7 Page migration allows moving the physical location of pages between
8 nodes in a NUMA system while the process is running. This means that the
9 virtual addresses that the process sees do not change. However, the
10 system rearranges the physical location of those pages.
12 Also see :ref:`Heterogeneous Memory Management (HMM) <hmm>`
13 for migrating pages to or from device private memory.
15 The main intent of page migration is to reduce the latency of memory accesses
16 by moving pages near to the processor where the process accessing that memory
19 Page migration allows a process to manually relocate the node on which its
20 pages are located through the MF_MOVE and MF_MOVE_ALL options while setting
21 a new memory policy via mbind(). The pages of a process can also be relocated
22 from another process using the sys_migrate_pages() function call. The
23 migrate_pages() function call takes two sets of nodes and moves pages of a
24 process that are located on the from nodes to the destination nodes.
25 Page migration functions are provided by the numactl package by Andi Kleen
26 (a version later than 0.9.3 is required. Get it from
27 https://github.com/numactl/numactl.git). numactl provides libnuma
28 which provides an interface similar to other NUMA functionality for page
29 migration. cat ``/proc/<pid>/numa_maps`` allows an easy review of where the
30 pages of a process are located. See also the numa_maps documentation in the
33 Manual migration is useful if for example the scheduler has relocated
34 a process to a processor on a distant node. A batch scheduler or an
35 administrator may detect the situation and move the pages of the process
36 nearer to the new processor. The kernel itself only provides
37 manual page migration support. Automatic page migration may be implemented
38 through user space processes that move pages. A special function call
39 "move_pages" allows the moving of individual pages within a process.
40 For example, A NUMA profiler may obtain a log showing frequent off-node
41 accesses and may use the result to move pages to more advantageous
44 Larger installations usually partition the system using cpusets into
45 sections of nodes. Paul Jackson has equipped cpusets with the ability to
46 move pages when a task is moved to another cpuset (See
47 :ref:`CPUSETS <cpusets>`).
48 Cpusets allow the automation of process locality. If a task is moved to
49 a new cpuset then also all its pages are moved with it so that the
50 performance of the process does not sink dramatically. Also the pages
51 of processes in a cpuset are moved if the allowed memory nodes of a
54 Page migration allows the preservation of the relative location of pages
55 within a group of nodes for all migration techniques which will preserve a
56 particular memory allocation pattern generated even after migrating a
57 process. This is necessary in order to preserve the memory latencies.
58 Processes will run with similar performance after migration.
60 Page migration occurs in several steps. First a high level
61 description for those trying to use migrate_pages() from the kernel
62 (for userspace usage see the Andi Kleen's numactl package mentioned above)
63 and then a low level description of how the low level details work.
65 In kernel use of migrate_pages()
66 ================================
68 1. Remove pages from the LRU.
70 Lists of pages to be migrated are generated by scanning over
71 pages and moving them into lists. This is done by
72 calling isolate_lru_page().
73 Calling isolate_lru_page() increases the references to the page
74 so that it cannot vanish while the page migration occurs.
75 It also prevents the swapper or other scans from encountering
78 2. We need to have a function of type new_page_t that can be
79 passed to migrate_pages(). This function should figure out
80 how to allocate the correct new page given the old page.
82 3. The migrate_pages() function is called which attempts
83 to do the migration. It will call the function to allocate
84 the new page for each page that is considered for
87 How migrate_pages() works
88 =========================
90 migrate_pages() does several passes over its list of pages. A page is moved
91 if all references to a page are removable at the time. The page has
92 already been removed from the LRU via isolate_lru_page() and the refcount
93 is increased so that the page cannot be freed while page migration occurs.
97 1. Lock the page to be migrated.
99 2. Ensure that writeback is complete.
101 3. Lock the new page that we want to move to. It is locked so that accesses to
102 this (not yet up-to-date) page immediately block while the move is in progress.
104 4. All the page table references to the page are converted to migration
105 entries. This decreases the mapcount of a page. If the resulting
106 mapcount is not zero then we do not migrate the page. All user space
107 processes that attempt to access the page will now wait on the page lock
108 or wait for the migration page table entry to be removed.
110 5. The i_pages lock is taken. This will cause all processes trying
111 to access the page via the mapping to block on the spinlock.
113 6. The refcount of the page is examined and we back out if references remain.
114 Otherwise, we know that we are the only one referencing this page.
116 7. The radix tree is checked and if it does not contain the pointer to this
117 page then we back out because someone else modified the radix tree.
119 8. The new page is prepped with some settings from the old page so that
120 accesses to the new page will discover a page with the correct settings.
122 9. The radix tree is changed to point to the new page.
124 10. The reference count of the old page is dropped because the address space
125 reference is gone. A reference to the new page is established because
126 the new page is referenced by the address space.
128 11. The i_pages lock is dropped. With that lookups in the mapping
129 become possible again. Processes will move from spinning on the lock
130 to sleeping on the locked new page.
132 12. The page contents are copied to the new page.
134 13. The remaining page flags are copied to the new page.
136 14. The old page flags are cleared to indicate that the page does
137 not provide any information anymore.
139 15. Queued up writeback on the new page is triggered.
141 16. If migration entries were inserted into the page table, then replace them
142 with real ptes. Doing so will enable access for user space processes not
143 already waiting for the page lock.
145 17. The page locks are dropped from the old and new page.
146 Processes waiting on the page lock will redo their page faults
147 and will reach the new page.
149 18. The new page is moved to the LRU and can be scanned by the swapper,
152 Non-LRU page migration
153 ======================
155 Although migration originally aimed for reducing the latency of memory accesses
156 for NUMA, compaction also uses migration to create high-order pages.
158 Current problem of the implementation is that it is designed to migrate only
159 *LRU* pages. However, there are potential non-LRU pages which can be migrated
160 in drivers, for example, zsmalloc, virtio-balloon pages.
162 For virtio-balloon pages, some parts of migration code path have been hooked
163 up and added virtio-balloon specific functions to intercept migration logics.
164 It's too specific to a driver so other drivers who want to make their pages
165 movable would have to add their own specific hooks in the migration path.
167 To overcome the problem, VM supports non-LRU page migration which provides
168 generic functions for non-LRU movable pages without driver specific hooks
169 in the migration path.
171 If a driver wants to make its pages movable, it should define three functions
172 which are function pointers of struct address_space_operations.
174 1. ``bool (*isolate_page) (struct page *page, isolate_mode_t mode);``
176 What VM expects from isolate_page() function of driver is to return *true*
177 if driver isolates the page successfully. On returning true, VM marks the page
178 as PG_isolated so concurrent isolation in several CPUs skip the page
179 for isolation. If a driver cannot isolate the page, it should return *false*.
181 Once page is successfully isolated, VM uses page.lru fields so driver
182 shouldn't expect to preserve values in those fields.
184 2. ``int (*migratepage) (struct address_space *mapping,``
185 | ``struct page *newpage, struct page *oldpage, enum migrate_mode);``
187 After isolation, VM calls migratepage() of driver with the isolated page.
188 The function of migratepage() is to move the contents of the old page to the
190 and set up fields of struct page newpage. Keep in mind that you should
191 indicate to the VM the oldpage is no longer movable via __ClearPageMovable()
192 under page_lock if you migrated the oldpage successfully and returned
193 MIGRATEPAGE_SUCCESS. If driver cannot migrate the page at the moment, driver
194 can return -EAGAIN. On -EAGAIN, VM will retry page migration in a short time
195 because VM interprets -EAGAIN as "temporary migration failure". On returning
196 any error except -EAGAIN, VM will give up the page migration without
199 Driver shouldn't touch the page.lru field while in the migratepage() function.
201 3. ``void (*putback_page)(struct page *);``
203 If migration fails on the isolated page, VM should return the isolated page
204 to the driver so VM calls the driver's putback_page() with the isolated page.
205 In this function, the driver should put the isolated page back into its own data
208 4. non-LRU movable page flags
210 There are two page flags for supporting non-LRU movable page.
214 Driver should use the function below to make page movable under page_lock::
216 void __SetPageMovable(struct page *page, struct address_space *mapping)
218 It needs argument of address_space for registering migration
219 family functions which will be called by VM. Exactly speaking,
220 PG_movable is not a real flag of struct page. Rather, VM
221 reuses the page->mapping's lower bits to represent it::
223 #define PAGE_MAPPING_MOVABLE 0x2
224 page->mapping = page->mapping | PAGE_MAPPING_MOVABLE;
226 so driver shouldn't access page->mapping directly. Instead, driver should
227 use page_mapping() which masks off the low two bits of page->mapping under
228 page lock so it can get the right struct address_space.
230 For testing of non-LRU movable pages, VM supports __PageMovable() function.
231 However, it doesn't guarantee to identify non-LRU movable pages because
232 the page->mapping field is unified with other variables in struct page.
233 If the driver releases the page after isolation by VM, page->mapping
234 doesn't have a stable value although it has PAGE_MAPPING_MOVABLE set
235 (look at __ClearPageMovable). But __PageMovable() is cheap to call whether
236 page is LRU or non-LRU movable once the page has been isolated because LRU
237 pages can never have PAGE_MAPPING_MOVABLE set in page->mapping. It is also
238 good for just peeking to test non-LRU movable pages before more expensive
239 checking with lock_page() in pfn scanning to select a victim.
241 For guaranteeing non-LRU movable page, VM provides PageMovable() function.
242 Unlike __PageMovable(), PageMovable() validates page->mapping and
243 mapping->a_ops->isolate_page under lock_page(). The lock_page() prevents
244 sudden destroying of page->mapping.
246 Drivers using __SetPageMovable() should clear the flag via
247 __ClearMovablePage() under page_lock() before the releasing the page.
251 To prevent concurrent isolation among several CPUs, VM marks isolated page
252 as PG_isolated under lock_page(). So if a CPU encounters PG_isolated
253 non-LRU movable page, it can skip it. Driver doesn't need to manipulate the
254 flag because VM will set/clear it automatically. Keep in mind that if the
255 driver sees a PG_isolated page, it means the page has been isolated by the
256 VM so it shouldn't touch the page.lru field.
257 The PG_isolated flag is aliased with the PG_reclaim flag so drivers
258 shouldn't use PG_isolated for its own purposes.
261 =====================
263 The following events (counters) can be used to monitor page migration.
265 1. PGMIGRATE_SUCCESS: Normal page migration success. Each count means that a
266 page was migrated. If the page was a non-THP page, then this counter is
267 increased by one. If the page was a THP, then this counter is increased by
268 the number of THP subpages. For example, migration of a single 2MB THP that
269 has 4KB-size base pages (subpages) will cause this counter to increase by
272 2. PGMIGRATE_FAIL: Normal page migration failure. Same counting rules as for
273 PGMIGRATE_SUCCESS, above: this will be increased by the number of subpages,
276 3. THP_MIGRATION_SUCCESS: A THP was migrated without being split.
278 4. THP_MIGRATION_FAIL: A THP could not be migrated nor it could be split.
280 5. THP_MIGRATION_SPLIT: A THP was migrated, but not as such: first, the THP had
281 to be split. After splitting, a migration retry was used for it's sub-pages.
283 THP_MIGRATION_* events also update the appropriate PGMIGRATE_SUCCESS or
284 PGMIGRATE_FAIL events. For example, a THP migration failure will cause both
285 THP_MIGRATION_FAIL and PGMIGRATE_FAIL to increase.
287 Christoph Lameter, May 8, 2006.
288 Minchan Kim, Mar 28, 2016.