Linux Audio

Check our new training course

Yocto / OpenEmbedded training

Mar 24-27, 2025, special US time zones
Register
Loading...
Note: File does not exist in v4.6.
   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#include <linux/rmap.h>
  39
  40struct nouveau_svm {
  41	struct nouveau_drm *drm;
  42	struct mutex mutex;
  43	struct list_head inst;
  44
  45	struct nouveau_svm_fault_buffer {
  46		int id;
  47		struct nvif_object object;
  48		u32 entries;
  49		u32 getaddr;
  50		u32 putaddr;
  51		u32 get;
  52		u32 put;
  53		struct nvif_notify notify;
  54
  55		struct nouveau_svm_fault {
  56			u64 inst;
  57			u64 addr;
  58			u64 time;
  59			u32 engine;
  60			u8  gpc;
  61			u8  hub;
  62			u8  access;
  63			u8  client;
  64			u8  fault;
  65			struct nouveau_svmm *svmm;
  66		} **fault;
  67		int fault_nr;
  68	} buffer[1];
  69};
  70
  71#define FAULT_ACCESS_READ 0
  72#define FAULT_ACCESS_WRITE 1
  73#define FAULT_ACCESS_ATOMIC 2
  74#define FAULT_ACCESS_PREFETCH 3
  75
  76#define SVM_DBG(s,f,a...) NV_DEBUG((s)->drm, "svm: "f"\n", ##a)
  77#define SVM_ERR(s,f,a...) NV_WARN((s)->drm, "svm: "f"\n", ##a)
  78
  79struct nouveau_pfnmap_args {
  80	struct nvif_ioctl_v0 i;
  81	struct nvif_ioctl_mthd_v0 m;
  82	struct nvif_vmm_pfnmap_v0 p;
  83};
  84
  85struct nouveau_ivmm {
  86	struct nouveau_svmm *svmm;
  87	u64 inst;
  88	struct list_head head;
  89};
  90
  91static struct nouveau_ivmm *
  92nouveau_ivmm_find(struct nouveau_svm *svm, u64 inst)
  93{
  94	struct nouveau_ivmm *ivmm;
  95	list_for_each_entry(ivmm, &svm->inst, head) {
  96		if (ivmm->inst == inst)
  97			return ivmm;
  98	}
  99	return NULL;
 100}
 101
 102#define SVMM_DBG(s,f,a...)                                                     \
 103	NV_DEBUG((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
 104#define SVMM_ERR(s,f,a...)                                                     \
 105	NV_WARN((s)->vmm->cli->drm, "svm-%p: "f"\n", (s), ##a)
 106
 107int
 108nouveau_svmm_bind(struct drm_device *dev, void *data,
 109		  struct drm_file *file_priv)
 110{
 111	struct nouveau_cli *cli = nouveau_cli(file_priv);
 112	struct drm_nouveau_svm_bind *args = data;
 113	unsigned target, cmd, priority;
 114	unsigned long addr, end;
 115	struct mm_struct *mm;
 116
 117	args->va_start &= PAGE_MASK;
 118	args->va_end = ALIGN(args->va_end, PAGE_SIZE);
 119
 120	/* Sanity check arguments */
 121	if (args->reserved0 || args->reserved1)
 122		return -EINVAL;
 123	if (args->header & (~NOUVEAU_SVM_BIND_VALID_MASK))
 124		return -EINVAL;
 125	if (args->va_start >= args->va_end)
 126		return -EINVAL;
 127
 128	cmd = args->header >> NOUVEAU_SVM_BIND_COMMAND_SHIFT;
 129	cmd &= NOUVEAU_SVM_BIND_COMMAND_MASK;
 130	switch (cmd) {
 131	case NOUVEAU_SVM_BIND_COMMAND__MIGRATE:
 132		break;
 133	default:
 134		return -EINVAL;
 135	}
 136
 137	priority = args->header >> NOUVEAU_SVM_BIND_PRIORITY_SHIFT;
 138	priority &= NOUVEAU_SVM_BIND_PRIORITY_MASK;
 139
 140	/* FIXME support CPU target ie all target value < GPU_VRAM */
 141	target = args->header >> NOUVEAU_SVM_BIND_TARGET_SHIFT;
 142	target &= NOUVEAU_SVM_BIND_TARGET_MASK;
 143	switch (target) {
 144	case NOUVEAU_SVM_BIND_TARGET__GPU_VRAM:
 145		break;
 146	default:
 147		return -EINVAL;
 148	}
 149
 150	/*
 151	 * FIXME: For now refuse non 0 stride, we need to change the migrate
 152	 * kernel function to handle stride to avoid to create a mess within
 153	 * each device driver.
 154	 */
 155	if (args->stride)
 156		return -EINVAL;
 157
 158	/*
 159	 * Ok we are ask to do something sane, for now we only support migrate
 160	 * commands but we will add things like memory policy (what to do on
 161	 * page fault) and maybe some other commands.
