1 /*
2  * Copyright 2017 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 #include "amdgpu_ids.h"
24 
25 #include <linux/idr.h>
26 #include <linux/dma-fence-array.h>
27 
28 
29 #include "amdgpu.h"
30 #include "amdgpu_trace.h"
31 
32 /*
33  * PASID manager
34  *
35  * PASIDs are global address space identifiers that can be shared
36  * between the GPU, an IOMMU and the driver. VMs on different devices
37  * may use the same PASID if they share the same address
38  * space. Therefore PASIDs are allocated using a global IDA. VMs are
39  * looked up from the PASID per amdgpu_device.
40  */
41 static DEFINE_IDA(amdgpu_pasid_ida);
42 
43 /* Helper to free pasid from a fence callback */
44 struct amdgpu_pasid_cb {
45 	struct dma_fence_cb cb;
46 	u32 pasid;
47 };
48 
49 /**
50  * amdgpu_pasid_alloc - Allocate a PASID
51  * @bits: Maximum width of the PASID in bits, must be at least 1
52  *
53  * Allocates a PASID of the given width while keeping smaller PASIDs
54  * available if possible.
55  *
56  * Returns a positive integer on success. Returns %-EINVAL if bits==0.
57  * Returns %-ENOSPC if no PASID was available. Returns %-ENOMEM on
58  * memory allocation failure.
59  */
60 int amdgpu_pasid_alloc(unsigned int bits)
61 {
62 	int pasid = -EINVAL;
63 
64 	for (bits = min(bits, 31U); bits > 0; bits--) {
65 		pasid = ida_simple_get(&amdgpu_pasid_ida,
66 				       1U << (bits - 1), 1U << bits,
67 				       GFP_KERNEL);
68 		if (pasid != -ENOSPC)
69 			break;
70 	}
71 
72 	if (pasid >= 0)
73 		trace_amdgpu_pasid_allocated(pasid);
74 
75 	return pasid;
76 }
77 
78 /**
79  * amdgpu_pasid_free - Free a PASID
80  * @pasid: PASID to free
81  */
82 void amdgpu_pasid_free(u32 pasid)
83 {
84 	trace_amdgpu_pasid_freed(pasid);
85 	ida_simple_remove(&amdgpu_pasid_ida, pasid);
86 }
87 
88 static void amdgpu_pasid_free_cb(struct dma_fence *fence,
89 				 struct dma_fence_cb *_cb)
90 {
91 	struct amdgpu_pasid_cb *cb =
92 		container_of(_cb, struct amdgpu_pasid_cb, cb);
93 
94 	amdgpu_pasid_free(cb->pasid);
95 	dma_fence_put(fence);
96 	kfree(cb);
97 }
98 
99 /**
100  * amdgpu_pasid_free_delayed - free pasid when fences signal
101  *
102  * @resv: reservation object with the fences to wait for
103  * @pasid: pasid to free
104  *
105  * Free the pasid only after all the fences in resv are signaled.
106  */
107 void amdgpu_pasid_free_delayed(struct dma_resv *resv,
108 			       u32 pasid)
109 {
110 	struct amdgpu_pasid_cb *cb;
111 	struct dma_fence *fence;
112 	int r;
113 
114 	r = dma_resv_get_singleton(resv, DMA_RESV_USAGE_BOOKKEEP, &fence);
115 	if (r)
116 		goto fallback;
117 
118 	if (!fence) {
119 		amdgpu_pasid_free(pasid);
120 		return;
121 	}
122 
123 	cb = kmalloc(sizeof(*cb), GFP_KERNEL);
124 	if (!cb) {
125 		/* Last resort when we are OOM */
126 		dma_fence_wait(fence, false);
127 		dma_fence_put(fence);
128 		amdgpu_pasid_free(pasid);
129 	} else {
130 		cb->pasid = pasid;
131 		if (dma_fence_add_callback(fence, &cb->cb,
132 					   amdgpu_pasid_free_cb))
133 			amdgpu_pasid_free_cb(fence, &cb->cb);
134 	}
135 
136 	return;
137 
138 fallback:
139 	/* Not enough memory for the delayed delete, as last resort
140 	 * block for all the fences to complete.
