xref: /linux/drivers/gpu/drm/xe/xe_gt_pagefault.c (revision 173b0b5b0e865348684c02bd9cb1d22b5d46e458)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #include "xe_gt_pagefault.h"
7 
8 #include <linux/bitfield.h>
9 #include <linux/circ_buf.h>
10 
11 #include <drm/drm_exec.h>
12 #include <drm/drm_managed.h>
13 #include <drm/ttm/ttm_execbuf_util.h>
14 
15 #include "abi/guc_actions_abi.h"
16 #include "xe_bo.h"
17 #include "xe_gt.h"
18 #include "xe_gt_tlb_invalidation.h"
19 #include "xe_guc.h"
20 #include "xe_guc_ct.h"
21 #include "xe_migrate.h"
22 #include "xe_pt.h"
23 #include "xe_trace.h"
24 #include "xe_vm.h"
25 
26 struct pagefault {
27 	u64 page_addr;
28 	u32 asid;
29 	u16 pdata;
30 	u8 vfid;
31 	u8 access_type;
32 	u8 fault_type;
33 	u8 fault_level;
34 	u8 engine_class;
35 	u8 engine_instance;
36 	u8 fault_unsuccessful;
37 	bool trva_fault;
38 };
39 
40 enum access_type {
41 	ACCESS_TYPE_READ = 0,
42 	ACCESS_TYPE_WRITE = 1,
43 	ACCESS_TYPE_ATOMIC = 2,
44 	ACCESS_TYPE_RESERVED = 3,
45 };
46 
47 enum fault_type {
48 	NOT_PRESENT = 0,
49 	WRITE_ACCESS_VIOLATION = 1,
50 	ATOMIC_ACCESS_VIOLATION = 2,
51 };
52 
53 struct acc {
54 	u64 va_range_base;
55 	u32 asid;
56 	u32 sub_granularity;
57 	u8 granularity;
58 	u8 vfid;
59 	u8 access_type;
60 	u8 engine_class;
61 	u8 engine_instance;
62 };
63 
64 static bool access_is_atomic(enum access_type access_type)
65 {
66 	return access_type == ACCESS_TYPE_ATOMIC;
67 }
68 
69 static bool vma_is_valid(struct xe_tile *tile, struct xe_vma *vma)
70 {
71 	return BIT(tile->id) & vma->tile_present &&
72 		!(BIT(tile->id) & vma->tile_invalidated);
73 }
74 
75 static bool vma_matches(struct xe_vma *vma, u64 page_addr)
76 {
77 	if (page_addr > xe_vma_end(vma) - 1 ||
78 	    page_addr + SZ_4K - 1 < xe_vma_start(vma))
79 		return false;
80 
81 	return true;
82 }
83 
84 static struct xe_vma *lookup_vma(struct xe_vm *vm, u64 page_addr)
85 {
86 	struct xe_vma *vma = NULL;
87 
88 	if (vm->usm.last_fault_vma) {   /* Fast lookup */
89 		if (vma_matches(vm->usm.last_fault_vma, page_addr))
90 			vma = vm->usm.last_fault_vma;
91 	}
92 	if (!vma)
93 		vma = xe_vm_find_overlapping_vma(vm, page_addr, SZ_4K);
94 
95 	return vma;
96 }
97 
98 static int xe_pf_begin(struct drm_exec *exec, struct xe_vma *vma,
99 		       bool atomic, unsigned int id)
100 {
101 	struct xe_bo *bo = xe_vma_bo(vma);
102 	struct xe_vm *vm = xe_vma_vm(vma);
103 	int err;
104 
105 	err = xe_vm_lock_vma(exec, vma);
106 	if (err)
107 		return err;
108 
109 	if (atomic && IS_DGFX(vm->xe)) {
110 		if (xe_vma_is_userptr(vma)) {
111 			err = -EACCES;
112 			return err;
113 		}
114 
115 		/* Migrate to VRAM, move should invalidate the VMA first */
116 		err = xe_bo_migrate(bo, XE_PL_VRAM0 + id);
117 		if (err)
118 			return err;
119 	} else if (bo) {
