drm_pagemap.c

  1// SPDX-License-Identifier: GPL-2.0-only OR MIT
  2/*
  3 * Copyright © 2024-2025 Intel Corporation
  4 */
  5
  6#include <linux/dma-fence.h>
  7#include <linux/dma-mapping.h>
  8#include <linux/migrate.h>
  9#include <linux/pagemap.h>
 10#include <drm/drm_drv.h>
 11#include <drm/drm_pagemap.h>
 12
 13/**
 14 * DOC: Overview
 15 *
 16 * The DRM pagemap layer is intended to augment the dev_pagemap functionality by
 17 * providing a way to populate a struct mm_struct virtual range with device
 18 * private pages and to provide helpers to abstract device memory allocations,
 19 * to migrate memory back and forth between device memory and system RAM and
 20 * to handle access (and in the future migration) between devices implementing
 21 * a fast interconnect that is not necessarily visible to the rest of the
 22 * system.
 23 *
 24 * Typically the DRM pagemap receives requests from one or more DRM GPU SVM
 25 * instances to populate struct mm_struct virtual ranges with memory, and the
 26 * migration is best effort only and may thus fail. The implementation should
 27 * also handle device unbinding by blocking (return an -ENODEV) error for new
 28 * population requests and after that migrate all device pages to system ram.
 29 */
 30
 31/**
 32 * DOC: Migration
 33 *
 34 * Migration granularity typically follows the GPU SVM range requests, but
 35 * if there are clashes, due to races or due to the fact that multiple GPU
 36 * SVM instances have different views of the ranges used, and because of that
 37 * parts of a requested range is already present in the requested device memory,
 38 * the implementation has a variety of options. It can fail and it can choose
 39 * to populate only the part of the range that isn't already in device memory,
 40 * and it can evict the range to system before trying to migrate. Ideally an
 41 * implementation would just try to migrate the missing part of the range and
 42 * allocate just enough memory to do so.
 43 *
 44 * When migrating to system memory as a response to a cpu fault or a device
 45 * memory eviction request, currently a full device memory allocation is
 46 * migrated back to system. Moving forward this might need improvement for
 47 * situations where a single page needs bouncing between system memory and
 48 * device memory due to, for example, atomic operations.
 49 *
 50 * Key DRM pagemap components:
 51 *
 52 * - Device Memory Allocations:
 53 *      Embedded structure containing enough information for the drm_pagemap to
 54 *      migrate to / from device memory.
 55 *
 56 * - Device Memory Operations:
 57 *      Define the interface for driver-specific device memory operations
 58 *      release memory, populate pfns, and copy to / from device memory.
 59 */
 60
 61/**
 62 * struct drm_pagemap_zdd - GPU SVM zone device data
 63 *
 64 * @refcount: Reference count for the zdd
 65 * @devmem_allocation: device memory allocation
 66 * @device_private_page_owner: Device private pages owner
 67 *
 68 * This structure serves as a generic wrapper installed in
 69 * page->zone_device_data. It provides infrastructure for looking up a device
 70 * memory allocation upon CPU page fault and asynchronously releasing device
 71 * memory once the CPU has no page references. Asynchronous release is useful
 72 * because CPU page references can be dropped in IRQ contexts, while releasing
 73 * device memory likely requires sleeping locks.
 74 */
 75struct drm_pagemap_zdd {
 76	struct kref refcount;
 77	struct drm_pagemap_devmem *devmem_allocation;
 78	void *device_private_page_owner;
 79};
 80
 81/**
 82 * drm_pagemap_zdd_alloc() - Allocate a zdd structure.
 83 * @device_private_page_owner: Device private pages owner
 84 *
 85 * This function allocates and initializes a new zdd structure. It sets up the
 86 * reference count and initializes the destroy work.
 87 *
 88 * Return: Pointer to the allocated zdd on success, ERR_PTR() on failure.
 89 */
 90static struct drm_pagemap_zdd *
 91drm_pagemap_zdd_alloc(void *device_private_page_owner)
 92{
 93	struct drm_pagemap_zdd *zdd;
 94
 95	zdd = kmalloc(sizeof(*zdd), GFP_KERNEL);
 96	if (!zdd)
 97		return NULL;
 98
 99	kref_init(&zdd->refcount);
100	zdd->devmem_allocation = NULL;
101	zdd->device_private_page_owner = device_private_page_owner;
102
103	return zdd;
104}
105
106/**
107 * drm_pagemap_zdd_get() - Get a reference to a zdd structure.
