1 // SPDX-License-Identifier: MIT
2 
3 /*
4  * Locking:
5  *
6  * The uvmm mutex protects any operations on the GPU VA space provided by the
7  * DRM GPU VA manager.
8  *
9  * The GEMs dma_resv lock protects the GEMs GPUVA list, hence link/unlink of a
10  * mapping to it's backing GEM must be performed under this lock.
11  *
12  * Actual map/unmap operations within the fence signalling critical path are
13  * protected by installing DMA fences to the corresponding GEMs DMA
14  * reservations, such that concurrent BO moves, which itself walk the GEMs GPUVA
15  * list in order to map/unmap it's entries, can't occur concurrently.
16  *
17  * Accessing the DRM_GPUVA_INVALIDATED flag doesn't need any separate
18  * protection, since there are no accesses other than from BO move callbacks
19  * and from the fence signalling critical path, which are already protected by
20  * the corresponding GEMs DMA reservation fence.
21  */
22 
23 #include "nouveau_drv.h"
24 #include "nouveau_gem.h"
25 #include "nouveau_mem.h"
26 #include "nouveau_uvmm.h"
27 
28 #include <nvif/vmm.h>
29 #include <nvif/mem.h>
30 
31 #include <nvif/class.h>
32 #include <nvif/if000c.h>
33 #include <nvif/if900d.h>
34 
35 #define NOUVEAU_VA_SPACE_BITS		47 /* FIXME */
36 #define NOUVEAU_VA_SPACE_START		0x0
37 #define NOUVEAU_VA_SPACE_END		(1ULL << NOUVEAU_VA_SPACE_BITS)
38 
39 #define list_last_op(_ops) list_last_entry(_ops, struct bind_job_op, entry)
40 #define list_prev_op(_op) list_prev_entry(_op, entry)
41 #define list_for_each_op(_op, _ops) list_for_each_entry(_op, _ops, entry)
42 #define list_for_each_op_from_reverse(_op, _ops) \
43 	list_for_each_entry_from_reverse(_op, _ops, entry)
44 #define list_for_each_op_safe(_op, _n, _ops) list_for_each_entry_safe(_op, _n, _ops, entry)
45 
46 enum vm_bind_op {
47 	OP_MAP = DRM_NOUVEAU_VM_BIND_OP_MAP,
48 	OP_UNMAP = DRM_NOUVEAU_VM_BIND_OP_UNMAP,
49 	OP_MAP_SPARSE,
50 	OP_UNMAP_SPARSE,
51 };
52 
53 struct nouveau_uvma_prealloc {
54 	struct nouveau_uvma *map;
55 	struct nouveau_uvma *prev;
56 	struct nouveau_uvma *next;
57 };
58 
59 struct bind_job_op {
60 	struct list_head entry;
61 
62 	enum vm_bind_op op;
63 	u32 flags;
64 
65 	struct {
66 		u64 addr;
67 		u64 range;
68 	} va;
69 
70 	struct {
71 		u32 handle;
72 		u64 offset;
73 		struct drm_gem_object *obj;
74 	} gem;
75 
76 	struct nouveau_uvma_region *reg;
77 	struct nouveau_uvma_prealloc new;
78 	struct drm_gpuva_ops *ops;
79 };
80 
81 struct uvmm_map_args {
82 	struct nouveau_uvma_region *region;
83 	u64 addr;
84 	u64 range;
85 	u8 kind;
86 };
87 
88 static int
89 nouveau_uvmm_vmm_sparse_ref(struct nouveau_uvmm *uvmm,
90 			    u64 addr, u64 range)
91 {
92 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
93 
94 	return nvif_vmm_raw_sparse(vmm, addr, range, true);
95 }
96 
97 static int
98 nouveau_uvmm_vmm_sparse_unref(struct nouveau_uvmm *uvmm,
99 			      u64 addr, u64 range)
100 {
101 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
102 
103 	return nvif_vmm_raw_sparse(vmm, addr, range, false);
104 }
105 
106 static int
107 nouveau_uvmm_vmm_get(struct nouveau_uvmm *uvmm,
108 		     u64 addr, u64 range)
109 {
110 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
111 
112 	return nvif_vmm_raw_get(vmm, addr, range, PAGE_SHIFT);
113 }
114 
115 static int
116 nouveau_uvmm_vmm_put(struct nouveau_uvmm *uvmm,
117 		     u64 addr, u64 range)
118 {
119 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
120 
121 	return nvif_vmm_raw_put(vmm, addr, range, PAGE_SHIFT);
122 }
123 
124 static int
125 nouveau_uvmm_vmm_unmap(struct nouveau_uvmm *uvmm,
126 		       u64 addr, u64 range, bool sparse)
127 {
128 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
129 
130 	return nvif_vmm_raw_unmap(vmm, addr, range, PAGE_SHIFT, sparse);
131 }
132 
133 static int
134 nouveau_uvmm_vmm_map(struct nouveau_uvmm *uvmm,
135 		     u64 addr, u64 range,
136 		     u64 bo_offset, u8 kind,
137 		     struct nouveau_mem *mem)
138 {
139 	struct nvif_vmm *vmm = &uvmm->vmm.vmm;
140 	union {
141 		struct gf100_vmm_map_v0 gf100;
142 	} args;
143 	u32 argc = 0;
144 
145 	switch (vmm->object.oclass) {
146 	case NVIF_CLASS_VMM_GF100:
147 	case NVIF_CLASS_VMM_GM200:
148 	case NVIF_CLASS_VMM_GP100:
149 		args.gf100.version = 0;
150 		if (mem->mem.type & NVIF_MEM_VRAM)
151 			args.gf100.vol = 0;
152 		else
153 			args.gf100.vol = 1;
154 		args.gf100.ro = 0;
155 		args.gf100.priv = 0;
156 		args.gf100.kind = kind;
157 		argc = sizeof(args.gf100);
158 		break;
159 	default:
160 		WARN_ON(1);
161 		return -ENOSYS;
162 	}
163 
164 	return nvif_vmm_raw_map(vmm, addr, range, PAGE_SHIFT,
165 				&args, argc,
166 				&mem->mem, bo_offset);
167 }
168 
169 static int
170 nouveau_uvma_region_sparse_unref(struct nouveau_uvma_region *reg)
171 {
172 	u64 addr = reg->va.addr;
173 	u64 range = reg->va.range;
174 
175 	return nouveau_uvmm_vmm_sparse_unref(reg->uvmm, addr, range);
176 }
177 
178 static int
179 nouveau_uvma_vmm_put(struct nouveau_uvma *uvma)
180 {
181 	u64 addr = uvma->va.va.addr;
182 	u64 range = uvma->va.va.range;
183 
184 	return nouveau_uvmm_vmm_put(to_uvmm(uvma), addr, range);
185 }
186 
187 static int
188 nouveau_uvma_map(struct nouveau_uvma *uvma,
189 		 struct nouveau_mem *mem)
190 {
191 	u64 addr = uvma->va.va.addr;
192 	u64 offset = uvma->va.gem.offset;
193 	u64 range = uvma->va.va.range;
194 
195 	return nouveau_uvmm_vmm_map(to_uvmm(uvma), addr, range,
196 				    offset, uvma->kind, mem);
197 }
198 
199 static int
200 nouveau_uvma_unmap(struct nouveau_uvma *uvma)
201 {
202 	u64 addr = uvma->va.va.addr;
203 	u64 range = uvma->va.va.range;
204 	bool sparse = !!uvma->region;
205 
206 	if (drm_gpuva_invalidated(&uvma->va))
207 		return 0;
208 
209 	return nouveau_uvmm_vmm_unmap(to_uvmm(uvma), addr, range, sparse);
210 }
211 
212 static int
213 nouveau_uvma_alloc(struct nouveau_uvma **puvma)
214 {
215 	*puvma = kzalloc(sizeof(**puvma), GFP_KERNEL);
216 	if (!*puvma)
217 		return -ENOMEM;
218 
219 	return 0;
220 }
221 
222 static void
223 nouveau_uvma_free(struct nouveau_uvma *uvma)
224 {
225 	kfree(uvma);
226 }
227 
228 static void
229 nouveau_uvma_gem_get(struct nouveau_uvma *uvma)
