1 .. SPDX-License-Identifier: GPL-2.0
7 Paged virtual memory was invented along with virtual memory as a concept in
8 1962 on the Ferranti Atlas Computer which was the first computer with paged
9 virtual memory. The feature migrated to newer computers and became a de facto
10 feature of all Unix-like systems as time went by. In 1985 the feature was
11 included in the Intel 80386, which was the CPU Linux 1.0 was developed on.
13 Page tables map virtual addresses as seen by the CPU into physical addresses
14 as seen on the external memory bus.
16 Linux defines page tables as a hierarchy which is currently five levels in
17 height. The architecture code for each supported architecture will then
18 map this to the restrictions of the hardware.
20 The physical address corresponding to the virtual address is often referenced
21 by the underlying physical page frame. The **page frame number** or **pfn**
22 is the physical address of the page (as seen on the external memory bus)
23 divided by `PAGE_SIZE`.
25 Physical memory address 0 will be *pfn 0* and the highest pfn will be
26 the last page of physical memory the external address bus of the CPU can
29 With a page granularity of 4KB and a address range of 32 bits, pfn 0 is at
30 address 0x00000000, pfn 1 is at address 0x00001000, pfn 2 is at 0x00002000
31 and so on until we reach pfn 0xfffff at 0xfffff000. With 16KB pages pfs are
32 at 0x00004000, 0x00008000 ... 0xffffc000 and pfn goes from 0 to 0x3fffff.
34 As you can see, with 4KB pages the page base address uses bits 12-31 of the
35 address, and this is why `PAGE_SHIFT` in this case is defined as 12 and
36 `PAGE_SIZE` is usually defined in terms of the page shift as `(1 << PAGE_SHIFT)`
38 Over time a deeper hierarchy has been developed in response to increasing memory
39 sizes. When Linux was created, 4KB pages and a single page table called
40 `swapper_pg_dir` with 1024 entries was used, covering 4MB which coincided with
41 the fact that Torvald's first computer had 4MB of physical memory. Entries in
42 this single table were referred to as *PTE*:s - page table entries.
44 The software page table hierarchy reflects the fact that page table hardware has
45 become hierarchical and that in turn is done to save page table memory and
48 One could of course imagine a single, linear page table with enormous amounts
49 of entries, breaking down the whole memory into single pages. Such a page table
50 would be very sparse, because large portions of the virtual memory usually
51 remains unused. By using hierarchical page tables large holes in the virtual
52 address space does not waste valuable page table memory, because it will suffice
53 to mark large areas as unmapped at a higher level in the page table hierarchy.
55 Additionally, on modern CPUs, a higher level page table entry can point directly
56 to a physical memory range, which allows mapping a contiguous range of several
57 megabytes or even gigabytes in a single high-level page table entry, taking
58 shortcuts in mapping virtual memory to physical memory: there is no need to
59 traverse deeper in the hierarchy when you find a large mapped range like this.
61 The page table hierarchy has now developed into this::
84 Symbols on the different levels of the page table hierarchy have the following
85 meaning beginning from the bottom:
87 - **pte**, `pte_t`, `pteval_t` = **Page Table Entry** - mentioned earlier.
88 The *pte* is an array of `PTRS_PER_PTE` elements of the `pteval_t` type, each
89 mapping a single page of virtual memory to a single page of physical memory.
90 The architecture defines the size and contents of `pteval_t`.
92 A typical example is that the `pteval_t` is a 32- or 64-bit value with the
93 upper bits being a **pfn** (page frame number), and the lower bits being some
94 architecture-specific bits such as memory protection.
96 The **entry** part of the name is a bit confusing because while in Linux 1.0
97 this did refer to a single page table entry in the single top level page
98 table, it was retrofitted to be an array of mapping elements when two-level
99 page tables were first introduced, so the *pte* is the lowermost page
100 *table*, not a page table *entry*.
