xref: /linux/drivers/gpu/drm/nouveau/nouveau_svm.c (revision dbcedec3a31119d7594baacc743300d127c99c56)
1 /*
2  * Copyright 2018 Red Hat 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 #include "nouveau_svm.h"
23 #include "nouveau_drv.h"
24 #include "nouveau_chan.h"
25 #include "nouveau_dmem.h"
26 
27 #include <nvif/event.h>
28 #include <nvif/object.h>
29 #include <nvif/vmm.h>
30 
31 #include <nvif/class.h>
32 #include <nvif/clb069.h>
33 #include <nvif/ifc00d.h>
34 
35 #include <linux/sched/mm.h>
36 #include <linux/sort.h>
37 #include <linux/hmm.h>
38 #include <linux/memremap.h>
39 #include <linux/rmap.h>
40 
41 struct nouveau_svm {
42 	struct nouveau_drm *drm;
43 	struct mutex mutex;
44 	struct list_head inst;
45 
46 	struct nouveau_svm_fault_buffer {
47 		int id;
48 		struct nvif_object object;
49 		u32 entries;
50 		u32 getaddr;
51 		u32 putaddr;
52 		u32 get;
53 		u32 put;
54 		struct nvif_event notify;
55 		struct work_struct work;
56 
57 		struct nouveau_svm_fault {
58 			u64 inst;
59 			u64 addr;
60 			u64 time;
61 			u32 engine;
62 			u8  gpc;
63 			u8  hub;
64 			u8  access;
65 			u8  client;
66 			u8  fault;
67 			struct nouveau_svmm *svmm;
68 		} **fault;
69 		int fault_nr;
70 	} buffer[];
71 };
72 
73 #define FAULT_ACCESS_READ 0
74 #define FAULT_ACCESS_WRITE 1
75 #define FAULT_ACCESS_ATOMIC 2
76 #define FAULT_ACCESS_PREFETCH 3
77 
78 #define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
79 #define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)
80 
81 struct nouveau_pfnmap_args {
82 	struct nvif_ioctl_v0 i;
83 	struct nvif_ioctl_mthd_v0 m;
84 	struct nvif_vmm_pfnmap_v0 p;
85 };
86 
87 struct nouveau_ivmm {
88 	struct nouveau_svmm *svmm;
89 	u64 inst;
90 	struct list_head head;
91 };
92 
93 static struct nouveau_ivmm *
94 nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
95 {
96 	struct nouveau_ivmm *ivmm;
97 	list_for_each_entry(ivmm, &svm->inst, head) {
98 		if (ivmm->inst == inst)
99 			return ivmm;
100 	}
101 	return NULL;
102 }
103 
104 #define SVMM_DBG(s,f,a...)                                                     \
105 	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
106 #define SVMM_ERR(s,f,a...)                                                     \
107 	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
108 
109 int
110 nouveau_svmm_bind(struct drm_device *dev, void *data,
111 		  struct drm_file *file_priv)
112 {
113 	struct nouveau_cli *cli = nouveau_cli(file_priv);
114 	struct drm_nouveau_svm_bind *args = data;
115 	unsigned target, cmd;
116 	unsigned long addr, end;
117 	struct mm_struct *mm;
118 
119 	args->va_start &= PAGE_MASK;
120 	args->va_end = ALIGN(args->va_end, PAGE_SIZE);
121 
122 	/* Sanity check arguments */
123 	if (args->reserved0 || args->reserved1)
124 		return -EINVAL;
125 	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
126 		return -EINVAL;
127 	if (args->va_start >= args->va_end)
128 		return -EINVAL;
129 
130 	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
131 	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
132 	switch (cmd) {
133 	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
134 		break;
135 	default:
136 		return -EINVAL;
137 	}
138 
139 	/* FIXME support CPU target ie all target value < GPU_VRAM */
140 	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
141 	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
142 	switch (target) {
143 	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
144 		break;
145 	default:
146 		return -EINVAL;
147 	}
148 
149 	/*
150 	 * FIXME: For now refuse non 0 stride, we need to change the migrate
151 	 * kernel function to handle stride to avoid to create a mess within
152 	 * each device driver.
153 	 */
154 	if (args->stride)
155 		return -EINVAL;
156 
157 	/*
158 	 * Ok we are ask to do something sane, for now we only support migrate
159 	 * commands but we will add things like memory policy (what to do on
160 	 * page fault) and maybe some other commands.
