Loading...
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 */
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/device.h>
15#include <linux/types.h>
16#include <linux/pfn_t.h>
17#include <linux/io.h>
18#include <linux/mm.h>
19#include <linux/memory_hotplug.h>
20#include <linux/swap.h>
21#include <linux/swapops.h>
22
23#ifndef ioremap_cache
24/* temporary while we convert existing ioremap_cache users to memremap */
25__weak void __iomem *ioremap_cache(resource_size_t offset, unsigned long size)
26{
27 return ioremap(offset, size);
28}
29#endif
30
31#ifndef arch_memremap_wb
32static void *arch_memremap_wb(resource_size_t offset, unsigned long size)
33{
34 return (__force void *)ioremap_cache(offset, size);
35}
36#endif
37
38#ifndef arch_memremap_can_ram_remap
39static bool arch_memremap_can_ram_remap(resource_size_t offset, size_t size,
40 unsigned long flags)
41{
42 return true;
43}
44#endif
45
46static void *try_ram_remap(resource_size_t offset, size_t size,
47 unsigned long flags)
48{
49 unsigned long pfn = PHYS_PFN(offset);
50
51 /* In the simple case just return the existing linear address */
52 if (pfn_valid(pfn) && !PageHighMem(pfn_to_page(pfn)) &&
53 arch_memremap_can_ram_remap(offset, size, flags))
54 return __va(offset);
55
56 return NULL; /* fallback to arch_memremap_wb */
57}
58
59/**
60 * memremap() - remap an iomem_resource as cacheable memory
61 * @offset: iomem resource start address
62 * @size: size of remap
63 * @flags: any of MEMREMAP_WB, MEMREMAP_WT, MEMREMAP_WC,
64 * MEMREMAP_ENC, MEMREMAP_DEC
65 *
66 * memremap() is "ioremap" for cases where it is known that the resource
67 * being mapped does not have i/o side effects and the __iomem
68 * annotation is not applicable. In the case of multiple flags, the different
69 * mapping types will be attempted in the order listed below until one of
70 * them succeeds.
71 *
72 * MEMREMAP_WB - matches the default mapping for System RAM on
73 * the architecture. This is usually a read-allocate write-back cache.
74 * Morever, if MEMREMAP_WB is specified and the requested remap region is RAM
75 * memremap() will bypass establishing a new mapping and instead return
76 * a pointer into the direct map.
77 *
78 * MEMREMAP_WT - establish a mapping whereby writes either bypass the
79 * cache or are written through to memory and never exist in a
80 * cache-dirty state with respect to program visibility. Attempts to
81 * map System RAM with this mapping type will fail.
82 *
83 * MEMREMAP_WC - establish a writecombine mapping, whereby writes may
84 * be coalesced together (e.g. in the CPU's write buffers), but is otherwise
85 * uncached. Attempts to map System RAM with this mapping type will fail.
86 */
87void *memremap(resource_size_t offset, size_t size, unsigned long flags)
88{
89 int is_ram = region_intersects(offset, size,
90 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
91 void *addr = NULL;
92
93 if (!flags)
94 return NULL;
95
96 if (is_ram == REGION_MIXED) {
97 WARN_ONCE(1, "memremap attempted on mixed range %pa size: %#lx\n",
98 &offset, (unsigned long) size);
99 return NULL;
100 }
101
102 /* Try all mapping types requested until one returns non-NULL */
103 if (flags & MEMREMAP_WB) {
104 /*
105 * MEMREMAP_WB is special in that it can be satisifed
106 * from the direct map. Some archs depend on the
107 * capability of memremap() to autodetect cases where
108 * the requested range is potentially in System RAM.
109 */
110 if (is_ram == REGION_INTERSECTS)
111 addr = try_ram_remap(offset, size, flags);
112 if (!addr)
113 addr = arch_memremap_wb(offset, size);
114 }
115
116 /*
117 * If we don't have a mapping yet and other request flags are
118 * present then we will be attempting to establish a new virtual
119 * address mapping. Enforce that this mapping is not aliasing
120 * System RAM.
