transhuge.txt 19.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
= Transparent Hugepage Support =

== Objective ==

Performance critical computing applications dealing with large memory
working sets are already running on top of libhugetlbfs and in turn
hugetlbfs. Transparent Hugepage Support is an alternative means of
using huge pages for the backing of virtual memory with huge pages
that supports the automatic promotion and demotion of page sizes and
without the shortcomings of hugetlbfs.

Currently it only works for anonymous memory mappings but in the
future it can expand over the pagecache layer starting with tmpfs.

The reason applications are running faster is because of two
factors. The first factor is almost completely irrelevant and it's not
of significant interest because it'll also have the downside of
requiring larger clear-page copy-page in page faults which is a
potentially negative effect. The first factor consists in taking a
single page fault for each 2M virtual region touched by userland (so
reducing the enter/exit kernel frequency by a 512 times factor). This
only matters the first time the memory is accessed for the lifetime of
a memory mapping. The second long lasting and much more important
factor will affect all subsequent accesses to the memory for the whole
runtime of the application. The second factor consist of two
components: 1) the TLB miss will run faster (especially with
virtualization using nested pagetables but almost always also on bare
metal without virtualization) and 2) a single TLB entry will be
mapping a much larger amount of virtual memory in turn reducing the
number of TLB misses. With virtualization and nested pagetables the
TLB can be mapped of larger size only if both KVM and the Linux guest
are using hugepages but a significant speedup already happens if only
one of the two is using hugepages just because of the fact the TLB
miss is going to run faster.

== Design ==

K
Kirill A. Shutemov 已提交
38 39 40 41
- "graceful fallback": mm components which don't have transparent hugepage
  knowledge fall back to breaking huge pmd mapping into table of ptes and,
  if necessary, split a transparent hugepage. Therefore these components
  can continue working on the regular pages or regular pte mappings.
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118

- if a hugepage allocation fails because of memory fragmentation,
  regular pages should be gracefully allocated instead and mixed in
  the same vma without any failure or significant delay and without
  userland noticing

- if some task quits and more hugepages become available (either
  immediately in the buddy or through the VM), guest physical memory
  backed by regular pages should be relocated on hugepages
  automatically (with khugepaged)

- it doesn't require memory reservation and in turn it uses hugepages
  whenever possible (the only possible reservation here is kernelcore=
  to avoid unmovable pages to fragment all the memory but such a tweak
  is not specific to transparent hugepage support and it's a generic
  feature that applies to all dynamic high order allocations in the
  kernel)

- this initial support only offers the feature in the anonymous memory
  regions but it'd be ideal to move it to tmpfs and the pagecache
  later

Transparent Hugepage Support maximizes the usefulness of free memory
if compared to the reservation approach of hugetlbfs by allowing all
unused memory to be used as cache or other movable (or even unmovable
entities). It doesn't require reservation to prevent hugepage
allocation failures to be noticeable from userland. It allows paging
and all other advanced VM features to be available on the
hugepages. It requires no modifications for applications to take
advantage of it.

Applications however can be further optimized to take advantage of
this feature, like for example they've been optimized before to avoid
a flood of mmap system calls for every malloc(4k). Optimizing userland
is by far not mandatory and khugepaged already can take care of long
lived page allocations even for hugepage unaware applications that
deals with large amounts of memory.

In certain cases when hugepages are enabled system wide, application
may end up allocating more memory resources. An application may mmap a
large region but only touch 1 byte of it, in that case a 2M page might
be allocated instead of a 4k page for no good. This is why it's
possible to disable hugepages system-wide and to only have them inside
MADV_HUGEPAGE madvise regions.

Embedded systems should enable hugepages only inside madvise regions
to eliminate any risk of wasting any precious byte of memory and to
only run faster.

Applications that gets a lot of benefit from hugepages and that don't
risk to lose memory by using hugepages, should use
madvise(MADV_HUGEPAGE) on their critical mmapped regions.