120 		/* Create backing store if needed */
121 		err = xe_bo_validate(bo, vm, true);
122 		if (err)
123 			return err;
124 	}
125 
126 	return 0;
127 }
128 
129 static int handle_pagefault(struct xe_gt *gt, struct pagefault *pf)
130 {
131 	struct xe_device *xe = gt_to_xe(gt);
132 	struct xe_tile *tile = gt_to_tile(gt);
133 	struct drm_exec exec;
134 	struct xe_vm *vm;
135 	struct xe_vma *vma = NULL;
136 	struct dma_fence *fence;
137 	bool write_locked;
138 	int ret = 0;
139 	bool atomic;
140 
141 	/* SW isn't expected to handle TRTT faults */
142 	if (pf->trva_fault)
143 		return -EFAULT;
144 
145 	/* ASID to VM */
146 	mutex_lock(&xe->usm.lock);
147 	vm = xa_load(&xe->usm.asid_to_vm, pf->asid);
148 	if (vm && xe_vm_in_fault_mode(vm))
149 		xe_vm_get(vm);
150 	else
151 		vm = NULL;
152 	mutex_unlock(&xe->usm.lock);
153 	if (!vm)
154 		return -EINVAL;
155 
156 retry_userptr:
157 	/*
158 	 * TODO: Avoid exclusive lock if VM doesn't have userptrs, or
159 	 * start out read-locked?
160 	 */
161 	down_write(&vm->lock);
162 	write_locked = true;
163 	vma = lookup_vma(vm, pf->page_addr);
164 	if (!vma) {
165 		ret = -EINVAL;
166 		goto unlock_vm;
167 	}
168 
169 	if (!xe_vma_is_userptr(vma) ||
170 	    !xe_vma_userptr_check_repin(to_userptr_vma(vma))) {
171 		downgrade_write(&vm->lock);
172 		write_locked = false;
173 	}
174 
175 	trace_xe_vma_pagefault(vma);
176 
177 	atomic = access_is_atomic(pf->access_type);
178 
179 	/* Check if VMA is valid */
180 	if (vma_is_valid(tile, vma) && !atomic)
181 		goto unlock_vm;
182 
183 	/* TODO: Validate fault */
184 
185 	if (xe_vma_is_userptr(vma) && write_locked) {
186 		struct xe_userptr_vma *uvma = to_userptr_vma(vma);
187 
188 		spin_lock(&vm->userptr.invalidated_lock);
189 		list_del_init(&uvma->userptr.invalidate_link);
190 		spin_unlock(&vm->userptr.invalidated_lock);
191 
192 		ret = xe_vma_userptr_pin_pages(uvma);
193 		if (ret)
194 			goto unlock_vm;
195 
196 		downgrade_write(&vm->lock);
197 		write_locked = false;
198 	}
199 
200 	/* Lock VM and BOs dma-resv */
201 	drm_exec_init(&exec, 0, 0);
202 	drm_exec_until_all_locked(&exec) {
203 		ret = xe_pf_begin(&exec, vma, atomic, tile->id);
204 		drm_exec_retry_on_contention(&exec);
205 		if (ret)
206 			goto unlock_dma_resv;
207 	}
208 
209 	/* Bind VMA only to the GT that has faulted */
210 	trace_xe_vma_pf_bind(vma);
211 	fence = __xe_pt_bind_vma(tile, vma, xe_tile_migrate_engine(tile), NULL, 0,
212 				 vma->tile_present & BIT(tile->id));
213 	if (IS_ERR(fence)) {
214 		ret = PTR_ERR(fence);
215 		goto unlock_dma_resv;
216 	}
217 
218 	/*
219 	 * XXX: Should we drop the lock before waiting? This only helps if doing
220 	 * GPU binds which is currently only done if we have to wait for more
221 	 * than 10ms on a move.
222 	 */
223 	dma_fence_wait(fence, false);
224 	dma_fence_put(fence);
225 
226 	if (xe_vma_is_userptr(vma))
227 		ret = xe_vma_userptr_check_repin(to_userptr_vma(vma));
228 	vma->tile_invalidated &= ~BIT(tile->id);
229 
230 unlock_dma_resv:
231 	drm_exec_fini(&exec);
232 unlock_vm:
233 	if (!ret)
234 		vm->usm.last_fault_vma = vma;
235 	if (write_locked)
236 		up_write(&vm->lock);
237 	else
238 		up_read(&vm->lock);
239 	if (ret == -EAGAIN)
240 		goto retry_userptr;
241 
242 	if (!ret) {
243 		ret = xe_gt_tlb_invalidation_vma(gt, NULL, vma);
244 		if (ret >= 0)
245 			ret = 0;
246 	}
247 	xe_vm_put(vm);
248 
249 	return ret;
250 }
251 
252 static int send_pagefault_reply(struct xe_guc *guc,
253 				struct xe_guc_pagefault_reply *reply)
254 {
255 	u32 action[] = {
256 		XE_GUC_ACTION_PAGE_FAULT_RES_DESC,
257 		reply->dw0,
258 		reply->dw1,
259 	};
260 
261 	return xe_guc_ct_send(&guc->ct, action, ARRAY_SIZE(action), 0, 0);
262 }
263 
264 static void print_pagefault(struct xe_device *xe, struct pagefault *pf)
265 {
266 	drm_dbg(&xe->drm, "\n\tASID: %d\n"
267 		 "\tVFID: %d\n"
268 		 "\tPDATA: 0x%04x\n"
269 		 "\tFaulted Address: 0x%08x%08x\n"
270 		 "\tFaultType: %d\n"
271 		 "\tAccessType: %d\n"
272 		 "\tFaultLevel: %d\n"
273 		 "\tEngineClass: %d\n"
274 		 "\tEngineInstance: %d\n",
275 		 pf->asid, pf->vfid, pf->pdata, upper_32_bits(pf->page_addr),
276 		 lower_32_bits(pf->page_addr),
277 		 pf->fault_type, pf->access_type, pf->fault_level,
278 		 pf->engine_class, pf->engine_instance);
279 }
280 
281 #define PF_MSG_LEN_DW	4
282 
283 static bool get_pagefault(struct pf_queue *pf_queue, struct pagefault *pf)
284 {
285 	const struct xe_guc_pagefault_desc *desc;
286 	bool ret = false;
287 
288 	spin_lock_irq(&pf_queue->lock);
289 	if (pf_queue->tail != pf_queue->head) {
290 		desc = (const struct xe_guc_pagefault_desc *)
291 			(pf_queue->data + pf_queue->tail);
292 
293 		pf->fault_level = FIELD_GET(PFD_FAULT_LEVEL, desc->dw0);
294 		pf->trva_fault = FIELD_GET(XE2_PFD_TRVA_FAULT, desc->dw0);
295 		pf->engine_class = FIELD_GET(PFD_ENG_CLASS, desc->dw0);
296 		pf->engine_instance = FIELD_GET(PFD_ENG_INSTANCE, desc->dw0);
297 		pf->pdata = FIELD_GET(PFD_PDATA_HI, desc->dw1) <<
298 			PFD_PDATA_HI_SHIFT;
299 		pf->pdata |= FIELD_GET(PFD_PDATA_LO, desc->dw0);
300 		pf->asid = FIELD_GET(PFD_ASID, desc->dw1);
301 		pf->vfid = FIELD_GET(PFD_VFID, desc->dw2);
302 		pf->access_type = FIELD_GET(PFD_ACCESS_TYPE, desc->dw2);
303 		pf->fault_type = FIELD_GET(PFD_FAULT_TYPE, desc->dw2);
304 		pf->page_addr = (u64)(FIELD_GET(PFD_VIRTUAL_ADDR_HI, desc->dw3)) <<
305 			PFD_VIRTUAL_ADDR_HI_SHIFT;
306 		pf->page_addr |= FIELD_GET(PFD_VIRTUAL_ADDR_LO, desc->dw2) <<
307 			PFD_VIRTUAL_ADDR_LO_SHIFT;
308 
309 		pf_queue->tail = (pf_queue->tail + PF_MSG_LEN_DW) %
310 			PF_QUEUE_NUM_DW;
311 		ret = true;
312 	}
313 	spin_unlock_irq(&pf_queue->lock);
314 
315 	return ret;
316 }
317 
318 static bool pf_queue_full(struct pf_queue *pf_queue)
319 {
320 	lockdep_assert_held(&pf_queue->lock);
321 
322 	return CIRC_SPACE(pf_queue->head, pf_queue->tail, PF_QUEUE_NUM_DW) <=
323 		PF_MSG_LEN_DW;
324 }
325 
326 int xe_guc_pagefault_handler(struct xe_guc *guc, u32 *msg, u32 len)
327 {
328 	struct xe_gt *gt = guc_to_gt(guc);
329 	struct xe_device *xe = gt_to_xe(gt);
330 	struct pf_queue *pf_queue;
331 	unsigned long flags;
332 	u32 asid;
333 	bool full;
334 
335 	/*
336 	 * The below logic doesn't work unless PF_QUEUE_NUM_DW % PF_MSG_LEN_DW == 0
337 	 */
338 	BUILD_BUG_ON(PF_QUEUE_NUM_DW % PF_MSG_LEN_DW);
339 
340 	if (unlikely(len != PF_MSG_LEN_DW))
341 		return -EPROTO;
342 
343 	asid = FIELD_GET(PFD_ASID, msg[1]);
344 	pf_queue = gt->usm.pf_queue + (asid % NUM_PF_QUEUE);
345 
346 	spin_lock_irqsave(&pf_queue->lock, flags);
347 	full = pf_queue_full(pf_queue);
348 	if (!full) {
349 		memcpy(pf_queue->data + pf_queue->head, msg, len * sizeof(u32));
350 		pf_queue->head = (pf_queue->head + len) % PF_QUEUE_NUM_DW;
351 		queue_work(gt->usm.pf_wq, &pf_queue->worker);
352 	} else {
353 		drm_warn(&xe->drm, "PF Queue full, shouldn't be possible");
354 	}
355 	spin_unlock_irqrestore(&pf_queue->lock, flags);
356 
357 	return full ? -ENOSPC : 0;
358 }
359 
360 #define USM_QUEUE_MAX_RUNTIME_MS	20
361 
362 static void pf_queue_work_func(struct work_struct *w)
363 {
364 	struct pf_queue *pf_queue = container_of(w, struct pf_queue, worker);
365 	struct xe_gt *gt = pf_queue->gt;
366 	struct xe_device *xe = gt_to_xe(gt);
367 	struct xe_guc_pagefault_reply reply = {};
368 	struct pagefault pf = {};
369 	unsigned long threshold;
370 	int ret;
371 
372 	threshold = jiffies + msecs_to_jiffies(USM_QUEUE_MAX_RUNTIME_MS);
373 
374 	while (get_pagefault(pf_queue, &pf)) {
375 		ret = handle_pagefault(gt, &pf);
376 		if (unlikely(ret)) {
377 			print_pagefault(xe, &pf);
378 			pf.fault_unsuccessful = 1;
379 			drm_dbg(&xe->drm, "Fault response: Unsuccessful %d\n", ret);
380 		}
381 
382 		reply.dw0 = FIELD_PREP(PFR_VALID, 1) |
383 			FIELD_PREP(PFR_SUCCESS, pf.fault_unsuccessful) |
384 			FIELD_PREP(PFR_REPLY, PFR_ACCESS) |
385 			FIELD_PREP(PFR_DESC_TYPE, FAULT_RESPONSE_DESC) |
386 			FIELD_PREP(PFR_ASID, pf.asid);
387 
388 		reply.dw1 = FIELD_PREP(PFR_VFID, pf.vfid) |
