1/* SPDX-License-Identifier: GPL-2.0 */
  2/* Copyright(c) 2015 Intel Corporation. All rights reserved. */
  3#include <linux/device.h>
  4#include <linux/io.h>
  5#include <linux/kasan.h>
  6#include <linux/memory_hotplug.h>
  7#include <linux/mm.h>
  8#include <linux/pfn_t.h>
  9#include <linux/swap.h>
 10#include <linux/mmzone.h>
 11#include <linux/swapops.h>
 12#include <linux/types.h>
 13#include <linux/wait_bit.h>
 14#include <linux/xarray.h>
 15
 16static DEFINE_XARRAY(pgmap_array);
 17
 18/*
 19 * The memremap() and memremap_pages() interfaces are alternately used
 20 * to map persistent memory namespaces. These interfaces place different
 21 * constraints on the alignment and size of the mapping (namespace).
 22 * memremap() can map individual PAGE_SIZE pages. memremap_pages() can
 23 * only map subsections (2MB), and at least one architecture (PowerPC)
 24 * the minimum mapping granularity of memremap_pages() is 16MB.
 25 *
 26 * The role of memremap_compat_align() is to communicate the minimum
 27 * arch supported alignment of a namespace such that it can freely
 28 * switch modes without violating the arch constraint. Namely, do not
 29 * allow a namespace to be PAGE_SIZE aligned since that namespace may be
 30 * reconfigured into a mode that requires SUBSECTION_SIZE alignment.
 31 */
 32#ifndef CONFIG_ARCH_HAS_MEMREMAP_COMPAT_ALIGN
 33unsigned long memremap_compat_align(void)
 34{
 35	return SUBSECTION_SIZE;
 36}
 37EXPORT_SYMBOL_GPL(memremap_compat_align);
 38#endif
 39
 40#ifdef CONFIG_DEV_PAGEMAP_OPS
 41DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
 42EXPORT_SYMBOL(devmap_managed_key);
 43static atomic_t devmap_managed_enable;
 44
 45static void devmap_managed_enable_put(void)
 46{
 47	if (atomic_dec_and_test(&devmap_managed_enable))
 48		static_branch_disable(&devmap_managed_key);
 49}
 50
 51static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
 52{
 53	if (pgmap->type == MEMORY_DEVICE_PRIVATE &&
 54	    (!pgmap->ops || !pgmap->ops->page_free)) {
 55		WARN(1, "Missing page_free method\n");
 56		return -EINVAL;
 57	}
 58
 59	if (atomic_inc_return(&devmap_managed_enable) == 1)
 60		static_branch_enable(&devmap_managed_key);
 61	return 0;
 62}
 63#else
 64static int devmap_managed_enable_get(struct dev_pagemap *pgmap)
 65{
 66	return -EINVAL;
 67}
 68static void devmap_managed_enable_put(void)
 69{
 70}
 71#endif /* CONFIG_DEV_PAGEMAP_OPS */
 72
 73static void pgmap_array_delete(struct resource *res)
 74{
 75	xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
 76			NULL, GFP_KERNEL);
 77	synchronize_rcu();
 78}
 79
 80static unsigned long pfn_first(struct dev_pagemap *pgmap)
 81{
 82	return PHYS_PFN(pgmap->res.start) +
 83		vmem_altmap_offset(pgmap_altmap(pgmap));
 84}
 85
 86static unsigned long pfn_end(struct dev_pagemap *pgmap)
 87{
 88	const struct resource *res = &pgmap->res;
 89
 90	return (res->start + resource_size(res)) >> PAGE_SHIFT;
 91}
 92
 93static unsigned long pfn_next(unsigned long pfn)
 94{
 95	if (pfn % 1024 == 0)
 96		cond_resched();
 97	return pfn + 1;
 98}
 99
100#define for_each_device_pfn(pfn, map) \
101	for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
102
103static void dev_pagemap_kill(struct dev_pagemap *pgmap)
104{
105	if (pgmap->ops && pgmap->ops->kill)
106		pgmap->ops->kill(pgmap);
107	else
108		percpu_ref_kill(pgmap->ref);
109}
110
111static void dev_pagemap_cleanup(struct dev_pagemap *pgmap)
112{
113	if (pgmap->ops && pgmap->ops->cleanup) {
114		pgmap->ops->cleanup(pgmap);
115	} else {
116		wait_for_completion(&pgmap->done);
117		percpu_ref_exit(pgmap->ref);
118	}
119	/*
120	 * Undo the pgmap ref assignment for the internal case as the
121	 * caller may re-enable the same pgmap.
