Date: Mon, 7 Oct 2013 12:09:56 -0400 From: John Baldwin <jhb@freebsd.org> To: Alexander Motin <mav@freebsd.org> Cc: freebsd-hackers@freebsd.org, freebsd-current@freebsd.org, freebsd-geom@freebsd.org Subject: Re: [RFC][CFT] GEOM direct dispatch and fine-grained CAM locking Message-ID: <201310071209.56312.jhb@freebsd.org> In-Reply-To: <525111A2.1020106@FreeBSD.org> References: <5224511D.4090503@FreeBSD.org> <201310021330.23251.jhb@freebsd.org> <525111A2.1020106@FreeBSD.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sunday, October 06, 2013 3:30:42 am Alexander Motin wrote: > On 02.10.2013 20:30, John Baldwin wrote: > > On Saturday, September 07, 2013 2:32:45 am Alexander Motin wrote: > >> On 07.09.2013 02:02, Jeremie Le Hen wrote: > >>> On Fri, Sep 06, 2013 at 11:29:11AM +0300, Alexander Motin wrote: > >>>> On 06.09.2013 11:06, Jeremie Le Hen wrote: > >>>>> On Fri, Sep 06, 2013 at 12:46:27AM +0200, Olivier Cochard-Labb=E9 w= rote: > >>>>>> On Thu, Sep 5, 2013 at 11:38 PM, Alexander Motin <mav@freebsd.org> > > wrote: > >>>>>>> I've found and fixed possible double request completion, that cou= ld > > cause > >>>>>>> such symptoms if happened. Updated patch located as usual: > >>>>>>>=20 http://people.freebsd.org/~mav/camlock_patches/camlock_20130905.patch > >>>>>>> > >>>>> With this new one I cannot boot any more (I also updated the source > >>>>> tree). This is a hand transcripted version: > >>>>> > >>>>> Trying to mount root from zfs:zroot/root []... > >>>>> panic: Batch flag already set > >>>>> cpuid =3D 1 > >>>>> KDB: stack backtrace: > >>>>> db_trace_self_wrapper() > >>>>> kdb_backtrace() > >>>>> vpanic() > >>>>> kassert_panic() > >>>>> xpt_batch_start() > >>>>> ata_interrupt() > >>>>> softclock_call_cc() > >>>>> softclock() > >>>>> ithread_loop() > >>>>> fork_exit() > >>>>> fork_trampoline() > >>>> > >>>> Thank you for the report. I see my fault. It is probably specific to > >>>> ata(4) driver only. I've workarounded that in new patch version, but > >>>> probably that area needs some rethinking. > >>>> > >>>> http://people.freebsd.org/~mav/camlock_patches/camlock_20130906.patch > >>> > >>> I'm not sure you needed a confirmation, but it boots. Thanks :). > >>> > >>> I didn't quite understand the thread; is direct dispatch enabled for > >>> amd64? ISTR you said only i386 but someone else posted the macro for > >>> amd64. > >> > >> Yes, it is enabled for amd64. I've said x86, meaning both i386 and amd= 64. > > > > FYI, I tested mfi with this patch set and mfid worked fine for handling= =20 g_up > > directly: > > > > Index: dev/mfi/mfi_disk.c > > =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D > > --- dev/mfi/mfi_disk.c (revision 257407) > > +++ dev/mfi/mfi_disk.c (working copy) > > @@ -162,6 +162,7 @@ > > sc->ld_disk->d_unit =3D sc->ld_unit; > > sc->ld_disk->d_sectorsize =3D secsize; > > sc->ld_disk->d_mediasize =3D sectors * secsize; > > + sc->ld_disk->d_flags =3D DISKFLAG_DIRECT_COMPLETION; > > if (sc->ld_disk->d_mediasize >=3D (1 * 1024 * 1024)) { > > sc->ld_disk->d_fwheads =3D 255; > > sc->ld_disk->d_fwsectors =3D 63; > > >=20 > Thank you for the feedback. But looking on mfi driver sources I would=20 > say that it calls biodone() from mfi_disk_complete() from cm_complete()=20 > method, which is called while holding mfi_io_lock mutex. I guess that if= =20 > on top of mfi device would be some GEOM class, supporting direct=20 > dispatch and sending new requests down on previous request completion=20 > (or retrying requests), that could cause recursive mfi_io_lock=20 > acquisition. That is exactly the cause why I've added this flag. May be=20 > it is a bit paranoid, but it is better to be safe then sorry. >=20 > Another good reason to drop the lock before calling biodone() would be=20 > reducing the lock hold time. Otherwise it may just increase lock=20 > congestion there and destroy all benefits of the direct dispatch. Ah, interesting. What is your policy for such drivers? Should they be left using g_up, should they drop the lock around biodone when completeing multiple requests in an interrupt? Should they try to batch them by waiting and doing biodone at the end after dropping the lock? =2D-=20 John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201310071209.56312.jhb>