230 {
231 	drm_gem_object_get(uvma->va.gem.obj);
232 }
233 
234 static void
235 nouveau_uvma_gem_put(struct nouveau_uvma *uvma)
236 {
237 	drm_gem_object_put(uvma->va.gem.obj);
238 }
239 
240 static int
241 nouveau_uvma_region_alloc(struct nouveau_uvma_region **preg)
242 {
243 	*preg = kzalloc(sizeof(**preg), GFP_KERNEL);
244 	if (!*preg)
245 		return -ENOMEM;
246 
247 	kref_init(&(*preg)->kref);
248 
249 	return 0;
250 }
251 
252 static void
253 nouveau_uvma_region_free(struct kref *kref)
254 {
255 	struct nouveau_uvma_region *reg =
256 		container_of(kref, struct nouveau_uvma_region, kref);
257 
258 	kfree(reg);
259 }
260 
261 static void
262 nouveau_uvma_region_get(struct nouveau_uvma_region *reg)
263 {
264 	kref_get(&reg->kref);
265 }
266 
267 static void
268 nouveau_uvma_region_put(struct nouveau_uvma_region *reg)
269 {
270 	kref_put(&reg->kref, nouveau_uvma_region_free);
271 }
272 
273 static int
274 __nouveau_uvma_region_insert(struct nouveau_uvmm *uvmm,
275 			     struct nouveau_uvma_region *reg)
276 {
277 	u64 addr = reg->va.addr;
278 	u64 range = reg->va.range;
279 	u64 last = addr + range - 1;
280 	MA_STATE(mas, &uvmm->region_mt, addr, addr);
281 
282 	if (unlikely(mas_walk(&mas)))
283 		return -EEXIST;
284 
285 	if (unlikely(mas.last < last))
286 		return -EEXIST;
287 
288 	mas.index = addr;
289 	mas.last = last;
290 
291 	mas_store_gfp(&mas, reg, GFP_KERNEL);
292 
293 	reg->uvmm = uvmm;
294 
295 	return 0;
296 }
297 
298 static int
299 nouveau_uvma_region_insert(struct nouveau_uvmm *uvmm,
300 			   struct nouveau_uvma_region *reg,
301 			   u64 addr, u64 range)
302 {
303 	int ret;
304 
305 	reg->uvmm = uvmm;
306 	reg->va.addr = addr;
307 	reg->va.range = range;
308 
309 	ret = __nouveau_uvma_region_insert(uvmm, reg);
310 	if (ret)
311 		return ret;
312 
313 	return 0;
314 }
315 
316 static void
317 nouveau_uvma_region_remove(struct nouveau_uvma_region *reg)
318 {
319 	struct nouveau_uvmm *uvmm = reg->uvmm;
320 	MA_STATE(mas, &uvmm->region_mt, reg->va.addr, 0);
321 
322 	mas_erase(&mas);
323 }
324 
325 static int
326 nouveau_uvma_region_create(struct nouveau_uvmm *uvmm,
327 			   u64 addr, u64 range)
328 {
329 	struct nouveau_uvma_region *reg;
330 	int ret;
331 
332 	if (!drm_gpuva_interval_empty(&uvmm->umgr, addr, range))
333 		return -ENOSPC;
334 
335 	ret = nouveau_uvma_region_alloc(&reg);
336 	if (ret)
337 		return ret;
338 
339 	ret = nouveau_uvma_region_insert(uvmm, reg, addr, range);
340 	if (ret)
341 		goto err_free_region;
342 
343 	ret = nouveau_uvmm_vmm_sparse_ref(uvmm, addr, range);
344 	if (ret)
345 		goto err_region_remove;
346 
347 	return 0;
348 
349 err_region_remove:
350 	nouveau_uvma_region_remove(reg);
351 err_free_region:
352 	nouveau_uvma_region_put(reg);
353 	return ret;
354 }
355 
356 static struct nouveau_uvma_region *
357 nouveau_uvma_region_find_first(struct nouveau_uvmm *uvmm,
358 			       u64 addr, u64 range)
359 {
360 	MA_STATE(mas, &uvmm->region_mt, addr, 0);
361 
362 	return mas_find(&mas, addr + range - 1);
363 }
364 
365 static struct nouveau_uvma_region *
366 nouveau_uvma_region_find(struct nouveau_uvmm *uvmm,
367 			 u64 addr, u64 range)
368 {
369 	struct nouveau_uvma_region *reg;
370 
371 	reg = nouveau_uvma_region_find_first(uvmm, addr, range);
372 	if (!reg)
373 		return NULL;
374 
375 	if (reg->va.addr != addr ||
376 	    reg->va.range != range)
377 		return NULL;
378 
379 	return reg;
380 }
381 
382 static bool
383 nouveau_uvma_region_empty(struct nouveau_uvma_region *reg)
384 {
385 	struct nouveau_uvmm *uvmm = reg->uvmm;
386 
387 	return drm_gpuva_interval_empty(&uvmm->umgr,
388 					reg->va.addr,
389 					reg->va.range);
390 }
391 
392 static int
393 __nouveau_uvma_region_destroy(struct nouveau_uvma_region *reg)
394 {
395 	struct nouveau_uvmm *uvmm = reg->uvmm;
396 	u64 addr = reg->va.addr;
397 	u64 range = reg->va.range;
398 
399 	if (!nouveau_uvma_region_empty(reg))
400 		return -EBUSY;
401 
402 	nouveau_uvma_region_remove(reg);
403 	nouveau_uvmm_vmm_sparse_unref(uvmm, addr, range);
404 	nouveau_uvma_region_put(reg);
405 
406 	return 0;
407 }
408 
409 static int
410 nouveau_uvma_region_destroy(struct nouveau_uvmm *uvmm,
411 			    u64 addr, u64 range)
412 {
413 	struct nouveau_uvma_region *reg;
414 
415 	reg = nouveau_uvma_region_find(uvmm, addr, range);
416 	if (!reg)
417 		return -ENOENT;
418 
419 	return __nouveau_uvma_region_destroy(reg);
420 }
421 
422 static void
423 nouveau_uvma_region_dirty(struct nouveau_uvma_region *reg)
424 {
425 
426 	init_completion(&reg->complete);
427 	reg->dirty = true;
428 }
429 
430 static void
431 nouveau_uvma_region_complete(struct nouveau_uvma_region *reg)
432 {
433 	complete_all(&reg->complete);
434 }
435 
436 static void
437 op_map_prepare_unwind(struct nouveau_uvma *uvma)
438 {
439 	nouveau_uvma_gem_put(uvma);
440 	drm_gpuva_remove(&uvma->va);
441 	nouveau_uvma_free(uvma);
442 }
443 
444 static void
445 op_unmap_prepare_unwind(struct drm_gpuva *va)
446 {
447 	drm_gpuva_insert(va->mgr, va);
448 }
449 
450 static void
451 nouveau_uvmm_sm_prepare_unwind(struct nouveau_uvmm *uvmm,
452 			       struct nouveau_uvma_prealloc *new,
453 			       struct drm_gpuva_ops *ops,
454 			       struct drm_gpuva_op *last,
455 			       struct uvmm_map_args *args)
456 {
457 	struct drm_gpuva_op *op = last;
458 	u64 vmm_get_start = args ? args->addr : 0;
459 	u64 vmm_get_end = args ? args->addr + args->range : 0;
460 
461 	/* Unwind GPUVA space. */
462 	drm_gpuva_for_each_op_from_reverse(op, ops) {
463 		switch (op->op) {
464 		case DRM_GPUVA_OP_MAP:
465 			op_map_prepare_unwind(new->map);
466 			break;
467 		case DRM_GPUVA_OP_REMAP: {
468 			struct drm_gpuva_op_remap *r = &op->remap;
469 
470 			if (r->next)
471 				op_map_prepare_unwind(new->next);
472 
473 			if (r->prev)
474 				op_map_prepare_unwind(new->prev);
475 
476 			op_unmap_prepare_unwind(r->unmap->va);
477 			break;
478 		}
479 		case DRM_GPUVA_OP_UNMAP:
480 			op_unmap_prepare_unwind(op->unmap.va);
481 			break;
482 		default:
483 			break;
484 		}
485 	}
486 
487 	/* Unmap operation don't allocate page tables, hence skip the following
488 	 * page table unwind.
