1/*
  2 * Copyright(c) 2015 Intel Corporation. All rights reserved.
  3 *
  4 * This program is free software; you can redistribute it and/or modify
  5 * it under the terms of version 2 of the GNU General Public License as
  6 * published by the Free Software Foundation.
  7 *
  8 * This program is distributed in the hope that it will be useful, but
  9 * WITHOUT ANY WARRANTY; without even the implied warranty of
 10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 11 * General Public License for more details.
 12 */
 13#include <linux/radix-tree.h>
 14#include <linux/memremap.h>
 15#include <linux/device.h>
 16#include <linux/types.h>
 17#include <linux/pfn_t.h>
 18#include <linux/io.h>
 19#include <linux/mm.h>
 20#include <linux/memory_hotplug.h>
 21
 22#ifndef ioremap_cache
 23/* temporary while we convert existing ioremap_cache users to memremap */
 24__weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
 25{
 26	return ioremap(offset, size);
 27}
 28#endif
 29
 30#ifndef arch_memremap_wb
 31static void *arch_memremap_wb(resource_size_t offset, unsigned long size)
 32{
 33	return (__force void *)ioremap_cache(offset, size);
 34}
 35#endif
 36
 37static void *try_ram_remap(resource_size_t offset, size_t size)
 38{
 39	unsigned long pfn = PHYS_PFN(offset);
 40
 41	/* In the simple case just return the existing linear address */
 42	if (pfn_valid(pfn) && !PageHighMem(pfn_to_page(pfn)))
 43		return __va(offset);
 44	return NULL; /* fallback to arch_memremap_wb */
 45}
 46
 47/**
 48 * memremap() - remap an iomem_resource as cacheable memory
 49 * @offset: iomem resource start address
 50 * @size: size of remap
 51 * @flags: any of MEMREMAP_WB, MEMREMAP_WT and MEMREMAP_WC
 52 *
 53 * memremap() is "ioremap" for cases where it is known that the resource
 54 * being mapped does not have i/o side effects and the __iomem
 55 * annotation is not applicable. In the case of multiple flags, the different
 56 * mapping types will be attempted in the order listed below until one of
 57 * them succeeds.
 58 *
 59 * MEMREMAP_WB - matches the default mapping for System RAM on
 60 * the architecture.  This is usually a read-allocate write-back cache.
 61 * Morever, if MEMREMAP_WB is specified and the requested remap region is RAM
 62 * memremap() will bypass establishing a new mapping and instead return
 63 * a pointer into the direct map.
 64 *
 65 * MEMREMAP_WT - establish a mapping whereby writes either bypass the
 66 * cache or are written through to memory and never exist in a
 67 * cache-dirty state with respect to program visibility.  Attempts to
 68 * map System RAM with this mapping type will fail.
 69 *
 70 * MEMREMAP_WC - establish a writecombine mapping, whereby writes may
 71 * be coalesced together (e.g. in the CPU's write buffers), but is otherwise
 72 * uncached. Attempts to map System RAM with this mapping type will fail.
 73 */
 74void *memremap(resource_size_t offset, size_t size, unsigned long flags)
 75{
 76	int is_ram = region_intersects(offset, size,
 77				       IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
 78	void *addr = NULL;
 79
 80	if (!flags)
 81		return NULL;
 82
 83	if (is_ram == REGION_MIXED) {
 84		WARN_ONCE(1, "memremap attempted on mixed range %pa size: %#lx\n",
 85				&offset, (unsigned long) size);
 86		return NULL;
 87	}
 88
 89	/* Try all mapping types requested until one returns non-NULL */
 90	if (flags & MEMREMAP_WB) {
 91		/*
 92		 * MEMREMAP_WB is special in that it can be satisifed
 93		 * from the direct map.  Some archs depend on the
 94		 * capability of memremap() to autodetect cases where
 95		 * the requested range is potentially in System RAM.
