summaryrefslogtreecommitdiff
path: root/Documentation/vm/transhuge.rst
diff options
context:
space:
mode:
Diffstat (limited to 'Documentation/vm/transhuge.rst')
-rw-r--r--Documentation/vm/transhuge.rst573
1 files changed, 573 insertions, 0 deletions
diff --git a/Documentation/vm/transhuge.rst b/Documentation/vm/transhuge.rst
new file mode 100644
index 000000000000..569d182cc973
--- /dev/null
+++ b/Documentation/vm/transhuge.rst
@@ -0,0 +1,573 @@
+.. _transhuge:
+
+============================
+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 and tmpfs/shmem.
+But in the future it can expand to other filesystems.
+
+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 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.
+
+- 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)
+
+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 for anonymous memory 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
+anonymous 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 defer >/sys/kernel/mm/transparent_hugepage/defrag
+ echo defer+madvise >/sys/kernel/mm/transparent_hugepage/defrag
+ echo madvise >/sys/kernel/mm/transparent_hugepage/defrag
+ echo never >/sys/kernel/mm/transparent_hugepage/defrag
+
+always
+ means that an application requesting THP will stall on
+ allocation failure and directly reclaim pages and compact
+ memory in an effort to allocate a THP immediately. This may be
+ desirable for virtual machines that benefit heavily from THP
+ use and are willing to delay the VM start to utilise them.
+
+defer
+ means that an application will wake kswapd in the background
+ to reclaim pages and wake kcompactd to compact memory so that
+ THP is available in the near future. It's the responsibility
+ of khugepaged to then install the THP pages later.
+
+defer+madvise
+ will enter direct reclaim and compaction like ``always``, but
+ only for regions that have used madvise(MADV_HUGEPAGE); all
+ other regions will wake kswapd in the background to reclaim
+ pages and wake kcompactd to compact memory so that THP is
+ available in the near future.
+
+madvise
+ will enter direct reclaim like ``always`` but only for regions
+ that are have used madvise(MADV_HUGEPAGE). This is the default
+ behaviour.
+
+never
+ should be self-explanatory.
+
+By default kernel tries to use huge zero page on read page fault to
+anonymous mapping. It's possible to disable huge zero page by writing 0
+or enable it back by writing 1::
+
+ echo 0 >/sys/kernel/mm/transparent_hugepage/use_zero_page
+ echo 1 >/sys/kernel/mm/transparent_hugepage/use_zero_page
+
+Some userspace (such as a test program, or an optimized memory allocation
+library) may want to know the size (in bytes) of a transparent hugepage::
+
+ cat /sys/kernel/mm/transparent_hugepage/hpage_pmd_size
+
+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 by writing 0 or enable
+defrag in khugepaged by writing 1::
+
+ echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag
+ echo 1 >/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
+
+``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.
+
+``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.
+
+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``
+to the kernel command line.
+
+Hugepages in tmpfs/shmem
+========================
+
+You can control hugepage allocation policy in tmpfs with mount option
+``huge=``. It can have following values:
+
+always
+ Attempt to allocate huge pages every time we need a new page;
+
+never
+ Do not allocate huge pages;
+
+within_size
+ Only allocate huge page if it will be fully within i_size.
+ Also respect fadvise()/madvise() hints;
+
+advise
+ Only allocate huge pages if requested with fadvise()/madvise();
+
+The default policy is ``never``.
+
+``mount -o remount,huge= /mountpoint`` works fine after mount: remounting
+``huge=never`` will not attempt to break up huge pages at all, just stop more
+from being allocated.
+
+There's also sysfs knob to control hugepage allocation policy for internal
+shmem mount: /sys/kernel/mm/transparent_hugepage/shmem_enabled. The mount
+is used for SysV SHM, memfds, shared anonymous mmaps (of /dev/zero or
+MAP_ANONYMOUS), GPU drivers' DRM objects, Ashmem.
+
+In addition to policies listed above, shmem_enabled allows two further
+values:
+
+deny
+ For use in emergencies, to force the huge option off from
+ all mounts;
+force
+ Force the huge option on for all - very useful for testing;
+
+Need of application restart
+===========================
+
+The transparent_hugepage/enabled values and tmpfs mount option 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.
+
+Monitoring usage
+================
+
+The number of anonymous transparent huge pages currently used by the
+system is available by reading the AnonHugePages field in ``/proc/meminfo``.
+To identify what applications are using anonymous transparent huge pages,
+it is necessary to read ``/proc/PID/smaps`` and count the AnonHugePages fields
+for each mapping.
+
+The number of file transparent huge pages mapped to userspace is available
+by reading ShmemPmdMapped and ShmemHugePages fields in ``/proc/meminfo``.
+To identify what applications are mapping file transparent huge pages, it
+is necessary to read ``/proc/PID/smaps`` and count the FileHugeMapped 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.
+
+thp_file_alloc
+ is incremented every time a file huge page is successfully
+ allocated.
+
+thp_file_mapped
+ is incremented every time a file huge page is mapped into
+ user address space.
+
+thp_split_page
+ is incremented every time a huge page is split into base
+ pages. This can happen for a variety of reasons but a common
+ reason is that a huge page is old and is being reclaimed.
+ This action implies splitting all PMD the page mapped with.
+
+thp_split_page_failed
+ is incremented if kernel fails to split huge
+ page. This can happen if the page was pinned by somebody.
+
+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.
+
+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.
+
+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.
+
+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.
+
+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. Taking reference on any head/tail page would
+prevent page from being split by anyone.
+
+.. 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 unaware about huge pmds can simply call
+split_huge_pmd(vma, pmd, addr) 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_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. split_huge_page() can fail
+if the page is pinned and you must handle this correctly.
+
+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_pmd(vma, pmd, addr);
+ 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_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
+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
+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 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 ->_refcount.
+
+ - ->_refcount 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). For file huge pages, we also increment
+ ->_mapcount of all sub-pages in order to have race-free detection of
+ last unmap of subpages.
+
+PageDoubleMap() indicates that the page is *possibly* mapped with PTEs.
+
+For anonymous pages PageDoubleMap() also indicates ->_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.
+
+For anonymous pages, we set PG_double_map when a PMD of the page got split
+for the first time, but still have PMD mapping. The additional references
+go away with last compound_mapcount.
+
+File pages get PG_double_map set on first map of the page with PTE and
+goes away when the page gets evicted from page cache.
+
+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 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->_refcount and
+page->_mapcount of anonymous pages. File pages just got unmapped.
+
+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 have zero ->_refcount until atomic_add(). This prevents the
+scanner from getting a reference to the tail page up to that point. After the
+atomic_add() we don't care about the ->_refcount value. We already known how
+many references should be uncharged from the 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 happens during exit(2) if
+a THP crosses a 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.