include/linux/memremap.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _LINUX_MEMREMAP_H_
   3 #define _LINUX_MEMREMAP_H_
   4 #include <linux/mm.h>
   5 #include <linux/ioport.h>
   6 #include <linux/percpu-refcount.h>
   7
   8 #include <asm/pgtable.h>
   9
  10 struct resource;
  11 struct device;
  12
  13 /**
  14  * struct vmem_altmap - pre-allocated storage for vmemmap_populate
  15  * @base_pfn: base of the entire dev_pagemap mapping
  16  * @reserve: pages mapped, but reserved for driver use (relative to @base)
  17  * @free: free pages set aside in the mapping for memmap storage
  18  * @align: pages reserved to meet allocation alignments
  19  * @alloc: track pages consumed, private to vmemmap_populate()
  20  */
  21 struct vmem_altmap {
  22         const unsigned long base_pfn;
  23         const unsigned long reserve;
  24         unsigned long free;
  25         unsigned long align;
  26         unsigned long alloc;
  27 };
  28
  29 unsigned long vmem_altmap_offset(struct vmem_altmap *altmap);
  30 void vmem_altmap_free(struct vmem_altmap *altmap, unsigned long nr_pfns);
  31
  32 #ifdef CONFIG_ZONE_DEVICE
  33 struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start);
  34 #else
  35 static inline struct vmem_altmap *to_vmem_altmap(unsigned long memmap_start)
  36 {
  37         return NULL;
  38 }
  39 #endif
  40
  41 /*
  42  * Specialize ZONE_DEVICE memory into multiple types each having differents
  43  * usage.
  44  *
  45  * MEMORY_DEVICE_HOST:
  46  * Persistent device memory (pmem): struct page might be allocated in different
  47  * memory and architecture might want to perform special actions. It is similar
  48  * to regular memory, in that the CPU can access it transparently. However,
  49  * it is likely to have different bandwidth and latency than regular memory.
  50  * See Documentation/nvdimm/nvdimm.txt for more information.
  51  *
  52  * MEMORY_DEVICE_PRIVATE:
  53  * Device memory that is not directly addressable by the CPU: CPU can neither
  54  * read nor write private memory. In this case, we do still have struct pages
  55  * backing the device memory. Doing so simplifies the implementation, but it is
  56  * important to remember that there are certain points at which the struct page
  57  * must be treated as an opaque object, rather than a "normal" struct page.
  58  *
  59  * A more complete discussion of unaddressable memory may be found in
  60  * include/linux/hmm.h and Documentation/vm/hmm.txt.
  61  *
  62  * MEMORY_DEVICE_PUBLIC:
  63  * Device memory that is cache coherent from device and CPU point of view. This
  64  * is use on platform that have an advance system bus (like CAPI or CCIX). A
  65  * driver can hotplug the device memory using ZONE_DEVICE and with that memory
  66  * type. Any page of a process can be migrated to such memory. However no one
  67  * should be allow to pin such memory so that it can always be evicted.
  68  */
  69 enum memory_type {
  70         MEMORY_DEVICE_HOST = 0,
  71         MEMORY_DEVICE_PRIVATE,
  72         MEMORY_DEVICE_PUBLIC,
  73 };
  74
  75 /*
  76  * For MEMORY_DEVICE_PRIVATE we use ZONE_DEVICE and extend it with two
  77  * callbacks:
  78  *   page_fault()
  79  *   page_free()
  80  *
  81  * Additional notes about MEMORY_DEVICE_PRIVATE may be found in
  82  * include/linux/hmm.h and Documentation/vm/hmm.txt. There is also a brief
  83  * explanation in include/linux/memory_hotplug.h.
  84  *
  85  * The page_fault() callback must migrate page back, from device memory to
  86  * system memory, so that the CPU can access it. This might fail for various
  87  * reasons (device issues,  device have been unplugged, ...). When such error
  88  * conditions happen, the page_fault() callback must return VM_FAULT_SIGBUS and
  89  * set the CPU page table entry to "poisoned".
  90  *
  91  * Note that because memory cgroup charges are transferred to the device memory,
  92  * this should never fail due to memory restrictions. However, allocation
  93  * of a regular system page might still fail because we are out of memory. If
  94  * that happens, the page_fault() callback must return VM_FAULT_OOM.
  95  *
  96  * The page_fault() callback can also try to migrate back multiple pages in one
  97  * chunk, as an optimization. It must, however, prioritize the faulting address
  98  * over all the others.
  99  *
 100  *
 101  * The page_free() callback is called once the page refcount reaches 1
 102  * (ZONE_DEVICE pages never reach 0 refcount unless there is a refcount bug.