 162	 */
 163
 164	mm = get_task_mm(current);
 165	mmap_read_lock(mm);
 166
 167	if (!cli->svm.svmm) {
 168		mmap_read_unlock(mm);
 169		return -EINVAL;
 170	}
 171
 172	for (addr = args->va_start, end = args->va_end; addr < end;) {
 173		struct vm_area_struct *vma;
 174		unsigned long next;
 175
 176		vma = find_vma_intersection(mm, addr, end);
 177		if (!vma)
 178			break;
 179
 180		addr = max(addr, vma->vm_start);
 181		next = min(vma->vm_end, end);
 182		/* This is a best effort so we ignore errors */
 183		nouveau_dmem_migrate_vma(cli->drm, cli->svm.svmm, vma, addr,
 184					 next);
 185		addr = next;
 186	}
 187
 188	/*
 189	 * FIXME Return the number of page we have migrated, again we need to
 190	 * update the migrate API to return that information so that we can
 191	 * report it to user space.
 192	 */
 193	args->result = 0;
 194
 195	mmap_read_unlock(mm);
 196	mmput(mm);
 197
 198	return 0;
 199}
 200
 201/* Unlink channel instance from SVMM. */
 202void
 203nouveau_svmm_part(struct nouveau_svmm *svmm, u64 inst)
 204{
 205	struct nouveau_ivmm *ivmm;
 206	if (svmm) {
 207		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
 208		ivmm = nouveau_ivmm_find(svmm->vmm->cli->drm->svm, inst);
 209		if (ivmm) {
 210			list_del(&ivmm->head);
 211			kfree(ivmm);
 212		}
 213		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
 214	}
 215}
 216
 217/* Link channel instance to SVMM. */
 218int
 219nouveau_svmm_join(struct nouveau_svmm *svmm, u64 inst)
 220{
 221	struct nouveau_ivmm *ivmm;
 222	if (svmm) {
 223		if (!(ivmm = kmalloc(sizeof(*ivmm), GFP_KERNEL)))
 224			return -ENOMEM;
 225		ivmm->svmm = svmm;
 226		ivmm->inst = inst;
 227
 228		mutex_lock(&svmm->vmm->cli->drm->svm->mutex);
 229		list_add(&ivmm->head, &svmm->vmm->cli->drm->svm->inst);
 230		mutex_unlock(&svmm->vmm->cli->drm->svm->mutex);
 231	}
 232	return 0;
 233}
 234
 235/* Invalidate SVMM address-range on GPU. */
 236void
 237nouveau_svmm_invalidate(struct nouveau_svmm *svmm, u64 start, u64 limit)
 238{
 239	if (limit > start) {
 240		nvif_object_mthd(&svmm->vmm->vmm.object, NVIF_VMM_V0_PFNCLR,
 241				 &(struct nvif_vmm_pfnclr_v0) {
 242					.addr = start,
 243					.size = limit - start,
 244				 }, sizeof(struct nvif_vmm_pfnclr_v0));
 245	}
 246}
 247
 248static int
 249nouveau_svmm_invalidate_range_start(struct mmu_notifier *mn,
 250				    const struct mmu_notifier_range *update)
 251{
 252	struct nouveau_svmm *svmm =
 253		container_of(mn, struct nouveau_svmm, notifier);
 254	unsigned long start = update->start;
 255	unsigned long limit = update->end;
 256
 257	if (!mmu_notifier_range_blockable(update))
 258		return -EAGAIN;
 259
 260	SVMM_DBG(svmm, "invalidate %016lx-%016lx", start, limit);
 261
 262	mutex_lock(&svmm->mutex);
 263	if (unlikely(!svmm->vmm))
 264		goto out;
 265
 266	/*
 267	 * Ignore invalidation callbacks for device private pages since
 268	 * the invalidation is handled as part of the migration process.
 269	 */
 270	if (update->event == MMU_NOTIFY_MIGRATE &&
 271	    update->owner == svmm->vmm->cli->drm->dev)
 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
 284out:
 285	mutex_unlock(&svmm->mutex);
 286	return 0;
 287}
 288
 289static void nouveau_svmm_free_notifier(struct mmu_notifier *mn)
 290{
 291	kfree(container_of(mn, struct nouveau_svmm, notifier));
 292}
 293
 294static 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
 299void
 300nouveau_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		*psvmm = NULL;
 309	}
 310}
 311
 312int
 313nouveau_svmm_init(struct drm_device *dev, void *data,
 314		  struct drm_file *file_priv)
 315{
 316	struct nouveau_cli *cli = nouveau_cli(file_priv);
 317	struct nouveau_svmm *svmm;
 318	struct drm_nouveau_svm_init *args = data;
 319	int ret;
 320
 321	/* We need to fail if svm is disabled */
 322	if (!cli->drm->svm)
 323		return -ENOSYS;
 324
 325	/* Allocate tracking for SVM-enabled VMM. */
 326	if (!(svmm = kzalloc(sizeof(*svmm), GFP_KERNEL)))
 327		return -ENOMEM;
 328	svmm->vmm = &cli->svm;
 329	svmm->unmanaged.start = args->unmanaged_addr;
 330	svmm->unmanaged.limit = args->unmanaged_addr + args->unmanaged_size;
 331	mutex_init(&svmm->mutex);
 332
 333	/* Check that SVM isn't already enabled for the client. */
 334	mutex_lock(&cli->mutex);
 335	if (cli->svm.cli) {
 336		ret = -EBUSY;
 337		goto out_free;
 338	}
 339
 340	/* Allocate a new GPU VMM that can support SVM (managed by the
 341	 * client, with replayable faults enabled).