389 			FIELD_PREP(PFR_ENG_INSTANCE, pf.engine_instance) |
390 			FIELD_PREP(PFR_ENG_CLASS, pf.engine_class) |
391 			FIELD_PREP(PFR_PDATA, pf.pdata);
392 
393 		send_pagefault_reply(&gt->uc.guc, &reply);
394 
395 		if (time_after(jiffies, threshold) &&
396 		    pf_queue->tail != pf_queue->head) {
397 			queue_work(gt->usm.pf_wq, w);
398 			break;
399 		}
400 	}
401 }
402 
403 static void acc_queue_work_func(struct work_struct *w);
404 
405 int xe_gt_pagefault_init(struct xe_gt *gt)
406 {
407 	struct xe_device *xe = gt_to_xe(gt);
408 	int i;
409 
410 	if (!xe->info.has_usm)
411 		return 0;
412 
413 	for (i = 0; i < NUM_PF_QUEUE; ++i) {
414 		gt->usm.pf_queue[i].gt = gt;
415 		spin_lock_init(&gt->usm.pf_queue[i].lock);
416 		INIT_WORK(&gt->usm.pf_queue[i].worker, pf_queue_work_func);
417 	}
418 	for (i = 0; i < NUM_ACC_QUEUE; ++i) {
419 		gt->usm.acc_queue[i].gt = gt;
420 		spin_lock_init(&gt->usm.acc_queue[i].lock);
421 		INIT_WORK(&gt->usm.acc_queue[i].worker, acc_queue_work_func);
422 	}
423 
424 	gt->usm.pf_wq = alloc_workqueue("xe_gt_page_fault_work_queue",
425 					WQ_UNBOUND | WQ_HIGHPRI, NUM_PF_QUEUE);
426 	if (!gt->usm.pf_wq)
427 		return -ENOMEM;
428 
429 	gt->usm.acc_wq = alloc_workqueue("xe_gt_access_counter_work_queue",
430 					 WQ_UNBOUND | WQ_HIGHPRI,
431 					 NUM_ACC_QUEUE);
432 	if (!gt->usm.acc_wq)
433 		return -ENOMEM;
434 
435 	return 0;
436 }
437 
438 void xe_gt_pagefault_reset(struct xe_gt *gt)
439 {
440 	struct xe_device *xe = gt_to_xe(gt);
441 	int i;
442 
443 	if (!xe->info.has_usm)
444 		return;
445 
446 	for (i = 0; i < NUM_PF_QUEUE; ++i) {
447 		spin_lock_irq(&gt->usm.pf_queue[i].lock);
448 		gt->usm.pf_queue[i].head = 0;
449 		gt->usm.pf_queue[i].tail = 0;
450 		spin_unlock_irq(&gt->usm.pf_queue[i].lock);
451 	}
452 
453 	for (i = 0; i < NUM_ACC_QUEUE; ++i) {
454 		spin_lock(&gt->usm.acc_queue[i].lock);
455 		gt->usm.acc_queue[i].head = 0;
456 		gt->usm.acc_queue[i].tail = 0;
457 		spin_unlock(&gt->usm.acc_queue[i].lock);
458 	}
459 }
460 
461 static int granularity_in_byte(int val)
462 {
463 	switch (val) {
464 	case 0:
465 		return SZ_128K;
466 	case 1:
467 		return SZ_2M;
468 	case 2:
469 		return SZ_16M;
470 	case 3:
471 		return SZ_64M;
472 	default:
473 		return 0;
474 	}
475 }
476 
477 static int sub_granularity_in_byte(int val)
478 {
479 	return (granularity_in_byte(val) / 32);
480 }
481 
482 static void print_acc(struct xe_device *xe, struct acc *acc)
483 {
484 	drm_warn(&xe->drm, "Access counter request:\n"
485 		 "\tType: %s\n"
486 		 "\tASID: %d\n"
487 		 "\tVFID: %d\n"
488 		 "\tEngine: %d:%d\n"
