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 	unsigned int 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(unsigned int 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 			       unsigned int pasid)
109 {
110 	struct dma_fence *fence, **fences;
111 	struct amdgpu_pasid_cb *cb;
112 	unsigned count;
113 	int r;
114 
115 	r = dma_resv_get_fences_rcu(resv, NULL, &count, &fences);
116 	if (r)
117 		goto fallback;
118 
119 	if (count == 0) {
120 		amdgpu_pasid_free(pasid);
121 		return;
122 	}
123 
124 	if (count == 1) {
125 		fence = fences[0];
126 		kfree(fences);
127 	} else {
128 		uint64_t context = dma_fence_context_alloc(1);
129 		struct dma_fence_array *array;
130 
131 		array = dma_fence_array_create(count, fences, context,
132 					       1, false);
133 		if (!array) {
134 			kfree(fences);
135 			goto fallback;
136 		}
137 		fence = &array->base;
138 	}
139 
140 	cb = kmalloc(sizeof(*cb), GFP_KERNEL);
141 	if (!cb) {
142 		/* Last resort when we are OOM */
143 		dma_fence_wait(fence, false);
144 		dma_fence_put(fence);
145 		amdgpu_pasid_free(pasid);
146 	} else {
147 		cb->pasid = pasid;
148 		if (dma_fence_add_callback(fence, &cb->cb,
149 					   amdgpu_pasid_free_cb))
150 			amdgpu_pasid_free_cb(fence, &cb->cb);
151 	}
152 
153 	return;
154 
155 fallback:
156 	/* Not enough memory for the delayed delete, as last resort
157 	 * block for all the fences to complete.
158 	 */
159 	dma_resv_wait_timeout_rcu(resv, true, false,
160 					    MAX_SCHEDULE_TIMEOUT);
161 	amdgpu_pasid_free(pasid);
162 }
163 
164 /*
165  * VMID manager
166  *
167  * VMIDs are a per VMHUB identifier for page tables handling.
168  */
169 
170 /**
171  * amdgpu_vmid_had_gpu_reset - check if reset occured since last use
172  *
173  * @adev: amdgpu_device pointer
174  * @id: VMID structure
175  *
176  * Check if GPU reset occured since last use of the VMID.
177  */
178 bool amdgpu_vmid_had_gpu_reset(struct amdgpu_device *adev,
179 			       struct amdgpu_vmid *id)
180 {
181 	return id->current_gpu_reset_count !=
182 		atomic_read(&adev->gpu_reset_counter);
183 }
184 
185 /**
186  * amdgpu_vm_grab_idle - grab idle VMID
187  *
188  * @vm: vm to allocate id for
189  * @ring: ring we want to submit job to
190  * @sync: sync object where we add dependencies
191  * @idle: resulting idle VMID
192  *
193  * Try to find an idle VMID, if none is idle add a fence to wait to the sync
194  * object. Returns -ENOMEM when we are out of memory.
195  */
196 static int amdgpu_vmid_grab_idle(struct amdgpu_vm *vm,
197 				 struct amdgpu_ring *ring,
198 				 struct amdgpu_sync *sync,
199 				 struct amdgpu_vmid **idle)
200 {
201 	struct amdgpu_device *adev = ring->adev;
202 	unsigned vmhub = ring->funcs->vmhub;
203 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
204 	struct dma_fence **fences;
205 	unsigned i;
206 	int r;
207 
208 	if (ring->vmid_wait && !dma_fence_is_signaled(ring->vmid_wait))
209 		return amdgpu_sync_fence(sync, ring->vmid_wait, false);
210 
211 	fences = kmalloc_array(sizeof(void *), id_mgr->num_ids, GFP_KERNEL);
212 	if (!fences)
213 		return -ENOMEM;
214 
215 	/* Check if we have an idle VMID */
216 	i = 0;
217 	list_for_each_entry((*idle), &id_mgr->ids_lru, list) {
218 		fences[i] = amdgpu_sync_peek_fence(&(*idle)->active, ring);
219 		if (!fences[i])
220 			break;
221 		++i;
222 	}
223 
224 	/* If we can't find a idle VMID to use, wait till one becomes available */
225 	if (&(*idle)->list == &id_mgr->ids_lru) {
226 		u64 fence_context = adev->vm_manager.fence_context + ring->idx;
227 		unsigned seqno = ++adev->vm_manager.seqno[ring->idx];
228 		struct dma_fence_array *array;
229 		unsigned j;
230 
231 		*idle = NULL;
232 		for (j = 0; j < i; ++j)
233 			dma_fence_get(fences[j]);
234 
235 		array = dma_fence_array_create(i, fences, fence_context,
236 					       seqno, true);
237 		if (!array) {
238 			for (j = 0; j < i; ++j)
239 				dma_fence_put(fences[j]);
240 			kfree(fences);
241 			return -ENOMEM;
242 		}
243 
244 		r = amdgpu_sync_fence(sync, &array->base, false);
245 		dma_fence_put(ring->vmid_wait);
246 		ring->vmid_wait = &array->base;
247 		return r;
248 	}
249 	kfree(fences);
250 
251 	return 0;
252 }
253 
254 /**
255  * amdgpu_vm_grab_reserved - try to assign reserved VMID
256  *
257  * @vm: vm to allocate id for
258  * @ring: ring we want to submit job to
259  * @sync: sync object where we add dependencies
260  * @fence: fence protecting ID from reuse
261  * @job: job who wants to use the VMID
262  *
263  * Try to assign a reserved VMID.
