1/* SPDX-License-Identifier: GPL-2.0 */
2/* Copyright(c) 2015 Intel Corporation. All rights reserved. */
3#include <linux/radix-tree.h>
4#include <linux/device.h>
5#include <linux/types.h>
6#include <linux/pfn_t.h>
7#include <linux/io.h>
8#include <linux/kasan.h>
9#include <linux/mm.h>
10#include <linux/memory_hotplug.h>
11#include <linux/swap.h>
12#include <linux/swapops.h>
13#include <linux/wait_bit.h>
14
15static DEFINE_MUTEX(pgmap_lock);
16static RADIX_TREE(pgmap_radix, GFP_KERNEL);
17#define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
18#define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
19
20static unsigned long order_at(struct resource *res, unsigned long pgoff)
21{
22 unsigned long phys_pgoff = PHYS_PFN(res->start) + pgoff;
23 unsigned long nr_pages, mask;
24
25 nr_pages = PHYS_PFN(resource_size(res));
26 if (nr_pages == pgoff)
27 return ULONG_MAX;
28
29 /*
30 * What is the largest aligned power-of-2 range available from
31 * this resource pgoff to the end of the resource range,
32 * considering the alignment of the current pgoff?
33 */
34 mask = phys_pgoff | rounddown_pow_of_two(nr_pages - pgoff);
35 if (!mask)
36 return ULONG_MAX;
37
38 return find_first_bit(&mask, BITS_PER_LONG);
39}
40
41#define foreach_order_pgoff(res, order, pgoff) \
42 for (pgoff = 0, order = order_at((res), pgoff); order < ULONG_MAX; \
43 pgoff += 1UL << order, order = order_at((res), pgoff))
44
45#if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
46int device_private_entry_fault(struct vm_area_struct *vma,
47 unsigned long addr,
48 swp_entry_t entry,
49 unsigned int flags,
50 pmd_t *pmdp)
51{
52 struct page *page = device_private_entry_to_page(entry);
53
54 /*
55 * The page_fault() callback must migrate page back to system memory
56 * so that CPU can access it. This might fail for various reasons
57 * (device issue, device was unsafely unplugged, ...). When such
58 * error conditions happen, the callback must return VM_FAULT_SIGBUS.
59 *
60 * Note that because memory cgroup charges are accounted to the device
61 * memory, this should never fail because of memory restrictions (but
62 * allocation of regular system page might still fail because we are
63 * out of memory).
64 *
65 * There is a more in-depth description of what that callback can and
66 * cannot do, in include/linux/memremap.h
67 */
68 return page->pgmap->page_fault(vma, addr, page, flags, pmdp);
69}
70EXPORT_SYMBOL(device_private_entry_fault);
71#endif /* CONFIG_DEVICE_PRIVATE */
72
73static void pgmap_radix_release(struct resource *res, unsigned long end_pgoff)
74{
75 unsigned long pgoff, order;
76
77 mutex_lock(&pgmap_lock);
78 foreach_order_pgoff(res, order, pgoff) {
79 if (pgoff >= end_pgoff)
80 break;
81 radix_tree_delete(&pgmap_radix, PHYS_PFN(res->start) + pgoff);
82 }
83 mutex_unlock(&pgmap_lock);
84
85 synchronize_rcu();
86}
87
88static unsigned long pfn_first(struct dev_pagemap *pgmap)
89{
90 const struct resource *res = &pgmap->res;
91 struct vmem_altmap *altmap = &pgmap->altmap;
92 unsigned long pfn;
93
94 pfn = res->start >> PAGE_SHIFT;
95 if (pgmap->altmap_valid)
96 pfn += vmem_altmap_offset(altmap);
97 return pfn;
98}
99
100static unsigned long pfn_end(struct dev_pagemap *pgmap)
101{
102 const struct resource *res = &pgmap->res;
103
104 return (res->start + resource_size(res)) >> PAGE_SHIFT;
105}
106
107static unsigned long pfn_next(unsigned long pfn)
108{
109 if (pfn % 1024 == 0)
110 cond_resched();
111 return pfn + 1;
112}
113
114#define for_each_device_pfn(pfn, map) \
115 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
116
117static void devm_memremap_pages_release(void *data)
118{
119 struct dev_pagemap *pgmap = data;
120 struct device *dev = pgmap->dev;
121 struct resource *res = &pgmap->res;
122 resource_size_t align_start, align_size;
123 unsigned long pfn;
124
125 for_each_device_pfn(pfn, pgmap)
126 put_page(pfn_to_page(pfn));
127
128 if (percpu_ref_tryget_live(pgmap->ref)) {
129 dev_WARN(dev, "%s: page mapping is still live!\n", __func__);
130 percpu_ref_put(pgmap->ref);
131 }
132
133 /* pages are dead and unused, undo the arch mapping */
134 align_start = res->start & ~(SECTION_SIZE - 1);
135 align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
136 - align_start;
137
138 mem_hotplug_begin();
139 arch_remove_memory(align_start, align_size, pgmap->altmap_valid ?
140 &pgmap->altmap : NULL);
141 kasan_remove_zero_shadow(__va(align_start), align_size);
142 mem_hotplug_done();
143
144 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
145 pgmap_radix_release(res, -1);
146 dev_WARN_ONCE(dev, pgmap->altmap.alloc,
147 "%s: failed to free all reserved pages\n", __func__);
148}
149
150/**
151 * devm_memremap_pages - remap and provide memmap backing for the given resource
152 * @dev: hosting device for @res
153 * @pgmap: pointer to a struct dev_pgmap
154 *
155 * Notes:
156 * 1/ At a minimum the res, ref and type members of @pgmap must be initialized
157 * by the caller before passing it to this function
158 *
159 * 2/ The altmap field may optionally be initialized, in which case altmap_valid
160 * must be set to true
161 *
162 * 3/ pgmap.ref must be 'live' on entry and 'dead' before devm_memunmap_pages()
163 * time (or devm release event). The expected order of events is that ref has
164 * been through percpu_ref_kill() before devm_memremap_pages_release(). The
165 * wait for the completion of all references being dropped and
166 * percpu_ref_exit() must occur after devm_memremap_pages_release().
167 *
168 * 4/ res is expected to be a host memory range that could feasibly be
169 * treated as a "System RAM" range, i.e. not a device mmio range, but
170 * this is not enforced.
171 */
172void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
173{
174 resource_size_t align_start, align_size, align_end;
175 struct vmem_altmap *altmap = pgmap->altmap_valid ?
176 &pgmap->altmap : NULL;
177 struct resource *res = &pgmap->res;
178 unsigned long pfn, pgoff, order;
179 pgprot_t pgprot = PAGE_KERNEL;
180 int error, nid, is_ram;
181 struct dev_pagemap *conflict_pgmap;
182
183 align_start = res->start & ~(SECTION_SIZE - 1);
184 align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
185 - align_start;
186 align_end = align_start + align_size - 1;
187
188 conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
189 if (conflict_pgmap) {
190 dev_WARN(dev, "Conflicting mapping in same section\n");
191 put_dev_pagemap(conflict_pgmap);
192 return ERR_PTR(-ENOMEM);
193 }
194
195 conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
196 if (conflict_pgmap) {
197 dev_WARN(dev, "Conflicting mapping in same section\n");
198 put_dev_pagemap(conflict_pgmap);
199 return ERR_PTR(-ENOMEM);
200 }
201
202 is_ram = region_intersects(align_start, align_size,
203 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
204
205 if (is_ram == REGION_MIXED) {
206 WARN_ONCE(1, "%s attempted on mixed region %pr\n",
207 __func__, res);
208 return ERR_PTR(-ENXIO);
209 }
210
211 if (is_ram == REGION_INTERSECTS)
212 return __va(res->start);
213
214 if (!pgmap->ref)
215 return ERR_PTR(-EINVAL);
216
217 pgmap->dev = dev;
218
219 mutex_lock(&pgmap_lock);
220 error = 0;
221
222 foreach_order_pgoff(res, order, pgoff) {
223 error = __radix_tree_insert(&pgmap_radix,
224 PHYS_PFN(res->start) + pgoff, order, pgmap);
225 if (error) {
226 dev_err(dev, "%s: failed: %d\n", __func__, error);
227 break;
228 }
229 }
230 mutex_unlock(&pgmap_lock);
231 if (error)
232 goto err_radix;
233
234 nid = dev_to_node(dev);
235 if (nid < 0)
236 nid = numa_mem_id();
237
238 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
239 align_size);
240 if (error)
241 goto err_pfn_remap;
242
243 mem_hotplug_begin();
244 error = kasan_add_zero_shadow(__va(align_start), align_size);
245 if (error) {
246 mem_hotplug_done();
247 goto err_kasan;
248 }
249
250 error = arch_add_memory(nid, align_start, align_size, altmap, false);
251 if (!error)
252 move_pfn_range_to_zone(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
253 align_start >> PAGE_SHIFT,
254 align_size >> PAGE_SHIFT, altmap);
255 mem_hotplug_done();
256 if (error)
257 goto err_add_memory;
258
259 for_each_device_pfn(pfn, pgmap) {
260 struct page *page = pfn_to_page(pfn);
261
262 /*
263 * ZONE_DEVICE pages union ->lru with a ->pgmap back
264 * pointer. It is a bug if a ZONE_DEVICE page is ever
265 * freed or placed on a driver-private list. Seed the
266 * storage with LIST_POISON* values.
