Messages in this thread | | | Date | Sat, 6 Dec 2008 14:12:23 +0530 | From | Balbir Singh <> | Subject | Re: [PATCH] introduce get_mm_hiwater_xxx(), fix taskstats->hiwater_xxx accounting |
| |
* Jay Lan <jlan@sgi.com> [2008-12-05 10:41:02]:
> Hugh Dickins wrote: > > On Thu, 4 Dec 2008, Oleg Nesterov wrote: > >> On 12/03, Hugh Dickins wrote: > >>> On Wed, 3 Dec 2008, Oleg Nesterov wrote: > >>> > >>>> Unless we are going to decrease rss/vm there is no point to call the > >>>> (racy) update_hiwater_xxx() helpers. Still do_exit() does this, and > >>> I'm puzzled by this comment. exit() _is_ about to decrease rss/vm, > >>> so isn't it right to be calling update_hiwater_xxx()? > >> Do you mean exit_mm()->...->exit_mmap() ? But this doesn't matter, this > >> ->mm is going away. Nobody can read these counters when ->mm_users == 0, > >> no? > > > > I was conceding that point in my next sentence... > > > >>> There is a question of who's going to be able to see the result from > >>> this point on: I forget whether I was doing it for my own satisfaction, > >>> or for a real observer. > > > > I've looked back at it now: the hiwater fields already existed, their > > updates were slowing things down, and nothing in tree was even looking > > at the results; I changed the way they were updated to minimize the > > slowdown. > > > > So my principal reason for putting the updates in do_exit() was simply > > that there had been an update_mem_hiwater(tsk) there in do_exit(), and > > no examples of how the results would get used, so I thought I'd better > > update them there too. > > Add Jonathan Lim to cc. He is responsible for CSA. The applications > using those data are running at user space through the taskstats > interface. That was the whole point of creating taskstats interface: > move accounting modules to user space.
Yes, true and getdelays can display all the exported information.
The race does seem concerning, I would vote for keeping the update in there and disabling preemption around the update, so that hiwater cannot swing back and forth.
-- Balbir
| |