161 	 */
162 
163 	mm = get_task_mm(current);
164 	if (!mm) {
165 		return -EINVAL;
166 	}
167 	mmap_read_lock(mm);
168 
169 	if (!cli->svm.svmm) {
170 		mmap_read_unlock(mm);
171 		mmput(mm);
172 		return -EINVAL;
173 	}
174 
175 	for (addr = args->va_start, end = args->va_end; addr < end;) {
176 		struct vm_area_struct *vma;
177 		unsigned long next;
178 
179 		vma = find_vma_intersection(mm, addr, end);
180 		if (!vma)
181 			break;
182 
183 		addr = max(addr, vma->vm_start);
184 		next = min(vma->vm_end, end);
185 		/* This is a best effort so we ignore errors */
186 		nouveau_dmem_migrate_vma(cli->drm, cli->svm.svmm, vma, addr,
187 					 next);
188 		addr = next;
189 	}
190 
191 	/*
192 	 * FIXME Return the number of page we have migrated, again we need to
193 	 * update the migrate API to return that information so that we can
194 	 * report it to user space.
195 	 */
196 	args->result = 0;
197 
198 	mmap_read_unlock(mm);
199 	mmput(mm);
200 
201 	return 0;
202 }
203 
204 /* Unlink channel instance from SVMM. */
205 void
206 nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
207 {
208 	struct nouveau_ivmm *ivmm;
209 	if (svmm) {
210 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
211 		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
212 		if (ivmm) {
213 			list_del(&ivmm->head);
214 			kfree(ivmm);
215 		}
216 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
217 	}
218 }
219 
220 /* Link channel instance to SVMM. */
221 int
222 nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
223 {
224 	struct nouveau_ivmm *ivmm;
225 	if (svmm) {
226 		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
227 			return -ENOMEM;
228 		ivmm->svmm = svmm;
229 		ivmm->inst = inst;
230 
231 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
232 		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
233 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
234 	}
235 	return 0;
236 }
237 
238 /* Invalidate SVMM address-range on GPU. */
239 void
240 nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
241 {
242 	if (limit > start) {
243 		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
244 				 &(struct nvif_vmm_pfnclr_v0) {
245 					.addr = start,
246 					.size = limit - start,
247 				 }, sizeof(struct nvif_vmm_pfnclr_v0));
248 	}
249 }
250 
251 static int
252 nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
253 				    const struct mmu_notifier_range *update)
254 {
255 	struct nouveau_svmm *svmm =
256 		container_of(mn, struct nouveau_svmm, notifier);
257 	unsigned long start = update->start;
258 	unsigned long limit = update->end;
259 
260 	if (!mmu_notifier_range_blockable(update))
261 		return -EAGAIN;
262 
263 	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
264 
265 	mutex_lock(&svmm->mutex);
266 	if (unlikely(!svmm->vmm))
267 		goto out;
268 
269 	/*
270 	 * Ignore invalidation callbacks for device private pages since
271 	 * the invalidation is handled as part of the migration process.
272 	 */
273 	if (update->event == MMU_NOTIFY_MIGRATE &&
274 	    update->owner == svmm->vmm->cli->drm->dev)
275 		goto out;
276 
277 	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
278 		if (start < svmm->unmanaged.start) {
279 			nouveau_svmm_invalidate(svmm, start,
280 						svmm->unmanaged.limit);
281 		}
282 		start = svmm->unmanaged.limit;
283 	}
284 
285 	nouveau_svmm_invalidate(svmm, start, limit);
286 
287 out:
288 	mutex_unlock(&svmm->mutex);
289 	return 0;
290 }
291 
292 static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
293 {
294 	kfree(container_of(mn, struct nouveau_svmm, notifier));
295 }
296 
297 static const struct mmu_notifier_ops nouveau_mn_ops = {
298 	.invalidate_range_start = nouveau_svmm_invalidate_range_start,
299 	.free_notifier = nouveau_svmm_free_notifier,
300 };
301 
302 void
303 nouveau_svmm_fini(struct nouveau_svmm **psvmm)
304 {
305 	struct nouveau_svmm *svmm = *psvmm;
306 	if (svmm) {
307 		mutex_lock(&svmm->mutex);
308 		svmm->vmm = NULL;
309 		mutex_unlock(&svmm->mutex);
310 		mmu_notifier_put(&svmm->notifier);
311 		*psvmm = NULL;
312 	}
313 }
314 
315 int
316 nouveau_svmm_init(struct drm_device *dev, void *data,
317 		  struct drm_file *file_priv)
318 {
319 	struct nouveau_cli *cli = nouveau_cli(file_priv);
320 	struct nouveau_svmm *svmm;
321 	struct drm_nouveau_svm_init *args = data;
322 	int ret;
323 
324 	/* We need to fail if svm is disabled */
325 	if (!cli->drm->svm)
326 		return -ENOSYS;
327 
328 	/* Allocate tracking for SVM-enabled VMM. */
329 	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
330 		return -ENOMEM;
331 	svmm->vmm = &cli->svm;
332 	svmm->unmanaged.start = args->unmanaged_addr;
333 	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
334 	mutex_init(&svmm->mutex);
335 
336 	/* Check that SVM isn't already enabled for the client. */
337 	mutex_lock(&cli->mutex);
338 	if (cli->svm.cli) {
339 		ret = -EBUSY;
340 		goto out_free;
341 	}
342 
343 	/* Allocate a new GPU VMM that can support SVM (managed by the
344 	 * client, with replayable faults enabled).
