5 By: Peter Zijlstra <a.p.zijlstra@chello.nl>
12 High memory (highmem) is used when the size of physical memory approaches or
13 exceeds the maximum size of virtual memory. At that point it becomes
14 impossible for the kernel to keep all of the available physical memory mapped
15 at all times. This means the kernel needs to start using temporary mappings of
16 the pieces of physical memory that it wants to access.
18 The part of (physical) memory not covered by a permanent mapping is what we
19 refer to as 'highmem'. There are various architecture dependent constraints on
20 where exactly that border lies.
22 In the i386 arch, for example, we choose to map the kernel into every process's
23 VM space so that we don't have to pay the full TLB invalidation costs for
24 kernel entry/exit. This means the available virtual memory space (4GiB on
25 i386) has to be divided between user and kernel space.
27 The traditional split for architectures using this approach is 3:1, 3GiB for
28 userspace and the top 1GiB for kernel space::
38 This means that the kernel can at most map 1GiB of physical memory at any one
39 time, but because we need virtual address space for other things - including
40 temporary maps to access the rest of the physical memory - the actual direct
41 map will typically be less (usually around ~896MiB).
43 Other architectures that have mm context tagged TLBs can have separate kernel
44 and user maps. Some hardware (like some ARMs), however, have limited virtual
45 space when they use mm context tags.
48 Temporary Virtual Mappings
49 ==========================
51 The kernel contains several ways of creating temporary mappings. The following
52 list shows them in order of preference of use.
54 * kmap_local_page(), kmap_local_folio() - These functions are used to create
55 short term mappings. They can be invoked from any context (including
56 interrupts) but the mappings can only be used in the context which acquired
57 them. The only differences between them consist in the first taking a pointer
58 to a struct page and the second taking a pointer to struct folio and the byte
59 offset within the folio which identifies the page.
61 These functions should always be used, whereas kmap_atomic() and kmap() have
64 These mappings are thread-local and CPU-local, meaning that the mapping
65 can only be accessed from within this thread and the thread is bound to the
66 CPU while the mapping is active. Although preemption is never disabled by
67 this function, the CPU can not be unplugged from the system via
68 CPU-hotplug until the mapping is disposed.
70 It's valid to take pagefaults in a local kmap region, unless the context
71 in which the local mapping is acquired does not allow it for other reasons.
73 As said, pagefaults and preemption are never disabled. There is no need to
74 disable preemption because, when context switches to a different task, the
75 maps of the outgoing task are saved and those of the incoming one are
78 kmap_local_page(), as well as kmap_local_folio() always returns valid virtual
79 kernel addresses and it is assumed that kunmap_local() will never fail.
81 On CONFIG_HIGHMEM=n kernels and for low memory pages they return the
82 virtual address of the direct mapping. Only real highmem pages are
83 temporarily mapped. Therefore, users may call a plain page_address()
84 for pages which are known to not come from ZONE_HIGHMEM. However, it is
85 always safe to use kmap_local_{page,folio}() / kunmap_local().
87 While they are significantly faster than kmap(), for the highmem case they
88 come with restrictions about the pointers validity. Contrary to kmap()
89 mappings, the local mappings are only valid in the context of the caller
90 and cannot be handed to other contexts. This implies that users must
91 be absolutely sure to keep the use of the return address local to the
92 thread which mapped it.
94 Most code can be designed to use thread local mappings. User should
95 therefore try to design their code to avoid the use of kmap() by mapping
96 pages in the same thread the address will be used and prefer
97 kmap_local_page() or kmap_local_folio().
99 Nesting kmap_local_page() and kmap_atomic() mappings is allowed to a certain
100 extent (up to KMAP_TYPE_NR) but their invocations have to be strictly ordered
101 because the map implementation is stack based. See kmap_local_page() kdocs
102 (included in the "Functions" section) for details on how to manage nested
105 * kmap_atomic(). This function has been deprecated; use kmap_local_page().
107 NOTE: Conversions to kmap_local_page() must take care to follow the mapping
108 restrictions imposed on kmap_local_page(). Furthermore, the code between
109 calls to kmap_atomic() and kunmap_atomic() may implicitly depend on the side
110 effects of atomic mappings, i.e. disabling page faults or preemption, or both.
111 In that case, explicit calls to pagefault_disable() or preempt_disable() or
112 both must be made in conjunction with the use of kmap_local_page().
114 [Legacy documentation]
116 This permits a very short duration mapping of a single page. Since the
117 mapping is restricted to the CPU that issued it, it performs well, but
118 the issuing task is therefore required to stay on that CPU until it has
119 finished, lest some other task displace its mappings.
121 kmap_atomic() may also be used by interrupt contexts, since it does not
122 sleep and the callers too may not sleep until after kunmap_atomic() is
125 Each call of kmap_atomic() in the kernel creates a non-preemptible section
126 and disable pagefaults. This could be a source of unwanted latency. Therefore
127 users should prefer kmap_local_page() instead of kmap_atomic().
129 It is assumed that k[un]map_atomic() won't fail.
131 * kmap(). This function has been deprecated; use kmap_local_page().
133 NOTE: Conversions to kmap_local_page() must take care to follow the mapping
134 restrictions imposed on kmap_local_page(). In particular, it is necessary to
135 make sure that the kernel virtual memory pointer is only valid in the thread
138 [Legacy documentation]
140 This should be used to make short duration mapping of a single page with no
141 restrictions on preemption or migration. It comes with an overhead as mapping
142 space is restricted and protected by a global lock for synchronization. When
143 mapping is no longer needed, the address that the page was mapped to must be
144 released with kunmap().
146 Mapping changes must be propagated across all the CPUs. kmap() also
147 requires global TLB invalidation when the kmap's pool wraps and it might
148 block when the mapping space is fully utilized until a slot becomes
149 available. Therefore, kmap() is only callable from preemptible context.
151 All the above work is necessary if a mapping must last for a relatively
152 long time but the bulk of high-memory mappings in the kernel are
153 short-lived and only used in one place. This means that the cost of
154 kmap() is mostly wasted in such cases. kmap() was not intended for long
155 term mappings but it has morphed in that direction and its use is
156 strongly discouraged in newer code and the set of the preceding functions
159 On 64-bit systems, calls to kmap_local_page(), kmap_atomic() and kmap() have
160 no real work to do because a 64-bit address space is more than sufficient to
161 address all the physical memory whose pages are permanently mapped.
163 * vmap(). This can be used to make a long duration mapping of multiple
164 physical pages into a contiguous virtual space. It needs global
165 synchronization to unmap.
168 Cost of Temporary Mappings
169 ==========================
171 The cost of creating temporary mappings can be quite high. The arch has to
172 manipulate the kernel's page tables, the data TLB and/or the MMU's registers.
174 If CONFIG_HIGHMEM is not set, then the kernel will try and create a mapping
175 simply with a bit of arithmetic that will convert the page struct address into
176 a pointer to the page contents rather than juggling mappings about. In such a
177 case, the unmap operation may be a null operation.
179 If CONFIG_MMU is not set, then there can be no temporary mappings and no
180 highmem. In such a case, the arithmetic approach will also be used.
186 The i386 arch, under some circumstances, will permit you to stick up to 64GiB
187 of RAM into your 32-bit machine. This has a number of consequences:
189 * Linux needs a page-frame structure for each page in the system and the
190 pageframes need to live in the permanent mapping, which means:
192 * you can have 896M/sizeof(struct page) page-frames at most; with struct
193 page being 32-bytes that would end up being something in the order of 112G
194 worth of pages; the kernel, however, needs to store more than just
195 page-frames in that memory...
197 * PAE makes your page tables larger - which slows the system down as more
198 data has to be accessed to traverse in TLB fills and the like. One
199 advantage is that PAE has more PTE bits and can provide advanced features
202 The general recommendation is that you don't use more than 8GiB on a 32-bit
203 machine - although more might work for you and your workload, you're pretty
204 much on your own - don't expect kernel developers to really care much if things
211 .. kernel-doc:: include/linux/highmem.h
212 .. kernel-doc:: mm/highmem.c
213 .. kernel-doc:: include/linux/highmem-internal.h