xref: /openbmc/linux/drivers/gpu/drm/nouveau/nouveau_svm.c (revision 06ff634c0dae791c17ceeeb60c74e14470d76898)
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 struct nouveau_svmm {
97 	struct mmu_notifier notifier;
98 	struct nouveau_vmm *vmm;
99 	struct {
100 		unsigned long start;
101 		unsigned long limit;
102 	} unmanaged;
103 
104 	struct mutex mutex;
105 };
106 
107 #define SVMM_DBG(s,f,a...)                                                     \
108 	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
109 #define SVMM_ERR(s,f,a...)                                                     \
110 	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
111 
112 int
113 nouveau_svmm_bind(struct drm_device *dev, void *data,
114 		  struct drm_file *file_priv)
115 {
116 	struct nouveau_cli *cli = nouveau_cli(file_priv);
117 	struct drm_nouveau_svm_bind *args = data;
118 	unsigned target, cmd, priority;
119 	unsigned long addr, end, size;
120 	struct mm_struct *mm;
121 
122 	args->va_start &= PAGE_MASK;
123 	args->va_end &= PAGE_MASK;
124 
125 	/* Sanity check arguments */
126 	if (args->reserved0 || args->reserved1)
127 		return -EINVAL;
128 	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
129 		return -EINVAL;
130 	if (args->va_start >= args->va_end)
131 		return -EINVAL;
132 	if (!args->npages)
133 		return -EINVAL;
134 
135 	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
136 	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
137 	switch (cmd) {
138 	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
139 		break;
140 	default:
141 		return -EINVAL;
142 	}
143 
144 	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
145 	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
146 
147 	/* FIXME support CPU target ie all target value < GPU_VRAM */
148 	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
149 	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
150 	switch (target) {
151 	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
152 		break;
153 	default:
154 		return -EINVAL;
155 	}
156 
157 	/*
158 	 * FIXME: For now refuse non 0 stride, we need to change the migrate
159 	 * kernel function to handle stride to avoid to create a mess within
160 	 * each device driver.
161 	 */
162 	if (args->stride)
163 		return -EINVAL;
164 
165 	size = ((unsigned long)args->npages) << PAGE_SHIFT;
166 	if ((args->va_start + size) <= args->va_start)
167 		return -EINVAL;
168 	if ((args->va_start + size) > args->va_end)
169 		return -EINVAL;
170 
171 	/*
172 	 * Ok we are ask to do something sane, for now we only support migrate
173 	 * commands but we will add things like memory policy (what to do on
174 	 * page fault) and maybe some other commands.
175 	 */
176 
177 	mm = get_task_mm(current);
178 	down_read(&mm->mmap_sem);
179 
180 	if (!cli->svm.svmm) {
181 		up_read(&mm->mmap_sem);
182 		return -EINVAL;
183 	}
184 
185 	for (addr = args->va_start, end = args->va_start + size; addr < end;) {
186 		struct vm_area_struct *vma;
187 		unsigned long next;
188 
189 		vma = find_vma_intersection(mm, addr, end);
190 		if (!vma)
191 			break;
192 
193 		addr = max(addr, vma->vm_start);
194 		next = min(vma->vm_end, end);
195 		/* This is a best effort so we ignore errors */
196 		nouveau_dmem_migrate_vma(cli->drm, cli->svm.svmm, vma, addr,
197 					 next);
198 		addr = next;
199 	}
200 
201 	/*
202 	 * FIXME Return the number of page we have migrated, again we need to
203 	 * update the migrate API to return that information so that we can
204 	 * report it to user space.