264  */
265 static int amdgpu_vmid_grab_reserved(struct amdgpu_vm *vm,
266 				     struct amdgpu_ring *ring,
267 				     struct amdgpu_sync *sync,
268 				     struct dma_fence *fence,
269 				     struct amdgpu_job *job,
270 				     struct amdgpu_vmid **id)
271 {
272 	struct amdgpu_device *adev = ring->adev;
273 	unsigned vmhub = ring->funcs->vmhub;
274 	uint64_t fence_context = adev->fence_context + ring->idx;
275 	struct dma_fence *updates = sync->last_vm_update;
276 	bool needs_flush = vm->use_cpu_for_update;
277 	int r = 0;
278 
279 	*id = vm->reserved_vmid[vmhub];
280 	if (updates && (*id)->flushed_updates &&
281 	    updates->context == (*id)->flushed_updates->context &&
282 	    !dma_fence_is_later(updates, (*id)->flushed_updates))
283 	    updates = NULL;
284 
285 	if ((*id)->owner != vm->direct.fence_context ||
286 	    job->vm_pd_addr != (*id)->pd_gpu_addr ||
287 	    updates || !(*id)->last_flush ||
288 	    ((*id)->last_flush->context != fence_context &&
289 	     !dma_fence_is_signaled((*id)->last_flush))) {
290 		struct dma_fence *tmp;
291 
292 		/* to prevent one context starved by another context */
293 		(*id)->pd_gpu_addr = 0;
294 		tmp = amdgpu_sync_peek_fence(&(*id)->active, ring);
295 		if (tmp) {
296 			*id = NULL;
297 			r = amdgpu_sync_fence(sync, tmp, false);
298 			return r;
299 		}
300 		needs_flush = true;
301 	}
302 
303 	/* Good we can use this VMID. Remember this submission as
304 	* user of the VMID.
305 	*/
306 	r = amdgpu_sync_fence(&(*id)->active, fence, false);
307 	if (r)
308 		return r;
309 
310 	if (updates) {
311 		dma_fence_put((*id)->flushed_updates);
312 		(*id)->flushed_updates = dma_fence_get(updates);
313 	}
314 	job->vm_needs_flush = needs_flush;
315 	return 0;
316 }
317 
318 /**
319  * amdgpu_vm_grab_used - try to reuse a VMID
320  *
321  * @vm: vm to allocate id for
322  * @ring: ring we want to submit job to
323  * @sync: sync object where we add dependencies
324  * @fence: fence protecting ID from reuse
325  * @job: job who wants to use the VMID
326  * @id: resulting VMID
327  *
328  * Try to reuse a VMID for this submission.
329  */
330 static int amdgpu_vmid_grab_used(struct amdgpu_vm *vm,
331 				 struct amdgpu_ring *ring,
332 				 struct amdgpu_sync *sync,
333 				 struct dma_fence *fence,
334 				 struct amdgpu_job *job,
335 				 struct amdgpu_vmid **id)
336 {
337 	struct amdgpu_device *adev = ring->adev;
338 	unsigned vmhub = ring->funcs->vmhub;
339 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
340 	uint64_t fence_context = adev->fence_context + ring->idx;
341 	struct dma_fence *updates = sync->last_vm_update;
342 	int r;
343 
344 	job->vm_needs_flush = vm->use_cpu_for_update;
345 
346 	/* Check if we can use a VMID already assigned to this VM */
347 	list_for_each_entry_reverse((*id), &id_mgr->ids_lru, list) {
348 		bool needs_flush = vm->use_cpu_for_update;
349 		struct dma_fence *flushed;
350 
351 		/* Check all the prerequisites to using this VMID */
352 		if ((*id)->owner != vm->direct.fence_context)
353 			continue;
354 
355 		if ((*id)->pd_gpu_addr != job->vm_pd_addr)
356 			continue;
357 
358 		if (!(*id)->last_flush ||
359 		    ((*id)->last_flush->context != fence_context &&
360 		     !dma_fence_is_signaled((*id)->last_flush)))
361 			needs_flush = true;
362 
363 		flushed  = (*id)->flushed_updates;
364 		if (updates && (!flushed || dma_fence_is_later(updates, flushed)))
365 			needs_flush = true;
366 
367 		/* Concurrent flushes are only possible starting with Vega10 and
368 		 * are broken on Navi10 and Navi14.
