x86: Speed up PV-guest superpage mapping

The current version of superpage mapping takes a PGT_writable
reference to every page in a superpage each time it is mapped.  This
is extremely slow, so slow that applications become unusable.

My solution for this is to introduce a superpage table in the
hypervisor, similar to the frametable structure for pages.  Currently
this table only has a type_info element.  There are three types a
superpage can have, SGT_mark, SGT_dynamic, or SGT_none.

In normal operation, the first time a superpage is mapped, a
PGT_writable reference is taken to each page in the superpage, and the
superpage is set to type SGT_dynamic and the superpage typecount is
incremented.  On subsequent mappings and unmappings, only the
superpage typecount changes.  On the last unmap, the PGT_writable
reference on each page is removed.

The SGT_mark type is set and cleared through two new MMUEXT
hypercalls, mark_super and unmark_super.  When the hypercall is made,
the superpage's type is set to SGT_mark and a PGT_writable reference
is taken to its pages.  On unmark, the type is cleared and the
reference removed.

If a page is already set to SGT_dynamic when mark_super is called, the
type is changed to SGT_mark and no additional PGT_writable reference
is taken.  If there are still outstanding mappings of this superpage
when unmark_super is called, the type is set to SGT_dynamic and the
PGT_writable reference is not removed.

Fast superpage mapping is only supported on 64 bit hypervisors.  For
32 bit hyperviors, superpage mapping is supported but will be
extremely slow.

Signed-off-by: Dave McCracken <dave.mccracken@oracle.com>

0 files changed, 0 insertions, 0 deletions