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_ivmm {
74 	struct nouveau_svmm *svmm;
75 	u64 inst;
76 	struct list_head head;
77 };
78 
79 static struct nouveau_ivmm *
80 nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
81 {
82 	struct nouveau_ivmm *ivmm;
83 	list_for_each_entry(ivmm, &svm->inst, head) {
84 		if (ivmm->inst == inst)
85 			return ivmm;
86 	}
87 	return NULL;
88 }
89 
90 struct nouveau_svmm {
91 	struct mmu_notifier notifier;
92 	struct nouveau_vmm *vmm;
93 	struct {
94 		unsigned long start;
95 		unsigned long limit;
96 	} unmanaged;
97 
98 	struct mutex mutex;
99 };
100 
101 #define SVMM_DBG(s,f,a...)                                                     \
102 	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
103 #define SVMM_ERR(s,f,a...)                                                     \
104 	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
105 
106 int
107 nouveau_svmm_bind(struct drm_device *dev, void *data,
108 		  struct drm_file *file_priv)
109 {
110 	struct nouveau_cli *cli = nouveau_cli(file_priv);
111 	struct drm_nouveau_svm_bind *args = data;
112 	unsigned target, cmd, priority;
113 	unsigned long addr, end, size;
114 	struct mm_struct *mm;
115 
116 	args->va_start &= PAGE_MASK;
117 	args->va_end &= PAGE_MASK;
118 
119 	/* Sanity check arguments */
120 	if (args->reserved0 || args->reserved1)
121 		return -EINVAL;
122 	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
123 		return -EINVAL;
124 	if (args->va_start >= args->va_end)
125 		return -EINVAL;
126 	if (!args->npages)
127 		return -EINVAL;
128 
129 	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
130 	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
131 	switch (cmd) {
132 	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
133 		break;
134 	default:
135 		return -EINVAL;
136 	}
137 
138 	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
139 	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
140 
141 	/* FIXME support CPU target ie all target value < GPU_VRAM */
142 	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
143 	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
144 	switch (target) {
145 	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
146 		break;
147 	default:
148 		return -EINVAL;
149 	}
150 
151 	/*
152 	 * FIXME: For now refuse non 0 stride, we need to change the migrate
153 	 * kernel function to handle stride to avoid to create a mess within
154 	 * each device driver.
155 	 */
156 	if (args->stride)
157 		return -EINVAL;
158 
159 	size = ((unsigned long)args->npages) << PAGE_SHIFT;
160 	if ((args->va_start + size) <= args->va_start)
161 		return -EINVAL;
162 	if ((args->va_start + size) > args->va_end)
163 		return -EINVAL;
164 
165 	/*
166 	 * Ok we are ask to do something sane, for now we only support migrate
167 	 * commands but we will add things like memory policy (what to do on
168 	 * page fault) and maybe some other commands.
169 	 */
170 
171 	mm = get_task_mm(current);
172 	down_read(&mm->mmap_sem);
173 
174 	for (addr = args->va_start, end = args->va_start + size; addr < end;) {
175 		struct vm_area_struct *vma;
176 		unsigned long next;
177 
178 		vma = find_vma_intersection(mm, addr, end);
179 		if (!vma)
180 			break;
181 
182 		next = min(vma->vm_end, end);
183 		/* This is a best effort so we ignore errors */
184 		nouveau_dmem_migrate_vma(cli->drm, vma, addr, next);
185 		addr = next;
186 	}
187 
188 	/*
189 	 * FIXME Return the number of page we have migrated, again we need to
190 	 * update the migrate API to return that information so that we can
191 	 * report it to user space.