205 	 */
206 	args->result = 0;
207 
208 	up_read(&mm->mmap_sem);
209 	mmput(mm);
210 
211 	return 0;
212 }
213 
214 /* Unlink channel instance from SVMM. */
215 void
216 nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
217 {
218 	struct nouveau_ivmm *ivmm;
219 	if (svmm) {
220 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
221 		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
222 		if (ivmm) {
223 			list_del(&ivmm->head);
224 			kfree(ivmm);
225 		}
226 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
227 	}
228 }
229 
230 /* Link channel instance to SVMM. */
231 int
232 nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
233 {
234 	struct nouveau_ivmm *ivmm;
235 	if (svmm) {
236 		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
237 			return -ENOMEM;
238 		ivmm->svmm = svmm;
239 		ivmm->inst = inst;
240 
241 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
242 		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
243 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
244 	}
245 	return 0;
246 }
247 
248 /* Invalidate SVMM address-range on GPU. */
249 static void
250 nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
251 {
252 	if (limit > start) {
253 		bool super = svmm->vmm->vmm.object.client->super;
254 		svmm->vmm->vmm.object.client->super = true;
255 		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
256 				 &(struct nvif_vmm_pfnclr_v0) {
257 					.addr = start,
258 					.size = limit - start,
259 				 }, sizeof(struct nvif_vmm_pfnclr_v0));
260 		svmm->vmm->vmm.object.client->super = super;
261 	}
262 }
263 
264 static int
265 nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
266 				    const struct mmu_notifier_range *update)
267 {
268 	struct nouveau_svmm *svmm =
269 		container_of(mn, struct nouveau_svmm, notifier);
270 	unsigned long start = update->start;
271 	unsigned long limit = update->end;
272 
273 	if (!mmu_notifier_range_blockable(update))
274 		return -EAGAIN;
275 
276 	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
277 
278 	mutex_lock(&svmm->mutex);
279 	if (unlikely(!svmm->vmm))
280 		goto out;
281 
282 	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
283 		if (start < svmm->unmanaged.start) {
284 			nouveau_svmm_invalidate(svmm, start,
285 						svmm->unmanaged.limit);
286 		}
287 		start = svmm->unmanaged.limit;
288 	}
289 
290 	nouveau_svmm_invalidate(svmm, start, limit);
291 
292 out:
293 	mutex_unlock(&svmm->mutex);
294 	return 0;
295 }
296 
297 static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
298 {
299 	kfree(container_of(mn, struct nouveau_svmm, notifier));
300 }
301 
302 static const struct mmu_notifier_ops nouveau_mn_ops = {
303 	.invalidate_range_start = nouveau_svmm_invalidate_range_start,
304 	.free_notifier = nouveau_svmm_free_notifier,
305 };
306 
307 void
308 nouveau_svmm_fini(struct nouveau_svmm **psvmm)
309 {
310 	struct nouveau_svmm *svmm = *psvmm;
311 	if (svmm) {
312 		mutex_lock(&svmm->mutex);
313 		svmm->vmm = NULL;
314 		mutex_unlock(&svmm->mutex);
315 		mmu_notifier_put(&svmm->notifier);
316 		*psvmm = NULL;
317 	}
318 }
319 
320 int
321 nouveau_svmm_init(struct drm_device *dev, void *data,
322 		  struct drm_file *file_priv)
323 {
324 	struct nouveau_cli *cli = nouveau_cli(file_priv);
325 	struct nouveau_svmm *svmm;
326 	struct drm_nouveau_svm_init *args = data;
327 	int ret;
328 
329 	/* Allocate tracking for SVM-enabled VMM. */
330 	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
331 		return -ENOMEM;
332 	svmm->vmm = &cli->svm;
333 	svmm->unmanaged.start = args->unmanaged_addr;
334 	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
335 	mutex_init(&svmm->mutex);
336 
337 	/* Check that SVM isn't already enabled for the client. */
338 	mutex_lock(&cli->mutex);
339 	if (cli->svm.cli) {
340 		ret = -EBUSY;
341 		goto out_free;
342 	}
343 
344 	/* Allocate a new GPU VMM that can support SVM (managed by the
345 	 * client, with replayable faults enabled).
346 	 *
347 	 * All future channel/memory allocations will make use of this
348 	 * VMM instead of the standard one.
349 	 */
350 	ret = nvif_vmm_init(&cli->mmu, cli->vmm.vmm.object.oclass, true,
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 	down_write(&current->mm->mmap_sem);
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 	up_write(&current->mm->mmap_sem);
368 	mutex_unlock(&cli->mutex);
369 	return 0;
370 
371 out_mm_unlock:
372 	up_write(&current->mm->mmap_sem);
373 out_free:
374 	mutex_unlock(&cli->mutex);
375 	kfree(svmm);
376 	return ret;
377 }
378 
379 static const u64
380 nouveau_svm_pfn_flags[HMM_PFN_FLAG_MAX] = {
381 	[HMM_PFN_VALID         ] = NVIF_VMM_PFNMAP_V0_V,
382 	[HMM_PFN_WRITE         ] = NVIF_VMM_PFNMAP_V0_W,
383 };
384 
385 static const u64
386 nouveau_svm_pfn_values[HMM_PFN_VALUE_MAX] = {
387 	[HMM_PFN_ERROR  ] = ~NVIF_VMM_PFNMAP_V0_V,
388 	[HMM_PFN_NONE   ] =  NVIF_VMM_PFNMAP_V0_NONE,
389 	[HMM_PFN_SPECIAL] = ~NVIF_VMM_PFNMAP_V0_V,
390 };
391 
392 /* Issue fault replay for GPU to retry accesses that faulted previously. */
393 static void
394 nouveau_svm_fault_replay(struct nouveau_svm *svm)
395 {
396 	SVM_DBG(svm, "replay");
397 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
398 				 GP100_VMM_VN_FAULT_REPLAY,
399 				 &(struct gp100_vmm_fault_replay_vn) {},
400 				 sizeof(struct gp100_vmm_fault_replay_vn)));
401 }
402 
403 /* Cancel a replayable fault that could not be handled.
404  *
405  * Cancelling the fault will trigger recovery to reset the engine
406  * and kill the offending channel (ie. GPU SIGSEGV).
407  */
408 static void
409 nouveau_svm_fault_cancel(struct nouveau_svm *svm,
410 			 u64 inst, u8 hub, u8 gpc, u8 client)
411 {
412 	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
413 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
414 				 GP100_VMM_VN_FAULT_CANCEL,
415 				 &(struct gp100_vmm_fault_cancel_v0) {
416 					.hub = hub,
417 					.gpc = gpc,
418 					.client = client,
419 					.inst = inst,
420 				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
421 }
422 
423 static void
424 nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
425 			       struct nouveau_svm_fault *fault)
426 {
427 	nouveau_svm_fault_cancel(svm, fault->inst,
428 				      fault->hub,
429 				      fault->gpc,
430 				      fault->client);
431 }
432 
433 static int
434 nouveau_svm_fault_cmp(const void *a, const void *b)
435 {
436 	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
437 	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
438 	int ret;
439 	if ((ret = (s64)fa->inst - fb->inst))
440 		return ret;
441 	if ((ret = (s64)fa->addr - fb->addr))
442 		return ret;
443 	/*XXX: atomic? */
444 	return (fa->access == 0 || fa->access == 3) -
445 	       (fb->access == 0 || fb->access == 3);
446 }
447 
448 static void
449 nouveau_svm_fault_cache(struct nouveau_svm *svm,
450 			struct nouveau_svm_fault_buffer *buffer, u32 offset)
451 {
452 	struct nvif_object *memory = &buffer->object;
453 	const u32 instlo = nvif_rd32(memory, offset + 0x00);
454 	const u32 insthi = nvif_rd32(memory, offset + 0x04);
455 	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
456 	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
457 	const u32 timelo = nvif_rd32(memory, offset + 0x10);
458 	const u32 timehi = nvif_rd32(memory, offset + 0x14);
459 	const u32 engine = nvif_rd32(memory, offset + 0x18);
460 	const u32   info = nvif_rd32(memory, offset + 0x1c);
461 	const u64   inst = (u64)insthi << 32 | instlo;
462 	const u8     gpc = (info & 0x1f000000) >> 24;
463 	const u8     hub = (info & 0x00100000) >> 20;
464 	const u8  client = (info & 0x00007f00) >> 8;
465 	struct nouveau_svm_fault *fault;
466 
467 	//XXX: i think we're supposed to spin waiting */
468 	if (WARN_ON(!(info & 0x80000000)))
469 		return;
470 
471 	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
472 
473 	if (!buffer->fault[buffer->fault_nr]) {
474 		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
475 		if (WARN_ON(!fault)) {
476 			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
477 			return;
478 		}
479 		buffer->fault[buffer->fault_nr] = fault;
480 	}
481 
482 	fault = buffer->fault[buffer->fault_nr++];
483 	fault->inst   = inst;
484 	fault->addr   = (u64)addrhi << 32 | addrlo;
485 	fault->time   = (u64)timehi << 32 | timelo;
486 	fault->engine = engine;
487 	fault->gpc    = gpc;
488 	fault->hub    = hub;
489 	fault->access = (info & 0x000f0000) >> 16;
490 	fault->client = client;
491 	fault->fault  = (info & 0x0000001f);
492 
493 	SVM_DBG(svm, "fault %016llx %016llx %02x",
494 		fault->inst, fault->addr, fault->access);
495 }
496 
497 struct svm_notifier {
498 	struct mmu_interval_notifier notifier;
499 	struct nouveau_svmm *svmm;
500 };
501 
502 static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
503 					 const struct mmu_notifier_range *range,
504 					 unsigned long cur_seq)
505 {
506 	struct svm_notifier *sn =
507 		container_of(mni, struct svm_notifier, notifier);
508 
509 	/*
510 	 * serializes the update to mni->invalidate_seq done by caller and
511 	 * prevents invalidation of the PTE from progressing while HW is being
512 	 * programmed. This is very hacky and only works because the normal
513 	 * notifier that does invalidation is always called after the range
514 	 * notifier.
515 	 */
516 	if (mmu_notifier_range_blockable(range))
517 		mutex_lock(&sn->svmm->mutex);
518 	else if (!mutex_trylock(&sn->svmm->mutex))
519 		return false;
520 	mmu_interval_set_seq(mni, cur_seq);
521 	mutex_unlock(&sn->svmm->mutex);
522 	return true;
523 }
524 
525 static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
526 	.invalidate = nouveau_svm_range_invalidate,
527 };
528 
529 static int nouveau_range_fault(struct nouveau_svmm *svmm,
530 			       struct nouveau_drm *drm, void *data, u32 size,
531 			       u64 *pfns, struct svm_notifier *notifier)
532 {
533 	unsigned long timeout =
534 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
535 	/* Have HMM fault pages within the fault window to the GPU. */
536 	struct hmm_range range = {
537 		.notifier = &notifier->notifier,
538 		.start = notifier->notifier.interval_tree.start,
539 		.end = notifier->notifier.interval_tree.last + 1,
540 		.pfns = pfns,
541 		.flags = nouveau_svm_pfn_flags,
542 		.values = nouveau_svm_pfn_values,
543 		.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT,
544 	};
545 	struct mm_struct *mm = notifier->notifier.mm;
546 	long ret;
547 
548 	while (true) {
549 		if (time_after(jiffies, timeout))
550 			return -EBUSY;
551 
552 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
553 		range.default_flags = 0;
554 		range.pfn_flags_mask = -1UL;
555 		down_read(&mm->mmap_sem);
556 		ret = hmm_range_fault(&range);
557 		up_read(&mm->mmap_sem);
558 		if (ret <= 0) {
559 			if (ret == 0 || ret == -EBUSY)
560 				continue;
561 			return ret;
562 		}
563 
564 		mutex_lock(&svmm->mutex);
565 		if (mmu_interval_read_retry(range.notifier,
566 					    range.notifier_seq)) {
567 			mutex_unlock(&svmm->mutex);
568 			continue;
569 		}
570 		break;
571 	}
572 
573 	nouveau_dmem_convert_pfn(drm, &range);
574 
575 	svmm->vmm->vmm.object.client->super = true;
576 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, data, size, NULL);
577 	svmm->vmm->vmm.object.client->super = false;
578 	mutex_unlock(&svmm->mutex);
579 
580 	return ret;
581 }
582 
583 static int
584 nouveau_svm_fault(struct nvif_notify *notify)
585 {
586 	struct nouveau_svm_fault_buffer *buffer =
587 		container_of(notify, typeof(*buffer), notify);
588 	struct nouveau_svm *svm =
589 		container_of(buffer, typeof(*svm), buffer[buffer->id]);
590 	struct nvif_object *device = &svm->drm->client.device.object;
591 	struct nouveau_svmm *svmm;
592 	struct {
593 		struct {
594 			struct nvif_ioctl_v0 i;
595 			struct nvif_ioctl_mthd_v0 m;
596 			struct nvif_vmm_pfnmap_v0 p;
597 		} i;
598 		u64 phys[16];
599 	} args;
600 	struct vm_area_struct *vma;
601 	u64 inst, start, limit;
602 	int fi, fn, pi, fill;
603 	int replay = 0, ret;
604 
605 	/* Parse available fault buffer entries into a cache, and update
606 	 * the GET pointer so HW can reuse the entries.
