Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 01 Feb 2008 12:00:16 +0200
From:      Andriy Gapon <avg@icyb.net.ua>
To:        Remko Lodder <remko@FreeBSD.org>, scottl@FreeBSD.org,  freebsd-fs@freebsd.org
Cc:        Bruce Evans <bde@zeta.org.au>
Subject:   Re: kern/84983: [udf] [patch] udf filesystem: stat-ting files could randomly fail
Message-ID:  <47A2EDB0.8000801@icyb.net.ua>
In-Reply-To: <200612221824.kBMIOhfM049471@freefall.freebsd.org>
References:  <200612221824.kBMIOhfM049471@freefall.freebsd.org>

next in thread | previous in thread | raw e-mail | index | archive | help
on 22/12/2006 20:24 Pav Lucistnik said the following:
> Synopsis: [udf] [patch] udf filesystem: stat-ting files could randomly fail
> 
> State-Changed-From-To: open->closed
> State-Changed-By: pav
> State-Changed-When: Fri Dec 22 18:24:14 UTC 2006
> State-Changed-Why: 
> Fixed in 6.1 and up
> 
> http://www.freebsd.org/cgi/query-pr.cgi?pr=84983

I found a bug in the patch. I got a panic in real situation when testing
a UDF fs with a directory with a huge number of files (~10^4), but this
can be easily shown in the code too:

static int
udf_readatoffset(struct udf_node *node, int *size, off_t offset,
    struct buf **bp, uint8_t **data)
{
...
*size = min(*size, MAXBSIZE);

if ((error = udf_readlblks(udfmp, sector, *size + (offset &
udfmp->bmask), bp))) {

If it so happens that *size gets MAXBSIZ value and (offset &
udfmp->bmask) is not zero, then a value > MAXBSIZ would be passed to
udf_readlblks->bread->breadn->getblk and the latter will panic because
it has an explicit assert for size <= MAXBSIZ.

I think there should be something in the code like the following:
*size = min(*size, MAXBSIZE - (offset & udfmp->bmask));

---- a different, under-debugged problem -----
BTW, on some smaller directories (but still large ones) I get some very
strange problems with reading a directory too. It seems like some bad
interaction between udf and buffer cache system. I added a lot of
debugging prints and the problems looks like the following:

read starting at physical sector (2048-byte one) N, size is ~20K, N%4=0
	bread(4 * N, some_big_size)
		correct data is read
repeat the above couple dozen times
read starting at physical sector (2048-byte one) N+1, size is ~20K
	bread(4 * (N+1), some_big_size)
		data is read from physical sector N+4 (instead of N+1)

I remember that Bruce Evance warned me that something like this could
happen but I couldn't understand him, because I don't understand
VM/buffer subsystem. I'll try to dig up the email.


-- 
Andriy Gapon



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?47A2EDB0.8000801>