1 .. SPDX-License-Identifier: GPL-2.0
3 .. _physical_memory_model:
9 Physical memory in a system may be addressed in different ways. The
10 simplest case is when the physical memory starts at address 0 and
11 spans a contiguous range up to the maximal address. It could be,
12 however, that this range contains small holes that are not accessible
13 for the CPU. Then there could be several contiguous ranges at
14 completely distinct addresses. And, don't forget about NUMA, where
15 different memory banks are attached to different CPUs.
17 Linux abstracts this diversity using one of the three memory models:
18 FLATMEM, DISCONTIGMEM and SPARSEMEM. Each architecture defines what
19 memory models it supports, what the default memory model is and
20 whether it is possible to manually override that default.
23 At time of this writing, DISCONTIGMEM is considered deprecated,
24 although it is still in use by several architectures.
26 All the memory models track the status of physical page frames using
27 struct page arranged in one or more arrays.
29 Regardless of the selected memory model, there exists one-to-one
30 mapping between the physical page frame number (PFN) and the
31 corresponding `struct page`.
33 Each memory model defines :c:func:`pfn_to_page` and :c:func:`page_to_pfn`
34 helpers that allow the conversion from PFN to `struct page` and vice
40 The simplest memory model is FLATMEM. This model is suitable for
41 non-NUMA systems with contiguous, or mostly contiguous, physical
44 In the FLATMEM memory model, there is a global `mem_map` array that
45 maps the entire physical memory. For most architectures, the holes
46 have entries in the `mem_map` array. The `struct page` objects
47 corresponding to the holes are never fully initialized.
49 To allocate the `mem_map` array, architecture specific setup code should
50 call :c:func:`free_area_init` function. Yet, the mappings array is not
51 usable until the call to :c:func:`memblock_free_all` that hands all the
52 memory to the page allocator.
54 An architecture may free parts of the `mem_map` array that do not cover the
55 actual physical pages. In such case, the architecture specific
56 :c:func:`pfn_valid` implementation should take the holes in the
57 `mem_map` into account.
59 With FLATMEM, the conversion between a PFN and the `struct page` is
60 straightforward: `PFN - ARCH_PFN_OFFSET` is an index to the
63 The `ARCH_PFN_OFFSET` defines the first page frame number for
64 systems with physical memory starting at address different from 0.
69 The DISCONTIGMEM model treats the physical memory as a collection of
70 `nodes` similarly to how Linux NUMA support does. For each node Linux
71 constructs an independent memory management subsystem represented by
72 `struct pglist_data` (or `pg_data_t` for short). Among other
73 things, `pg_data_t` holds the `node_mem_map` array that maps
74 physical pages belonging to that node. The `node_start_pfn` field of
75 `pg_data_t` is the number of the first page frame belonging to that
78 The architecture setup code should call :c:func:`free_area_init_node` for
79 each node in the system to initialize the `pg_data_t` object and its
82 Every `node_mem_map` behaves exactly as FLATMEM's `mem_map` -
83 every physical page frame in a node has a `struct page` entry in the
84 `node_mem_map` array. When DISCONTIGMEM is enabled, a portion of the
85 `flags` field of the `struct page` encodes the node number of the
86 node hosting that page.
88 The conversion between a PFN and the `struct page` in the
89 DISCONTIGMEM model became slightly more complex as it has to determine
90 which node hosts the physical page and which `pg_data_t` object
91 holds the `struct page`.
93 Architectures that support DISCONTIGMEM provide :c:func:`pfn_to_nid`
94 to convert PFN to the node number. The opposite conversion helper
95 :c:func:`page_to_nid` is generic as it uses the node number encoded in
98 Once the node number is known, the PFN can be used to index
99 appropriate `node_mem_map` array to access the `struct page` and
100 the offset of the `struct page` from the `node_mem_map` plus
101 `node_start_pfn` is the PFN of that page.
106 SPARSEMEM is the most versatile memory model available in Linux and it
107 is the only memory model that supports several advanced features such
108 as hot-plug and hot-remove of the physical memory, alternative memory
109 maps for non-volatile memory devices and deferred initialization of
110 the memory map for larger systems.
112 The SPARSEMEM model presents the physical memory as a collection of
113 sections. A section is represented with struct mem_section
114 that contains `section_mem_map` that is, logically, a pointer to an
115 array of struct pages. However, it is stored with some other magic
116 that aids the sections management. The section size and maximal number
117 of section is specified using `SECTION_SIZE_BITS` and
118 `MAX_PHYSMEM_BITS` constants defined by each architecture that
119 supports SPARSEMEM. While `MAX_PHYSMEM_BITS` is an actual width of a
120 physical address that an architecture supports, the
121 `SECTION_SIZE_BITS` is an arbitrary value.