121 */
122 if (!addr && is_ram == REGION_INTERSECTS && flags != MEMREMAP_WB) {
123 WARN_ONCE(1, "memremap attempted on ram %pa size: %#lx\n",
124 &offset, (unsigned long) size);
125 return NULL;
126 }
127
128 if (!addr && (flags & MEMREMAP_WT))
129 addr = ioremap_wt(offset, size);
130
131 if (!addr && (flags & MEMREMAP_WC))
132 addr = ioremap_wc(offset, size);
133
134 return addr;
135}
136EXPORT_SYMBOL(memremap);
137
138void memunmap(void *addr)
139{
140 if (is_vmalloc_addr(addr))
141 iounmap((void __iomem *) addr);
142}
143EXPORT_SYMBOL(memunmap);
144
145static void devm_memremap_release(struct device *dev, void *res)
146{
147 memunmap(*(void **)res);
148}
149
150static int devm_memremap_match(struct device *dev, void *res, void *match_data)
151{
152 return *(void **)res == match_data;
153}
154
155void *devm_memremap(struct device *dev, resource_size_t offset,
156 size_t size, unsigned long flags)
157{
158 void **ptr, *addr;
159
160 ptr = devres_alloc_node(devm_memremap_release, sizeof(*ptr), GFP_KERNEL,
161 dev_to_node(dev));
162 if (!ptr)
163 return ERR_PTR(-ENOMEM);
164
165 addr = memremap(offset, size, flags);
166 if (addr) {
167 *ptr = addr;
168 devres_add(dev, ptr);
169 } else {
170 devres_free(ptr);
171 return ERR_PTR(-ENXIO);
172 }
173
174 return addr;
175}
176EXPORT_SYMBOL(devm_memremap);
177
178void devm_memunmap(struct device *dev, void *addr)
179{
180 WARN_ON(devres_release(dev, devm_memremap_release,
181 devm_memremap_match, addr));
182}
183EXPORT_SYMBOL(devm_memunmap);
184
185#ifdef CONFIG_ZONE_DEVICE
186static DEFINE_MUTEX(pgmap_lock);
187static RADIX_TREE(pgmap_radix, GFP_KERNEL);
188#define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
189#define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
190
191static unsigned long order_at(struct resource *res, unsigned long pgoff)
192{
193 unsigned long phys_pgoff = PHYS_PFN(res->start) + pgoff;
194 unsigned long nr_pages, mask;
195
196 nr_pages = PHYS_PFN(resource_size(res));
197 if (nr_pages == pgoff)
198 return ULONG_MAX;
199
200 /*
201 * What is the largest aligned power-of-2 range available from
202 * this resource pgoff to the end of the resource range,
203 * considering the alignment of the current pgoff?
204 */
205 mask = phys_pgoff | rounddown_pow_of_two(nr_pages - pgoff);
206 if (!mask)
207 return ULONG_MAX;
208
209 return find_first_bit(&mask, BITS_PER_LONG);
210}
211
212#define foreach_order_pgoff(res, order, pgoff) \
213 for (pgoff = 0, order = order_at((res), pgoff); order < ULONG_MAX; \
214 pgoff += 1UL << order, order = order_at((res), pgoff))
215
216#if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
217int device_private_entry_fault(struct vm_area_struct *vma,
218 unsigned long addr,
219 swp_entry_t entry,
220 unsigned int flags,
221 pmd_t *pmdp)
222{
223 struct page *page = device_private_entry_to_page(entry);
224
225 /*
226 * The page_fault() callback must migrate page back to system memory
227 * so that CPU can access it. This might fail for various reasons
228 * (device issue, device was unsafely unplugged, ...). When such
229 * error conditions happen, the callback must return VM_FAULT_SIGBUS.
230 *
231 * Note that because memory cgroup charges are accounted to the device
232 * memory, this should never fail because of memory restrictions (but
233 * allocation of regular system page might still fail because we are
234 * out of memory).