 96		 */
 97		if (is_ram == REGION_INTERSECTS)
 98			addr = try_ram_remap(offset, size);
 99		if (!addr)
100			addr = arch_memremap_wb(offset, size);
101	}
102
103	/*
104	 * If we don't have a mapping yet and other request flags are
105	 * present then we will be attempting to establish a new virtual
106	 * address mapping.  Enforce that this mapping is not aliasing
107	 * System RAM.
108	 */
109	if (!addr && is_ram == REGION_INTERSECTS && flags != MEMREMAP_WB) {
110		WARN_ONCE(1, "memremap attempted on ram %pa size: %#lx\n",
111				&offset, (unsigned long) size);
112		return NULL;
113	}
114
115	if (!addr && (flags & MEMREMAP_WT))
116		addr = ioremap_wt(offset, size);
117
118	if (!addr && (flags & MEMREMAP_WC))
119		addr = ioremap_wc(offset, size);
120
121	return addr;
122}
123EXPORT_SYMBOL(memremap);
124
125void memunmap(void *addr)
126{
127	if (is_vmalloc_addr(addr))
128		iounmap((void __iomem *) addr);
129}
130EXPORT_SYMBOL(memunmap);
131
132static void devm_memremap_release(struct device *dev, void *res)
133{
134	memunmap(*(void **)res);
135}
136
137static int devm_memremap_match(struct device *dev, void *res, void *match_data)
138{
139	return *(void **)res == match_data;
140}
141
142void *devm_memremap(struct device *dev, resource_size_t offset,
143		size_t size, unsigned long flags)
144{
145	void **ptr, *addr;
146
147	ptr = devres_alloc_node(devm_memremap_release, sizeof(*ptr), GFP_KERNEL,
148			dev_to_node(dev));
149	if (!ptr)
150		return ERR_PTR(-ENOMEM);
151
152	addr = memremap(offset, size, flags);
153	if (addr) {
154		*ptr = addr;
155		devres_add(dev, ptr);
156	} else {
157		devres_free(ptr);
158		return ERR_PTR(-ENXIO);
159	}
160
161	return addr;
162}
163EXPORT_SYMBOL(devm_memremap);
164
165void devm_memunmap(struct device *dev, void *addr)
166{
167	WARN_ON(devres_release(dev, devm_memremap_release,
168				devm_memremap_match, addr));
169}
170EXPORT_SYMBOL(devm_memunmap);
171
172#ifdef CONFIG_ZONE_DEVICE
173static DEFINE_MUTEX(pgmap_lock);
174static RADIX_TREE(pgmap_radix, GFP_KERNEL);
175#define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
176#define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
177
178struct page_map {
179	struct resource res;
180	struct percpu_ref *ref;
181	struct dev_pagemap pgmap;
182	struct vmem_altmap altmap;
183};
184
185void get_zone_device_page(struct page *page)
186{
187	percpu_ref_get(page->pgmap->ref);
188}
189EXPORT_SYMBOL(get_zone_device_page);
190
191void put_zone_device_page(struct page *page)
192{
193	put_dev_pagemap(page->pgmap);
194}
195EXPORT_SYMBOL(put_zone_device_page);
196
197static void pgmap_radix_release(struct resource *res)
198{
199	resource_size_t key, align_start, align_size, align_end;
200
201	align_start = res->start & ~(SECTION_SIZE - 1);
202	align_size = ALIGN(resource_size(res), SECTION_SIZE);