== sysfs ==

Transparent Hugepage Support can be entirely disabled (mostly for
debugging purposes) or only enabled inside MADV_HUGEPAGE regions (to
avoid the risk of consuming more memory resources) or enabled system
wide. This can be achieved with one of:

echo always >/sys/kernel/mm/transparent_hugepage/enabled
echo madvise >/sys/kernel/mm/transparent_hugepage/enabled
echo never >/sys/kernel/mm/transparent_hugepage/enabled

It's also possible to limit defrag efforts in the VM to generate
hugepages in case they're not immediately free to madvise regions or
to never try to defrag memory and simply fallback to regular pages
unless hugepages are immediately available. Clearly if we spend CPU
time to defrag memory, we would expect to gain even more by the fact
we use hugepages later instead of regular pages. This isn't always
guaranteed, but it may be more likely in case the allocation is for a
MADV_HUGEPAGE region.

echo always >/sys/kernel/mm/transparent_hugepage/defrag
echo madvise >/sys/kernel/mm/transparent_hugepage/defrag
echo never >/sys/kernel/mm/transparent_hugepage/defrag

119 120 121 122
By default kernel tries to use huge zero page on read page fault.
It's possible to disable huge zero page by writing 0 or enable it
back by writing 1:

123 124
echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page
echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page
125

126 127 128 129 130 131 132
khugepaged will be automatically started when
transparent_hugepage/enabled is set to "always" or "madvise, and it'll
be automatically shutdown if it's set to "never".

khugepaged runs usually at low frequency so while one may not want to
invoke defrag algorithms synchronously during the page faults, it
should be worth invoking defrag at least in khugepaged. However it's
133 134
also possible to disable defrag in khugepaged by writing 0 or enable
defrag in khugepaged by writing 1:
135

136 137
echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161

You can also control how many pages khugepaged should scan at each
pass:

/sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan

and how many milliseconds to wait in khugepaged between each pass (you
can set this to 0 to run khugepaged at 100% utilization of one core):

/sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs

and how many milliseconds to wait in khugepaged if there's an hugepage
allocation failure to throttle the next allocation attempt.

/sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs

The khugepaged progress can be seen in the number of pages collapsed:

/sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed

for each pass:

/sys/kernel/mm/transparent_hugepage/khugepaged/full_scans

162 163 164 165 166 167 168 169 170 171 172
max_ptes_none specifies how many extra small pages (that are
not already mapped) can be allocated when collapsing a group
of small pages into one large page.

/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_none

A higher value leads to use additional memory for programs.
A lower value leads to gain less thp performance. Value of
max_ptes_none can waste cpu time very little, you can
ignore it.

173 174 175 176 177 178 179 180 181 182
max_ptes_swap specifies how many pages can be brought in from
swap when collapsing a group of pages into a transparent huge page.

/sys/kernel/mm/transparent_hugepage/khugepaged/max_ptes_swap

A higher value can cause excessive swap IO and waste
memory. A lower value can prevent THPs from being
collapsed, resulting fewer pages being collapsed into
THPs, and lower memory access performance.

183 184 185 186 187 188 189 190 191 192 193 194 195 196
== Boot parameter ==

You can change the sysfs boot time defaults of Transparent Hugepage
Support by passing the parameter "transparent_hugepage=always" or
"transparent_hugepage=madvise" or "transparent_hugepage=never"
(without "") to the kernel command line.

== Need of application restart ==

The transparent_hugepage/enabled values only affect future
behavior. So to make them effective you need to restart any
application that could have been using hugepages. This also applies to
the regions registered in khugepaged.

197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
== Monitoring usage ==

The number of transparent huge pages currently used by the system is
available by reading the AnonHugePages field in /proc/meminfo. To
identify what applications are using transparent huge pages, it is
necessary to read /proc/PID/smaps and count the AnonHugePages fields
for each mapping. Note that reading the smaps file is expensive and
reading it frequently will incur overhead.

There are a number of counters in /proc/vmstat that may be used to
monitor how successfully the system is providing huge pages for use.

thp_fault_alloc is incremented every time a huge page is successfully
	allocated to handle a page fault. This applies to both the
	first time a page is faulted and for COW faults.

thp_collapse_alloc is incremented by khugepaged when it has found
	a range of pages to collapse into one huge page and has
	successfully allocated a new huge page to store the data.

thp_fault_fallback is incremented if a page fault fails to allocate
	a huge page and instead falls back to using small pages.

thp_collapse_alloc_failed is incremented if khugepaged found a range
	of pages that should be collapsed into one huge page but failed
	the allocation.

K
Kirill A. Shutemov 已提交
224
thp_split_page is incremented every time a huge page is split into base
225 226
	pages. This can happen for a variety of reasons but a common
	reason is that a huge page is old and is being reclaimed.
K
Kirill A. Shutemov 已提交
227 228 229 230 231
	This action implies splitting all PMD the page mapped with.

thp_split_page_failed is is incremented if kernel fails to split huge
	page. This can happen if the page was pinned by somebody.