235 *
236 * There is a more in-depth description of what that callback can and
237 * cannot do, in include/linux/memremap.h
238 */
239 return page->pgmap->page_fault(vma, addr, page, flags, pmdp);
240}
241EXPORT_SYMBOL(device_private_entry_fault);
242#endif /* CONFIG_DEVICE_PRIVATE */
243
244static void pgmap_radix_release(struct resource *res, unsigned long end_pgoff)
245{
246 unsigned long pgoff, order;
247
248 mutex_lock(&pgmap_lock);
249 foreach_order_pgoff(res, order, pgoff) {
250 if (pgoff >= end_pgoff)
251 break;
252 radix_tree_delete(&pgmap_radix, PHYS_PFN(res->start) + pgoff);
253 }
254 mutex_unlock(&pgmap_lock);
255
256 synchronize_rcu();
257}
258
259static unsigned long pfn_first(struct dev_pagemap *pgmap)
260{
261 const struct resource *res = &pgmap->res;
262 struct vmem_altmap *altmap = &pgmap->altmap;
263 unsigned long pfn;
264
265 pfn = res->start >> PAGE_SHIFT;
266 if (pgmap->altmap_valid)
267 pfn += vmem_altmap_offset(altmap);
268 return pfn;
269}
270
271static unsigned long pfn_end(struct dev_pagemap *pgmap)
272{
273 const struct resource *res = &pgmap->res;
274
275 return (res->start + resource_size(res)) >> PAGE_SHIFT;
276}
277
278static unsigned long pfn_next(unsigned long pfn)
279{
280 if (pfn % 1024 == 0)
281 cond_resched();
282 return pfn + 1;
283}
284
285#define for_each_device_pfn(pfn, map) \
286 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
287
288static void devm_memremap_pages_release(void *data)
289{
290 struct dev_pagemap *pgmap = data;
291 struct device *dev = pgmap->dev;
292 struct resource *res = &pgmap->res;
293 resource_size_t align_start, align_size;
294 unsigned long pfn;
295
296 for_each_device_pfn(pfn, pgmap)
297 put_page(pfn_to_page(pfn));
298
299 if (percpu_ref_tryget_live(pgmap->ref)) {
300 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
301 percpu_ref_put(pgmap->ref);
302 }
303
304 /* pages are dead and unused, undo the arch mapping */
305 align_start = res->start & ~(SECTION_SIZE - 1);
306 align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
307 - align_start;
308
309 mem_hotplug_begin();
310 arch_remove_memory(align_start, align_size, pgmap->altmap_valid ?
311 &pgmap->altmap : NULL);
312 mem_hotplug_done();
313
314 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
315 pgmap_radix_release(res, -1);
316 dev_WARN_ONCE(dev, pgmap->altmap.alloc,
317 "%s: failed to free all reserved pages\n", __func__);
318}
319
320/**
321 * devm_memremap_pages - remap and provide memmap backing for the given resource
322 * @dev: hosting device for @res
323 * @pgmap: pointer to a struct dev_pgmap
324 *
325 * Notes:
326 * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
327 * by the caller before passing it to this function
328 *
329 * 2/ The altmap field may optionally be initialized, in which case altmap_valid
330 * must be set to true
331 *
332 * 3/ pgmap.ref must be 'live' on entry and 'dead' before devm_memunmap_pages()
333 * time (or devm release event). The expected order of events is that ref has
334 * been through percpu_ref_kill() before devm_memremap_pages_release(). The
335 * wait for the completion of all references being dropped and
336 * percpu_ref_exit() must occur after devm_memremap_pages_release().
337 *
338 * 4/ res is expected to be a host memory range that could feasibly be
339 * treated as a "System RAM" range, i.e. not a device mmio range, but
340 * this is not enforced.
341 */
342void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
343{
344 resource_size_t align_start, align_size, align_end;
345 struct vmem_altmap *altmap = pgmap->altmap_valid ?