108 * @zdd: Pointer to the zdd structure.
109 *
110 * This function increments the reference count of the provided zdd structure.
111 *
112 * Return: Pointer to the zdd structure.
113 */
114static struct drm_pagemap_zdd *drm_pagemap_zdd_get(struct drm_pagemap_zdd *zdd)
115{
116	kref_get(&zdd->refcount);
117	return zdd;
118}
119
120/**
121 * drm_pagemap_zdd_destroy() - Destroy a zdd structure.
122 * @ref: Pointer to the reference count structure.
123 *
124 * This function queues the destroy_work of the zdd for asynchronous destruction.
125 */
126static void drm_pagemap_zdd_destroy(struct kref *ref)
127{
128	struct drm_pagemap_zdd *zdd =
129		container_of(ref, struct drm_pagemap_zdd, refcount);
130	struct drm_pagemap_devmem *devmem = zdd->devmem_allocation;
131
132	if (devmem) {
133		complete_all(&devmem->detached);
134		if (devmem->ops->devmem_release)
135			devmem->ops->devmem_release(devmem);
136	}
137	kfree(zdd);
138}
139
140/**
141 * drm_pagemap_zdd_put() - Put a zdd reference.
142 * @zdd: Pointer to the zdd structure.
143 *
144 * This function decrements the reference count of the provided zdd structure
145 * and schedules its destruction if the count drops to zero.
146 */
147static void drm_pagemap_zdd_put(struct drm_pagemap_zdd *zdd)
148{
149	kref_put(&zdd->refcount, drm_pagemap_zdd_destroy);
150}
151
152/**
153 * drm_pagemap_migration_unlock_put_page() - Put a migration page
154 * @page: Pointer to the page to put
155 *
156 * This function unlocks and puts a page.
157 */
158static void drm_pagemap_migration_unlock_put_page(struct page *page)
159{
160	unlock_page(page);
161	put_page(page);
162}
163
164/**
165 * drm_pagemap_migration_unlock_put_pages() - Put migration pages
166 * @npages: Number of pages
167 * @migrate_pfn: Array of migrate page frame numbers
168 *
169 * This function unlocks and puts an array of pages.
170 */
171static void drm_pagemap_migration_unlock_put_pages(unsigned long npages,
172						   unsigned long *migrate_pfn)
173{
174	unsigned long i;
175
176	for (i = 0; i < npages; ++i) {
177		struct page *page;
178
179		if (!migrate_pfn[i])
180			continue;
181
182		page = migrate_pfn_to_page(migrate_pfn[i]);
183		drm_pagemap_migration_unlock_put_page(page);
184		migrate_pfn[i] = 0;
185	}
186}
187
188/**
189 * drm_pagemap_get_devmem_page() - Get a reference to a device memory page
190 * @page: Pointer to the page
191 * @zdd: Pointer to the GPU SVM zone device data
192 *
193 * This function associates the given page with the specified GPU SVM zone
194 * device data and initializes it for zone device usage.
195 */
196static void drm_pagemap_get_devmem_page(struct page *page,
197					struct drm_pagemap_zdd *zdd)
198{
199	page->zone_device_data = drm_pagemap_zdd_get(zdd);
200	zone_device_page_init(page, page_pgmap(page), 0);
201}
202
203/**
204 * drm_pagemap_migrate_map_pages() - Map migration pages for GPU SVM migration
205 * @dev: The device for which the pages are being mapped
206 * @pagemap_addr: Array to store DMA information corresponding to mapped pages
207 * @migrate_pfn: Array of migrate page frame numbers to map
208 * @npages: Number of pages to map
209 * @dir: Direction of data transfer (e.g., DMA_BIDIRECTIONAL)
210 *
211 * This function maps pages of memory for migration usage in GPU SVM. It
212 * iterates over each page frame number provided in @migrate_pfn, maps the
213 * corresponding page, and stores the DMA address in the provided @dma_addr
214 * array.