 342	 *
 343	 * All future channel/memory allocations will make use of this
 344	 * VMM instead of the standard one.
 345	 */
 346	ret = nvif_vmm_ctor(&cli->mmu, "svmVmm",
 347			    cli->vmm.vmm.object.oclass, true,
 348			    args->unmanaged_addr, args->unmanaged_size,
 349			    &(struct gp100_vmm_v0) {
 350				.fault_replay = true,
 351			    }, sizeof(struct gp100_vmm_v0), &cli->svm.vmm);
 352	if (ret)
 353		goto out_free;
 354
 355	mmap_write_lock(current->mm);
 356	svmm->notifier.ops = &nouveau_mn_ops;
 357	ret = __mmu_notifier_register(&svmm->notifier, current->mm);
 358	if (ret)
 359		goto out_mm_unlock;
 360	/* Note, ownership of svmm transfers to mmu_notifier */
 361
 362	cli->svm.svmm = svmm;
 363	cli->svm.cli = cli;
 364	mmap_write_unlock(current->mm);
 365	mutex_unlock(&cli->mutex);
 366	return 0;
 367
 368out_mm_unlock:
 369	mmap_write_unlock(current->mm);
 370out_free:
 371	mutex_unlock(&cli->mutex);
 372	kfree(svmm);
 373	return ret;
 374}
 375
 376/* Issue fault replay for GPU to retry accesses that faulted previously. */
 377static void
 378nouveau_svm_fault_replay(struct nouveau_svm *svm)
 379{
 380	SVM_DBG(svm, "replay");
 381	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
 382				 GP100_VMM_VN_FAULT_REPLAY,
 383				 &(struct gp100_vmm_fault_replay_vn) {},
 384				 sizeof(struct gp100_vmm_fault_replay_vn)));
 385}
 386
 387/* Cancel a replayable fault that could not be handled.
 388 *
 389 * Cancelling the fault will trigger recovery to reset the engine
 390 * and kill the offending channel (ie. GPU SIGSEGV).
 391 */
 392static void
 393nouveau_svm_fault_cancel(struct nouveau_svm *svm,
 394			 u64 inst, u8 hub, u8 gpc, u8 client)
 395{
 396	SVM_DBG(svm, "cancel %016llx %d %02x %02x", inst, hub, gpc, client);
 397	WARN_ON(nvif_object_mthd(&svm->drm->client.vmm.vmm.object,
 398				 GP100_VMM_VN_FAULT_CANCEL,
 399				 &(struct gp100_vmm_fault_cancel_v0) {
 400					.hub = hub,
 401					.gpc = gpc,
 402					.client = client,
 403					.inst = inst,
 404				 }, sizeof(struct gp100_vmm_fault_cancel_v0)));
 405}
 406
 407static void
 408nouveau_svm_fault_cancel_fault(struct nouveau_svm *svm,
 409			       struct nouveau_svm_fault *fault)
 410{
 411	nouveau_svm_fault_cancel(svm, fault->inst,
 412				      fault->hub,
 413				      fault->gpc,
 414				      fault->client);
 415}
 416
 417static int
 418nouveau_svm_fault_priority(u8 fault)
 419{
 420	switch (fault) {
 421	case FAULT_ACCESS_PREFETCH:
 422		return 0;
 423	case FAULT_ACCESS_READ:
 424		return 1;
 425	case FAULT_ACCESS_WRITE:
 426		return 2;
 427	case FAULT_ACCESS_ATOMIC:
 428		return 3;
 429	default:
 430		WARN_ON_ONCE(1);
 431		return -1;
 432	}
 433}
 434
 435static int
 436nouveau_svm_fault_cmp(const void *a, const void *b)
 437{
 438	const struct nouveau_svm_fault *fa = *(struct nouveau_svm_fault **)a;
 439	const struct nouveau_svm_fault *fb = *(struct nouveau_svm_fault **)b;
 440	int ret;
 441	if ((ret = (s64)fa->inst - fb->inst))
 442		return ret;
 443	if ((ret = (s64)fa->addr - fb->addr))
 444		return ret;
 445	return nouveau_svm_fault_priority(fa->access) -
 446		nouveau_svm_fault_priority(fb->access);
 447}
 448
 449static void
 450nouveau_svm_fault_cache(struct nouveau_svm *svm,
 451			struct nouveau_svm_fault_buffer *buffer, u32 offset)
 452{
 453	struct nvif_object *memory = &buffer->object;
 454	const u32 instlo = nvif_rd32(memory, offset + 0x00);
 455	const u32 insthi = nvif_rd32(memory, offset + 0x04);
 456	const u32 addrlo = nvif_rd32(memory, offset + 0x08);
 457	const u32 addrhi = nvif_rd32(memory, offset + 0x0c);
 458	const u32 timelo = nvif_rd32(memory, offset + 0x10);
