1 .. SPDX-License-Identifier: GPL-2.0
2
3 =====================
4 Physical Memory Model
5 =====================
6
7 Physical memory in a system may be addressed in different ways. The
8 simplest case is when the physical memory starts at address 0 and
9 spans a contiguous range up to the maximal address. It could be,
10 however, that this range contains small holes that are not accessible
11 for the CPU. Then there could be several contiguous ranges at
12 completely distinct addresses. And, don't forget about NUMA, where
13 different memory banks are attached to different CPUs.
14
15 Linux abstracts this diversity using one of the two memory models:
16 FLATMEM and SPARSEMEM. Each architecture defines what
17 memory models it supports, what the default memory model is and
18 whether it is possible to manually override that default.
19
20 All the memory models track the status of physical page frames using
21 struct page arranged in one or more arrays.
22
23 Regardless of the selected memory model, there exists one-to-one
24 mapping between the physical page frame number (PFN) and the
25 corresponding `struct page`.
26
27 Each memory model defines :c:func:`pfn_to_page` and :c:func:`page_to_pfn`
28 helpers that allow the conversion from PFN to `struct page` and vice
29 versa.
30
31 FLATMEM
32 =======
33
34 The simplest memory model is FLATMEM. This model is suitable for
35 non-NUMA systems with contiguous, or mostly contiguous, physical
36 memory.
37
38 In the FLATMEM memory model, there is a global `mem_map` array that
39 maps the entire physical memory. For most architectures, the holes
40 have entries in the `mem_map` array. The `struct page` objects
41 corresponding to the holes are never fully initialized.
42
43 To allocate the `mem_map` array, architecture specific setup code should
44 call :c:func:`free_area_init` function. Yet, the mappings array is not
45 usable until the call to :c:func:`memblock_free_all` that hands all the
46 memory to the page allocator.
47
48 An architecture may free parts of the `mem_map` array that do not cover the
49 actual physical pages. In such case, the architecture specific
50 :c:func:`pfn_valid` implementation should take the holes in the
51 `mem_map` into account.
52
53 With FLATMEM, the conversion between a PFN and the `struct page` is
54 straightforward: `PFN - ARCH_PFN_OFFSET` is an index to the
55 `mem_map` array.
56
57 The `ARCH_PFN_OFFSET` defines the first page frame number for
58 systems with physical memory starting at address different from 0.
59
60 SPARSEMEM
61 =========
62
63 SPARSEMEM is the most versatile memory model available in Linux and it
64 is the only memory model that supports several advanced features such
65 as hot-plug and hot-remove of the physical memory, alternative memory
66 maps for non-volatile memory devices and deferred initialization of
67 the memory map for larger systems.
68
69 The SPARSEMEM model presents the physical memory as a collection of
70 sections. A section is represented with struct mem_section
71 that contains `section_mem_map` that is, logically, a pointer to an
72 array of struct pages. However, it is stored with some other magic
73 that aids the sections management. The section size and maximal number
74 of section is specified using `SECTION_SIZE_BITS` and
75 `MAX_PHYSMEM_BITS` constants defined by each architecture that
76 supports SPARSEMEM. While `MAX_PHYSMEM_BITS` is an actual width of a
77 physical address that an architecture supports, the
78 `SECTION_SIZE_BITS` is an arbitrary value.
79
80 The maximal number of sections is denoted `NR_MEM_SECTIONS` and
81 defined as
82
83 .. math::
84
85 NR\_MEM\_SECTIONS = 2 ^ {(MAX\_PHYSMEM\_BITS - SECTION\_SIZE\_BITS)}
86
87 The `mem_section` objects are arranged in a two-dimensional array
88 called `mem_sections`. The size and placement of this array depend
89 on `CONFIG_SPARSEMEM_EXTREME` and the maximal possible number of
90 sections:
91
92 * When `CONFIG_SPARSEMEM_EXTREME` is disabled, the `mem_sections`
93 array is static and has `NR_MEM_SECTIONS` rows. Each row holds a
94 single `mem_section` object.
95 * When `CONFIG_SPARSEMEM_EXTREME` is enabled, the `mem_sections`
96 array is dynamically allocated. Each row contains PAGE_SIZE worth of
97 `mem_section` objects and the number of rows is calculated to fit
98 all the memory sections.
99
100 The architecture setup code should call sparse_init() to
101 initialize the memory sections and the memory maps.
102
103 With SPARSEMEM there are two possible ways to convert a PFN to the
104 corresponding `struct page` - a "classic sparse" and "sparse
105 vmemmap". The selection is made at build time and it is determined by
106 the value of `CONFIG_SPARSEMEM_VMEMMAP`.
107
108 The classic sparse encodes the section number of a page in page->flags
109 and uses high bits of a PFN to access the section that maps that page
110 frame. Inside a section, the PFN is the index to the array of pages.
111
112 The sparse vmemmap uses a virtually mapped memory map to optimize
113 pfn_to_page and page_to_pfn operations. There is a global `struct
114 page *vmemmap` pointer that points to a virtually contiguous array of
115 `struct page` objects. A PFN is an index to that array and the
116 offset of the `struct page` from `vmemmap` is the PFN of that
117 page.
118
119 To use vmemmap, an architecture has to reserve a range of virtual
120 addresses that will map the physical pages containing the memory
121 map and make sure that `vmemmap` points to that range. In addition,
122 the architecture should implement :c:func:`vmemmap_populate` method
123 that will allocate the physical memory and create page tables for the
124 virtual memory map. If an architecture does not have any special
125 requirements for the vmemmap mappings, it can use default
126 :c:func:`vmemmap_populate_basepages` provided by the generic memory
127 management.
128
129 The virtually mapped memory map allows storing `struct page` objects
130 for persistent memory devices in pre-allocated storage on those
131 devices. This storage is represented with struct vmem_altmap
132 that is eventually passed to vmemmap_populate() through a long chain
133 of function calls. The vmemmap_populate() implementation may use the
134 `vmem_altmap` along with :c:func:`vmemmap_alloc_block_buf` helper to
135 allocate memory map on the persistent memory device.
136
137 ZONE_DEVICE
138 ===========
139 The `ZONE_DEVICE` facility builds upon `SPARSEMEM_VMEMMAP` to offer
140 `struct page` `mem_map` services for device driver identified physical
141 address ranges. The "device" aspect of `ZONE_DEVICE` relates to the fact
142 that the page objects for these address ranges are never marked online,
143 and that a reference must be taken against the device, not just the page
144 to keep the memory pinned for active use. `ZONE_DEVICE`, via
145 :c:func:`devm_memremap_pages`, performs just enough memory hotplug to
146 turn on :c:func:`pfn_to_page`, :c:func:`page_to_pfn`, and
147 :c:func:`get_user_pages` service for the given range of pfns. Since the
148 page reference count never drops below 1 the page is never tracked as
149 free memory and the page's `struct list_head lru` space is repurposed
150 for back referencing to the host device / driver that mapped the memory.
151
152 While `SPARSEMEM` presents memory as a collection of sections,
153 optionally collected into memory blocks, `ZONE_DEVICE` users have a need
154 for smaller granularity of populating the `mem_map`. Given that
155 `ZONE_DEVICE` memory is never marked online it is subsequently never
156 subject to its memory ranges being exposed through the sysfs memory
157 hotplug api on memory block boundaries. The implementation relies on
158 this lack of user-api constraint to allow sub-section sized memory
159 ranges to be specified to :c:func:`arch_add_memory`, the top-half of
160 memory hotplug. Sub-section support allows for 2MB as the cross-arch
161 common alignment granularity for :c:func:`devm_memremap_pages`.
162
163 The users of `ZONE_DEVICE` are:
164
165 * pmem: Map platform persistent memory to be used as a direct-I/O target
166 via DAX mappings.
167
168 * hmm: Extend `ZONE_DEVICE` with `->page_fault()` and `->page_free()`
169 event callbacks to allow a device-driver to coordinate memory management
170 events related to device-memory, typically GPU memory. See
171 Documentation/mm/hmm.rst.
172
173 * p2pdma: Create `struct page` objects to allow peer devices in a
174 PCI/-E topology to coordinate direct-DMA operations between themselves,
175 i.e. bypass host memory.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.