Date: Wed, 5 Jul 2017 18:12:16 +0200 From: Willem Jan Withagen <wjw@digiware.nl> To: Bruce Evans <brde@optusnet.com.au>, Konstantin Belousov <kostikbel@gmail.com> Cc: FreeBSD Filesystems <freebsd-fs@freebsd.org> Subject: Re: newfs returns cg 0: bad magic number Message-ID: <9aa4dd5c-7ec9-6fdf-420b-e31bbd5217c3@digiware.nl> In-Reply-To: <20170705154533.M1171@besplex.bde.org> References: <c98c813c-c393-9ba5-5c70-b9575fe59553@digiware.nl> <20170705051458.GU1935@kib.kiev.ua> <20170705154533.M1171@besplex.bde.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On 5-7-2017 08:55, Bruce Evans wrote:
> On Wed, 5 Jul 2017, Konstantin Belousov wrote:
>
>> On Wed, Jul 05, 2017 at 02:00:43AM +0200, Willem Jan Withagen wrote:
>>> Hi,
>>>
>>> I'm able to create a Ceph RBD backed ggate disk, in /dev/ggate0.
>>> It looks like I can:
>>> run dd on it
>>> gpart the disk
>>> create a zpool on it
>>>
>>> But when I try to create a UFS file system on it, newfs complains
>>> straight from the bat.
>>>
>>> # sudo newfs -E /dev/ggate0p1
>>> /dev/ggate0p1: 1022.0MB (2093056 sectors) block size 32768, fragment
>>> size 4096
>>> using 4 cylinder groups of 255.53MB, 8177 blks, 32768 inodes.
>>> Erasing sectors [128...2093055]
>>> super-block backups (for fsck_ffs -b #) at:
>>> 192, 523520, 1046848, 1570176
>>> cg 0: bad magic number
>>>
>>> Googling returns that this is on and off a problem with new devices, but
>>> there is no generic suggestion on how to debug this....
>>>
>>> Any/all suggestions are welcome,
>> Typically this error means that the drive returns wrong data, not the
>> bytes that were written to it and expected to be read.
>
> This might be for writing to a nonexistent sector. Checking for write
> errors was broken by libufs, so some write errors are only sometimes
> detected as a side effect of reading back garbage.
>
> I use the following quick fix (the patch also fixes some style bugs).
>
> X Index: mkfs.c
> X ===================================================================
> X RCS file: /home/ncvs/src/sbin/newfs/mkfs.c,v
> X retrieving revision 1.85
> X diff -u -1 -r1.85 mkfs.c
> X --- mkfs.c 9 Apr 2004 19:58:33 -0000 1.85
> X +++ mkfs.c 7 Apr 2005 23:51:56 -0000
> X @@ -437,16 +441,19 @@
> X if (!Nflag && Oflag != 1) {
> X - i = bread(&disk, SBLOCK_UFS1 / disk.d_bsize, chdummy,
> SBLOCKSIZE);
> X + i = bread(&disk, SBLOCK_UFS1 / disk.d_bsize, chdummy,
> X + SBLOCKSIZE);
> X if (i == -1)
> X - err(1, "can't read old UFS1 superblock: %s", disk.d_error);
> X -
> X + err(1, "can't read old UFS1 superblock: %s",
> X + disk.d_error);
> X if (fsdummy.fs_magic == FS_UFS1_MAGIC) {
> X fsdummy.fs_magic = 0;
> X - bwrite(&disk, SBLOCK_UFS1 / disk.d_bsize, chdummy,
> SBLOCKSIZE);
> X + bwrite(&disk, SBLOCK_UFS1 / disk.d_bsize, chdummy,
> X + SBLOCKSIZE);
> X for (i = 0; i < fsdummy.fs_ncg; i++)
> X - bwrite(&disk, fsbtodb(&fsdummy, cgsblock(&fsdummy, i)),
> X - chdummy, SBLOCKSIZE);
> X + bwrite(&disk,
> X + fsbtodb(&fsdummy, cgsblock(&fsdummy, i)),
> X + chdummy, SBLOCKSIZE);
> X }
> X }
> X - if (!Nflag)
> X - sbwrite(&disk, 0);
> X + if (!Nflag && sbwrite(&disk, 0) != 0)
> X + err(1, "sbwrite: %s", disk.d_error);
> X if (Eflag == 1) {
> X @@ -518,4 +525,4 @@
> X }
> X - if (!Nflag)
> X - sbwrite(&disk, 0);
> X + if (!Nflag && sbwrite(&disk, 0) != 0)
> X + err(1, "sbwrite: %s", disk.d_error);
> X for (i = 0; i < sblock.fs_cssize; i += sblock.fs_bsize)
>
> libufs broke the error handling for the most important writes -- to
> the superblock. Error handling is still done almost correctly in
> wtfs(), and most writes are still done using wtfs() which is now
> just a wrapper which adds error handling to libufs's bwrite(3), but
> writes to superblock are (were) now done internally by libufs's
> sbwrite(3) which (like most of libufs) is too hard to use.
>
> Note that -current needs a slightly different fix. Part of libufs
> being too hard to use is that it is a library so it can't just exit
> for errors. It returns errors in the string disk.d_error and the
> fix uses that for newfs, unlike for most other calls to sbwrite(3).
> However, newfs no longer uses sbwrite(3). It uses a wrapper
> do_sbwrite() which reduces to pwrite(2). The wrapper doesn't set
> d_error, so it is incompatible with sbwrite(3).
>
> This is an example that libufs is even harder to use than might first
> appear. The version with the do_sbwrite() wrapper fixes a previous
> version which replaced bwrite(3) instead of wrapping it. bwrite()
> in the application conflicted with bwrite(3) in libufs, since libufs
> is not designed to have its internals replaced by inconsistent parts
> like that. Apparently, a special case is only needed for superblock
> writes, and do_sbwrite() does that, and since libufs doesn't call any
> sbwrite() function internally there is no need to replace sbwrite(3);
> sbwrite(3) is just useless for its main application. All that the
> bwrite(3) and sbwrite(3) library functions do is handle the block
> size implicitly in a way that makes them harder to use than just
> multiplying by the block size like wtfs() used to do and do_sbwrite()
> now does.
the 11.1RC code reads:
if (!Nflag)
do_sbwrite(&disk);
So i'm not certain from the above if I now replace that with:
if (!Nflag && do_sbwrite(&disk) != 0)
err(1, "sbwrite: %s", disk.d_error);
Or to stick the sbwrite() back in again?
I guess the later, because of the error return value in sbwrite().
--WjW
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?9aa4dd5c-7ec9-6fdf-420b-e31bbd5217c3>