215 *
216 * Returns: 0 on success, -EFAULT if an error occurs during mapping.
217 */
218static int drm_pagemap_migrate_map_pages(struct device *dev,
219					 struct drm_pagemap_addr *pagemap_addr,
220					 unsigned long *migrate_pfn,
221					 unsigned long npages,
222					 enum dma_data_direction dir)
223{
224	unsigned long i;
225
226	for (i = 0; i < npages;) {
227		struct page *page = migrate_pfn_to_page(migrate_pfn[i]);
228		dma_addr_t dma_addr;
229		struct folio *folio;
230		unsigned int order = 0;
231
232		if (!page)
233			goto next;
234
235		if (WARN_ON_ONCE(is_zone_device_page(page)))
236			return -EFAULT;
237
238		folio = page_folio(page);
239		order = folio_order(folio);
240
241		dma_addr = dma_map_page(dev, page, 0, page_size(page), dir);
242		if (dma_mapping_error(dev, dma_addr))
243			return -EFAULT;
244
245		pagemap_addr[i] =
246			drm_pagemap_addr_encode(dma_addr,
247						DRM_INTERCONNECT_SYSTEM,
248						order, dir);
249
250next:
251		i += NR_PAGES(order);
252	}
253
254	return 0;
255}
256
257/**
258 * drm_pagemap_migrate_unmap_pages() - Unmap pages previously mapped for GPU SVM migration
259 * @dev: The device for which the pages were mapped
260 * @pagemap_addr: Array of DMA information corresponding to mapped pages
261 * @npages: Number of pages to unmap
262 * @dir: Direction of data transfer (e.g., DMA_BIDIRECTIONAL)
263 *
264 * This function unmaps previously mapped pages of memory for GPU Shared Virtual
265 * Memory (SVM). It iterates over each DMA address provided in @dma_addr, checks
266 * if it's valid and not already unmapped, and unmaps the corresponding page.
267 */
268static void drm_pagemap_migrate_unmap_pages(struct device *dev,
269					    struct drm_pagemap_addr *pagemap_addr,
270					    unsigned long npages,
271					    enum dma_data_direction dir)
272{
273	unsigned long i;
274
275	for (i = 0; i < npages;) {
276		if (!pagemap_addr[i].addr || dma_mapping_error(dev, pagemap_addr[i].addr))
277			goto next;
278
279		dma_unmap_page(dev, pagemap_addr[i].addr, PAGE_SIZE << pagemap_addr[i].order, dir);
280
281next:
282		i += NR_PAGES(pagemap_addr[i].order);
283	}
284}
285
286static unsigned long
287npages_in_range(unsigned long start, unsigned long end)
288{
289	return (end - start) >> PAGE_SHIFT;
290}
291
292/**
293 * drm_pagemap_migrate_to_devmem() - Migrate a struct mm_struct range to device memory
294 * @devmem_allocation: The device memory allocation to migrate to.
295 * The caller should hold a reference to the device memory allocation,
296 * and the reference is consumed by this function unless it returns with
297 * an error.
298 * @mm: Pointer to the struct mm_struct.
299 * @start: Start of the virtual address range to migrate.
300 * @end: End of the virtual address range to migrate.
301 * @timeslice_ms: The time requested for the migrated pagemap pages to
302 * be present in @mm before being allowed to be migrated back.
303 * @pgmap_owner: Not used currently, since only system memory is considered.
304 *
305 * This function migrates the specified virtual address range to device memory.
306 * It performs the necessary setup and invokes the driver-specific operations for
307 * migration to device memory. Expected to be called while holding the mmap lock in
308 * at least read mode.
309 *
310 * Note: The @timeslice_ms parameter can typically be used to force data to
311 * remain in pagemap pages long enough for a GPU to perform a task and to prevent
312 * a migration livelock. One alternative would be for the GPU driver to block
313 * in a mmu_notifier for the specified amount of time, but adding the
314 * functionality to the pagemap is likely nicer to the system as a whole.
315 *
316 * Return: %0 on success, negative error code on failure.