489 	 */
490 	if (!args)
491 		return;
492 
493 	drm_gpuva_for_each_op(op, ops) {
494 		switch (op->op) {
495 		case DRM_GPUVA_OP_MAP: {
496 			u64 vmm_get_range = vmm_get_end - vmm_get_start;
497 
498 			if (vmm_get_range)
499 				nouveau_uvmm_vmm_put(uvmm, vmm_get_start,
500 						     vmm_get_range);
501 			break;
502 		}
503 		case DRM_GPUVA_OP_REMAP: {
504 			struct drm_gpuva_op_remap *r = &op->remap;
505 			struct drm_gpuva *va = r->unmap->va;
506 			u64 ustart = va->va.addr;
507 			u64 urange = va->va.range;
508 			u64 uend = ustart + urange;
509 
510 			if (r->prev)
511 				vmm_get_start = uend;
512 
513 			if (r->next)
514 				vmm_get_end = ustart;
515 
516 			if (r->prev && r->next)
517 				vmm_get_start = vmm_get_end = 0;
518 
519 			break;
520 		}
521 		case DRM_GPUVA_OP_UNMAP: {
522 			struct drm_gpuva_op_unmap *u = &op->unmap;
523 			struct drm_gpuva *va = u->va;
524 			u64 ustart = va->va.addr;
525 			u64 urange = va->va.range;
526 			u64 uend = ustart + urange;
527 
528 			/* Nothing to do for mappings we merge with. */
529 			if (uend == vmm_get_start ||
530 			    ustart == vmm_get_end)
531 				break;
532 
533 			if (ustart > vmm_get_start) {
534 				u64 vmm_get_range = ustart - vmm_get_start;
535 
536 				nouveau_uvmm_vmm_put(uvmm, vmm_get_start,
537 						     vmm_get_range);
538 			}
539 			vmm_get_start = uend;
540 			break;
541 		}
542 		default:
543 			break;
544 		}
545 
546 		if (op == last)
547 			break;
548 	}
549 }
550 
551 static void
552 nouveau_uvmm_sm_map_prepare_unwind(struct nouveau_uvmm *uvmm,
553 				   struct nouveau_uvma_prealloc *new,
554 				   struct drm_gpuva_ops *ops,
555 				   u64 addr, u64 range)
556 {
557 	struct drm_gpuva_op *last = drm_gpuva_last_op(ops);
558 	struct uvmm_map_args args = {
559 		.addr = addr,
560 		.range = range,
561 	};
562 
563 	nouveau_uvmm_sm_prepare_unwind(uvmm, new, ops, last, &args);
564 }
565 
566 static void
567 nouveau_uvmm_sm_unmap_prepare_unwind(struct nouveau_uvmm *uvmm,
568 				     struct nouveau_uvma_prealloc *new,
569 				     struct drm_gpuva_ops *ops)
570 {
571 	struct drm_gpuva_op *last = drm_gpuva_last_op(ops);
572 
573 	nouveau_uvmm_sm_prepare_unwind(uvmm, new, ops, last, NULL);
574 }
575 
576 static int
577 op_map_prepare(struct nouveau_uvmm *uvmm,
578 	       struct nouveau_uvma **puvma,
579 	       struct drm_gpuva_op_map *op,
580 	       struct uvmm_map_args *args)
581 {
582 	struct nouveau_uvma *uvma;
583 	int ret;
584 
585 	ret = nouveau_uvma_alloc(&uvma);
586 	if (ret)
587 		return ret;
588 
589 	uvma->region = args->region;
590 	uvma->kind = args->kind;
591 
592 	drm_gpuva_map(&uvmm->umgr, &uvma->va, op);
593 
594 	/* Keep a reference until this uvma is destroyed. */
595 	nouveau_uvma_gem_get(uvma);
596 
597 	*puvma = uvma;
598 	return 0;
599 }
600 
601 static void
602 op_unmap_prepare(struct drm_gpuva_op_unmap *u)
603 {
604 	drm_gpuva_unmap(u);
605 }
606 
607 static int
608 nouveau_uvmm_sm_prepare(struct nouveau_uvmm *uvmm,
609 			struct nouveau_uvma_prealloc *new,
610 			struct drm_gpuva_ops *ops,
611 			struct uvmm_map_args *args)
612 {
613 	struct drm_gpuva_op *op;
614 	u64 vmm_get_start = args ? args->addr : 0;
615 	u64 vmm_get_end = args ? args->addr + args->range : 0;
616 	int ret;
617 
618 	drm_gpuva_for_each_op(op, ops) {
619 		switch (op->op) {
620 		case DRM_GPUVA_OP_MAP: {
621 			u64 vmm_get_range = vmm_get_end - vmm_get_start;
622 
623 			ret = op_map_prepare(uvmm, &new->map, &op->map, args);
624 			if (ret)
625 				goto unwind;
626 
627 			if (args && vmm_get_range) {
628 				ret = nouveau_uvmm_vmm_get(uvmm, vmm_get_start,
629 							   vmm_get_range);
630 				if (ret) {
631 					op_map_prepare_unwind(new->map);
632 					goto unwind;
633 				}
634 			}
635 			break;
636 		}
637 		case DRM_GPUVA_OP_REMAP: {
638 			struct drm_gpuva_op_remap *r = &op->remap;
639 			struct drm_gpuva *va = r->unmap->va;
640 			struct uvmm_map_args remap_args = {
641 				.kind = uvma_from_va(va)->kind,
642 				.region = uvma_from_va(va)->region,
643 			};
644 			u64 ustart = va->va.addr;
645 			u64 urange = va->va.range;
646 			u64 uend = ustart + urange;
647 
648 			op_unmap_prepare(r->unmap);
649 
650 			if (r->prev) {
651 				ret = op_map_prepare(uvmm, &new->prev, r->prev,
652 						     &remap_args);
653 				if (ret)
654 					goto unwind;
655 
656 				if (args)
657 					vmm_get_start = uend;
658 			}
659 
660 			if (r->next) {
661 				ret = op_map_prepare(uvmm, &new->next, r->next,
662 						     &remap_args);
663 				if (ret) {
664 					if (r->prev)
665 						op_map_prepare_unwind(new->prev);
666 					goto unwind;
667 				}
668 
669 				if (args)
670 					vmm_get_end = ustart;
671 			}
672 
673 			if (args && (r->prev && r->next))
674 				vmm_get_start = vmm_get_end = 0;
675 
676 			break;
677 		}
678 		case DRM_GPUVA_OP_UNMAP: {
679 			struct drm_gpuva_op_unmap *u = &op->unmap;
680 			struct drm_gpuva *va = u->va;
681 			u64 ustart = va->va.addr;
682 			u64 urange = va->va.range;
683 			u64 uend = ustart + urange;
684 
685 			op_unmap_prepare(u);
686 
687 			if (!args)
688 				break;
689 
690 			/* Nothing to do for mappings we merge with. */
691 			if (uend == vmm_get_start ||
692 			    ustart == vmm_get_end)
693 				break;
694 
695 			if (ustart > vmm_get_start) {
696 				u64 vmm_get_range = ustart - vmm_get_start;
697 
698 				ret = nouveau_uvmm_vmm_get(uvmm, vmm_get_start,
699 							   vmm_get_range);
700 				if (ret) {
701 					op_unmap_prepare_unwind(va);
702 					goto unwind;
703 				}
704 			}
705 			vmm_get_start = uend;
706 
707 			break;
708 		}
709 		default:
710 			ret = -EINVAL;
711 			goto unwind;
712 		}
713 	}
714 
715 	return 0;
716 
717 unwind:
718 	if (op != drm_gpuva_first_op(ops))
719 		nouveau_uvmm_sm_prepare_unwind(uvmm, new, ops,
720 					       drm_gpuva_prev_op(op),
721 					       args);
722 	return ret;
723 }
724 
725 static int
726 nouveau_uvmm_sm_map_prepare(struct nouveau_uvmm *uvmm,
727 			    struct nouveau_uvma_prealloc *new,
728 			    struct nouveau_uvma_region *region,
729 			    struct drm_gpuva_ops *ops,
730 			    u64 addr, u64 range, u8 kind)
731 {
732 	struct uvmm_map_args args = {
733 		.region = region,
734 		.addr = addr,
735 		.range = range,
736 		.kind = kind,
737 	};
738 
739 	return nouveau_uvmm_sm_prepare(uvmm, new, ops, &args);
740 }
741 
742 static int
743 nouveau_uvmm_sm_unmap_prepare(struct nouveau_uvmm *uvmm,
744 			      struct nouveau_uvma_prealloc *new,
745 			      struct drm_gpuva_ops *ops)
746 {
747 	return nouveau_uvmm_sm_prepare(uvmm, new, ops, NULL);
748 }
749 
750 static struct drm_gem_object *
751 op_gem_obj(struct drm_gpuva_op *op)
752 {
753 	switch (op->op) {
754 	case DRM_GPUVA_OP_MAP:
755 		return op->map.gem.obj;
756 	case DRM_GPUVA_OP_REMAP:
757 		/* Actually, we're looking for the GEMs backing remap.prev and
758 		 * remap.next, but since this is a remap they're identical to
759 		 * the GEM backing the unmapped GPUVA.