607 	 */
608 	SVM_DBG(svm, "fault handler");
609 	if (buffer->get == buffer->put) {
610 		buffer->put = nvif_rd32(device, buffer->putaddr);
611 		buffer->get = nvif_rd32(device, buffer->getaddr);
612 		if (buffer->get == buffer->put)
613 			return NVIF_NOTIFY_KEEP;
614 	}
615 	buffer->fault_nr = 0;
616 
617 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
618 	while (buffer->get != buffer->put) {
619 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
620 		if (++buffer->get == buffer->entries)
621 			buffer->get = 0;
622 	}
623 	nvif_wr32(device, buffer->getaddr, buffer->get);
624 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
625 
626 	/* Sort parsed faults by instance pointer to prevent unnecessary
627 	 * instance to SVMM translations, followed by address and access
628 	 * type to reduce the amount of work when handling the faults.
629 	 */
630 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
631 	     nouveau_svm_fault_cmp, NULL);
632 
633 	/* Lookup SVMM structure for each unique instance pointer. */
634 	mutex_lock(&svm->mutex);
635 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
636 		if (!svmm || buffer->fault[fi]->inst != inst) {
637 			struct nouveau_ivmm *ivmm =
638 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
639 			svmm = ivmm ? ivmm->svmm : NULL;
640 			inst = buffer->fault[fi]->inst;
641 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
642 		}
643 		buffer->fault[fi]->svmm = svmm;
644 	}
645 	mutex_unlock(&svm->mutex);
646 
647 	/* Process list of faults. */
648 	args.i.i.version = 0;
649 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
650 	args.i.m.version = 0;
651 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
652 	args.i.p.version = 0;
653 
654 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
655 		struct svm_notifier notifier;
656 		struct mm_struct *mm;
657 
658 		/* Cancel any faults from non-SVM channels. */
659 		if (!(svmm = buffer->fault[fi]->svmm)) {
660 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
661 			continue;
662 		}
663 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
664 
665 		/* We try and group handling of faults within a small
666 		 * window into a single update.
667 		 */
668 		start = buffer->fault[fi]->addr;
669 		limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT);
670 		if (start < svmm->unmanaged.limit)
671 			limit = min_t(u64, limit, svmm->unmanaged.start);
672 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
673 
674 		mm = svmm->notifier.mm;
675 		if (!mmget_not_zero(mm)) {
676 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
677 			continue;
678 		}
679 
680 		/* Intersect fault window with the CPU VMA, cancelling
681 		 * the fault if the address is invalid.
682 		 */
683 		down_read(&mm->mmap_sem);
684 		vma = find_vma_intersection(mm, start, limit);
685 		if (!vma) {
686 			SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit);
687 			up_read(&mm->mmap_sem);
688 			mmput(mm);
689 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
690 			continue;
691 		}
692 		start = max_t(u64, start, vma->vm_start);
693 		limit = min_t(u64, limit, vma->vm_end);
694 		up_read(&mm->mmap_sem);
695 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
696 
697 		if (buffer->fault[fi]->addr != start) {
698 			SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr);
699 			mmput(mm);
700 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
701 			continue;
702 		}
703 
704 		/* Prepare the GPU-side update of all pages within the
705 		 * fault window, determining required pages and access
706 		 * permissions based on pending faults.
707 		 */
708 		args.i.p.page = PAGE_SHIFT;
709 		args.i.p.addr = start;
710 		for (fn = fi, pi = 0;;) {
711 			/* Determine required permissions based on GPU fault
712 			 * access flags.
713 			 *XXX: atomic?