192 	 */
193 	args->result = 0;
194 
195 	up_read(&mm->mmap_sem);
196 	mmput(mm);
197 
198 	return 0;
199 }
200 
201 /* Unlink channel instance from SVMM. */
202 void
203 nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
204 {
205 	struct nouveau_ivmm *ivmm;
206 	if (svmm) {
207 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
208 		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
209 		if (ivmm) {
210 			list_del(&ivmm->head);
211 			kfree(ivmm);
212 		}
213 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
214 	}
215 }
216 
217 /* Link channel instance to SVMM. */
218 int
219 nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
220 {
221 	struct nouveau_ivmm *ivmm;
222 	if (svmm) {
223 		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
224 			return -ENOMEM;
225 		ivmm->svmm = svmm;
226 		ivmm->inst = inst;
227 
228 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
229 		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
230 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
231 	}
232 	return 0;
233 }
234 
235 /* Invalidate SVMM address-range on GPU. */
236 static void
237 nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
238 {
239 	if (limit > start) {
240 		bool super = svmm->vmm->vmm.object.client->super;
241 		svmm->vmm->vmm.object.client->super = true;
242 		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
243 				 &(struct nvif_vmm_pfnclr_v0) {
244 					.addr = start,
245 					.size = limit - start,
246 				 }, sizeof(struct nvif_vmm_pfnclr_v0));
247 		svmm->vmm->vmm.object.client->super = super;
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 	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
270 		if (start < svmm->unmanaged.start) {
271 			nouveau_svmm_invalidate(svmm, start,
272 						svmm->unmanaged.limit);
273 		}
274 		start = svmm->unmanaged.limit;
275 	}
276 
277 	nouveau_svmm_invalidate(svmm, start, limit);
278 
279 out:
280 	mutex_unlock(&svmm->mutex);
281 	return 0;
282 }
283 
284 static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
285 {
286 	kfree(container_of(mn, struct nouveau_svmm, notifier));
287 }
288 
289 static const struct mmu_notifier_ops nouveau_mn_ops = {
290 	.invalidate_range_start = nouveau_svmm_invalidate_range_start,
291 	.free_notifier = nouveau_svmm_free_notifier,
292 };
293 
294 void
295 nouveau_svmm_fini(struct nouveau_svmm **psvmm)
296 {
297 	struct nouveau_svmm *svmm = *psvmm;
298 	if (svmm) {
299 		mutex_lock(&svmm->mutex);
300 		svmm->vmm = NULL;
301 		mutex_unlock(&svmm->mutex);
302 		mmu_notifier_put(&svmm->notifier);
303 		*psvmm = NULL;
304 	}
305 }
306 
307 int
308 nouveau_svmm_init(struct drm_device *dev, void *data,
309 		  struct drm_file *file_priv)
310 {
311 	struct nouveau_cli *cli = nouveau_cli(file_priv);
312 	struct nouveau_svmm *svmm;
313 	struct drm_nouveau_svm_init *args = data;
314 	int ret;
315 
316 	/* Allocate tracking for SVM-enabled VMM. */
317 	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
318 		return -ENOMEM;
319 	svmm->vmm = &cli->svm;
320 	svmm->unmanaged.start = args->unmanaged_addr;
321 	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
322 	mutex_init(&svmm->mutex);
323 
324 	/* Check that SVM isn't already enabled for the client. */
325 	mutex_lock(&cli->mutex);
326 	if (cli->svm.cli) {
327 		ret = -EBUSY;
328 		goto out_free;
329 	}
330 
331 	/* Allocate a new GPU VMM that can support SVM (managed by the
332 	 * client, with replayable faults enabled).
333 	 *
334 	 * All future channel/memory allocations will make use of this
335 	 * VMM instead of the standard one.
