lkml.org 
[lkml]   [2002]   [Feb]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [RFC] Page table sharing
Date
On February 19, 2002 04:22 am, Linus Torvalds wrote:
> On Mon, 18 Feb 2002, Linus Torvalds wrote:
> >
> > We can, of course, introduce a "pmd-rmap" thing, with a pointer to a
> > circular list of all mm's using that pmd inside the "struct page *" of the
> > pmd. Right now the rmap patches just make the pointer point directly to
> > the one exclusive mm that holds the pmd, right?
>
> There's another approach:
> - get rid of "page_table_lock"
> - replace it with a "per-pmd lock"
> - notice that we already _have_ such a lock
>
> The lock we have is the lock that we've always had in "struct page".

Yes, I even have an earlier version of the patch that implements a spinlock
on that bit. It doesn't use the normal lock_page of course.

> There are some interesting advantages from this:
> - we allow even more parallelism from threads across different CPU's.
> - we already have the cacheline for the pmd "struct page" because we
> needed it for the pmd count.Y
>
> That still leaves the TLB invalidation issue, but we could handle that
> with an alternate approach: use the same "free_pte_ctx" kind of gathering
> that the zap_page_range() code uses for similar reasons (ie gather up the
> pte entries that you're going to free first, and then do a global
> invalidate later).
>
> Note that this is likely to speed things up anyway (whether the pages are
> gathered by rmap or by the current linear walk), by virtue of being able
> to do just _one_ TLB invalidate (potentially cross-CPU) rather than having
> to do it once for each page we free.
>
> At that point you might as well make the TLB shootdown global (ie you keep
> track of a mask of CPU's whose TLB's you want to kill, and any pmd that
> has count > 1 just makes that mask be "all CPU's").
>
> I'm a bit worried about the "lock each mm on the pmd-rmap list" approach,
> because I think we need to lock them _all_ to be safe (as opposed to
> locking them one at a time), which always implies all the nasty potential
> deadlocks you get for doing multiple locking.
>
> The "page-lock + potentially one global TLB flush" approach looks a lot
> safer in this respect.

How do we know when to do the global tlb flush?

--
Daniel
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:24    [W:0.284 / U:0.112 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site