141 	 */
142 	dma_resv_wait_timeout(resv, DMA_RESV_USAGE_BOOKKEEP,
143 			      false, MAX_SCHEDULE_TIMEOUT);
144 	amdgpu_pasid_free(pasid);
145 }
146 
147 /*
148  * VMID manager
149  *
150  * VMIDs are a per VMHUB identifier for page tables handling.
151  */
152 
153 /**
154  * amdgpu_vmid_had_gpu_reset - check if reset occured since last use
155  *
156  * @adev: amdgpu_device pointer
157  * @id: VMID structure
158  *
159  * Check if GPU reset occured since last use of the VMID.
160  */
161 bool amdgpu_vmid_had_gpu_reset(struct amdgpu_device *adev,
162 			       struct amdgpu_vmid *id)
163 {
164 	return id->current_gpu_reset_count !=
165 		atomic_read(&adev->gpu_reset_counter);
166 }
167 
168 /* Check if we need to switch to another set of resources */
169 static bool amdgpu_vmid_gds_switch_needed(struct amdgpu_vmid *id,
170 					  struct amdgpu_job *job)
171 {
172 	return id->gds_base != job->gds_base ||
173 		id->gds_size != job->gds_size ||
174 		id->gws_base != job->gws_base ||
175 		id->gws_size != job->gws_size ||
176 		id->oa_base != job->oa_base ||
177 		id->oa_size != job->oa_size;
178 }
179 
180 /* Check if the id is compatible with the job */
181 static bool amdgpu_vmid_compatible(struct amdgpu_vmid *id,
182 				   struct amdgpu_job *job)
183 {
184 	return  id->pd_gpu_addr == job->vm_pd_addr &&
185 		!amdgpu_vmid_gds_switch_needed(id, job);
186 }
187 
188 /**
189  * amdgpu_vmid_grab_idle - grab idle VMID
190  *
191  * @vm: vm to allocate id for
192  * @ring: ring we want to submit job to
193  * @idle: resulting idle VMID
194  * @fence: fence to wait for if no id could be grabbed
195  *
196  * Try to find an idle VMID, if none is idle add a fence to wait to the sync
197  * object. Returns -ENOMEM when we are out of memory.
198  */
199 static int amdgpu_vmid_grab_idle(struct amdgpu_vm *vm,
200 				 struct amdgpu_ring *ring,
201 				 struct amdgpu_vmid **idle,
202 				 struct dma_fence **fence)
203 {
204 	struct amdgpu_device *adev = ring->adev;
205 	unsigned vmhub = ring->funcs->vmhub;
206 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
207 	struct dma_fence **fences;
208 	unsigned i;
209 
210 	if (!dma_fence_is_signaled(ring->vmid_wait)) {
211 		*fence = dma_fence_get(ring->vmid_wait);
212 		return 0;
213 	}
214 
215 	fences = kmalloc_array(id_mgr->num_ids, sizeof(void *), GFP_KERNEL);
216 	if (!fences)
217 		return -ENOMEM;
218 
219 	/* Check if we have an idle VMID */
220 	i = 0;
221 	list_for_each_entry((*idle), &id_mgr->ids_lru, list) {
222 		/* Don't use per engine and per process VMID at the same time */
223 		struct amdgpu_ring *r = adev->vm_manager.concurrent_flush ?
224 			NULL : ring;
225 
226 		fences[i] = amdgpu_sync_peek_fence(&(*idle)->active, r);
227 		if (!fences[i])
228 			break;
229 		++i;
230 	}
231 
232 	/* If we can't find a idle VMID to use, wait till one becomes available */
233 	if (&(*idle)->list == &id_mgr->ids_lru) {
234 		u64 fence_context = adev->vm_manager.fence_context + ring->idx;
235 		unsigned seqno = ++adev->vm_manager.seqno[ring->idx];
236 		struct dma_fence_array *array;
237 		unsigned j;
238 
239 		*idle = NULL;
240 		for (j = 0; j < i; ++j)
241 			dma_fence_get(fences[j]);
242 
243 		array = dma_fence_array_create(i, fences, fence_context,
244 					       seqno, true);
245 		if (!array) {
246 			for (j = 0; j < i; ++j)
247 				dma_fence_put(fences[j]);
248 			kfree(fences);
249 			return -ENOMEM;
250 		}
251 
252 		*fence = dma_fence_get(&array->base);
253 		dma_fence_put(ring->vmid_wait);
254 		ring->vmid_wait = &array->base;
255 		return 0;
256 	}
257 	kfree(fences);
258 
259 	return 0;
260 }
261 
262 /**
263  * amdgpu_vmid_grab_reserved - try to assign reserved VMID
264  *
265  * @vm: vm to allocate id for
266  * @ring: ring we want to submit job to
267  * @job: job who wants to use the VMID
268  * @id: resulting VMID
269  * @fence: fence to wait for if no id could be grabbed
270  *
271  * Try to assign a reserved VMID.
