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 	mmap_read_lock(mm);
179 
180 	if (!cli->svm.svmm) {
181 		mmap_read_unlock(mm);
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 	mmap_read_unlock(mm);
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_ctor(&cli->mmu, "svmVmm",
351 			    cli->vmm.vmm.object.oclass, true,
352 			    args->unmanaged_addr, args->unmanaged_size,
353 			    &(struct gp100_vmm_v0) {
354 				.fault_replay = true,
355 			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
356 	if (ret)
357 		goto out_free;
358 
359 	mmap_write_lock(current->mm);
360 	svmm->notifier.ops = &nouveau_mn_ops;
361 	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
362 	if (ret)
363 		goto out_mm_unlock;
364 	/* Note, ownership of svmm transfers to mmu_notifier */
365 
366 	cli->svm.svmm = svmm;
367 	cli->svm.cli = cli;
368 	mmap_write_unlock(current->mm);
369 	mutex_unlock(&cli->mutex);
370 	return 0;
371 
372 out_mm_unlock:
373 	mmap_write_unlock(current->mm);
374 out_free:
375 	mutex_unlock(&cli->mutex);
376 	kfree(svmm);
377 	return ret;
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 void nouveau_hmm_convert_pfn(struct nouveau_drm *drm,
518 				    struct hmm_range *range, u64 *ioctl_addr)
519 {
520 	unsigned long i, npages;
521 
522 	/*
523 	 * The ioctl_addr prepared here is passed through nvif_object_ioctl()
524 	 * to an eventual DMA map in something like gp100_vmm_pgt_pfn()
525 	 *
526 	 * This is all just encoding the internal hmm representation into a
527 	 * different nouveau internal representation.
528 	 */
529 	npages = (range->end - range->start) >> PAGE_SHIFT;
530 	for (i = 0; i < npages; ++i) {
531 		struct page *page;
532 
533 		if (!(range->hmm_pfns[i] & HMM_PFN_VALID)) {
534 			ioctl_addr[i] = 0;
535 			continue;
536 		}
537 
538 		page = hmm_pfn_to_page(range->hmm_pfns[i]);
539 		if (is_device_private_page(page))
540 			ioctl_addr[i] = nouveau_dmem_page_addr(page) |
541 					NVIF_VMM_PFNMAP_V0_V |
542 					NVIF_VMM_PFNMAP_V0_VRAM;
543 		else
544 			ioctl_addr[i] = page_to_phys(page) |
545 					NVIF_VMM_PFNMAP_V0_V |
546 					NVIF_VMM_PFNMAP_V0_HOST;
547 		if (range->hmm_pfns[i] & HMM_PFN_WRITE)
548 			ioctl_addr[i] |= NVIF_VMM_PFNMAP_V0_W;
549 	}
550 }
551 
552 static int nouveau_range_fault(struct nouveau_svmm *svmm,
553 			       struct nouveau_drm *drm, void *data, u32 size,
554 			       unsigned long hmm_pfns[], u64 *ioctl_addr,
555 			       struct svm_notifier *notifier)
556 {
557 	unsigned long timeout =
558 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
559 	/* Have HMM fault pages within the fault window to the GPU. */
560 	struct hmm_range range = {
561 		.notifier = &notifier->notifier,
562 		.start = notifier->notifier.interval_tree.start,
563 		.end = notifier->notifier.interval_tree.last + 1,
564 		.pfn_flags_mask = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE,
565 		.hmm_pfns = hmm_pfns,
566 		.dev_private_owner = drm->dev,
567 	};
568 	struct mm_struct *mm = notifier->notifier.mm;
569 	int ret;
570 
571 	while (true) {
572 		if (time_after(jiffies, timeout))
573 			return -EBUSY;
574 
575 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
576 		mmap_read_lock(mm);
577 		ret = hmm_range_fault(&range);
578 		mmap_read_unlock(mm);
579 		if (ret) {
580 			/*
581 			 * FIXME: the input PFN_REQ flags are destroyed on
582 			 * -EBUSY, we need to regenerate them, also for the
583 			 * other continue below
584 			 */
585 			if (ret == -EBUSY)
586 				continue;
587 			return ret;
588 		}
589 
590 		mutex_lock(&svmm->mutex);
591 		if (mmu_interval_read_retry(range.notifier,
592 					    range.notifier_seq)) {
593 			mutex_unlock(&svmm->mutex);
594 			continue;
595 		}
596 		break;
597 	}
598 
599 	nouveau_hmm_convert_pfn(drm, &range, ioctl_addr);
600 
601 	svmm->vmm->vmm.object.client->super = true;
602 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, data, size, NULL);
603 	svmm->vmm->vmm.object.client->super = false;
604 	mutex_unlock(&svmm->mutex);
605 
606 	return ret;
607 }
608 
609 static int
610 nouveau_svm_fault(struct nvif_notify *notify)
611 {
612 	struct nouveau_svm_fault_buffer *buffer =
613 		container_of(notify, typeof(*buffer), notify);
614 	struct nouveau_svm *svm =
615 		container_of(buffer, typeof(*svm), buffer[buffer->id]);
616 	struct nvif_object *device = &svm->drm->client.device.object;
617 	struct nouveau_svmm *svmm;
618 	struct {
619 		struct {
620 			struct nvif_ioctl_v0 i;
621 			struct nvif_ioctl_mthd_v0 m;
622 			struct nvif_vmm_pfnmap_v0 p;
623 		} i;
624 		u64 phys[16];
625 	} args;
626 	unsigned long hmm_pfns[ARRAY_SIZE(args.phys)];
627 	struct vm_area_struct *vma;
628 	u64 inst, start, limit;
629 	int fi, fn, pi, fill;
630 	int replay = 0, ret;
631 
632 	/* Parse available fault buffer entries into a cache, and update
633 	 * the GET pointer so HW can reuse the entries.