267 */
268 list_del(&page->lru);
269 page->pgmap = pgmap;
270 percpu_ref_get(pgmap->ref);
271 }
272
273 devm_add_action(dev, devm_memremap_pages_release, pgmap);
274
275 return __va(res->start);
276
277 err_add_memory:
278 kasan_remove_zero_shadow(__va(align_start), align_size);
279 err_kasan:
280 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
281 err_pfn_remap:
282 err_radix:
283 pgmap_radix_release(res, pgoff);
284 return ERR_PTR(error);
285}
286EXPORT_SYMBOL(devm_memremap_pages);
287
288unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
289{
290 /* number of pfns from base where pfn_to_page() is valid */
291 return altmap->reserve + altmap->free;
292}
293
294void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
295{
296 altmap->alloc -= nr_pfns;
297}
298
299/**
300 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
301 * @pfn: page frame number to lookup page_map
302 * @pgmap: optional known pgmap that already has a reference
303 *
304 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
305 * is non-NULL but does not cover @pfn the reference to it will be released.
306 */
307struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
308 struct dev_pagemap *pgmap)
309{
310 resource_size_t phys = PFN_PHYS(pfn);
311
312 /*
313 * In the cached case we're already holding a live reference.
314 */
315 if (pgmap) {
316 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
317 return pgmap;
318 put_dev_pagemap(pgmap);
319 }
320
321 /* fall back to slow path lookup */
322 rcu_read_lock();
323 pgmap = radix_tree_lookup(&pgmap_radix, PHYS_PFN(phys));
324 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
325 pgmap = NULL;
326 rcu_read_unlock();
327
328 return pgmap;
329}
330EXPORT_SYMBOL_GPL(get_dev_pagemap);
331
332#ifdef CONFIG_DEV_PAGEMAP_OPS
333DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
334EXPORT_SYMBOL(devmap_managed_key);
335static atomic_t devmap_enable;
336
337/*
338 * Toggle the static key for ->page_free() callbacks when dev_pagemap
339 * pages go idle.
340 */
341void dev_pagemap_get_ops(void)
342{
343 if (atomic_inc_return(&devmap_enable) == 1)
344 static_branch_enable(&devmap_managed_key);
345}
346EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
347
348void dev_pagemap_put_ops(void)
349{
350 if (atomic_dec_and_test(&devmap_enable))
351 static_branch_disable(&devmap_managed_key);
352}
353EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
354
355void __put_devmap_managed_page(struct page *page)
356{
357 int count = page_ref_dec_return(page);
358
359 /*
360 * If refcount is 1 then page is freed and refcount is stable as nobody
361 * holds a reference on the page.
362 */
363 if (count == 1) {
364 /* Clear Active bit in case of parallel mark_page_accessed */
365 __ClearPageActive(page);
366 __ClearPageWaiters(page);
367
368 page->mapping = NULL;
369 mem_cgroup_uncharge(page);
370
371 page->pgmap->page_free(page, page->pgmap->data);
372 } else if (!count)
373 __put_page(page);
374}
375EXPORT_SYMBOL(__put_devmap_managed_page);
376#endif /* CONFIG_DEV_PAGEMAP_OPS */
377