Re: [PATCH 0/3] extend get/setrlimit to support setting rlimits external to a process (v7)

[Jiri Slaby <jirislaby@gmail.com>]

On 11/19/2009 12:15 AM, Oleg Nesterov wrote:
> On 11/18, Jiri Slaby wrote:
>>
>> On the last point: I added explicit ACCESS_ONCE all over there. If you
>> don't like I will trash it.
> 
> Just curious, why?
> 
> I mean, do you have any example of the "bad" behaviour which is
> fixed by ACCESS_ONCE() ?

Hi,

no, I haven't seen any errors caused by that yet. I added those in the
"just in case compiler starts to do weird things" manner.

>> The only remaining weird user is in
>> kernel/acct.c:
>>         /*
>>          * Accounting records are not subject to resource limits.
>>          */
>>         flim = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
>>         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
>>         file->f_op->write(file, (char *)&ac,
>>                                sizeof(acct_t), &file->f_pos);
>>         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = flim;
>>
>> It means that threads of the process with PACCT caps have unlimited file
>> size for a short while. If there is setrlimit in between, it gets wiped
>> out as well.
> 
> This is called when the whole thread-group exits, there are no
> live threads except current.

Not really, it is called from umount, sys_acct and other paths.

> At least, given that do_acct_process() does override_creds(), an
> ordinary user can't fool the accounting.

Agreed. That path can be executed only by a user with (at least) PACCT
or SYS_ADMIN caps. Hopefully.

Thanks.