 459	const u32 timehi = nvif_rd32(memory, offset + 0x14);
 460	const u32 engine = nvif_rd32(memory, offset + 0x18);
 461	const u32   info = nvif_rd32(memory, offset + 0x1c);
 462	const u64   inst = (u64)insthi << 32 | instlo;
 463	const u8     gpc = (info & 0x1f000000) >> 24;
 464	const u8     hub = (info & 0x00100000) >> 20;
 465	const u8  client = (info & 0x00007f00) >> 8;
 466	struct nouveau_svm_fault *fault;
 467
 468	//XXX: i think we're supposed to spin waiting */
 469	if (WARN_ON(!(info & 0x80000000)))
 470		return;
 471
 472	nvif_mask(memory, offset + 0x1c, 0x80000000, 0x00000000);
 473
 474	if (!buffer->fault[buffer->fault_nr]) {
 475		fault = kmalloc(sizeof(*fault), GFP_KERNEL);
 476		if (WARN_ON(!fault)) {
 477			nouveau_svm_fault_cancel(svm, inst, hub, gpc, client);
 478			return;
 479		}
 480		buffer->fault[buffer->fault_nr] = fault;
 481	}
 482
 483	fault = buffer->fault[buffer->fault_nr++];
 484	fault->inst   = inst;
 485	fault->addr   = (u64)addrhi << 32 | addrlo;
 486	fault->time   = (u64)timehi << 32 | timelo;
 487	fault->engine = engine;
 488	fault->gpc    = gpc;
 489	fault->hub    = hub;
 490	fault->access = (info & 0x000f0000) >> 16;
 491	fault->client = client;
 492	fault->fault  = (info & 0x0000001f);
 493
 494	SVM_DBG(svm, "fault %016llx %016llx %02x",
 495		fault->inst, fault->addr, fault->access);
 496}
 497
 498struct svm_notifier {
 499	struct mmu_interval_notifier notifier;
 500	struct nouveau_svmm *svmm;
 501};
 502
 503static bool nouveau_svm_range_invalidate(struct mmu_interval_notifier *mni,
 504					 const struct mmu_notifier_range *range,
 505					 unsigned long cur_seq)
 506{
 507	struct svm_notifier *sn =
 508		container_of(mni, struct svm_notifier, notifier);
 509
 510	if (range->event == MMU_NOTIFY_EXCLUSIVE &&
 511	    range->owner == sn->svmm->vmm->cli->drm->dev)
 512		return true;
 513
 514	/*
 515	 * serializes the update to mni->invalidate_seq done by caller and
 516	 * prevents invalidation of the PTE from progressing while HW is being
 517	 * programmed. This is very hacky and only works because the normal
 518	 * notifier that does invalidation is always called after the range
 519	 * notifier.
 520	 */
 521	if (mmu_notifier_range_blockable(range))
 522		mutex_lock(&sn->svmm->mutex);
 523	else if (!mutex_trylock(&sn->svmm->mutex))
 524		return false;
 525	mmu_interval_set_seq(mni, cur_seq);
 526	mutex_unlock(&sn->svmm->mutex);
 527	return true;
 528}
 529
 530static const struct mmu_interval_notifier_ops nouveau_svm_mni_ops = {
 531	.invalidate = nouveau_svm_range_invalidate,
 532};
 533
 534static void nouveau_hmm_convert_pfn(struct nouveau_drm *drm,
 535				    struct hmm_range *range,
 536				    struct nouveau_pfnmap_args *args)
 537{
 538	struct page *page;
 539
 540	/*
 541	 * The address prepared here is passed through nvif_object_ioctl()
 542	 * to an eventual DMA map in something like gp100_vmm_pgt_pfn()
 543	 *
 544	 * This is all just encoding the internal hmm representation into a
 545	 * different nouveau internal representation.
 546	 */
 547	if (!(range->hmm_pfns[0] & HMM_PFN_VALID)) {
 548		args->p.phys[0] = 0;
 549		return;
 550	}
 551
 552	page = hmm_pfn_to_page(range->hmm_pfns[0]);
 553	/*
 554	 * Only map compound pages to the GPU if the CPU is also mapping the
 555	 * page as a compound page. Otherwise, the PTE protections might not be
 556	 * consistent (e.g., CPU only maps part of a compound page).
 557	 * Note that the underlying page might still be larger than the
 558	 * CPU mapping (e.g., a PUD sized compound page partially mapped with
 559	 * a PMD sized page table entry).