345 	 *
346 	 * All future channel/memory allocations will make use of this
347 	 * VMM instead of the standard one.
348 	 */
349 	ret = nvif_vmm_ctor(&cli->mmu, "svmVmm",
350 			    cli->vmm.vmm.object.oclass, MANAGED,
351 			    args->unmanaged_addr, args->unmanaged_size,
352 			    &(struct gp100_vmm_v0) {
353 				.fault_replay = true,
354 			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
355 	if (ret)
356 		goto out_free;
357 
358 	mmap_write_lock(current->mm);
359 	svmm->notifier.ops = &nouveau_mn_ops;
360 	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
361 	if (ret)
362 		goto out_mm_unlock;
363 	/* Note, ownership of svmm transfers to mmu_notifier */
364 
365 	cli->svm.svmm = svmm;
366 	cli->svm.cli = cli;
367 	mmap_write_unlock(current->mm);
368 	mutex_unlock(&cli->mutex);
369 	return 0;
370 
371 out_mm_unlock:
372 	mmap_write_unlock(current->mm);
373 out_free:
374 	mutex_unlock(&cli->mutex);
375 	kfree(svmm);
376 	return ret;
377 }
378 
379 /* Issue fault replay for GPU to retry accesses that faulted previously. */
380 static void
381 nouveau_svm_fault_replay(struct nouveau_svm *svm)
382 {
383 	SVM_DBG(svm, "replay");
384 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
385 				 GP100_VMM_VN_FAULT_REPLAY,
386 				 &(struct gp100_vmm_fault_replay_vn) {},
387 				 sizeof(struct gp100_vmm_fault_replay_vn)));
388 }
389 
390 /* Cancel a replayable fault that could not be handled.
391  *
392  * Cancelling the fault will trigger recovery to reset the engine
393  * and kill the offending channel (ie. GPU SIGSEGV).
394  */
395 static void
396 nouveau_svm_fault_cancel(struct nouveau_svm *svm,
397 			 u64 inst, u8 hub, u8 gpc, u8 client)
398 {
399 	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
400 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
401 				 GP100_VMM_VN_FAULT_CANCEL,
402 				 &(struct gp100_vmm_fault_cancel_v0) {
403 					.hub = hub,
404 					.gpc = gpc,
405 					.client = client,
406 					.inst = inst,
407 				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
408 }
409 
410 static void
411 nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
412 			       struct nouveau_svm_fault *fault)
413 {
414 	nouveau_svm_fault_cancel(svm, fault->inst,
415 				      fault->hub,
416 				      fault->gpc,
417 				      fault->client);
418 }
419 
420 static int
421 nouveau_svm_fault_priority(u8 fault)
422 {
423 	switch (fault) {
424 	case FAULT_ACCESS_PREFETCH:
425 		return 0;
426 	case FAULT_ACCESS_READ:
427 		return 1;
428 	case FAULT_ACCESS_WRITE:
429 		return 2;
430 	case FAULT_ACCESS_ATOMIC:
431 		return 3;
432 	default:
433 		WARN_ON_ONCE(1);
434 		return -1;
435 	}
436 }
437 
438 static int
439 nouveau_svm_fault_cmp(const void *a, const void *b)
440 {
441 	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
442 	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
443 	int ret;
444 	if ((ret = (s64)fa->inst - fb->inst))
445 		return ret;
446 	if ((ret = (s64)fa->addr - fb->addr))
447 		return ret;
448 	return nouveau_svm_fault_priority(fa->access) -
449 		nouveau_svm_fault_priority(fb->access);
450 }
451 
452 static void
453 nouveau_svm_fault_cache(struct nouveau_svm *svm,
454 			struct nouveau_svm_fault_buffer *buffer, u32 offset)
455 {
456 	struct nvif_object *memory = &buffer->object;
457 	const u32 instlo = nvif_rd32(memory, offset + 0x00);
458 	const u32 insthi = nvif_rd32(memory, offset + 0x04);
459 	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
460 	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
461 	const u32 timelo = nvif_rd32(memory, offset + 0x10);