489 		 "\tGranularity: 0x%x KB Region/ %d KB sub-granularity\n"
490 		 "\tSub_Granularity Vector: 0x%08x\n"
491 		 "\tVA Range base: 0x%016llx\n",
492 		 acc->access_type ? "AC_NTFY_VAL" : "AC_TRIG_VAL",
493 		 acc->asid, acc->vfid, acc->engine_class, acc->engine_instance,
494 		 granularity_in_byte(acc->granularity) / SZ_1K,
495 		 sub_granularity_in_byte(acc->granularity) / SZ_1K,
496 		 acc->sub_granularity, acc->va_range_base);
497 }
498 
499 static struct xe_vma *get_acc_vma(struct xe_vm *vm, struct acc *acc)
500 {
501 	u64 page_va = acc->va_range_base + (ffs(acc->sub_granularity) - 1) *
502 		sub_granularity_in_byte(acc->granularity);
503 
504 	return xe_vm_find_overlapping_vma(vm, page_va, SZ_4K);
505 }
506 
507 static int handle_acc(struct xe_gt *gt, struct acc *acc)
508 {
509 	struct xe_device *xe = gt_to_xe(gt);
510 	struct xe_tile *tile = gt_to_tile(gt);
511 	struct drm_exec exec;
512 	struct xe_vm *vm;
513 	struct xe_vma *vma;
514 	int ret = 0;
515 
516 	/* We only support ACC_TRIGGER at the moment */
517 	if (acc->access_type != ACC_TRIGGER)
518 		return -EINVAL;
519 
520 	/* ASID to VM */
521 	mutex_lock(&xe->usm.lock);
522 	vm = xa_load(&xe->usm.asid_to_vm, acc->asid);
523 	if (vm)
524 		xe_vm_get(vm);
525 	mutex_unlock(&xe->usm.lock);
526 	if (!vm || !xe_vm_in_fault_mode(vm))
527 		return -EINVAL;
528 
529 	down_read(&vm->lock);
530 
531 	/* Lookup VMA */
532 	vma = get_acc_vma(vm, acc);
533 	if (!vma) {
534 		ret = -EINVAL;
535 		goto unlock_vm;
536 	}
537 
538 	trace_xe_vma_acc(vma);
539 
540 	/* Userptr or null can't be migrated, nothing to do */
541 	if (xe_vma_has_no_bo(vma))
542 		goto unlock_vm;
543 
544 	/* Lock VM and BOs dma-resv */
545 	drm_exec_init(&exec, 0, 0);
546 	drm_exec_until_all_locked(&exec) {
547 		ret = xe_pf_begin(&exec, vma, true, tile->id);
548 		drm_exec_retry_on_contention(&exec);
549 		if (ret)
550 			break;
551 	}
552 
553 	drm_exec_fini(&exec);
554 unlock_vm:
555 	up_read(&vm->lock);
556 	xe_vm_put(vm);
557 
558 	return ret;
559 }
560 
561 #define make_u64(hi__, low__)  ((u64)(hi__) << 32 | (u64)(low__))
562 
563 #define ACC_MSG_LEN_DW        4
564 
565 static bool get_acc(struct acc_queue *acc_queue, struct acc *acc)
566 {
567 	const struct xe_guc_acc_desc *desc;
568 	bool ret = false;
569 
570 	spin_lock(&acc_queue->lock);
571 	if (acc_queue->tail != acc_queue->head) {
572 		desc = (const struct xe_guc_acc_desc *)
573 			(acc_queue->data + acc_queue->tail);
574 
575 		acc->granularity = FIELD_GET(ACC_GRANULARITY, desc->dw2);
576 		acc->sub_granularity = FIELD_GET(ACC_SUBG_HI, desc->dw1) << 31 |
577 			FIELD_GET(ACC_SUBG_LO, desc->dw0);
