Date: Thu, 24 Dec 2009 07:55:11 -0500 From: John Baldwin <jhb@freebsd.org> To: David Xu <davidxu@freebsd.org> Cc: freebsd-threads@freebsd.org Subject: Re: first patch for process-shared semaphore Message-ID: <200912240755.11841.jhb@freebsd.org> In-Reply-To: <4B32C1DC.9080308@freebsd.org> References: <4B317741.8080004@freebsd.org> <200912230936.35998.jhb@freebsd.org> <4B32C1DC.9080308@freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Wednesday 23 December 2009 8:20:28 pm David Xu wrote: > John Baldwin wrote: > > On Tuesday 22 December 2009 8:49:53 pm David Xu wrote: > >> This is my first attempt to make process-shared mutex work, this means > >> you can mmap(MAP_SHARED) a memory area, and put semaphore there, > >> or you can sem_open a named semaphore, and just use it between > >> processes, the named semaphore uses file system and mmap(), directory > >> /tmp/.semaphore is used as IPC directory, any named semaphore > >> locates in the directory. old semaphore implementation still exists > >> to make it binary compatible, it uses symbol version. > >> > >> http://people.freebsd.org/~davidxu/patch/shared_semaphore_1.patch > > > > I would suggest that you leave named semaphores as they currently exist and > > follow this approach instead: > > > > 1) Named semaphores use ksem_*() still. > > 2) sem_init/sem_destroy operate on UTMX-backed semaphores identical to the > > ones used in the current libthr code. The semid_t structure now becomes the > > full structure that libthr currently allocates with a flag to indicate if it > > is a "system" semaphore or otherwise. The pshared flag passed to sem_init() > > can be used to set the sharing properties of the UMTX. > > 3) All of sem_init/sem_destroy is just in libc. Just move the libthr > > implementation bits into libc. > > > > ksem base shared semaphore is slow because whenever you call > sem_wait(), it always enters kernel even if count is non-zero, > sem_post() also always enters kernel even if there is no waiter. > but the new implementation is as simple as just an atomic operation > in these cases, I know another competitor OS is doing things in > this way. Yes, Solaris uses files in /tmp and Darwin uses special file descriptors similar to what we do. However, you will have to restrict the namespace if you go the /tmp route to be safe I think similar to what Solaris does (no path separators, just simple names like 'foo'). You might also want to use the same naming convention as Solaris if you go the /tmp route (I think they use a path other than .semaphore under /tmp IIRC). Not sure if we want to do anything special to ensure that those particular set of files in /tmp always get purged on reboot to avoid weird bugs with semaphores unexpectedly persisting across reboots. -- John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200912240755.11841.jhb>