336 	 */
337 	ret = nvif_vmm_init(&cli->mmu, cli->vmm.vmm.object.oclass, true,
338 			    args->unmanaged_addr, args->unmanaged_size,
339 			    &(struct gp100_vmm_v0) {
340 				.fault_replay = true,
341 			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
342 	if (ret)
343 		goto out_free;
344 
345 	down_write(&current->mm->mmap_sem);
346 	svmm->notifier.ops = &nouveau_mn_ops;
347 	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
348 	if (ret)
349 		goto out_mm_unlock;
350 	/* Note, ownership of svmm transfers to mmu_notifier */
351 
352 	cli->svm.svmm = svmm;
353 	cli->svm.cli = cli;
354 	up_write(&current->mm->mmap_sem);
355 	mutex_unlock(&cli->mutex);
356 	return 0;
357 
358 out_mm_unlock:
359 	up_write(&current->mm->mmap_sem);
360 out_free:
361 	mutex_unlock(&cli->mutex);
362 	kfree(svmm);
363 	return ret;
364 }
365 
366 static const u64
367 nouveau_svm_pfn_flags[HMM_PFN_FLAG_MAX] = {
368 	[HMM_PFN_VALID         ] = NVIF_VMM_PFNMAP_V0_V,
369 	[HMM_PFN_WRITE         ] = NVIF_VMM_PFNMAP_V0_W,
370 	[HMM_PFN_DEVICE_PRIVATE] = NVIF_VMM_PFNMAP_V0_VRAM,
371 };
372 
373 static const u64
374 nouveau_svm_pfn_values[HMM_PFN_VALUE_MAX] = {
375 	[HMM_PFN_ERROR  ] = ~NVIF_VMM_PFNMAP_V0_V,
376 	[HMM_PFN_NONE   ] =  NVIF_VMM_PFNMAP_V0_NONE,
377 	[HMM_PFN_SPECIAL] = ~NVIF_VMM_PFNMAP_V0_V,
378 };
379 
380 /* Issue fault replay for GPU to retry accesses that faulted previously. */
381 static void
382 nouveau_svm_fault_replay(struct nouveau_svm *svm)
383 {
384 	SVM_DBG(svm, "replay");
385 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
386 				 GP100_VMM_VN_FAULT_REPLAY,
387 				 &(struct gp100_vmm_fault_replay_vn) {},
388 				 sizeof(struct gp100_vmm_fault_replay_vn)));
389 }
390 
391 /* Cancel a replayable fault that could not be handled.
392  *
393  * Cancelling the fault will trigger recovery to reset the engine
394  * and kill the offending channel (ie. GPU SIGSEGV).
395  */
396 static void
397 nouveau_svm_fault_cancel(struct nouveau_svm *svm,
398 			 u64 inst, u8 hub, u8 gpc, u8 client)
399 {
400 	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
401 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
402 				 GP100_VMM_VN_FAULT_CANCEL,
403 				 &(struct gp100_vmm_fault_cancel_v0) {
404 					.hub = hub,
405 					.gpc = gpc,
406 					.client = client,
407 					.inst = inst,
408 				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
409 }
410 
411 static void
412 nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
413 			       struct nouveau_svm_fault *fault)
414 {
415 	nouveau_svm_fault_cancel(svm, fault->inst,
416 				      fault->hub,
417 				      fault->gpc,
418 				      fault->client);
419 }
420 
421 static int
422 nouveau_svm_fault_cmp(const void *a, const void *b)
423 {
424 	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
425 	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
426 	int ret;
427 	if ((ret = (s64)fa->inst - fb->inst))
428 		return ret;
429 	if ((ret = (s64)fa->addr - fb->addr))
430 		return ret;
431 	/*XXX: atomic? */
432 	return (fa->access == 0 || fa->access == 3) -
433 	       (fb->access == 0 || fb->access == 3);
434 }
435 
436 static void
437 nouveau_svm_fault_cache(struct nouveau_svm *svm,
438 			struct nouveau_svm_fault_buffer *buffer, u32 offset)
439 {
440 	struct nvif_object *memory = &buffer->object;
441 	const u32 instlo = nvif_rd32(memory, offset + 0x00);
442 	const u32 insthi = nvif_rd32(memory, offset + 0x04);
443 	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
444 	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
445 	const u32 timelo = nvif_rd32(memory, offset + 0x10);
446 	const u32 timehi = nvif_rd32(memory, offset + 0x14);
447 	const u32 engine = nvif_rd32(memory, offset + 0x18);
448 	const u32   info = nvif_rd32(memory, offset + 0x1c);
449 	const u64   inst = (u64)insthi << 32 | instlo;
450 	const u8     gpc = (info & 0x1f000000) >> 24;
451 	const u8     hub = (info & 0x00100000) >> 20;
452 	const u8  client = (info & 0x00007f00) >> 8;
453 	struct nouveau_svm_fault *fault;
454 
455 	//XXX: i think we're supposed to spin waiting */
456 	if (WARN_ON(!(info & 0x80000000)))
457 		return;
458 
459 	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
460 
461 	if (!buffer->fault[buffer->fault_nr]) {
462 		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
463 		if (WARN_ON(!fault)) {
464 			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
465 			return;
466 		}
467 		buffer->fault[buffer->fault_nr] = fault;
468 	}
469 
470 	fault = buffer->fault[buffer->fault_nr++];
471 	fault->inst   = inst;
472 	fault->addr   = (u64)addrhi << 32 | addrlo;
473 	fault->time   = (u64)timehi << 32 | timelo;
474 	fault->engine = engine;
475 	fault->gpc    = gpc;
476 	fault->hub    = hub;
477 	fault->access = (info & 0x000f0000) >> 16;
478 	fault->client = client;
479 	fault->fault  = (info & 0x0000001f);
480 
481 	SVM_DBG(svm, "fault %016llx %016llx %02x",
482 		fault->inst, fault->addr, fault->access);
483 }
484 
485 struct svm_notifier {
486 	struct mmu_interval_notifier notifier;
487 	struct nouveau_svmm *svmm;
488 };
489 
490 static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
491 					 const struct mmu_notifier_range *range,
492 					 unsigned long cur_seq)
493 {
494 	struct svm_notifier *sn =
495 		container_of(mni, struct svm_notifier, notifier);
496 
497 	/*
498 	 * serializes the update to mni->invalidate_seq done by caller and
499 	 * prevents invalidation of the PTE from progressing while HW is being
500 	 * programmed. This is very hacky and only works because the normal
501 	 * notifier that does invalidation is always called after the range
502 	 * notifier.
503 	 */
504 	if (mmu_notifier_range_blockable(range))
505 		mutex_lock(&sn->svmm->mutex);
506 	else if (!mutex_trylock(&sn->svmm->mutex))
507 		return false;
508 	mmu_interval_set_seq(mni, cur_seq);
509 	mutex_unlock(&sn->svmm->mutex);
510 	return true;
511 }
512 
513 static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
514 	.invalidate = nouveau_svm_range_invalidate,
515 };
516 
517 static int nouveau_range_fault(struct nouveau_svmm *svmm,
518 			       struct nouveau_drm *drm, void *data, u32 size,
519 			       u64 *pfns, struct svm_notifier *notifier)
520 {
521 	unsigned long timeout =
522 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
523 	/* Have HMM fault pages within the fault window to the GPU. */
524 	struct hmm_range range = {
525 		.notifier = &notifier->notifier,
526 		.start = notifier->notifier.interval_tree.start,
527 		.end = notifier->notifier.interval_tree.last + 1,
528 		.pfns = pfns,
529 		.flags = nouveau_svm_pfn_flags,
530 		.values = nouveau_svm_pfn_values,
531 		.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT,
532 	};
533 	struct mm_struct *mm = notifier->notifier.mm;
534 	long ret;
535 
536 	while (true) {
537 		if (time_after(jiffies, timeout))
538 			return -EBUSY;
539 
540 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
541 		range.default_flags = 0;
542 		range.pfn_flags_mask = -1UL;
543 		down_read(&mm->mmap_sem);
544 		ret = hmm_range_fault(&range, 0);
545 		up_read(&mm->mmap_sem);
546 		if (ret <= 0) {
547 			if (ret == 0 || ret == -EBUSY)
548 				continue;
549 			return ret;
550 		}
551 
552 		mutex_lock(&svmm->mutex);
553 		if (mmu_interval_read_retry(range.notifier,
554 					    range.notifier_seq)) {
555 			mutex_unlock(&svmm->mutex);
556 			continue;
557 		}
558 		break;
559 	}
560 
561 	nouveau_dmem_convert_pfn(drm, &range);
562 
563 	svmm->vmm->vmm.object.client->super = true;
564 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, data, size, NULL);