760 		 */
761 		return op->remap.unmap->va->gem.obj;
762 	case DRM_GPUVA_OP_UNMAP:
763 		return op->unmap.va->gem.obj;
764 	default:
765 		WARN(1, "Unknown operation.\n");
766 		return NULL;
767 	}
768 }
769 
770 static void
771 op_map(struct nouveau_uvma *uvma)
772 {
773 	struct nouveau_bo *nvbo = nouveau_gem_object(uvma->va.gem.obj);
774 
775 	nouveau_uvma_map(uvma, nouveau_mem(nvbo->bo.resource));
776 }
777 
778 static void
779 op_unmap(struct drm_gpuva_op_unmap *u)
780 {
781 	struct drm_gpuva *va = u->va;
782 	struct nouveau_uvma *uvma = uvma_from_va(va);
783 
784 	/* nouveau_uvma_unmap() does not unmap if backing BO is evicted. */
785 	if (!u->keep)
786 		nouveau_uvma_unmap(uvma);
787 }
788 
789 static void
790 op_unmap_range(struct drm_gpuva_op_unmap *u,
791 	       u64 addr, u64 range)
792 {
793 	struct nouveau_uvma *uvma = uvma_from_va(u->va);
794 	bool sparse = !!uvma->region;
795 
796 	if (!drm_gpuva_invalidated(u->va))
797 		nouveau_uvmm_vmm_unmap(to_uvmm(uvma), addr, range, sparse);
798 }
799 
800 static void
801 op_remap(struct drm_gpuva_op_remap *r,
802 	 struct nouveau_uvma_prealloc *new)
803 {
804 	struct drm_gpuva_op_unmap *u = r->unmap;
805 	struct nouveau_uvma *uvma = uvma_from_va(u->va);
806 	u64 addr = uvma->va.va.addr;
807 	u64 range = uvma->va.va.range;
808 
809 	if (r->prev)
810 		addr = r->prev->va.addr + r->prev->va.range;
811 
812 	if (r->next)
813 		range = r->next->va.addr - addr;
814 
815 	op_unmap_range(u, addr, range);
816 }
817 
818 static int
819 nouveau_uvmm_sm(struct nouveau_uvmm *uvmm,
820 		struct nouveau_uvma_prealloc *new,
821 		struct drm_gpuva_ops *ops)
822 {
823 	struct drm_gpuva_op *op;
824 
825 	drm_gpuva_for_each_op(op, ops) {
826 		switch (op->op) {
827 		case DRM_GPUVA_OP_MAP:
828 			op_map(new->map);
829 			break;
830 		case DRM_GPUVA_OP_REMAP:
831 			op_remap(&op->remap, new);
832 			break;
833 		case DRM_GPUVA_OP_UNMAP:
834 			op_unmap(&op->unmap);
835 			break;
836 		default:
837 			break;
838 		}
839 	}
840 
841 	return 0;
842 }
843 
844 static int
845 nouveau_uvmm_sm_map(struct nouveau_uvmm *uvmm,
846 		    struct nouveau_uvma_prealloc *new,
847 		    struct drm_gpuva_ops *ops)
848 {
849 	return nouveau_uvmm_sm(uvmm, new, ops);
850 }
851 
852 static int
853 nouveau_uvmm_sm_unmap(struct nouveau_uvmm *uvmm,
854 		      struct nouveau_uvma_prealloc *new,
855 		      struct drm_gpuva_ops *ops)
856 {
857 	return nouveau_uvmm_sm(uvmm, new, ops);
858 }
859 
860 static void
861 nouveau_uvmm_sm_cleanup(struct nouveau_uvmm *uvmm,
862 			struct nouveau_uvma_prealloc *new,
863 			struct drm_gpuva_ops *ops, bool unmap)
864 {
865 	struct drm_gpuva_op *op;
866 
867 	drm_gpuva_for_each_op(op, ops) {
868 		switch (op->op) {
869 		case DRM_GPUVA_OP_MAP:
870 			break;
871 		case DRM_GPUVA_OP_REMAP: {
872 			struct drm_gpuva_op_remap *r = &op->remap;
873 			struct drm_gpuva_op_map *p = r->prev;
874 			struct drm_gpuva_op_map *n = r->next;
875 			struct drm_gpuva *va = r->unmap->va;
876 			struct nouveau_uvma *uvma = uvma_from_va(va);
877 
878 			if (unmap) {
879 				u64 addr = va->va.addr;
880 				u64 end = addr + va->va.range;
881 
882 				if (p)
883 					addr = p->va.addr + p->va.range;
884 
885 				if (n)
886 					end = n->va.addr;
887 
888 				nouveau_uvmm_vmm_put(uvmm, addr, end - addr);
889 			}
890 
891 			nouveau_uvma_gem_put(uvma);
892 			nouveau_uvma_free(uvma);
893 			break;
894 		}
895 		case DRM_GPUVA_OP_UNMAP: {
896 			struct drm_gpuva_op_unmap *u = &op->unmap;
897 			struct drm_gpuva *va = u->va;
898 			struct nouveau_uvma *uvma = uvma_from_va(va);
899 
900 			if (unmap)
901 				nouveau_uvma_vmm_put(uvma);
902 
903 			nouveau_uvma_gem_put(uvma);
904 			nouveau_uvma_free(uvma);
905 			break;
906 		}
907 		default:
908 			break;
909 		}
910 	}
911 }
912 
913 static void
914 nouveau_uvmm_sm_map_cleanup(struct nouveau_uvmm *uvmm,
915 			    struct nouveau_uvma_prealloc *new,
916 			    struct drm_gpuva_ops *ops)
917 {
918 	nouveau_uvmm_sm_cleanup(uvmm, new, ops, false);
919 }
920 
921 static void
922 nouveau_uvmm_sm_unmap_cleanup(struct nouveau_uvmm *uvmm,
923 			      struct nouveau_uvma_prealloc *new,
924 			      struct drm_gpuva_ops *ops)
925 {
926 	nouveau_uvmm_sm_cleanup(uvmm, new, ops, true);
927 }
928 
929 static int
930 nouveau_uvmm_validate_range(struct nouveau_uvmm *uvmm, u64 addr, u64 range)
931 {
932 	u64 end = addr + range;
933 	u64 kernel_managed_end = uvmm->kernel_managed_addr +
934 				 uvmm->kernel_managed_size;
935 
936 	if (addr & ~PAGE_MASK)
937 		return -EINVAL;
938 
939 	if (range & ~PAGE_MASK)
940 		return -EINVAL;
941 
942 	if (end <= addr)
943 		return -EINVAL;
944 
945 	if (addr < NOUVEAU_VA_SPACE_START ||
946 	    end > NOUVEAU_VA_SPACE_END)
947 		return -EINVAL;
948 
949 	if (addr < kernel_managed_end &&
950 	    end > uvmm->kernel_managed_addr)
951 		return -EINVAL;
952 
953 	return 0;
954 }
955 
956 static int
957 nouveau_uvmm_bind_job_alloc(struct nouveau_uvmm_bind_job **pjob)
958 {
959 	*pjob = kzalloc(sizeof(**pjob), GFP_KERNEL);
960 	if (!*pjob)
961 		return -ENOMEM;
962 
963 	kref_init(&(*pjob)->kref);
964 
965 	return 0;
966 }
967 
968 static void
969 nouveau_uvmm_bind_job_free(struct kref *kref)
970 {
971 	struct nouveau_uvmm_bind_job *job =
972 		container_of(kref, struct nouveau_uvmm_bind_job, kref);
973 
974 	nouveau_job_free(&job->base);
975 	kfree(job);
976 }
977 
978 static void
979 nouveau_uvmm_bind_job_get(struct nouveau_uvmm_bind_job *job)
980 {
981 	kref_get(&job->kref);
982 }
983 
984 static void
985 nouveau_uvmm_bind_job_put(struct nouveau_uvmm_bind_job *job)
986 {
987 	kref_put(&job->kref, nouveau_uvmm_bind_job_free);
988 }
989 
990 static int
991 bind_validate_op(struct nouveau_job *job,
992 		 struct bind_job_op *op)
993 {
994 	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(job->cli);
995 	struct drm_gem_object *obj = op->gem.obj;
996 
997 	if (op->op == OP_MAP) {
998 		if (op->gem.offset & ~PAGE_MASK)
999 			return -EINVAL;
1000 
1001 		if (obj->size <= op->gem.offset)
1002 			return -EINVAL;
1003 
1004 		if (op->va.range > (obj->size - op->gem.offset))
1005 			return -EINVAL;
1006 	}
1007 
1008 	return nouveau_uvmm_validate_range(uvmm, op->va.addr, op->va.range);
1009 }
1010 
1011 static void
1012 bind_validate_map_sparse(struct nouveau_job *job, u64 addr, u64 range)
1013 {
1014 	struct nouveau_uvmm_bind_job *bind_job;
1015 	struct nouveau_sched_entity *entity = job->entity;
1016 	struct bind_job_op *op;
1017 	u64 end = addr + range;
1018 
1019 again:
1020 	spin_lock(&entity->job.list.lock);
1021 	list_for_each_entry(bind_job, &entity->job.list.head, entry) {
1022 		list_for_each_op(op, &bind_job->ops) {
1023 			if (op->op == OP_UNMAP) {
1024 				u64 op_addr = op->va.addr;
1025 				u64 op_end = op_addr + op->va.range;
1026 
1027 				if (!(end <= op_addr || addr >= op_end)) {
1028 					nouveau_uvmm_bind_job_get(bind_job);
1029 					spin_unlock(&entity->job.list.lock);
1030 					wait_for_completion(&bind_job->complete);
1031 					nouveau_uvmm_bind_job_put(bind_job);
1032 					goto again;
1033 				}
1034 			}
1035 		}
1036 	}
1037 	spin_unlock(&entity->job.list.lock);
1038 }
1039 
1040 static int
1041 bind_validate_map_common(struct nouveau_job *job, u64 addr, u64 range,
1042 			 bool sparse)
1043 {
1044 	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(job->cli);
1045 	struct nouveau_uvma_region *reg;
1046 	u64 reg_addr, reg_end;
1047 	u64 end = addr + range;
1048 
1049 again:
1050 	nouveau_uvmm_lock(uvmm);
1051 	reg = nouveau_uvma_region_find_first(uvmm, addr, range);
1052 	if (!reg) {
1053 		nouveau_uvmm_unlock(uvmm);
1054 		return 0;
1055 	}
1056 
1057 	/* Generally, job submits are serialized, hence only
1058 	 * dirty regions can be modified concurrently.
