1Page migration
  2--------------
  3
  4Page migration allows the moving of the physical location of pages between
  5nodes in a numa system while the process is running. This means that the
  6virtual addresses that the process sees do not change. However, the
  7system rearranges the physical location of those pages.
  8
  9The main intend of page migration is to reduce the latency of memory access
 10by moving pages near to the processor where the process accessing that memory
 11is running.
 12
 13Page migration allows a process to manually relocate the node on which its
 14pages are located through the MF_MOVE and MF_MOVE_ALL options while setting
 15a new memory policy via mbind(). The pages of process can also be relocated
 16from another process using the sys_migrate_pages() function call. The
 17migrate_pages function call takes two sets of nodes and moves pages of a
 18process that are located on the from nodes to the destination nodes.
 19Page migration functions are provided by the numactl package by Andi Kleen
 20(a version later than 0.9.3 is required. Get it from
 21ftp://oss.sgi.com/www/projects/libnuma/download/). numactl provides libnuma
 22which provides an interface similar to other numa functionality for page
 23migration.  cat /proc/<pid>/numa_maps allows an easy review of where the
 24pages of a process are located. See also the numa_maps documentation in the
 25proc(5) man page.
 26
 27Manual migration is useful if for example the scheduler has relocated
 28a process to a processor on a distant node. A batch scheduler or an
 29administrator may detect the situation and move the pages of the process
 30nearer to the new processor. The kernel itself does only provide
 31manual page migration support. Automatic page migration may be implemented
 32through user space processes that move pages. A special function call
 33"move_pages" allows the moving of individual pages within a process.
 34A NUMA profiler may f.e. obtain a log showing frequent off node
 35accesses and may use the result to move pages to more advantageous
 36locations.
 37
 38Larger installations usually partition the system using cpusets into
 39sections of nodes. Paul Jackson has equipped cpusets with the ability to
 40move pages when a task is moved to another cpuset (See
 41Documentation/cgroups/cpusets.txt).
 42Cpusets allows the automation of process locality. If a task is moved to
 43a new cpuset then also all its pages are moved with it so that the
 44performance of the process does not sink dramatically. Also the pages
 45of processes in a cpuset are moved if the allowed memory nodes of a
 46cpuset are changed.
 47
 48Page migration allows the preservation of the relative location of pages
 49within a group of nodes for all migration techniques which will preserve a
 50particular memory allocation pattern generated even after migrating a
 51process. This is necessary in order to preserve the memory latencies.
 52Processes will run with similar performance after migration.
 53
 54Page migration occurs in several steps. First a high level
 55description for those trying to use migrate_pages() from the kernel
 56(for userspace usage see the Andi Kleen's numactl package mentioned above)
 57and then a low level description of how the low level details work.
 58
 59A. In kernel use of migrate_pages()
 60-----------------------------------
 61
 621. Remove pages from the LRU.
 63
 64   Lists of pages to be migrated are generated by scanning over
 65   pages and moving them into lists. This is done by
 66   calling isolate_lru_page().
 67   Calling isolate_lru_page increases the references to the page
 68   so that it cannot vanish while the page migration occurs.
 69   It also prevents the swapper or other scans to encounter
 70   the page.
 71
 722. We need to have a function of type new_page_t that can be
 73   passed to migrate_pages(). This function should figure out
 74   how to allocate the correct new page given the old page.
 75
 763. The migrate_pages() function is called which attempts
 77   to do the migration. It will call the function to allocate
 78   the new page for each page that is considered for
 79   moving.
 80
 81B. How migrate_pages() works
 82----------------------------
 83
 84migrate_pages() does several passes over its list of pages. A page is moved
 85if all references to a page are removable at the time. The page has
 86already been removed from the LRU via isolate_lru_page() and the refcount
 87is increased so that the page cannot be freed while page migration occurs.
 88
 89Steps:
 90
 911. Lock the page to be migrated
 92
 932. Insure that writeback is complete.
 94
 953. Lock the new page that we want to move to. It is locked so that accesses to
 96   this (not yet uptodate) page immediately lock while the move is in progress.
 97
 984. All the page table references to the page are converted to migration
 99   entries. This decreases the mapcount of a page. If the resulting
100   mapcount is not zero then we do not migrate the page. All user space
101   processes that attempt to access the page will now wait on the page lock.
102
1035. The radix tree lock is taken. This will cause all processes trying
104   to access the page via the mapping to block on the radix tree spinlock.
105
1066. The refcount of the page is examined and we back out if references remain
107   otherwise we know that we are the only one referencing this page.
108
1097. The radix tree is checked and if it does not contain the pointer to this
110   page then we back out because someone else modified the radix tree.
111
1128. The new page is prepped with some settings from the old page so that
113   accesses to the new page will discover a page with the correct settings.
114
1159. The radix tree is changed to point to the new page.
116
11710. The reference count of the old page is dropped because the radix tree
118    reference is gone. A reference to the new page is established because
119    the new page is referenced to by the radix tree.
120
12111. The radix tree lock is dropped. With that lookups in the mapping
122    become possible again. Processes will move from spinning on the tree_lock
123    to sleeping on the locked new page.
124
12512. The page contents are copied to the new page.
126
12713. The remaining page flags are copied to the new page.
128
12914. The old page flags are cleared to indicate that the page does
130    not provide any information anymore.
131
13215. Queued up writeback on the new page is triggered.
133
13416. If migration entries were page then replace them with real ptes. Doing
135    so will enable access for user space processes not already waiting for
136    the page lock.
137
13819. The page locks are dropped from the old and new page.
139    Processes waiting on the page lock will redo their page faults
140    and will reach the new page.
141
14220. The new page is moved to the LRU and can be scanned by the swapper
143    etc again.
144
145Christoph Lameter, May 8, 2006.
146