Date: Mon, 9 Dec 2013 15:17:10 +0100 From: Peter Holm <peter@holm.cc> To: Konstantin Belousov <kostikbel@gmail.com> Cc: mav@freebsd.org, freebsd-current@freebsd.org, Stefan Hegnauer <stefan.hegnauer@gmx.ch> Subject: Re: nanobsd / dd problem? Message-ID: <20131209141710.GA18608@x2.osted.lan> In-Reply-To: <20131209044239.GS59496@kib.kiev.ua> References: <52a4ad05.892ee50a.41cd.084aSMTPIN_ADDED_BROKEN@mx.google.com> <20131209044239.GS59496@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Dec 09, 2013 at 06:42:39AM +0200, Konstantin Belousov wrote: > On Sun, Dec 08, 2013 at 06:31:36PM +0100, Stefan Hegnauer wrote: > > Hi, > > > > > > > > I am using freebsd-current (FreeBSD BUILDMASTER 11.0-CURRENT FreeBSD > > 11.0-CURRENT #0 r259095: Sun Dec 8 10:20:40 CET 2013 > > root@BUILDMASTER:/usr/obj/usr/src/sys/ASUS i386) in a VirtualBox as a build > > machine for nanobsd images to be used on pc-engines.ch alix boards. The only > > difference to GENERIC is the inclusion of 'march=geode' and disabling of > > most debugging switches (malloc, Witness etc). Worked like a charm in the > > past. > > > > > > > > Since late summer - sorry, no exact date / svn revision - nanobsd.sh fails > > at the last stage when building the disk image, e.g. with > > > > ... > > > > 00:00:25 ### log: /usr/obj/nanobsd.alixpf//_.di > > > > # > > > > > > > > Looking a bit closer it seems that dd(1) returns with an I/O error whenever > > the input is a file created with mdconfig(8): > > > > # dd if=/dev/zero of=somebackingfile bs=1k count=5k > > > > # mdconfig -f somebackingfile -u md0 > > > > # newfs -U /dev/md0 > > > > # dd if=/dev/md0 of=/dev/null > > > > dd: /dev/md0: Input/output error > > > > 10241+0 records in > > > > 10241+0 records out > > > > 5243392 bytes transferred in 3.240345 secs (1618159 bytes/sec) > > > > > > > > The outputfile in nanobsd.sh seems to be error-free. > It should be one block larger than the right size. > \ > > > > Anyone else seen similar behaviour? How to proceed/fix it? > > > > The following patch should clear the error. > > The issue is that kern_physio() incorrectly detects EOF due to incorrect > calculation of bio bio_resid after the bio_length was clipped by the > 'excess' code in g_io_check(). Both bio_length and bio_resid appear > to be 0 in the pre-last dd transfer, which starts exactly and the > mediasize, and kern_physio() thinks that it transferred one more block > than was transferred. > > I _suspect_ that it was caused by 'excess' code moving in r256880, > but I am really not in the right condition to analyze it. If somebody > could try the same dd experiment to confirm or deny my suspicion, it > would be useful. > > The patch below should be a right thing to do anyway. > > diff --git a/sys/kern/vfs_bio.c b/sys/kern/vfs_bio.c > index c23a74b..b7c4d60 100644 > --- a/sys/kern/vfs_bio.c > +++ b/sys/kern/vfs_bio.c > @@ -3679,7 +3679,6 @@ bufdonebio(struct bio *bip) > > bp = bip->bio_caller2; > bp->b_resid = bp->b_bcount - bip->bio_completed; > - bp->b_resid = bip->bio_resid; /* XXX: remove */ > bp->b_ioflags = bip->bio_flags; > bp->b_error = bip->bio_error; > if (bp->b_error) I have tested this patch with a buildworld + selected other tests. No problems seen (and problem fixed, of cause). - Peter
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20131209141710.GA18608>