1============= 2HugeTLB Pages 3============= 4 5Overview 6======== 7 8The intent of this file is to give a brief summary of hugetlbpage support in 9the Linux kernel. This support is built on top of multiple page size support 10that is provided by most modern architectures. For example, x86 CPUs normally 11support 4K and 2M (1G if architecturally supported) page sizes, ia64 12architecture supports multiple page sizes 4K, 8K, 64K, 256K, 1M, 4M, 16M, 13256M and ppc64 supports 4K and 16M. A TLB is a cache of virtual-to-physical 14translations. Typically this is a very scarce resource on processor. 15Operating systems try to make best use of limited number of TLB resources. 16This optimization is more critical now as bigger and bigger physical memories 17(several GBs) are more readily available. 18 19Users can use the huge page support in Linux kernel by either using the mmap 20system call or standard SYSV shared memory system calls (shmget, shmat). 21 22First the Linux kernel needs to be built with the CONFIG_HUGETLBFS 23(present under "File systems") and CONFIG_HUGETLB_PAGE (selected 24automatically when CONFIG_HUGETLBFS is selected) configuration 25options. 26 27The ``/proc/meminfo`` file provides information about the total number of 28persistent hugetlb pages in the kernel's huge page pool. It also displays 29default huge page size and information about the number of free, reserved 30and surplus huge pages in the pool of huge pages of default size. 31The huge page size is needed for generating the proper alignment and 32size of the arguments to system calls that map huge page regions. 33 34The output of ``cat /proc/meminfo`` will include lines like:: 35 36 HugePages_Total: uuu 37 HugePages_Free: vvv 38 HugePages_Rsvd: www 39 HugePages_Surp: xxx 40 Hugepagesize: yyy kB 41 Hugetlb: zzz kB 42 43where: 44 45HugePages_Total 46 is the size of the pool of huge pages. 47HugePages_Free 48 is the number of huge pages in the pool that are not yet 49 allocated. 50HugePages_Rsvd 51 is short for "reserved," and is the number of huge pages for 52 which a commitment to allocate from the pool has been made, 53 but no allocation has yet been made. Reserved huge pages 54 guarantee that an application will be able to allocate a 55 huge page from the pool of huge pages at fault time. 56HugePages_Surp 57 is short for "surplus," and is the number of huge pages in 58 the pool above the value in ``/proc/sys/vm/nr_hugepages``. The 59 maximum number of surplus huge pages is controlled by 60 ``/proc/sys/vm/nr_overcommit_hugepages``. 61 Note: When the feature of freeing unused vmemmap pages associated 62 with each hugetlb page is enabled, the number of surplus huge pages 63 may be temporarily larger than the maximum number of surplus huge 64 pages when the system is under memory pressure. 65Hugepagesize 66 is the default hugepage size (in kB). 67Hugetlb 68 is the total amount of memory (in kB), consumed by huge 69 pages of all sizes. 70 If huge pages of different sizes are in use, this number 71 will exceed HugePages_Total \* Hugepagesize. To get more 72 detailed information, please, refer to 73 ``/sys/kernel/mm/hugepages`` (described below). 74 75 76``/proc/filesystems`` should also show a filesystem of type "hugetlbfs" 77configured in the kernel. 78 79``/proc/sys/vm/nr_hugepages`` indicates the current number of "persistent" huge 80pages in the kernel's huge page pool. "Persistent" huge pages will be 81returned to the huge page pool when freed by a task. A user with root 82privileges can dynamically allocate more or free some persistent huge pages 83by increasing or decreasing the value of ``nr_hugepages``. 84 85Note: When the feature of freeing unused vmemmap pages associated with each 86hugetlb page is enabled, we can fail to free the huge pages triggered by 87the user when the system is under memory pressure. Please try again later. 88 89Pages that are used as huge pages are reserved inside the kernel and cannot 90be used for other purposes. Huge pages cannot be swapped out under 91memory pressure. 92 93Once a number of huge pages have been pre-allocated to the kernel huge page 94pool, a user with appropriate privilege can use either the mmap system call 95or shared memory system calls to use the huge pages. See the discussion of 96:ref:`Using Huge Pages <using_huge_pages>`, below. 97 98The administrator can allocate persistent huge pages on the kernel boot 99command line by specifying the "hugepages=N" parameter, where 'N' = the 100number of huge pages requested. This is the most reliable method of 101allocating huge pages as memory has not yet become fragmented. 