123 The maximal number of sections is denoted `NR_MEM_SECTIONS` and
128 NR\_MEM\_SECTIONS = 2 ^ {(MAX\_PHYSMEM\_BITS - SECTION\_SIZE\_BITS)}
130 The `mem_section` objects are arranged in a two-dimensional array
131 called `mem_sections`. The size and placement of this array depend
132 on `CONFIG_SPARSEMEM_EXTREME` and the maximal possible number of
135 * When `CONFIG_SPARSEMEM_EXTREME` is disabled, the `mem_sections`
136 array is static and has `NR_MEM_SECTIONS` rows. Each row holds a
137 single `mem_section` object.
138 * When `CONFIG_SPARSEMEM_EXTREME` is enabled, the `mem_sections`
139 array is dynamically allocated. Each row contains PAGE_SIZE worth of
140 `mem_section` objects and the number of rows is calculated to fit
141 all the memory sections.
143 The architecture setup code should call sparse_init() to
144 initialize the memory sections and the memory maps.
146 With SPARSEMEM there are two possible ways to convert a PFN to the
147 corresponding `struct page` - a "classic sparse" and "sparse
148 vmemmap". The selection is made at build time and it is determined by
149 the value of `CONFIG_SPARSEMEM_VMEMMAP`.
151 The classic sparse encodes the section number of a page in page->flags
152 and uses high bits of a PFN to access the section that maps that page
153 frame. Inside a section, the PFN is the index to the array of pages.
155 The sparse vmemmap uses a virtually mapped memory map to optimize
156 pfn_to_page and page_to_pfn operations. There is a global `struct
157 page *vmemmap` pointer that points to a virtually contiguous array of
158 `struct page` objects. A PFN is an index to that array and the
159 offset of the `struct page` from `vmemmap` is the PFN of that
162 To use vmemmap, an architecture has to reserve a range of virtual
163 addresses that will map the physical pages containing the memory
164 map and make sure that `vmemmap` points to that range. In addition,
165 the architecture should implement :c:func:`vmemmap_populate` method
166 that will allocate the physical memory and create page tables for the
167 virtual memory map. If an architecture does not have any special
168 requirements for the vmemmap mappings, it can use default
169 :c:func:`vmemmap_populate_basepages` provided by the generic memory
172 The virtually mapped memory map allows storing `struct page` objects
173 for persistent memory devices in pre-allocated storage on those
174 devices. This storage is represented with struct vmem_altmap
175 that is eventually passed to vmemmap_populate() through a long chain
176 of function calls. The vmemmap_populate() implementation may use the
177 `vmem_altmap` along with :c:func:`vmemmap_alloc_block_buf` helper to
178 allocate memory map on the persistent memory device.
182 The `ZONE_DEVICE` facility builds upon `SPARSEMEM_VMEMMAP` to offer
183 `struct page` `mem_map` services for device driver identified physical
184 address ranges. The "device" aspect of `ZONE_DEVICE` relates to the fact
185 that the page objects for these address ranges are never marked online,
186 and that a reference must be taken against the device, not just the page
187 to keep the memory pinned for active use. `ZONE_DEVICE`, via
188 :c:func:`devm_memremap_pages`, performs just enough memory hotplug to
189 turn on :c:func:`pfn_to_page`, :c:func:`page_to_pfn`, and
190 :c:func:`get_user_pages` service for the given range of pfns. Since the
191 page reference count never drops below 1 the page is never tracked as
192 free memory and the page's `struct list_head lru` space is repurposed
193 for back referencing to the host device / driver that mapped the memory.
195 While `SPARSEMEM` presents memory as a collection of sections,
196 optionally collected into memory blocks, `ZONE_DEVICE` users have a need
197 for smaller granularity of populating the `mem_map`. Given that
198 `ZONE_DEVICE` memory is never marked online it is subsequently never
199 subject to its memory ranges being exposed through the sysfs memory
200 hotplug api on memory block boundaries. The implementation relies on
201 this lack of user-api constraint to allow sub-section sized memory
202 ranges to be specified to :c:func:`arch_add_memory`, the top-half of
203 memory hotplug. Sub-section support allows for 2MB as the cross-arch
204 common alignment granularity for :c:func:`devm_memremap_pages`.
206 The users of `ZONE_DEVICE` are:
208 * pmem: Map platform persistent memory to be used as a direct-I/O target
211 * hmm: Extend `ZONE_DEVICE` with `->page_fault()` and `->page_free()`
212 event callbacks to allow a device-driver to coordinate memory management
213 events related to device-memory, typically GPU memory. See
214 Documentation/vm/hmm.rst.
216 * p2pdma: Create `struct page` objects to allow peer devices in a
217 PCI/-E topology to coordinate direct-DMA operations between themselves,
218 i.e. bypass host memory.