1059 	 */
1060 	if (reg->dirty) {
1061 		nouveau_uvma_region_get(reg);
1062 		nouveau_uvmm_unlock(uvmm);
1063 		wait_for_completion(&reg->complete);
1064 		nouveau_uvma_region_put(reg);
1065 		goto again;
1066 	}
1067 	nouveau_uvmm_unlock(uvmm);
1068 
1069 	if (sparse)
1070 		return -ENOSPC;
1071 
1072 	reg_addr = reg->va.addr;
1073 	reg_end = reg_addr + reg->va.range;
1074 
1075 	/* Make sure the mapping is either outside of a
1076 	 * region or fully enclosed by a region.
1077 	 */
1078 	if (reg_addr > addr || reg_end < end)
1079 		return -ENOSPC;
1080 
1081 	return 0;
1082 }
1083 
1084 static int
1085 bind_validate_region(struct nouveau_job *job)
1086 {
1087 	struct nouveau_uvmm_bind_job *bind_job = to_uvmm_bind_job(job);
1088 	struct bind_job_op *op;
1089 	int ret;
1090 
1091 	list_for_each_op(op, &bind_job->ops) {
1092 		u64 op_addr = op->va.addr;
1093 		u64 op_range = op->va.range;
1094 		bool sparse = false;
1095 
1096 		switch (op->op) {
1097 		case OP_MAP_SPARSE:
1098 			sparse = true;
1099 			bind_validate_map_sparse(job, op_addr, op_range);
1100 			fallthrough;
1101 		case OP_MAP:
1102 			ret = bind_validate_map_common(job, op_addr, op_range,
1103 						       sparse);
1104 			if (ret)
1105 				return ret;
1106 			break;
1107 		default:
1108 			break;
1109 		}
1110 	}
1111 
1112 	return 0;
1113 }
1114 
1115 static void
1116 bind_link_gpuvas(struct drm_gpuva_ops *ops, struct nouveau_uvma_prealloc *new)
1117 {
1118 	struct drm_gpuva_op *op;
1119 
1120 	drm_gpuva_for_each_op(op, ops) {
1121 		switch (op->op) {
1122 		case DRM_GPUVA_OP_MAP:
1123 			drm_gpuva_link(&new->map->va);
1124 			break;
1125 		case DRM_GPUVA_OP_REMAP:
1126 			if (op->remap.prev)
1127 				drm_gpuva_link(&new->prev->va);
1128 			if (op->remap.next)
1129 				drm_gpuva_link(&new->next->va);
1130 			drm_gpuva_unlink(op->remap.unmap->va);
1131 			break;
1132 		case DRM_GPUVA_OP_UNMAP:
1133 			drm_gpuva_unlink(op->unmap.va);
1134 			break;
1135 		default:
1136 			break;
1137 		}
1138 	}
1139 }
1140 
1141 static int
1142 nouveau_uvmm_bind_job_submit(struct nouveau_job *job)
1143 {
1144 	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(job->cli);
1145 	struct nouveau_uvmm_bind_job *bind_job = to_uvmm_bind_job(job);
1146 	struct nouveau_sched_entity *entity = job->entity;
1147 	struct drm_exec *exec = &job->exec;
1148 	struct bind_job_op *op;
1149 	int ret;
1150 
1151 	list_for_each_op(op, &bind_job->ops) {
1152 		if (op->op == OP_MAP) {
1153 			op->gem.obj = drm_gem_object_lookup(job->file_priv,
1154 							    op->gem.handle);
1155 			if (!op->gem.obj)
1156 				return -ENOENT;
1157 		}
1158 
1159 		ret = bind_validate_op(job, op);
1160 		if (ret)
1161 			return ret;
1162 	}
1163 
1164 	/* If a sparse region or mapping overlaps a dirty region, we need to
1165 	 * wait for the region to complete the unbind process. This is due to
1166 	 * how page table management is currently implemented. A future
1167 	 * implementation might change this.
1168 	 */
1169 	ret = bind_validate_region(job);
1170 	if (ret)
1171 		return ret;
1172 
1173 	/* Once we start modifying the GPU VA space we need to keep holding the
1174 	 * uvmm lock until we can't fail anymore. This is due to the set of GPU
1175 	 * VA space changes must appear atomically and we need to be able to
1176 	 * unwind all GPU VA space changes on failure.
1177 	 */
1178 	nouveau_uvmm_lock(uvmm);
1179 	list_for_each_op(op, &bind_job->ops) {
1180 		switch (op->op) {
1181 		case OP_MAP_SPARSE:
1182 			ret = nouveau_uvma_region_create(uvmm,
1183 							 op->va.addr,
1184 							 op->va.range);
1185 			if (ret)
1186 				goto unwind_continue;
1187 
1188 			break;
1189 		case OP_UNMAP_SPARSE:
1190 			op->reg = nouveau_uvma_region_find(uvmm, op->va.addr,
1191 							   op->va.range);
1192 			if (!op->reg || op->reg->dirty) {
1193 				ret = -ENOENT;
1194 				goto unwind_continue;
1195 			}
1196 
1197 			op->ops = drm_gpuva_sm_unmap_ops_create(&uvmm->umgr,
1198 								op->va.addr,
1199 								op->va.range);
1200 			if (IS_ERR(op->ops)) {
1201 				ret = PTR_ERR(op->ops);
1202 				goto unwind_continue;
1203 			}
1204 
1205 			ret = nouveau_uvmm_sm_unmap_prepare(uvmm, &op->new,
1206 							    op->ops);
1207 			if (ret) {
1208 				drm_gpuva_ops_free(&uvmm->umgr, op->ops);
1209 				op->ops = NULL;
1210 				op->reg = NULL;
1211 				goto unwind_continue;
1212 			}
1213 
1214 			nouveau_uvma_region_dirty(op->reg);
1215 
1216 			break;
1217 		case OP_MAP: {
1218 			struct nouveau_uvma_region *reg;
1219 
1220 			reg = nouveau_uvma_region_find_first(uvmm,
1221 							     op->va.addr,
1222 							     op->va.range);
1223 			if (reg) {
1224 				u64 reg_addr = reg->va.addr;
1225 				u64 reg_end = reg_addr + reg->va.range;
1226 				u64 op_addr = op->va.addr;
1227 				u64 op_end = op_addr + op->va.range;
1228 
1229 				if (unlikely(reg->dirty)) {
1230 					ret = -EINVAL;
1231 					goto unwind_continue;
1232 				}
1233 
1234 				/* Make sure the mapping is either outside of a
1235 				 * region or fully enclosed by a region.
