Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 1 | = Transparent Hugepage Support = |
| 2 | |
| 3 | == Objective == |
| 4 | |
| 5 | Performance critical computing applications dealing with large memory |
| 6 | working sets are already running on top of libhugetlbfs and in turn |
| 7 | hugetlbfs. Transparent Hugepage Support is an alternative means of |
| 8 | using huge pages for the backing of virtual memory with huge pages |
| 9 | that supports the automatic promotion and demotion of page sizes and |
| 10 | without the shortcomings of hugetlbfs. |
| 11 | |
| 12 | Currently it only works for anonymous memory mappings but in the |
| 13 | future it can expand over the pagecache layer starting with tmpfs. |
| 14 | |
| 15 | The reason applications are running faster is because of two |
| 16 | factors. The first factor is almost completely irrelevant and it's not |
| 17 | of significant interest because it'll also have the downside of |
| 18 | requiring larger clear-page copy-page in page faults which is a |
| 19 | potentially negative effect. The first factor consists in taking a |
| 20 | single page fault for each 2M virtual region touched by userland (so |
| 21 | reducing the enter/exit kernel frequency by a 512 times factor). This |
| 22 | only matters the first time the memory is accessed for the lifetime of |
| 23 | a memory mapping. The second long lasting and much more important |
| 24 | factor will affect all subsequent accesses to the memory for the whole |
| 25 | runtime of the application. The second factor consist of two |
| 26 | components: 1) the TLB miss will run faster (especially with |
| 27 | virtualization using nested pagetables but almost always also on bare |
| 28 | metal without virtualization) and 2) a single TLB entry will be |
| 29 | mapping a much larger amount of virtual memory in turn reducing the |
| 30 | number of TLB misses. With virtualization and nested pagetables the |
| 31 | TLB can be mapped of larger size only if both KVM and the Linux guest |
| 32 | are using hugepages but a significant speedup already happens if only |
| 33 | one of the two is using hugepages just because of the fact the TLB |
| 34 | miss is going to run faster. |
| 35 | |
| 36 | == Design == |
| 37 | |
| 38 | - "graceful fallback": mm components which don't have transparent |
| 39 | hugepage knowledge fall back to breaking a transparent hugepage and |
| 40 | working on the regular pages and their respective regular pmd/pte |
| 41 | mappings |
| 42 | |
| 43 | - if a hugepage allocation fails because of memory fragmentation, |
| 44 | regular pages should be gracefully allocated instead and mixed in |
| 45 | the same vma without any failure or significant delay and without |
| 46 | userland noticing |
| 47 | |
| 48 | - if some task quits and more hugepages become available (either |
| 49 | immediately in the buddy or through the VM), guest physical memory |
| 50 | backed by regular pages should be relocated on hugepages |
| 51 | automatically (with khugepaged) |
| 52 | |
| 53 | - it doesn't require memory reservation and in turn it uses hugepages |
| 54 | whenever possible (the only possible reservation here is kernelcore= |
| 55 | to avoid unmovable pages to fragment all the memory but such a tweak |
| 56 | is not specific to transparent hugepage support and it's a generic |
| 57 | feature that applies to all dynamic high order allocations in the |
| 58 | kernel) |
| 59 | |
| 60 | - this initial support only offers the feature in the anonymous memory |
| 61 | regions but it'd be ideal to move it to tmpfs and the pagecache |
| 62 | later |
| 63 | |
| 64 | Transparent Hugepage Support maximizes the usefulness of free memory |
| 65 | if compared to the reservation approach of hugetlbfs by allowing all |
| 66 | unused memory to be used as cache or other movable (or even unmovable |
| 67 | entities). It doesn't require reservation to prevent hugepage |
| 68 | allocation failures to be noticeable from userland. It allows paging |
| 69 | and all other advanced VM features to be available on the |
| 70 | hugepages. It requires no modifications for applications to take |
| 71 | advantage of it. |