565 	svmm->vmm->vmm.object.client->super = false;
566 	mutex_unlock(&svmm->mutex);
567 
568 	return ret;
569 }
570 
571 static int
572 nouveau_svm_fault(struct nvif_notify *notify)
573 {
574 	struct nouveau_svm_fault_buffer *buffer =
575 		container_of(notify, typeof(*buffer), notify);
576 	struct nouveau_svm *svm =
577 		container_of(buffer, typeof(*svm), buffer[buffer->id]);
578 	struct nvif_object *device = &svm->drm->client.device.object;
579 	struct nouveau_svmm *svmm;
580 	struct {
581 		struct {
582 			struct nvif_ioctl_v0 i;
583 			struct nvif_ioctl_mthd_v0 m;
584 			struct nvif_vmm_pfnmap_v0 p;
585 		} i;
586 		u64 phys[16];
587 	} args;
588 	struct vm_area_struct *vma;
589 	u64 inst, start, limit;
590 	int fi, fn, pi, fill;
591 	int replay = 0, ret;
592 
593 	/* Parse available fault buffer entries into a cache, and update
594 	 * the GET pointer so HW can reuse the entries.
595 	 */
596 	SVM_DBG(svm, "fault handler");
597 	if (buffer->get == buffer->put) {
598 		buffer->put = nvif_rd32(device, buffer->putaddr);
599 		buffer->get = nvif_rd32(device, buffer->getaddr);
600 		if (buffer->get == buffer->put)
601 			return NVIF_NOTIFY_KEEP;
602 	}
603 	buffer->fault_nr = 0;
604 
605 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
606 	while (buffer->get != buffer->put) {
607 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
608 		if (++buffer->get == buffer->entries)
609 			buffer->get = 0;
610 	}
611 	nvif_wr32(device, buffer->getaddr, buffer->get);
612 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
613 
614 	/* Sort parsed faults by instance pointer to prevent unnecessary
615 	 * instance to SVMM translations, followed by address and access
616 	 * type to reduce the amount of work when handling the faults.
617 	 */
618 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
619 	     nouveau_svm_fault_cmp, NULL);
620 
621 	/* Lookup SVMM structure for each unique instance pointer. */
622 	mutex_lock(&svm->mutex);
623 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
624 		if (!svmm || buffer->fault[fi]->inst != inst) {
625 			struct nouveau_ivmm *ivmm =
626 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
627 			svmm = ivmm ? ivmm->svmm : NULL;
628 			inst = buffer->fault[fi]->inst;
629 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
630 		}
631 		buffer->fault[fi]->svmm = svmm;
632 	}
633 	mutex_unlock(&svm->mutex);
634 
635 	/* Process list of faults. */
636 	args.i.i.version = 0;
637 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
638 	args.i.m.version = 0;
639 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
640 	args.i.p.version = 0;
641 
642 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
643 		struct svm_notifier notifier;
644 		struct mm_struct *mm;
645 
646 		/* Cancel any faults from non-SVM channels. */
647 		if (!(svmm = buffer->fault[fi]->svmm)) {
648 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
649 			continue;
650 		}
651 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
652 
653 		/* We try and group handling of faults within a small
654 		 * window into a single update.
655 		 */
656 		start = buffer->fault[fi]->addr;
657 		limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT);
658 		if (start < svmm->unmanaged.limit)
659 			limit = min_t(u64, limit, svmm->unmanaged.start);
660 		else
661 		if (limit > svmm->unmanaged.start)
662 			start = max_t(u64, start, svmm->unmanaged.limit);
663 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
664 
665 		mm = svmm->notifier.mm;
666 		if (!mmget_not_zero(mm)) {
667 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
668 			continue;
669 		}
670 
671 		/* Intersect fault window with the CPU VMA, cancelling
672 		 * the fault if the address is invalid.
