1.. _admin_guide_transhuge:
2
3============================
4Transparent Hugepage Support
5============================
6
7Objective
8=========
9
10Performance critical computing applications dealing with large memory
11working sets are already running on top of libhugetlbfs and in turn
12hugetlbfs. Transparent HugePage Support (THP) is an alternative mean of
13using huge pages for the backing of virtual memory with huge pages
14that supports the automatic promotion and demotion of page sizes and
15without the shortcomings of hugetlbfs.
16
17Currently THP only works for anonymous memory mappings and tmpfs/shmem.
18But in the future it can expand to other filesystems.
19
20.. note::
21   in the examples below we presume that the basic page size is 4K and
22   the huge page size is 2M, although the actual numbers may vary
23   depending on the CPU architecture.
24
25The reason applications are running faster is because of two
26factors. The first factor is almost completely irrelevant and it's not
27of significant interest because it'll also have the downside of
28requiring larger clear-page copy-page in page faults which is a
29potentially negative effect. The first factor consists in taking a
30single page fault for each 2M virtual region touched by userland (so
31reducing the enter/exit kernel frequency by a 512 times factor). This
32only matters the first time the memory is accessed for the lifetime of
33a memory mapping. The second long lasting and much more important
34factor will affect all subsequent accesses to the memory for the whole
35runtime of the application. The second factor consist of two
36components:
37
381) the TLB miss will run faster (especially with virtualization using
39   nested pagetables but almost always also on bare metal without
40   virtualization)
41
422) a single TLB entry will be mapping a much larger amount of virtual
43   memory in turn reducing the number of TLB misses. With
44   virtualization and nested pagetables the TLB can be mapped of
45   larger size only if both KVM and the Linux guest are using
46   hugepages but a significant speedup already happens if only one of
47   the two is using hugepages just because of the fact the TLB miss is
48   going to run faster.
49
50THP can be enabled system wide or restricted to certain tasks or even
51memory ranges inside task's address space. Unless THP is completely
52disabled, there is ``khugepaged`` daemon that scans memory and
53collapses sequences of basic pages into huge pages.
54
55The THP behaviour is controlled via :ref:`sysfs <thp_sysfs>`
56interface and using madivse(2) and prctl(2) system calls.
57
58Transparent Hugepage Support maximizes the usefulness of free memory
59if compared to the reservation approach of hugetlbfs by allowing all
60unused memory to be used as cache or other movable (or even unmovable
61entities). It doesn't require reservation to prevent hugepage
62allocation failures to be noticeable from userland. It allows paging
63and all other advanced VM features to be available on the
64hugepages. It requires no modifications for applications to take
65advantage of it.
66
67Applications however can be further optimized to take advantage of
68this feature, like for example they've been optimized before to avoid
69a flood of mmap system calls for every malloc(4k). Optimizing userland
70is by far not mandatory and khugepaged already can take care of long
71lived page allocations even for hugepage unaware applications that
72deals with large amounts of memory.
73
74In certain cases when hugepages are enabled system wide, application
75may end up allocating more memory resources. An application may mmap a
76large region but only touch 1 byte of it, in that case a 2M page might
77be allocated instead of a 4k page for no good. This is why it's
78possible to disable hugepages system-wide and to only have them inside
79MADV_HUGEPAGE madvise regions.
80
81Embedded systems should enable hugepages only inside madvise regions
82to eliminate any risk of wasting any precious byte of memory and to
83only run faster.
84
85Applications that gets a lot of benefit from hugepages and that don't
86risk to lose memory by using hugepages, should use
87madvise(MADV_HUGEPAGE) on their critical mmapped regions.
88
89.. _thp_sysfs:
90
91sysfs
92=====
93
94Global THP controls
95-------------------
96
97Transparent Hugepage Support for anonymous memory can be entirely disabled
98(mostly for debugging purposes) or only enabled inside MADV_HUGEPAGE
99regions (to avoid the risk of consuming more memory resources) or enabled
100system wide. This can be achieved with one of::
101
102	echo always >/sys/kernel/mm/transparent_hugepage/enabled
103	echo madvise >/sys/kernel/mm/transparent_hugepage/enabled
104	echo never >/sys/kernel/mm/transparent_hugepage/enabled
105
106It's also possible to limit defrag efforts in the VM to generate
107anonymous hugepages in case they're not immediately free to madvise
108regions or to never try to defrag memory and simply fallback to regular
109pages unless hugepages are immediately available. Clearly if we spend CPU
110time to defrag memory, we would expect to gain even more by the fact we
111use hugepages later instead of regular pages. This isn't always
112guaranteed, but it may be more likely in case the allocation is for a
113MADV_HUGEPAGE region.
