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