634 	 */
635 	SVM_DBG(svm, "fault handler");
636 	if (buffer->get == buffer->put) {
637 		buffer->put = nvif_rd32(device, buffer->putaddr);
638 		buffer->get = nvif_rd32(device, buffer->getaddr);
639 		if (buffer->get == buffer->put)
640 			return NVIF_NOTIFY_KEEP;
641 	}
642 	buffer->fault_nr = 0;
643 
644 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
645 	while (buffer->get != buffer->put) {
646 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
647 		if (++buffer->get == buffer->entries)
648 			buffer->get = 0;
649 	}
650 	nvif_wr32(device, buffer->getaddr, buffer->get);
651 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
652 
653 	/* Sort parsed faults by instance pointer to prevent unnecessary
654 	 * instance to SVMM translations, followed by address and access
655 	 * type to reduce the amount of work when handling the faults.
656 	 */
657 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
658 	     nouveau_svm_fault_cmp, NULL);
659 
660 	/* Lookup SVMM structure for each unique instance pointer. */
661 	mutex_lock(&svm->mutex);
662 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
663 		if (!svmm || buffer->fault[fi]->inst != inst) {
664 			struct nouveau_ivmm *ivmm =
665 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
666 			svmm = ivmm ? ivmm->svmm : NULL;
667 			inst = buffer->fault[fi]->inst;
668 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
669 		}
670 		buffer->fault[fi]->svmm = svmm;
671 	}
672 	mutex_unlock(&svm->mutex);
673 
674 	/* Process list of faults. */
675 	args.i.i.version = 0;
676 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
677 	args.i.m.version = 0;
678 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
679 	args.i.p.version = 0;
680 
681 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
682 		struct svm_notifier notifier;
683 		struct mm_struct *mm;
684 
685 		/* Cancel any faults from non-SVM channels. */
686 		if (!(svmm = buffer->fault[fi]->svmm)) {
687 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
688 			continue;
689 		}
690 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
691 
692 		/* We try and group handling of faults within a small
693 		 * window into a single update.
694 		 */
695 		start = buffer->fault[fi]->addr;
696 		limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT);
697 		if (start < svmm->unmanaged.limit)
698 			limit = min_t(u64, limit, svmm->unmanaged.start);
699 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
700 
701 		mm = svmm->notifier.mm;
702 		if (!mmget_not_zero(mm)) {
703 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
704 			continue;
705 		}
706 
707 		/* Intersect fault window with the CPU VMA, cancelling
708 		 * the fault if the address is invalid.
709 		 */
710 		mmap_read_lock(mm);
711 		vma = find_vma_intersection(mm, start, limit);
712 		if (!vma) {
713 			SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit);
714 			mmap_read_unlock(mm);
715 			mmput(mm);
716 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
717 			continue;
718 		}
719 		start = max_t(u64, start, vma->vm_start);
720 		limit = min_t(u64, limit, vma->vm_end);
721 		mmap_read_unlock(mm);
722 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
723 
724 		if (buffer->fault[fi]->addr != start) {
725 			SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr);
726 			mmput(mm);
727 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
728 			continue;
729 		}
730 
731 		/* Prepare the GPU-side update of all pages within the
732 		 * fault window, determining required pages and access
733 		 * permissions based on pending faults.