462 	const u32 timehi = nvif_rd32(memory, offset + 0x14);
463 	const u32 engine = nvif_rd32(memory, offset + 0x18);
464 	const u32   info = nvif_rd32(memory, offset + 0x1c);
465 	const u64   inst = (u64)insthi << 32 | instlo;
466 	const u8     gpc = (info & 0x1f000000) >> 24;
467 	const u8     hub = (info & 0x00100000) >> 20;
468 	const u8  client = (info & 0x00007f00) >> 8;
469 	struct nouveau_svm_fault *fault;
470 
471 	//XXX: i think we're supposed to spin waiting */
472 	if (WARN_ON(!(info & 0x80000000)))
473 		return;
474 
475 	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
476 
477 	if (!buffer->fault[buffer->fault_nr]) {
478 		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
479 		if (WARN_ON(!fault)) {
480 			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
481 			return;
482 		}
483 		buffer->fault[buffer->fault_nr] = fault;
484 	}
485 
486 	fault = buffer->fault[buffer->fault_nr++];
487 	fault->inst   = inst;
488 	fault->addr   = (u64)addrhi << 32 | addrlo;
489 	fault->time   = (u64)timehi << 32 | timelo;
490 	fault->engine = engine;
491 	fault->gpc    = gpc;
492 	fault->hub    = hub;
493 	fault->access = (info & 0x000f0000) >> 16;
494 	fault->client = client;
495 	fault->fault  = (info & 0x0000001f);
496 
497 	SVM_DBG(svm, "fault %016llx %016llx %02x",
498 		fault->inst, fault->addr, fault->access);
499 }
500 
501 struct svm_notifier {
502 	struct mmu_interval_notifier notifier;
503 	struct nouveau_svmm *svmm;
504 };
505 
506 static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
507 					 const struct mmu_notifier_range *range,
508 					 unsigned long cur_seq)
509 {
510 	struct svm_notifier *sn =
511 		container_of(mni, struct svm_notifier, notifier);
512 
513 	if (range->event == MMU_NOTIFY_EXCLUSIVE &&
514 	    range->owner == sn->svmm->vmm->cli->drm->dev)
515 		return true;
516 
517 	/*
518 	 * serializes the update to mni->invalidate_seq done by caller and
519 	 * prevents invalidation of the PTE from progressing while HW is being
520 	 * programmed. This is very hacky and only works because the normal
521 	 * notifier that does invalidation is always called after the range
522 	 * notifier.
523 	 */
524 	if (mmu_notifier_range_blockable(range))
525 		mutex_lock(&sn->svmm->mutex);
526 	else if (!mutex_trylock(&sn->svmm->mutex))
527 		return false;
528 	mmu_interval_set_seq(mni, cur_seq);
529 	mutex_unlock(&sn->svmm->mutex);
530 	return true;
531 }
532 
533 static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
534 	.invalidate = nouveau_svm_range_invalidate,
535 };
536 
537 static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm,
538 				    struct hmm_range *range,
539 				    struct nouveau_pfnmap_args *args)
540 {
541 	struct page *page;
542 
543 	/*
544 	 * The address prepared here is passed through nvif_object_ioctl()
545 	 * to an eventual DMA map in something like gp100_vmm_pgt_pfn()
546 	 *
547 	 * This is all just encoding the internal hmm representation into a
548 	 * different nouveau internal representation.
549 	 */
550 	if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) {
551 		args->p.phys[0] = 0;
552 		return;
553 	}
554 
555 	page = hmm_pfn_to_page(range->hmm_pfns[0]);
556 	/*
557 	 * Only map compound pages to the GPU if the CPU is also mapping the
558 	 * page as a compound page. Otherwise, the PTE protections might not be
559 	 * consistent (e.g., CPU only maps part of a compound page).
560 	 * Note that the underlying page might still be larger than the
561 	 * CPU mapping (e.g., a PUD sized compound page partially mapped with
562 	 * a PMD sized page table entry).
