On Wed, Feb 12, 2014 at 12:13:19PM -0800, Linus Torvalds wrote: > On Wed, Feb 12, 2014 at 3:39 AM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > On Tue, Feb 11, 2014 at 10:28:12PM -0800, Linus Torvalds wrote: > > > >> It looks like just "do_signal()" has a stack frame that is about 230 > >> bytes even under normal circumstancs (largely due to "struct ksignal" > >> - which in turn is largely due to the insane 128-byte padding in > >> siginfo_t). Add a few other frames in there, and I guess that if it > >> was close before, the coredump path just makes it go off. > > > > We could, in principle, put it into task_struct and make get_signal() > > return its address - do_signal() is called only in the code that does > > assorted returns to userland... > > We have better uses for random buffers in "struct task_struct", I'd > hate to put a siginfo_t there. *nod* > The thing is, siginfo_t has that idiotic 128-byte area, but it's all > "for future expansion". I think it's some damn glibc disease - we've > seen these kinds of insane paddings before. > > The actual *useful* part of siginfo_t is on the order of 32 bytes. If that. > > Sad. Umm... What if we delay __sigqueue_free()? After all, that's where the fat sucker normally comes from. That way we might get away with much smaller structure on stack... Just introduce a small structure that would contain signr, uid, pid and pointer to struct sigqueue. And pass a pointer to _that_ all the way down to collect_signal(). Pointer's NULL == it's SI_USER with signr/uid/pid from the small struct and all other fields are zero. Pointer isn't NULL - use &small_struct->p->info. And have struct sigqueue actually freed via task_work_add() in that case. Do you see any fundamental problems with that? Looks like it would be faster as well - less copying involved... _______________________________________________ xfs mailing list xfs@xxxxxxxxxxx http://oss.sgi.com/mailman/listinfo/xfs