734 		 */
735 		args.i.p.page = PAGE_SHIFT;
736 		args.i.p.addr = start;
737 		for (fn = fi, pi = 0;;) {
738 			/* Determine required permissions based on GPU fault
739 			 * access flags.
740 			 *XXX: atomic?
741 			 */
742 			switch (buffer->fault[fn]->access) {
743 			case 0: /* READ. */
744 				hmm_pfns[pi++] = HMM_PFN_REQ_FAULT;
745 				break;
746 			case 3: /* PREFETCH. */
747 				hmm_pfns[pi++] = 0;
748 				break;
749 			default:
750 				hmm_pfns[pi++] = HMM_PFN_REQ_FAULT |
751 						 HMM_PFN_REQ_WRITE;
752 				break;
753 			}
754 			args.i.p.size = pi << PAGE_SHIFT;
755 
756 			/* It's okay to skip over duplicate addresses from the
757 			 * same SVMM as faults are ordered by access type such
758 			 * that only the first one needs to be handled.
759 			 *
760 			 * ie. WRITE faults appear first, thus any handling of
761 			 * pending READ faults will already be satisfied.
762 			 */
763 			while (++fn < buffer->fault_nr &&
764 			       buffer->fault[fn]->svmm == svmm &&
765 			       buffer->fault[fn    ]->addr ==
766 			       buffer->fault[fn - 1]->addr);
767 
768 			/* If the next fault is outside the window, or all GPU
769 			 * faults have been dealt with, we're done here.
770 			 */
771 			if (fn >= buffer->fault_nr ||
772 			    buffer->fault[fn]->svmm != svmm ||
773 			    buffer->fault[fn]->addr >= limit)
774 				break;
775 
776 			/* Fill in the gap between this fault and the next. */
777 			fill = (buffer->fault[fn    ]->addr -
778 				buffer->fault[fn - 1]->addr) >> PAGE_SHIFT;
779 			while (--fill)
780 				hmm_pfns[pi++] = 0;
781 		}
782 
783 		SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)",
784 			 args.i.p.addr,
785 			 args.i.p.addr + args.i.p.size, fn - fi);
786 
787 		notifier.svmm = svmm;
788 		ret = mmu_interval_notifier_insert(&notifier.notifier,
789 						   svmm->notifier.mm,
790 						   args.i.p.addr, args.i.p.size,
791 						   &nouveau_svm_mni_ops);
792 		if (!ret) {
793 			ret = nouveau_range_fault(
794 				svmm, svm->drm, &args,
795 				sizeof(args.i) + pi * sizeof(args.phys[0]),
796 				hmm_pfns, args.phys, &notifier);
797 			mmu_interval_notifier_remove(&notifier.notifier);
798 		}
799 		mmput(mm);
800 
801 		/* Cancel any faults in the window whose pages didn't manage
802 		 * to keep their valid bit, or stay writeable when required.
803 		 *
804 		 * If handling failed completely, cancel all faults.
805 		 */
806 		while (fi < fn) {
807 			struct nouveau_svm_fault *fault = buffer->fault[fi++];
808 			pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT;
809 			if (ret ||
810 			     !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) ||
811 			    (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) &&
812 			     fault->access != 0 && fault->access != 3)) {
813 				nouveau_svm_fault_cancel_fault(svm, fault);
814 				continue;
815 			}
816 			replay++;
817 		}
818 	}
819 
820 	/* Issue fault replay to the GPU. */
821 	if (replay)
822 		nouveau_svm_fault_replay(svm);
823 	return NVIF_NOTIFY_KEEP;
824 }
825 
826 static struct nouveau_pfnmap_args *
827 nouveau_pfns_to_args(void *pfns)
828 {
829 	return container_of(pfns, struct nouveau_pfnmap_args, p.phys);
830 }
831 
832 u64 *
833 nouveau_pfns_alloc(unsigned long npages)
834 {
835 	struct nouveau_pfnmap_args *args;
836 
837 	args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL);
838 	if (!args)
839 		return NULL;
840 
841 	args->i.type = NVIF_IOCTL_V0_MTHD;
842 	args->m.method = NVIF_VMM_V0_PFNMAP;
843 	args->p.page = PAGE_SHIFT;
844 
845 	return args->p.phys;
846 }
847 
848 void
849 nouveau_pfns_free(u64 *pfns)
850 {
851 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
852 
853 	kfree(args);
854 }
855 
856 void
857 nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm,
858 		 unsigned long addr, u64 *pfns, unsigned long npages)
859 {
860 	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
861 	int ret;
862 
863 	args->p.addr = addr;
864 	args->p.size = npages << PAGE_SHIFT;
865 
866 	mutex_lock(&svmm->mutex);
867 