563 	 */
564 	if (hmm_pfn_to_map_order(range->hmm_pfns[0])) {
565 		unsigned long addr = args->p.addr;
566 
567 		args->p.page = hmm_pfn_to_map_order(range->hmm_pfns[0]) +
568 				PAGE_SHIFT;
569 		args->p.size = 1UL << args->p.page;
570 		args->p.addr &= ~(args->p.size - 1);
571 		page -= (addr - args->p.addr) >> PAGE_SHIFT;
572 	}
573 	if (is_device_private_page(page))
574 		args->p.phys[0] = nouveau_dmem_page_addr(page) |
575 				NVIF_VMM_PFNMAP_V0_V |
576 				NVIF_VMM_PFNMAP_V0_VRAM;
577 	else
578 		args->p.phys[0] = page_to_phys(page) |
579 				NVIF_VMM_PFNMAP_V0_V |
580 				NVIF_VMM_PFNMAP_V0_HOST;
581 	if (range->hmm_pfns[0] & HMM_PFN_WRITE)
582 		args->p.phys[0] |= NVIF_VMM_PFNMAP_V0_W;
583 }
584 
585 static int nouveau_atomic_range_fault(struct nouveau_svmm *svmm,
586 			       struct nouveau_drm *drm,
587 			       struct nouveau_pfnmap_args *args, u32 size,
588 			       struct svm_notifier *notifier)
589 {
590 	unsigned long timeout =
591 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
592 	struct mm_struct *mm = svmm->notifier.mm;
593 	struct page *page;
594 	unsigned long start = args->p.addr;
595 	unsigned long notifier_seq;
596 	int ret = 0;
597 
598 	ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
599 					args->p.addr, args->p.size,
600 					&nouveau_svm_mni_ops);
601 	if (ret)
602 		return ret;
603 
604 	while (true) {
605 		if (time_after(jiffies, timeout)) {
606 			ret = -EBUSY;
607 			goto out;
608 		}
609 
610 		notifier_seq = mmu_interval_read_begin(&notifier->notifier);
611 		mmap_read_lock(mm);
612 		ret = make_device_exclusive_range(mm, start, start + PAGE_SIZE,
613 					    &page, drm->dev);
614 		mmap_read_unlock(mm);
615 		if (ret <= 0 || !page) {
616 			ret = -EINVAL;
617 			goto out;
618 		}
619 
620 		mutex_lock(&svmm->mutex);
621 		if (!mmu_interval_read_retry(&notifier->notifier,
622 					     notifier_seq))
623 			break;
624 		mutex_unlock(&svmm->mutex);
625 	}
626 
627 	/* Map the page on the GPU. */
628 	args->p.page = 12;
629 	args->p.size = PAGE_SIZE;
630 	args->p.addr = start;
631 	args->p.phys[0] = page_to_phys(page) |
632 		NVIF_VMM_PFNMAP_V0_V |
633 		NVIF_VMM_PFNMAP_V0_W |
634 		NVIF_VMM_PFNMAP_V0_A |
635 		NVIF_VMM_PFNMAP_V0_HOST;
636 
637 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
638 	mutex_unlock(&svmm->mutex);
639 
640 	unlock_page(page);
641 	put_page(page);
642 
643 out:
644 	mmu_interval_notifier_remove(&notifier->notifier);
645 	return ret;
646 }
647 
648 static int nouveau_range_fault(struct nouveau_svmm *svmm,
649 			       struct nouveau_drm *drm,
650 			       struct nouveau_pfnmap_args *args, u32 size,
651 			       unsigned long hmm_flags,
652 			       struct svm_notifier *notifier)
653 {
654 	unsigned long timeout =
655 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
656 	/* Have HMM fault pages within the fault window to the GPU. */
657 	unsigned long hmm_pfns[1];
658 	struct hmm_range range = {
659 		.notifier = &notifier->notifier,
660 		.default_flags = hmm_flags,
661 		.hmm_pfns = hmm_pfns,
662 		.dev_private_owner = drm->dev,
663 	};
664 	struct mm_struct *mm = svmm->notifier.mm;
665 	int ret;
666 
667 	ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
668 					args->p.addr, args->p.size,
669 					&nouveau_svm_mni_ops);
670 	if (ret)
671 		return ret;
672 
673 	range.start = notifier->notifier.interval_tree.start;
674 	range.end = notifier->notifier.interval_tree.last + 1;
675 
676 	while (true) {
677 		if (time_after(jiffies, timeout)) {
678 			ret = -EBUSY;
679 			goto out;
680 		}
681 
682 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
683 		mmap_read_lock(mm);
684 		ret = hmm_range_fault(&range);
685 		mmap_read_unlock(mm);
686 		if (ret) {
687 			if (ret == -EBUSY)
688 				continue;
689 			goto out;
690 		}
691 
692 		mutex_lock(&svmm->mutex);