 103  * This allows the device driver to implement its own memory management.)
 104  *
 105  * For MEMORY_DEVICE_PUBLIC only the page_free() callback matter.
 106  */
 107 typedef int (*dev_page_fault_t)(struct vm_area_struct *vma,
 108                                 unsigned long addr,
 109                                 const struct page *page,
 110                                 unsigned int flags,
 111                                 pmd_t *pmdp);
 112 typedef void (*dev_page_free_t)(struct page *page, void *data);
 113
 114 /**
 115  * struct dev_pagemap - metadata for ZONE_DEVICE mappings
 116  * @page_fault: callback when CPU fault on an unaddressable device page
 117  * @page_free: free page callback when page refcount reaches 1
 118  * @altmap: pre-allocated/reserved memory for vmemmap allocations
 119  * @res: physical address range covered by @ref
 120  * @ref: reference count that pins the devm_memremap_pages() mapping
 121  * @dev: host device of the mapping for debug
 122  * @data: private data pointer for page_free()
 123  * @type: memory type: see MEMORY_* in memory_hotplug.h
 124  */
 125 struct dev_pagemap {
 126         dev_page_fault_t page_fault;
 127         dev_page_free_t page_free;
 128         struct vmem_altmap *altmap;
 129         const struct resource *res;
 130         struct percpu_ref *ref;
 131         struct device *dev;
 132         void *data;
 133         enum memory_type type;
 134 };
 135
 136 #ifdef CONFIG_ZONE_DEVICE
 137 void *devm_memremap_pages(struct device *dev, struct resource *res,
 138                 struct percpu_ref *ref, struct vmem_altmap *altmap);
 139 struct dev_pagemap *find_dev_pagemap(resource_size_t phys);
 140
 141 static inline bool is_zone_device_page(const struct page *page);
 142 #else
 143 static inline void *devm_memremap_pages(struct device *dev,
 144                 struct resource *res, struct percpu_ref *ref,
 145                 struct vmem_altmap *altmap)
 146 {
 147         /*
 148          * Fail attempts to call devm_memremap_pages() without
 149          * ZONE_DEVICE support enabled, this requires callers to fall
 150          * back to plain devm_memremap() based on config
 151          */
 152         WARN_ON_ONCE(1);
 153         return ERR_PTR(-ENXIO);
 154 }
 155
 156 static inline struct dev_pagemap *find_dev_pagemap(resource_size_t phys)
 157 {
 158         return NULL;
 159 }
 160 #endif
 161
 162 #if defined(CONFIG_DEVICE_PRIVATE) || defined(CONFIG_DEVICE_PUBLIC)
 163 static inline bool is_device_private_page(const struct page *page)
 164 {
 165         return is_zone_device_page(page) &&
 166                 page->pgmap->type == MEMORY_DEVICE_PRIVATE;
 167 }
 168
 169 static inline bool is_device_public_page(const struct page *page)
 170 {
 171         return is_zone_device_page(page) &&
 172                 page->pgmap->type == MEMORY_DEVICE_PUBLIC;
 173 }
 174 #endif /* CONFIG_DEVICE_PRIVATE || CONFIG_DEVICE_PUBLIC */
 175
 176 /**
 177  * get_dev_pagemap() - take a new live reference on the dev_pagemap for @pfn
 178  * @pfn: page frame number to lookup page_map
 179  * @pgmap: optional known pgmap that already has a reference
 180  *
 181  * @pgmap allows the overhead of a lookup to be bypassed when @pfn lands in the
 182  * same mapping.
 183  */
 184 static inline struct dev_pagemap *get_dev_pagemap(unsigned long pfn,
 185                 struct dev_pagemap *pgmap)
 186 {
 187         const struct resource *res = pgmap ? pgmap->res : NULL;
 188         resource_size_t phys = PFN_PHYS(pfn);
 189
 190         /*
 191          * In the cached case we're already holding a live reference so
 192          * we can simply do a blind increment
 193          */
 194         if (res && phys >= res->start && phys <= res->end) {
 195                 percpu_ref_get(pgmap->ref);
 196                 return pgmap;
 197         }
 198
 199         /* fall back to slow path lookup */
 200         rcu_read_lock();
 201         pgmap = find_dev_pagemap(phys);
 202         if (pgmap && !percpu_ref_tryget_live(pgmap->ref))
 203                 pgmap = NULL;
 204         rcu_read_unlock();
 205
 206         return pgmap;
 207 }
 208
 209 static inline void put_dev_pagemap(struct dev_pagemap *pgmap)
 210 {
 211         if (pgmap)
 212                 percpu_ref_put(pgmap->ref);
 213 }
 214 #endif /* _LINUX_MEMREMAP_H_ */