317 */
318int drm_pagemap_migrate_to_devmem(struct drm_pagemap_devmem *devmem_allocation,
319				  struct mm_struct *mm,
320				  unsigned long start, unsigned long end,
321				  unsigned long timeslice_ms,
322				  void *pgmap_owner)
323{
324	const struct drm_pagemap_devmem_ops *ops = devmem_allocation->ops;
325	struct migrate_vma migrate = {
326		.start		= start,
327		.end		= end,
328		.pgmap_owner	= pgmap_owner,
329		.flags		= MIGRATE_VMA_SELECT_SYSTEM,
330	};
331	unsigned long i, npages = npages_in_range(start, end);
332	struct vm_area_struct *vas;
333	struct drm_pagemap_zdd *zdd = NULL;
334	struct page **pages;
335	struct drm_pagemap_addr *pagemap_addr;
336	void *buf;
337	int err;
338
339	mmap_assert_locked(mm);
340
341	if (!ops->populate_devmem_pfn || !ops->copy_to_devmem ||
342	    !ops->copy_to_ram)
343		return -EOPNOTSUPP;
344
345	vas = vma_lookup(mm, start);
346	if (!vas) {
347		err = -ENOENT;
348		goto err_out;
349	}
350
351	if (end > vas->vm_end || start < vas->vm_start) {
352		err = -EINVAL;
353		goto err_out;
354	}
355
356	if (!vma_is_anonymous(vas)) {
357		err = -EBUSY;
358		goto err_out;
359	}
360
361	buf = kvcalloc(npages, 2 * sizeof(*migrate.src) + sizeof(*pagemap_addr) +
362		       sizeof(*pages), GFP_KERNEL);
363	if (!buf) {
364		err = -ENOMEM;
365		goto err_out;
366	}
367	pagemap_addr = buf + (2 * sizeof(*migrate.src) * npages);
368	pages = buf + (2 * sizeof(*migrate.src) + sizeof(*pagemap_addr)) * npages;
369
370	zdd = drm_pagemap_zdd_alloc(pgmap_owner);
371	if (!zdd) {
372		err = -ENOMEM;
373		goto err_free;
374	}
375
376	migrate.vma = vas;
377	migrate.src = buf;
378	migrate.dst = migrate.src + npages;
379
380	err = migrate_vma_setup(&migrate);
381	if (err)
382		goto err_free;
383
384	if (!migrate.cpages) {
385		err = -EFAULT;
386		goto err_free;
387	}
388
389	if (migrate.cpages != npages) {
390		err = -EBUSY;
391		goto err_finalize;
392	}
393
394	err = ops->populate_devmem_pfn(devmem_allocation, npages, migrate.dst);
395	if (err)
396		goto err_finalize;
397
398	err = drm_pagemap_migrate_map_pages(devmem_allocation->dev, pagemap_addr,
399					    migrate.src, npages, DMA_TO_DEVICE);
400
401	if (err)
402		goto err_finalize;
403
404	for (i = 0; i < npages; ++i) {
405		struct page *page = pfn_to_page(migrate.dst[i]);
406
407		pages[i] = page;
408		migrate.dst[i] = migrate_pfn(migrate.dst[i]);
409		drm_pagemap_get_devmem_page(page, zdd);
410	}
411
412	err = ops->copy_to_devmem(pages, pagemap_addr, npages,
413				  devmem_allocation->pre_migrate_fence);
414	if (err)
415		goto err_finalize;
416
417	dma_fence_put(devmem_allocation->pre_migrate_fence);
418	devmem_allocation->pre_migrate_fence = NULL;
419
420	/* Upon success bind devmem allocation to range and zdd */
421	devmem_allocation->timeslice_expiration = get_jiffies_64() +
422		msecs_to_jiffies(timeslice_ms);
423	zdd->devmem_allocation = devmem_allocation;	/* Owns ref */
424
425err_finalize:
426	if (err)
427		drm_pagemap_migration_unlock_put_pages(npages, migrate.dst);
428	migrate_vma_pages(&migrate);
429	migrate_vma_finalize(&migrate);
430	drm_pagemap_migrate_unmap_pages(devmem_allocation->dev, pagemap_addr, npages,
431					DMA_TO_DEVICE);
432err_free:
433	if (zdd)
434		drm_pagemap_zdd_put(zdd);
435	kvfree(buf);
436err_out:
437	return err;
438}
439EXPORT_SYMBOL_GPL(drm_pagemap_migrate_to_devmem);
440
441/**
442 * drm_pagemap_migrate_populate_ram_pfn() - Populate RAM PFNs for a VM area
443 * @vas: Pointer to the VM area structure, can be NULL
444 * @fault_page: Fault page
445 * @npages: Number of pages to populate
446 * @mpages: Number of pages to migrate
447 * @src_mpfn: Source array of migrate PFNs
448 * @mpfn: Array of migrate PFNs to populate
449 * @addr: Start address for PFN allocation
450 *
451 * This function populates the RAM migrate page frame numbers (PFNs) for the
452 * specified VM area structure. It allocates and locks pages in the VM area for
453 * RAM usage. If vas is non-NULL use alloc_page_vma for allocation, if NULL use
454 * alloc_page for allocation.
