Date: Sun, 17 Aug 2003 13:19:36 -0400 (EDT) From: Daniel Eischen <eischen@vigrid.com> To: Marcel Moolenaar <marcel@xcllnt.net> Cc: threads@freebsd.org Subject: Re: KSE/ia64: NULL thread pointer in _thr_sig_add() Message-ID: <Pine.GSO.4.10.10308171316040.12360-100000@pcnet5.pcnet.com> In-Reply-To: <20030817161255.GA567@dhcp42.pn.xcllnt.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, 17 Aug 2003, Marcel Moolenaar wrote: > On Sun, Aug 17, 2003 at 08:54:56AM -0400, Daniel Eischen wrote: > > > > > > > > One question. When you call (the syscall) setcontext() from > > > > _ia64_break_setcontext(), do you ignore the signal mask (uc.uc_sigmask)? > > > > > > Good point. I don't think so. > > > > OK, this will cause signal handling problems. You can > > use another flag, but that would dirty up MI setcontext(). > > Perhaps a separate ia64-specific system call? > > I actually have been thinking about using a flag. One that simply > indicates that the sigmask field is valid or not. The advantage > of such flag is that it makes the general interaction with user > land versions of getcontext() much more predictable and increases > the overall applicability of the *context syscalls. > > For example: libthr uses getcontext(2) but is not interested in > the sigmask. It simply uses the syscall as an initialization for > new threads. The context is passed to thr_create where we only > use the mcontext part to setup the registers. libkse tries to > avoid the syscalls. The userland versions don't save or restore > signal masks. There too we use contexts without actually being > interested in sigmasks. Also in libkse we see that the contexts > created in the kernel and given to userland are not contexts that > contains sigmasks. We simply copyout the mcontext part of the > ucontext. So, it appears that most of our uses of contexts are > such that we don't need to save and/or restore sigmasks. > > I think an UCF_SIGMASK flag to indicate that the context has a > valid sigmask is a logical or natural way to have setcontext(2) > operate well with getcontext(2), which always defines contexts > with sigmasks and userland versions of getcontext() or other > more specific uses of contexts that may or may not have a need > to save the sigmask. > > Reasoning differently: if we were to introduce a new syscall, > then the syscall would in effect be a clone of setcontext(2) > with the difference that we don't restore a part of the context. > This is duplication to avoid variation. Also, with a new syscall > we introduce the notion of having 2 seperate contexts. This not > only begs for a getcontext() clone that does not save the sigmask, > but also begs for a flag to differentiate between those 2 contexts. OK. I mostly wanted to avoid #if defined(ia64) in kern_context.c thinking the flag would have been in mc_flags. Please get rid of UCF_SWAPPED while you're at it; it's no longer used. -- Dan Eischen
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.GSO.4.10.10308171316040.12360-100000>