1 2 ==================== 3 HIGH MEMORY HANDLING 4 ==================== 5 6By: Peter Zijlstra <a.p.zijlstra@chello.nl> 7 8Contents: 9 10 (*) What is high memory? 11 12 (*) Temporary virtual mappings. 13 14 (*) Using kmap_atomic. 15 16 (*) Cost of temporary mappings. 17 18 (*) i386 PAE. 19 20 21==================== 22WHAT IS HIGH MEMORY? 23==================== 24 25High memory (highmem) is used when the size of physical memory approaches or 26exceeds the maximum size of virtual memory. At that point it becomes 27impossible for the kernel to keep all of the available physical memory mapped 28at all times. This means the kernel needs to start using temporary mappings of 29the pieces of physical memory that it wants to access. 30 31The part of (physical) memory not covered by a permanent mapping is what we 32refer to as 'highmem'. There are various architecture dependent constraints on 33where exactly that border lies. 34 35In the i386 arch, for example, we choose to map the kernel into every process's 36VM space so that we don't have to pay the full TLB invalidation costs for 37kernel entry/exit. This means the available virtual memory space (4GiB on 38i386) has to be divided between user and kernel space. 39 40The traditional split for architectures using this approach is 3:1, 3GiB for 41userspace and the top 1GiB for kernel space: 42 43 +--------+ 0xffffffff 44 | Kernel | 45 +--------+ 0xc0000000 46 | | 47 | User | 48 | | 49 +--------+ 0x00000000 50 51This means that the kernel can at most map 1GiB of physical memory at any one 52time, but because we need virtual address space for other things - including 53temporary maps to access the rest of the physical memory - the actual direct 54map will typically be less (usually around ~896MiB). 55 56Other architectures that have mm context tagged TLBs can have separate kernel 57and user maps. Some hardware (like some ARMs), however, have limited virtual 58space when they use mm context tags. 59 60 61========================== 62TEMPORARY VIRTUAL MAPPINGS 63========================== 64 65The kernel contains several ways of creating temporary mappings: 66 67 (*) vmap(). This can be used to make a long duration mapping of multiple 68 physical pages into a contiguous virtual space. It needs global 69 synchronization to unmap. 70 71 (*) kmap(). This permits a short duration mapping of a single page. It needs 72 global synchronization, but is amortized somewhat. It is also prone to 73 deadlocks when using in a nested fashion, and so it is not recommended for 74 new code. 75 76 (*) kmap_atomic(). This permits a very short duration mapping of a single 77 page. Since the mapping is restricted to the CPU that issued it, it 78 performs well, but the issuing task is therefore required to stay on that 79 CPU until it has finished, lest some other task displace its mappings. 80 81 kmap_atomic() may also be used by interrupt contexts, since it is does not 82 sleep and the caller may not sleep until after kunmap_atomic() is called. 83 84 It may be assumed that k[un]map_atomic() won't fail. 85 86 87================= 88USING KMAP_ATOMIC 89================= 90 91When and where to use kmap_atomic() is straightforward. It is used when code 92wants to access the contents of a page that might be allocated from high memory 93(see __GFP_HIGHMEM), for example a page in the pagecache. The API has two 94functions, and they can be used in a manner similar to the following: 95 96 /* Find the page of interest. */ 97 struct page *page = find_get_page(mapping, offset); 98 99 /* Gain access to the contents of that page. */ 100 void *vaddr = kmap_atomic(page); 101 102 /* Do something to the contents of that page. */ 103 memset(vaddr, 0, PAGE_SIZE); 104 105 /* Unmap that page. */ 106 kunmap_atomic(vaddr); 107 108Note that the kunmap_atomic() call takes the result of the kmap_atomic() call 109not the argument. 110 111If you need to map two pages because you want to copy from one page to 112another you need to keep the kmap_atomic calls strictly nested, like: 113 114 vaddr1 = kmap_atomic(page1); 115 vaddr2 = kmap_atomic(page2); 116 117 memcpy(vaddr1, vaddr2, PAGE_SIZE); 118 119 kunmap_atomic(vaddr2); 120 kunmap_atomic(vaddr1); 121 122 123========================== 124COST OF TEMPORARY MAPPINGS 125========================== 126 127The cost of creating temporary mappings can be quite high. The arch has to 128manipulate the kernel's page tables, the data TLB and/or the MMU's registers. 129 130If CONFIG_HIGHMEM is not set, then the kernel will try and create a mapping 131simply with a bit of arithmetic that will convert the page struct address into 132a pointer to the page contents rather than juggling mappings about. In such a 133case, the unmap operation may be a null operation. 134 135If CONFIG_MMU is not set, then there can be no temporary mappings and no 136highmem. In such a case, the arithmetic approach will also be used. 137 138 139======== 140i386 PAE 141======== 142 143The i386 arch, under some circumstances, will permit you to stick up to 64GiB 144of RAM into your 32-bit machine. This has a number of consequences: 145 146 (*) Linux needs a page-frame structure for each page in the system and the 147 pageframes need to live in the permanent mapping, which means: 148 149 (*) you can have 896M/sizeof(struct page) page-frames at most; with struct 150 page being 32-bytes that would end up being something in the order of 112G 151 worth of pages; the kernel, however, needs to store more than just 152 page-frames in that memory... 153 154 (*) PAE makes your page tables larger - which slows the system down as more 155 data has to be accessed to traverse in TLB fills and the like. One 156 advantage is that PAE has more PTE bits and can provide advanced features 157 like NX and PAT. 158 159The general recommendation is that you don't use more than 8GiB on a 32-bit 160machine - although more might work for you and your workload, you're pretty 161much on your own - don't expect kernel developers to really care much if things 162come apart. 163