xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/nouveau/nouveau_svm.c (revision 798b8d11ecd8257a8e35c3396210f98abf3d9ade)
1 /*	$NetBSD: nouveau_svm.c,v 1.3 2021/12/19 11:34:44 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2018 Red Hat Inc.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  */
24 #include <sys/cdefs.h>
25 __KERNEL_RCSID(0, "$NetBSD: nouveau_svm.c,v 1.3 2021/12/19 11:34:44 riastradh Exp $");
26 
27 #include "nouveau_svm.h"
28 #include "nouveau_drv.h"
29 #include "nouveau_chan.h"
30 #include "nouveau_dmem.h"
31 
32 #include <nvif/notify.h>
33 #include <nvif/object.h>
34 #include <nvif/vmm.h>
35 
36 #include <nvif/class.h>
37 #include <nvif/clb069.h>
38 #include <nvif/ifc00d.h>
39 
40 #include <linux/sched/mm.h>
41 #include <linux/sort.h>
42 #include <linux/hmm.h>
43 
44 struct nouveau_svm {
45 	struct nouveau_drm *drm;
46 	struct mutex mutex;
47 	struct list_head inst;
48 
49 	struct nouveau_svm_fault_buffer {
50 		int id;
51 		struct nvif_object object;
52 		u32 entries;
53 		u32 getaddr;
54 		u32 putaddr;
55 		u32 get;
56 		u32 put;
57 		struct nvif_notify notify;
58 
59 		struct nouveau_svm_fault {
60 			u64 inst;
61 			u64 addr;
62 			u64 time;
63 			u32 engine;
64 			u8  gpc;
65 			u8  hub;
66 			u8  access;
67 			u8  client;
68 			u8  fault;
69 			struct nouveau_svmm *svmm;
70 		} **fault;
71 		int fault_nr;
72 	} buffer[1];
73 };
74 
75 #define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
76 #define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)
77 
78 struct nouveau_ivmm {
79 	struct nouveau_svmm *svmm;
80 	u64 inst;
81 	struct list_head head;
82 };
83 
84 static struct nouveau_ivmm *
nouveau_ivmm_find(struct nouveau_svm * svm,u64 inst)85 nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
86 {
87 	struct nouveau_ivmm *ivmm;
88 	list_for_each_entry(ivmm, &svm->inst, head) {
89 		if (ivmm->inst == inst)
90 			return ivmm;
91 	}
92 	return NULL;
93 }
94 
95 struct nouveau_svmm {
96 	struct mmu_notifier notifier;
97 	struct nouveau_vmm *vmm;
98 	struct {
99 		unsigned long start;
100 		unsigned long limit;
101 	} unmanaged;
102 
103 	struct mutex mutex;
104 };
105 
106 #define SVMM_DBG(s,f,a...)                                                     \
107 	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
108 #define SVMM_ERR(s,f,a...)                                                     \
109 	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
110 
111 int
nouveau_svmm_bind(struct drm_device * dev,void * data,struct drm_file * file_priv)112 nouveau_svmm_bind(struct drm_device *dev, void *data,
113 		  struct drm_file *file_priv)
114 {
115 	struct nouveau_cli *cli = nouveau_cli(file_priv);
116 	struct drm_nouveau_svm_bind *args = data;
117 	unsigned target, cmd, priority;
118 	unsigned long addr, end, size;
119 	struct mm_struct *mm;
120 
121 	args->va_start &= PAGE_MASK;
122 	args->va_end &= PAGE_MASK;
123 
124 	/* Sanity check arguments */
125 	if (args->reserved0 || args->reserved1)
126 		return -EINVAL;
127 	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
128 		return -EINVAL;
129 	if (args->va_start >= args->va_end)
130 		return -EINVAL;
131 	if (!args->npages)
132 		return -EINVAL;
133 
134 	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
135 	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
136 	switch (cmd) {
137 	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
138 		break;
139 	default:
140 		return -EINVAL;
141 	}
142 
143 	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
144 	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
145 
146 	/* FIXME support CPU target ie all target value < GPU_VRAM */
147 	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
148 	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
149 	switch (target) {
150 	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
151 		break;
152 	default:
153 		return -EINVAL;
154 	}
155 
156 	/*
157 	 * FIXME: For now refuse non 0 stride, we need to change the migrate
158 	 * kernel function to handle stride to avoid to create a mess within
159 	 * each device driver.