102 103Some platforms support multiple huge page sizes. To allocate huge pages 104of a specific size, one must precede the huge pages boot command parameters 105with a huge page size selection parameter "hugepagesz=<size>". <size> must 106be specified in bytes with optional scale suffix [kKmMgG]. The default huge 107page size may be selected with the "default_hugepagesz=<size>" boot parameter. 108 109Hugetlb boot command line parameter semantics 110 111hugepagesz 112 Specify a huge page size. Used in conjunction with hugepages 113 parameter to preallocate a number of huge pages of the specified 114 size. Hence, hugepagesz and hugepages are typically specified in 115 pairs such as:: 116 117 hugepagesz=2M hugepages=512 118 119 hugepagesz can only be specified once on the command line for a 120 specific huge page size. Valid huge page sizes are architecture 121 dependent. 122hugepages 123 Specify the number of huge pages to preallocate. This typically 124 follows a valid hugepagesz or default_hugepagesz parameter. However, 125 if hugepages is the first or only hugetlb command line parameter it 126 implicitly specifies the number of huge pages of default size to 127 allocate. If the number of huge pages of default size is implicitly 128 specified, it can not be overwritten by a hugepagesz,hugepages 129 parameter pair for the default size. This parameter also has a 130 node format. The node format specifies the number of huge pages 131 to allocate on specific nodes. 132 133 For example, on an architecture with 2M default huge page size:: 134 135 hugepages=256 hugepagesz=2M hugepages=512 136 137 will result in 256 2M huge pages being allocated and a warning message 138 indicating that the hugepages=512 parameter is ignored. If a hugepages 139 parameter is preceded by an invalid hugepagesz parameter, it will 140 be ignored. 141 142 Node format example:: 143 144 hugepagesz=2M hugepages=0:1,1:2 145 146 It will allocate 1 2M hugepage on node0 and 2 2M hugepages on node1. 147 If the node number is invalid, the parameter will be ignored. 148 149default_hugepagesz 150 Specify the default huge page size. This parameter can 151 only be specified once on the command line. default_hugepagesz can 152 optionally be followed by the hugepages parameter to preallocate a 153 specific number of huge pages of default size. The number of default 154 sized huge pages to preallocate can also be implicitly specified as 155 mentioned in the hugepages section above. Therefore, on an 156 architecture with 2M default huge page size:: 157 158 hugepages=256 159 default_hugepagesz=2M hugepages=256 160 hugepages=256 default_hugepagesz=2M 161 162 will all result in 256 2M huge pages being allocated. Valid default 163 huge page size is architecture dependent. 164hugetlb_free_vmemmap 165 When CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP is set, this enables HugeTLB 166 Vmemmap Optimization (HVO). 167 168When multiple huge page sizes are supported, ``/proc/sys/vm/nr_hugepages`` 169indicates the current number of pre-allocated huge pages of the default size. 170Thus, one can use the following command to dynamically allocate/deallocate 171default sized persistent huge pages:: 172 173 echo 20 > /proc/sys/vm/nr_hugepages 174 175This command will try to adjust the number of default sized huge pages in the 176huge page pool to 20, allocating or freeing huge pages, as required. 177 178On a NUMA platform, the kernel will attempt to distribute the huge page pool 179over all the set of allowed nodes specified by the NUMA memory policy of the 180task that modifies ``nr_hugepages``. The default for the allowed nodes--when the 181task has default memory policy--is all on-line nodes with memory. Allowed 182nodes with insufficient available, contiguous memory for a huge page will be 183silently skipped when allocating persistent huge pages. See the 184:ref:`discussion below <mem_policy_and_hp_alloc>` 185of the interaction of task memory policy, cpusets and per node attributes 186with the allocation and freeing of persistent huge pages. 