272  */
273 static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
274 				     struct amdgpu_ring *ring,
275 				     struct amdgpu_job *job,
276 				     struct amdgpu_vmid **id,
277 				     struct dma_fence **fence)
278 {
279 	struct amdgpu_device *adev = ring->adev;
280 	unsigned vmhub = ring->funcs->vmhub;
281 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
282 	uint64_t fence_context = adev->fence_context + ring->idx;
283 	bool needs_flush = vm->use_cpu_for_update;
284 	uint64_t updates = amdgpu_vm_tlb_seq(vm);
285 	int r;
286 
287 	*id = id_mgr->reserved;
288 	if ((*id)->owner != vm->immediate.fence_context ||
289 	    !amdgpu_vmid_compatible(*id, job) ||
290 	    (*id)->flushed_updates < updates ||
291 	    !(*id)->last_flush ||
292 	    ((*id)->last_flush->context != fence_context &&
293 	     !dma_fence_is_signaled((*id)->last_flush))) {
294 		struct dma_fence *tmp;
295 
296 		/* Don't use per engine and per process VMID at the same time */
297 		if (adev->vm_manager.concurrent_flush)
298 			ring = NULL;
299 
300 		/* to prevent one context starved by another context */
301 		(*id)->pd_gpu_addr = 0;
302 		tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
303 		if (tmp) {
304 			*id = NULL;
305 			*fence = dma_fence_get(tmp);
306 			return 0;
307 		}
308 		needs_flush = true;
309 	}
310 
311 	/* Good we can use this VMID. Remember this submission as
312 	* user of the VMID.
313 	*/
314 	r = amdgpu_sync_fence(&(*id)->active, &job->base.s_fence->finished);
315 	if (r)
316 		return r;
317 
318 	job->vm_needs_flush = needs_flush;
319 	job->spm_update_needed = true;
320 	return 0;
321 }
322 
323 /**
324  * amdgpu_vmid_grab_used - try to reuse a VMID
325  *
326  * @vm: vm to allocate id for
327  * @ring: ring we want to submit job to
328  * @job: job who wants to use the VMID
329  * @id: resulting VMID
330  * @fence: fence to wait for if no id could be grabbed
331  *
332  * Try to reuse a VMID for this submission.
333  */
334 static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm,
335 				 struct amdgpu_ring *ring,
336 				 struct amdgpu_job *job,
337 				 struct amdgpu_vmid **id,
338 				 struct dma_fence **fence)
339 {
340 	struct amdgpu_device *adev = ring->adev;
341 	unsigned vmhub = ring->funcs->vmhub;
342 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
343 	uint64_t fence_context = adev->fence_context + ring->idx;
344 	uint64_t updates = amdgpu_vm_tlb_seq(vm);
345 	int r;
346 
347 	job->vm_needs_flush = vm->use_cpu_for_update;
348 
349 	/* Check if we can use a VMID already assigned to this VM */
350 	list_for_each_entry_reverse((*id), &id_mgr->ids_lru, list) {
351 		bool needs_flush = vm->use_cpu_for_update;
352 
353 		/* Check all the prerequisites to using this VMID */
354 		if ((*id)->owner != vm->immediate.fence_context)
355 			continue;
356 
357 		if (!amdgpu_vmid_compatible(*id, job))
358 			continue;
359 
360 		if (!(*id)->last_flush ||
361 		    ((*id)->last_flush->context != fence_context &&
362 		     !dma_fence_is_signaled((*id)->last_flush)))
363 			needs_flush = true;
364 
365 		if ((*id)->flushed_updates < updates)
366 			needs_flush = true;
367 
368 		if (needs_flush && !adev->vm_manager.concurrent_flush)
369 			continue;
370 
371 		/* Good, we can use this VMID. Remember this submission as
372 		 * user of the VMID.