455 *
456 * Return: 0 on success, negative error code on failure.
457 */
458static int drm_pagemap_migrate_populate_ram_pfn(struct vm_area_struct *vas,
459						struct page *fault_page,
460						unsigned long npages,
461						unsigned long *mpages,
462						unsigned long *src_mpfn,
463						unsigned long *mpfn,
464						unsigned long addr)
465{
466	unsigned long i;
467
468	for (i = 0; i < npages;) {
469		struct page *page = NULL, *src_page;
470		struct folio *folio;
471		unsigned int order = 0;
472
473		if (!(src_mpfn[i] & MIGRATE_PFN_MIGRATE))
474			goto next;
475
476		src_page = migrate_pfn_to_page(src_mpfn[i]);
477		if (!src_page)
478			goto next;
479
480		if (fault_page) {
481			if (src_page->zone_device_data !=
482			    fault_page->zone_device_data)
483				goto next;
484		}
485
486		order = folio_order(page_folio(src_page));
487
488		/* TODO: Support fallback to single pages if THP allocation fails */
489		if (vas)
490			folio = vma_alloc_folio(GFP_HIGHUSER, order, vas, addr);
491		else
492			folio = folio_alloc(GFP_HIGHUSER, order);
493
494		if (!folio)
495			goto free_pages;
496
497		page = folio_page(folio, 0);
498		mpfn[i] = migrate_pfn(page_to_pfn(page));
499
500next:
501		if (page)
502			addr += page_size(page);
503		else
504			addr += PAGE_SIZE;
505
506		i += NR_PAGES(order);
507	}
508
509	for (i = 0; i < npages;) {
510		struct page *page = migrate_pfn_to_page(mpfn[i]);
511		unsigned int order = 0;
512
513		if (!page)
514			goto next_lock;
515
516		WARN_ON_ONCE(!folio_trylock(page_folio(page)));
517
518		order = folio_order(page_folio(page));
519		*mpages += NR_PAGES(order);
520
521next_lock:
522		i += NR_PAGES(order);
523	}
524
525	return 0;
526
527free_pages:
528	for (i = 0; i < npages;) {
529		struct page *page = migrate_pfn_to_page(mpfn[i]);
530		unsigned int order = 0;
531
532		if (!page)
533			goto next_put;
534
535		put_page(page);
536		mpfn[i] = 0;
537
538		order = folio_order(page_folio(page));
539
540next_put:
541		i += NR_PAGES(order);
542	}
543	return -ENOMEM;
544}
545
546/**
547 * drm_pagemap_evict_to_ram() - Evict GPU SVM range to RAM
548 * @devmem_allocation: Pointer to the device memory allocation
549 *
550 * Similar to __drm_pagemap_migrate_to_ram but does not require mmap lock and
551 * migration done via migrate_device_* functions.
552 *
553 * Return: 0 on success, negative error code on failure.