102 - **pmd**, `pmd_t`, `pmdval_t` = **Page Middle Directory**, the hierarchy right
103 above the *pte*, with `PTRS_PER_PMD` references to the *pte*:s.
105 - **pud**, `pud_t`, `pudval_t` = **Page Upper Directory** was introduced after
106 the other levels to handle 4-level page tables. It is potentially unused,
107 or *folded* as we will discuss later.
109 - **p4d**, `p4d_t`, `p4dval_t` = **Page Level 4 Directory** was introduced to
110 handle 5-level page tables after the *pud* was introduced. Now it was clear
111 that we needed to replace *pgd*, *pmd*, *pud* etc with a figure indicating the
112 directory level and that we cannot go on with ad hoc names any more. This
113 is only used on systems which actually have 5 levels of page tables, otherwise
116 - **pgd**, `pgd_t`, `pgdval_t` = **Page Global Directory** - the Linux kernel
117 main page table handling the PGD for the kernel memory is still found in
118 `swapper_pg_dir`, but each userspace process in the system also has its own
119 memory context and thus its own *pgd*, found in `struct mm_struct` which
120 in turn is referenced to in each `struct task_struct`. So tasks have memory
121 context in the form of a `struct mm_struct` and this in turn has a
122 `struct pgt_t *pgd` pointer to the corresponding page global directory.
124 To repeat: each level in the page table hierarchy is a *array of pointers*, so
125 the **pgd** contains `PTRS_PER_PGD` pointers to the next level below, **p4d**
126 contains `PTRS_PER_P4D` pointers to **pud** items and so on. The number of
127 pointers on each level is architecture-defined.::
131 | ptr |-------> +-----+
132 | ptr |- | ptr |-------> PAGE
137 +-----+ +----> +-----+
146 If the architecture does not use all the page table levels, they can be *folded*
147 which means skipped, and all operations performed on page tables will be
148 compile-time augmented to just skip a level when accessing the next lower
151 Page table handling code that wishes to be architecture-neutral, such as the
152 virtual memory manager, will need to be written so that it traverses all of the
153 currently five levels. This style should also be preferred for
154 architecture-specific code, so as to be robust to future changes.
157 MMU, TLB, and Page Faults
158 =========================
160 The `Memory Management Unit (MMU)` is a hardware component that handles virtual
161 to physical address translations. It may use relatively small caches in hardware
162 called `Translation Lookaside Buffers (TLBs)` and `Page Walk Caches` to speed up
165 When CPU accesses a memory location, it provides a virtual address to the MMU,
166 which checks if there is the existing translation in the TLB or in the Page
167 Walk Caches (on architectures that support them). If no translation is found,
168 MMU uses the page walks to determine the physical address and create the map.
170 The dirty bit for a page is set (i.e., turned on) when the page is written to.
171 Each page of memory has associated permission and dirty bits. The latter
172 indicate that the page has been modified since it was loaded into memory.
174 If nothing prevents it, eventually the physical memory can be accessed and the
175 requested operation on the physical frame is performed.
177 There are several reasons why the MMU can't find certain translations. It could
178 happen because the CPU is trying to access memory that the current task is not
179 permitted to, or because the data is not present into physical memory.
181 When these conditions happen, the MMU triggers page faults, which are types of
182 exceptions that signal the CPU to pause the current execution and run a special
183 function to handle the mentioned exceptions.
185 There are common and expected causes of page faults. These are triggered by
186 process management optimization techniques called "Lazy Allocation" and
187 "Copy-on-Write". Page faults may also happen when frames have been swapped out
188 to persistent storage (swap partition or file) and evicted from their physical
191 These techniques improve memory efficiency, reduce latency, and minimize space
192 occupation. This document won't go deeper into the details of "Lazy Allocation"
193 and "Copy-on-Write" because these subjects are out of scope as they belong to
194 Process Address Management.
196 Swapping differentiates itself from the other mentioned techniques because it's
197 undesirable since it's performed as a means to reduce memory under heavy
200 Swapping can't work for memory mapped by kernel logical addresses. These are a
201 subset of the kernel virtual space that directly maps a contiguous range of
202 physical memory. Given any logical address, its physical address is determined
203 with simple arithmetic on an offset. Accesses to logical addresses are fast
204 because they avoid the need for complex page table lookups at the expenses of
205 frames not being evictable and pageable out.
207 If the kernel fails to make room for the data that must be present in the
208 physical frames, the kernel invokes the out-of-memory (OOM) killer to make room
209 by terminating lower priority processes until pressure reduces under a safe
212 Additionally, page faults may be also caused by code bugs or by maliciously
213 crafted addresses that the CPU is instructed to access. A thread of a process
214 could use instructions to address (non-shared) memory which does not belong to
215 its own address space, or could try to execute an instruction that want to write
216 to a read-only location.
218 If the above-mentioned conditions happen in user-space, the kernel sends a
219 `Segmentation Fault` (SIGSEGV) signal to the current thread. That signal usually
220 causes the termination of the thread and of the process it belongs to.
222 This document is going to simplify and show an high altitude view of how the
223 Linux kernel handles these page faults, creates tables and tables' entries,
224 check if memory is present and, if not, requests to load data from persistent
225 storage or from other devices, and updates the MMU and its caches.
227 The first steps are architecture dependent. Most architectures jump to
228 `do_page_fault()`, whereas the x86 interrupt handler is defined by the
229 `DEFINE_IDTENTRY_RAW_ERRORCODE()` macro which calls `handle_page_fault()`.
231 Whatever the routes, all architectures end up to the invocation of
232 `handle_mm_fault()` which, in turn, (likely) ends up calling
233 `__handle_mm_fault()` to carry out the actual work of allocating the page
236 The unfortunate case of not being able to call `__handle_mm_fault()` means
237 that the virtual address is pointing to areas of physical memory which are not
238 permitted to be accessed (at least from the current context). This
239 condition resolves to the kernel sending the above-mentioned SIGSEGV signal
240 to the process and leads to the consequences already explained.
242 `__handle_mm_fault()` carries out its work by calling several functions to
243 find the entry's offsets of the upper layers of the page tables and allocate
244 the tables that it may need.
246 The functions that look for the offset have names like `*_offset()`, where the
247 "*" is for pgd, p4d, pud, pmd, pte; instead the functions to allocate the
248 corresponding tables, layer by layer, are called `*_alloc`, using the
249 above-mentioned convention to name them after the corresponding types of tables
252 The page table walk may end at one of the middle or upper layers (PMD, PUD).
254 Linux supports larger page sizes than the usual 4KB (i.e., the so called
255 `huge pages`). When using these kinds of larger pages, higher level pages can
256 directly map them, with no need to use lower level page entries (PTE). Huge
257 pages contain large contiguous physical regions that usually span from 2MB to
258 1GB. They are respectively mapped by the PMD and PUD page entries.
260 The huge pages bring with them several benefits like reduced TLB pressure,
261 reduced page table overhead, memory allocation efficiency, and performance
262 improvement for certain workloads. However, these benefits come with
263 trade-offs, like wasted memory and allocation challenges.
265 At the very end of the walk with allocations, if it didn't return errors,
266 `__handle_mm_fault()` finally calls `handle_pte_fault()`, which via `do_fault()`
267 performs one of `do_read_fault()`, `do_cow_fault()`, `do_shared_fault()`.
268 "read", "cow", "shared" give hints about the reasons and the kind of fault it's
271 The actual implementation of the workflow is very complex. Its design allows
272 Linux to handle page faults in a way that is tailored to the specific
273 characteristics of each architecture, while still sharing a common overall
276 To conclude this high altitude view of how Linux handles page faults, let's
277 add that the page faults handler can be disabled and enabled respectively with
278 `pagefault_disable()` and `pagefault_enable()`.
280 Several code path make use of the latter two functions because they need to
281 disable traps into the page faults handler, mostly to prevent deadlocks.