693 		if (mmu_interval_read_retry(range.notifier,
694 					    range.notifier_seq)) {
695 			mutex_unlock(&svmm->mutex);
696 			continue;
697 		}
698 		break;
699 	}
700 
701 	nouveau_hmm_convert_pfn(drm, &range, args);
702 
703 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
704 	mutex_unlock(&svmm->mutex);
705 
706 out:
707 	mmu_interval_notifier_remove(&notifier->notifier);
708 
709 	return ret;
710 }
711 
712 static void
713 nouveau_svm_fault(struct work_struct *work)
714 {
715 	struct nouveau_svm_fault_buffer *buffer = container_of(work, typeof(*buffer), work);
716 	struct nouveau_svm *svm = container_of(buffer, typeof(*svm), buffer[buffer->id]);
717 	struct nvif_object *device = &svm->drm->client.device.object;
718 	struct nouveau_svmm *svmm;
719 	struct {
720 		struct nouveau_pfnmap_args i;
721 		u64 phys[1];
722 	} args;
723 	unsigned long hmm_flags;
724 	u64 inst, start, limit;
725 	int fi, fn;
726 	int replay = 0, atomic = 0, ret;
727 
728 	/* Parse available fault buffer entries into a cache, and update
729 	 * the GET pointer so HW can reuse the entries.
730 	 */
731 	SVM_DBG(svm, "fault handler");
732 	if (buffer->get == buffer->put) {
733 		buffer->put = nvif_rd32(device, buffer->putaddr);
734 		buffer->get = nvif_rd32(device, buffer->getaddr);
735 		if (buffer->get == buffer->put)
736 			return;
737 	}
738 	buffer->fault_nr = 0;
739 
740 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
741 	while (buffer->get != buffer->put) {
742 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
743 		if (++buffer->get == buffer->entries)
744 			buffer->get = 0;
745 	}
746 	nvif_wr32(device, buffer->getaddr, buffer->get);
747 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
748 
749 	/* Sort parsed faults by instance pointer to prevent unnecessary
750 	 * instance to SVMM translations, followed by address and access
751 	 * type to reduce the amount of work when handling the faults.
752 	 */
753 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
754 	     nouveau_svm_fault_cmp, NULL);
755 
756 	/* Lookup SVMM structure for each unique instance pointer. */
757 	mutex_lock(&svm->mutex);
758 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
759 		if (!svmm || buffer->fault[fi]->inst != inst) {
760 			struct nouveau_ivmm *ivmm =
761 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
762 			svmm = ivmm ? ivmm->svmm : NULL;
763 			inst = buffer->fault[fi]->inst;
764 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
765 		}
766 		buffer->fault[fi]->svmm = svmm;
767 	}
768 	mutex_unlock(&svm->mutex);
769 
770 	/* Process list of faults. */
771 	args.i.i.version = 0;
772 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
773 	args.i.m.version = 0;
774 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
775 	args.i.p.version = 0;
776 
777 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
778 		struct svm_notifier notifier;
779 		struct mm_struct *mm;
780 
781 		/* Cancel any faults from non-SVM channels. */
782 		if (!(svmm = buffer->fault[fi]->svmm)) {
783 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
784 			continue;
785 		}
786 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
787 
788 		/* We try and group handling of faults within a small
789 		 * window into a single update.
790 		 */
791 		start = buffer->fault[fi]->addr;
792 		limit = start + PAGE_SIZE;
793 		if (start < svmm->unmanaged.limit)
794 			limit = min_t(u64, limit, svmm->unmanaged.start);
795 
796 		/*
797 		 * Prepare the GPU-side update of all pages within the
798 		 * fault window, determining required pages and access
799 		 * permissions based on pending faults.
800 		 */
801 		args.i.p.addr = start;
802 		args.i.p.page = PAGE_SHIFT;
803 		args.i.p.size = PAGE_SIZE;
804 		/*
805 		 * Determine required permissions based on GPU fault
806 		 * access flags.
807 		 */
808 		switch (buffer->fault[fi]->access) {
809 		case 0: /* READ. */
810 			hmm_flags = HMM_PFN_REQ_FAULT;
811 			break;
812 		case 2: /* ATOMIC. */
813 			atomic = true;
814 			break;
815 		case 3: /* PREFETCH. */
816 			hmm_flags = 0;
817 			break;
818 		default:
819 			hmm_flags = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE;
820 			break;
821 		}
822 
823 		mm = svmm->notifier.mm;
824 		if (!mmget_not_zero(mm)) {
825 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
826 			continue;
827 		}
828 
829 		notifier.svmm = svmm;
830 		if (atomic)
831 			ret = nouveau_atomic_range_fault(svmm, svm->drm,
832 							 &args.i, sizeof(args),
833 							 &notifier);
834 		else
835 			ret = nouveau_range_fault(svmm, svm->drm, &args.i,
836 						  sizeof(args), hmm_flags,
837 						  &notifier);
838 		mmput(mm);
839 
840 		limit = args.i.p.addr + args.i.p.size;
841 		for (fn = fi; ++fn < buffer->fault_nr; ) {
842 			/* It's okay to skip over duplicate addresses from the
843 			 * same SVMM as faults are ordered by access type such
844 			 * that only the first one needs to be handled.
845 			 *
846 			 * ie. WRITE faults appear first, thus any handling of
847 			 * pending READ faults will already be satisfied.
848 			 * But if a large page is mapped, make sure subsequent
849 			 * fault addresses have sufficient access permission.
850 			 */
851 			if (buffer->fault[fn]->svmm != svmm ||
852 			    buffer->fault[fn]->addr >= limit ||
853 			    (buffer->fault[fi]->access == FAULT_ACCESS_READ &&
854 			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_V)) ||
855 			    (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
856 			     buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
857 			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W)) ||
858 			    (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
859 			     buffer->fault[fi]->access != FAULT_ACCESS_WRITE &&
860 			     buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
861 			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_A)))
862 				break;
863 		}
864 
865 		/* If handling failed completely, cancel all faults. */
866 		if (ret) {
867 			while (fi < fn) {
868 				struct nouveau_svm_fault *fault =
869 					buffer->fault[fi++];
870 
871 				nouveau_svm_fault_cancel_fault(svm, fault);
872 			}
873 		} else
874 			replay++;
875 	}
876 
877 	/* Issue fault replay to the GPU. */
878 	if (replay)
879 		nouveau_svm_fault_replay(svm);
880 }
881 
882 static int
883 nouveau_svm_event(struct nvif_event *event, void *argv, u32 argc)
884 {
885 	struct nouveau_svm_fault_buffer *buffer = container_of(event, typeof(*buffer), notify);
886 
887 	schedule_work(&buffer->work);
888 	return NVIF_EVENT_KEEP;
889 }
890 
891 static struct nouveau_pfnmap_args *
892 nouveau_pfns_to_args(void *pfns)
893 {
894 	return container_of(pfns, struct nouveau_pfnmap_args, p.phys);
895 }
896 
897 u64 *
898 nouveau_pfns_alloc(unsigned long npages)
899 {
900 	struct nouveau_pfnmap_args *args;
901 
902 	args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL);
903 	if (!args)
904 		return NULL;
905 
906 	args->i.type = NVIF_IOCTL_V0_MTHD;
907 	args->m.method = NVIF_VMM_V0_PFNMAP;
908 	args->p.page = PAGE_SHIFT;
909 
910 	return args->p.phys;
911 }
912 
913 void
914 nouveau_pfns_free(u64 *pfns)
915 {
916 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
917 
918 	kfree(args);
919 }
920 
921 void
922 nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm,
923 		 unsigned long addr, u64 *pfns, unsigned long npages)
924 {
925 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
926 
927 	args->p.addr = addr;
928 	args->p.size = npages << PAGE_SHIFT;
929 
930 	mutex_lock(&svmm->mutex);
931 
932 	nvif_object_ioctl(&svmm->vmm->vmm.object, args,
933 			  struct_size(args, p.phys, npages), NULL);
934 
935 	mutex_unlock(&svmm->mutex);
936 }
937 
938 static void
939 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
940 {
