DRM_XE_ENGINE_CLASS_COMPUTE] = XE_ENGINE_CLASS_COMPUTE,
46 };
47 
calc_hw_engine_info_size(struct xe_device * xe)48 static size_t calc_hw_engine_info_size(struct xe_device *xe)
49 {
50 	struct xe_hw_engine *hwe;
51 	enum xe_hw_engine_id id;
52 	struct xe_gt *gt;
53 	u8 gt_id;
54 	int i = 0;
55 
56 	for_each_gt(gt, xe, gt_id)
57 		for_each_hw_engine(hwe, gt, id) {
58 			if (xe_hw_engine_is_reserved(hwe))
59 				continue;
60 			i++;
61 		}
62 
63 	return sizeof(struct drm_xe_query_engines) +
64 		i * sizeof(struct drm_xe_engine);
65 }
66 
67 typedef u64 (*__ktime_func_t)(void);
__clock_id_to_func(clockid_t clk_id)68 static __ktime_func_t __clock_id_to_func(clockid_t clk_id)
69 {
70 	/*
71 	 * Use logic same as the perf subsystem to allow user to select the
72 	 * reference clock id to be used for timestamps.
73 	 */
74 	switch (clk_id) {
75 	case CLOCK_MONOTONIC:
76 		return &ktime_get_ns;
77 	case CLOCK_MONOTONIC_RAW:
78 		return &ktime_get_raw_ns;
79 	case CLOCK_REALTIME:
80 		return &ktime_get_real_ns;
81 	case CLOCK_BOOTTIME:
82 		return &ktime_get_boottime_ns;
83 	case CLOCK_TAI:
84 		return &ktime_get_clocktai_ns;
85 	default:
86 		return NULL;
87 	}
88 }
89 
90 static void
hwe_read_timestamp(struct xe_hw_engine * hwe,u64 * engine_ts,u64 * cpu_ts,u64 * cpu_delta,__ktime_func_t cpu_clock)91 hwe_read_timestamp(struct xe_hw_engine *hwe, u64 *engine_ts, u64 *cpu_ts,
92 		   u64 *cpu_delta, __ktime_func_t cpu_clock)
93 {
94 	struct xe_mmio *mmio = &hwe->gt->mmio;
95 	u32 upper, lower, old_upper, loop = 0;
96 	struct xe_reg upper_reg = RING_TIMESTAMP_UDW(hwe->mmio_base),
97 		      lower_reg = RING_TIMESTAMP(hwe->mmio_base);
98 
99 	upper = xe_mmio_read32(mmio, upper_reg);
100 	do {
101 		*cpu_delta = local_clock();
102 		*cpu_ts = cpu_clock();
103 		lower = xe_mmio_read32(mmio, lower_reg);
104 		*cpu_delta = local_clock() - *cpu_delta;
105 		old_upper = upper;
106 		upper = xe_mmio_read32(mmio, upper_reg);
107 	} while (upper != old_upper && loop++ < 2);
108 
109 	*engine_ts = (u64)upper << 32 | lower;
110 }
111 
112 static int
query_engine_cycles(struct xe_device * xe,struct drm_xe_device_query * query)113 query_engine_cycles(struct xe_device *xe,
114 		    struct drm_xe_device_query *query)
115 {
116 	struct drm_xe_query_engine_cycles __user *query_ptr;
117 	struct drm_xe_engine_class_instance *eci;
118 	struct drm_xe_query_engine_cycles resp;
119 	size_t size = sizeof(resp);
120 	__ktime_func_t cpu_clock;
121 	struct xe_hw_engine *hwe;
122 	struct xe_gt *gt;
123 	unsigned int fw_ref;
124 
125 	if (IS_SRIOV_VF(xe))
126 		return -EOPNOTSUPP;
127 
128 	if (query->size == 0) {
129 		query->size = size;
130 		return 0;
131 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
132 		return -EINVAL;
133 	}
134 
135 	query_ptr = u64_to_user_ptr(query->data);
136 	if (copy_from_user(&resp, query_ptr, size))
137 		return -EFAULT;
138 
139 	cpu_clock = __clock_id_to_func(resp.clockid);
140 	if (!cpu_clock)
141 		return -EINVAL;
142 
143 	eci = &resp.eci;
144 	if (eci->gt_id >= xe->info.max_gt_per_tile)
145 		return -EINVAL;
146 
147 	gt = xe_device_get_gt(xe, eci->gt_id);
148 	if (!gt)
149 		return -EINVAL;
150 
151 	if (eci->engine_class >= ARRAY_SIZE(user_to_xe_engine_class))
152 		return -EINVAL;
153 
154 	hwe = xe_gt_hw_engine(gt, user_to_xe_engine_class[eci->engine_class],
155 			      eci->engine_instance, true);
156 	if (!hwe)
157 		return -EINVAL;
158 
159 	fw_ref = xe_force_wake_get(gt_to_fw(gt), XE_FORCEWAKE_ALL);
160 	if (!xe_force_wake_ref_has_domain(fw_ref, XE_FORCEWAKE_ALL))  {
161 		xe_force_wake_put(gt_to_fw(gt), fw_ref);
162 		return -EIO;
163 	}
164 
165 	hwe_read_timestamp(hwe, &resp.engine_cycles, &resp.cpu_timestamp,
166 			   &resp.cpu_delta, cpu_clock);
167 
168 	xe_force_wake_put(gt_to_fw(gt), fw_ref);
169 
170 	if (GRAPHICS_VER(xe) >= 20)
171 		resp.width = 64;
172 	else
173 		resp.width = 36;
174 
175 	/* Only write to the output fields of user query */
176 	if (put_user(resp.cpu_timestamp, &query_ptr->cpu_timestamp) ||
177 	    put_user(resp.cpu_delta, &query_ptr->cpu_delta) ||
178 	    put_user(resp.engine_cycles, &query_ptr->engine_cycles) ||
179 	    put_user(resp.width, &query_ptr->width))
180 		return -EFAULT;
181 
182 	return 0;
183 }
184 
query_engines(struct xe_device * xe,struct drm_xe_device_query * query)185 static int query_engines(struct xe_device *xe,
186 			 struct drm_xe_device_query *query)
187 {
188 	size_t size = calc_hw_engine_info_size(xe);
189 	struct drm_xe_query_engines __user *query_ptr =
190 		u64_to_user_ptr(query->data);
191 	struct drm_xe_query_engines *engines;
192 	struct xe_hw_engine *hwe;
193 	enum xe_hw_engine_id id;
194 	struct xe_gt *gt;
195 	u8 gt_id;
196 	int i = 0;
197 
198 	if (query->size == 0) {
199 		query->size = size;
200 		return 0;
201 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
202 		return -EINVAL;
203 	}
204 
205 	engines = kzalloc(size, GFP_KERNEL);
206 	if (!engines)
207 		return -ENOMEM;
208 
209 	for_each_gt(gt, xe, gt_id)
210 		for_each_hw_engine(hwe, gt, id) {
211 			if (xe_hw_engine_is_reserved(hwe))
212 				continue;
213 
214 			engines->engines[i].instance.engine_class =
215 				xe_to_user_engine_class[hwe->class];
216 			engines->engines[i].instance.engine_instance =
217 				hwe->logical_instance;
218 			engines->engines[i].instance.gt_id = gt->info.id;
219 
220 			i++;
221 		}
222 
223 	engines->num_engines = i;
224 
225 	if (copy_to_user(query_ptr, engines, size)) {
226 		kfree(engines);
227 		return -EFAULT;
228 	}
229 	kfree(engines);
230 
231 	return 0;
232 }
233 
calc_mem_regions_size(struct xe_device * xe)234 static size_t calc_mem_regions_size(struct xe_device *xe)
235 {
236 	u32 num_managers = 1;
237 	int i;
238 
239 	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i)
240 		if (ttm_manager_type(&xe->ttm, i))
241 			num_managers++;
242 
243 	return offsetof(struct drm_xe_query_mem_regions, mem_regions[num_managers]);
244 }
245 
query_mem_regions(struct xe_device * xe,struct drm_xe_device_query * query)246 static int query_mem_regions(struct xe_device *xe,
247 			    struct drm_xe_device_query *query)
248 {
249 	size_t size = calc_mem_regions_size(xe);
250 	struct drm_xe_query_mem_regions *mem_regions;
251 	struct drm_xe_query_mem_regions __user *query_ptr =
252 		u64_to_user_ptr(query->data);
253 	struct ttm_resource_manager *man;
254 	int ret, i;
255 
256 	if (query->size == 0) {
257 		query->size = size;
258 		return 0;
259 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
260 		return -EINVAL;
261 	}
262 
263 	mem_regions = kzalloc(size, GFP_KERNEL);
264 	if (XE_IOCTL_DBG(xe, !mem_regions))
265 		return -ENOMEM;
266 
267 	man = ttm_manager_type(&xe->ttm, XE_PL_TT);
268 	mem_regions->mem_regions[0].mem_class = DRM_XE_MEM_REGION_CLASS_SYSMEM;
269 	/*
270 	 * The instance needs to be a unique number that represents the index
271 	 * in the placement mask used at xe_gem_create_ioctl() for the
272 	 * xe_bo_create() placement.
273 	 */
274 	mem_regions->mem_regions[0].instance = 0;
275 	mem_regions->mem_regions[0].min_page_size = PAGE_SIZE;
276 	mem_regions->mem_regions[0].total_size = man->size << PAGE_SHIFT;
277 	if (perfmon_capable())
278 		mem_regions->mem_regions[0].used = ttm_resource_manager_usage(man);
279 	mem_regions->num_mem_regions = 1;
280 
281 	for (i = XE_PL_VRAM0; i <= XE_PL_VRAM1; ++i) {
282 		man = ttm_manager_type(&xe->ttm, i);
283 		if (man) {
284 			mem_regions->mem_regions[mem_regions->num_mem_regions].mem_class =
285 				DRM_XE_MEM_REGION_CLASS_VRAM;
286 			mem_regions->mem_regions[mem_regions->num_mem_regions].instance =
287 				mem_regions->num_mem_regions;
288 			mem_regions->mem_regions[mem_regions->num_mem_regions].min_page_size =
289 				xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K ?
290 				SZ_64K : PAGE_SIZE;
291 			mem_regions->mem_regions[mem_regions->num_mem_regions].total_size =
292 				man->size;
293 
294 			if (perfmon_capable()) {
295 				xe_ttm_vram_get_used(man,
296 					&mem_regions->mem_regions
297 					[mem_regions->num_mem_regions].used,
298 					&mem_regions->mem_regions
299 					[mem_regions->num_mem_regions].cpu_visible_used);
300 			}
301 
302 			mem_regions->mem_regions[mem_regions->num_mem_regions].cpu_visible_size =
303 				xe_ttm_vram_get_cpu_visible_size(man);
304 			mem_regions->num_mem_regions++;
305 		}
306 	}
307 
308 	if (!copy_to_user(query_ptr, mem_regions, size))
309 		ret = 0;
310 	else
311 		ret = -ENOSPC;
312 
313 	kfree(mem_regions);
314 	return ret;
315 }
316 
query_config(struct xe_device * xe,struct drm_xe_device_query * query)317 static int query_config(struct xe_device *xe, struct drm_xe_device_query *query)
318 {
319 	const u32 num_params = DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY + 1;
320 	size_t size =
321 		sizeof(struct drm_xe_query_config) + num_params * sizeof(u64);
322 	struct drm_xe_query_config __user *query_ptr =
323 		u64_to_user_ptr(query->data);
324 	struct drm_xe_query_config *config;
325 
326 	if (query->size == 0) {
327 		query->size = size;
328 		return 0;
329 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
330 		return -EINVAL;
331 	}
332 
333 	config = kzalloc(size, GFP_KERNEL);
334 	if (!config)
335 		return -ENOMEM;
336 
337 	config->num_params = num_params;
338 	config->info[DRM_XE_QUERY_CONFIG_REV_AND_DEVICE_ID] =
339 		xe->info.devid | (xe->info.revid << 16);
340 	if (xe_device_get_root_tile(xe)->mem.vram.usable_size)
341 		config->info[DRM_XE_QUERY_CONFIG_FLAGS] |=
342 			DRM_XE_QUERY_CONFIG_FLAG_HAS_VRAM;
343 	if (xe->info.has_usm && IS_ENABLED(CONFIG_DRM_XE_GPUSVM))
344 		config->info[DRM_XE_QUERY_CONFIG_FLAGS] |=
345 			DRM_XE_QUERY_CONFIG_FLAG_HAS_CPU_ADDR_MIRROR;
346 	config->info[DRM_XE_QUERY_CONFIG_FLAGS] |=
347 			DRM_XE_QUERY_CONFIG_FLAG_HAS_LOW_LATENCY;
348 	config->info[DRM_XE_QUERY_CONFIG_MIN_ALIGNMENT] =
349 		xe->info.vram_flags & XE_VRAM_FLAGS_NEED64K ? SZ_64K : SZ_4K;
350 	config->info[DRM_XE_QUERY_CONFIG_VA_BITS] = xe->info.va_bits;
351 	config->info[DRM_XE_QUERY_CONFIG_MAX_EXEC_QUEUE_PRIORITY] =
352 		xe_exec_queue_device_get_max_priority(xe);
353 
354 	if (copy_to_user(query_ptr, config, size)) {
355 		kfree(config);
356 		return -EFAULT;
357 	}
358 	kfree(config);
359 
360 	return 0;
361 }
362 
query_gt_list(struct xe_device * xe,struct drm_xe_device_query * query)363 static int query_gt_list(struct xe_device *xe, struct drm_xe_device_query *query)
364 {
365 	struct xe_gt *gt;
366 	size_t size = sizeof(struct drm_xe_query_gt_list) +
367 		xe->info.gt_count * sizeof(struct drm_xe_gt);
368 	struct drm_xe_query_gt_list __user *query_ptr =
369 		u64_to_user_ptr(query->data);
370 	struct drm_xe_query_gt_list *gt_list;
371 	int iter = 0;
372 	u8 id;
373 
374 	if (query->size == 0) {
375 		query->size = size;
376 		return 0;
377 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
378 		return -EINVAL;
379 	}
380 
381 	gt_list = kzalloc(size, GFP_KERNEL);
382 	if (!gt_list)
383 		return -ENOMEM;
384 
385 	gt_list->num_gt = xe->info.gt_count;
386 
387 	for_each_gt(gt, xe, id) {
388 		if (xe_gt_is_media_type(gt))
389 			gt_list->gt_list[iter].type = DRM_XE_QUERY_GT_TYPE_MEDIA;
390 		else
391 			gt_list->gt_list[iter].type = DRM_XE_QUERY_GT_TYPE_MAIN;
392 		gt_list->gt_list[iter].tile_id = gt_to_tile(gt)->id;
393 		gt_list->gt_list[iter].gt_id = gt->info.id;
394 		gt_list->gt_list[iter].reference_clock = gt->info.reference_clock;
395 		/*
396 		 * The mem_regions indexes in the mask below need to
397 		 * directly identify the struct
398 		 * drm_xe_query_mem_regions' instance constructed at
399 		 * query_mem_regions()
400 		 *
401 		 * For our current platforms:
402 		 * Bit 0 -> System Memory
403 		 * Bit 1 -> VRAM0 on Tile0
404 		 * Bit 2 -> VRAM1 on Tile1
405 		 * However the uAPI is generic and it's userspace's
406 		 * responsibility to check the mem_class, without any
407 		 * assumption.
408 		 */
409 		if (!IS_DGFX(xe))
410 			gt_list->gt_list[iter].near_mem_regions = 0x1;
411 		else
412 			gt_list->gt_list[iter].near_mem_regions =
413 				BIT(gt_to_tile(gt)->id) << 1;
414 		gt_list->gt_list[iter].far_mem_regions = xe->info.mem_region_mask ^
415 			gt_list->gt_list[iter].near_mem_regions;
416 
417 		gt_list->gt_list[iter].ip_ver_major =
418 			REG_FIELD_GET(GMD_ID_ARCH_MASK, gt->info.gmdid);
419 		gt_list->gt_list[iter].ip_ver_minor =
420 			REG_FIELD_GET(GMD_ID_RELEASE_MASK, gt->info.gmdid);
421 		gt_list->gt_list[iter].ip_ver_rev =
422 			REG_FIELD_GET(GMD_ID_REVID, gt->info.gmdid);
423 
424 		iter++;
425 	}
426 
427 	if (copy_to_user(query_ptr, gt_list, size)) {
428 		kfree(gt_list);
429 		return -EFAULT;
430 	}
431 	kfree(gt_list);
432 
433 	return 0;
434 }
435 
query_hwconfig(struct xe_device * xe,struct drm_xe_device_query * query)436 static int query_hwconfig(struct xe_device *xe,
437 			  struct drm_xe_device_query *query)
438 {
439 	struct xe_gt *gt = xe_root_mmio_gt(xe);
440 	size_t size = xe_guc_hwconfig_size(&gt->uc.guc);
441 	void __user *query_ptr = u64_to_user_ptr(query->data);
442 	void *hwconfig;
443 
444 	if (query->size == 0) {
445 		query->size = size;
446 		return 0;
447 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
448 		return -EINVAL;
449 	}
450 
451 	hwconfig = kzalloc(size, GFP_KERNEL);
452 	if (!hwconfig)
453 		return -ENOMEM;
454 
455 	xe_guc_hwconfig_copy(&gt->uc.guc, hwconfig);
456 
457 	if (copy_to_user(query_ptr, hwconfig, size)) {
458 		kfree(hwconfig);
459 		return -EFAULT;
460 	}
461 	kfree(hwconfig);
462 
463 	return 0;
464 }
465 
calc_topo_query_size(struct xe_device * xe)466 static size_t calc_topo_query_size(struct xe_device *xe)
467 {
468 	struct xe_gt *gt;
469 	size_t query_size = 0;
470 	int id;
471 
472 	for_each_gt(gt, xe, id) {
473 		query_size += 3 * sizeof(struct drm_xe_query_topology_mask) +
474 			sizeof_field(struct xe_gt, fuse_topo.g_dss_mask) +
475 			sizeof_field(struct xe_gt, fuse_topo.c_dss_mask) +
476 			sizeof_field(struct xe_gt, fuse_topo.eu_mask_per_dss);
477 
478 		/* L3bank mask may not be available for some GTs */
479 		if (!XE_WA(gt, no_media_l3))
480 			query_size += sizeof(struct drm_xe_query_topology_mask) +
481 				sizeof_field(struct xe_gt, fuse_topo.l3_bank_mask);
482 	}
483 
484 	return query_size;
485 }
486 
copy_mask(void __user ** ptr,struct drm_xe_query_topology_mask * topo,void * mask,size_t mask_size)487 static int copy_mask(void __user **ptr,
488 		     struct drm_xe_query_topology_mask *topo,
489 		     void *mask, size_t mask_size)
490 {
491 	topo->num_bytes = mask_size;
492 
493 	if (copy_to_user(*ptr, topo, sizeof(*topo)))
494 		return -EFAULT;
495 	*ptr += sizeof(topo);
496 
497 	if (copy_to_user(*ptr, mask, mask_size))
498 		return -EFAULT;
499 	*ptr += mask_size;
500 
501 	return 0;
502 }
503 
query_gt_topology(struct xe_device * xe,struct drm_xe_device_query * query)504 static int query_gt_topology(struct xe_device *xe,
505 			     struct drm_xe_device_query *query)
506 {
507 	void __user *query_ptr = u64_to_user_ptr(query->data);
508 	size_t size = calc_topo_query_size(xe);
509 	struct drm_xe_query_topology_mask topo;
510 	struct xe_gt *gt;
511 	int id;
512 
513 	if (query->size == 0) {
514 		query->size = size;
515 		return 0;
516 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
517 		return -EINVAL;
518 	}
519 
520 	for_each_gt(gt, xe, id) {
521 		int err;
522 
523 		topo.gt_id = id;
524 
525 		topo.type = DRM_XE_TOPO_DSS_GEOMETRY;
526 		err = copy_mask(&query_ptr, &topo, gt->fuse_topo.g_dss_mask,
527 				sizeof(gt->fuse_topo.g_dss_mask));
528 		if (err)
529 			return err;
530 
531 		topo.type = DRM_XE_TOPO_DSS_COMPUTE;
532 		err = copy_mask(&query_ptr, &topo, gt->fuse_topo.c_dss_mask,
533 				sizeof(gt->fuse_topo.c_dss_mask));
534 		if (err)
535 			return err;
536 
537 		/*
538 		 * If the kernel doesn't have a way to obtain a correct L3bank
539 		 * mask, then it's better to omit L3 from the query rather than
540 		 * reporting bogus or zeroed information to userspace.