673 		 */
674 		down_read(&mm->mmap_sem);
675 		vma = find_vma_intersection(mm, start, limit);
676 		if (!vma) {
677 			SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit);
678 			up_read(&mm->mmap_sem);
679 			mmput(mm);
680 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
681 			continue;
682 		}
683 		start = max_t(u64, start, vma->vm_start);
684 		limit = min_t(u64, limit, vma->vm_end);
685 		up_read(&mm->mmap_sem);
686 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
687 
688 		if (buffer->fault[fi]->addr != start) {
689 			SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr);
690 			mmput(mm);
691 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
692 			continue;
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.page = PAGE_SHIFT;
700 		args.i.p.addr = start;
701 		for (fn = fi, pi = 0;;) {
702 			/* Determine required permissions based on GPU fault
703 			 * access flags.
704 			 *XXX: atomic?
705 			 */
706 			if (buffer->fault[fn]->access != 0 /* READ. */ &&
707 			    buffer->fault[fn]->access != 3 /* PREFETCH. */) {
708 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V |
709 						  NVIF_VMM_PFNMAP_V0_W;
710 			} else {
711 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V;
712 			}
713 			args.i.p.size = pi << PAGE_SHIFT;
714 
715 			/* It's okay to skip over duplicate addresses from the
716 			 * same SVMM as faults are ordered by access type such
717 			 * that only the first one needs to be handled.
718 			 *
719 			 * ie. WRITE faults appear first, thus any handling of
720 			 * pending READ faults will already be satisfied.
721 			 */
722 			while (++fn < buffer->fault_nr &&
723 			       buffer->fault[fn]->svmm == svmm &&
724 			       buffer->fault[fn    ]->addr ==
725 			       buffer->fault[fn - 1]->addr);
726 
727 			/* If the next fault is outside the window, or all GPU
728 			 * faults have been dealt with, we're done here.
729 			 */
730 			if (fn >= buffer->fault_nr ||
731 			    buffer->fault[fn]->svmm != svmm ||
732 			    buffer->fault[fn]->addr >= limit)
733 				break;
734 
735 			/* Fill in the gap between this fault and the next. */
736 			fill = (buffer->fault[fn    ]->addr -
737 				buffer->fault[fn - 1]->addr) >> PAGE_SHIFT;
738 			while (--fill)
739 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_NONE;
740 		}
741 
742 		SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)",
743 			 args.i.p.addr,
744 			 args.i.p.addr + args.i.p.size, fn - fi);
745 
746 		notifier.svmm = svmm;
747 		ret = mmu_interval_notifier_insert(&notifier.notifier,
748 						   svmm->notifier.mm,
749 						   args.i.p.addr, args.i.p.size,
750 						   &nouveau_svm_mni_ops);
751 		if (!ret) {
752 			ret = nouveau_range_fault(
753 				svmm, svm->drm, &args,
754 				sizeof(args.i) + pi * sizeof(args.phys[0]),
755 				args.phys, &notifier);
756 			mmu_interval_notifier_remove(&notifier.notifier);
757 		}
758 		mmput(mm);
759 
760 		/* Cancel any faults in the window whose pages didn't manage
761 		 * to keep their valid bit, or stay writeable when required.
762 		 *
763 		 * If handling failed completely, cancel all faults.