554 */
555int drm_pagemap_evict_to_ram(struct drm_pagemap_devmem *devmem_allocation)
556{
557	const struct drm_pagemap_devmem_ops *ops = devmem_allocation->ops;
558	unsigned long npages, mpages = 0;
559	struct page **pages;
560	unsigned long *src, *dst;
561	struct drm_pagemap_addr *pagemap_addr;
562	void *buf;
563	int i, err = 0;
564	unsigned int retry_count = 2;
565
566	npages = devmem_allocation->size >> PAGE_SHIFT;
567
568retry:
569	if (!mmget_not_zero(devmem_allocation->mm))
570		return -EFAULT;
571
572	buf = kvcalloc(npages, 2 * sizeof(*src) + sizeof(*pagemap_addr) +
573		       sizeof(*pages), GFP_KERNEL);
574	if (!buf) {
575		err = -ENOMEM;
576		goto err_out;
577	}
578	src = buf;
579	dst = buf + (sizeof(*src) * npages);
580	pagemap_addr = buf + (2 * sizeof(*src) * npages);
581	pages = buf + (2 * sizeof(*src) + sizeof(*pagemap_addr)) * npages;
582
583	err = ops->populate_devmem_pfn(devmem_allocation, npages, src);
584	if (err)
585		goto err_free;
586
587	err = migrate_device_pfns(src, npages);
588	if (err)
589		goto err_free;
590
591	err = drm_pagemap_migrate_populate_ram_pfn(NULL, NULL, npages, &mpages,
592						   src, dst, 0);
593	if (err || !mpages)
594		goto err_finalize;
595
596	err = drm_pagemap_migrate_map_pages(devmem_allocation->dev, pagemap_addr,
597					    dst, npages, DMA_FROM_DEVICE);
598	if (err)
599		goto err_finalize;
600
601	for (i = 0; i < npages; ++i)
602		pages[i] = migrate_pfn_to_page(src[i]);
603
604	err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
605	if (err)
606		goto err_finalize;
607
608err_finalize:
609	if (err)
610		drm_pagemap_migration_unlock_put_pages(npages, dst);
611	migrate_device_pages(src, dst, npages);
612	migrate_device_finalize(src, dst, npages);
613	drm_pagemap_migrate_unmap_pages(devmem_allocation->dev, pagemap_addr, npages,
614					DMA_FROM_DEVICE);
615err_free:
616	kvfree(buf);
617err_out:
618	mmput_async(devmem_allocation->mm);
619
620	if (completion_done(&devmem_allocation->detached))
621		return 0;
622
623	if (retry_count--) {
624		cond_resched();
625		goto retry;
626	}
627
628	return err ?: -EBUSY;
629}
630EXPORT_SYMBOL_GPL(drm_pagemap_evict_to_ram);
631
632/**
633 * __drm_pagemap_migrate_to_ram() - Migrate GPU SVM range to RAM (internal)
634 * @vas: Pointer to the VM area structure
635 * @device_private_page_owner: Device private pages owner
636 * @page: Pointer to the page for fault handling (can be NULL)
637 * @fault_addr: Fault address
638 * @size: Size of migration
639 *
640 * This internal function performs the migration of the specified GPU SVM range
641 * to RAM. It sets up the migration, populates + dma maps RAM PFNs, and
642 * invokes the driver-specific operations for migration to RAM.
643 *
644 * Return: 0 on success, negative error code on failure.