 560	 */
 561	if (hmm_pfn_to_map_order(range->hmm_pfns[0])) {
 562		unsigned long addr = args->p.addr;
 563
 564		args->p.page = hmm_pfn_to_map_order(range->hmm_pfns[0]) +
 565				PAGE_SHIFT;
 566		args->p.size = 1UL << args->p.page;
 567		args->p.addr &= ~(args->p.size - 1);
 568		page -= (addr - args->p.addr) >> PAGE_SHIFT;
 569	}
 570	if (is_device_private_page(page))
 571		args->p.phys[0] = nouveau_dmem_page_addr(page) |
 572				NVIF_VMM_PFNMAP_V0_V |
 573				NVIF_VMM_PFNMAP_V0_VRAM;
 574	else
 575		args->p.phys[0] = page_to_phys(page) |
 576				NVIF_VMM_PFNMAP_V0_V |
 577				NVIF_VMM_PFNMAP_V0_HOST;
 578	if (range->hmm_pfns[0] & HMM_PFN_WRITE)
 579		args->p.phys[0] |= NVIF_VMM_PFNMAP_V0_W;
 580}
 581
 582static int nouveau_atomic_range_fault(struct nouveau_svmm *svmm,
 583			       struct nouveau_drm *drm,
 584			       struct nouveau_pfnmap_args *args, u32 size,
 585			       struct svm_notifier *notifier)
 586{
 587	unsigned long timeout =
 588		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 589	struct mm_struct *mm = svmm->notifier.mm;
 590	struct page *page;
 591	unsigned long start = args->p.addr;
 592	unsigned long notifier_seq;
 593	int ret = 0;
 594
 595	ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
 596					args->p.addr, args->p.size,
 597					&nouveau_svm_mni_ops);
 598	if (ret)
 599		return ret;
 600
 601	while (true) {
 602		if (time_after(jiffies, timeout)) {
 603			ret = -EBUSY;
 604			goto out;
 605		}
 606
 607		notifier_seq = mmu_interval_read_begin(&notifier->notifier);
 608		mmap_read_lock(mm);
 609		ret = make_device_exclusive_range(mm, start, start + PAGE_SIZE,
 610					    &page, drm->dev);
 611		mmap_read_unlock(mm);
 612		if (ret <= 0 || !page) {
 613			ret = -EINVAL;
 614			goto out;
 615		}
 616
 617		mutex_lock(&svmm->mutex);
 618		if (!mmu_interval_read_retry(&notifier->notifier,
 619					     notifier_seq))
 620			break;
 621		mutex_unlock(&svmm->mutex);
 622	}
 623
 624	/* Map the page on the GPU. */
 625	args->p.page = 12;
 626	args->p.size = PAGE_SIZE;
 627	args->p.addr = start;
 628	args->p.phys[0] = page_to_phys(page) |
 629		NVIF_VMM_PFNMAP_V0_V |
 630		NVIF_VMM_PFNMAP_V0_W |
 631		NVIF_VMM_PFNMAP_V0_A |
 632		NVIF_VMM_PFNMAP_V0_HOST;
 633
 634	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
 635	mutex_unlock(&svmm->mutex);
 636
 637	unlock_page(page);
 638	put_page(page);
 639
 640out:
 641	mmu_interval_notifier_remove(&notifier->notifier);
 642	return ret;
 643}
 644
 645static int nouveau_range_fault(struct nouveau_svmm *svmm,
 646			       struct nouveau_drm *drm,
 647			       struct nouveau_pfnmap_args *args, u32 size,
 648			       unsigned long hmm_flags,
 649			       struct svm_notifier *notifier)
 650{
 651	unsigned long timeout =
 652		jiffies + msecs_to_jiffies(HMM_RANGE_DEFAULT_TIMEOUT);
 653	/* Have HMM fault pages within the fault window to the GPU. */
 654	unsigned long hmm_pfns[1];
 655	struct hmm_range range = {
 656		.notifier = &notifier->notifier,
 657		.default_flags = hmm_flags,
 658		.hmm_pfns = hmm_pfns,
 659		.dev_private_owner = drm->dev,
 660	};
 661	struct mm_struct *mm = svmm->notifier.mm;
 662	int ret;
 663
 664	ret = mmu_interval_notifier_insert(&notifier->notifier, mm,
 665					args->p.addr, args->p.size,
 666					&nouveau_svm_mni_ops);
 667	if (ret)
 668		return ret;
 669
 670	range.start = notifier->notifier.interval_tree.start;
 671	range.end = notifier->notifier.interval_tree.last + 1;
 672
 673	while (true) {
 674		if (time_after(jiffies, timeout)) {
 675			ret = -EBUSY;
 676			goto out;
 677		}
 678
 679		range.notifier_seq = mmu_interval_read_begin(range.notifier);
 680		mmap_read_lock(mm);
 681		ret = hmm_range_fault(&range);
 682		mmap_read_unlock(mm);
 683		if (ret) {
 684			if (ret == -EBUSY)
 685				continue;