| 72 | |
| 73 | Applications however can be further optimized to take advantage of |
| 74 | this feature, like for example they've been optimized before to avoid |
| 75 | a flood of mmap system calls for every malloc(4k). Optimizing userland |
| 76 | is by far not mandatory and khugepaged already can take care of long |
| 77 | lived page allocations even for hugepage unaware applications that |
| 78 | deals with large amounts of memory. |
| 79 | |
| 80 | In certain cases when hugepages are enabled system wide, application |
| 81 | may end up allocating more memory resources. An application may mmap a |
| 82 | large region but only touch 1 byte of it, in that case a 2M page might |
| 83 | be allocated instead of a 4k page for no good. This is why it's |
| 84 | possible to disable hugepages system-wide and to only have them inside |
| 85 | MADV_HUGEPAGE madvise regions. |
| 86 | |
| 87 | Embedded systems should enable hugepages only inside madvise regions |
| 88 | to eliminate any risk of wasting any precious byte of memory and to |
| 89 | only run faster. |
| 90 | |
| 91 | Applications that gets a lot of benefit from hugepages and that don't |
| 92 | risk to lose memory by using hugepages, should use |
| 93 | madvise(MADV_HUGEPAGE) on their critical mmapped regions. |
| 94 | |
| 95 | == sysfs == |
| 96 | |
| 97 | Transparent Hugepage Support can be entirely disabled (mostly for |
| 98 | debugging purposes) or only enabled inside MADV_HUGEPAGE regions (to |
| 99 | avoid the risk of consuming more memory resources) or enabled system |
| 100 | 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 | |
| 106 | It's also possible to limit defrag efforts in the VM to generate |
| 107 | hugepages in case they're not immediately free to madvise regions or |
| 108 | to never try to defrag memory and simply fallback to regular pages |
| 109 | unless hugepages are immediately available. Clearly if we spend CPU |
| 110 | time to defrag memory, we would expect to gain even more by the fact |
| 111 | we use hugepages later instead of regular pages. This isn't always |
| 112 | guaranteed, but it may be more likely in case the allocation is for a |
| 113 | MADV_HUGEPAGE region. |
| 114 | |
| 115 | echo always >/sys/kernel/mm/transparent_hugepage/defrag |
| 116 | echo madvise >/sys/kernel/mm/transparent_hugepage/defrag |
| 117 | echo never >/sys/kernel/mm/transparent_hugepage/defrag |
| 118 | |
| 119 | khugepaged will be automatically started when |
| 120 | transparent_hugepage/enabled is set to "always" or "madvise, and it'll |
| 121 | be automatically shutdown if it's set to "never". |
| 122 | |
| 123 | khugepaged runs usually at low frequency so while one may not want to |
| 124 | invoke defrag algorithms synchronously during the page faults, it |
| 125 | should be worth invoking defrag at least in khugepaged. However it's |
David Rientjes | e369fde | 2011-09-22 14:11:38 -0700 | [diff] [blame] | 126 | also possible to disable defrag in khugepaged by writing 0 or enable |
| 127 | defrag in khugepaged by writing 1: |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 128 | |
David Rientjes | e369fde | 2011-09-22 14:11:38 -0700 | [diff] [blame] | 129 | echo 0 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag |
| 130 | echo 1 >/sys/kernel/mm/transparent_hugepage/khugepaged/defrag |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 131 | |
| 132 | You can also control how many pages khugepaged should scan at each |
| 133 | pass: |
| 134 | |
| 135 | /sys/kernel/mm/transparent_hugepage/khugepaged/pages_to_scan |
| 136 | |
| 137 | and how many milliseconds to wait in khugepaged between each pass (you |
| 138 | can set this to 0 to run khugepaged at 100% utilization of one core): |
| 139 | |
| 140 | /sys/kernel/mm/transparent_hugepage/khugepaged/scan_sleep_millisecs |
| 141 | |
| 142 | and how many milliseconds to wait in khugepaged if there's an hugepage |
| 143 | allocation failure to throttle the next allocation attempt. |
| 144 | |
| 145 | /sys/kernel/mm/transparent_hugepage/khugepaged/alloc_sleep_millisecs |
| 146 | |
| 147 | The khugepaged progress can be seen in the number of pages collapsed: |