541 		 */
542 		if (!XE_WA(gt, no_media_l3)) {
543 			topo.type = DRM_XE_TOPO_L3_BANK;
544 			err = copy_mask(&query_ptr, &topo, gt->fuse_topo.l3_bank_mask,
545 					sizeof(gt->fuse_topo.l3_bank_mask));
546 			if (err)
547 				return err;
548 		}
549 
550 		topo.type = gt->fuse_topo.eu_type == XE_GT_EU_TYPE_SIMD16 ?
551 			DRM_XE_TOPO_SIMD16_EU_PER_DSS :
552 			DRM_XE_TOPO_EU_PER_DSS;
553 		err = copy_mask(&query_ptr, &topo,
554 				gt->fuse_topo.eu_mask_per_dss,
555 				sizeof(gt->fuse_topo.eu_mask_per_dss));
556 		if (err)
557 			return err;
558 	}
559 
560 	return 0;
561 }
562 
563 static int
query_uc_fw_version(struct xe_device * xe,struct drm_xe_device_query * query)564 query_uc_fw_version(struct xe_device *xe, struct drm_xe_device_query *query)
565 {
566 	struct drm_xe_query_uc_fw_version __user *query_ptr = u64_to_user_ptr(query->data);
567 	size_t size = sizeof(struct drm_xe_query_uc_fw_version);
568 	struct drm_xe_query_uc_fw_version resp;
569 	struct xe_uc_fw_version *version = NULL;
570 
571 	if (query->size == 0) {
572 		query->size = size;
573 		return 0;
574 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
575 		return -EINVAL;
576 	}
577 
578 	if (copy_from_user(&resp, query_ptr, size))
579 		return -EFAULT;
580 
581 	if (XE_IOCTL_DBG(xe, resp.pad || resp.pad2 || resp.reserved))
582 		return -EINVAL;
583 
584 	switch (resp.uc_type) {
585 	case XE_QUERY_UC_TYPE_GUC_SUBMISSION: {
586 		struct xe_guc *guc = &xe->tiles[0].primary_gt->uc.guc;
587 
588 		version = &guc->fw.versions.found[XE_UC_FW_VER_COMPATIBILITY];
589 		break;
590 	}
591 	case XE_QUERY_UC_TYPE_HUC: {
592 		struct xe_gt *media_gt = NULL;
593 		struct xe_huc *huc;
594 
595 		if (MEDIA_VER(xe) >= 13) {
596 			struct xe_tile *tile;
597 			u8 gt_id;
598 
599 			for_each_tile(tile, xe, gt_id) {
600 				if (tile->media_gt) {
601 					media_gt = tile->media_gt;
602 					break;
603 				}
604 			}
605 		} else {
606 			media_gt = xe->tiles[0].primary_gt;
607 		}
608 
609 		if (!media_gt)
610 			break;
611 
612 		huc = &media_gt->uc.huc;
613 		if (huc->fw.status == XE_UC_FIRMWARE_RUNNING)
614 			version = &huc->fw.versions.found[XE_UC_FW_VER_RELEASE];
615 		break;
616 	}
617 	default:
618 		return -EINVAL;
619 	}
620 
621 	if (version) {
622 		resp.branch_ver = 0;
623 		resp.major_ver = version->major;
624 		resp.minor_ver = version->minor;
625 		resp.patch_ver = version->patch;
626 	} else {
627 		return -ENODEV;
628 	}
629 
630 	if (copy_to_user(query_ptr, &resp, size))
631 		return -EFAULT;
632 
633 	return 0;
634 }
635 
calc_oa_unit_query_size(struct xe_device * xe)636 static size_t calc_oa_unit_query_size(struct xe_device *xe)
637 {
638 	size_t size = sizeof(struct drm_xe_query_oa_units);
639 	struct xe_gt *gt;
640 	int i, id;
641 
642 	for_each_gt(gt, xe, id) {
643 		for (i = 0; i < gt->oa.num_oa_units; i++) {
644 			size += sizeof(struct drm_xe_oa_unit);
645 			size += gt->oa.oa_unit[i].num_engines *
646 				sizeof(struct drm_xe_engine_class_instance);
647 		}
648 	}
649 
650 	return size;
651 }
652 
query_oa_units(struct xe_device * xe,struct drm_xe_device_query * query)653 static int query_oa_units(struct xe_device *xe,
654 			  struct drm_xe_device_query *query)
655 {
656 	void __user *query_ptr = u64_to_user_ptr(query->data);
657 	size_t size = calc_oa_unit_query_size(xe);
658 	struct drm_xe_query_oa_units *qoa;
659 	enum xe_hw_engine_id hwe_id;
660 	struct drm_xe_oa_unit *du;
661 	struct xe_hw_engine *hwe;
662 	struct xe_oa_unit *u;
663 	int gt_id, i, j, ret;
664 	struct xe_gt *gt;
665 	u8 *pdu;
666 
667 	if (query->size == 0) {
668 		query->size = size;
669 		return 0;
670 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
671 		return -EINVAL;
672 	}
673 
674 	qoa = kzalloc(size, GFP_KERNEL);
675 	if (!qoa)
676 		return -ENOMEM;
677 
678 	pdu = (u8 *)&qoa->oa_units[0];
679 	for_each_gt(gt, xe, gt_id) {
680 		for (i = 0; i < gt->oa.num_oa_units; i++) {
681 			u = &gt->oa.oa_unit[i];
682 			du = (struct drm_xe_oa_unit *)pdu;
683 
684 			du->oa_unit_id = u->oa_unit_id;
685 			du->oa_unit_type = u->type;
686 			du->oa_timestamp_freq = xe_oa_timestamp_frequency(gt);
687 			du->capabilities = DRM_XE_OA_CAPS_BASE | DRM_XE_OA_CAPS_SYNCS |
688 					   DRM_XE_OA_CAPS_OA_BUFFER_SIZE |
689 					   DRM_XE_OA_CAPS_WAIT_NUM_REPORTS |
690 					   DRM_XE_OA_CAPS_OAM;
691 			j = 0;
692 			for_each_hw_engine(hwe, gt, hwe_id) {
693 				if (!xe_hw_engine_is_reserved(hwe) &&
694 				    xe_oa_unit_id(hwe) == u->oa_unit_id) {
695 					du->eci[j].engine_class =
696 						xe_to_user_engine_class[hwe->class];
697 					du->eci[j].engine_instance = hwe->logical_instance;
698 					du->eci[j].gt_id = gt->info.id;
699 					j++;
700 				}
701 			}
702 			du->num_engines = j;
703 			pdu += sizeof(*du) + j * sizeof(du->eci[0]);
704 			qoa->num_oa_units++;
705 		}
706 	}
707 
708 	ret = copy_to_user(query_ptr, qoa, size);
709 	kfree(qoa);
710 
711 	return ret ? -EFAULT : 0;
712 }
713 
query_pxp_status(struct xe_device * xe,struct drm_xe_device_query * query)714 static int query_pxp_status(struct xe_device *xe, struct drm_xe_device_query *query)
715 {
716 	struct drm_xe_query_pxp_status __user *query_ptr = u64_to_user_ptr(query->data);
717 	size_t size = sizeof(struct drm_xe_query_pxp_status);
718 	struct drm_xe_query_pxp_status resp = { 0 };
719 	int ret;
720 
721 	if (query->size == 0) {
722 		query->size = size;
723 		return 0;
724 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
