On Mar 7, 2012, at 7:25 PM, David Gibson wrote:
On Sat, Mar 03, 2012 at 10:39:34AM -0600, Nathan Whitehorn wrote:
Fix large page support in TCG. The old code would overwrite the
large page table entry with the fake 4 KB
one generated here whenever the ref/change bits were updated,
causing it to point to the wrong area of memory. Instead of creating
a fake PTE, just update the real address at the end.
Signed-off-by: Nathan Whitehorn<address@hidden>
Hrm. This looks like a cleaner way of handling things, but I don't
really follow what exactly was going wrong in the old way. Can you
spell out in more detail where the modified pte1 value caused
problems?
The problem was that pte1 would get extra bits added into it in
_find_pte() to produce a new, fake 4KB page table entry. When the
ref/change bits were updated, pte1 would be written back to the page
table -- *including* the bits added to make a fake 4K page. At the
next access, since this function does not clear the low bits of
large pages (which is probably itself a bug) when it interprets
them, the generated address would be the large page base, ored with
the large page remainder for this access, ored with the large page
remainder from the *previous* access, etc. and you would get a
progressively more bogus address in the end.
0005-Fix-large-page-support-in-TCG.-The-old-code-would-ov.patch