| 148 | |
| 149 | /sys/kernel/mm/transparent_hugepage/khugepaged/pages_collapsed |
| 150 | |
| 151 | for each pass: |
| 152 | |
| 153 | /sys/kernel/mm/transparent_hugepage/khugepaged/full_scans |
| 154 | |
| 155 | == Boot parameter == |
| 156 | |
| 157 | You can change the sysfs boot time defaults of Transparent Hugepage |
| 158 | Support by passing the parameter "transparent_hugepage=always" or |
| 159 | "transparent_hugepage=madvise" or "transparent_hugepage=never" |
| 160 | (without "") to the kernel command line. |
| 161 | |
| 162 | == Need of application restart == |
| 163 | |
| 164 | The transparent_hugepage/enabled values only affect future |
| 165 | behavior. So to make them effective you need to restart any |
| 166 | application that could have been using hugepages. This also applies to |
| 167 | the regions registered in khugepaged. |
| 168 | |
Mel Gorman | 6925699 | 2012-05-29 15:06:45 -0700 | [diff] [blame] | 169 | == Monitoring usage == |
| 170 | |
| 171 | The number of transparent huge pages currently used by the system is |
| 172 | available by reading the AnonHugePages field in /proc/meminfo. To |
| 173 | identify what applications are using transparent huge pages, it is |
| 174 | necessary to read /proc/PID/smaps and count the AnonHugePages fields |
| 175 | for each mapping. Note that reading the smaps file is expensive and |
| 176 | reading it frequently will incur overhead. |
| 177 | |
| 178 | There are a number of counters in /proc/vmstat that may be used to |
| 179 | monitor how successfully the system is providing huge pages for use. |
| 180 | |
| 181 | thp_fault_alloc is incremented every time a huge page is successfully |
| 182 | allocated to handle a page fault. This applies to both the |
| 183 | first time a page is faulted and for COW faults. |
| 184 | |
| 185 | thp_collapse_alloc is incremented by khugepaged when it has found |
| 186 | a range of pages to collapse into one huge page and has |
| 187 | successfully allocated a new huge page to store the data. |
| 188 | |
| 189 | thp_fault_fallback is incremented if a page fault fails to allocate |
| 190 | a huge page and instead falls back to using small pages. |
| 191 | |
| 192 | thp_collapse_alloc_failed is incremented if khugepaged found a range |
| 193 | of pages that should be collapsed into one huge page but failed |
| 194 | the allocation. |
| 195 | |
| 196 | thp_split is incremented every time a huge page is split into base |
| 197 | pages. This can happen for a variety of reasons but a common |
| 198 | reason is that a huge page is old and is being reclaimed. |
| 199 | |
| 200 | As the system ages, allocating huge pages may be expensive as the |
| 201 | system uses memory compaction to copy data around memory to free a |
| 202 | huge page for use. There are some counters in /proc/vmstat to help |
| 203 | monitor this overhead. |
| 204 | |
| 205 | compact_stall is incremented every time a process stalls to run |
| 206 | memory compaction so that a huge page is free for use. |
| 207 | |
| 208 | compact_success is incremented if the system compacted memory and |
| 209 | freed a huge page for use. |
| 210 | |
| 211 | compact_fail is incremented if the system tries to compact memory |
| 212 | but failed. |
| 213 | |
| 214 | compact_pages_moved is incremented each time a page is moved. If |
| 215 | this value is increasing rapidly, it implies that the system |
| 216 | is copying a lot of data to satisfy the huge page allocation. |
| 217 | It is possible that the cost of copying exceeds any savings |
| 218 | from reduced TLB misses. |
| 219 | |
| 220 | compact_pagemigrate_failed is incremented when the underlying mechanism |
| 221 | for moving a page failed. |
| 222 | |
| 223 | compact_blocks_moved is incremented each time memory compaction examines |
| 224 | a huge page aligned range of pages. |
| 225 | |
| 226 | It is possible to establish how long the stalls were using the function |
| 227 | tracer to record how long was spent in __alloc_pages_nodemask and |
| 228 | using the mm_page_alloc tracepoint to identify which allocations were |
| 229 | for huge pages. |
