Messages in this thread | | | Date | Thu, 29 Jan 2009 10:46:04 -0800 (PST) | From | Davide Libenzi <> | Subject | Re: [patch] drop epoll max_user_instances and rely only on max_user_watches |
| |
On Thu, 29 Jan 2009, Andrew Morton wrote:
> > > Subject: [patch] drop epoll max_user_instances and rely only on max_user_watches > > nanonit: please prepare titles in the form "subsystem-id: > what-i-did-to-it", so a suitable name here would be > > epoll: drop max_user_instances and rely only on max_user_watches
Sorry, I forgot, again. I sent those by hand, while I have that logic in my scripts.
> On Wed, 28 Jan 2009 20:56:07 -0800 (PST) Davide Libenzi <davidel@xmailserver.org> wrote: > > > Linus suggested to put limits where the money is, and max_user_watches > > already does that w/out the need of max_user_instances. That has the > > advantage to mitigate the potential DoS while allowing pretty generous > > default behavior. > > A reader of this changelog would be wondering what this DoS is.
for i 1..1021 fd[i] = epoll_create() for i 1..1021 for j 1..1021, j != i epoll_ctl(fd[i], ADD, fd[j])
I'm not sure we want to advertise it too much though.
> > Allowing top 4% of low memory (per user) to be allocated in epoll > > watches, we have: > > > > LOMEM MAX_WATCHES (per user) > > 512MB ~178000 > > 1GB ~356000 > > 2GB ~712000 > > > > A box with 512MB of lomem, will meet some challenge in hitting 180K > > watches, socket buffers math teaches us. > > No more max_user_instances limits then. > > So the max consumable memory is > > number-of-users * max_user_watches * sizeof(whatever) > > ? > > So if enough users gang up (or if one person has access to a lot of > UIDs), there's still a DoS? > > I suspect we can live with that.
I think so, given that we've seen that putting too restrictive limits creates more problems. A similar thing, even though milder, exist on sys_poll too. sys_poll eats (on a 32bit box) 28 bytes of kernel memory for each 12 bytes of user memory, worse figures on a 64bit box.
> I assume that because you based all this on all the other patches, you > view it as 2.6.30 material? > > > @@ -581,10 +570,6 @@ > > please use `diff -p'. It helps.
Need to add to my quilt diff options :)
> The code I have is > > if (error < 0) > ep_free(ep); > else > atomic_inc(&ep->user->epoll_devs); > > so I obviously nuked the `else' as well.
That's the right thing.
- Davide
| |