 686			goto out;
 687		}
 688
 689		mutex_lock(&svmm->mutex);
 690		if (mmu_interval_read_retry(range.notifier,
 691					    range.notifier_seq)) {
 692			mutex_unlock(&svmm->mutex);
 693			continue;
 694		}
 695		break;
 696	}
 697
 698	nouveau_hmm_convert_pfn(drm, &range, args);
 699
 700	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, size, NULL);
 701	mutex_unlock(&svmm->mutex);
 702
 703out:
 704	mmu_interval_notifier_remove(&notifier->notifier);
 705
 706	return ret;
 707}
 708
 709static int
 710nouveau_svm_fault(struct nvif_notify *notify)
 711{
 712	struct nouveau_svm_fault_buffer *buffer =
 713		container_of(notify, typeof(*buffer), notify);
 714	struct nouveau_svm *svm =
 715		container_of(buffer, typeof(*svm), buffer[buffer->id]);
 716	struct nvif_object *device = &svm->drm->client.device.object;
 717	struct nouveau_svmm *svmm;
 718	struct {
 719		struct nouveau_pfnmap_args i;
 720		u64 phys[1];
 721	} args;
 722	unsigned long hmm_flags;
 723	u64 inst, start, limit;
 724	int fi, fn;
 725	int replay = 0, atomic = 0, ret;
 726
 727	/* Parse available fault buffer entries into a cache, and update
 728	 * the GET pointer so HW can reuse the entries.
 729	 */
 730	SVM_DBG(svm, "fault handler");
 731	if (buffer->get == buffer->put) {
 732		buffer->put = nvif_rd32(device, buffer->putaddr);
 733		buffer->get = nvif_rd32(device, buffer->getaddr);
 734		if (buffer->get == buffer->put)
 735			return NVIF_NOTIFY_KEEP;
 736	}
 737	buffer->fault_nr = 0;
 738
 739	SVM_DBG(svm, "get %08x put %08x", buffer->get, buffer->put);
 740	while (buffer->get != buffer->put) {
 741		nouveau_svm_fault_cache(svm, buffer, buffer->get * 0x20);
 742		if (++buffer->get == buffer->entries)
 743			buffer->get = 0;
 744	}
 745	nvif_wr32(device, buffer->getaddr, buffer->get);
 746	SVM_DBG(svm, "%d fault(s) pending", buffer->fault_nr);
 747
 748	/* Sort parsed faults by instance pointer to prevent unnecessary
 749	 * instance to SVMM translations, followed by address and access
 750	 * type to reduce the amount of work when handling the faults.
 751	 */
 752	sort(buffer->fault, buffer->fault_nr, sizeof(*buffer->fault),
 753	     nouveau_svm_fault_cmp, NULL);
 754
 755	/* Lookup SVMM structure for each unique instance pointer. */
 756	mutex_lock(&svm->mutex);
 757	for (fi = 0, svmm = NULL; fi < buffer->fault_nr; fi++) {
 758		if (!svmm || buffer->fault[fi]->inst != inst) {
 759			struct nouveau_ivmm *ivmm =
 760				nouveau_ivmm_find(svm, buffer->fault[fi]->inst);
 761			svmm = ivmm ? ivmm->svmm : NULL;
 762			inst = buffer->fault[fi]->inst;
 763			SVM_DBG(svm, "inst %016llx -> svm-%p", inst, svmm);
 764		}
 765		buffer->fault[fi]->svmm = svmm;
 766	}
 767	mutex_unlock(&svm->mutex);
 768
 769	/* Process list of faults. */
 770	args.i.i.version = 0;
 771	args.i.i.type = NVIF_IOCTL_V0_MTHD;
 772	args.i.m.version = 0;
 773	args.i.m.method = NVIF_VMM_V0_PFNMAP;
 774	args.i.p.version = 0;
 775
 776	for (fi = 0; fn = fi + 1, fi < buffer->fault_nr; fi = fn) {
 777		struct svm_notifier notifier;
 778		struct mm_struct *mm;
 779
 780		/* Cancel any faults from non-SVM channels. */
 781		if (!(svmm = buffer->fault[fi]->svmm)) {
 782			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
 783			continue;
 784		}
 785		SVMM_DBG(svmm, "addr %016llx", buffer->fault[fi]->addr);
 786
 787		/* We try and group handling of faults within a small
 788		 * window into a single update.
 789		 */
 790		start = buffer->fault[fi]->addr;
 791		limit = start + PAGE_SIZE;
 792		if (start < svmm->unmanaged.limit)
 793			limit = min_t(u64, limit, svmm->unmanaged.start);
 794
 795		/*
 796		 * Prepare the GPU-side update of all pages within the
 797		 * fault window, determining required pages and access
 798		 * permissions based on pending faults.