122	 */
123	if (pgmap->ref == &pgmap->internal_ref)
124		pgmap->ref = NULL;
125}
126
127void memunmap_pages(struct dev_pagemap *pgmap)
128{
129	struct resource *res = &pgmap->res;
130	struct page *first_page;
131	unsigned long pfn;
132	int nid;
133
134	dev_pagemap_kill(pgmap);
135	for_each_device_pfn(pfn, pgmap)
136		put_page(pfn_to_page(pfn));
137	dev_pagemap_cleanup(pgmap);
138
139	/* make sure to access a memmap that was actually initialized */
140	first_page = pfn_to_page(pfn_first(pgmap));
141
142	/* pages are dead and unused, undo the arch mapping */
143	nid = page_to_nid(first_page);
144
145	mem_hotplug_begin();
146	remove_pfn_range_from_zone(page_zone(first_page), PHYS_PFN(res->start),
147				   PHYS_PFN(resource_size(res)));
148	if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
149		__remove_pages(PHYS_PFN(res->start),
150			       PHYS_PFN(resource_size(res)), NULL);
151	} else {
152		arch_remove_memory(nid, res->start, resource_size(res),
153				pgmap_altmap(pgmap));
154		kasan_remove_zero_shadow(__va(res->start), resource_size(res));
155	}
156	mem_hotplug_done();
157
158	untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
159	pgmap_array_delete(res);
160	WARN_ONCE(pgmap->altmap.alloc, "failed to free all reserved pages\n");
161	devmap_managed_enable_put();
162}
163EXPORT_SYMBOL_GPL(memunmap_pages);
164
165static void devm_memremap_pages_release(void *data)
166{
167	memunmap_pages(data);
168}
169
170static void dev_pagemap_percpu_release(struct percpu_ref *ref)
171{
172	struct dev_pagemap *pgmap =
173		container_of(ref, struct dev_pagemap, internal_ref);
174
175	complete(&pgmap->done);
176}
177
178/*
179 * Not device managed version of dev_memremap_pages, undone by
180 * memunmap_pages().  Please use dev_memremap_pages if you have a struct
181 * device available.
182 */
183void *memremap_pages(struct dev_pagemap *pgmap, int nid)
184{
185	struct resource *res = &pgmap->res;
186	struct dev_pagemap *conflict_pgmap;
187	struct mhp_params params = {
188		/*
189		 * We do not want any optional features only our own memmap
190		 */
191		.altmap = pgmap_altmap(pgmap),
192		.pgprot = PAGE_KERNEL,
193	};
194	int error, is_ram;
195	bool need_devmap_managed = true;
196
197	switch (pgmap->type) {
198	case MEMORY_DEVICE_PRIVATE:
199		if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
200			WARN(1, "Device private memory not supported\n");
201			return ERR_PTR(-EINVAL);
202		}
203		if (!pgmap->ops || !pgmap->ops->migrate_to_ram) {
204			WARN(1, "Missing migrate_to_ram method\n");
205			return ERR_PTR(-EINVAL);
206		}
207		if (!pgmap->owner) {
208			WARN(1, "Missing owner\n");
209			return ERR_PTR(-EINVAL);
210		}
211		break;
212	case MEMORY_DEVICE_FS_DAX:
213		if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
214		    IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
215			WARN(1, "File system DAX not supported\n");
216			return ERR_PTR(-EINVAL);
217		}
218		break;
219	case MEMORY_DEVICE_GENERIC:
220		need_devmap_managed = false;
221		break;
222	case MEMORY_DEVICE_PCI_P2PDMA:
223		params.pgprot = pgprot_noncached(params.pgprot);
224		need_devmap_managed = false;
225		break;
226	default:
227		WARN(1, "Invalid pgmap type %d\n", pgmap->type);
228		break;
229	}
230
231	if (!pgmap->ref) {
232		if (pgmap->ops && (pgmap->ops->kill || pgmap->ops->cleanup))
233			return ERR_PTR(-EINVAL);
234
235		init_completion(&pgmap->done);
236		error = percpu_ref_init(&pgmap->internal_ref,
237				dev_pagemap_percpu_release, 0, GFP_KERNEL);
238		if (error)
239			return ERR_PTR(error);
240		pgmap->ref = &pgmap->internal_ref;
241	} else {
242		if (!pgmap->ops || !pgmap->ops->kill || !pgmap->ops->cleanup) {
243			WARN(1, "Missing reference count teardown definition\n");
244			return ERR_PTR(-EINVAL);
245		}
246	}
247
248	if (need_devmap_managed) {
249		error = devmap_managed_enable_get(pgmap);
250		if (error)
251			return ERR_PTR(error);
252	}
253
254	conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->start), NULL);
255	if (conflict_pgmap) {
256		WARN(1, "Conflicting mapping in same section\n");
257		put_dev_pagemap(conflict_pgmap);
258		error = -ENOMEM;
259		goto err_array;
260	}
261
262	conflict_pgmap = get_dev_pagemap(PHYS_PFN(res->end), NULL);
263	if (conflict_pgmap) {
264		WARN(1, "Conflicting mapping in same section\n");
265		put_dev_pagemap(conflict_pgmap);
266		error = -ENOMEM;
267		goto err_array;
268	}
269
270	is_ram = region_intersects(res->start, resource_size(res),
271		IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
272
273	if (is_ram != REGION_DISJOINT) {
274		WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
275				is_ram == REGION_MIXED ? "mixed" : "ram", res);
276		error = -ENXIO;
277		goto err_array;
278	}
279
280	error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
281				PHYS_PFN(res->end), pgmap, GFP_KERNEL));
282	if (error)
283		goto err_array;
284
285	if (nid < 0)
286		nid = numa_mem_id();
287
288	error = track_pfn_remap(NULL, &params.pgprot, PHYS_PFN(res->start),
289				0, resource_size(res));
290	if (error)
291		goto err_pfn_remap;
292
293	mem_hotplug_begin();
294
295	/*
296	 * For device private memory we call add_pages() as we only need to
297	 * allocate and initialize struct page for the device memory. More-
298	 * over the device memory is un-accessible thus we do not want to
299	 * create a linear mapping for the memory like arch_add_memory()
300	 * would do.
301	 *
302	 * For all other device memory types, which are accessible by
303	 * the CPU, we do want the linear mapping and thus use
304	 * arch_add_memory().
305	 */
306	if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
307		error = add_pages(nid, PHYS_PFN(res->start),
308				PHYS_PFN(resource_size(res)), &params);
309	} else {
310		error = kasan_add_zero_shadow(__va(res->start), resource_size(res));
311		if (error) {
312			mem_hotplug_done();
313			goto err_kasan;
314		}
315
316		error = arch_add_memory(nid, res->start, resource_size(res),
317					&params);
318	}
319
320	if (!error) {
321		struct zone *zone;
322
323		zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
324		move_pfn_range_to_zone(zone, PHYS_PFN(res->start),
325				PHYS_PFN(resource_size(res)), params.altmap);
326	}
327
328	mem_hotplug_done();
329	if (error)
330		goto err_add_memory;
331
332	/*
333	 * Initialization of the pages has been deferred until now in order
334	 * to allow us to do the work while not holding the hotplug lock.
335	 */
336	memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
337				PHYS_PFN(res->start),
338				PHYS_PFN(resource_size(res)), pgmap);
339	percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
340	return __va(res->start);
341
342 err_add_memory:
343	kasan_remove_zero_shadow(__va(res->start), resource_size(res));
344 err_kasan:
345	untrack_pfn(NULL, PHYS_PFN(res->start), resource_size(res));
346 err_pfn_remap:
347	pgmap_array_delete(res);
348 err_array:
349	dev_pagemap_kill(pgmap);
350	dev_pagemap_cleanup(pgmap);
351	devmap_managed_enable_put();
352	return ERR_PTR(error);
353}
354EXPORT_SYMBOL_GPL(memremap_pages);
355
356/**
357 * devm_memremap_pages - remap and provide memmap backing for the given resource
358 * @dev: hosting device for @res
359 * @pgmap: pointer to a struct dev_pagemap
360 *
361 * Notes:
362 * 1/ At a minimum the res and type members of @pgmap must be initialized
363 *    by the caller before passing it to this function
364 *
365 * 2/ The altmap field may optionally be initialized, in which case
366 *    PGMAP_ALTMAP_VALID must be set in pgmap->flags.