114
115::
116
117	echo always >/sys/kernel/mm/transparent_hugepage/defrag
118	echo defer >/sys/kernel/mm/transparent_hugepage/defrag
119	echo defer+madvise >/sys/kernel/mm/transparent_hugepage/defrag
120	echo madvise >/sys/kernel/mm/transparent_hugepage/defrag
121	echo never >/sys/kernel/mm/transparent_hugepage/defrag
122
123always
124	means that an application requesting THP will stall on
125	allocation failure and directly reclaim pages and compact
126	memory in an effort to allocate a THP immediately. This may be
127	desirable for virtual machines that benefit heavily from THP
128	use and are willing to delay the VM start to utilise them.
129
130defer
131	means that an application will wake kswapd in the background
132	to reclaim pages and wake kcompactd to compact memory so that
133	THP is available in the near future. It's the responsibility
134	of khugepaged to then install the THP pages later.
135
136defer+madvise
137	will enter direct reclaim and compaction like ``always``, but
138	only for regions that have used madvise(MADV_HUGEPAGE); all
139	other regions will wake kswapd in the background to reclaim
140	pages and wake kcompactd to compact memory so that THP is
141	available in the near future.
142
143madvise
144	will enter direct reclaim like ``always`` but only for regions
145	that are have used madvise(MADV_HUGEPAGE). This is the default
146	behaviour.
147
148never
149	should be self-explanatory.
150
151By default kernel tries to use huge zero page on read page fault to
152anonymous mapping. It's possible to disable huge zero page by writing 0
153or enable it back by writing 1::
154
155	echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page
156	echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page
157
158Some userspace (such as a test program, or an optimized memory allocation
159library) may want to know the size (in bytes) of a transparent hugepage::
160
161	cat /sys/kernel/mm/transparent_hugepage/hpage_pmd_size
162
163khugepaged will be automatically started when
164transparent_hugepage/enabled is set to "always" or "madvise, and it'll
165be automatically shutdown if it's set to "never".
166
167Khugepaged controls
168-------------------
169
170khugepaged runs usually at low frequency so while one may not want to
171invoke defrag algorithms synchronously during the page faults, it
172should be worth invoking defrag at least in khugepaged. However it's
173also possible to disable defrag in khugepaged by writing 0 or enable
174defrag in khugepaged by writing 1::
175
176	echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
177	echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
178
179You can also control how many pages khugepaged should scan at each
180pass::
181
182	/sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan
183
184and how many milliseconds to wait in khugepaged between each pass (you
185can set this to 0 to run khugepaged at 100% utilization of one core)::
186
187	/sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs
188
189and how many milliseconds to wait in khugepaged if there's an hugepage
190allocation failure to throttle the next allocation attempt::
191
192	/sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs
193
194The khugepaged progress can be seen in the number of pages collapsed::
195
196	/sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed
197
198for each pass::
199
200	/sys/kernel/mm/transparent_hugepage/khugepaged/full_scans
201
202``max_ptes_none`` specifies how many extra small pages (that are
203not already mapped) can be allocated when collapsing a group
204of small pages into one large page::
205
206	/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none
207
208A higher value leads to use additional memory for programs.
209A lower value leads to gain less thp performance. Value of
210max_ptes_none can waste cpu time very little, you can
211ignore it.
212
213``max_ptes_swap`` specifies how many pages can be brought in from
214swap when collapsing a group of pages into a transparent huge page::
215
216	/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap
217
218A higher value can cause excessive swap IO and waste
219memory. A lower value can prevent THPs from being
220collapsed, resulting fewer pages being collapsed into
221THPs, and lower memory access performance.
222
223Boot parameter
224==============
225
226You can change the sysfs boot time defaults of Transparent Hugepage
227Support by passing the parameter ``transparent_hugepage=always`` or
228``transparent_hugepage=madvise`` or ``transparent_hugepage=never``
229to the kernel command line.
230
231Hugepages in tmpfs/shmem
232========================
233
234You can control hugepage allocation policy in tmpfs with mount option
235``huge=``. It can have following values:
236
237always
238    Attempt to allocate huge pages every time we need a new page;
239
240never
241    Do not allocate huge pages;
242
243within_size
244    Only allocate huge page if it will be fully within i_size.
245    Also respect fadvise()/madvise() hints;
246
247advise
248    Only allocate huge pages if requested with fadvise()/madvise();
249
250The default policy is ``never``.
251
252``mount -o remount,huge= /mountpoint`` works fine after mount: remounting
253``huge=never`` will not attempt to break up huge pages at all, just stop more
254from being allocated.
255
256There's also sysfs knob to control hugepage allocation policy for internal
257shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount
258is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or
259MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem.
260
261In addition to policies listed above, shmem_enabled allows two further
262values:
263
264deny
265    For use in emergencies, to force the huge option off from
266    all mounts;
267force
268    Force the huge option on for all - very useful for testing;
269
270Need of application restart
271===========================
272
273The transparent_hugepage/enabled values and tmpfs mount option only affect
274future behavior. So to make them effective you need to restart any
275application that could have been using hugepages. This also applies to the
276regions registered in khugepaged.