232 233 234 235 236
thp_deferred_split_page is incremented when a huge page is put onto split
	queue. This happens when a huge page is partially unmapped and
	splitting it would free up some memory. Pages on split queue are
	going to be split under memory pressure.

K
Kirill A. Shutemov 已提交
237 238 239 240
thp_split_pmd is incremented every time a PMD split into table of PTEs.
	This can happen, for instance, when application calls mprotect() or
	munmap() on part of huge page. It doesn't split huge page, only
	page table entry.
241

242 243 244 245 246 247 248 249
thp_zero_page_alloc is incremented every time a huge zero page is
	successfully allocated. It includes allocations which where
	dropped due race with other allocation. Note, it doesn't count
	every map of the huge zero page, only its allocation.

thp_zero_page_alloc_failed is incremented if kernel fails to allocate
	huge zero page and falls back to using small pages.

250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
As the system ages, allocating huge pages may be expensive as the
system uses memory compaction to copy data around memory to free a
huge page for use. There are some counters in /proc/vmstat to help
monitor this overhead.

compact_stall is incremented every time a process stalls to run
	memory compaction so that a huge page is free for use.

compact_success is incremented if the system compacted memory and
	freed a huge page for use.

compact_fail is incremented if the system tries to compact memory
	but failed.

compact_pages_moved is incremented each time a page is moved. If
	this value is increasing rapidly, it implies that the system
	is copying a lot of data to satisfy the huge page allocation.
	It is possible that the cost of copying exceeds any savings
	from reduced TLB misses.

compact_pagemigrate_failed is incremented when the underlying mechanism
	for moving a page failed.

compact_blocks_moved is incremented each time memory compaction examines
	a huge page aligned range of pages.

It is possible to establish how long the stalls were using the function
tracer to record how long was spent in __alloc_pages_nodemask and
using the mm_page_alloc tracepoint to identify which allocations were
for huge pages.

281 282 283 284 285 286 287 288 289 290
== get_user_pages and follow_page ==

get_user_pages and follow_page if run on a hugepage, will return the
head or tail pages as usual (exactly as they would do on
hugetlbfs). Most gup users will only care about the actual physical
address of the page and its temporary pinning to release after the I/O
is complete, so they won't ever notice the fact the page is huge. But
if any driver is going to mangle over the page structure of the tail
page (like for checking page->mapping or other bits that are relevant
for the head page and not the tail page), it should be updated to jump
K
Kirill A. Shutemov 已提交
291 292
to check head page instead. Taking reference on any head/tail page would
prevent page from being split by anyone.
293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326

NOTE: these aren't new constraints to the GUP API, and they match the
same constrains that applies to hugetlbfs too, so any driver capable
of handling GUP on hugetlbfs will also work fine on transparent
hugepage backed mappings.

In case you can't handle compound pages if they're returned by
follow_page, the FOLL_SPLIT bit can be specified as parameter to
follow_page, so that it will split the hugepages before returning
them. Migration for example passes FOLL_SPLIT as parameter to
follow_page because it's not hugepage aware and in fact it can't work
at all on hugetlbfs (but it instead works fine on transparent
hugepages thanks to FOLL_SPLIT). migration simply can't deal with
hugepages being returned (as it's not only checking the pfn of the
page and pinning it during the copy but it pretends to migrate the
memory in regular page sizes and with regular pte/pmd mappings).

== Optimizing the applications ==

To be guaranteed that the kernel will map a 2M page immediately in any
memory region, the mmap region has to be hugepage naturally
aligned. posix_memalign() can provide that guarantee.

== Hugetlbfs ==

You can use hugetlbfs on a kernel that has transparent hugepage
support enabled just fine as always. No difference can be noted in
hugetlbfs other than there will be less overall fragmentation. All
usual features belonging to hugetlbfs are preserved and
unaffected. libhugetlbfs will also work fine as usual.

== Graceful fallback ==

Code walking pagetables but unware about huge pmds can simply call
K
Kirill A. Shutemov 已提交
327
split_huge_pmd(vma, pmd, addr) where the pmd is the one returned by
328
pmd_offset. It's trivial to make the code transparent hugepage aware
K
Kirill A. Shutemov 已提交
329
by just grepping for "pmd_offset" and adding split_huge_pmd where
330 331 332 333 334 335 336 337
missing after pmd_offset returns the pmd. Thanks to the graceful
fallback design, with a one liner change, you can avoid to write
hundred if not thousand of lines of complex code to make your code
hugepage aware.

If you're not walking pagetables but you run into a physical hugepage
but you can't handle it natively in your code, you can split it by
calling split_huge_page(page). This is what the Linux VM does before
K
Kirill A. Shutemov 已提交
338 339
it tries to swapout the hugepage for example. split_huge_page() can fail
if the page is pinned and you must handle this correctly.
340 341 342 343 344 345 346 347 348 349 350

Example to make mremap.c transparent hugepage aware with a one liner
change:

diff --git a/mm/mremap.c b/mm/mremap.c
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru
		return NULL;

	pmd = pmd_offset(pud, addr);
K
Kirill A. Shutemov 已提交
351
+	split_huge_pmd(vma, pmd, addr);
352 353 354 355 356 357
	if (pmd_none_or_clear_bad(pmd))
		return NULL;

== Locking in hugepage aware code ==

We want as much code as possible hugepage aware, as calling
K
Kirill A. Shutemov 已提交
358
split_huge_page() or split_huge_pmd() has a cost.
359 360 361 362 363 364 365 366

To make pagetable walks huge pmd aware, all you need to do is to call
pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the
mmap_sem in read (or write) mode to be sure an huge pmd cannot be
created from under you by khugepaged (khugepaged collapse_huge_page
takes the mmap_sem in write mode in addition to the anon_vma lock). If
pmd_trans_huge returns false, you just fallback in the old code
paths. If instead pmd_trans_huge returns true, you have to take the
K
Kirill A. Shutemov 已提交
367 368 369
page table lock (pmd_lock()) and re-run pmd_trans_huge. Taking the
page table lock will prevent the huge pmd to be converted into a
regular pmd from under you (split_huge_pmd can run in parallel to the
370
pagetable walk). If the second pmd_trans_huge returns false, you
K
Kirill A. Shutemov 已提交
371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401
should just drop the page table lock and fallback to the old code as
before. Otherwise you can proceed to process the huge pmd and the
hugepage natively. Once finished you can drop the page table lock.

== Refcounts and transparent huge pages ==

Refcounting on THP is mostly consistent with refcounting on other compound
pages:

  - get_page()/put_page() and GUP operate in head page's ->_count.

  - ->_count in tail pages is always zero: get_page_unless_zero() never
    succeed on tail pages.

  - map/unmap of the pages with PTE entry increment/decrement ->_mapcount
    on relevant sub-page of the compound page.

  - map/unmap of the whole compound page accounted in compound_mapcount
    (stored in first tail page).

PageDoubleMap() indicates that ->_mapcount in all subpages is offset up by one.
This additional reference is required to get race-free detection of unmap of
subpages when we have them mapped with both PMDs and PTEs.

This is optimization required to lower overhead of per-subpage mapcount
tracking. The alternative is alter ->_mapcount in all subpages on each
map/unmap of the whole compound page.

We set PG_double_map when a PMD of the page got split for the first time,
but still have PMD mapping. The addtional references go away with last
compound_mapcount.
402 403

split_huge_page internally has to distribute the refcounts in the head
K
Kirill A. Shutemov 已提交
404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
page to the tail pages before clearing all PG_head/tail bits from the page
structures. It can be done easily for refcounts taken by page table
entries. But we don't have enough information on how to distribute any
additional pins (i.e. from get_user_pages). split_huge_page() fails any
requests to split pinned huge page: it expects page count to be equal to
sum of mapcount of all sub-pages plus one (split_huge_page caller must
have reference for head page).

split_huge_page uses migration entries to stabilize page->_count and
page->_mapcount.

We safe against physical memory scanners too: the only legitimate way
scanner can get reference to a page is get_page_unless_zero().

All tail pages has zero ->_count until atomic_add(). It prevent scanner
from geting reference to tail page up to the point. After the atomic_add()
we don't care about ->_count value.  We already known how many references
with should uncharge from head page.

For head page get_page_unless_zero() will succeed and we don't mind. It's
clear where reference should go after split: it will stay on head page.

Note that split_huge_pmd() doesn't have any limitation on refcounting:
pmd can be split at any point and never fails.

== Partial unmap and deferred_split_huge_page() ==

Unmapping part of THP (with munmap() or other way) is not going to free
memory immediately. Instead, we detect that a subpage of THP is not in use
in page_remove_rmap() and queue the THP for splitting if memory pressure
comes. Splitting will free up unused subpages.

Splitting the page right away is not an option due to locking context in
the place where we can detect partial unmap. It's also might be
counterproductive since in many cases partial unmap unmap happens during
exit(2) if an THP crosses VMA boundary.

Function deferred_split_huge_page() is used to queue page for splitting.
The splitting itself will happen when we get memory pressure via shrinker
interface.