thp: transparent hugepage support documentation

= 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 ==

- "graceful fallback": mm components which don't have transparent

  hugepage knowledge fall back to breaking a transparent hugepage and

  working on the regular pages and their respective regular pmd/pte

  mappings

- 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

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

also possible to disable defrag in khugepaged:

echo yes >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag

echo no >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag

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

== 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.

== 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

to check head page instead (while serializing properly against

split_huge_page() to avoid the head and tail pages to disappear from

under it, see the futex code to see an example of that, hugetlbfs also

needed special handling in futex code for similar reasons).

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

split_huge_page_pmd(mm, pmd) where the pmd is the one returned by

pmd_offset. It's trivial to make the code transparent hugepage aware

by just grepping for "pmd_offset" and adding split_huge_page_pmd where

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

it tries to swapout the hugepage for example.

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);

+	split_huge_page_pmd(mm, pmd);

	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

split_huge_page() or split_huge_page_pmd() has a cost.

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

mm->page_table_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_page can run in parallel to the

pagetable walk). If the second pmd_trans_huge returns false, you

should just drop the page_table_lock and fallback to the old code as

before. Otherwise you should run pmd_trans_splitting on the pmd. In

case pmd_trans_splitting returns true, it means split_huge_page is

already in the middle of splitting the page. So if pmd_trans_splitting

returns true it's enough to drop the page_table_lock and call

wait_split_huge_page and then fallback the old code paths. You are

guaranteed by the time wait_split_huge_page returns, the pmd isn't

huge anymore. If pmd_trans_splitting returns false, you can proceed to

process the huge pmd and the hugepage natively. Once finished you can

drop the page_table_lock.

== compound_lock, get_user_pages and put_page ==

split_huge_page internally has to distribute the refcounts in the head

page to the tail pages before clearing all PG_head/tail bits from the

page structures. It can do that easily for refcounts taken by huge pmd

mappings. But the GUI API as created by hugetlbfs (that returns head

and tail pages if running get_user_pages on an address backed by any

hugepage), requires the refcount to be accounted on the tail pages and

not only in the head pages, if we want to be able to run

split_huge_page while there are gup pins established on any tail

page. Failure to be able to run split_huge_page if there's any gup pin

on any tail page, would mean having to split all hugepages upfront in

get_user_pages which is unacceptable as too many gup users are

performance critical and they must work natively on hugepages like

they work natively on hugetlbfs already (hugetlbfs is simpler because

hugetlbfs pages cannot be splitted so there wouldn't be requirement of

accounting the pins on the tail pages for hugetlbfs). If we wouldn't

account the gup refcounts on the tail pages during gup, we won't know

anymore which tail page is pinned by gup and which is not while we run

split_huge_page. But we still have to add the gup pin to the head page

too, to know when we can free the compound page in case it's never

splitted during its lifetime. That requires changing not just

get_page, but put_page as well so that when put_page runs on a tail

page (and only on a tail page) it will find its respective head page,

and then it will decrease the head page refcount in addition to the

tail page refcount. To obtain a head page reliably and to decrease its

refcount without race conditions, put_page has to serialize against

__split_huge_page_refcount using a special per-page lock called

compound_lock.