| 230 | |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 231 | == get_user_pages and follow_page == |
| 232 | |
| 233 | get_user_pages and follow_page if run on a hugepage, will return the |
| 234 | head or tail pages as usual (exactly as they would do on |
| 235 | hugetlbfs). Most gup users will only care about the actual physical |
| 236 | address of the page and its temporary pinning to release after the I/O |
| 237 | is complete, so they won't ever notice the fact the page is huge. But |
| 238 | if any driver is going to mangle over the page structure of the tail |
| 239 | page (like for checking page->mapping or other bits that are relevant |
| 240 | for the head page and not the tail page), it should be updated to jump |
| 241 | to check head page instead (while serializing properly against |
| 242 | split_huge_page() to avoid the head and tail pages to disappear from |
| 243 | under it, see the futex code to see an example of that, hugetlbfs also |
| 244 | needed special handling in futex code for similar reasons). |
| 245 | |
| 246 | NOTE: these aren't new constraints to the GUP API, and they match the |
| 247 | same constrains that applies to hugetlbfs too, so any driver capable |
| 248 | of handling GUP on hugetlbfs will also work fine on transparent |
| 249 | hugepage backed mappings. |
| 250 | |
| 251 | In case you can't handle compound pages if they're returned by |
| 252 | follow_page, the FOLL_SPLIT bit can be specified as parameter to |
| 253 | follow_page, so that it will split the hugepages before returning |
| 254 | them. Migration for example passes FOLL_SPLIT as parameter to |
| 255 | follow_page because it's not hugepage aware and in fact it can't work |
| 256 | at all on hugetlbfs (but it instead works fine on transparent |
| 257 | hugepages thanks to FOLL_SPLIT). migration simply can't deal with |
| 258 | hugepages being returned (as it's not only checking the pfn of the |
| 259 | page and pinning it during the copy but it pretends to migrate the |
| 260 | memory in regular page sizes and with regular pte/pmd mappings). |
| 261 | |
| 262 | == Optimizing the applications == |
| 263 | |
| 264 | To be guaranteed that the kernel will map a 2M page immediately in any |
| 265 | memory region, the mmap region has to be hugepage naturally |
| 266 | aligned. posix_memalign() can provide that guarantee. |
| 267 | |
| 268 | == Hugetlbfs == |
| 269 | |
| 270 | You can use hugetlbfs on a kernel that has transparent hugepage |
| 271 | support enabled just fine as always. No difference can be noted in |
| 272 | hugetlbfs other than there will be less overall fragmentation. All |
| 273 | usual features belonging to hugetlbfs are preserved and |
| 274 | unaffected. libhugetlbfs will also work fine as usual. |
| 275 | |
| 276 | == Graceful fallback == |
| 277 | |
| 278 | Code walking pagetables but unware about huge pmds can simply call |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 279 | split_huge_page_pmd(vma, addr, pmd) where the pmd is the one returned by |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 280 | pmd_offset. It's trivial to make the code transparent hugepage aware |
| 281 | by just grepping for "pmd_offset" and adding split_huge_page_pmd where |
| 282 | missing after pmd_offset returns the pmd. Thanks to the graceful |
| 283 | fallback design, with a one liner change, you can avoid to write |
| 284 | hundred if not thousand of lines of complex code to make your code |
| 285 | hugepage aware. |
| 286 | |
| 287 | If you're not walking pagetables but you run into a physical hugepage |
| 288 | but you can't handle it natively in your code, you can split it by |
| 289 | calling split_huge_page(page). This is what the Linux VM does before |
| 290 | it tries to swapout the hugepage for example. |
| 291 | |
| 292 | Example to make mremap.c transparent hugepage aware with a one liner |
| 293 | change: |
| 294 | |
| 295 | diff --git a/mm/mremap.c b/mm/mremap.c |
| 296 | --- a/mm/mremap.c |
| 297 | +++ b/mm/mremap.c |
| 298 | @@ -41,6 +41,7 @@ static pmd_t *get_old_pmd(struct mm_stru |
| 299 | return NULL; |
| 300 | |
| 301 | pmd = pmd_offset(pud, addr); |