764 		 */
765 		while (fi < fn) {
766 			struct nouveau_svm_fault *fault = buffer->fault[fi++];
767 			pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT;
768 			if (ret ||
769 			     !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) ||
770 			    (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) &&
771 			     fault->access != 0 && fault->access != 3)) {
772 				nouveau_svm_fault_cancel_fault(svm, fault);
773 				continue;
774 			}
775 			replay++;
776 		}
777 	}
778 
779 	/* Issue fault replay to the GPU. */
780 	if (replay)
781 		nouveau_svm_fault_replay(svm);
782 	return NVIF_NOTIFY_KEEP;
783 }
784 
785 static void
786 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
787 {
788 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
789 	nvif_notify_put(&buffer->notify);
790 }
791 
792 static int
793 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
794 {
795 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
796 	struct nvif_object *device = &svm->drm->client.device.object;
797 	buffer->get = nvif_rd32(device, buffer->getaddr);
798 	buffer->put = nvif_rd32(device, buffer->putaddr);
799 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
800 	return nvif_notify_get(&buffer->notify);
801 }
802 
803 static void
804 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
805 {
806 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
807 	int i;
808 
809 	if (buffer->fault) {
810 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
811 			kfree(buffer->fault[i]);
812 		kvfree(buffer->fault);
813 	}
814 
815 	nouveau_svm_fault_buffer_fini(svm, id);
816 
817 	nvif_notify_fini(&buffer->notify);
818 	nvif_object_fini(&buffer->object);
819 }
820 
821 static int
822 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
823 {
824 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
825 	struct nouveau_drm *drm = svm->drm;
826 	struct nvif_object *device = &drm->client.device.object;
827 	struct nvif_clb069_v0 args = {};
828 	int ret;
829 
830 	buffer->id = id;
831 
832 	ret = nvif_object_init(device, 0, oclass, &args, sizeof(args),
833 			       &buffer->object);
834 	if (ret < 0) {
835 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
836 		return ret;
837 	}
838 
839 	nvif_object_map(&buffer->object, NULL, 0);
840 	buffer->entries = args.entries;
841 	buffer->getaddr = args.get;
842 	buffer->putaddr = args.put;
843 
844 	ret = nvif_notify_init(&buffer->object, nouveau_svm_fault, true,
845 			       NVB069_V0_NTFY_FAULT, NULL, 0, 0,
846 			       &buffer->notify);
847 	if (ret)
848 		return ret;
849 
850 	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
851 	if (!buffer->fault)
852 		return -ENOMEM;
853 
854 	return nouveau_svm_fault_buffer_init(svm, id);
855 }
856 
857 void
858 nouveau_svm_resume(struct nouveau_drm *drm)
859 {
860 	struct nouveau_svm *svm = drm->svm;
861 	if (svm)
862 		nouveau_svm_fault_buffer_init(svm, 0);
863 }
864 
865 void
866 nouveau_svm_suspend(struct nouveau_drm *drm)
867 {
868 	struct nouveau_svm *svm = drm->svm;
869 	if (svm)
870 		nouveau_svm_fault_buffer_fini(svm, 0);
871 }
872 
873 void
874 nouveau_svm_fini(struct nouveau_drm *drm)
875 {
876 	struct nouveau_svm *svm = drm->svm;
877 	if (svm) {
878 		nouveau_svm_fault_buffer_dtor(svm, 0);
879 		kfree(drm->svm);
880 		drm->svm = NULL;
881 	}
882 }
883 
884 void
885 nouveau_svm_init(struct nouveau_drm *drm)
886 {
887 	static const struct nvif_mclass buffers[] = {
888 		{   VOLTA_FAULT_BUFFER_A, 0 },
889 		{ MAXWELL_FAULT_BUFFER_A, 0 },
890 		{}
891 	};
892 	struct nouveau_svm *svm;
893 	int ret;
894 
895 	/* Disable on Volta and newer until channel recovery is fixed,
896 	 * otherwise clients will have a trivial way to trash the GPU
897 	 * for everyone.
898 	 */
899 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
900 		return;
901 
902 	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
903 		return;
904 
905 	drm->svm->drm = drm;
906 	mutex_init(&drm->svm->mutex);
907 	INIT_LIST_HEAD(&drm->svm->inst);
908 
909 	ret = nvif_mclass(&drm->client.device.object, buffers);
910 	if (ret < 0) {
911 		SVM_DBG(svm, "No supported fault buffer class");
912 		nouveau_svm_fini(drm);
913 		return;
914 	}
915 
916 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
917 	if (ret) {
918 		nouveau_svm_fini(drm);
919 		return;
920 	}
921 
922 	SVM_DBG(svm, "Initialised");
923 }
924