346 &pgmap->altmap : NULL;
347 struct resource *res = &pgmap->res;
348 unsigned long pfn, pgoff, order;
349 pgprot_t pgprot = PAGE_KERNEL;
350 int error, nid, is_ram;
351
352 align_start = res->start & ~(SECTION_SIZE - 1);
353 align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
354 - align_start;
355 is_ram = region_intersects(align_start, align_size,
356 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
357
358 if (is_ram == REGION_MIXED) {
359 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
360 __func__, res);
361 return ERR_PTR(-ENXIO);
362 }
363
364 if (is_ram == REGION_INTERSECTS)
365 return __va(res->start);
366
367 if (!pgmap->ref)
368 return ERR_PTR(-EINVAL);
369
370 pgmap->dev = dev;
371
372 mutex_lock(&pgmap_lock);
373 error = 0;
374 align_end = align_start + align_size - 1;
375
376 foreach_order_pgoff(res, order, pgoff) {
377 error = __radix_tree_insert(&pgmap_radix,
378 PHYS_PFN(res->start) + pgoff, order, pgmap);
379 if (error) {
380 dev_err(dev, "%s: failed: %d\n", __func__, error);
381 break;
382 }
383 }
384 mutex_unlock(&pgmap_lock);
385 if (error)
386 goto err_radix;
387
388 nid = dev_to_node(dev);
389 if (nid < 0)
390 nid = numa_mem_id();
391
392 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
393 align_size);
394 if (error)
395 goto err_pfn_remap;
396
397 mem_hotplug_begin();
398 error = arch_add_memory(nid, align_start, align_size, altmap, false);
399 if (!error)
400 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
401 align_start >> PAGE_SHIFT,
402 align_size >> PAGE_SHIFT, altmap);
403 mem_hotplug_done();
404 if (error)
405 goto err_add_memory;
406
407 for_each_device_pfn(pfn, pgmap) {
408 struct page *page = pfn_to_page(pfn);
409
410 /*
411 * ZONE_DEVICE pages union ->lru with a ->pgmap back
412 * pointer. It is a bug if a ZONE_DEVICE page is ever
413 * freed or placed on a driver-private list. Seed the
414 * storage with LIST_POISON* values.
415 */
416 list_del(&page->lru);
417 page->pgmap = pgmap;
418 percpu_ref_get(pgmap->ref);
419 }
420
421 devm_add_action(dev, devm_memremap_pages_release, pgmap);
422
423 return __va(res->start);
424
425 err_add_memory:
426 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
427 err_pfn_remap:
428 err_radix:
429 pgmap_radix_release(res, pgoff);
430 return ERR_PTR(error);
431}
432EXPORT_SYMBOL(devm_memremap_pages);
433
434unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
435{
436 /* number of pfns from base where pfn_to_page() is valid */
437 return altmap->reserve + altmap->free;
438}
439
440void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
441{
442 altmap->alloc -= nr_pfns;
443}
444
445/**
446 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
447 * @pfn: page frame number to lookup page_map
448 * @pgmap: optional known pgmap that already has a reference
449 *
450 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
451 * is non-NULL but does not cover @pfn the reference to it will be released.
452 */
453struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
454 struct dev_pagemap *pgmap)
455{
456 resource_size_t phys = PFN_PHYS(pfn);
457
458 /*
459 * In the cached case we're already holding a live reference.
460 */
461 if (pgmap) {
462 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
463 return pgmap;
464 put_dev_pagemap(pgmap);
465 }
466
467 /* fall back to slow path lookup */
468 rcu_read_lock();
469 pgmap = radix_tree_lookup(&pgmap_radix, PHYS_PFN(phys));
470 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
471 pgmap = NULL;
472 rcu_read_unlock();
473
474 return pgmap;
475}
476#endif /* CONFIG_ZONE_DEVICE */
477
478#if IS_ENABLED(CONFIG_DEVICE_PRIVATE) || IS_ENABLED(CONFIG_DEVICE_PUBLIC)
479void put_zone_device_private_or_public_page(struct page *page)
480{
481 int count = page_ref_dec_return(page);
482
483 /*
484 * If refcount is 1 then page is freed and refcount is stable as nobody
485 * holds a reference on the page.
486 */
487 if (count == 1) {
488 /* Clear Active bit in case of parallel mark_page_accessed */
489 __ClearPageActive(page);
490 __ClearPageWaiters(page);
491
492 page->mapping = NULL;
493 mem_cgroup_uncharge(page);
494
495 page->pgmap->page_free(page, page->pgmap->data);
496 } else if (!count)
497 __put_page(page);
498}
499EXPORT_SYMBOL(put_zone_device_private_or_public_page);
500#endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */