Based on kernel version 2.6.26. Page generated on 2008-07-16 21:13 EST.
1 Page migration 2 -------------- 3 4 Page migration allows the moving of the physical location of pages between 5 nodes in a numa system while the process is running. This means that the 6 virtual addresses that the process sees do not change. However, the 7 system rearranges the physical location of those pages. 8 9 The main intend of page migration is to reduce the latency of memory access 10 by moving pages near to the processor where the process accessing that memory 11 is running. 12 13 Page migration allows a process to manually relocate the node on which its 14 pages are located through the MF_MOVE and MF_MOVE_ALL options while setting 15 a new memory policy via mbind(). The pages of process can also be relocated 16 from another process using the sys_migrate_pages() function call. The 17 migrate_pages function call takes two sets of nodes and moves pages of a 18 process that are located on the from nodes to the destination nodes. 19 Page migration functions are provided by the numactl package by Andi Kleen 20 (a version later than 0.9.3 is required. Get it from 21 ftp://ftp.suse.com/pub/people/ak). numactl provided libnuma which 22 provides an interface similar to other numa functionality for page migration. 23 cat /proc/<pid>/numa_maps allows an easy review of where the pages of 24 a process are located. See also the numa_maps manpage in the numactl package. 25 26 Manual migration is useful if for example the scheduler has relocated 27 a process to a processor on a distant node. A batch scheduler or an 28 administrator may detect the situation and move the pages of the process 29 nearer to the new processor. The kernel itself does only provide 30 manual page migration support. Automatic page migration may be implemented 31 through user space processes that move pages. A special function call 32 "move_pages" allows the moving of individual pages within a process. 33 A NUMA profiler may f.e. obtain a log showing frequent off node 34 accesses and may use the result to move pages to more advantageous 35 locations. 36 37 Larger installations usually partition the system using cpusets into 38 sections of nodes. Paul Jackson has equipped cpusets with the ability to 39 move pages when a task is moved to another cpuset (See ../cpusets.txt). 40 Cpusets allows the automation of process locality. If a task is moved to 41 a new cpuset then also all its pages are moved with it so that the 42 performance of the process does not sink dramatically. Also the pages 43 of processes in a cpuset are moved if the allowed memory nodes of a 44 cpuset are changed. 45 46 Page migration allows the preservation of the relative location of pages 47 within a group of nodes for all migration techniques which will preserve a 48 particular memory allocation pattern generated even after migrating a 49 process. This is necessary in order to preserve the memory latencies. 50 Processes will run with similar performance after migration. 51 52 Page migration occurs in several steps. First a high level 53 description for those trying to use migrate_pages() from the kernel 54 (for userspace usage see the Andi Kleen's numactl package mentioned above) 55 and then a low level description of how the low level details work. 56 57 A. In kernel use of migrate_pages() 58 ----------------------------------- 59 60 1. Remove pages from the LRU. 61 62 Lists of pages to be migrated are generated by scanning over 63 pages and moving them into lists. This is done by 64 calling isolate_lru_page(). 65 Calling isolate_lru_page increases the references to the page 66 so that it cannot vanish while the page migration occurs. 67 It also prevents the swapper or other scans to encounter 68 the page. 69 70 2. We need to have a function of type new_page_t that can be 71 passed to migrate_pages(). This function should figure out 72 how to allocate the correct new page given the old page. 73 74 3. The migrate_pages() function is called which attempts 75 to do the migration. It will call the function to allocate 76 the new page for each page that is considered for 77 moving. 78 79 B. How migrate_pages() works 80 ---------------------------- 81 82 migrate_pages() does several passes over its list of pages. A page is moved 83 if all references to a page are removable at the time. The page has 84 already been removed from the LRU via isolate_lru_page() and the refcount 85 is increased so that the page cannot be freed while page migration occurs. 86 87 Steps: 88 89 1. Lock the page to be migrated 90 91 2. Insure that writeback is complete. 92 93 3. Prep the new page that we want to move to. It is locked 94 and set to not being uptodate so that all accesses to the new 95 page immediately lock while the move is in progress. 96 97 4. The new page is prepped with some settings from the old page so that 98 accesses to the new page will discover a page with the correct settings. 99 100 5. All the page table references to the page are converted 101 to migration entries or dropped (nonlinear vmas). 102 This decrease the mapcount of a page. If the resulting 103 mapcount is not zero then we do not migrate the page. 104 All user space processes that attempt to access the page 105 will now wait on the page lock. 106 107 6. The radix tree lock is taken. This will cause all processes trying 108 to access the page via the mapping to block on the radix tree spinlock. 109 110 7. The refcount of the page is examined and we back out if references remain 111 otherwise we know that we are the only one referencing this page. 112 113 8. The radix tree is checked and if it does not contain the pointer to this 114 page then we back out because someone else modified the radix tree. 115 116 9. The radix tree is changed to point to the new page. 117 118 10. The reference count of the old page is dropped because the radix tree 119 reference is gone. A reference to the new page is established because 120 the new page is referenced to by the radix tree. 121 122 11. The radix tree lock is dropped. With that lookups in the mapping 123 become possible again. Processes will move from spinning on the tree_lock 124 to sleeping on the locked new page. 125 126 12. The page contents are copied to the new page. 127 128 13. The remaining page flags are copied to the new page. 129 130 14. The old page flags are cleared to indicate that the page does 131 not provide any information anymore. 132 133 15. Queued up writeback on the new page is triggered. 134 135 16. If migration entries were page then replace them with real ptes. Doing 136 so will enable access for user space processes not already waiting for 137 the page lock. 138 139 19. The page locks are dropped from the old and new page. 140 Processes waiting on the page lock will redo their page faults 141 and will reach the new page. 142 143 20. The new page is moved to the LRU and can be scanned by the swapper 144 etc again. 145 146 Christoph Lameter, May 8, 2006.