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/notify.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 
39 struct nouveau_svm {
40 	struct nouveau_drm *drm;
41 	struct mutex mutex;
42 	struct list_head inst;
43 
44 	struct nouveau_svm_fault_buffer {
45 		int id;
46 		struct nvif_object object;
47 		u32 entries;
48 		u32 getaddr;
49 		u32 putaddr;
50 		u32 get;
51 		u32 put;
52 		struct nvif_notify notify;
53 
54 		struct nouveau_svm_fault {
55 			u64 inst;
56 			u64 addr;
57 			u64 time;
58 			u32 engine;
59 			u8  gpc;
60 			u8  hub;
61 			u8  access;
62 			u8  client;
63 			u8  fault;
64 			struct nouveau_svmm *svmm;
65 		} **fault;
66 		int fault_nr;
67 	} buffer[1];
68 };
69 
70 #define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
71 #define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)
72 
73 struct nouveau_pfnmap_args {
74 	struct nvif_ioctl_v0 i;
75 	struct nvif_ioctl_mthd_v0 m;
76 	struct nvif_vmm_pfnmap_v0 p;
77 };
78 
79 struct nouveau_ivmm {
80 	struct nouveau_svmm *svmm;
81 	u64 inst;
82 	struct list_head head;
83 };
84 
85 static struct nouveau_ivmm *
86 nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
87 {
88 	struct nouveau_ivmm *ivmm;
89 	list_for_each_entry(ivmm, &svm->inst, head) {
90 		if (ivmm->inst == inst)
91 			return ivmm;
92 	}
93 	return NULL;
94 }
95 
96 #define SVMM_DBG(s,f,a...)                                                     \
97 	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
98 #define SVMM_ERR(s,f,a...)                                                     \
99 	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
100 
101 int
102 nouveau_svmm_bind(struct drm_device *dev, void *data,
103 		  struct drm_file *file_priv)
104 {
105 	struct nouveau_cli *cli = nouveau_cli(file_priv);
106 	struct drm_nouveau_svm_bind *args = data;
107 	unsigned target, cmd, priority;
108 	unsigned long addr, end;
109 	struct mm_struct *mm;
110 
111 	args->va_start &= PAGE_MASK;
112 	args->va_end = ALIGN(args->va_end, PAGE_SIZE);
113 
114 	/* Sanity check arguments */
115 	if (args->reserved0 || args->reserved1)
116 		return -EINVAL;
117 	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
118 		return -EINVAL;
119 	if (args->va_start >= args->va_end)
120 		return -EINVAL;
121 
122 	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
123 	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
124 	switch (cmd) {
125 	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
126 		break;
127 	default:
128 		return -EINVAL;
129 	}
130 
131 	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
132 	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
133 
134 	/* FIXME support CPU target ie all target value < GPU_VRAM */
135 	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
136 	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
137 	switch (target) {
138 	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
139 		break;
140 	default:
141 		return -EINVAL;
142 	}
143 
144 	/*
145 	 * FIXME: For now refuse non 0 stride, we need to change the migrate
146 	 * kernel function to handle stride to avoid to create a mess within
147 	 * each device driver.
148 	 */
149 	if (args->stride)
150 		return -EINVAL;
151 
152 	/*
153 	 * Ok we are ask to do something sane, for now we only support migrate
154 	 * commands but we will add things like memory policy (what to do on
155 	 * page fault) and maybe some other commands.
156 	 */
157 
158 	mm = get_task_mm(current);
159 	mmap_read_lock(mm);
160 
161 	if (!cli->svm.svmm) {
162 		mmap_read_unlock(mm);
163 		return -EINVAL;
164 	}
165 
166 	for (addr = args->va_start, end = args->va_end; addr < end;) {
167 		struct vm_area_struct *vma;
168 		unsigned long next;
169 
170 		vma = find_vma_intersection(mm, addr, end);
171 		if (!vma)
172 			break;
173 
174 		addr = max(addr, vma->vm_start);
175 		next = min(vma->vm_end, end);
176 		/* This is a best effort so we ignore errors */
177 		nouveau_dmem_migrate_vma(cli->drm, cli->svm.svmm, vma, addr,
178 					 next);
179 		addr = next;
180 	}
181 
182 	/*
183 	 * FIXME Return the number of page we have migrated, again we need to
184 	 * update the migrate API to return that information so that we can
185 	 * report it to user space.
186 	 */
187 	args->result = 0;
188 
189 	mmap_read_unlock(mm);
190 	mmput(mm);
191 
192 	return 0;
193 }
194 
195 /* Unlink channel instance from SVMM. */
196 void
197 nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
198 {
199 	struct nouveau_ivmm *ivmm;
200 	if (svmm) {
201 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
202 		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
203 		if (ivmm) {
204 			list_del(&ivmm->head);
205 			kfree(ivmm);
206 		}
207 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
208 	}
209 }
210 
211 /* Link channel instance to SVMM. */
212 int
213 nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
214 {
215 	struct nouveau_ivmm *ivmm;
216 	if (svmm) {
217 		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
218 			return -ENOMEM;
219 		ivmm->svmm = svmm;
220 		ivmm->inst = inst;
221 
222 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
223 		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
224 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
225 	}
226 	return 0;
227 }
228 
229 /* Invalidate SVMM address-range on GPU. */
230 void
231 nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
232 {
233 	if (limit > start) {
234 		bool super = svmm->vmm->vmm.object.client->super;
235 		svmm->vmm->vmm.object.client->super = true;
236 		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
237 				 &(struct nvif_vmm_pfnclr_v0) {
238 					.addr = start,
239 					.size = limit - start,
240 				 }, sizeof(struct nvif_vmm_pfnclr_v0));
241 		svmm->vmm->vmm.object.client->super = super;
242 	}
243 }
244 
245 static int
246 nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
247 				    const struct mmu_notifier_range *update)
248 {
249 	struct nouveau_svmm *svmm =
250 		container_of(mn, struct nouveau_svmm, notifier);
251 	unsigned long start = update->start;
252 	unsigned long limit = update->end;
253 
254 	if (!mmu_notifier_range_blockable(update))
255 		return -EAGAIN;
256 
257 	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
258 
259 	mutex_lock(&svmm->mutex);
260 	if (unlikely(!svmm->vmm))
261 		goto out;
262 
263 	/*
264 	 * Ignore invalidation callbacks for device private pages since
265 	 * the invalidation is handled as part of the migration process.
266 	 */
267 	if (update->event == MMU_NOTIFY_MIGRATE &&
268 	    update->migrate_pgmap_owner == svmm->vmm->cli->drm->dev)
269 		goto out;
270 
271 	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
272 		if (start < svmm->unmanaged.start) {
273 			nouveau_svmm_invalidate(svmm, start,
274 						svmm->unmanaged.limit);
275 		}
276 		start = svmm->unmanaged.limit;
277 	}
278 
279 	nouveau_svmm_invalidate(svmm, start, limit);
280 
281 out:
282 	mutex_unlock(&svmm->mutex);
283 	return 0;
284 }
285 
286 static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
287 {
288 	kfree(container_of(mn, struct nouveau_svmm, notifier));
289 }
290 
291 static const struct mmu_notifier_ops nouveau_mn_ops = {
292 	.invalidate_range_start = nouveau_svmm_invalidate_range_start,
293 	.free_notifier = nouveau_svmm_free_notifier,
294 };
295 
296 void
297 nouveau_svmm_fini(struct nouveau_svmm **psvmm)
298 {
299 	struct nouveau_svmm *svmm = *psvmm;
300 	if (svmm) {
301 		mutex_lock(&svmm->mutex);
302 		svmm->vmm = NULL;
303 		mutex_unlock(&svmm->mutex);
304 		mmu_notifier_put(&svmm->notifier);
305 		*psvmm = NULL;
306 	}
307 }
308 
309 int
310 nouveau_svmm_init(struct drm_device *dev, void *data,
311 		  struct drm_file *file_priv)
312 {
313 	struct nouveau_cli *cli = nouveau_cli(file_priv);
314 	struct nouveau_svmm *svmm;
315 	struct drm_nouveau_svm_init *args = data;
316 	int ret;
317 
318 	/* Allocate tracking for SVM-enabled VMM. */
319 	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
320 		return -ENOMEM;
321 	svmm->vmm = &cli->svm;
322 	svmm->unmanaged.start = args->unmanaged_addr;
323 	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
324 	mutex_init(&svmm->mutex);
325 
326 	/* Check that SVM isn't already enabled for the client. */
327 	mutex_lock(&cli->mutex);
328 	if (cli->svm.cli) {
329 		ret = -EBUSY;
330 		goto out_free;
331 	}
332 
333 	/* Allocate a new GPU VMM that can support SVM (managed by the
334 	 * client, with replayable faults enabled).
335 	 *
336 	 * All future channel/memory allocations will make use of this
337 	 * VMM instead of the standard one.
338 	 */
339 	ret = nvif_vmm_ctor(&cli->mmu, "svmVmm",
340 			    cli->vmm.vmm.object.oclass, true,
341 			    args->unmanaged_addr, args->unmanaged_size,
342 			    &(struct gp100_vmm_v0) {
343 				.fault_replay = true,
344 			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
345 	if (ret)
346 		goto out_free;
347 
348 	mmap_write_lock(current->mm);
349 	svmm->notifier.ops = &nouveau_mn_ops;
350 	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
351 	if (ret)
352 		goto out_mm_unlock;
353 	/* Note, ownership of svmm transfers to mmu_notifier */
354 
355 	cli->svm.svmm = svmm;
356 	cli->svm.cli = cli;
357 	mmap_write_unlock(current->mm);
358 	mutex_unlock(&cli->mutex);
359 	return 0;
360 
361 out_mm_unlock:
362 	mmap_write_unlock(current->mm);
363 out_free:
364 	mutex_unlock(&cli->mutex);
365 	kfree(svmm);
366 	return ret;
367 }
368 
369 /* Issue fault replay for GPU to retry accesses that faulted previously. */
370 static void
371 nouveau_svm_fault_replay(struct nouveau_svm *svm)
372 {
373 	SVM_DBG(svm, "replay");
374 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
375 				 GP100_VMM_VN_FAULT_REPLAY,
376 				 &(struct gp100_vmm_fault_replay_vn) {},
377 				 sizeof(struct gp100_vmm_fault_replay_vn)));
378 }
379 
380 /* Cancel a replayable fault that could not be handled.
381  *
382  * Cancelling the fault will trigger recovery to reset the engine
383  * and kill the offending channel (ie. GPU SIGSEGV).
384  */
385 static void
386 nouveau_svm_fault_cancel(struct nouveau_svm *svm,
387 			 u64 inst, u8 hub, u8 gpc, u8 client)
388 {
389 	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
390 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
391 				 GP100_VMM_VN_FAULT_CANCEL,
392 				 &(struct gp100_vmm_fault_cancel_v0) {
393 					.hub = hub,
394 					.gpc = gpc,
395 					.client = client,
396 					.inst = inst,
397 				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
398 }
399 
400 static void
401 nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
402 			       struct nouveau_svm_fault *fault)
403 {
404 	nouveau_svm_fault_cancel(svm, fault->inst,
405 				      fault->hub,
406 				      fault->gpc,
407 				      fault->client);
408 }
409 
410 static int
411 nouveau_svm_fault_cmp(const void *a, const void *b)
412 {
413 	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
414 	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
415 	int ret;
416 	if ((ret = (s64)fa->inst - fb->inst))
417 		return ret;
418 	if ((ret = (s64)fa->addr - fb->addr))
419 		return ret;
420 	/*XXX: atomic? */
421 	return (fa->access == 0 || fa->access == 3) -
422 	       (fb->access == 0 || fb->access == 3);
423 }
424 
425 static void
426 nouveau_svm_fault_cache(struct nouveau_svm *svm,
427 			struct nouveau_svm_fault_buffer *buffer, u32 offset)
428 {
429 	struct nvif_object *memory = &buffer->object;
430 	const u32 instlo = nvif_rd32(memory, offset + 0x00);
431 	const u32 insthi = nvif_rd32(memory, offset + 0x04);
432 	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
433 	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
434 	const u32 timelo = nvif_rd32(memory, offset + 0x10);
435 	const u32 timehi = nvif_rd32(memory, offset + 0x14);
436 	const u32 engine = nvif_rd32(memory, offset + 0x18);
437 	const u32   info = nvif_rd32(memory, offset + 0x1c);
438 	const u64   inst = (u64)insthi << 32 | instlo;
439 	const u8     gpc = (info & 0x1f000000) >> 24;
440 	const u8     hub = (info & 0x00100000) >> 20;
441 	const u8  client = (info & 0x00007f00) >> 8;
442 	struct nouveau_svm_fault *fault;
443 
444 	//XXX: i think we're supposed to spin waiting */
445 	if (WARN_ON(!(info & 0x80000000)))
446 		return;
447 
448 	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
449 
450 	if (!buffer->fault[buffer->fault_nr]) {
451 		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
452 		if (WARN_ON(!fault)) {
453 			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
454 			return;
455 		}
456 		buffer->fault[buffer->fault_nr] = fault;
457 	}
458 
459 	fault = buffer->fault[buffer->fault_nr++];
460 	fault->inst   = inst;
461 	fault->addr   = (u64)addrhi << 32 | addrlo;
462 	fault->time   = (u64)timehi << 32 | timelo;
463 	fault->engine = engine;
464 	fault->gpc    = gpc;
465 	fault->hub    = hub;
466 	fault->access = (info & 0x000f0000) >> 16;
467 	fault->client = client;
468 	fault->fault  = (info & 0x0000001f);
469 
470 	SVM_DBG(svm, "fault %016llx %016llx %02x",
471 		fault->inst, fault->addr, fault->access);
472 }
473 
474 struct svm_notifier {
475 	struct mmu_interval_notifier notifier;
476 	struct nouveau_svmm *svmm;
477 };
478 
479 static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
480 					 const struct mmu_notifier_range *range,
481 					 unsigned long cur_seq)
482 {
483 	struct svm_notifier *sn =
484 		container_of(mni, struct svm_notifier, notifier);
485 
486 	/*
487 	 * serializes the update to mni->invalidate_seq done by caller and
488 	 * prevents invalidation of the PTE from progressing while HW is being
489 	 * programmed. This is very hacky and only works because the normal
490 	 * notifier that does invalidation is always called after the range
491 	 * notifier.
492 	 */
493 	if (mmu_notifier_range_blockable(range))
494 		mutex_lock(&sn->svmm->mutex);
495 	else if (!mutex_trylock(&sn->svmm->mutex))
496 		return false;
497 	mmu_interval_set_seq(mni, cur_seq);
498 	mutex_unlock(&sn->svmm->mutex);
499 	return true;
500 }
501 
502 static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
503 	.invalidate = nouveau_svm_range_invalidate,
504 };
505 
506 static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm,
507 				    struct hmm_range *range,
508 				    struct nouveau_pfnmap_args *args)
509 {
510 	struct page *page;
511 
512 	/*
513 	 * The address prepared here is passed through nvif_object_ioctl()
514 	 * to an eventual DMA map in something like gp100_vmm_pgt_pfn()
515 	 *
516 	 * This is all just encoding the internal hmm representation into a
517 	 * different nouveau internal representation.
518 	 */
519 	if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) {
520 		args->p.phys[0] = 0;
521 		return;
522 	}
523 
524 	page = hmm_pfn_to_page(range->hmm_pfns[0]);
525 	/*
526 	 * Only map compound pages to the GPU if the CPU is also mapping the
527 	 * page as a compound page. Otherwise, the PTE protections might not be
528 	 * consistent (e.g., CPU only maps part of a compound page).
529 	 * Note that the underlying page might still be larger than the
530 	 * CPU mapping (e.g., a PUD sized compound page partially mapped with
531 	 * a PMD sized page table entry).
532 	 */
533 	if (hmm_pfn_to_map_order(range->hmm_pfns[0])) {
534 		unsigned long addr = args->p.addr;
535 
536 		args->p.page = hmm_pfn_to_map_order(range->hmm_pfns[0]) +
537 				PAGE_SHIFT;
538 		args->p.size = 1UL << args->p.page;
539 		args->p.addr &= ~(args->p.size - 1);
540 		page -= (addr - args->p.addr) >> PAGE_SHIFT;
541 	}
542 	if (is_device_private_page(page))
543 		args->p.phys[0] = nouveau_dmem_page_addr(page) |
544 				NVIF_VMM_PFNMAP_V0_V |
545 				NVIF_VMM_PFNMAP_V0_VRAM;
546 	else
547 		args->p.phys[0] = page_to_phys(page) |
548 				NVIF_VMM_PFNMAP_V0_V |
549 				NVIF_VMM_PFNMAP_V0_HOST;
550 	if (range->hmm_pfns[0] & HMM_PFN_WRITE)
551 		args->p.phys[0] |= NVIF_VMM_PFNMAP_V0_W;
552 }
553 
554 static int nouveau_range_fault(struct nouveau_svmm *svmm,
555 			       struct nouveau_drm *drm,
556 			       struct nouveau_pfnmap_args *args, u32 size,
557 			       unsigned long hmm_flags,
558 			       struct svm_notifier *notifier)
559 {
560 	unsigned long timeout =
561 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
562 	/* Have HMM fault pages within the fault window to the GPU. */
563 	unsigned long hmm_pfns[1];
564 	struct hmm_range range = {
565 		.notifier = &notifier->notifier,
566 		.start = notifier->notifier.interval_tree.start,
567 		.end = notifier->notifier.interval_tree.last + 1,
568 		.default_flags = hmm_flags,
569 		.hmm_pfns = hmm_pfns,
570 		.dev_private_owner = drm->dev,
571 	};
572 	struct mm_struct *mm = notifier->notifier.mm;
573 	int ret;
574 
575 	while (true) {
576 		if (time_after(jiffies, timeout))
577 			return -EBUSY;
578 
579 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
580 		mmap_read_lock(mm);
581 		ret = hmm_range_fault(&range);
582 		mmap_read_unlock(mm);
583 		if (ret) {
584 			if (ret == -EBUSY)
585 				continue;
586 			return ret;
587 		}
588 
589 		mutex_lock(&svmm->mutex);
590 		if (mmu_interval_read_retry(range.notifier,
591 					    range.notifier_seq)) {
592 			mutex_unlock(&svmm->mutex);
593 			continue;
594 		}
595 		break;
596 	}
597 
598 	nouveau_hmm_convert_pfn(drm, &range, args);
599 
600 	svmm->vmm->vmm.object.client->super = true;
601 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
602 	svmm->vmm->vmm.object.client->super = false;
603 	mutex_unlock(&svmm->mutex);
604 
605 	return ret;
606 }
607 
608 static int
609 nouveau_svm_fault(struct nvif_notify *notify)
610 {
611 	struct nouveau_svm_fault_buffer *buffer =
612 		container_of(notify, typeof(*buffer), notify);
613 	struct nouveau_svm *svm =
614 		container_of(buffer, typeof(*svm), buffer[buffer->id]);
615 	struct nvif_object *device = &svm->drm->client.device.object;
616 	struct nouveau_svmm *svmm;
617 	struct {
618 		struct nouveau_pfnmap_args i;
619 		u64 phys[1];
620 	} args;
621 	unsigned long hmm_flags;
622 	u64 inst, start, limit;
623 	int fi, fn;
624 	int replay = 0, ret;
625 
626 	/* Parse available fault buffer entries into a cache, and update
627 	 * the GET pointer so HW can reuse the entries.
628 	 */
629 	SVM_DBG(svm, "fault handler");
630 	if (buffer->get == buffer->put) {
631 		buffer->put = nvif_rd32(device, buffer->putaddr);
632 		buffer->get = nvif_rd32(device, buffer->getaddr);
633 		if (buffer->get == buffer->put)
634 			return NVIF_NOTIFY_KEEP;
635 	}
636 	buffer->fault_nr = 0;
637 
638 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
639 	while (buffer->get != buffer->put) {
640 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
641 		if (++buffer->get == buffer->entries)
642 			buffer->get = 0;
643 	}
644 	nvif_wr32(device, buffer->getaddr, buffer->get);
645 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
646 
647 	/* Sort parsed faults by instance pointer to prevent unnecessary
648 	 * instance to SVMM translations, followed by address and access
649 	 * type to reduce the amount of work when handling the faults.
650 	 */
651 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
652 	     nouveau_svm_fault_cmp, NULL);
653 
654 	/* Lookup SVMM structure for each unique instance pointer. */
655 	mutex_lock(&svm->mutex);
656 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
657 		if (!svmm || buffer->fault[fi]->inst != inst) {
658 			struct nouveau_ivmm *ivmm =
659 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
660 			svmm = ivmm ? ivmm->svmm : NULL;
661 			inst = buffer->fault[fi]->inst;
662 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
663 		}
664 		buffer->fault[fi]->svmm = svmm;
665 	}
666 	mutex_unlock(&svm->mutex);
667 
668 	/* Process list of faults. */
669 	args.i.i.version = 0;
670 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
671 	args.i.m.version = 0;
672 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
673 	args.i.p.version = 0;
674 
675 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
676 		struct svm_notifier notifier;
677 		struct mm_struct *mm;
678 
679 		/* Cancel any faults from non-SVM channels. */
680 		if (!(svmm = buffer->fault[fi]->svmm)) {
681 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
682 			continue;
683 		}
684 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
685 
686 		/* We try and group handling of faults within a small
687 		 * window into a single update.
688 		 */
689 		start = buffer->fault[fi]->addr;
690 		limit = start + PAGE_SIZE;
691 		if (start < svmm->unmanaged.limit)
692 			limit = min_t(u64, limit, svmm->unmanaged.start);
693 
694 		/*
695 		 * Prepare the GPU-side update of all pages within the
696 		 * fault window, determining required pages and access
697 		 * permissions based on pending faults.
698 		 */
699 		args.i.p.addr = start;
700 		args.i.p.page = PAGE_SHIFT;
701 		args.i.p.size = PAGE_SIZE;
702 		/*
703 		 * Determine required permissions based on GPU fault
704 		 * access flags.
705 		 * XXX: atomic?
706 		 */
707 		switch (buffer->fault[fi]->access) {
708 		case 0: /* READ. */
709 			hmm_flags = HMM_PFN_REQ_FAULT;
710 			break;
711 		case 3: /* PREFETCH. */
712 			hmm_flags = 0;
713 			break;
714 		default:
715 			hmm_flags = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE;
716 			break;
717 		}
718 
719 		mm = svmm->notifier.mm;
720 		if (!mmget_not_zero(mm)) {
721 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
722 			continue;
723 		}
724 
725 		notifier.svmm = svmm;
726 		ret = mmu_interval_notifier_insert(&notifier.notifier, mm,
727 						   args.i.p.addr, args.i.p.size,
728 						   &nouveau_svm_mni_ops);
729 		if (!ret) {
730 			ret = nouveau_range_fault(svmm, svm->drm, &args.i,
731 				sizeof(args), hmm_flags, &notifier);
732 			mmu_interval_notifier_remove(&notifier.notifier);
733 		}
734 		mmput(mm);
735 
736 		limit = args.i.p.addr + args.i.p.size;
737 		for (fn = fi; ++fn < buffer->fault_nr; ) {
738 			/* It's okay to skip over duplicate addresses from the
739 			 * same SVMM as faults are ordered by access type such
740 			 * that only the first one needs to be handled.
741 			 *
742 			 * ie. WRITE faults appear first, thus any handling of
743 			 * pending READ faults will already be satisfied.
744 			 * But if a large page is mapped, make sure subsequent
745 			 * fault addresses have sufficient access permission.
746 			 */
747 			if (buffer->fault[fn]->svmm != svmm ||
748 			    buffer->fault[fn]->addr >= limit ||
749 			    (buffer->fault[fi]->access == 0 /* READ. */ &&
750 			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_V)) ||
751 			    (buffer->fault[fi]->access != 0 /* READ. */ &&
752 			     buffer->fault[fi]->access != 3 /* PREFETCH. */ &&
753 			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W)))
754 				break;
755 		}
756 
757 		/* If handling failed completely, cancel all faults. */
758 		if (ret) {
759 			while (fi < fn) {
760 				struct nouveau_svm_fault *fault =
761 					buffer->fault[fi++];
762 
763 				nouveau_svm_fault_cancel_fault(svm, fault);
764 			}
765 		} else
766 			replay++;
767 	}
768 
769 	/* Issue fault replay to the GPU. */
770 	if (replay)
771 		nouveau_svm_fault_replay(svm);
772 	return NVIF_NOTIFY_KEEP;
773 }
774 
775 static struct nouveau_pfnmap_args *
776 nouveau_pfns_to_args(void *pfns)
777 {
778 	return container_of(pfns, struct nouveau_pfnmap_args, p.phys);
779 }
780 
781 u64 *
782 nouveau_pfns_alloc(unsigned long npages)
783 {
784 	struct nouveau_pfnmap_args *args;
785 
786 	args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL);
787 	if (!args)
788 		return NULL;
789 
790 	args->i.type = NVIF_IOCTL_V0_MTHD;
791 	args->m.method = NVIF_VMM_V0_PFNMAP;
792 	args->p.page = PAGE_SHIFT;
793 
794 	return args->p.phys;
795 }
796 
797 void
798 nouveau_pfns_free(u64 *pfns)
799 {
800 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
801 
802 	kfree(args);
803 }
804 
805 void
806 nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm,
807 		 unsigned long addr, u64 *pfns, unsigned long npages)
808 {
809 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
810 	int ret;
811 
812 	args->p.addr = addr;
813 	args->p.size = npages << PAGE_SHIFT;
814 
815 	mutex_lock(&svmm->mutex);
816 
817 	svmm->vmm->vmm.object.client->super = true;
818 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, sizeof(*args) +
819 				npages * sizeof(args->p.phys[0]), NULL);
820 	svmm->vmm->vmm.object.client->super = false;
821 
822 	mutex_unlock(&svmm->mutex);
823 }
824 
825 static void
826 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
827 {
828 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
829 	nvif_notify_put(&buffer->notify);
830 }
831 
832 static int
833 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
834 {
835 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
836 	struct nvif_object *device = &svm->drm->client.device.object;
837 	buffer->get = nvif_rd32(device, buffer->getaddr);
838 	buffer->put = nvif_rd32(device, buffer->putaddr);
839 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
840 	return nvif_notify_get(&buffer->notify);
841 }
842 
843 static void
844 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
845 {
846 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
847 	int i;
848 
849 	if (buffer->fault) {
850 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
851 			kfree(buffer->fault[i]);
852 		kvfree(buffer->fault);
853 	}
854 
855 	nouveau_svm_fault_buffer_fini(svm, id);
856 
857 	nvif_notify_dtor(&buffer->notify);
858 	nvif_object_dtor(&buffer->object);
859 }
860 
861 static int
862 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
863 {
864 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
865 	struct nouveau_drm *drm = svm->drm;
866 	struct nvif_object *device = &drm->client.device.object;
867 	struct nvif_clb069_v0 args = {};
868 	int ret;
869 
870 	buffer->id = id;
871 
872 	ret = nvif_object_ctor(device, "svmFaultBuffer", 0, oclass, &args,
873 			       sizeof(args), &buffer->object);
874 	if (ret < 0) {
875 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
876 		return ret;
877 	}
878 
879 	nvif_object_map(&buffer->object, NULL, 0);
880 	buffer->entries = args.entries;
881 	buffer->getaddr = args.get;
882 	buffer->putaddr = args.put;
883 
884 	ret = nvif_notify_ctor(&buffer->object, "svmFault", nouveau_svm_fault,
885 			       true, NVB069_V0_NTFY_FAULT, NULL, 0, 0,
886 			       &buffer->notify);
887 	if (ret)
888 		return ret;
889 
890 	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
891 	if (!buffer->fault)
892 		return -ENOMEM;
893 
894 	return nouveau_svm_fault_buffer_init(svm, id);
895 }
896 
897 void
898 nouveau_svm_resume(struct nouveau_drm *drm)
899 {
900 	struct nouveau_svm *svm = drm->svm;
901 	if (svm)
902 		nouveau_svm_fault_buffer_init(svm, 0);
903 }
904 
905 void
906 nouveau_svm_suspend(struct nouveau_drm *drm)
907 {
908 	struct nouveau_svm *svm = drm->svm;
909 	if (svm)
910 		nouveau_svm_fault_buffer_fini(svm, 0);
911 }
912 
913 void
914 nouveau_svm_fini(struct nouveau_drm *drm)
915 {
916 	struct nouveau_svm *svm = drm->svm;
917 	if (svm) {
918 		nouveau_svm_fault_buffer_dtor(svm, 0);
919 		kfree(drm->svm);
920 		drm->svm = NULL;
921 	}
922 }
923 
924 void
925 nouveau_svm_init(struct nouveau_drm *drm)
926 {
927 	static const struct nvif_mclass buffers[] = {
928 		{   VOLTA_FAULT_BUFFER_A, 0 },
929 		{ MAXWELL_FAULT_BUFFER_A, 0 },
930 		{}
931 	};
932 	struct nouveau_svm *svm;
933 	int ret;
934 
935 	/* Disable on Volta and newer until channel recovery is fixed,
936 	 * otherwise clients will have a trivial way to trash the GPU
937 	 * for everyone.
938 	 */
939 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
940 		return;
941 
942 	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
943 		return;
944 
945 	drm->svm->drm = drm;
946 	mutex_init(&drm->svm->mutex);
947 	INIT_LIST_HEAD(&drm->svm->inst);
948 
949 	ret = nvif_mclass(&drm->client.device.object, buffers);
950 	if (ret < 0) {
951 		SVM_DBG(svm, "No supported fault buffer class");
952 		nouveau_svm_fini(drm);
953 		return;
954 	}
955 
956 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
957 	if (ret) {
958 		nouveau_svm_fini(drm);
959 		return;
960 	}
961 
962 	SVM_DBG(svm, "Initialised");
963 }
964