373 		 */
374 		r = amdgpu_sync_fence(&(*id)->active,
375 				      &job->base.s_fence->finished);
376 		if (r)
377 			return r;
378 
379 		job->vm_needs_flush |= needs_flush;
380 		return 0;
381 	}
382 
383 	*id = NULL;
384 	return 0;
385 }
386 
387 /**
388  * amdgpu_vmid_grab - allocate the next free VMID
389  *
390  * @vm: vm to allocate id for
391  * @ring: ring we want to submit job to
392  * @job: job who wants to use the VMID
393  * @fence: fence to wait for if no id could be grabbed
394  *
395  * Allocate an id for the vm, adding fences to the sync obj as necessary.
396  */
397 int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
398 		     struct amdgpu_job *job, struct dma_fence **fence)
399 {
400 	struct amdgpu_device *adev = ring->adev;
401 	unsigned vmhub = ring->funcs->vmhub;
402 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
403 	struct amdgpu_vmid *idle = NULL;
404 	struct amdgpu_vmid *id = NULL;
405 	int r = 0;
406 
407 	mutex_lock(&id_mgr->lock);
408 	r = amdgpu_vmid_grab_idle(vm, ring, &idle, fence);
409 	if (r || !idle)
410 		goto error;
411 
412 	if (vm->reserved_vmid[vmhub]) {
413 		r = amdgpu_vmid_grab_reserved(vm, ring, job, &id, fence);
414 		if (r || !id)
415 			goto error;
416 	} else {
417 		r = amdgpu_vmid_grab_used(vm, ring, job, &id, fence);
418 		if (r)
419 			goto error;
420 
421 		if (!id) {
422 			/* Still no ID to use? Then use the idle one found earlier */
423 			id = idle;
424 
425 			/* Remember this submission as user of the VMID */
426 			r = amdgpu_sync_fence(&id->active,
427 					      &job->base.s_fence->finished);
428 			if (r)
429 				goto error;
430 
431 			job->vm_needs_flush = true;
432 		}
433 
434 		list_move_tail(&id->list, &id_mgr->ids_lru);
435 	}
436 
437 	job->gds_switch_needed = amdgpu_vmid_gds_switch_needed(id, job);
438 	if (job->vm_needs_flush) {
439 		id->flushed_updates = amdgpu_vm_tlb_seq(vm);
440 		dma_fence_put(id->last_flush);
441 		id->last_flush = NULL;
442 	}
443 	job->vmid = id - id_mgr->ids;
444 	job->pasid = vm->pasid;
445 
446 	id->gds_base = job->gds_base;
447 	id->gds_size = job->gds_size;
448 	id->gws_base = job->gws_base;
449 	id->gws_size = job->gws_size;
450 	id->oa_base = job->oa_base;
451 	id->oa_size = job->oa_size;
452 	id->pd_gpu_addr = job->vm_pd_addr;
453 	id->owner = vm->immediate.fence_context;
454 
455 	trace_amdgpu_vm_grab_id(vm, ring, job);
456 
457 error:
458 	mutex_unlock(&id_mgr->lock);
459 	return r;
460 }
461 
462 int amdgpu_vmid_alloc_reserved(struct amdgpu_device *adev,
463 			       struct amdgpu_vm *vm,
464 			       unsigned vmhub)
465 {
466 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
467 
468 	mutex_lock(&id_mgr->lock);
469 	if (vm->reserved_vmid[vmhub])
470 		goto unlock;
471 
472 	++id_mgr->reserved_use_count;
473 	if (!id_mgr->reserved) {
474 		struct amdgpu_vmid *id;
475 
476 		id = list_first_entry(&id_mgr->ids_lru, struct amdgpu_vmid,
477 				      list);
478 		/* Remove from normal round robin handling */
479 		list_del_init(&id->list);
480 		id_mgr->reserved = id;
481 	}
482 	vm->reserved_vmid[vmhub] = true;
483 
484 unlock:
485 	mutex_unlock(&id_mgr->lock);
486 	return 0;
487 }
488 
489 void amdgpu_vmid_free_reserved(struct amdgpu_device *adev,
490 			       struct amdgpu_vm *vm,
491 			       unsigned vmhub)
492 {
493 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
494 
495 	mutex_lock(&id_mgr->lock);
496 	if (vm->reserved_vmid[vmhub] &&
497 	    !--id_mgr->reserved_use_count) {
498 		/* give the reserved ID back to normal round robin */
499 		list_add(&id_mgr->reserved->list, &id_mgr->ids_lru);
500 		id_mgr->reserved = NULL;
501 	}
502 	vm->reserved_vmid[vmhub] = false;
503 	mutex_unlock(&id_mgr->lock);
504 }
505 
506 /**
507  * amdgpu_vmid_reset - reset VMID to zero
508  *
509  * @adev: amdgpu device structure
510  * @vmhub: vmhub type
511  * @vmid: vmid number to use
512  *
513  * Reset saved GDW, GWS and OA to force switch on next flush.
