Re: Deadlock on the mm->mmap_sem

["Manfred Spraul" <manfred@colorfullife.com>]

>
> I also don't think the hack is that bad. All it's doing is taking a
> copy of the process's VM decription so that it knows that
> nobody is going to modify it whilst a coredump is in progress.

You break the locking scheme of the mm structure.
Right now the rules are

1 get a mm_struct pointer by whatever means (walk the process list and
    read task->mm, walk the mm_list)
2 increase mm_users
3 release the spinlock you acquired for 1
4 you can do with the result what you want.

With your patch applied, we would have to restrict rule 4 - at least
modifying the vma list is not possible anymore, probably further
changes.
AFAIK right now no external mm_struct user modifies the vma list, but it
could be a problem in the future.

>
> However, if you don't like that, how about just changing the lock on
> mm_struct to a special mm_struct-only type lock that has a
> recursive lock operation for use by the pagefault handler (and
> _only_ the pagefault handler)? I've attached  a patch to do just that.
> This introduces five operations:

Does that solve the latency problem? That problem is pagefaults vs.
another operation.

--
    Manfred