868 	svmm->vmm->vmm.object.client->super = true;
869 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, sizeof(*args) +
870 				npages * sizeof(args->p.phys[0]), NULL);
871 	svmm->vmm->vmm.object.client->super = false;
872 
873 	mutex_unlock(&svmm->mutex);
874 }
875 
876 static void
877 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
878 {
879 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
880 	nvif_notify_put(&buffer->notify);
881 }
882 
883 static int
884 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
885 {
886 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
887 	struct nvif_object *device = &svm->drm->client.device.object;
888 	buffer->get = nvif_rd32(device, buffer->getaddr);
889 	buffer->put = nvif_rd32(device, buffer->putaddr);
890 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
891 	return nvif_notify_get(&buffer->notify);
892 }
893 
894 static void
895 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
896 {
897 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
898 	int i;
899 
900 	if (buffer->fault) {
901 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
902 			kfree(buffer->fault[i]);
903 		kvfree(buffer->fault);
904 	}
905 
906 	nouveau_svm_fault_buffer_fini(svm, id);
907 
908 	nvif_notify_dtor(&buffer->notify);
909 	nvif_object_dtor(&buffer->object);
910 }
911 
912 static int
913 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
914 {
915 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
916 	struct nouveau_drm *drm = svm->drm;
917 	struct nvif_object *device = &drm->client.device.object;
918 	struct nvif_clb069_v0 args = {};
919 	int ret;
920 
921 	buffer->id = id;
922 
923 	ret = nvif_object_ctor(device, "svmFaultBuffer", 0, oclass, &args,
924 			       sizeof(args), &buffer->object);
925 	if (ret < 0) {
926 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
927 		return ret;
928 	}
929 
930 	nvif_object_map(&buffer->object, NULL, 0);
931 	buffer->entries = args.entries;
932 	buffer->getaddr = args.get;
933 	buffer->putaddr = args.put;
934 
935 	ret = nvif_notify_ctor(&buffer->object, "svmFault", nouveau_svm_fault,
936 			       true, NVB069_V0_NTFY_FAULT, NULL, 0, 0,
937 			       &buffer->notify);
938 	if (ret)
939 		return ret;
940 
941 	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
942 	if (!buffer->fault)
943 		return -ENOMEM;
944 
945 	return nouveau_svm_fault_buffer_init(svm, id);
946 }
947 
948 void
949 nouveau_svm_resume(struct nouveau_drm *drm)
950 {
951 	struct nouveau_svm *svm = drm->svm;
952 	if (svm)
953 		nouveau_svm_fault_buffer_init(svm, 0);
954 }
955 
956 void
957 nouveau_svm_suspend(struct nouveau_drm *drm)
958 {
959 	struct nouveau_svm *svm = drm->svm;
960 	if (svm)
961 		nouveau_svm_fault_buffer_fini(svm, 0);
962 }
963 
964 void
965 nouveau_svm_fini(struct nouveau_drm *drm)
966 {
967 	struct nouveau_svm *svm = drm->svm;
968 	if (svm) {
969 		nouveau_svm_fault_buffer_dtor(svm, 0);
970 		kfree(drm->svm);
971 		drm->svm = NULL;
972 	}
973 }
974 
975 void
976 nouveau_svm_init(struct nouveau_drm *drm)
977 {
978 	static const struct nvif_mclass buffers[] = {
979 		{   VOLTA_FAULT_BUFFER_A, 0 },
980 		{ MAXWELL_FAULT_BUFFER_A, 0 },
981 		{}
982 	};
983 	struct nouveau_svm *svm;
984 	int ret;
985 
986 	/* Disable on Volta and newer until channel recovery is fixed,
987 	 * otherwise clients will have a trivial way to trash the GPU
988 	 * for everyone.
989 	 */
990 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
991 		return;
992 
993 	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
994 		return;
995 
996 	drm->svm->drm = drm;
997 	mutex_init(&drm->svm->mutex);
998 	INIT_LIST_HEAD(&drm->svm->inst);
999 
1000 	ret = nvif_mclass(&drm->client.device.object, buffers);
1001 	if (ret < 0) {
1002 		SVM_DBG(svm, "No supported fault buffer class");
1003 		nouveau_svm_fini(drm);
1004 		return;
1005 	}
1006 
1007 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
1008 	if (ret) {
1009 		nouveau_svm_fini(drm);
1010 		return;
1011 	}
1012 
1013 	SVM_DBG(svm, "Initialised");
1014 }
1015