203	align_end = align_start + align_size - 1;
204
205	mutex_lock(&pgmap_lock);
206	for (key = res->start; key <= res->end; key += SECTION_SIZE)
207		radix_tree_delete(&pgmap_radix, key >> PA_SECTION_SHIFT);
208	mutex_unlock(&pgmap_lock);
209}
210
211static unsigned long pfn_first(struct page_map *page_map)
212{
213	struct dev_pagemap *pgmap = &page_map->pgmap;
214	const struct resource *res = &page_map->res;
215	struct vmem_altmap *altmap = pgmap->altmap;
216	unsigned long pfn;
217
218	pfn = res->start >> PAGE_SHIFT;
219	if (altmap)
220		pfn += vmem_altmap_offset(altmap);
221	return pfn;
222}
223
224static unsigned long pfn_end(struct page_map *page_map)
225{
226	const struct resource *res = &page_map->res;
227
228	return (res->start + resource_size(res)) >> PAGE_SHIFT;
229}
230
231#define for_each_device_pfn(pfn, map) \
232	for (pfn = pfn_first(map); pfn < pfn_end(map); pfn++)
233
234static void devm_memremap_pages_release(struct device *dev, void *data)
235{
236	struct page_map *page_map = data;
237	struct resource *res = &page_map->res;
238	resource_size_t align_start, align_size;
239	struct dev_pagemap *pgmap = &page_map->pgmap;
240
241	if (percpu_ref_tryget_live(pgmap->ref)) {
242		dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
243		percpu_ref_put(pgmap->ref);
244	}
245
246	/* pages are dead and unused, undo the arch mapping */
247	align_start = res->start & ~(SECTION_SIZE - 1);
248	align_size = ALIGN(resource_size(res), SECTION_SIZE);
249
250	lock_device_hotplug();
251	mem_hotplug_begin();
252	arch_remove_memory(align_start, align_size);
253	mem_hotplug_done();
254	unlock_device_hotplug();
255
256	untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
257	pgmap_radix_release(res);
258	dev_WARN_ONCE(dev, pgmap->altmap && pgmap->altmap->alloc,
259			"%s: failed to free all reserved pages\n", __func__);
260}
261
262/* assumes rcu_read_lock() held at entry */
263struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
264{
265	struct page_map *page_map;
266
267	WARN_ON_ONCE(!rcu_read_lock_held());
268
269	page_map = radix_tree_lookup(&pgmap_radix, phys >> PA_SECTION_SHIFT);
270	return page_map ? &page_map->pgmap : NULL;
271}
272
273/**
274 * devm_memremap_pages - remap and provide memmap backing for the given resource
275 * @dev: hosting device for @res
276 * @res: "host memory" address range
277 * @ref: a live per-cpu reference count
278 * @altmap: optional descriptor for allocating the memmap from @res
279 *
280 * Notes:
281 * 1/ @ref must be 'live' on entry and 'dead' before devm_memunmap_pages() time
282 *    (or devm release event).
283 *
284 * 2/ @res is expected to be a host memory range that could feasibly be
285 *    treated as a "System RAM" range, i.e. not a device mmio range, but
286 *    this is not enforced.