1236 				 */
1237 				if (reg_addr > op_addr || reg_end < op_end) {
1238 					ret = -ENOSPC;
1239 					goto unwind_continue;
1240 				}
1241 			}
1242 
1243 			op->ops = drm_gpuva_sm_map_ops_create(&uvmm->umgr,
1244 							      op->va.addr,
1245 							      op->va.range,
1246 							      op->gem.obj,
1247 							      op->gem.offset);
1248 			if (IS_ERR(op->ops)) {
1249 				ret = PTR_ERR(op->ops);
1250 				goto unwind_continue;
1251 			}
1252 
1253 			ret = nouveau_uvmm_sm_map_prepare(uvmm, &op->new,
1254 							  reg, op->ops,
1255 							  op->va.addr,
1256 							  op->va.range,
1257 							  op->flags & 0xff);
1258 			if (ret) {
1259 				drm_gpuva_ops_free(&uvmm->umgr, op->ops);
1260 				op->ops = NULL;
1261 				goto unwind_continue;
1262 			}
1263 
1264 			break;
1265 		}
1266 		case OP_UNMAP:
1267 			op->ops = drm_gpuva_sm_unmap_ops_create(&uvmm->umgr,
1268 								op->va.addr,
1269 								op->va.range);
1270 			if (IS_ERR(op->ops)) {
1271 				ret = PTR_ERR(op->ops);
1272 				goto unwind_continue;
1273 			}
1274 
1275 			ret = nouveau_uvmm_sm_unmap_prepare(uvmm, &op->new,
1276 							    op->ops);
1277 			if (ret) {
1278 				drm_gpuva_ops_free(&uvmm->umgr, op->ops);
1279 				op->ops = NULL;
1280 				goto unwind_continue;
1281 			}
1282 
1283 			break;
1284 		default:
1285 			ret = -EINVAL;
1286 			goto unwind_continue;
1287 		}
1288 	}
1289 
1290 	drm_exec_init(exec, DRM_EXEC_INTERRUPTIBLE_WAIT |
1291 			    DRM_EXEC_IGNORE_DUPLICATES);
1292 	drm_exec_until_all_locked(exec) {
1293 		list_for_each_op(op, &bind_job->ops) {
1294 			struct drm_gpuva_op *va_op;
1295 
1296 			if (IS_ERR_OR_NULL(op->ops))
1297 				continue;
1298 
1299 			drm_gpuva_for_each_op(va_op, op->ops) {
1300 				struct drm_gem_object *obj = op_gem_obj(va_op);
1301 
1302 				if (unlikely(!obj))
1303 					continue;
1304 
1305 				ret = drm_exec_prepare_obj(exec, obj, 1);
1306 				drm_exec_retry_on_contention(exec);
1307 				if (ret) {
1308 					op = list_last_op(&bind_job->ops);
1309 					goto unwind;
1310 				}
1311 			}
1312 		}
1313 	}
1314 
1315 	list_for_each_op(op, &bind_job->ops) {
1316 		struct drm_gpuva_op *va_op;
1317 
1318 		if (IS_ERR_OR_NULL(op->ops))
1319 			continue;
1320 
1321 		drm_gpuva_for_each_op(va_op, op->ops) {
1322 			struct drm_gem_object *obj = op_gem_obj(va_op);
1323 
1324 			if (unlikely(!obj))
1325 				continue;
1326 
1327 			/* Don't validate GEMs backing mappings we're about to
1328 			 * unmap, it's not worth the effort.
1329 			 */
1330 			if (unlikely(va_op->op == DRM_GPUVA_OP_UNMAP))
1331 				continue;
1332 
1333 			ret = nouveau_bo_validate(nouveau_gem_object(obj),
1334 						  true, false);
1335 			if (ret) {
1336 				op = list_last_op(&bind_job->ops);
1337 				goto unwind;
1338 			}
1339 		}
1340 	}
1341 
1342 	/* Link and unlink GPUVAs while holding the dma_resv lock.
1343 	 *
1344 	 * As long as we validate() all GEMs and add fences to all GEMs DMA
1345 	 * reservations backing map and remap operations we can be sure there
1346 	 * won't be any concurrent (in)validations during job execution, hence
1347 	 * we're safe to check drm_gpuva_invalidated() within the fence
1348 	 * signalling critical path without holding a separate lock.
1349 	 *
1350 	 * GPUVAs about to be unmapped are safe as well, since they're unlinked
1351 	 * already.
1352 	 *
1353 	 * GEMs from map and remap operations must be validated before linking
1354 	 * their corresponding mappings to prevent the actual PT update to
1355 	 * happen right away in validate() rather than asynchronously as
1356 	 * intended.
1357 	 *
1358 	 * Note that after linking and unlinking the GPUVAs in this loop this
1359 	 * function cannot fail anymore, hence there is no need for an unwind
1360 	 * path.
1361 	 */
1362 	list_for_each_op(op, &bind_job->ops) {
1363 		switch (op->op) {
1364 		case OP_UNMAP_SPARSE:
1365 		case OP_MAP:
1366 		case OP_UNMAP:
1367 			bind_link_gpuvas(op->ops, &op->new);
1368 			break;
1369 		default:
1370 			break;
1371 		}
1372 	}
1373 	nouveau_uvmm_unlock(uvmm);
1374 
1375 	spin_lock(&entity->job.list.lock);
1376 	list_add(&bind_job->entry, &entity->job.list.head);
1377 	spin_unlock(&entity->job.list.lock);
1378 
1379 	return 0;
1380 
1381 unwind_continue:
1382 	op = list_prev_op(op);
1383 unwind:
1384 	list_for_each_op_from_reverse(op, &bind_job->ops) {
1385 		switch (op->op) {
1386 		case OP_MAP_SPARSE:
1387 			nouveau_uvma_region_destroy(uvmm, op->va.addr,
1388 						    op->va.range);
1389 			break;
1390 		case OP_UNMAP_SPARSE:
1391 			__nouveau_uvma_region_insert(uvmm, op->reg);
1392 			nouveau_uvmm_sm_unmap_prepare_unwind(uvmm, &op->new,
1393 							     op->ops);
1394 			break;
1395 		case OP_MAP:
1396 			nouveau_uvmm_sm_map_prepare_unwind(uvmm, &op->new,
1397 							   op->ops,
1398 							   op->va.addr,
1399 							   op->va.range);
1400 			break;
1401 		case OP_UNMAP:
1402 			nouveau_uvmm_sm_unmap_prepare_unwind(uvmm, &op->new,
1403 							     op->ops);
1404 			break;
1405 		}
1406 
1407 		drm_gpuva_ops_free(&uvmm->umgr, op->ops);
1408 		op->ops = NULL;
1409 		op->reg = NULL;
1410 	}
1411 
1412 	nouveau_uvmm_unlock(uvmm);
1413 	drm_exec_fini(exec);
1414 	return ret;
1415 }
1416 
1417 static void
1418 nouveau_uvmm_bind_job_armed_submit(struct nouveau_job *job)
1419 {
1420 	struct drm_exec *exec = &job->exec;
1421 	struct drm_gem_object *obj;
1422 	unsigned long index;
1423 
1424 	drm_exec_for_each_locked_object(exec, index, obj)
1425 		dma_resv_add_fence(obj->resv, job->done_fence, job->resv_usage);
1426 
1427 	drm_exec_fini(exec);
1428 }
1429 
1430 static struct dma_fence *
1431 nouveau_uvmm_bind_job_run(struct nouveau_job *job)
1432 {
1433 	struct nouveau_uvmm_bind_job *bind_job = to_uvmm_bind_job(job);
1434 	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(job->cli);
1435 	struct bind_job_op *op;
1436 	int ret = 0;
1437 
1438 	list_for_each_op(op, &bind_job->ops) {
1439 		switch (op->op) {
1440 		case OP_MAP_SPARSE:
1441 			/* noop */
1442 			break;
1443 		case OP_MAP:
1444 			ret = nouveau_uvmm_sm_map(uvmm, &op->new, op->ops);
1445 			if (ret)
1446 				goto out;
1447 			break;
1448 		case OP_UNMAP_SPARSE:
1449 			fallthrough;
1450 		case OP_UNMAP:
1451 			ret = nouveau_uvmm_sm_unmap(uvmm, &op->new, op->ops);
1452 			if (ret)
1453 				goto out;
1454 			break;
1455 		}
1456 	}
1457 
1458 out:
1459 	if (ret)
1460 		NV_PRINTK(err, job->cli, "bind job failed: %d\n", ret);
1461 	return ERR_PTR(ret);
1462 }
1463 
1464 static void
1465 nouveau_uvmm_bind_job_free_work_fn(struct work_struct *work)
1466 {
1467 	struct nouveau_uvmm_bind_job *bind_job =
1468 		container_of(work, struct nouveau_uvmm_bind_job, work);
1469 	struct nouveau_job *job = &bind_job->base;
1470 	struct nouveau_uvmm *uvmm = nouveau_cli_uvmm(job->cli);
1471 	struct nouveau_sched_entity *entity = job->entity;
1472 	struct bind_job_op *op, *next;
1473 
1474 	list_for_each_op(op, &bind_job->ops) {
1475 		struct drm_gem_object *obj = op->gem.obj;
1476 
1477 		/* When nouveau_uvmm_bind_job_submit() fails op->ops and op->reg
1478 		 * will be NULL, hence skip the cleanup.