369 		 */
370 		if (needs_flush && (adev->asic_type < CHIP_VEGA10 ||
371 				    adev->asic_type == CHIP_NAVI10 ||
372 				    adev->asic_type == CHIP_NAVI14))
373 			continue;
374 
375 		/* Good, we can use this VMID. Remember this submission as
376 		 * user of the VMID.
377 		 */
378 		r = amdgpu_sync_fence(&(*id)->active, fence, false);
379 		if (r)
380 			return r;
381 
382 		if (updates && (!flushed || dma_fence_is_later(updates, flushed))) {
383 			dma_fence_put((*id)->flushed_updates);
384 			(*id)->flushed_updates = dma_fence_get(updates);
385 		}
386 
387 		job->vm_needs_flush |= needs_flush;
388 		return 0;
389 	}
390 
391 	*id = NULL;
392 	return 0;
393 }
394 
395 /**
396  * amdgpu_vm_grab_id - allocate the next free VMID
397  *
398  * @vm: vm to allocate id for
399  * @ring: ring we want to submit job to
400  * @sync: sync object where we add dependencies
401  * @fence: fence protecting ID from reuse
402  * @job: job who wants to use the VMID
403  *
404  * Allocate an id for the vm, adding fences to the sync obj as necessary.
405  */
406 int amdgpu_vmid_grab(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
407 		     struct amdgpu_sync *sync, struct dma_fence *fence,
408 		     struct amdgpu_job *job)
409 {
410 	struct amdgpu_device *adev = ring->adev;
411 	unsigned vmhub = ring->funcs->vmhub;
412 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
413 	struct amdgpu_vmid *idle = NULL;
414 	struct amdgpu_vmid *id = NULL;
415 	int r = 0;
416 
417 	mutex_lock(&id_mgr->lock);
418 	r = amdgpu_vmid_grab_idle(vm, ring, sync, &idle);
419 	if (r || !idle)
420 		goto error;
421 
422 	if (vm->reserved_vmid[vmhub]) {
423 		r = amdgpu_vmid_grab_reserved(vm, ring, sync, fence, job, &id);
424 		if (r || !id)
425 			goto error;
426 	} else {
427 		r = amdgpu_vmid_grab_used(vm, ring, sync, fence, job, &id);
428 		if (r)
429 			goto error;
430 
431 		if (!id) {
432 			struct dma_fence *updates = sync->last_vm_update;
433 
434 			/* Still no ID to use? Then use the idle one found earlier */
435 			id = idle;
436 
437 			/* Remember this submission as user of the VMID */
438 			r = amdgpu_sync_fence(&id->active, fence, false);
439 			if (r)
440 				goto error;
441 
442 			dma_fence_put(id->flushed_updates);
443 			id->flushed_updates = dma_fence_get(updates);
444 			job->vm_needs_flush = true;
445 		}
446 
447 		list_move_tail(&id->list, &id_mgr->ids_lru);
448 	}
449 
450 	id->pd_gpu_addr = job->vm_pd_addr;
451 	id->owner = vm->direct.fence_context;
452 
453 	if (job->vm_needs_flush) {
454 		dma_fence_put(id->last_flush);
455 		id->last_flush = NULL;
456 	}
457 	job->vmid = id - id_mgr->ids;
458 	job->pasid = vm->pasid;
459 	trace_amdgpu_vm_grab_id(vm, ring, job);
460 
461 error:
462 	mutex_unlock(&id_mgr->lock);
463 	return r;
464 }
465 
466 int amdgpu_vmid_alloc_reserved(struct amdgpu_device *adev,
467 			       struct amdgpu_vm *vm,
468 			       unsigned vmhub)
469 {
470 	struct amdgpu_vmid_mgr *id_mgr;
471 	struct amdgpu_vmid *idle;
472 	int r = 0;
473 
474 	id_mgr = &adev->vm_manager.id_mgr[vmhub];
475 	mutex_lock(&id_mgr->lock);
476 	if (vm->reserved_vmid[vmhub])
477 		goto unlock;
478 	if (atomic_inc_return(&id_mgr->reserved_vmid_num) >
479 	    AMDGPU_VM_MAX_RESERVED_VMID) {
480 		DRM_ERROR("Over limitation of reserved vmid\n");
481 		atomic_dec(&id_mgr->reserved_vmid_num);
482 		r = -EINVAL;
483 		goto unlock;
484 	}
485 	/* Select the first entry VMID */
486 	idle = list_first_entry(&id_mgr->ids_lru, struct amdgpu_vmid, list);
487 	list_del_init(&idle->list);
488 	vm->reserved_vmid[vmhub] = idle;
489 	mutex_unlock(&id_mgr->lock);
490 
491 	return 0;
492 unlock:
493 	mutex_unlock(&id_mgr->lock);
494 	return r;
495 }
496 
497 void amdgpu_vmid_free_reserved(struct amdgpu_device *adev,
498 			       struct amdgpu_vm *vm,
499 			       unsigned vmhub)
500 {
501 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
502 
503 	mutex_lock(&id_mgr->lock);
504 	if (vm->reserved_vmid[vmhub]) {
505 		list_add(&vm->reserved_vmid[vmhub]->list,
506 			&id_mgr->ids_lru);
507 		vm->reserved_vmid[vmhub] = NULL;
508 		atomic_dec(&id_mgr->reserved_vmid_num);
509 	}
510 	mutex_unlock(&id_mgr->lock);
511 }
512 
513 /**
514  * amdgpu_vmid_reset - reset VMID to zero
515  *
516  * @adev: amdgpu device structure
517  * @vmid: vmid number to use
518  *
519  * Reset saved GDW, GWS and OA to force switch on next flush.