287 */
288void *devm_memremap_pages(struct device *dev, struct resource *res,
289		struct percpu_ref *ref, struct vmem_altmap *altmap)
290{
291	resource_size_t key, align_start, align_size, align_end;
292	pgprot_t pgprot = PAGE_KERNEL;
293	struct dev_pagemap *pgmap;
294	struct page_map *page_map;
295	int error, nid, is_ram;
296	unsigned long pfn;
297
298	align_start = res->start & ~(SECTION_SIZE - 1);
299	align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
300		- align_start;
301	is_ram = region_intersects(align_start, align_size,
302		IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
303
304	if (is_ram == REGION_MIXED) {
305		WARN_ONCE(1, "%s attempted on mixed region %pr\n",
306				__func__, res);
307		return ERR_PTR(-ENXIO);
308	}
309
310	if (is_ram == REGION_INTERSECTS)
311		return __va(res->start);
312
313	if (!ref)
314		return ERR_PTR(-EINVAL);
315
316	page_map = devres_alloc_node(devm_memremap_pages_release,
317			sizeof(*page_map), GFP_KERNEL, dev_to_node(dev));
318	if (!page_map)
319		return ERR_PTR(-ENOMEM);
320	pgmap = &page_map->pgmap;
321
322	memcpy(&page_map->res, res, sizeof(*res));
323
324	pgmap->dev = dev;
325	if (altmap) {
326		memcpy(&page_map->altmap, altmap, sizeof(*altmap));
327		pgmap->altmap = &page_map->altmap;
328	}
329	pgmap->ref = ref;
330	pgmap->res = &page_map->res;
331
332	mutex_lock(&pgmap_lock);
333	error = 0;
334	align_end = align_start + align_size - 1;
335	for (key = align_start; key <= align_end; key += SECTION_SIZE) {
336		struct dev_pagemap *dup;
337
338		rcu_read_lock();
339		dup = find_dev_pagemap(key);
340		rcu_read_unlock();
341		if (dup) {
342			dev_err(dev, "%s: %pr collides with mapping for %s\n",
343					__func__, res, dev_name(dup->dev));
344			error = -EBUSY;
345			break;
346		}
347		error = radix_tree_insert(&pgmap_radix, key >> PA_SECTION_SHIFT,
348				page_map);
349		if (error) {
350			dev_err(dev, "%s: failed: %d\n", __func__, error);
351			break;
352		}
353	}
354	mutex_unlock(&pgmap_lock);
355	if (error)
356		goto err_radix;
357
358	nid = dev_to_node(dev);
359	if (nid < 0)
360		nid = numa_mem_id();
361
362	error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
363			align_size);
364	if (error)
365		goto err_pfn_remap;
366
367	lock_device_hotplug();
368	mem_hotplug_begin();
369	error = arch_add_memory(nid, align_start, align_size, true);
370	mem_hotplug_done();
371	unlock_device_hotplug();
372	if (error)
373		goto err_add_memory;
374
375	for_each_device_pfn(pfn, page_map) {
376		struct page *page = pfn_to_page(pfn);
377
378		/*
379		 * ZONE_DEVICE pages union ->lru with a ->pgmap back
380		 * pointer.  It is a bug if a ZONE_DEVICE page is ever
381		 * freed or placed on a driver-private list.  Seed the
382		 * storage with LIST_POISON* values.
383		 */
384		list_del(&page->lru);
385		page->pgmap = pgmap;
386	}
387	devres_add(dev, page_map);
388	return __va(res->start);
389
390 err_add_memory:
391	untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
392 err_pfn_remap:
393 err_radix:
394	pgmap_radix_release(res);
395	devres_free(page_map);
396	return ERR_PTR(error);
397}
398EXPORT_SYMBOL(devm_memremap_pages);
399
400unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
401{
402	/* number of pfns from base where pfn_to_page() is valid */
403	return altmap->reserve + altmap->free;
404}
405
406void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
407{
408	altmap->alloc -= nr_pfns;
409}
410
411struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
412{
413	/*
414	 * 'memmap_start' is the virtual address for the first "struct
415	 * page" in this range of the vmemmap array.  In the case of
416	 * CONFIG_SPARSEMEM_VMEMMAP a page_to_pfn conversion is simple
417	 * pointer arithmetic, so we can perform this to_vmem_altmap()
418	 * conversion without concern for the initialization state of
419	 * the struct page fields.
420	 */
421	struct page *page = (struct page *) memmap_start;
422	struct dev_pagemap *pgmap;
423
424	/*
425	 * Unconditionally retrieve a dev_pagemap associated with the
426	 * given physical address, this is only for use in the
427	 * arch_{add|remove}_memory() for setting up and tearing down
428	 * the memmap.
429	 */
430	rcu_read_lock();
431	pgmap = find_dev_pagemap(__pfn_to_phys(page_to_pfn(page)));
432	rcu_read_unlock();
433
434	return pgmap ? pgmap->altmap : NULL;
435}
436#endif /* CONFIG_ZONE_DEVICE */