514  */
515 void amdgpu_vmid_reset(struct amdgpu_device *adev, unsigned vmhub,
516 		       unsigned vmid)
517 {
518 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
519 	struct amdgpu_vmid *id = &id_mgr->ids[vmid];
520 
521 	mutex_lock(&id_mgr->lock);
522 	id->owner = 0;
523 	id->gds_base = 0;
524 	id->gds_size = 0;
525 	id->gws_base = 0;
526 	id->gws_size = 0;
527 	id->oa_base = 0;
528 	id->oa_size = 0;
529 	mutex_unlock(&id_mgr->lock);
530 }
531 
532 /**
533  * amdgpu_vmid_reset_all - reset VMID to zero
534  *
535  * @adev: amdgpu device structure
536  *
537  * Reset VMID to force flush on next use
538  */
539 void amdgpu_vmid_reset_all(struct amdgpu_device *adev)
540 {
541 	unsigned i, j;
542 
543 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
544 		struct amdgpu_vmid_mgr *id_mgr =
545 			&adev->vm_manager.id_mgr[i];
546 
547 		for (j = 1; j < id_mgr->num_ids; ++j)
548 			amdgpu_vmid_reset(adev, i, j);
549 	}
550 }
551 
552 /**
553  * amdgpu_vmid_mgr_init - init the VMID manager
554  *
555  * @adev: amdgpu_device pointer
556  *
557  * Initialize the VM manager structures
558  */
559 void amdgpu_vmid_mgr_init(struct amdgpu_device *adev)
560 {
561 	unsigned i, j;
562 
563 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
564 		struct amdgpu_vmid_mgr *id_mgr =
565 			&adev->vm_manager.id_mgr[i];
566 
567 		mutex_init(&id_mgr->lock);
568 		INIT_LIST_HEAD(&id_mgr->ids_lru);
569 		id_mgr->reserved_use_count = 0;
570 
571 		/* manage only VMIDs not used by KFD */
572 		id_mgr->num_ids = adev->vm_manager.first_kfd_vmid;
573 
574 		/* skip over VMID 0, since it is the system VM */
575 		for (j = 1; j < id_mgr->num_ids; ++j) {
576 			amdgpu_vmid_reset(adev, i, j);
577 			amdgpu_sync_create(&id_mgr->ids[j].active);
578 			list_add_tail(&id_mgr->ids[j].list, &id_mgr->ids_lru);
579 		}
580 	}
581 }
582 
583 /**
584  * amdgpu_vmid_mgr_fini - cleanup VM manager
585  *
586  * @adev: amdgpu_device pointer
587  *
588  * Cleanup the VM manager and free resources.
589  */
590 void amdgpu_vmid_mgr_fini(struct amdgpu_device *adev)
591 {
592 	unsigned i, j;
593 
594 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
595 		struct amdgpu_vmid_mgr *id_mgr =
596 			&adev->vm_manager.id_mgr[i];
597 
598 		mutex_destroy(&id_mgr->lock);
599 		for (j = 0; j < AMDGPU_NUM_VMID; ++j) {
600 			struct amdgpu_vmid *id = &id_mgr->ids[j];
601 
602 			amdgpu_sync_free(&id->active);
603 			dma_fence_put(id->last_flush);
604 			dma_fence_put(id->pasid_mapping);
605 		}
606 	}
607 }
608