941 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
942 
943 	nvif_event_block(&buffer->notify);
944 	flush_work(&buffer->work);
945 }
946 
947 static int
948 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
949 {
950 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
951 	struct nvif_object *device = &svm->drm->client.device.object;
952 
953 	buffer->get = nvif_rd32(device, buffer->getaddr);
954 	buffer->put = nvif_rd32(device, buffer->putaddr);
955 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
956 
957 	return nvif_event_allow(&buffer->notify);
958 }
959 
960 static void
961 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
962 {
963 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
964 	int i;
965 
966 	if (!nvif_object_constructed(&buffer->object))
967 		return;
968 
969 	nouveau_svm_fault_buffer_fini(svm, id);
970 
971 	if (buffer->fault) {
972 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
973 			kfree(buffer->fault[i]);
974 		kvfree(buffer->fault);
975 	}
976 
977 	nvif_event_dtor(&buffer->notify);
978 	nvif_object_dtor(&buffer->object);
979 }
980 
981 static int
982 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
983 {
984 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
985 	struct nouveau_drm *drm = svm->drm;
986 	struct nvif_object *device = &drm->client.device.object;
987 	struct nvif_clb069_v0 args = {};
988 	int ret;
989 
990 	buffer->id = id;
991 
992 	ret = nvif_object_ctor(device, "svmFaultBuffer", 0, oclass, &args,
993 			       sizeof(args), &buffer->object);
994 	if (ret < 0) {
995 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
996 		return ret;
997 	}
998 
999 	nvif_object_map(&buffer->object, NULL, 0);
1000 	buffer->entries = args.entries;
1001 	buffer->getaddr = args.get;
1002 	buffer->putaddr = args.put;
1003 	INIT_WORK(&buffer->work, nouveau_svm_fault);
1004 
1005 	ret = nvif_event_ctor(&buffer->object, "svmFault", id, nouveau_svm_event, true, NULL, 0,
1006 			      &buffer->notify);
1007 	if (ret)
1008 		return ret;
1009 
1010 	buffer->fault = kvcalloc(buffer->entries, sizeof(*buffer->fault), GFP_KERNEL);
1011 	if (!buffer->fault)
1012 		return -ENOMEM;
1013 
1014 	return nouveau_svm_fault_buffer_init(svm, id);
1015 }
1016 
1017 void
1018 nouveau_svm_resume(struct nouveau_drm *drm)
1019 {
1020 	struct nouveau_svm *svm = drm->svm;
1021 	if (svm)
1022 		nouveau_svm_fault_buffer_init(svm, 0);
1023 }
1024 
1025 void
1026 nouveau_svm_suspend(struct nouveau_drm *drm)
1027 {
1028 	struct nouveau_svm *svm = drm->svm;
1029 	if (svm)
1030 		nouveau_svm_fault_buffer_fini(svm, 0);
1031 }
1032 
1033 void
1034 nouveau_svm_fini(struct nouveau_drm *drm)
1035 {
1036 	struct nouveau_svm *svm = drm->svm;
1037 	if (svm) {
1038 		nouveau_svm_fault_buffer_dtor(svm, 0);
1039 		kfree(drm->svm);
1040 		drm->svm = NULL;
1041 	}
1042 }
1043 
1044 void
1045 nouveau_svm_init(struct nouveau_drm *drm)
1046 {
1047 	static const struct nvif_mclass buffers[] = {
1048 		{   VOLTA_FAULT_BUFFER_A, 0 },
1049 		{ MAXWELL_FAULT_BUFFER_A, 0 },
1050 		{}
1051 	};
1052 	struct nouveau_svm *svm;
1053 	int ret;
1054 
1055 	/* Disable on Volta and newer until channel recovery is fixed,
1056 	 * otherwise clients will have a trivial way to trash the GPU
1057 	 * for everyone.
1058 	 */
1059 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
1060 		return;
1061 
1062 	drm->svm = svm = kzalloc(struct_size(drm->svm, buffer, 1), GFP_KERNEL);
1063 	if (!drm->svm)
1064 		return;
1065 
1066 	drm->svm->drm = drm;
1067 	mutex_init(&drm->svm->mutex);
1068 	INIT_LIST_HEAD(&drm->svm->inst);
1069 
1070 	ret = nvif_mclass(&drm->client.device.object, buffers);
1071 	if (ret < 0) {
1072 		SVM_DBG(svm, "No supported fault buffer class");
1073 		nouveau_svm_fini(drm);
1074 		return;
1075 	}
1076 
1077 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
1078 	if (ret) {
1079 		nouveau_svm_fini(drm);
1080 		return;
1081 	}
1082 
1083 	SVM_DBG(svm, "Initialised");
1084 }
1085