Date: Sun, 4 Jan 2009 18:00:43 +0200 From: Kostik Belousov <kostikbel@gmail.com> To: Jaakko Heinonen <jh@saunalahti.fi> Cc: freebsd-fs@freebsd.org Subject: Re: [patch] ext2fs freeze/corruption on amd64 Message-ID: <20090104160043.GI93900@deviant.kiev.zoral.com.ua> In-Reply-To: <20090103182107.GA4119@a91-153-125-115.elisa-laajakaista.fi> References: <20090103182107.GA4119@a91-153-125-115.elisa-laajakaista.fi>
next in thread | previous in thread | raw e-mail | index | archive | help
--wZdghQXYJzyo6AGC Content-Type: text/plain; charset=us-ascii Content-Disposition: inline Content-Transfer-Encoding: quoted-printable On Sat, Jan 03, 2009 at 08:21:08PM +0200, Jaakko Heinonen wrote: >=20 > Hi, >=20 > I can quite easily reproduce a hard freeze while writing to an ext2fs > filesystem on amd64. I can reliably reproduce it by building a kernel > with "make -j2" on ext2fs. >=20 > This has been reported by other people too: >=20 > http://lists.freebsd.org/pipermail/freebsd-fs/2008-March/004504.html > http://lists.freebsd.org/pipermail/freebsd-bugs/2008-January/027555.html > http://www.freebsd.org/cgi/query-pr.cgi?pr=3Dkern/119529 > http://www.freebsd.org/cgi/query-pr.cgi?pr=3Damd64/99561 >=20 > I tracked the problem down to find_next_zero_bit() in ext2_bitops.h. > The bug doesn't affect i386 because it uses asm bitops from > i386-bitops.h. (There's also sparc64-bitops.h file but it's not used > anywhere.) >=20 > The problem is that if the offset (ofs) has one or more of it's five low > bits set those bits are incorrectly re-added to the result. With > following patch applied I can't reproduce the problem. >=20 > %%% > Index: sys/gnu/fs/ext2fs/ext2_bitops.h > =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 > --- sys/gnu/fs/ext2fs/ext2_bitops.h (revision 186639) > +++ sys/gnu/fs/ext2fs/ext2_bitops.h (working copy) > @@ -84,7 +84,7 @@ find_next_zero_bit(void *data, size_t sz > mask =3D ~0U << (ofs & 31); > bit =3D *p | ~mask; > if (bit !=3D ~0U) > - return (ffs(~bit) + ofs - 1); > + return (ffs(~bit) + (ofs & ~31U) - 1); > p++; > ofs =3D (ofs + 31U) & ~31U; > } > %%% >=20 > I hope that someone would commit this or similar fix. It might be also > reasonable to delete the unused sparc64-bitops.h file. Please note that find_next_zero_bit() uses gcc extension to perform void-pointer arithmetic. Also, at the +2 line, it performs unaligned access to the uint32_t entities. --wZdghQXYJzyo6AGC Content-Type: application/pgp-signature Content-Disposition: inline -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.9 (FreeBSD) iEYEARECAAYFAklg3SoACgkQC3+MBN1Mb4jasACgnVBp7ehnPSciZTa4jT3R771M NfEAn2bUN458Uav6daqeGf5NOEjrTXPg =3Dqp -----END PGP SIGNATURE----- --wZdghQXYJzyo6AGC--
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20090104160043.GI93900>