1.. _highmem: 2 3==================== 4High Memory Handling 5==================== 6 7By: Peter Zijlstra <a.p.zijlstra@chello.nl> 8 9.. contents:: :local: 10 11What Is High Memory? 12==================== 13 14High memory (highmem) is used when the size of physical memory approaches or 15exceeds the maximum size of virtual memory. At that point it becomes 16impossible for the kernel to keep all of the available physical memory mapped 17at all times. This means the kernel needs to start using temporary mappings of 18the pieces of physical memory that it wants to access. 19 20The part of (physical) memory not covered by a permanent mapping is what we 21refer to as 'highmem'. There are various architecture dependent constraints on 22where exactly that border lies. 23 24In the i386 arch, for example, we choose to map the kernel into every process's 25VM space so that we don't have to pay the full TLB invalidation costs for 26kernel entry/exit. This means the available virtual memory space (4GiB on 27i386) has to be divided between user and kernel space. 28 29The traditional split for architectures using this approach is 3:1, 3GiB for 30userspace and the top 1GiB for kernel space:: 31 32 +--------+ 0xffffffff 33 | Kernel | 34 +--------+ 0xc0000000 35 | | 36 | User | 37 | | 38 +--------+ 0x00000000 39 40This means that the kernel can at most map 1GiB of physical memory at any one 41time, but because we need virtual address space for other things - including 42temporary maps to access the rest of the physical memory - the actual direct 43map will typically be less (usually around ~896MiB). 44 45Other architectures that have mm context tagged TLBs can have separate kernel 46and user maps. Some hardware (like some ARMs), however, have limited virtual 47space when they use mm context tags. 48 49 50Temporary Virtual Mappings 51========================== 52 53The kernel contains several ways of creating temporary mappings. The following 54list shows them in order of preference of use. 55 56* kmap_local_page(). This function is used to require short term mappings. 57 It can be invoked from any context (including interrupts) but the mappings 58 can only be used in the context which acquired them. 59 60 This function should be preferred, where feasible, over all the others. 61 62 These mappings are thread-local and CPU-local, meaning that the mapping 63 can only be accessed from within this thread and the thread is bound to the 64 CPU while the mapping is active. Although preemption is never disabled by 65 this function, the CPU can not be unplugged from the system via 66 CPU-hotplug until the mapping is disposed. 67 68 It's valid to take pagefaults in a local kmap region, unless the context 69 in which the local mapping is acquired does not allow it for other reasons. 70 71 As said, pagefaults and preemption are never disabled. There is no need to 72 disable preemption because, when context switches to a different task, the 73 maps of the outgoing task are saved and those of the incoming one are 74 restored. 75 76 kmap_local_page() always returns a valid virtual address and it is assumed 77 that kunmap_local() will never fail. 78 79 On CONFIG_HIGHMEM=n kernels and for low memory pages this returns the 80 virtual address of the direct mapping. Only real highmem pages are 81 temporarily mapped. Therefore, users may call a plain page_address() 82 for pages which are known to not come from ZONE_HIGHMEM. However, it is 83 always safe to use kmap_local_page() / kunmap_local(). 84 85 While it is significantly faster than kmap(), for the higmem case it 86 comes with restrictions about the pointers validity. Contrary to kmap() 87 mappings, the local mappings are only valid in the context of the caller 88 and cannot be handed to other contexts. This implies that users must 89 be absolutely sure to keep the use of the return address local to the 90 thread which mapped it. 91 92 Most code can be designed to use thread local mappings. User should 93 therefore try to design their code to avoid the use of kmap() by mapping 94 pages in the same thread the address will be used and prefer 95 kmap_local_page(). 96 97 Nesting kmap_local_page() and kmap_atomic() mappings is allowed to a certain 98 extent (up to KMAP_TYPE_NR) but their invocations have to be strictly ordered 99 because the map implementation is stack based. See kmap_local_page() kdocs 100 (included in the "Functions" section) for details on how to manage nested 101 mappings. 102 103* kmap_atomic(). This permits a very short duration mapping of a single 104 page. Since the mapping is restricted to the CPU that issued it, it 105 performs well, but the issuing task is therefore required to stay on that 106 CPU until it has finished, lest some other task displace its mappings. 107 108 kmap_atomic() may also be used by interrupt contexts, since it does not 109 sleep and the callers too may not sleep until after kunmap_atomic() is 110 called. 111 112 Each call of kmap_atomic() in the kernel creates a non-preemptible section 113 and disable pagefaults. This could be a source of unwanted latency. Therefore 114 users should prefer kmap_local_page() instead of kmap_atomic(). 115 116 It is assumed that k[un]map_atomic() won't fail. 117 118* kmap(). This should be used to make short duration mapping of a single 119 page with no restrictions on preemption or migration. It comes with an 120 overhead as mapping space is restricted and protected by a global lock 121 for synchronization. When mapping is no longer needed, the address that 122 the page was mapped to must be released with kunmap(). 123 124 Mapping changes must be propagated across all the CPUs. kmap() also 125 requires global TLB invalidation when the kmap's pool wraps and it might 126 block when the mapping space is fully utilized until a slot becomes 127 available. Therefore, kmap() is only callable from preemptible context. 128 129 All the above work is necessary if a mapping must last for a relatively 130 long time but the bulk of high-memory mappings in the kernel are 131 short-lived and only used in one place. This means that the cost of 132 kmap() is mostly wasted in such cases. kmap() was not intended for long 133 term mappings but it has morphed in that direction and its use is 134 strongly discouraged in newer code and the set of the preceding functions 135 should be preferred. 136 137 On 64-bit systems, calls to kmap_local_page(), kmap_atomic() and kmap() have 138 no real work to do because a 64-bit address space is more than sufficient to 139 address all the physical memory whose pages are permanently mapped. 140 141* vmap(). This can be used to make a long duration mapping of multiple 142 physical pages into a contiguous virtual space. It needs global 143 synchronization to unmap. 144 145 146Cost of Temporary Mappings 147========================== 148 149The cost of creating temporary mappings can be quite high. The arch has to 150manipulate the kernel's page tables, the data TLB and/or the MMU's registers. 151 152If CONFIG_HIGHMEM is not set, then the kernel will try and create a mapping 153simply with a bit of arithmetic that will convert the page struct address into 154a pointer to the page contents rather than juggling mappings about. In such a 155case, the unmap operation may be a null operation. 156 157If CONFIG_MMU is not set, then there can be no temporary mappings and no 158highmem. In such a case, the arithmetic approach will also be used. 159 160 161i386 PAE 162======== 163 164The i386 arch, under some circumstances, will permit you to stick up to 64GiB 165of RAM into your 32-bit machine. This has a number of consequences: 166 167* Linux needs a page-frame structure for each page in the system and the 168 pageframes need to live in the permanent mapping, which means: 169 170* you can have 896M/sizeof(struct page) page-frames at most; with struct 171 page being 32-bytes that would end up being something in the order of 112G 172 worth of pages; the kernel, however, needs to store more than just 173 page-frames in that memory... 174 175* PAE makes your page tables larger - which slows the system down as more 176 data has to be accessed to traverse in TLB fills and the like. One 177 advantage is that PAE has more PTE bits and can provide advanced features 178 like NX and PAT. 179 180The general recommendation is that you don't use more than 8GiB on a 32-bit 181machine - although more might work for you and your workload, you're pretty 182much on your own - don't expect kernel developers to really care much if things 183come apart. 184 185 186Functions 187========= 188 189.. kernel-doc:: include/linux/highmem.h 190.. kernel-doc:: include/linux/highmem-internal.h 191