1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* Copyright(c) 2015 Intel Corporation. All rights reserved. */
3 #include <linux/device.h>
5 #include <linux/kasan.h>
6 #include <linux/memory_hotplug.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>
16 static DEFINE_XARRAY(pgmap_array);
17 #define SECTION_MASK ~((1UL << PA_SECTION_SHIFT) - 1)
18 #define SECTION_SIZE (1UL << PA_SECTION_SHIFT)
20 #if IS_ENABLED(CONFIG_DEVICE_PRIVATE)
21 vm_fault_t device_private_entry_fault(struct vm_area_struct *vma,
27 struct page *page = device_private_entry_to_page(entry);
28 struct hmm_devmem *devmem;
30 devmem = container_of(page->pgmap, typeof(*devmem), pagemap);
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.
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
43 * There is a more in-depth description of what that callback can and
44 * cannot do, in include/linux/memremap.h
46 return devmem->page_fault(vma, addr, page, flags, pmdp);
48 #endif /* CONFIG_DEVICE_PRIVATE */
50 static void pgmap_array_delete(struct resource *res)
52 xa_store_range(&pgmap_array, PHYS_PFN(res->start), PHYS_PFN(res->end),
57 static unsigned long pfn_first(struct dev_pagemap *pgmap)
59 const struct resource *res = &pgmap->res;
60 struct vmem_altmap *altmap = &pgmap->altmap;
63 pfn = res->start >> PAGE_SHIFT;
64 if (pgmap->altmap_valid)
65 pfn += vmem_altmap_offset(altmap);
69 static unsigned long pfn_end(struct dev_pagemap *pgmap)
71 const struct resource *res = &pgmap->res;
73 return (res->start + resource_size(res)) >> PAGE_SHIFT;
76 static unsigned long pfn_next(unsigned long pfn)
83 #define for_each_device_pfn(pfn, map) \
84 for (pfn = pfn_first(map); pfn < pfn_end(map); pfn = pfn_next(pfn))
86 static void devm_memremap_pages_release(void *data)
88 struct dev_pagemap *pgmap = data;
89 struct device *dev = pgmap->dev;
90 struct resource *res = &pgmap->res;
91 resource_size_t align_start, align_size;
95 pgmap->ops->kill(pgmap->ref);
96 for_each_device_pfn(pfn, pgmap)
97 put_page(pfn_to_page(pfn));
98 pgmap->ops->cleanup(pgmap->ref);
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)
105 nid = page_to_nid(pfn_to_page(align_start >> PAGE_SHIFT));
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);
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);
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__);
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
131 * 1/ At a minimum the res, ref and type and ops members of @pgmap must be
132 * initialized by the caller before passing it to this function
134 * 2/ The altmap field may optionally be initialized, in which case altmap_valid
135 * must be set to true
137 * 3/ pgmap->ref must be 'live' on entry and will be killed and reaped
138 * at devm_memremap_pages_release() time, or if this routine fails.
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.
144 void *devm_memremap_pages(struct device *dev, struct dev_pagemap *pgmap)
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 struct mhp_restrictions restrictions = {
153 * We do not want any optional features only our own memmap
157 pgprot_t pgprot = PAGE_KERNEL;
158 int error, nid, is_ram;
160 switch (pgmap->type) {
161 case MEMORY_DEVICE_PRIVATE:
162 if (!IS_ENABLED(CONFIG_DEVICE_PRIVATE)) {
163 WARN(1, "Device private memory not supported\n");
164 return ERR_PTR(-EINVAL);
167 case MEMORY_DEVICE_FS_DAX:
168 if (!IS_ENABLED(CONFIG_ZONE_DEVICE) ||
169 IS_ENABLED(CONFIG_FS_DAX_LIMITED)) {
170 WARN(1, "File system DAX not supported\n");
171 return ERR_PTR(-EINVAL);
174 case MEMORY_DEVICE_DEVDAX:
175 case MEMORY_DEVICE_PCI_P2PDMA:
178 WARN(1, "Invalid pgmap type %d\n", pgmap->type);
182 if (!pgmap->ref || !pgmap->ops || !pgmap->ops->kill ||
183 !pgmap->ops->cleanup) {
184 WARN(1, "Missing reference count teardown definition\n");
185 return ERR_PTR(-EINVAL);
188 align_start = res->start & ~(SECTION_SIZE - 1);
189 align_size = ALIGN(res->start + resource_size(res), SECTION_SIZE)
191 align_end = align_start + align_size - 1;
193 conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_start), NULL);
194 if (conflict_pgmap) {
195 dev_WARN(dev, "Conflicting mapping in same section\n");
196 put_dev_pagemap(conflict_pgmap);
201 conflict_pgmap = get_dev_pagemap(PHYS_PFN(align_end), NULL);
202 if (conflict_pgmap) {
203 dev_WARN(dev, "Conflicting mapping in same section\n");
204 put_dev_pagemap(conflict_pgmap);
209 is_ram = region_intersects(align_start, align_size,
210 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE);
212 if (is_ram != REGION_DISJOINT) {
213 WARN_ONCE(1, "%s attempted on %s region %pr\n", __func__,
214 is_ram == REGION_MIXED ? "mixed" : "ram", res);
221 error = xa_err(xa_store_range(&pgmap_array, PHYS_PFN(res->start),
222 PHYS_PFN(res->end), pgmap, GFP_KERNEL));
226 nid = dev_to_node(dev);
230 error = track_pfn_remap(NULL, &pgprot, PHYS_PFN(align_start), 0,
238 * For device private memory we call add_pages() as we only need to
239 * allocate and initialize struct page for the device memory. More-
240 * over the device memory is un-accessible thus we do not want to
241 * create a linear mapping for the memory like arch_add_memory()
244 * For all other device memory types, which are accessible by
245 * the CPU, we do want the linear mapping and thus use
248 if (pgmap->type == MEMORY_DEVICE_PRIVATE) {
249 error = add_pages(nid, align_start >> PAGE_SHIFT,
250 align_size >> PAGE_SHIFT, &restrictions);
252 error = kasan_add_zero_shadow(__va(align_start), align_size);
258 error = arch_add_memory(nid, align_start, align_size,
265 zone = &NODE_DATA(nid)->node_zones[ZONE_DEVICE];
266 move_pfn_range_to_zone(zone, align_start >> PAGE_SHIFT,
267 align_size >> PAGE_SHIFT, altmap);
275 * Initialization of the pages has been deferred until now in order
276 * to allow us to do the work while not holding the hotplug lock.
278 memmap_init_zone_device(&NODE_DATA(nid)->node_zones[ZONE_DEVICE],
279 align_start >> PAGE_SHIFT,
280 align_size >> PAGE_SHIFT, pgmap);
281 percpu_ref_get_many(pgmap->ref, pfn_end(pgmap) - pfn_first(pgmap));
283 error = devm_add_action_or_reset(dev, devm_memremap_pages_release,
286 return ERR_PTR(error);
288 return __va(res->start);
291 kasan_remove_zero_shadow(__va(align_start), align_size);
293 untrack_pfn(NULL, PHYS_PFN(align_start), align_size);
295 pgmap_array_delete(res);
297 pgmap->ops->kill(pgmap->ref);
298 pgmap->ops->cleanup(pgmap->ref);
299 return ERR_PTR(error);
301 EXPORT_SYMBOL_GPL(devm_memremap_pages);
303 void devm_memunmap_pages(struct device *dev, struct dev_pagemap *pgmap)
305 devm_release_action(dev, devm_memremap_pages_release, pgmap);
307 EXPORT_SYMBOL_GPL(devm_memunmap_pages);
309 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap)
311 /* number of pfns from base where pfn_to_page() is valid */
312 return altmap->reserve + altmap->free;
315 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns)
317 altmap->alloc -= nr_pfns;
321 * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
322 * @pfn: page frame number to lookup page_map
323 * @pgmap: optional known pgmap that already has a reference
325 * If @pgmap is non-NULL and covers @pfn it will be returned as-is. If @pgmap
326 * is non-NULL but does not cover @pfn the reference to it will be released.
328 struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
329 struct dev_pagemap *pgmap)
331 resource_size_t phys = PFN_PHYS(pfn);
334 * In the cached case we're already holding a live reference.
337 if (phys >= pgmap->res.start && phys <= pgmap->res.end)
339 put_dev_pagemap(pgmap);
342 /* fall back to slow path lookup */
344 pgmap = xa_load(&pgmap_array, PHYS_PFN(phys));
345 if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
351 EXPORT_SYMBOL_GPL(get_dev_pagemap);
353 #ifdef CONFIG_DEV_PAGEMAP_OPS
354 DEFINE_STATIC_KEY_FALSE(devmap_managed_key);
355 EXPORT_SYMBOL(devmap_managed_key);
356 static atomic_t devmap_enable;
359 * Toggle the static key for ->page_free() callbacks when dev_pagemap
362 void dev_pagemap_get_ops(void)
364 if (atomic_inc_return(&devmap_enable) == 1)
365 static_branch_enable(&devmap_managed_key);
367 EXPORT_SYMBOL_GPL(dev_pagemap_get_ops);
369 void dev_pagemap_put_ops(void)
371 if (atomic_dec_and_test(&devmap_enable))
372 static_branch_disable(&devmap_managed_key);
374 EXPORT_SYMBOL_GPL(dev_pagemap_put_ops);
376 void __put_devmap_managed_page(struct page *page)
378 int count = page_ref_dec_return(page);
381 * If refcount is 1 then page is freed and refcount is stable as nobody
382 * holds a reference on the page.
385 /* Clear Active bit in case of parallel mark_page_accessed */
386 __ClearPageActive(page);
387 __ClearPageWaiters(page);
389 mem_cgroup_uncharge(page);
391 page->pgmap->ops->page_free(page, page->pgmap->data);
395 EXPORT_SYMBOL(__put_devmap_managed_page);
396 #endif /* CONFIG_DEV_PAGEMAP_OPS */