187 188The success or failure of huge page allocation depends on the amount of 189physically contiguous memory that is present in system at the time of the 190allocation attempt. If the kernel is unable to allocate huge pages from 191some nodes in a NUMA system, it will attempt to make up the difference by 192allocating extra pages on other nodes with sufficient available contiguous 193memory, if any. 194 195System administrators may want to put this command in one of the local rc 196init files. This will enable the kernel to allocate huge pages early in 197the boot process when the possibility of getting physical contiguous pages 198is still very high. Administrators can verify the number of huge pages 199actually allocated by checking the sysctl or meminfo. To check the per node 200distribution of huge pages in a NUMA system, use:: 201 202 cat /sys/devices/system/node/node*/meminfo | fgrep Huge 203 204``/proc/sys/vm/nr_overcommit_hugepages`` specifies how large the pool of 205huge pages can grow, if more huge pages than ``/proc/sys/vm/nr_hugepages`` are 206requested by applications. Writing any non-zero value into this file 207indicates that the hugetlb subsystem is allowed to try to obtain that 208number of "surplus" huge pages from the kernel's normal page pool, when the 209persistent huge page pool is exhausted. As these surplus huge pages become 210unused, they are freed back to the kernel's normal page pool. 211 212When increasing the huge page pool size via ``nr_hugepages``, any existing 213surplus pages will first be promoted to persistent huge pages. Then, additional 214huge pages will be allocated, if necessary and if possible, to fulfill 215the new persistent huge page pool size. 216 217The administrator may shrink the pool of persistent huge pages for 218the default huge page size by setting the ``nr_hugepages`` sysctl to a 219smaller value. The kernel will attempt to balance the freeing of huge pages 220across all nodes in the memory policy of the task modifying ``nr_hugepages``. 221Any free huge pages on the selected nodes will be freed back to the kernel's 222normal page pool. 223 224Caveat: Shrinking the persistent huge page pool via ``nr_hugepages`` such that 225it becomes less than the number of huge pages in use will convert the balance 226of the in-use huge pages to surplus huge pages. This will occur even if 227the number of surplus pages would exceed the overcommit value. As long as 228this condition holds--that is, until ``nr_hugepages+nr_overcommit_hugepages`` is 229increased sufficiently, or the surplus huge pages go out of use and are freed-- 230no more surplus huge pages will be allowed to be allocated. 231 232With support for multiple huge page pools at run-time available, much of 233the huge page userspace interface in ``/proc/sys/vm`` has been duplicated in 234sysfs. 235The ``/proc`` interfaces discussed above have been retained for backwards 236compatibility. The root huge page control directory in sysfs is:: 237 238 /sys/kernel/mm/hugepages 239 240For each huge page size supported by the running kernel, a subdirectory 241will exist, of the form:: 242 243 hugepages-${size}kB 244 245Inside each of these directories, the set of files contained in ``/proc`` 246will exist. In addition, two additional interfaces for demoting huge 247pages may exist:: 248 249 demote 250 demote_size 251 nr_hugepages 252 nr_hugepages_mempolicy 253 nr_overcommit_hugepages 254 free_hugepages 255 resv_hugepages 256 surplus_hugepages 257 258The demote interfaces provide the ability to split a huge page into 259smaller huge pages. For example, the x86 architecture supports both 2601GB and 2MB huge pages sizes. A 1GB huge page can be split into 512 2612MB huge pages. Demote interfaces are not available for the smallest 262huge page size. The demote interfaces are: 263 264demote_size 265 is the size of demoted pages. When a page is demoted a corresponding 266 number of huge pages of demote_size will be created. By default, 267 demote_size is set to the next smaller huge page size. If there are 268 multiple smaller huge page sizes, demote_size can be set to any of 269 these smaller sizes. Only huge page sizes less than the current huge 270 pages size are allowed. 271 272demote 273 is used to demote a number of huge pages. A user with root privileges 274 can write to this file. It may not be possible to demote the 275 requested number of huge pages. To determine how many pages were 276 actually demoted, compare the value of nr_hugepages before and after 277 writing to the demote interface. demote is a write only interface. 278 279The interfaces which are the same as in ``/proc`` (all except demote and 280demote_size) function as described above for the default huge page-sized case. 281 282.. _mem_policy_and_hp_alloc: 283 284Interaction of Task Memory Policy with Huge Page Allocation/Freeing 285=================================================================== 286 287Whether huge pages are allocated and freed via the ``/proc`` interface or 288the ``/sysfs`` interface using the ``nr_hugepages_mempolicy`` attribute, the 289NUMA nodes from which huge pages are allocated or freed are controlled by the 290NUMA memory policy of the task that modifies the ``nr_hugepages_mempolicy`` 291sysctl or attribute. When the ``nr_hugepages`` attribute is used, mempolicy 292is ignored. 