714 			 */
715 			if (buffer->fault[fn]->access != 0 /* READ. */ &&
716 			    buffer->fault[fn]->access != 3 /* PREFETCH. */) {
717 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V |
718 						  NVIF_VMM_PFNMAP_V0_W;
719 			} else {
720 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V;
721 			}
722 			args.i.p.size = pi << PAGE_SHIFT;
723 
724 			/* It's okay to skip over duplicate addresses from the
725 			 * same SVMM as faults are ordered by access type such
726 			 * that only the first one needs to be handled.
727 			 *
728 			 * ie. WRITE faults appear first, thus any handling of
729 			 * pending READ faults will already be satisfied.
730 			 */
731 			while (++fn < buffer->fault_nr &&
732 			       buffer->fault[fn]->svmm == svmm &&
733 			       buffer->fault[fn    ]->addr ==
734 			       buffer->fault[fn - 1]->addr);
735 
736 			/* If the next fault is outside the window, or all GPU
737 			 * faults have been dealt with, we're done here.
738 			 */
739 			if (fn >= buffer->fault_nr ||
740 			    buffer->fault[fn]->svmm != svmm ||
741 			    buffer->fault[fn]->addr >= limit)
742 				break;
743 
744 			/* Fill in the gap between this fault and the next. */
745 			fill = (buffer->fault[fn    ]->addr -
746 				buffer->fault[fn - 1]->addr) >> PAGE_SHIFT;
747 			while (--fill)
748 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_NONE;
749 		}
750 
751 		SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)",
752 			 args.i.p.addr,
753 			 args.i.p.addr + args.i.p.size, fn - fi);
754 
755 		notifier.svmm = svmm;
756 		ret = mmu_interval_notifier_insert(&notifier.notifier,
757 						   svmm->notifier.mm,
758 						   args.i.p.addr, args.i.p.size,
759 						   &nouveau_svm_mni_ops);
760 		if (!ret) {
761 			ret = nouveau_range_fault(
762 				svmm, svm->drm, &args,
763 				sizeof(args.i) + pi * sizeof(args.phys[0]),
764 				args.phys, &notifier);
765 			mmu_interval_notifier_remove(&notifier.notifier);
766 		}
767 		mmput(mm);
768 
769 		/* Cancel any faults in the window whose pages didn't manage
770 		 * to keep their valid bit, or stay writeable when required.
771 		 *
772 		 * If handling failed completely, cancel all faults.
773 		 */
774 		while (fi < fn) {
775 			struct nouveau_svm_fault *fault = buffer->fault[fi++];
776 			pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT;
777 			if (ret ||
778 			     !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) ||
779 			    (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) &&
780 			     fault->access != 0 && fault->access != 3)) {
781 				nouveau_svm_fault_cancel_fault(svm, fault);
782 				continue;
783 			}
784 			replay++;
785 		}
786 	}
787 
788 	/* Issue fault replay to the GPU. */
789 	if (replay)
790 		nouveau_svm_fault_replay(svm);
791 	return NVIF_NOTIFY_KEEP;
792 }
793 
794 static struct nouveau_pfnmap_args *
795 nouveau_pfns_to_args(void *pfns)
796 {
797 	return container_of(pfns, struct nouveau_pfnmap_args, p.phys);
798 }
799 
800 u64 *
801 nouveau_pfns_alloc(unsigned long npages)
802 {
803 	struct nouveau_pfnmap_args *args;
804 
805 	args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL);
806 	if (!args)
807 		return NULL;
808 
809 	args->i.type = NVIF_IOCTL_V0_MTHD;
810 	args->m.method = NVIF_VMM_V0_PFNMAP;
811 	args->p.page = PAGE_SHIFT;
812 
813 	return args->p.phys;
814 }
815 
816 void
817 nouveau_pfns_free(u64 *pfns)
818 {
819 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
820 
821 	kfree(args);
822 }
823 
824 void
825 nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm,
826 		 unsigned long addr, u64 *pfns, unsigned long npages)
827 {
828 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
829 	int ret;
830 
831 	args->p.addr = addr;
832 	args->p.size = npages << PAGE_SHIFT;
833 
834 	mutex_lock(&svmm->mutex);
835 
836 	svmm->vmm->vmm.object.client->super = true;
837 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, sizeof(*args) +
838 				npages * sizeof(args->p.phys[0]), NULL);
839 	svmm->vmm->vmm.object.client->super = false;
840 
841 	mutex_unlock(&svmm->mutex);
842 }
843 
844 static void
845 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
846 {
847 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
848 	nvif_notify_put(&buffer->notify);
849 }
850 
851 static int
852 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
853 {
854 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
855 	struct nvif_object *device = &svm->drm->client.device.object;
856 	buffer->get = nvif_rd32(device, buffer->getaddr);
857 	buffer->put = nvif_rd32(device, buffer->putaddr);
858 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
859 	return nvif_notify_get(&buffer->notify);
860 }
861 
862 static void
863 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
864 {
865 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
866 	int i;
867 
868 	if (buffer->fault) {
869 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
870 			kfree(buffer->fault[i]);
871 		kvfree(buffer->fault);
872 	}
873 
874 	nouveau_svm_fault_buffer_fini(svm, id);
875 
876 	nvif_notify_fini(&buffer->notify);
877 	nvif_object_fini(&buffer->object);
878 }
879 
880 static int
881 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
882 {
883 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
884 	struct nouveau_drm *drm = svm->drm;
885 	struct nvif_object *device = &drm->client.device.object;
886 	struct nvif_clb069_v0 args = {};
887 	int ret;
888 
889 	buffer->id = id;
890 
891 	ret = nvif_object_init(device, 0, oclass, &args, sizeof(args),
892 			       &buffer->object);
893 	if (ret < 0) {
894 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
895 		return ret;
896 	}
897 
898 	nvif_object_map(&buffer->object, NULL, 0);
899 	buffer->entries = args.entries;
900 	buffer->getaddr = args.get;
901 	buffer->putaddr = args.put;
902 
903 	ret = nvif_notify_init(&buffer->object, nouveau_svm_fault, true,
904 			       NVB069_V0_NTFY_FAULT, NULL, 0, 0,
905 			       &buffer->notify);
906 	if (ret)
907 		return ret;
908 
909 	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
910 	if (!buffer->fault)
911 		return -ENOMEM;
912 
913 	return nouveau_svm_fault_buffer_init(svm, id);
914 }
915 
916 void
917 nouveau_svm_resume(struct nouveau_drm *drm)
918 {
919 	struct nouveau_svm *svm = drm->svm;
920 	if (svm)
921 		nouveau_svm_fault_buffer_init(svm, 0);
922 }
923 
924 void
925 nouveau_svm_suspend(struct nouveau_drm *drm)
926 {
927 	struct nouveau_svm *svm = drm->svm;
928 	if (svm)
929 		nouveau_svm_fault_buffer_fini(svm, 0);
930 }
931 
932 void
933 nouveau_svm_fini(struct nouveau_drm *drm)
934 {
935 	struct nouveau_svm *svm = drm->svm;
936 	if (svm) {
937 		nouveau_svm_fault_buffer_dtor(svm, 0);
938 		kfree(drm->svm);
939 		drm->svm = NULL;
940 	}
941 }
942 
943 void
944 nouveau_svm_init(struct nouveau_drm *drm)
945 {
946 	static const struct nvif_mclass buffers[] = {
947 		{   VOLTA_FAULT_BUFFER_A, 0 },
948 		{ MAXWELL_FAULT_BUFFER_A, 0 },
949 		{}
950 	};
951 	struct nouveau_svm *svm;
952 	int ret;
953 
954 	/* Disable on Volta and newer until channel recovery is fixed,
955 	 * otherwise clients will have a trivial way to trash the GPU
956 	 * for everyone.
957 	 */
958 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
959 		return;
960 
961 	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
962 		return;
963 
964 	drm->svm->drm = drm;
965 	mutex_init(&drm->svm->mutex);
966 	INIT_LIST_HEAD(&drm->svm->inst);
967 
968 	ret = nvif_mclass(&drm->client.device.object, buffers);
969 	if (ret < 0) {
970 		SVM_DBG(svm, "No supported fault buffer class");
971 		nouveau_svm_fini(drm);
972 		return;
973 	}
974 
975 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
976 	if (ret) {
977 		nouveau_svm_fini(drm);
978 		return;
979 	}
980 
981 	SVM_DBG(svm, "Initialised");
982 }
983