Date: Mon, 5 Jan 2004 18:22:31 +1100 (EST) From: Bruce Evans <bde@zeta.org.au> To: "Bjoern A. Zeeb" <bzeeb-lists@lists.zabbadoz.net> Cc: current@freebsd.org Subject: Re: Expensive timeout(9) function ? Message-ID: <20040105175530.K4057@gamplex.bde.org> In-Reply-To: <Pine.BSF.4.53.0401041820230.54854@e0-0.zab2.int.zabbadoz.net> References: <Pine.BSF.4.53.0401041820230.54854@e0-0.zab2.int.zabbadoz.net>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, 4 Jan 2004, Bjoern A. Zeeb wrote: > what reports do you expect with the > > "Expensive timeout(9) function" > > message ? What you reported (function names and timeout time) is interesting. Why do we see it ? Kernel bugs :-). > Expensive timeout(9) function: 0xc04885a0(0) 1.024846430 s [1] > Expensive timeout(9) function: 0xc04885a0(0) 1.024846430 s [1] > Expensive timeout(9) function: 0xc04b3940(0) 0.008629758 s [2] > Expensive timeout(9) function: 0xc04b39a0(0) 0.004333781 s [2] > Expensive timeout(9) function: 0xc04f71f0(0) 0.027004551 s [3] > Expensive timeout(9) function: 0xc04f71f0(0) 0.027004551 s [3] > Expensive timeout(9) function: 0xc04f71f0(0) 0.027004551 s [3] > > [1] sys/kern/kern_synch.c:loadav() > [2] sys/kern/uipc_domain.c:pfslowtimo() > [3] sys/netinet/ip_fw2.c:ipfw_tick() [1] is easiest to understand. loadav() is obviously broken since it uses sleep locks. Apparently it sometimes sleeps for more than 1 second altogether! There is a check for sleeping in timeouts under DIAGNOSTIC. I would expect to complaints from this too if you just used DIAGNOSTIC to get the above. [3] ipfw_tick() is obviously broken in the same way. This is from blind conversion of splimp() to a sleep lock. Mutexes work quite differently from spl's. A quick fix for timeout routines that only lock things once might be to use mtx_trylock() and not do anything in the timeout routine (except re-arm the timeout, perhaps with a smaller interval) if the mutex cannot be acquired immediately. This depends on the exact timing of timeout routines not being critical (not that we have exact timing -- the above shows all timeouts being delayed by a factor of at least 100 (1 second instead of 1/100 seconds)). This should work expecially well in loadav() -- loadav() intentionally adds jitter to the interval. This might have worked in schedcpu() too (schedcpu() was converted to a thread). Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20040105175530.K4057>