293 294The recommended method to allocate or free huge pages to/from the kernel 295huge page pool, using the ``nr_hugepages`` example above, is:: 296 297 numactl --interleave <node-list> echo 20 \ 298 >/proc/sys/vm/nr_hugepages_mempolicy 299 300or, more succinctly:: 301 302 numactl -m <node-list> echo 20 >/proc/sys/vm/nr_hugepages_mempolicy 303 304This will allocate or free ``abs(20 - nr_hugepages)`` to or from the nodes 305specified in <node-list>, depending on whether number of persistent huge pages 306is initially less than or greater than 20, respectively. No huge pages will be 307allocated nor freed on any node not included in the specified <node-list>. 308 309When adjusting the persistent hugepage count via ``nr_hugepages_mempolicy``, any 310memory policy mode--bind, preferred, local or interleave--may be used. The 311resulting effect on persistent huge page allocation is as follows: 312 313#. Regardless of mempolicy mode [see 314 Documentation/admin-guide/mm/numa_memory_policy.rst], 315 persistent huge pages will be distributed across the node or nodes 316 specified in the mempolicy as if "interleave" had been specified. 317 However, if a node in the policy does not contain sufficient contiguous 318 memory for a huge page, the allocation will not "fallback" to the nearest 319 neighbor node with sufficient contiguous memory. To do this would cause 320 undesirable imbalance in the distribution of the huge page pool, or 321 possibly, allocation of persistent huge pages on nodes not allowed by 322 the task's memory policy. 323 324#. One or more nodes may be specified with the bind or interleave policy. 325 If more than one node is specified with the preferred policy, only the 326 lowest numeric id will be used. Local policy will select the node where 327 the task is running at the time the nodes_allowed mask is constructed. 328 For local policy to be deterministic, the task must be bound to a cpu or 329 cpus in a single node. Otherwise, the task could be migrated to some 330 other node at any time after launch and the resulting node will be 331 indeterminate. Thus, local policy is not very useful for this purpose. 332 Any of the other mempolicy modes may be used to specify a single node. 333 334#. The nodes allowed mask will be derived from any non-default task mempolicy, 335 whether this policy was set explicitly by the task itself or one of its 336 ancestors, such as numactl. This means that if the task is invoked from a 337 shell with non-default policy, that policy will be used. One can specify a 338 node list of "all" with numactl --interleave or --membind [-m] to achieve 339 interleaving over all nodes in the system or cpuset. 340 341#. Any task mempolicy specified--e.g., using numactl--will be constrained by 342 the resource limits of any cpuset in which the task runs. Thus, there will 343 be no way for a task with non-default policy running in a cpuset with a 344 subset of the system nodes to allocate huge pages outside the cpuset 345 without first moving to a cpuset that contains all of the desired nodes. 346 347#. Boot-time huge page allocation attempts to distribute the requested number 348 of huge pages over all on-lines nodes with memory. 349 350Per Node Hugepages Attributes 351============================= 352 353A subset of the contents of the root huge page control directory in sysfs, 354described above, will be replicated under each the system device of each 355NUMA node with memory in:: 356 357 /sys/devices/system/node/node[0-9]*/hugepages/ 358 359Under this directory, the subdirectory for each supported huge page size 360contains the following attribute files:: 361 362 nr_hugepages 363 free_hugepages 364 surplus_hugepages 365 366The free\_' and surplus\_' attribute files are read-only. They return the number 367of free and surplus [overcommitted] huge pages, respectively, on the parent 368node. 369 370The ``nr_hugepages`` attribute returns the total number of huge pages on the 371specified node. When this attribute is written, the number of persistent huge 372pages on the parent node will be adjusted to the specified value, if sufficient 373resources exist, regardless of the task's mempolicy or cpuset constraints. 374 375Note that the number of overcommit and reserve pages remain global quantities, 376as we don't know until fault time, when the faulting task's mempolicy is 377applied, from which node the huge page allocation will be attempted. 378 379The hugetlb may be migrated between the per-node hugepages pool in the following 380scenarios: memory offline, memory failure, longterm pinning, syscalls(mbind, 381migrate_pages and move_pages), alloc_contig_range() and alloc_contig_pages(). 