[Xen-devel] [PATCH v2 0/5] xen mapcache fixes and improvements

[Stefano Stabellini <Stefano.Stabellini@xxxxxxxxxxxxx>]

Hi all,
this patch series introduces a series of fixes and improvements to the
xen mapcache in qemu.

Changes compared to v1:
    - remove the two includes from xen-mapcache.h.


The list of patches with a diffstat follows:

Stefano Stabellini (5):
      xen: fix qemu_map_cache with size != MCACHE_BUCKET_SIZE
      xen: remove qemu_map_cache_unlock
      xen: remove xen_map_block and xen_unmap_block
      exec.c: refactor cpu_physical_memory_map
      xen: mapcache performance improvements

 cpu-common.h        |    1 +
 exec.c              |   88 ++++++++++++++++----------------
 xen-mapcache-stub.c |    8 ---
 xen-mapcache.c      |  141 +++++++++++++++++++++++---------------------------
 xen-mapcache.h      |   16 ------
 5 files changed, 109 insertions(+), 145 deletions(-)


Most of them are just straightforward cleanups and only touch Xen code.

The first patch fixes the current mapcache implementation that even
though provides a size parameter isn't actually able to cope with sizes
other than 0 or the bucket size.

The second and the third patch remove some not very useful mapcache
related functions and replace them with proper calls to qemu_map_cache
and qemu_invalidate_entry. In particular in the case of xen_map_block
and xen_unmap_block it wasn't possible before because of the size
problem describe above.

The fourth patch refactors cpu_physical_memory_map to simplify the
implementation and replace multiple calls to qemu_get_ram_ptr with a
single call to a new function that takes an address and a size a
parameters. Hopefully the changes make the function easier to understand
and more efficient.
Comments and reviews on this patch are very very welcome.

The last patch introduces few interesting performance improvements:
assuming that qemu_get_ram_ptr is only called to perform one of the
following two actions:

1) map an entire block other than the main ram block (offset == 0);

2) map a single page in the main ram block to issue a read or a write
straight after the call;

we can make the conscious decision of avoid locking the mapcache entries
for case 2).
Then considering that qemu_put_ram_ptr is called to unmap pages in the
same two cases as before, and considering that we don't need to unmap
unlocked mapcache entries, we can avoid calling qemu_invalidate_entry
completely.

A git tree with the patch series applied to the latest qemu is available
here:

git://xenbits.xen.org/people/sstabellini/qemu-dm.git mapcache_fixes_2

- Stefano

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel