xref: /linux-6.15/include/linux/memremap.h (revision bb7e5ce7)
1 #ifndef _LINUX_MEMREMAP_H_
2 #define _LINUX_MEMREMAP_H_
3 #include <linux/mm.h>
4 #include <linux/ioport.h>
5 #include <linux/percpu-refcount.h>
6 
7 #include <asm/pgtable.h>
8 
9 struct resource;
10 struct device;
11 
12 /**
13  * struct vmem_altmap - pre-allocated storage for vmemmap_populate
14  * @base_pfn: base of the entire dev_pagemap mapping
15  * @reserve: pages mapped, but reserved for driver use (relative to @base)
16  * @free: free pages set aside in the mapping for memmap storage
17  * @align: pages reserved to meet allocation alignments
18  * @alloc: track pages consumed, private to vmemmap_populate()
19  */
20 struct vmem_altmap {
21 	const unsigned long base_pfn;
22 	const unsigned long reserve;
23 	unsigned long free;
24 	unsigned long align;
25 	unsigned long alloc;
26 };
27 
28 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap);
29 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns);
30 
31 #ifdef CONFIG_ZONE_DEVICE
32 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start);
33 #else
34 static inline struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
35 {
36 	return NULL;
37 }
38 #endif
39 
40 /*
41  * Specialize ZONE_DEVICE memory into multiple types each having differents
42  * usage.
43  *
44  * MEMORY_DEVICE_HOST:
45  * Persistent device memory (pmem): struct page might be allocated in different
46  * memory and architecture might want to perform special actions. It is similar
47  * to regular memory, in that the CPU can access it transparently. However,
48  * it is likely to have different bandwidth and latency than regular memory.
49  * See Documentation/nvdimm/nvdimm.txt for more information.
50  *
51  * MEMORY_DEVICE_PRIVATE:
52  * Device memory that is not directly addressable by the CPU: CPU can neither
53  * read nor write private memory. In this case, we do still have struct pages
54  * backing the device memory. Doing so simplifies the implementation, but it is
55  * important to remember that there are certain points at which the struct page
56  * must be treated as an opaque object, rather than a "normal" struct page.
57  *
58  * A more complete discussion of unaddressable memory may be found in
59  * include/linux/hmm.h and Documentation/vm/hmm.txt.
60  *
61  * MEMORY_DEVICE_PUBLIC:
62  * Device memory that is cache coherent from device and CPU point of view. This
63  * is use on platform that have an advance system bus (like CAPI or CCIX). A
64  * driver can hotplug the device memory using ZONE_DEVICE and with that memory
65  * type. Any page of a process can be migrated to such memory. However no one
66  * should be allow to pin such memory so that it can always be evicted.
67  */
68 enum memory_type {
69 	MEMORY_DEVICE_HOST = 0,
70 	MEMORY_DEVICE_PRIVATE,
71 	MEMORY_DEVICE_PUBLIC,
72 };
73 
74 /*
75  * For MEMORY_DEVICE_PRIVATE we use ZONE_DEVICE and extend it with two
76  * callbacks:
77  *   page_fault()
78  *   page_free()
79  *
80  * Additional notes about MEMORY_DEVICE_PRIVATE may be found in
81  * include/linux/hmm.h and Documentation/vm/hmm.txt. There is also a brief
82  * explanation in include/linux/memory_hotplug.h.
83  *
84  * The page_fault() callback must migrate page back, from device memory to
85  * system memory, so that the CPU can access it. This might fail for various
86  * reasons (device issues,  device have been unplugged, ...). When such error
87  * conditions happen, the page_fault() callback must return VM_FAULT_SIGBUS and
88  * set the CPU page table entry to "poisoned".
89  *
90  * Note that because memory cgroup charges are transferred to the device memory,
91  * this should never fail due to memory restrictions. However, allocation
92  * of a regular system page might still fail because we are out of memory. If
93  * that happens, the page_fault() callback must return VM_FAULT_OOM.
94  *
95  * The page_fault() callback can also try to migrate back multiple pages in one
96  * chunk, as an optimization. It must, however, prioritize the faulting address
97  * over all the others.