382Now only memory offline, memory failure and syscalls allow fallbacking to allocate 383a new hugetlb on a different node if the current node is unable to allocate during 384hugetlb migration, that means these 3 cases can break the per-node hugepages pool. 385 386.. _using_huge_pages: 387 388Using Huge Pages 389================ 390 391If the user applications are going to request huge pages using mmap system 392call, then it is required that system administrator mount a file system of 393type hugetlbfs:: 394 395 mount -t hugetlbfs \ 396 -o uid=<value>,gid=<value>,mode=<value>,pagesize=<value>,size=<value>,\ 397 min_size=<value>,nr_inodes=<value> none /mnt/huge 398 399This command mounts a (pseudo) filesystem of type hugetlbfs on the directory 400``/mnt/huge``. Any file created on ``/mnt/huge`` uses huge pages. 401 402The ``uid`` and ``gid`` options sets the owner and group of the root of the 403file system. By default the ``uid`` and ``gid`` of the current process 404are taken. 405 406The ``mode`` option sets the mode of root of file system to value & 01777. 407This value is given in octal. By default the value 0755 is picked. 408 409If the platform supports multiple huge page sizes, the ``pagesize`` option can 410be used to specify the huge page size and associated pool. ``pagesize`` 411is specified in bytes. If ``pagesize`` is not specified the platform's 412default huge page size and associated pool will be used. 413 414The ``size`` option sets the maximum value of memory (huge pages) allowed 415for that filesystem (``/mnt/huge``). The ``size`` option can be specified 416in bytes, or as a percentage of the specified huge page pool (``nr_hugepages``). 417The size is rounded down to HPAGE_SIZE boundary. 418 419The ``min_size`` option sets the minimum value of memory (huge pages) allowed 420for the filesystem. ``min_size`` can be specified in the same way as ``size``, 421either bytes or a percentage of the huge page pool. 422At mount time, the number of huge pages specified by ``min_size`` are reserved 423for use by the filesystem. 424If there are not enough free huge pages available, the mount will fail. 425As huge pages are allocated to the filesystem and freed, the reserve count 426is adjusted so that the sum of allocated and reserved huge pages is always 427at least ``min_size``. 428 429The option ``nr_inodes`` sets the maximum number of inodes that ``/mnt/huge`` 430can use. 431 432If the ``size``, ``min_size`` or ``nr_inodes`` option is not provided on 433command line then no limits are set. 434 435For ``pagesize``, ``size``, ``min_size`` and ``nr_inodes`` options, you can 436use [G|g]/[M|m]/[K|k] to represent giga/mega/kilo. 437For example, size=2K has the same meaning as size=2048. 438 439While read system calls are supported on files that reside on hugetlb 440file systems, write system calls are not. 441 442Regular chown, chgrp, and chmod commands (with right permissions) could be 443used to change the file attributes on hugetlbfs. 444 445Also, it is important to note that no such mount command is required if 446applications are going to use only shmat/shmget system calls or mmap with 447MAP_HUGETLB. For an example of how to use mmap with MAP_HUGETLB see 448:ref:`map_hugetlb <map_hugetlb>` below. 449 450Users who wish to use hugetlb memory via shared memory segment should be 451members of a supplementary group and system admin needs to configure that gid 452into ``/proc/sys/vm/hugetlb_shm_group``. It is possible for same or different 453applications to use any combination of mmaps and shm* calls, though the mount of 454filesystem will be required for using mmap calls without MAP_HUGETLB. 455 456Syscalls that operate on memory backed by hugetlb pages only have their lengths 457aligned to the native page size of the processor; they will normally fail with 458errno set to EINVAL or exclude hugetlb pages that extend beyond the length if 459not hugepage aligned. For example, munmap(2) will fail if memory is backed by 460a hugetlb page and the length is smaller than the hugepage size. 461 462 463Examples 464======== 465 466.. _map_hugetlb: 467 468``map_hugetlb`` 469 see tools/testing/selftests/mm/map_hugetlb.c 470 471``hugepage-shm`` 472 see tools/testing/selftests/mm/hugepage-shm.c 473 474``hugepage-mmap`` 475 see tools/testing/selftests/mm/hugepage-mmap.c 476 477The `libhugetlbfs`_ library provides a wide range of userspace tools 478to help with huge page usability, environment setup, and control. 479 480.. _libhugetlbfs: https://github.com/libhugetlbfs/libhugetlbfs 481