160 	 */
161 	if (args->stride)
162 		return -EINVAL;
163 
164 	size = ((unsigned long)args->npages) << PAGE_SHIFT;
165 	if ((args->va_start + size) <= args->va_start)
166 		return -EINVAL;
167 	if ((args->va_start + size) > args->va_end)
168 		return -EINVAL;
169 
170 	/*
171 	 * Ok we are ask to do something sane, for now we only support migrate
172 	 * commands but we will add things like memory policy (what to do on
173 	 * page fault) and maybe some other commands.
174 	 */
175 
176 	mm = get_task_mm(current);
177 	down_read(&mm->mmap_sem);
178 
179 	for (addr = args->va_start, end = args->va_start + size; addr < end;) {
180 		struct vm_area_struct *vma;
181 		unsigned long next;
182 
183 		vma = find_vma_intersection(mm, addr, end);
184 		if (!vma)
185 			break;
186 
187 		next = min(vma->vm_end, end);
188 		/* This is a best effort so we ignore errors */
189 		nouveau_dmem_migrate_vma(cli->drm, vma, addr, next);
190 		addr = next;
191 	}
192 
193 	/*
194 	 * FIXME Return the number of page we have migrated, again we need to
195 	 * update the migrate API to return that information so that we can
196 	 * report it to user space.
197 	 */
198 	args->result = 0;
199 
200 	up_read(&mm->mmap_sem);
201 	mmput(mm);
202 
203 	return 0;
204 }
205 
206 /* Unlink channel instance from SVMM. */
207 void
nouveau_svmm_part(struct nouveau_svmm * svmm,u64 inst)208 nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
209 {
210 	struct nouveau_ivmm *ivmm;
211 	if (svmm) {
212 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
213 		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
214 		if (ivmm) {
215 			list_del(&ivmm->head);
216 			kfree(ivmm);
217 		}
218 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
219 	}
220 }
221 
222 /* Link channel instance to SVMM. */
223 int
nouveau_svmm_join(struct nouveau_svmm * svmm,u64 inst)224 nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
225 {
226 	struct nouveau_ivmm *ivmm;
227 	if (svmm) {
228 		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
229 			return -ENOMEM;
230 		ivmm->svmm = svmm;
231 		ivmm->inst = inst;
232 
233 		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
234 		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
235 		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
236 	}
237 	return 0;
238 }
239 
240 /* Invalidate SVMM address-range on GPU. */
241 static void
nouveau_svmm_invalidate(struct nouveau_svmm * svmm,u64 start,u64 limit)242 nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
243 {
244 	if (limit > start) {
245 		bool super = svmm->vmm->vmm.object.client->super;
246 		svmm->vmm->vmm.object.client->super = true;
247 		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
248 				 &(struct nvif_vmm_pfnclr_v0) {
249 					.addr = start,
250 					.size = limit - start,
251 				 }, sizeof(struct nvif_vmm_pfnclr_v0));
252 		svmm->vmm->vmm.object.client->super = super;
253 	}
254 }
255 
256 static int
nouveau_svmm_invalidate_range_start(struct mmu_notifier * mn,const struct mmu_notifier_range * update)257 nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
258 				    const struct mmu_notifier_range *update)
259 {
260 	struct nouveau_svmm *svmm =
261 		container_of(mn, struct nouveau_svmm, notifier);
262 	unsigned long start = update->start;
263 	unsigned long limit = update->end;
264 
265 	if (!mmu_notifier_range_blockable(update))
266 		return -EAGAIN;
267 
268 	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
269 
270 	mutex_lock(&svmm->mutex);
271 	if (unlikely(!svmm->vmm))
272 		goto out;
273 
274 	if (limit > svmm->unmanaged.start && start < svmm->unmanaged.limit) {
275 		if (start < svmm->unmanaged.start) {
276 			nouveau_svmm_invalidate(svmm, start,
277 						svmm->unmanaged.limit);
278 		}
279 		start = svmm->unmanaged.limit;
