Hugh Dickens was right. mremap was broken wrt shared page tables. Here's
the fix.
Dave McCracken
======================================================================
Dave McCracken IBM Linux Base Kernel Team 1-512-838-3059
[email protected] T/L 678-3059
On Tue, 15 Oct 2002, Dave McCracken wrote:
>
> Hugh Dickens was right. mremap was broken wrt shared page tables.
i
Sorry, Dave, I only mentioned mremap as one _example_. You have
changed the locking rules for ptes in page tables in two or three
sources, but there's a number more which need to be updated too.
Please don't ask me to list which, it's a matter of looking at all
the uses of ->page_table_lock and whether those places need to be
changed to fit the new locking scheme - some of the uses are quite
irrelevant, some may be obscure, some may be historic, some must
be changed.
I wouldn't mind if it was just broken wrt shared page tables,
CONFIG_SHAREPTE is easily made experimental. But the locking rules
have (rightly) been changed in all cases, so -mm tree is now unsafe.
A stopgap measure might be to hold on to page_table_lock for as
long as before, then fix sources up piecemeal - but that approach
would still leave CONFIG_SHAREPTE y broken for a while.
Let me say again, that I have _not_ actually observed any problem
yet: we're talking about races and correctness, I don't mean to be
alarmist.
Hugh