1479 		 */
1480 		switch (op->op) {
1481 		case OP_MAP_SPARSE:
1482 			/* noop */
1483 			break;
1484 		case OP_UNMAP_SPARSE:
1485 			if (!IS_ERR_OR_NULL(op->ops))
1486 				nouveau_uvmm_sm_unmap_cleanup(uvmm, &op->new,
1487 							      op->ops);
1488 
1489 			if (op->reg) {
1490 				nouveau_uvma_region_sparse_unref(op->reg);
1491 				nouveau_uvmm_lock(uvmm);
1492 				nouveau_uvma_region_remove(op->reg);
1493 				nouveau_uvmm_unlock(uvmm);
1494 				nouveau_uvma_region_complete(op->reg);
1495 				nouveau_uvma_region_put(op->reg);
1496 			}
1497 
1498 			break;
1499 		case OP_MAP:
1500 			if (!IS_ERR_OR_NULL(op->ops))
1501 				nouveau_uvmm_sm_map_cleanup(uvmm, &op->new,
1502 							    op->ops);
1503 			break;
1504 		case OP_UNMAP:
1505 			if (!IS_ERR_OR_NULL(op->ops))
1506 				nouveau_uvmm_sm_unmap_cleanup(uvmm, &op->new,
1507 							      op->ops);
1508 			break;
1509 		}
1510 
1511 		if (!IS_ERR_OR_NULL(op->ops))
1512 			drm_gpuva_ops_free(&uvmm->umgr, op->ops);
1513 
1514 		if (obj)
1515 			drm_gem_object_put(obj);
1516 	}
1517 
1518 	spin_lock(&entity->job.list.lock);
1519 	list_del(&bind_job->entry);
1520 	spin_unlock(&entity->job.list.lock);
1521 
1522 	complete_all(&bind_job->complete);
1523 	wake_up(&entity->job.wq);
1524 
1525 	/* Remove and free ops after removing the bind job from the job list to
1526 	 * avoid races against bind_validate_map_sparse().
1527 	 */
1528 	list_for_each_op_safe(op, next, &bind_job->ops) {
1529 		list_del(&op->entry);
1530 		kfree(op);
1531 	}
1532 
1533 	nouveau_uvmm_bind_job_put(bind_job);
1534 }
1535 
1536 static void
1537 nouveau_uvmm_bind_job_free_qwork(struct nouveau_job *job)
1538 {
1539 	struct nouveau_uvmm_bind_job *bind_job = to_uvmm_bind_job(job);
1540 	struct nouveau_sched_entity *entity = job->entity;
1541 
1542 	nouveau_sched_entity_qwork(entity, &bind_job->work);
1543 }
1544 
1545 static struct nouveau_job_ops nouveau_bind_job_ops = {
1546 	.submit = nouveau_uvmm_bind_job_submit,
1547 	.armed_submit = nouveau_uvmm_bind_job_armed_submit,
1548 	.run = nouveau_uvmm_bind_job_run,
1549 	.free = nouveau_uvmm_bind_job_free_qwork,
1550 };
1551 
1552 static int
1553 bind_job_op_from_uop(struct bind_job_op **pop,
1554 		     struct drm_nouveau_vm_bind_op *uop)
1555 {
1556 	struct bind_job_op *op;
1557 
1558 	op = *pop = kzalloc(sizeof(*op), GFP_KERNEL);
1559 	if (!op)
1560 		return -ENOMEM;
1561 
1562 	switch (uop->op) {
1563 	case OP_MAP:
1564 		op->op = uop->flags & DRM_NOUVEAU_VM_BIND_SPARSE ?
1565 			 OP_MAP_SPARSE : OP_MAP;
1566 		break;
1567 	case OP_UNMAP:
1568 		op->op = uop->flags & DRM_NOUVEAU_VM_BIND_SPARSE ?
1569 			 OP_UNMAP_SPARSE : OP_UNMAP;
1570 		break;
1571 	default:
1572 		op->op = uop->op;
1573 		break;
1574 	}
1575 
1576 	op->flags = uop->flags;
1577 	op->va.addr = uop->addr;
1578 	op->va.range = uop->range;
1579 	op->gem.handle = uop->handle;
1580 	op->gem.offset = uop->bo_offset;
1581 
1582 	return 0;
1583 }
1584 
1585 static void
1586 bind_job_ops_free(struct list_head *ops)
1587 {
1588 	struct bind_job_op *op, *next;
1589 
1590 	list_for_each_op_safe(op, next, ops) {
1591 		list_del(&op->entry);
1592 		kfree(op);
1593 	}
1594 }
1595 
1596 static int
1597 nouveau_uvmm_bind_job_init(struct nouveau_uvmm_bind_job **pjob,
1598 			   struct nouveau_uvmm_bind_job_args *__args)
1599 {
1600 	struct nouveau_uvmm_bind_job *job;
1601 	struct nouveau_job_args args = {};
1602 	struct bind_job_op *op;
1603 	int i, ret;
1604 
1605 	ret = nouveau_uvmm_bind_job_alloc(&job);
1606 	if (ret)
1607 		return ret;
1608 
1609 	INIT_LIST_HEAD(&job->ops);
1610 	INIT_LIST_HEAD(&job->entry);
1611 
1612 	for (i = 0; i < __args->op.count; i++) {
1613 		ret = bind_job_op_from_uop(&op, &__args->op.s[i]);
1614 		if (ret)
1615 			goto err_free;
1616 
1617 		list_add_tail(&op->entry, &job->ops);
1618 	}
1619 
1620 	init_completion(&job->complete);
1621 	INIT_WORK(&job->work, nouveau_uvmm_bind_job_free_work_fn);
1622 
1623 	args.sched_entity = __args->sched_entity;
1624 	args.file_priv = __args->file_priv;
1625 
1626 	args.in_sync.count = __args->in_sync.count;
1627 	args.in_sync.s = __args->in_sync.s;
1628 
1629 	args.out_sync.count = __args->out_sync.count;
1630 	args.out_sync.s = __args->out_sync.s;
1631 
1632 	args.sync = !(__args->flags & DRM_NOUVEAU_VM_BIND_RUN_ASYNC);
1633 	args.ops = &nouveau_bind_job_ops;
1634 	args.resv_usage = DMA_RESV_USAGE_BOOKKEEP;
1635 
1636 	ret = nouveau_job_init(&job->base, &args);
1637 	if (ret)
1638 		goto err_free;
1639 
1640 	*pjob = job;
1641 	return 0;
1642 
1643 err_free:
1644 	bind_job_ops_free(&job->ops);
1645 	kfree(job);
1646 	*pjob = NULL;
1647 
1648 	return ret;
1649 }
1650 
1651 int
1652 nouveau_uvmm_ioctl_vm_init(struct drm_device *dev,
1653 			   void *data,
1654 			   struct drm_file *file_priv)
1655 {
1656 	struct nouveau_cli *cli = nouveau_cli(file_priv);
1657 	struct drm_nouveau_vm_init *init = data;
1658 
1659 	return nouveau_uvmm_init(&cli->uvmm, cli, init->kernel_managed_addr,
1660 				 init->kernel_managed_size);
1661 }
1662 
1663 static int
1664 nouveau_uvmm_vm_bind(struct nouveau_uvmm_bind_job_args *args)
1665 {
1666 	struct nouveau_uvmm_bind_job *job;
1667 	int ret;
1668 
1669 	ret = nouveau_uvmm_bind_job_init(&job, args);
1670 	if (ret)
1671 		return ret;
1672 
1673 	ret = nouveau_job_submit(&job->base);
1674 	if (ret)
1675 		goto err_job_fini;
1676 
1677 	return 0;
1678 
1679 err_job_fini:
1680 	nouveau_job_fini(&job->base);
1681 	return ret;
1682 }
1683 
1684 static int
1685 nouveau_uvmm_vm_bind_ucopy(struct nouveau_uvmm_bind_job_args *args,
1686 			   struct drm_nouveau_vm_bind *req)
1687 {
1688 	struct drm_nouveau_sync **s;
1689 	u32 inc = req->wait_count;
1690 	u64 ins = req->wait_ptr;
1691 	u32 outc = req->sig_count;
1692 	u64 outs = req->sig_ptr;
1693 	u32 opc = req->op_count;
1694 	u64 ops = req->op_ptr;
1695 	int ret;
1696 