280 	}
281 
282 	nouveau_svmm_invalidate(svmm, start, limit);
283 
284 out:
285 	mutex_unlock(&svmm->mutex);
286 	return 0;
287 }
288 
nouveau_svmm_free_notifier(struct mmu_notifier * mn)289 static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
290 {
291 	kfree(container_of(mn, struct nouveau_svmm, notifier));
292 }
293 
294 static const struct mmu_notifier_ops nouveau_mn_ops = {
295 	.invalidate_range_start = nouveau_svmm_invalidate_range_start,
296 	.free_notifier = nouveau_svmm_free_notifier,
297 };
298 
299 void
nouveau_svmm_fini(struct nouveau_svmm ** psvmm)300 nouveau_svmm_fini(struct nouveau_svmm **psvmm)
301 {
302 	struct nouveau_svmm *svmm = *psvmm;
303 	if (svmm) {
304 		mutex_lock(&svmm->mutex);
305 		svmm->vmm = NULL;
306 		mutex_unlock(&svmm->mutex);
307 		mmu_notifier_put(&svmm->notifier);
308 		mutex_destroy(&svmm->mutex);
309 		*psvmm = NULL;
310 	}
311 }
312 
313 int
nouveau_svmm_init(struct drm_device * dev,void * data,struct drm_file * file_priv)314 nouveau_svmm_init(struct drm_device *dev, void *data,
315 		  struct drm_file *file_priv)
316 {
317 	struct nouveau_cli *cli = nouveau_cli(file_priv);
318 	struct nouveau_svmm *svmm;
319 	struct drm_nouveau_svm_init *args = data;
320 	int ret;
321 
322 	/* Allocate tracking for SVM-enabled VMM. */
323 	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
324 		return -ENOMEM;
325 	svmm->vmm = &cli->svm;
326 	svmm->unmanaged.start = args->unmanaged_addr;
327 	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
328 	mutex_init(&svmm->mutex);
329 
330 	/* Check that SVM isn't already enabled for the client. */
331 	mutex_lock(&cli->mutex);
332 	if (cli->svm.cli) {
333 		ret = -EBUSY;
334 		goto out_free;
335 	}
336 
337 	/* Allocate a new GPU VMM that can support SVM (managed by the
338 	 * client, with replayable faults enabled).
339 	 *
340 	 * All future channel/memory allocations will make use of this
341 	 * VMM instead of the standard one.
342 	 */
343 	ret = nvif_vmm_init(&cli->mmu, cli->vmm.vmm.object.oclass, true,
344 			    args->unmanaged_addr, args->unmanaged_size,
345 			    &(struct gp100_vmm_v0) {
346 				.fault_replay = true,
347 			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
348 	if (ret)
349 		goto out_free;
350 
351 	down_write(&current->mm->mmap_sem);
352 	svmm->notifier.ops = &nouveau_mn_ops;
353 	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
354 	if (ret)
355 		goto out_mm_unlock;
356 	/* Note, ownership of svmm transfers to mmu_notifier */
357 
358 	cli->svm.svmm = svmm;
359 	cli->svm.cli = cli;
360 	up_write(&current->mm->mmap_sem);
361 	mutex_unlock(&cli->mutex);
362 	return 0;
363 
364 out_mm_unlock:
365 	up_write(&current->mm->mmap_sem);
366 out_free:
367 	mutex_unlock(&cli->mutex);
368 	mutex_destroy(&svmm->mutex);
369 	kfree(svmm);
370 	return ret;
371 }
372 
373 static const u64
374 nouveau_svm_pfn_flags[HMM_PFN_FLAG_MAX] = {
375 	[HMM_PFN_VALID         ] = NVIF_VMM_PFNMAP_V0_V,
376 	[HMM_PFN_WRITE         ] = NVIF_VMM_PFNMAP_V0_W,
377 	[HMM_PFN_DEVICE_PRIVATE] = NVIF_VMM_PFNMAP_V0_VRAM,
378 };
379 
380 static const u64
381 nouveau_svm_pfn_values[HMM_PFN_VALUE_MAX] = {
382 	[HMM_PFN_ERROR  ] = ~NVIF_VMM_PFNMAP_V0_V,
383 	[HMM_PFN_NONE   ] =  NVIF_VMM_PFNMAP_V0_NONE,
384 	[HMM_PFN_SPECIAL] = ~NVIF_VMM_PFNMAP_V0_V,
385 };
386 
387 /* Issue fault replay for GPU to retry accesses that faulted previously. */
388 static void
nouveau_svm_fault_replay(struct nouveau_svm * svm)389 nouveau_svm_fault_replay(struct nouveau_svm *svm)
390 {
391 	SVM_DBG(svm, "replay");
392 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
393 				 GP100_VMM_VN_FAULT_REPLAY,
394 				 &(struct gp100_vmm_fault_replay_vn) {},
395 				 sizeof(struct gp100_vmm_fault_replay_vn)));
396 }
397 
398 /* Cancel a replayable fault that could not be handled.
399  *
400  * Cancelling the fault will trigger recovery to reset the engine
401  * and kill the offending channel (ie. GPU SIGSEGV).
402  */
403 static void
nouveau_svm_fault_cancel(struct nouveau_svm * svm,u64 inst,u8 hub,u8 gpc,u8 client)404 nouveau_svm_fault_cancel(struct nouveau_svm *svm,
405 			 u64 inst, u8 hub, u8 gpc, u8 client)
406 {
407 	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
408 	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
409 				 GP100_VMM_VN_FAULT_CANCEL,
410 				 &(struct gp100_vmm_fault_cancel_v0) {
411 					.hub = hub,
412 					.gpc = gpc,
413 					.client = client,
414 					.inst = inst,
415 				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
416 }
417 
418 static void
nouveau_svm_fault_cancel_fault(struct nouveau_svm * svm,struct nouveau_svm_fault * fault)419 nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
420 			       struct nouveau_svm_fault *fault)
421 {
422 	nouveau_svm_fault_cancel(svm, fault->inst,
423 				      fault->hub,
424 				      fault->gpc,
425 				      fault->client);
426 }
427 
428 static int
nouveau_svm_fault_cmp(const void * a,const void * b)429 nouveau_svm_fault_cmp(const void *a, const void *b)
430 {
431 	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
432 	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
433 	int ret;
434 	if ((ret = (s64)fa->inst - fb->inst))
435 		return ret;
436 	if ((ret = (s64)fa->addr - fb->addr))
437 		return ret;
438 	/*XXX: atomic? */
439 	return (fa->access == 0 || fa->access == 3) -
440 	       (fb->access == 0 || fb->access == 3);
441 }
442 
443 static void
nouveau_svm_fault_cache(struct nouveau_svm * svm,struct nouveau_svm_fault_buffer * buffer,u32 offset)444 nouveau_svm_fault_cache(struct nouveau_svm *svm,
445 			struct nouveau_svm_fault_buffer *buffer, u32 offset)
446 {
447 	struct nvif_object *memory = &buffer->object;
448 	const u32 instlo = nvif_rd32(memory, offset + 0x00);
449 	const u32 insthi = nvif_rd32(memory, offset + 0x04);
450 	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
451 	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
452 	const u32 timelo = nvif_rd32(memory, offset + 0x10);
453 	const u32 timehi = nvif_rd32(memory, offset + 0x14);
454 	const u32 engine = nvif_rd32(memory, offset + 0x18);
455 	const u32   info = nvif_rd32(memory, offset + 0x1c);
456 	const u64   inst = (u64)insthi << 32 | instlo;
457 	const u8     gpc = (info & 0x1f000000) >> 24;
458 	const u8     hub = (info & 0x00100000) >> 20;
459 	const u8  client = (info & 0x00007f00) >> 8;
460 	struct nouveau_svm_fault *fault;
461 
462 	//XXX: i think we're supposed to spin waiting */
463 	if (WARN_ON(!(info & 0x80000000)))
464 		return;
465 
466 	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
467 
468 	if (!buffer->fault[buffer->fault_nr]) {
469 		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
470 		if (WARN_ON(!fault)) {
471 			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
472 			return;
473 		}
474 		buffer->fault[buffer->fault_nr] = fault;
475 	}
476 
477 	fault = buffer->fault[buffer->fault_nr++];
478 	fault->inst   = inst;
479 	fault->addr   = (u64)addrhi << 32 | addrlo;
480 	fault->time   = (u64)timehi << 32 | timelo;
481 	fault->engine = engine;
482 	fault->gpc    = gpc;
483 	fault->hub    = hub;
484 	fault->access = (info & 0x000f0000) >> 16;
485 	fault->client = client;
486 	fault->fault  = (info & 0x0000001f);
487 
488 	SVM_DBG(svm, "fault %016llx %016llx %02x",
489 		fault->inst, fault->addr, fault->access);
490 }
491 
492 struct svm_notifier {
493 	struct mmu_interval_notifier notifier;
494 	struct nouveau_svmm *svmm;
495 };
496 
nouveau_svm_range_invalidate(struct mmu_interval_notifier * mni,const struct mmu_notifier_range * range,unsigned long cur_seq)497 static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
498 					 const struct mmu_notifier_range *range,
499 					 unsigned long cur_seq)
500 {
501 	struct svm_notifier *sn =
502 		container_of(mni, struct svm_notifier, notifier);
503 
504 	/*
505 	 * serializes the update to mni->invalidate_seq done by caller and
506 	 * prevents invalidation of the PTE from progressing while HW is being
507 	 * programmed. This is very hacky and only works because the normal
508 	 * notifier that does invalidation is always called after the range
509 	 * notifier.
510 	 */
511 	if (mmu_notifier_range_blockable(range))
512 		mutex_lock(&sn->svmm->mutex);
513 	else if (!mutex_trylock(&sn->svmm->mutex))
514 		return false;
515 	mmu_interval_set_seq(mni, cur_seq);
516 	mutex_unlock(&sn->svmm->mutex);
517 	return true;
518 }
519 
520 static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
521 	.invalidate = nouveau_svm_range_invalidate,
522 };
523 
nouveau_range_fault(struct nouveau_svmm * svmm,struct nouveau_drm * drm,void * data,u32 size,u64 * pfns,struct svm_notifier * notifier)524 static int nouveau_range_fault(struct nouveau_svmm *svmm,
525 			       struct nouveau_drm *drm, void *data, u32 size,
526 			       u64 *pfns, struct svm_notifier *notifier)
527 {
528 	unsigned long timeout =
529 		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
530 	/* Have HMM fault pages within the fault window to the GPU. */
531 	struct hmm_range range = {
532 		.notifier = &notifier->notifier,
533 		.start = notifier->notifier.interval_tree.start,
534 		.end = notifier->notifier.interval_tree.last + 1,
535 		.pfns = pfns,
536 		.flags = nouveau_svm_pfn_flags,
537 		.values = nouveau_svm_pfn_values,
538 		.pfn_shift = NVIF_VMM_PFNMAP_V0_ADDR_SHIFT,
539 	};
540 	struct mm_struct *mm = notifier->notifier.mm;
541 	long ret;
542 
543 	while (true) {
544 		if (time_after(jiffies, timeout))
545 			return -EBUSY;
546 
547 		range.notifier_seq = mmu_interval_read_begin(range.notifier);
548 		range.default_flags = 0;
549 		range.pfn_flags_mask = -1UL;
550 		down_read(&mm->mmap_sem);
551 		ret = hmm_range_fault(&range, 0);
552 		up_read(&mm->mmap_sem);
553 		if (ret <= 0) {
554 			if (ret == 0 || ret == -EBUSY)
555 				continue;
556 			return ret;
557 		}
558 
559 		mutex_lock(&svmm->mutex);
560 		if (mmu_interval_read_retry(range.notifier,
561 					    range.notifier_seq)) {
562 			mutex_unlock(&svmm->mutex);
563 			continue;
564 		}
565 		break;
566 	}
567 
568 	nouveau_dmem_convert_pfn(drm, &range);
569 
570 	svmm->vmm->vmm.object.client->super = true;
571 	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, data, size, NULL);
572 	svmm->vmm->vmm.object.client->super = false;
573 	mutex_unlock(&svmm->mutex);
574 
575 	return ret;
576 }
577 
578 static int
nouveau_svm_fault(struct nvif_notify * notify)579 nouveau_svm_fault(struct nvif_notify *notify)
580 {
581 	struct nouveau_svm_fault_buffer *buffer =
582 		container_of(notify, typeof(*buffer), notify);
583 	struct nouveau_svm *svm =
584 		container_of(buffer, typeof(*svm), buffer[buffer->id]);
585 	struct nvif_object *device = &svm->drm->client.device.object;
586 	struct nouveau_svmm *svmm;
587 	struct {
588 		struct {
589 			struct nvif_ioctl_v0 i;
590 			struct nvif_ioctl_mthd_v0 m;
591 			struct nvif_vmm_pfnmap_v0 p;
592 		} i;
593 		u64 phys[16];
594 	} args;
595 	struct vm_area_struct *vma;
596 	u64 inst, start, limit;
597 	int fi, fn, pi, fill;
598 	int replay = 0, ret;
599 
600 	/* Parse available fault buffer entries into a cache, and update
601 	 * the GET pointer so HW can reuse the entries.
602 	 */
603 	SVM_DBG(svm, "fault handler");
604 	if (buffer->get == buffer->put) {
605 		buffer->put = nvif_rd32(device, buffer->putaddr);
606 		buffer->get = nvif_rd32(device, buffer->getaddr);
607 		if (buffer->get == buffer->put)
608 			return NVIF_NOTIFY_KEEP;
609 	}
610 	buffer->fault_nr = 0;
611 
612 	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
613 	while (buffer->get != buffer->put) {
614 		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
615 		if (++buffer->get == buffer->entries)
616 			buffer->get = 0;
617 	}
618 	nvif_wr32(device, buffer->getaddr, buffer->get);
619 	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
620 
621 	/* Sort parsed faults by instance pointer to prevent unnecessary
622 	 * instance to SVMM translations, followed by address and access
623 	 * type to reduce the amount of work when handling the faults.
624 	 */
625 	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
626 	     nouveau_svm_fault_cmp, NULL);
627 
628 	/* Lookup SVMM structure for each unique instance pointer. */
629 	mutex_lock(&svm->mutex);
630 	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
631 		if (!svmm || buffer->fault[fi]->inst != inst) {
632 			struct nouveau_ivmm *ivmm =
633 				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
634 			svmm = ivmm ? ivmm->svmm : NULL;
635 			inst = buffer->fault[fi]->inst;
636 			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
637 		}
638 		buffer->fault[fi]->svmm = svmm;
639 	}
640 	mutex_unlock(&svm->mutex);
641 
642 	/* Process list of faults. */
643 	args.i.i.version = 0;
644 	args.i.i.type = NVIF_IOCTL_V0_MTHD;
645 	args.i.m.version = 0;
646 	args.i.m.method = NVIF_VMM_V0_PFNMAP;
647 	args.i.p.version = 0;
648 
649 	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
650 		struct svm_notifier notifier;
651 		struct mm_struct *mm;
652 
653 		/* Cancel any faults from non-SVM channels. */
654 		if (!(svmm = buffer->fault[fi]->svmm)) {
655 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
656 			continue;
657 		}
658 		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
659 
660 		/* We try and group handling of faults within a small
661 		 * window into a single update.
662 		 */
663 		start = buffer->fault[fi]->addr;
664 		limit = start + (ARRAY_SIZE(args.phys) << PAGE_SHIFT);
665 		if (start < svmm->unmanaged.limit)
666 			limit = min_t(u64, limit, svmm->unmanaged.start);
667 		else
668 		if (limit > svmm->unmanaged.start)
669 			start = max_t(u64, start, svmm->unmanaged.limit);
670 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
671 
672 		mm = svmm->notifier.mm;
673 		if (!mmget_not_zero(mm)) {
674 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
675 			continue;
676 		}
677 
678 		/* Intersect fault window with the CPU VMA, cancelling
679 		 * the fault if the address is invalid.
680 		 */
681 		down_read(&mm->mmap_sem);
682 		vma = find_vma_intersection(mm, start, limit);
683 		if (!vma) {
684 			SVMM_ERR(svmm, "wndw %016llx-%016llx", start, limit);
685 			up_read(&mm->mmap_sem);
686 			mmput(mm);
687 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
688 			continue;
689 		}
690 		start = max_t(u64, start, vma->vm_start);
691 		limit = min_t(u64, limit, vma->vm_end);
692 		up_read(&mm->mmap_sem);
693 		SVMM_DBG(svmm, "wndw %016llx-%016llx", start, limit);
694 
695 		if (buffer->fault[fi]->addr != start) {
696 			SVMM_ERR(svmm, "addr %016llx", buffer->fault[fi]->addr);
697 			mmput(mm);
698 			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
699 			continue;
700 		}
701 
702 		/* Prepare the GPU-side update of all pages within the
703 		 * fault window, determining required pages and access
704 		 * permissions based on pending faults.
705 		 */
706 		args.i.p.page = PAGE_SHIFT;
707 		args.i.p.addr = start;
708 		for (fn = fi, pi = 0;;) {
709 			/* Determine required permissions based on GPU fault
710 			 * access flags.
711 			 *XXX: atomic?
712 			 */
713 			if (buffer->fault[fn]->access != 0 /* READ. */ &&
714 			    buffer->fault[fn]->access != 3 /* PREFETCH. */) {
715 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V |
716 						  NVIF_VMM_PFNMAP_V0_W;
717 			} else {
718 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_V;
719 			}
720 			args.i.p.size = pi << PAGE_SHIFT;
721 
722 			/* It's okay to skip over duplicate addresses from the
723 			 * same SVMM as faults are ordered by access type such
724 			 * that only the first one needs to be handled.
725 			 *
726 			 * ie. WRITE faults appear first, thus any handling of
727 			 * pending READ faults will already be satisfied.
728 			 */
729 			while (++fn < buffer->fault_nr &&
730 			       buffer->fault[fn]->svmm == svmm &&
731 			       buffer->fault[fn    ]->addr ==
732 			       buffer->fault[fn - 1]->addr);
733 
734 			/* If the next fault is outside the window, or all GPU
735 			 * faults have been dealt with, we're done here.
736 			 */
737 			if (fn >= buffer->fault_nr ||
738 			    buffer->fault[fn]->svmm != svmm ||
739 			    buffer->fault[fn]->addr >= limit)
740 				break;
741 
742 			/* Fill in the gap between this fault and the next. */
743 			fill = (buffer->fault[fn    ]->addr -
744 				buffer->fault[fn - 1]->addr) >> PAGE_SHIFT;
745 			while (--fill)
746 				args.phys[pi++] = NVIF_VMM_PFNMAP_V0_NONE;
747 		}
748 
749 		SVMM_DBG(svmm, "wndw %016llx-%016llx covering %d fault(s)",
750 			 args.i.p.addr,
751 			 args.i.p.addr + args.i.p.size, fn - fi);
752 
753 		notifier.svmm = svmm;
754 		ret = mmu_interval_notifier_insert(&notifier.notifier,
755 						   svmm->notifier.mm,
756 						   args.i.p.addr, args.i.p.size,
757 						   &nouveau_svm_mni_ops);
758 		if (!ret) {
759 			ret = nouveau_range_fault(
760 				svmm, svm->drm, &args,
761 				sizeof(args.i) + pi * sizeof(args.phys[0]),
762 				args.phys, &notifier);
763 			mmu_interval_notifier_remove(&notifier.notifier);
764 		}
765 		mmput(mm);
766 
767 		/* Cancel any faults in the window whose pages didn't manage
768 		 * to keep their valid bit, or stay writeable when required.
769 		 *
770 		 * If handling failed completely, cancel all faults.
771 		 */
772 		while (fi < fn) {
773 			struct nouveau_svm_fault *fault = buffer->fault[fi++];
774 			pi = (fault->addr - args.i.p.addr) >> PAGE_SHIFT;
775 			if (ret ||
776 			     !(args.phys[pi] & NVIF_VMM_PFNMAP_V0_V) ||
777 			    (!(args.phys[pi] & NVIF_VMM_PFNMAP_V0_W) &&
778 			     fault->access != 0 && fault->access != 3)) {
779 				nouveau_svm_fault_cancel_fault(svm, fault);
780 				continue;
781 			}
782 			replay++;
783 		}
784 	}
785 
786 	/* Issue fault replay to the GPU. */
787 	if (replay)
788 		nouveau_svm_fault_replay(svm);
789 	return NVIF_NOTIFY_KEEP;
790 }
791 
792 static void
nouveau_svm_fault_buffer_fini(struct nouveau_svm * svm,int id)793 nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
794 {
795 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
796 	nvif_notify_put(&buffer->notify);
797 }
798 
799 static int
nouveau_svm_fault_buffer_init(struct nouveau_svm * svm,int id)800 nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
801 {
802 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
803 	struct nvif_object *device = &svm->drm->client.device.object;
804 	buffer->get = nvif_rd32(device, buffer->getaddr);
805 	buffer->put = nvif_rd32(device, buffer->putaddr);
806 	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
807 	return nvif_notify_get(&buffer->notify);
808 }
809 
810 static void
nouveau_svm_fault_buffer_dtor(struct nouveau_svm * svm,int id)811 nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
812 {
813 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
814 	int i;
815 
816 	if (buffer->fault) {
817 		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
818 			kfree(buffer->fault[i]);
819 		kvfree(buffer->fault);
820 	}
821 
822 	nouveau_svm_fault_buffer_fini(svm, id);
823 
824 	nvif_notify_fini(&buffer->notify);
825 	nvif_object_fini(&buffer->object);
826 }
827 
828 static int
nouveau_svm_fault_buffer_ctor(struct nouveau_svm * svm,s32 oclass,int id)829 nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
830 {
831 	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
832 	struct nouveau_drm *drm = svm->drm;
833 	struct nvif_object *device = &drm->client.device.object;
834 	struct nvif_clb069_v0 args = {};
835 	int ret;
836 
837 	buffer->id = id;
838 
839 	ret = nvif_object_init(device, 0, oclass, &args, sizeof(args),
840 			       &buffer->object);
841 	if (ret < 0) {
842 		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
843 		return ret;
844 	}
845 
846 	nvif_object_map(&buffer->object, NULL, 0);
847 	buffer->entries = args.entries;
848 	buffer->getaddr = args.get;
849 	buffer->putaddr = args.put;
850 
851 	ret = nvif_notify_init(&buffer->object, nouveau_svm_fault, true,
852 			       NVB069_V0_NTFY_FAULT, NULL, 0, 0,
853 			       &buffer->notify);
854 	if (ret)
855 		return ret;
856 
857 	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
858 	if (!buffer->fault)
859 		return -ENOMEM;
860 
861 	return nouveau_svm_fault_buffer_init(svm, id);
862 }
863 
864 void
nouveau_svm_resume(struct nouveau_drm * drm)865 nouveau_svm_resume(struct nouveau_drm *drm)
866 {
867 	struct nouveau_svm *svm = drm->svm;
868 	if (svm)
869 		nouveau_svm_fault_buffer_init(svm, 0);
870 }
871 
872 void
nouveau_svm_suspend(struct nouveau_drm * drm)873 nouveau_svm_suspend(struct nouveau_drm *drm)
874 {
875 	struct nouveau_svm *svm = drm->svm;
876 	if (svm)
877 		nouveau_svm_fault_buffer_fini(svm, 0);
878 }
879 
880 void
nouveau_svm_fini(struct nouveau_drm * drm)881 nouveau_svm_fini(struct nouveau_drm *drm)
882 {
883 	struct nouveau_svm *svm = drm->svm;
884 	if (svm) {
885 		nouveau_svm_fault_buffer_dtor(svm, 0);
886 		kfree(drm->svm);
887 		drm->svm = NULL;
888 	}
889 }
890 
891 void
nouveau_svm_init(struct nouveau_drm * drm)892 nouveau_svm_init(struct nouveau_drm *drm)
893 {
894 	static const struct nvif_mclass buffers[] = {
895 		{   VOLTA_FAULT_BUFFER_A, 0 },
896 		{ MAXWELL_FAULT_BUFFER_A, 0 },
897 		{}
898 	};
899 	struct nouveau_svm *svm;
900 	int ret;
901 
902 	/* Disable on Volta and newer until channel recovery is fixed,
903 	 * otherwise clients will have a trivial way to trash the GPU
904 	 * for everyone.
905 	 */
906 	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
907 		return;
908 
909 	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
910 		return;
911 
912 	drm->svm->drm = drm;
913 	mutex_init(&drm->svm->mutex);
914 	INIT_LIST_HEAD(&drm->svm->inst);
915 
916 	ret = nvif_mclass(&drm->client.device.object, buffers);
917 	if (ret < 0) {
918 		SVM_DBG(svm, "No supported fault buffer class");
919 		nouveau_svm_fini(drm);
920 		return;
921 	}
922 
923 	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
924 	if (ret) {
925 		nouveau_svm_fini(drm);
926 		return;
927 	}
928 
929 	SVM_DBG(svm, "Initialised");
930 }
931