725 		return -EINVAL;
726 	}
727 
728 	ret = xe_pxp_get_readiness_status(xe->pxp);
729 	if (ret < 0)
730 		return ret;
731 
732 	resp.status = ret;
733 	resp.supported_session_types = BIT(DRM_XE_PXP_TYPE_HWDRM);
734 
735 	if (copy_to_user(query_ptr, &resp, size))
736 		return -EFAULT;
737 
738 	return 0;
739 }
740 
query_eu_stall(struct xe_device * xe,struct drm_xe_device_query * query)741 static int query_eu_stall(struct xe_device *xe,
742 			  struct drm_xe_device_query *query)
743 {
744 	void __user *query_ptr = u64_to_user_ptr(query->data);
745 	struct drm_xe_query_eu_stall *info;
746 	size_t size, array_size;
747 	const u64 *rates;
748 	u32 num_rates;
749 	int ret;
750 
751 	if (!xe_eu_stall_supported_on_platform(xe)) {
752 		drm_dbg(&xe->drm, "EU stall monitoring is not supported on this platform\n");
753 		return -ENODEV;
754 	}
755 
756 	array_size = xe_eu_stall_get_sampling_rates(&num_rates, &rates);
757 	size = sizeof(struct drm_xe_query_eu_stall) + array_size;
758 
759 	if (query->size == 0) {
760 		query->size = size;
761 		return 0;
762 	} else if (XE_IOCTL_DBG(xe, query->size != size)) {
763 		return -EINVAL;
764 	}
765 
766 	info = kzalloc(size, GFP_KERNEL);
767 	if (!info)
768 		return -ENOMEM;
769 
770 	info->num_sampling_rates = num_rates;
771 	info->capabilities = DRM_XE_EU_STALL_CAPS_BASE;
772 	info->record_size = xe_eu_stall_data_record_size(xe);
773 	info->per_xecore_buf_size = xe_eu_stall_get_per_xecore_buf_size();
774 	memcpy(info->sampling_rates, rates, array_size);
775 
776 	ret = copy_to_user(query_ptr, info, size);
777 	kfree(info);
778 
779 	return ret ? -EFAULT : 0;
780 }
781 
782 static int (* const xe_query_funcs[])(struct xe_device *xe,
783 				      struct drm_xe_device_query *query) = {
784 	query_engines,
785 	query_mem_regions,
786 	query_config,
787 	query_gt_list,
788 	query_hwconfig,
789 	query_gt_topology,
790 	query_engine_cycles,
791 	query_uc_fw_version,
792 	query_oa_units,
793 	query_pxp_status,
794 	query_eu_stall,
795 };
796 
xe_query_ioctl(struct drm_device * dev,void * data,struct drm_file * file)797 int xe_query_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
798 {
799 	struct xe_device *xe = to_xe_device(dev);
800 	struct drm_xe_device_query *query = data;
801 	u32 idx;
802 
803 	if (XE_IOCTL_DBG(xe, query->extensions) ||
804 	    XE_IOCTL_DBG(xe, query->reserved[0] || query->reserved[1]))
805 		return -EINVAL;
806 
807 	if (XE_IOCTL_DBG(xe, query->query >= ARRAY_SIZE(xe_query_funcs)))
808 		return -EINVAL;
809 
810 	idx = array_index_nospec(query->query, ARRAY_SIZE(xe_query_funcs));
811 	if (XE_IOCTL_DBG(xe, !xe_query_funcs[idx]))
812 		return -EINVAL;
813 
814 	return xe_query_funcs[idx](xe, query);
815 }
816