Re: 2.6.24-rc5-mm1 - wonky disk cache and CDROM behavior...

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 18 Dec 2007 10:37:32 +0800 Dave Young <hidave.darkstar@gmail.com> wrote:

> On Mon, Dec 17, 2007 at 09:07:56PM -0500, Valdis.Kletnieks@vt.edu wrote:
> > On Mon, 17 Dec 2007 14:56:44 PST, Andrew Morton said:
> > 
> > (Adding Al Viro to the list, he's listed as "file systems" and MAINTAINERS
> > doesn't list 'isofs' anyplace.  Will Al or Andrew please vector to whoever
> > actually does that code?)
> > 
> > > > I try it again, and it reports it died at the same exact place, but in about
> > > > 2 seconds flat, and reports 91M/sec transfer.  OK, that's *weird*, I didn't
> > > > think that blocks read from /dev/cdrom would get cached, but OK.
> > > 
> > > It'll remain cached if something is holding the device open.
> > 
> > Does it need to be "device open", or are there other things as well? If the
> > drop_cache was hosed, that would result in the same symptoms, no?
> > 
> > > Something's holding s_umount for writing I guess.  Possibly busted error
> > > handling somewhere totally different.
> > 
> > Aha - found what was holding it - an attempt to loopback mount the truncated
> > file (before I realized it was truncated) had failed - I had gotten a 'Killed'
> > back from the mount, but I didn't realize it had pulled an actual oops:
> > 
> > Dec 17 15:54:33 turing-police kernel: [14503.402385] attempt to access beyond end of device
> > Dec 17 15:54:33 turing-police kernel: [14503.402391] loop1: rw=0, want=1284500, limit=314240
> > Dec 17 15:54:33 turing-police kernel: [14503.402395] ISOFS: unable to read i-node block
> > Dec 17 15:54:33 turing-police kernel: [14503.402428] Unable to handle kernel NULL pointer dereference at 000000000000010b RIP:
> > Dec 17 15:54:33 turing-police kernel: [14503.402440]  [<ffffffff802a096b>] iput+0x11/0x80
> > ...
> > Dec 17 15:54:33 turing-police kernel: [14503.403008] Call Trace:
> > Dec 17 15:54:33 turing-police kernel: [14503.403026]  [<ffffffff802ff73e>] isofs_fill_super+0x7e9/0xa6b
> > Dec 17 15:54:33 turing-police kernel: [14503.403045]  [<ffffffff80523d28>] __down_write_nested+0x3d/0xa1
> > Dec 17 15:54:33 turing-police kernel: [14503.403061]  [<ffffffff80523d97>] __down_write+0xb/0xd
> > Dec 17 15:54:33 turing-police kernel: [14503.403076]  [<ffffffff8028fb63>] sget+0x397/0x3a9
> > Dec 17 15:54:33 turing-police kernel: [14503.403090]  [<ffffffff8028f204>] set_bdev_super+0x0/0x14
> > Dec 17 15:54:33 turing-police kernel: [14503.403106]  [<ffffffff80290301>] get_sb_bdev+0x109/0x157
> > Dec 17 15:54:33 turing-police kernel: [14503.403120]  [<ffffffff802fef55>] isofs_fill_super+0x0/0xa6b
> > Dec 17 15:54:33 turing-police kernel: [14503.403138]  [<ffffffff802fe2e9>] isofs_get_sb+0x13/0x15
> > Dec 17 15:54:33 turing-police kernel: [14503.403151]  [<ffffffff80290075>] vfs_kern_mount+0x90/0x11a
> > Dec 17 15:54:33 turing-police kernel: [14503.403167]  [<ffffffff8029015c>] do_kern_mount+0x47/0xe3
> > Dec 17 15:54:33 turing-police kernel: [14503.403183]  [<ffffffff802a5012>] do_mount+0x717/0x78a
> > Dec 17 15:54:33 turing-police kernel: [14503.403199]  [<ffffffff805242fc>] _read_lock_irq+0x9/0xb
> > Dec 17 15:54:33 turing-police kernel: [14503.403212]  [<ffffffff8026cce0>] find_lock_page+0x8c/0x97
> > Dec 17 15:54:33 turing-police kernel: [14503.403227]  [<ffffffff8026ecb6>] filemap_fault+0x1fa/0x3c6
> > Dec 17 15:54:33 turing-police kernel: [14503.403241]  [<ffffffff8026cb6b>] unlock_page+0x2d/0x31
> > Dec 17 15:54:33 turing-police kernel: [14503.403254]  [<ffffffff8027925c>] __do_fault+0x38d/0x3c3
> > Dec 17 15:54:33 turing-police kernel: [14503.403274]  [<ffffffff8027ab68>] handle_mm_fault+0x36d/0x6e9
> > Dec 17 15:54:33 turing-police kernel: [14503.403293]  [<ffffffff80271903>] __alloc_pages+0x68/0x2f6
> > Dec 17 15:54:33 turing-police kernel: [14503.403314]  [<ffffffff802a510e>] sys_mount+0x89/0xcb
> > Dec 17 15:54:33 turing-police kernel: [14503.403328]  [<ffffffff80214f34>] syscall_trace_enter+0x97/0x9b
> > Dec 17 15:54:33 turing-police kernel: [14503.403344]  [<ffffffff8020c34c>] tracesys+0xdc/0xe1
> > Dec 17 15:54:33 turing-police kernel: [14503.403359]
> > Dec 17 15:54:33 turing-police kernel: [14503.403366]
> > Dec 17 15:54:33 turing-police kernel: [14503.403367] Code: 48 8b 87 10 01 00 00 48 83 bf 38 02 00 00 40 48 8b 40 38 75
> > 
> > I don't mind it failing the mount, but the oops seems excessive.  I suspect
> > that *somewhere* in that stack trace, we're wanting something like a
> > 
> > 	if (!foo_ptr)
> > 		return -EIO;
> > 
> > but I admit not being competent enough to decide where that should be.
> > 
> 
> Hi,
> Could you please try the below patch:
> 
> Signed-off-by: Dave Young <hidave.darkstar@gmail.com> 
> 
> ---
> fs/isofs/inode.c |    2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff -upr linux/fs/isofs/inode.c linux.new/fs/isofs/inode.c
> --- linux/fs/isofs/inode.c	2007-12-18 10:31:12.000000000 +0800
> +++ linux.new/fs/isofs/inode.c	2007-12-18 10:31:56.000000000 +0800
> @@ -1414,7 +1414,7 @@ struct inode *isofs_iget(struct super_bl
>  		ret = isofs_read_inode(inode);
>  		if (ret < 0) {
>  			iget_failed(inode);
> -			inode = ERR_PTR(ret);
> +			return NULL;
>  		} else {
>  			unlock_new_inode(inode);
>  		}
> 

Yup.

David, this is concerning.  More such error-path bugs in that code will
take years and years to get found and fixed.  The best way to eliminate
them is a line-by-line re-review of the patchset.