367 *
368 * 3/ The ref field may optionally be provided, in which pgmap->ref must be
369 *    'live' on entry and will be killed and reaped at
370 *    devm_memremap_pages_release() time, or if this routine fails.
371 *
372 * 4/ res is expected to be a host memory range that could feasibly be
373 *    treated as a "System RAM" range, i.e. not a device mmio range, but
374 *    this is not enforced.
375 */
376void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
377{
378	int error;
379	void *ret;
380
381	ret = memremap_pages(pgmap, dev_to_node(dev));
382	if (IS_ERR(ret))
383		return ret;
384
385	error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
386			pgmap);
387	if (error)
388		return ERR_PTR(error);
389	return ret;
390}
391EXPORT_SYMBOL_GPL(devm_memremap_pages);
392
393void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
394{
395	devm_release_action(dev, devm_memremap_pages_release, pgmap);
396}
397EXPORT_SYMBOL_GPL(devm_memunmap_pages);
398
399unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
400{
401	/* number of pfns from base where pfn_to_page() is valid */
402	if (altmap)
403		return altmap->reserve + altmap->free;
404	return 0;
405}
406
407void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
408{
409	altmap->alloc -= nr_pfns;
410}
411
412/**
413 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
414 * @pfn: page frame number to lookup page_map
415 * @pgmap: optional known pgmap that already has a reference
416 *
417 * If @pgmap is non-NULL and covers @pfn it will be returned as-is.  If @pgmap
418 * is non-NULL but does not cover @pfn the reference to it will be released.
419 */
420struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
421		struct dev_pagemap *pgmap)
422{
423	resource_size_t phys = PFN_PHYS(pfn);
424
425	/*
426	 * In the cached case we're already holding a live reference.
427	 */
428	if (pgmap) {
429		if (phys >= pgmap->res.start && phys <= pgmap->res.end)
430			return pgmap;
431		put_dev_pagemap(pgmap);
432	}
433
434	/* fall back to slow path lookup */
435	rcu_read_lock();
436	pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
437	if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
438		pgmap = NULL;
439	rcu_read_unlock();
440
441	return pgmap;
442}
443EXPORT_SYMBOL_GPL(get_dev_pagemap);
444
445#ifdef CONFIG_DEV_PAGEMAP_OPS
446void free_devmap_managed_page(struct page *page)
447{
448	/* notify page idle for dax */
449	if (!is_device_private_page(page)) {
450		wake_up_var(&page->_refcount);
451		return;
452	}
453
454	/* Clear Active bit in case of parallel mark_page_accessed */
455	__ClearPageActive(page);
456	__ClearPageWaiters(page);
457
458	mem_cgroup_uncharge(page);
459
460	/*
461	 * When a device_private page is freed, the page->mapping field
462	 * may still contain a (stale) mapping value. For example, the
463	 * lower bits of page->mapping may still identify the page as an
464	 * anonymous page. Ultimately, this entire field is just stale
465	 * and wrong, and it will cause errors if not cleared.  One
466	 * example is:
467	 *
468	 *  migrate_vma_pages()
469	 *    migrate_vma_insert_page()
470	 *      page_add_new_anon_rmap()
471	 *        __page_set_anon_rmap()
472	 *          ...checks page->mapping, via PageAnon(page) call,
473	 *            and incorrectly concludes that the page is an
474	 *            anonymous page. Therefore, it incorrectly,
475	 *            silently fails to set up the new anon rmap.
476	 *
477	 * For other types of ZONE_DEVICE pages, migration is either
478	 * handled differently or not done at all, so there is no need
479	 * to clear page->mapping.
480	 */
481	page->mapping = NULL;
482	page->pgmap->ops->page_free(page);
483}
484#endif /* CONFIG_DEV_PAGEMAP_OPS */