98  *
99  *
100  * The page_free() callback is called once the page refcount reaches 1
101  * (ZONE_DEVICE pages never reach 0 refcount unless there is a refcount bug.
102  * This allows the device driver to implement its own memory management.)
103  *
104  * For MEMORY_DEVICE_PUBLIC only the page_free() callback matter.
105  */
106 typedef int (*dev_page_fault_t)(struct vm_area_struct *vma,
107 				unsigned long addr,
108 				const struct page *page,
109 				unsigned int flags,
110 				pmd_t *pmdp);
111 typedef void (*dev_page_free_t)(struct page *page, void *data);
112 
113 /**
114  * struct dev_pagemap - metadata for ZONE_DEVICE mappings
115  * @page_fault: callback when CPU fault on an unaddressable device page
116  * @page_free: free page callback when page refcount reaches 1
117  * @altmap: pre-allocated/reserved memory for vmemmap allocations
118  * @res: physical address range covered by @ref
119  * @ref: reference count that pins the devm_memremap_pages() mapping
120  * @dev: host device of the mapping for debug
121  * @data: private data pointer for page_free()
122  * @type: memory type: see MEMORY_* in memory_hotplug.h
123  */
124 struct dev_pagemap {
125 	dev_page_fault_t page_fault;
126 	dev_page_free_t page_free;
127 	struct vmem_altmap *altmap;
128 	const struct resource *res;
129 	struct percpu_ref *ref;
130 	struct device *dev;
131 	void *data;
132 	enum memory_type type;
133 };
134 
135 #ifdef CONFIG_ZONE_DEVICE
136 void *devm_memremap_pages(struct device *dev, struct resource *res,
137 		struct percpu_ref *ref, struct vmem_altmap *altmap);
138 struct dev_pagemap *find_dev_pagemap(resource_size_t phys);
139 
140 static inline bool is_zone_device_page(const struct page *page);
141 #else
142 static inline void *devm_memremap_pages(struct device *dev,
143 		struct resource *res, struct percpu_ref *ref,
144 		struct vmem_altmap *altmap)
145 {
146 	/*
147 	 * Fail attempts to call devm_memremap_pages() without
148 	 * ZONE_DEVICE support enabled, this requires callers to fall
149 	 * back to plain devm_memremap() based on config
150 	 */
151 	WARN_ON_ONCE(1);
152 	return ERR_PTR(-ENXIO);
153 }
154 
155 static inline struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
156 {
157 	return NULL;
158 }
159 #endif
160 
161 #if defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
162 static inline bool is_device_private_page(const struct page *page)
163 {
164 	return is_zone_device_page(page) &&
165 		page->pgmap->type == MEMORY_DEVICE_PRIVATE;
166 }
167 
168 static inline bool is_device_public_page(const struct page *page)
169 {
170 	return is_zone_device_page(page) &&
171 		page->pgmap->type == MEMORY_DEVICE_PUBLIC;
172 }
173 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
174 
175 /**
176  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
177  * @pfn: page frame number to lookup page_map
178  * @pgmap: optional known pgmap that already has a reference
179  *
180  * @pgmap allows the overhead of a lookup to be bypassed when @pfn lands in the
181  * same mapping.
182  */
183 static inline struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
184 		struct dev_pagemap *pgmap)
185 {
186 	const struct resource *res = pgmap ? pgmap->res : NULL;
187 	resource_size_t phys = PFN_PHYS(pfn);
188 
189 	/*
190 	 * In the cached case we're already holding a live reference so
191 	 * we can simply do a blind increment
192 	 */
193 	if (res && phys >= res->start && phys <= res->end) {
194 		percpu_ref_get(pgmap->ref);
195 		return pgmap;
196 	}
197 
198 	/* fall back to slow path lookup */
199 	rcu_read_lock();
200 	pgmap = find_dev_pagemap(phys);
201 	if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
202 		pgmap = NULL;
203 	rcu_read_unlock();
204 
205 	return pgmap;
206 }
207 
208 static inline void put_dev_pagemap(struct dev_pagemap *pgmap)
209 {
210 	if (pgmap)
211 		percpu_ref_put(pgmap->ref);
212 }
213 #endif /* _LINUX_MEMREMAP_H_ */
214