645 */
646static int __drm_pagemap_migrate_to_ram(struct vm_area_struct *vas,
647					void *device_private_page_owner,
648					struct page *page,
649					unsigned long fault_addr,
650					unsigned long size)
651{
652	struct migrate_vma migrate = {
653		.vma		= vas,
654		.pgmap_owner	= device_private_page_owner,
655		.flags		= MIGRATE_VMA_SELECT_DEVICE_PRIVATE |
656		MIGRATE_VMA_SELECT_DEVICE_COHERENT,
657		.fault_page	= page,
658	};
659	struct drm_pagemap_zdd *zdd;
660	const struct drm_pagemap_devmem_ops *ops;
661	struct device *dev = NULL;
662	unsigned long npages, mpages = 0;
663	struct page **pages;
664	struct drm_pagemap_addr *pagemap_addr;
665	unsigned long start, end;
666	void *buf;
667	int i, err = 0;
668
669	if (page) {
670		zdd = page->zone_device_data;
671		if (time_before64(get_jiffies_64(),
672				  zdd->devmem_allocation->timeslice_expiration))
673			return 0;
674	}
675
676	start = ALIGN_DOWN(fault_addr, size);
677	end = ALIGN(fault_addr + 1, size);
678
679	/* Corner where VMA area struct has been partially unmapped */
680	if (start < vas->vm_start)
681		start = vas->vm_start;
682	if (end > vas->vm_end)
683		end = vas->vm_end;
684
685	migrate.start = start;
686	migrate.end = end;
687	npages = npages_in_range(start, end);
688
689	buf = kvcalloc(npages, 2 * sizeof(*migrate.src) + sizeof(*pagemap_addr) +
690		       sizeof(*pages), GFP_KERNEL);
691	if (!buf) {
692		err = -ENOMEM;
693		goto err_out;
694	}
695	pagemap_addr = buf + (2 * sizeof(*migrate.src) * npages);
696	pages = buf + (2 * sizeof(*migrate.src) + sizeof(*pagemap_addr)) * npages;
697
698	migrate.vma = vas;
699	migrate.src = buf;
700	migrate.dst = migrate.src + npages;
701
702	err = migrate_vma_setup(&migrate);
703	if (err)
704		goto err_free;
705
706	/* Raced with another CPU fault, nothing to do */
707	if (!migrate.cpages)
708		goto err_free;
709
710	if (!page) {
711		for (i = 0; i < npages; ++i) {
712			if (!(migrate.src[i] & MIGRATE_PFN_MIGRATE))
713				continue;
714
715			page = migrate_pfn_to_page(migrate.src[i]);
716			break;
717		}
718
719		if (!page)
720			goto err_finalize;
721	}
722	zdd = page->zone_device_data;
723	ops = zdd->devmem_allocation->ops;
724	dev = zdd->devmem_allocation->dev;
725
726	err = drm_pagemap_migrate_populate_ram_pfn(vas, page, npages, &mpages,
727						   migrate.src, migrate.dst,
728						   start);
729	if (err)
730		goto err_finalize;
731
732	err = drm_pagemap_migrate_map_pages(dev, pagemap_addr, migrate.dst, npages,
733					    DMA_FROM_DEVICE);
734	if (err)
735		goto err_finalize;
736
737	for (i = 0; i < npages; ++i)
738		pages[i] = migrate_pfn_to_page(migrate.src[i]);
739
740	err = ops->copy_to_ram(pages, pagemap_addr, npages, NULL);
741	if (err)
742		goto err_finalize;
743
744err_finalize:
745	if (err)
746		drm_pagemap_migration_unlock_put_pages(npages, migrate.dst);
747	migrate_vma_pages(&migrate);
748	migrate_vma_finalize(&migrate);
749	if (dev)
750		drm_pagemap_migrate_unmap_pages(dev, pagemap_addr, npages,
751						DMA_FROM_DEVICE);
752err_free:
753	kvfree(buf);
754err_out:
755
756	return err;
757}
758
759/**
760 * drm_pagemap_folio_free() - Put GPU SVM zone device data associated with a folio
761 * @folio: Pointer to the folio
762 *
763 * This function is a callback used to put the GPU SVM zone device data
764 * associated with a page when it is being released.
765 */
766static void drm_pagemap_folio_free(struct folio *folio)
767{
768	drm_pagemap_zdd_put(folio->page.zone_device_data);
769}
770
771/**
772 * drm_pagemap_migrate_to_ram() - Migrate a virtual range to RAM (page fault handler)
773 * @vmf: Pointer to the fault information structure
774 *
775 * This function is a page fault handler used to migrate a virtual range
776 * to ram. The device memory allocation in which the device page is found is
777 * migrated in its entirety.
778 *
779 * Returns:
780 * VM_FAULT_SIGBUS on failure, 0 on success.
781 */
782static vm_fault_t drm_pagemap_migrate_to_ram(struct vm_fault *vmf)
783{
784	struct drm_pagemap_zdd *zdd = vmf->page->zone_device_data;
785	int err;
786
787	err = __drm_pagemap_migrate_to_ram(vmf->vma,
788					   zdd->device_private_page_owner,
789					   vmf->page, vmf->address,
790					   zdd->devmem_allocation->size);
791
792	return err ? VM_FAULT_SIGBUS : 0;
793}
794
795static const struct dev_pagemap_ops drm_pagemap_pagemap_ops = {
796	.folio_free = drm_pagemap_folio_free,
797	.migrate_to_ram = drm_pagemap_migrate_to_ram,
798};
799
800/**
801 * drm_pagemap_pagemap_ops_get() - Retrieve GPU SVM device page map operations
802 *
803 * Returns:
804 * Pointer to the GPU SVM device page map operations structure.
