Date: Thu, 18 Aug 2005 00:57:39 -0700 From: Luigi Rizzo <rizzo@icir.org> To: Max Laier <max@love2party.net> Cc: arch@freebsd.org, net@freebsd.org Subject: Re: duplicate read/write locks in net/pfil.c and netinet/ip_fw2.c Message-ID: <20050818005739.A83776@xorpc.icir.org> In-Reply-To: <200508180332.34895.max@love2party.net>; from max@love2party.net on Thu, Aug 18, 2005 at 03:32:19AM %2B0200 References: <20050816170519.A74422@xorpc.icir.org> <200508170435.34688.max@love2party.net> <20050817170248.A70991@xorpc.icir.org> <200508180332.34895.max@love2party.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Aug 18, 2005 at 03:32:19AM +0200, Max Laier wrote:
> On Thursday 18 August 2005 02:02, Luigi Rizzo wrote:
...
> > could you guys look at the following code and see if it makes sense,
> > or tell me where i am wrong ?
> >
> > It should solve the starvation and blocking trylock problems,
> > because the active reader does not hold the mutex in the critical
^^^^^^
i meant 'writer', sorry... as max said even in the current implementation
the reader does not hold the lock.
> > int
> > RLOCK(struct rwlock *rwl, int try)
> > {
> > if (!try)
> > mtx_lock(&rwl->m);
> > else if (!mtx_trylock(&rwl->m))
> > return EBUSY;
> > if (rwl->writers == 0) /* no writer, pass */
> > rwl->readers++;
> > else {
> > rwl->br++;
> > cv_wait(&rwl->qr, &rwl->m);
> ^^^^^^^
>
> That we can't do. That's exactly the thing the existing sx(9) implementation
> does and where it breaks. The problem is that cv_wait() is an implicit sleep
> which breaks when we try to RLOCK() with other mutex already acquired.
but that is not a solvable problem given that the *LOCK may be blocking.
And the cv_wait is not an unconditioned sleep, it is one where you release
the lock right before ans wait for an event to wake you up.
In fact i don't understand why you consider spinning and sleeping
on a mutex two different things.
> Moreover will this break for recursive reads e.g.:
>
> Thread 1: RLOCK() ... RLOCK() -> cv_wait ...
> Thread 2: WLOCK() -> cv_wait ...
>
> This is exactly what pfil_hooks must be able to do as the packet filter may
> want to call back to the stack in order to send rejects etc.
that's another story (also in issue in ipfw) and the way it is addressed
elsewhere is by releasing and reaquiring the lock. In fact is the topic
that started this thread.
> change). The problem is, that we still have to invest 4 mutex operations for
> every access. The current implementation has the same basic problem (though
> it only uses 2 mutex operations for the WLOCK/UNLOCK). Ideally the RLOCK/
> UNLOCK should be free unless there is a writer waiting for the lock. In
"free unless" means not free - you always have to check, be it through
an atomic cmpswap or something else. But this is what mtx_lock does
anyways in the fast path.
cheers
luigi
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20050818005739.A83776>