 799		 */
 800		args.i.p.addr = start;
 801		args.i.p.page = PAGE_SHIFT;
 802		args.i.p.size = PAGE_SIZE;
 803		/*
 804		 * Determine required permissions based on GPU fault
 805		 * access flags.
 806		 */
 807		switch (buffer->fault[fi]->access) {
 808		case 0: /* READ. */
 809			hmm_flags = HMM_PFN_REQ_FAULT;
 810			break;
 811		case 2: /* ATOMIC. */
 812			atomic = true;
 813			break;
 814		case 3: /* PREFETCH. */
 815			hmm_flags = 0;
 816			break;
 817		default:
 818			hmm_flags = HMM_PFN_REQ_FAULT | HMM_PFN_REQ_WRITE;
 819			break;
 820		}
 821
 822		mm = svmm->notifier.mm;
 823		if (!mmget_not_zero(mm)) {
 824			nouveau_svm_fault_cancel_fault(svm, buffer->fault[fi]);
 825			continue;
 826		}
 827
 828		notifier.svmm = svmm;
 829		if (atomic)
 830			ret = nouveau_atomic_range_fault(svmm, svm->drm,
 831							 &args.i, sizeof(args),
 832							 &notifier);
 833		else
 834			ret = nouveau_range_fault(svmm, svm->drm, &args.i,
 835						  sizeof(args), hmm_flags,
 836						  &notifier);
 837		mmput(mm);
 838
 839		limit = args.i.p.addr + args.i.p.size;
 840		for (fn = fi; ++fn < buffer->fault_nr; ) {
 841			/* It's okay to skip over duplicate addresses from the
 842			 * same SVMM as faults are ordered by access type such
 843			 * that only the first one needs to be handled.
 844			 *
 845			 * ie. WRITE faults appear first, thus any handling of
 846			 * pending READ faults will already be satisfied.
 847			 * But if a large page is mapped, make sure subsequent
 848			 * fault addresses have sufficient access permission.
 849			 */
 850			if (buffer->fault[fn]->svmm != svmm ||
 851			    buffer->fault[fn]->addr >= limit ||
 852			    (buffer->fault[fi]->access == FAULT_ACCESS_READ &&
 853			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_V)) ||
 854			    (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
 855			     buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
 856			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_W)) ||
 857			    (buffer->fault[fi]->access != FAULT_ACCESS_READ &&
 858			     buffer->fault[fi]->access != FAULT_ACCESS_WRITE &&
 859			     buffer->fault[fi]->access != FAULT_ACCESS_PREFETCH &&
 860			     !(args.phys[0] & NVIF_VMM_PFNMAP_V0_A)))
 861				break;
 862		}
 863
 864		/* If handling failed completely, cancel all faults. */
 865		if (ret) {
 866			while (fi < fn) {
 867				struct nouveau_svm_fault *fault =
 868					buffer->fault[fi++];
 869
 870				nouveau_svm_fault_cancel_fault(svm, fault);
 871			}
 872		} else
 873			replay++;
 874	}
 875
 876	/* Issue fault replay to the GPU. */
 877	if (replay)
 878		nouveau_svm_fault_replay(svm);
 879	return NVIF_NOTIFY_KEEP;
 880}
 881
 882static struct nouveau_pfnmap_args *
 883nouveau_pfns_to_args(void *pfns)
 884{
 885	return container_of(pfns, struct nouveau_pfnmap_args, p.phys);
 886}
 887
 888u64 *
 889nouveau_pfns_alloc(unsigned long npages)
 890{
 891	struct nouveau_pfnmap_args *args;
 892
 893	args = kzalloc(struct_size(args, p.phys, npages), GFP_KERNEL);
 894	if (!args)
 895		return NULL;
 896
 897	args->i.type = NVIF_IOCTL_V0_MTHD;
 898	args->m.method = NVIF_VMM_V0_PFNMAP;
 899	args->p.page = PAGE_SHIFT;
 900
 901	return args->p.phys;
 902}
 903
 904void
 905nouveau_pfns_free(u64 *pfns)
 906{
 907	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
 908
 909	kfree(args);
 910}
 911
 912void
 913nouveau_pfns_map(struct nouveau_svmm *svmm, struct mm_struct *mm,
 914		 unsigned long addr, u64 *pfns, unsigned long npages)
 915{
 916	struct nouveau_pfnmap_args *args = nouveau_pfns_to_args(pfns);
 917	int ret;
 918
 919	args->p.addr = addr;
 920	args->p.size = npages << PAGE_SHIFT;
 921
 922	mutex_lock(&svmm->mutex);
 923
 924	ret = nvif_object_ioctl(&svmm->vmm->vmm.object, args, sizeof(*args) +
 925				npages * sizeof(args->p.phys[0]), NULL);
 926
 927	mutex_unlock(&svmm->mutex);
 928}
 929
 930static void