1697 	args->flags = req->flags;
1698 
1699 	if (opc) {
1700 		args->op.count = opc;
1701 		args->op.s = u_memcpya(ops, opc,
1702 				       sizeof(*args->op.s));
1703 		if (IS_ERR(args->op.s))
1704 			return PTR_ERR(args->op.s);
1705 	}
1706 
1707 	if (inc) {
1708 		s = &args->in_sync.s;
1709 
1710 		args->in_sync.count = inc;
1711 		*s = u_memcpya(ins, inc, sizeof(**s));
1712 		if (IS_ERR(*s)) {
1713 			ret = PTR_ERR(*s);
1714 			goto err_free_ops;
1715 		}
1716 	}
1717 
1718 	if (outc) {
1719 		s = &args->out_sync.s;
1720 
1721 		args->out_sync.count = outc;
1722 		*s = u_memcpya(outs, outc, sizeof(**s));
1723 		if (IS_ERR(*s)) {
1724 			ret = PTR_ERR(*s);
1725 			goto err_free_ins;
1726 		}
1727 	}
1728 
1729 	return 0;
1730 
1731 err_free_ops:
1732 	u_free(args->op.s);
1733 err_free_ins:
1734 	u_free(args->in_sync.s);
1735 	return ret;
1736 }
1737 
1738 static void
1739 nouveau_uvmm_vm_bind_ufree(struct nouveau_uvmm_bind_job_args *args)
1740 {
1741 	u_free(args->op.s);
1742 	u_free(args->in_sync.s);
1743 	u_free(args->out_sync.s);
1744 }
1745 
1746 int
1747 nouveau_uvmm_ioctl_vm_bind(struct drm_device *dev,
1748 			   void *data,
1749 			   struct drm_file *file_priv)
1750 {
1751 	struct nouveau_cli *cli = nouveau_cli(file_priv);
1752 	struct nouveau_uvmm_bind_job_args args = {};
1753 	struct drm_nouveau_vm_bind *req = data;
1754 	int ret = 0;
1755 
1756 	if (unlikely(!nouveau_cli_uvmm_locked(cli)))
1757 		return -ENOSYS;
1758 
1759 	ret = nouveau_uvmm_vm_bind_ucopy(&args, req);
1760 	if (ret)
1761 		return ret;
1762 
1763 	args.sched_entity = &cli->sched_entity;
1764 	args.file_priv = file_priv;
1765 
1766 	ret = nouveau_uvmm_vm_bind(&args);
1767 	if (ret)
1768 		goto out_free_args;
1769 
1770 out_free_args:
1771 	nouveau_uvmm_vm_bind_ufree(&args);
1772 	return ret;
1773 }
1774 
1775 void
1776 nouveau_uvmm_bo_map_all(struct nouveau_bo *nvbo, struct nouveau_mem *mem)
1777 {
1778 	struct drm_gem_object *obj = &nvbo->bo.base;
1779 	struct drm_gpuva *va;
1780 
1781 	dma_resv_assert_held(obj->resv);
1782 
1783 	drm_gem_for_each_gpuva(va, obj) {
1784 		struct nouveau_uvma *uvma = uvma_from_va(va);
1785 
1786 		nouveau_uvma_map(uvma, mem);
1787 		drm_gpuva_invalidate(va, false);
1788 	}
1789 }
1790 
1791 void
1792 nouveau_uvmm_bo_unmap_all(struct nouveau_bo *nvbo)
1793 {
1794 	struct drm_gem_object *obj = &nvbo->bo.base;
1795 	struct drm_gpuva *va;
1796 
1797 	dma_resv_assert_held(obj->resv);
1798 
1799 	drm_gem_for_each_gpuva(va, obj) {
1800 		struct nouveau_uvma *uvma = uvma_from_va(va);
1801 
1802 		nouveau_uvma_unmap(uvma);
1803 		drm_gpuva_invalidate(va, true);
1804 	}
1805 }
1806 
1807 int
1808 nouveau_uvmm_init(struct nouveau_uvmm *uvmm, struct nouveau_cli *cli,
1809 		  u64 kernel_managed_addr, u64 kernel_managed_size)
1810 {
1811 	int ret;
1812 	u64 kernel_managed_end = kernel_managed_addr + kernel_managed_size;
1813 
1814 	mutex_init(&uvmm->mutex);
1815 	dma_resv_init(&uvmm->resv);
1816 	mt_init_flags(&uvmm->region_mt, MT_FLAGS_LOCK_EXTERN);
1817 	mt_set_external_lock(&uvmm->region_mt, &uvmm->mutex);
1818 
1819 	mutex_lock(&cli->mutex);
1820 
1821 	if (unlikely(cli->uvmm.disabled)) {
1822 		ret = -ENOSYS;
1823 		goto out_unlock;
1824 	}
1825 
1826 	if (kernel_managed_end <= kernel_managed_addr) {
1827 		ret = -EINVAL;
1828 		goto out_unlock;
1829 	}
1830 
1831 	if (kernel_managed_end > NOUVEAU_VA_SPACE_END) {
1832 		ret = -EINVAL;
1833 		goto out_unlock;
1834 	}
1835 
1836 	uvmm->kernel_managed_addr = kernel_managed_addr;
1837 	uvmm->kernel_managed_size = kernel_managed_size;
1838 
1839 	drm_gpuva_manager_init(&uvmm->umgr, cli->name,
1840 			       NOUVEAU_VA_SPACE_START,
1841 			       NOUVEAU_VA_SPACE_END,
1842 			       kernel_managed_addr, kernel_managed_size,
1843 			       NULL);
1844 
1845 	ret = nvif_vmm_ctor(&cli->mmu, "uvmm",
1846 			    cli->vmm.vmm.object.oclass, RAW,
1847 			    kernel_managed_addr, kernel_managed_size,
1848 			    NULL, 0, &cli->uvmm.vmm.vmm);
1849 	if (ret)
1850 		goto out_free_gpuva_mgr;
1851 
1852 	cli->uvmm.vmm.cli = cli;
1853 	mutex_unlock(&cli->mutex);
1854 
1855 	return 0;
1856 
1857 out_free_gpuva_mgr:
1858 	drm_gpuva_manager_destroy(&uvmm->umgr);
1859 out_unlock:
1860 	mutex_unlock(&cli->mutex);
1861 	return ret;
1862 }
1863 
1864 void
1865 nouveau_uvmm_fini(struct nouveau_uvmm *uvmm)
1866 {
1867 	MA_STATE(mas, &uvmm->region_mt, 0, 0);
1868 	struct nouveau_uvma_region *reg;
1869 	struct nouveau_cli *cli = uvmm->vmm.cli;
1870 	struct nouveau_sched_entity *entity = &cli->sched_entity;
1871 	struct drm_gpuva *va, *next;
1872 
1873 	if (!cli)
1874 		return;
1875 
1876 	rmb(); /* for list_empty to work without lock */
1877 	wait_event(entity->job.wq, list_empty(&entity->job.list.head));
1878 
1879 	nouveau_uvmm_lock(uvmm);
1880 	drm_gpuva_for_each_va_safe(va, next, &uvmm->umgr) {
1881 		struct nouveau_uvma *uvma = uvma_from_va(va);
1882 		struct drm_gem_object *obj = va->gem.obj;
1883 
1884 		if (unlikely(va == &uvmm->umgr.kernel_alloc_node))
1885 			continue;
1886 
1887 		drm_gpuva_remove(va);
1888 
1889 		dma_resv_lock(obj->resv, NULL);
1890 		drm_gpuva_unlink(va);
1891 		dma_resv_unlock(obj->resv);
1892 
1893 		nouveau_uvma_unmap(uvma);
1894 		nouveau_uvma_vmm_put(uvma);
1895 
1896 		nouveau_uvma_gem_put(uvma);
1897 		nouveau_uvma_free(uvma);
1898 	}
1899 
1900 	mas_for_each(&mas, reg, ULONG_MAX) {
1901 		mas_erase(&mas);
1902 		nouveau_uvma_region_sparse_unref(reg);
1903 		nouveau_uvma_region_put(reg);
1904 	}
1905 
1906 	WARN(!mtree_empty(&uvmm->region_mt),
1907 	     "nouveau_uvma_region tree not empty, potentially leaking memory.");
1908 	__mt_destroy(&uvmm->region_mt);
1909 	nouveau_uvmm_unlock(uvmm);
1910 
1911 	mutex_lock(&cli->mutex);
1912 	nouveau_vmm_fini(&uvmm->vmm);
1913 	drm_gpuva_manager_destroy(&uvmm->umgr);
1914 	mutex_unlock(&cli->mutex);
1915 
1916 	dma_resv_fini(&uvmm->resv);
1917 }
1918