Messages in this thread | | | Date | Mon, 29 Nov 2010 12:33:57 +0100 | From | Oleg Nesterov <> | Subject | Re: [resend][PATCH 4/4] oom: don't ignore rss in nascent mm |
| |
On 11/29, KOSAKI Motohiro wrote: > > > The patch is not complete, compat_copy_strings() needs changes. > > But, shouldn't it use get_arg_page() too? Otherwise, where do > > we check RLIMIT_STACK? > > Because NOMMU doesn't have variable length argv. Instead it is still > using MAX_ARG_STRLEN as old MMU code. > > 32 pages hard coded argv limitation naturally prevent this nascent mm > issue.
Ah, I didn't mean NOMMU. I meant compat_execve()->compat_copy_strings(). If a 32bit process execs we seem to miss the RLIMIT_STACK check, no?
> > The patch asks for the cleanups. In particular, I think exec_mmap() > > should accept bprm, not mm. But I'd prefer to do this later. > > > > Oleg. > > General request. Please consider to keep Brad's reported-by tag.
Yes, yes, sure.
> > +static void acct_arg_size(struct linux_binprm *bprm, unsigned long pages)
OK.
> Please move this function into #ifdef CONFIG_MMU. nommu code doesn't use it.
Well it does, to revert the MM_ANONPAGES counter. I'll add the empty function for NOMMU.
> > +{ > > + struct mm_struct *mm = current->mm; > > + long diff = pages - bprm->vma_pages; > > I prefer to cast signed before assignment. It's safer more.
OK.
> > @@ -1003,6 +1024,7 @@ int flush_old_exec(struct linux_binprm * > > /* > > * Release all of the old mmap stuff > > */ > > + acct_arg_size(bprm, 0); > > Why do we need this unacct here? I mean 1) if exec_mmap() is success, > we don't need unaccount at all
Yes, we already killed all sub-threads. But this doesn't mean nobody else can use current->mm, think about CLONE_VM. The simplest example is vfork().
> 2) if exec_mmap() is failure, an epilogue of > do_execve() does unaccount thing.
Yes.
Thanks Kosaki!
I'll resend v2 today. I am still not sure about compat_copy_strings()...
Oleg.
| |