Kirill A. Shutemov | e180377 | 2012-12-12 13:50:59 -0800 | [diff] [blame] | 302 | + split_huge_page_pmd(vma, addr, pmd); |
Andrea Arcangeli | 1c9bf22 | 2011-01-13 15:46:30 -0800 | [diff] [blame] | 303 | if (pmd_none_or_clear_bad(pmd)) |
| 304 | return NULL; |
| 305 | |
| 306 | == Locking in hugepage aware code == |
| 307 | |
| 308 | We want as much code as possible hugepage aware, as calling |
| 309 | split_huge_page() or split_huge_page_pmd() has a cost. |
| 310 | |
| 311 | To make pagetable walks huge pmd aware, all you need to do is to call |
| 312 | pmd_trans_huge() on the pmd returned by pmd_offset. You must hold the |
| 313 | mmap_sem in read (or write) mode to be sure an huge pmd cannot be |
| 314 | created from under you by khugepaged (khugepaged collapse_huge_page |
| 315 | takes the mmap_sem in write mode in addition to the anon_vma lock). If |
| 316 | pmd_trans_huge returns false, you just fallback in the old code |
| 317 | paths. If instead pmd_trans_huge returns true, you have to take the |
| 318 | mm->page_table_lock and re-run pmd_trans_huge. Taking the |
| 319 | page_table_lock will prevent the huge pmd to be converted into a |
| 320 | regular pmd from under you (split_huge_page can run in parallel to the |
| 321 | pagetable walk). If the second pmd_trans_huge returns false, you |
| 322 | should just drop the page_table_lock and fallback to the old code as |
| 323 | before. Otherwise you should run pmd_trans_splitting on the pmd. In |
| 324 | case pmd_trans_splitting returns true, it means split_huge_page is |
| 325 | already in the middle of splitting the page. So if pmd_trans_splitting |
| 326 | returns true it's enough to drop the page_table_lock and call |
| 327 | wait_split_huge_page and then fallback the old code paths. You are |
| 328 | guaranteed by the time wait_split_huge_page returns, the pmd isn't |
| 329 | huge anymore. If pmd_trans_splitting returns false, you can proceed to |
| 330 | process the huge pmd and the hugepage natively. Once finished you can |
| 331 | drop the page_table_lock. |
| 332 | |
| 333 | == compound_lock, get_user_pages and put_page == |
| 334 | |
| 335 | split_huge_page internally has to distribute the refcounts in the head |
| 336 | page to the tail pages before clearing all PG_head/tail bits from the |
| 337 | page structures. It can do that easily for refcounts taken by huge pmd |
| 338 | mappings. But the GUI API as created by hugetlbfs (that returns head |
| 339 | and tail pages if running get_user_pages on an address backed by any |
| 340 | hugepage), requires the refcount to be accounted on the tail pages and |
| 341 | not only in the head pages, if we want to be able to run |
| 342 | split_huge_page while there are gup pins established on any tail |
| 343 | page. Failure to be able to run split_huge_page if there's any gup pin |
| 344 | on any tail page, would mean having to split all hugepages upfront in |
| 345 | get_user_pages which is unacceptable as too many gup users are |
| 346 | performance critical and they must work natively on hugepages like |
| 347 | they work natively on hugetlbfs already (hugetlbfs is simpler because |
| 348 | hugetlbfs pages cannot be splitted so there wouldn't be requirement of |
| 349 | accounting the pins on the tail pages for hugetlbfs). If we wouldn't |
| 350 | account the gup refcounts on the tail pages during gup, we won't know |
| 351 | anymore which tail page is pinned by gup and which is not while we run |
| 352 | split_huge_page. But we still have to add the gup pin to the head page |
| 353 | too, to know when we can free the compound page in case it's never |
| 354 | splitted during its lifetime. That requires changing not just |
| 355 | get_page, but put_page as well so that when put_page runs on a tail |
| 356 | page (and only on a tail page) it will find its respective head page, |
| 357 | and then it will decrease the head page refcount in addition to the |
| 358 | tail page refcount. To obtain a head page reliably and to decrease its |
| 359 | refcount without race conditions, put_page has to serialize against |
| 360 | __split_huge_page_refcount using a special per-page lock called |
| 361 | compound_lock. |