277
278Monitoring usage
279================
280
281The number of anonymous transparent huge pages currently used by the
282system is available by reading the AnonHugePages field in ``/proc/meminfo``.
283To identify what applications are using anonymous transparent huge pages,
284it is necessary to read ``/proc/PID/smaps`` and count the AnonHugePages fields
285for each mapping.
286
287The number of file transparent huge pages mapped to userspace is available
288by reading ShmemPmdMapped and ShmemHugePages fields in ``/proc/meminfo``.
289To identify what applications are mapping file transparent huge pages, it
290is necessary to read ``/proc/PID/smaps`` and count the FileHugeMapped fields
291for each mapping.
292
293Note that reading the smaps file is expensive and reading it
294frequently will incur overhead.
295
296There are a number of counters in ``/proc/vmstat`` that may be used to
297monitor how successfully the system is providing huge pages for use.
298
299thp_fault_alloc
300	is incremented every time a huge page is successfully
301	allocated to handle a page fault. This applies to both the
302	first time a page is faulted and for COW faults.
303
304thp_collapse_alloc
305	is incremented by khugepaged when it has found
306	a range of pages to collapse into one huge page and has
307	successfully allocated a new huge page to store the data.
308
309thp_fault_fallback
310	is incremented if a page fault fails to allocate
311	a huge page and instead falls back to using small pages.
312
313thp_collapse_alloc_failed
314	is incremented if khugepaged found a range
315	of pages that should be collapsed into one huge page but failed
316	the allocation.
317
318thp_file_alloc
319	is incremented every time a file huge page is successfully
320	allocated.
321
322thp_file_mapped
323	is incremented every time a file huge page is mapped into
324	user address space.
325
326thp_split_page
327	is incremented every time a huge page is split into base
328	pages. This can happen for a variety of reasons but a common
329	reason is that a huge page is old and is being reclaimed.
330	This action implies splitting all PMD the page mapped with.
331
332thp_split_page_failed
333	is incremented if kernel fails to split huge
334	page. This can happen if the page was pinned by somebody.
335
336thp_deferred_split_page
337	is incremented when a huge page is put onto split
338	queue. This happens when a huge page is partially unmapped and
339	splitting it would free up some memory. Pages on split queue are
340	going to be split under memory pressure.
341
342thp_split_pmd
343	is incremented every time a PMD split into table of PTEs.
344	This can happen, for instance, when application calls mprotect() or
345	munmap() on part of huge page. It doesn't split huge page, only
346	page table entry.
347
348thp_zero_page_alloc
349	is incremented every time a huge zero page is
350	successfully allocated. It includes allocations which where
351	dropped due race with other allocation. Note, it doesn't count
352	every map of the huge zero page, only its allocation.
353
354thp_zero_page_alloc_failed
355	is incremented if kernel fails to allocate
356	huge zero page and falls back to using small pages.
357
358thp_swpout
359	is incremented every time a huge page is swapout in one
360	piece without splitting.
361
362thp_swpout_fallback
363	is incremented if a huge page has to be split before swapout.
364	Usually because failed to allocate some continuous swap space
365	for the huge page.
366
367As the system ages, allocating huge pages may be expensive as the
368system uses memory compaction to copy data around memory to free a
369huge page for use. There are some counters in ``/proc/vmstat`` to help
370monitor this overhead.
371
372compact_stall
373	is incremented every time a process stalls to run
374	memory compaction so that a huge page is free for use.
375
376compact_success
377	is incremented if the system compacted memory and
378	freed a huge page for use.
379
380compact_fail
381	is incremented if the system tries to compact memory
382	but failed.
383
384compact_pages_moved
385	is incremented each time a page is moved. If
386	this value is increasing rapidly, it implies that the system
387	is copying a lot of data to satisfy the huge page allocation.
388	It is possible that the cost of copying exceeds any savings
389	from reduced TLB misses.
390
391compact_pagemigrate_failed
392	is incremented when the underlying mechanism
393	for moving a page failed.
394
395compact_blocks_moved
396	is incremented each time memory compaction examines
397	a huge page aligned range of pages.
398
399It is possible to establish how long the stalls were using the function
400tracer to record how long was spent in __alloc_pages_nodemask and
401using the mm_page_alloc tracepoint to identify which allocations were
402for huge pages.
403
404Optimizing the applications
405===========================
406
407To be guaranteed that the kernel will map a 2M page immediately in any
408memory region, the mmap region has to be hugepage naturally
409aligned. posix_memalign() can provide that guarantee.
410
411Hugetlbfs
412=========
413
414You can use hugetlbfs on a kernel that has transparent hugepage
415support enabled just fine as always. No difference can be noted in
416hugetlbfs other than there will be less overall fragmentation. All
417usual features belonging to hugetlbfs are preserved and
418unaffected. libhugetlbfs will also work fine as usual.
419