[PATCH 0/5] Direct Migration V9: Overview

[Christoph Lameter]

Swap migration is now in Linus tree. So this is the first direct page
migration patchset against his tree (2.6.15-git6 no changes apart
from the rediff). Also done on the off chance that we decide to have
the full thing in 2.6.16 instead of only swap migration.
Maybe this can now get into Andrew's tree?

-----

Page migration allows the moving of the physical location of pages between
nodes in a numa system while the process is running. This means that the
virtual addresses that the process sees do not change. However, the
system rearranges the physical location of those pages.

The main intend of page migration patches here is to reduce the latency of
memory access by moving pages near to the processor where the process
accessing that memory is running.

The migration patchsets allow a process to manually relocate the node
on which its pages are located through the MF_MOVE and MF_MOVE_ALL options
while setting a new memory policy. The pages of process can also be relocated
from another process using the sys_migrate_pages() function call. The
migrate_pages function call takes two sets of nodes and moves pages of a
process that are located on the from nodes to the destination nodes.

Manual migration is very useful if for example the scheduler has relocated
a process to a processor on a distant node. A batch scheduler or an
administrator can detect  the situation and move the pages of the process
nearer to the new processor.

Larger installations usually partition the system using cpusets into
sections of nodes. Paul Jackson has  equipped cpusets with the ability to
move pages when a task is moved to another cpuset. This allows automatic
control over locality of a process. If a task is moved to a new cpuset
then also all its pages are moved with it so that the performance of the
process does not sink dramatically (as is the case today).

The swap migration patchset in 2.6.16-git6 works by simply evicting
the page. The pages must be faulted back in. The pages are then typically
reallocated by the system near the node where the process is executing.
For swap migration the destination of the move is controlled by the
allocation policy. Cpusets set the allocation policy before calling
sys_migrate_pages() in order to move the pages as intended.

No allocation policy changes are performed for sys_migrate_pages(). This
means that the pages may not faulted in to the specified nodes if no
allocation policy was set by other means. The pages will just end up
near the node where the fault occurred.  The direct migration patchset
extends the migration functionality to avoid going through swap.
The destination node of the relation is controllable during the actual
moving of pages. The crutch of using the allocation policy to relocate
is not necessary any and the pages are moved directly to the target.

The direct migration patchset allows the preservation of the relative
location of pages within a group of nodes for all migration techniques which
will preserve a particular memory allocation pattern generated even after
migrating a process. This is necessary in order to preserve the memory
latencies. Processes will run with similar performance after migration.

This patch makes sys_migrate_pages() finally work as intended but does not
do any significant modifications to APIs.

Benefits over swap migration:

1. It makes migrates_pages() actually migrate pages instead of just
   swapping pages from a set of nodes out.

2. Its faster because the page does not need to be written to swap space.

3. It does not use swap space and therefore there is no danger of running
   out of swap space.

4. The need to write back a dirty page before migration is avoided through
   a file system specific method.

5. Direct migration allows the preservation of the relative location of a page
   within a set of nodes.

Many of the ideas for this code were originally developed in the memory
hotplug project and we hope that this code also will allow the hotplug
project to build on this patch in order to get to their goals.

The patchset consists of five patches (only the first two are necessary to
have basic direct migration support):

1. SwapCache patch

   SwapCache pages may have changed their type after lock_page() if the
   page was migrated. Check for this and retry lookup if the page is no
   longer a SwapCache page.

2. migrate_pages()

   Basic direct migration with fallback to swap if all other attempts
   fail.

3. remove_from_swap()

   Page migration installs swap ptes for anonymous pages in order to
   preserve the information contained in the page tables. This patch
   removes the swap ptes after migration and replaces them with regular
   ptes.

4. upgrade of MPOL_MF_MOVE and sys_migrate_pages()

   Add logic to mm/mempolicy.c to allow the policy layer to control
   direct page migration. Thanks to Paul Jackson for the interative
   logic to move between sets of nodes.

5. buffer_migrate_pages() patch

   Allow migration without writing back dirty pages. Add filesystem dependent
   migration support for ext2/ext3 and xfs. Use swapper space to setup a
   method to migrate anonymous pages without writeback.

Credits (also in mm/vmscan.c):

The idea for this scheme of page migration was first developed in the context
of the memory hotplug project. The main authors of the migration code from
the memory hotplug project are:

IWAMOTO Toshihiro <iwamoto@xxxxxxxxxxxxx>
Hirokazu Takahashi <taka@xxxxxxxxxxxxx>
Dave Hansen <haveblue@xxxxxxxxxx>

Changes V8->V9:
- Patchset against 2.6.15-git6

Changes V7->V8:
- Patchset against 2.6.15-rc5-mm3
- Export more functions so that filesystems are able to implement their own
  migrate_page() function.
- Fix remove_from_swap() to remove the page from the swap cache in addition
  to replacing swap ptes. Call with the page lock on the new page.
- Fix copying of struct page {} field to avoid using the macros that process
  field information.

Changes V7->V7:
- Rediff against 2.6.14-rc5-mm2

Changes V6->V7:
- Patchset agsinst 2.6.15-rc5-mm1
- Fix one occurence of page->mapping in migrate_page_remove_references()
- Update description]

Changes V5->V6:
- Patchset against 2.6.15-rc3-mm1
- Remove checks for page count increases while migrating after Andrew assured
  me that this cannot happen. Revise documentation to reflect that. If this is
  the case then we will have no need to include the unwind code from the
  hotplug project in the future.
- Wrong reference while calling remove_from_swap to page instead of newpage
  fixed.

Changes V4->V5:
- Patchset against 2.6.15-rc2-mm1
- Update policy layer patch to use the generic check_range in 2.6.15-rc2-mm1.
- Remove try_to_unmap patch since VM_RESERVED vanished under us and therefore
  there is no point anymore to distinguish between permament and transitional
  failures.

Changes V3->V4:
- Patchset against 2.6.15-rc1-mm2 + two swap migration fixes posted today.
- Remove what is already in 2.6.14-rc1-mm2 which results in a significant
  cleanup of the code.

Changes V2->V3:
- Patchset against 2.6.14-mm2
- Fix single processor build and builds without CONFIG_MIGRATION
- export symbols for filesystems that are modules and for
  modules using migrate_pages().
- Paul Jackson's cpuset migration support is in 2.6.14-mm2 so
  this patchset can be easily applied to -mm2 to get from swap
  based to direct page migration.

Changes V1->V2:
- Call node_remap with the right parameters in do_migrate_pages().
- Take radix tree lock while examining page count to avoid races with
  find_get_page() and various *_get_pages based on it.
- Convert direct ptes to swap ptes before radix tree update to avoid
  more races.
- Fix problem if CONFIG_MIGRATION is off for buffer_migrate_page
- Add documentation about page migration
- Change migrate_pages() api so that the caller can decide what
  to do about the migrated pages (badmem handling and hotplug
  have to remove those pages for good).
- Drop config patch (already in mm)
- Add try_to_unmap patch
- Patchset now against 2.6.14-mm1 without requiring additional patches.


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/