520  */
521 void amdgpu_vmid_reset(struct amdgpu_device *adev, unsigned vmhub,
522 		       unsigned vmid)
523 {
524 	struct amdgpu_vmid_mgr *id_mgr = &adev->vm_manager.id_mgr[vmhub];
525 	struct amdgpu_vmid *id = &id_mgr->ids[vmid];
526 
527 	mutex_lock(&id_mgr->lock);
528 	id->owner = 0;
529 	id->gds_base = 0;
530 	id->gds_size = 0;
531 	id->gws_base = 0;
532 	id->gws_size = 0;
533 	id->oa_base = 0;
534 	id->oa_size = 0;
535 	mutex_unlock(&id_mgr->lock);
536 }
537 
538 /**
539  * amdgpu_vmid_reset_all - reset VMID to zero
540  *
541  * @adev: amdgpu device structure
542  *
543  * Reset VMID to force flush on next use
544  */
545 void amdgpu_vmid_reset_all(struct amdgpu_device *adev)
546 {
547 	unsigned i, j;
548 
549 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
550 		struct amdgpu_vmid_mgr *id_mgr =
551 			&adev->vm_manager.id_mgr[i];
552 
553 		for (j = 1; j < id_mgr->num_ids; ++j)
554 			amdgpu_vmid_reset(adev, i, j);
555 	}
556 }
557 
558 /**
559  * amdgpu_vmid_mgr_init - init the VMID manager
560  *
561  * @adev: amdgpu_device pointer
562  *
563  * Initialize the VM manager structures
564  */
565 void amdgpu_vmid_mgr_init(struct amdgpu_device *adev)
566 {
567 	unsigned i, j;
568 
569 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
570 		struct amdgpu_vmid_mgr *id_mgr =
571 			&adev->vm_manager.id_mgr[i];
572 
573 		mutex_init(&id_mgr->lock);
574 		INIT_LIST_HEAD(&id_mgr->ids_lru);
575 		atomic_set(&id_mgr->reserved_vmid_num, 0);
576 
577 		/* skip over VMID 0, since it is the system VM */
578 		for (j = 1; j < id_mgr->num_ids; ++j) {
579 			amdgpu_vmid_reset(adev, i, j);
580 			amdgpu_sync_create(&id_mgr->ids[j].active);
581 			list_add_tail(&id_mgr->ids[j].list, &id_mgr->ids_lru);
582 		}
583 	}
584 }
585 
586 /**
587  * amdgpu_vmid_mgr_fini - cleanup VM manager
588  *
589  * @adev: amdgpu_device pointer
590  *
591  * Cleanup the VM manager and free resources.
592  */
593 void amdgpu_vmid_mgr_fini(struct amdgpu_device *adev)
594 {
595 	unsigned i, j;
596 
597 	for (i = 0; i < AMDGPU_MAX_VMHUBS; ++i) {
598 		struct amdgpu_vmid_mgr *id_mgr =
599 			&adev->vm_manager.id_mgr[i];
600 
601 		mutex_destroy(&id_mgr->lock);
602 		for (j = 0; j < AMDGPU_NUM_VMID; ++j) {
603 			struct amdgpu_vmid *id = &id_mgr->ids[j];
604 
605 			amdgpu_sync_free(&id->active);
606 			dma_fence_put(id->flushed_updates);
607 			dma_fence_put(id->last_flush);
608 			dma_fence_put(id->pasid_mapping);
609 		}
610 	}
611 }
612