Re: [PATCH/RFC] Shared page tables

["Ray Bryant" <raybry@mpdtxmail.amd.com>]

Dave,

Here's another one to keep you awake at night:

mmap a shared, anonymous region of 8MB (2MB aligned), and fork off some 
children (4 is good enough for me).    In each child, munmap() a 2 MB portion 
of the shared region (starting at offset 2MB in the shared region) and then 
mmap() a private, anonymous region in its place.   Have each child store some 
unique data in that region, sleep for a bit and then go look to see if its 
data is still there.

Under 2.6.15, this works just fine.   Under 2.6.15 + shpt patch, each child 
still points at the shared region and steps on the data of the other 
children.

I imagine the above gets even more interesting if the offset or length of the 
unmapped region is not a multiple of the 2MB alignment.

All of these test cases are for Opteron.   Your alignment may vary.

(Sorry about this... )

-- 
Ray Bryant
AMD Performance Labs                   Austin, Tx
512-602-0038 (o)                 512-507-7807 (c)