On 02/14, David Howells wrote:
Oleg Nesterov [off-list ref] wrote:
quoted
No, only current can change ->blocked. This is even documented in
sigprocmask(). And more, the only correct way to change ->blocked
is set_current_blocked(). OK, with a couple of "I know what I am
doing" exceptions in kernel/signal.c.
I was looking at force_sig_info() and derivatives. Is that what you refer to?
Ah, sorry, forgot to mention...
force_sig_info() (and its callers) need the cleanups and fixes. It
is almost always wrong if t != current.
For example, please look at
[PATCH 1/4] signal: give SEND_SIG_FORCED more power to beat SIGNAL_UNKILLABLE
http://marc.info/?l=linux-kernel&m=132890442717122
Hopefully we can fix all ab-users soon.
Just in case... if we race with force_sig_info() the task will be
killed anyway. But I agree, this is not nice and should be fixed.
And in any case, there are other places which assume it is safe
to read current->blocked lockless.
If so, is it worth providing a force_sig_info_current(),
force_sigsegv_current() and force_sig_current() to make things clearer to grep
for, I wonder?
Yes, I think the "task_struct *t" argument should die.
Oleg.