lkml.org 
[lkml]   [2011]   [Jun]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v4 3.0-rc2-tip 2/22] 2: uprobes: Breakground page replacement.
On 06/15, Srikar Dronamraju wrote:
>
> > > > > +
> > > > > + /* Read the page with vaddr into memory */
> > > > > + ret = get_user_pages(tsk, tsk->mm, vaddr, 1, 1, 1, &old_page, &vma);
> > > >
> > > > Sorry if this was already discussed... But why we are using FOLL_WRITE here?
> > > > We are not going to write into this page, and this provokes the unnecessary
> > > > cow, no?
> > >
> > > Yes, We are not going to write to the page returned by get_user_pages
> > > but a copy of that page.
> >
> > Yes I see. But the page returned by get_user_pages(write => 1) is already
> > a cow'ed copy (this mapping should be read-only).
> >
> > > The idea was if we cow the page then we dont
> > > need to cow it at the replace_page time
> >
> > Yes, replace_page() shouldn't cow.
> >
> > > and since get_user_pages knows
> > > the right way to cow the page, we dont have to write another routine to
> > > cow the page.
> >
> > Confused. write_opcode() allocs another page and does memcpy. This is
> > correct, but I don't understand the first cow.
> >
>
> we decided on get_user_pages(FOLL_WRITE|FOLL_FORCE) based on discussions
> in these threads https://lkml.org/lkml/2010/4/23/327 and
> https://lkml.org/lkml/2010/5/12/119

Failed to Connect.

> Summary of those two sub-threads as I understand was to have
> get_user_pages do the "real" cow for us.
>
> If I understand correctly, your concern is on the extra overhead added
> by the get_user_pages.

No. My main concern is that I do not understand why do we need an extra cow.
This is fine, I am not vm expert. But I think it is not fine that you can't
explain why your code needs it ;)

What this 'get_user_pages do the "real" cow for us' actually means? It does
not do cow for us, __replace_page() does the actual/final cow. It re-installs
the modified copy of the page returned by get_user_pages() at the same pte.

> > Probably I missed something... but could you please explain why we can't
> >
> > - ret = get_user_pages(tsk, tsk->mm, vaddr, 1, 1, 1, &old_page, &vma);
> > + ret = get_user_pages(tsk, tsk->mm, vaddr, 1, 0, 0, &old_page, &vma);
> >
> > ?
>
> I tried the code with this change and it works for regular cases.
> I am not sure if it affects cases where programs do mprotect

Hmm... How can mprotect make a difference? This mapping should be read
only, and we are not going to do pte_mkwrite.

> So I am okay to not force cow through get_user_pages.

I am okay either way ;) But, imho, if we use FOLL_WRITE|FOLL_FORCE then
it would be nice to document why it this needed.

Oleg.



\
 
 \ /
  Last update: 2011-06-15 19:59    [W:0.092 / U:2.876 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site