805 */
806const struct dev_pagemap_ops *drm_pagemap_pagemap_ops_get(void)
807{
808	return &drm_pagemap_pagemap_ops;
809}
810EXPORT_SYMBOL_GPL(drm_pagemap_pagemap_ops_get);
811
812/**
813 * drm_pagemap_devmem_init() - Initialize a drm_pagemap device memory allocation
814 *
815 * @devmem_allocation: The struct drm_pagemap_devmem to initialize.
816 * @dev: Pointer to the device structure which device memory allocation belongs to
817 * @mm: Pointer to the mm_struct for the address space
818 * @ops: Pointer to the operations structure for GPU SVM device memory
819 * @dpagemap: The struct drm_pagemap we're allocating from.
820 * @size: Size of device memory allocation
821 * @pre_migrate_fence: Fence to wait for or pipeline behind before migration starts.
822 * (May be NULL).
823 */
824void drm_pagemap_devmem_init(struct drm_pagemap_devmem *devmem_allocation,
825			     struct device *dev, struct mm_struct *mm,
826			     const struct drm_pagemap_devmem_ops *ops,
827			     struct drm_pagemap *dpagemap, size_t size,
828			     struct dma_fence *pre_migrate_fence)
829{
830	init_completion(&devmem_allocation->detached);
831	devmem_allocation->dev = dev;
832	devmem_allocation->mm = mm;
833	devmem_allocation->ops = ops;
834	devmem_allocation->dpagemap = dpagemap;
835	devmem_allocation->size = size;
836	devmem_allocation->pre_migrate_fence = pre_migrate_fence;
837}
838EXPORT_SYMBOL_GPL(drm_pagemap_devmem_init);
839
840/**
841 * drm_pagemap_page_to_dpagemap() - Return a pointer the drm_pagemap of a page
842 * @page: The struct page.
843 *
844 * Return: A pointer to the struct drm_pagemap of a device private page that
845 * was populated from the struct drm_pagemap. If the page was *not* populated
846 * from a struct drm_pagemap, the result is undefined and the function call
847 * may result in dereferencing and invalid address.
848 */
849struct drm_pagemap *drm_pagemap_page_to_dpagemap(struct page *page)
850{
851	struct drm_pagemap_zdd *zdd = page->zone_device_data;
852
853	return zdd->devmem_allocation->dpagemap;
854}
855EXPORT_SYMBOL_GPL(drm_pagemap_page_to_dpagemap);
856
857/**
858 * drm_pagemap_populate_mm() - Populate a virtual range with device memory pages
859 * @dpagemap: Pointer to the drm_pagemap managing the device memory
860 * @start: Start of the virtual range to populate.
861 * @end: End of the virtual range to populate.
862 * @mm: Pointer to the virtual address space.
863 * @timeslice_ms: The time requested for the migrated pagemap pages to
864 * be present in @mm before being allowed to be migrated back.
865 *
866 * Attempt to populate a virtual range with device memory pages,
867 * clearing them or migrating data from the existing pages if necessary.
868 * The function is best effort only, and implementations may vary
869 * in how hard they try to satisfy the request.
870 *
871 * Return: %0 on success, negative error code on error. If the hardware
872 * device was removed / unbound the function will return %-ENODEV.
873 */
874int drm_pagemap_populate_mm(struct drm_pagemap *dpagemap,
875			    unsigned long start, unsigned long end,
876			    struct mm_struct *mm,
877			    unsigned long timeslice_ms)
878{
879	int err;
880
881	if (!mmget_not_zero(mm))
882		return -EFAULT;
883	mmap_read_lock(mm);
884	err = dpagemap->ops->populate_mm(dpagemap, start, end, mm,
885					 timeslice_ms);
886	mmap_read_unlock(mm);
887	mmput(mm);
888
889	return err;
890}
891EXPORT_SYMBOL(drm_pagemap_populate_mm);