 931nouveau_svm_fault_buffer_fini(struct nouveau_svm *svm, int id)
 932{
 933	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
 934	nvif_notify_put(&buffer->notify);
 935}
 936
 937static int
 938nouveau_svm_fault_buffer_init(struct nouveau_svm *svm, int id)
 939{
 940	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
 941	struct nvif_object *device = &svm->drm->client.device.object;
 942	buffer->get = nvif_rd32(device, buffer->getaddr);
 943	buffer->put = nvif_rd32(device, buffer->putaddr);
 944	SVM_DBG(svm, "get %08x put %08x (init)", buffer->get, buffer->put);
 945	return nvif_notify_get(&buffer->notify);
 946}
 947
 948static void
 949nouveau_svm_fault_buffer_dtor(struct nouveau_svm *svm, int id)
 950{
 951	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
 952	int i;
 953
 954	if (buffer->fault) {
 955		for (i = 0; buffer->fault[i] && i < buffer->entries; i++)
 956			kfree(buffer->fault[i]);
 957		kvfree(buffer->fault);
 958	}
 959
 960	nouveau_svm_fault_buffer_fini(svm, id);
 961
 962	nvif_notify_dtor(&buffer->notify);
 963	nvif_object_dtor(&buffer->object);
 964}
 965
 966static int
 967nouveau_svm_fault_buffer_ctor(struct nouveau_svm *svm, s32 oclass, int id)
 968{
 969	struct nouveau_svm_fault_buffer *buffer = &svm->buffer[id];
 970	struct nouveau_drm *drm = svm->drm;
 971	struct nvif_object *device = &drm->client.device.object;
 972	struct nvif_clb069_v0 args = {};
 973	int ret;
 974
 975	buffer->id = id;
 976
 977	ret = nvif_object_ctor(device, "svmFaultBuffer", 0, oclass, &args,
 978			       sizeof(args), &buffer->object);
 979	if (ret < 0) {
 980		SVM_ERR(svm, "Fault buffer allocation failed: %d", ret);
 981		return ret;
 982	}
 983
 984	nvif_object_map(&buffer->object, NULL, 0);
 985	buffer->entries = args.entries;
 986	buffer->getaddr = args.get;
 987	buffer->putaddr = args.put;
 988
 989	ret = nvif_notify_ctor(&buffer->object, "svmFault", nouveau_svm_fault,
 990			       true, NVB069_V0_NTFY_FAULT, NULL, 0, 0,
 991			       &buffer->notify);
 992	if (ret)
 993		return ret;
 994
 995	buffer->fault = kvzalloc(sizeof(*buffer->fault) * buffer->entries, GFP_KERNEL);
 996	if (!buffer->fault)
 997		return -ENOMEM;
 998
 999	return nouveau_svm_fault_buffer_init(svm, id);
1000}
1001
1002void
1003nouveau_svm_resume(struct nouveau_drm *drm)
1004{
1005	struct nouveau_svm *svm = drm->svm;
1006	if (svm)
1007		nouveau_svm_fault_buffer_init(svm, 0);
1008}
1009
1010void
1011nouveau_svm_suspend(struct nouveau_drm *drm)
1012{
1013	struct nouveau_svm *svm = drm->svm;
1014	if (svm)
1015		nouveau_svm_fault_buffer_fini(svm, 0);
1016}
1017
1018void
1019nouveau_svm_fini(struct nouveau_drm *drm)
1020{
1021	struct nouveau_svm *svm = drm->svm;
1022	if (svm) {
1023		nouveau_svm_fault_buffer_dtor(svm, 0);
1024		kfree(drm->svm);
1025		drm->svm = NULL;
1026	}
1027}
1028
1029void
1030nouveau_svm_init(struct nouveau_drm *drm)
1031{
1032	static const struct nvif_mclass buffers[] = {
1033		{   VOLTA_FAULT_BUFFER_A, 0 },
1034		{ MAXWELL_FAULT_BUFFER_A, 0 },
1035		{}
1036	};
1037	struct nouveau_svm *svm;
1038	int ret;
1039
1040	/* Disable on Volta and newer until channel recovery is fixed,
1041	 * otherwise clients will have a trivial way to trash the GPU
1042	 * for everyone.
1043	 */
1044	if (drm->client.device.info.family > NV_DEVICE_INFO_V0_PASCAL)
1045		return;
1046
1047	if (!(drm->svm = svm = kzalloc(sizeof(*drm->svm), GFP_KERNEL)))
1048		return;
1049
1050	drm->svm->drm = drm;
1051	mutex_init(&drm->svm->mutex);
1052	INIT_LIST_HEAD(&drm->svm->inst);
1053
1054	ret = nvif_mclass(&drm->client.device.object, buffers);
1055	if (ret < 0) {
1056		SVM_DBG(svm, "No supported fault buffer class");
1057		nouveau_svm_fini(drm);
1058		return;
1059	}
1060
1061	ret = nouveau_svm_fault_buffer_ctor(svm, buffers[ret].oclass, 0);
1062	if (ret) {
1063		nouveau_svm_fini(drm);
1064		return;
1065	}
1066
1067	SVM_DBG(svm, "Initialised");
1068}