578 		acc->engine_class = FIELD_GET(ACC_ENG_CLASS, desc->dw1);
579 		acc->engine_instance = FIELD_GET(ACC_ENG_INSTANCE, desc->dw1);
580 		acc->asid =  FIELD_GET(ACC_ASID, desc->dw1);
581 		acc->vfid =  FIELD_GET(ACC_VFID, desc->dw2);
582 		acc->access_type = FIELD_GET(ACC_TYPE, desc->dw0);
583 		acc->va_range_base = make_u64(desc->dw3 & ACC_VIRTUAL_ADDR_RANGE_HI,
584 					      desc->dw2 & ACC_VIRTUAL_ADDR_RANGE_LO);
585 
586 		acc_queue->tail = (acc_queue->tail + ACC_MSG_LEN_DW) %
587 				  ACC_QUEUE_NUM_DW;
588 		ret = true;
589 	}
590 	spin_unlock(&acc_queue->lock);
591 
592 	return ret;
593 }
594 
595 static void acc_queue_work_func(struct work_struct *w)
596 {
597 	struct acc_queue *acc_queue = container_of(w, struct acc_queue, worker);
598 	struct xe_gt *gt = acc_queue->gt;
599 	struct xe_device *xe = gt_to_xe(gt);
600 	struct acc acc = {};
601 	unsigned long threshold;
602 	int ret;
603 
604 	threshold = jiffies + msecs_to_jiffies(USM_QUEUE_MAX_RUNTIME_MS);
605 
606 	while (get_acc(acc_queue, &acc)) {
607 		ret = handle_acc(gt, &acc);
608 		if (unlikely(ret)) {
609 			print_acc(xe, &acc);
610 			drm_warn(&xe->drm, "ACC: Unsuccessful %d\n", ret);
611 		}
612 
613 		if (time_after(jiffies, threshold) &&
614 		    acc_queue->tail != acc_queue->head) {
615 			queue_work(gt->usm.acc_wq, w);
616 			break;
617 		}
618 	}
619 }
620 
621 static bool acc_queue_full(struct acc_queue *acc_queue)
622 {
623 	lockdep_assert_held(&acc_queue->lock);
624 
625 	return CIRC_SPACE(acc_queue->head, acc_queue->tail, ACC_QUEUE_NUM_DW) <=
626 		ACC_MSG_LEN_DW;
627 }
628 
629 int xe_guc_access_counter_notify_handler(struct xe_guc *guc, u32 *msg, u32 len)
630 {
631 	struct xe_gt *gt = guc_to_gt(guc);
632 	struct acc_queue *acc_queue;
633 	u32 asid;
634 	bool full;
635 
636 	/*
637 	 * The below logic doesn't work unless ACC_QUEUE_NUM_DW % ACC_MSG_LEN_DW == 0
638 	 */
639 	BUILD_BUG_ON(ACC_QUEUE_NUM_DW % ACC_MSG_LEN_DW);
640 
641 	if (unlikely(len != ACC_MSG_LEN_DW))
642 		return -EPROTO;
643 
644 	asid = FIELD_GET(ACC_ASID, msg[1]);
645 	acc_queue = &gt->usm.acc_queue[asid % NUM_ACC_QUEUE];
646 
647 	spin_lock(&acc_queue->lock);
648 	full = acc_queue_full(acc_queue);
649 	if (!full) {
650 		memcpy(acc_queue->data + acc_queue->head, msg,
651 		       len * sizeof(u32));
652 		acc_queue->head = (acc_queue->head + len) % ACC_QUEUE_NUM_DW;
653 		queue_work(gt->usm.acc_wq, &acc_queue->worker);
654 	} else {
655 		drm_warn(&gt_to_xe(gt)->drm, "ACC Queue full, dropping ACC");
656 	}
657 	spin_unlock(&acc_queue->lock);
658 
659 	return full ? -ENOSPC : 0;
660 }
661