Date: Sat, 19 Mar 2005 11:46:31 -0700 From: Scott Long <scottl@samsco.org> To: Danny Braniss <danny@cs.huji.ac.il> Cc: net@freebsd.org Subject: Re: iSCSI initiator driver beta version, testers wanted Message-ID: <423C7387.8010804@samsco.org> In-Reply-To: <E1DCiL6-000MUy-6n@cs1.cs.huji.ac.il> References: <E1DCiL6-000MUy-6n@cs1.cs.huji.ac.il>
next in thread | previous in thread | raw e-mail | index | archive | help
Danny Braniss wrote: >>Scott Long wrote: >> >>>Danny Braniss wrote: >>> >>> >>>>>with tags enabled, iSCSI is much faster, but it also causes a >>>>>deadlock :-( >>>>>this is what i run: >>>>> newfs -U /<iscsi device> >>>>> cd /<iscsi device> >>>>> restore rf /home/file.dump >>>>> >>>>>on the same motherboard, a dual Xeon, with smp disabled all is OK >>>>>with smp enabled restore gets stuck usualy waiting on biord. >>>>>the iscsi driver shows that all requests have been done, the sniffing >>>>>shows the same(ie all request have been done). >>>>> >>>>>so this leads me to think that there is some race condition that i'm not >>>>>aware of in a SMP system, where xpt_done(ccb) is called while >>>>>another process is calling biowait. >>>>> >>>>>another lead is that after restore gets stuck, the system slowly gets >>>>>'stalled'. >>>>> >>>>>any insight is most welcome!, i'm also stuck. >>>> >>>> >>>> >>>>ahh, hate talking to myself :-) >>>> >>>>grabbing Giant before calling xpt_done solved it, so the problem is >>>>most probably in the CAM ... >>>> >>>>danny >>>> >>>> >>>> >>> >>>No, you need to grab Giant when calling xpt_done(). I even put an >>>assertion into CAM to make sure of that. Are you running with WITNESS >>>and/or INVARIANTS enabled? Those would have caught this problem. >>> > > they are off :-(, would have saved me some time. > >>>Scott >> >>Oops, I forgot to mention that I recently addressed this in 6-CURRENT. >>Now, much of the rest of cam API still requires Giant to be held, but >>xpt_done() does not. This only applies to 6-CURRENT, and I doubt that >>it will be backported to 5-STABLE. >> >>Scott > > > so what you are saying is that in 5.x it's a must to grab Giant before calling > xpt_done, and not in 6? > > danny > > > Correct. Again, turning on WITNESS and INVARIANTS is a very good thing to do during development. Scott
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?423C7387.8010804>