FreeBSD 9: fdisk -It crashes kernel
Guy Helmer
guy.helmer at gmail.com
Thu Apr 25 18:49:28 UTC 2013
On Apr 25, 2013, at 12:57 PM, Jeremy Chadwick <jdc at koitsu.org> wrote:
> On Thu, Apr 25, 2013 at 11:58:42AM -0500, Guy Helmer wrote:
>> On Apr 25, 2013, at 10:58 AM, Jeremy Chadwick <jdc at koitsu.org> wrote:
>>
>>> On Thu, Apr 25, 2013 at 09:06:49AM -0500, Guy Helmer wrote:
>>>> Encountered a surprise when my disk resizing rc.d script caused FreeBSD 9.1-STABLE to crash. I used "fdisk -It ada0" to determine what the available size of the disk (which happened to be the root disk), and on FreeBSD 9.1 the kernel comes crashing down:
>>>>
>>>> + fdisk -It ada0
>>>> + /rescue/sed -En 's,.*start ([0-9]+).*size ([0-9]+).*,\1 + \2,p'
>>>> vnode_pager_getpages: I/O read error
>>>> vm_fault: pager read error, pid 65 (fdisk)
>>>> pid 65 (fdisk), uid 0: exited on signal 11
>>>> eval: arithmetic expression: expecting primary: ""
>>>> Entropy harvesting: point_to_pointeval: date: Device not configured
>>>> eval: df: Device not configured
>>>> eval: dmesg: Device not configured
>>>> cat: /bin/ls: Device not configured
>>>> kickstart.
>>>> eval: cannot open /etc/fstab: Device not configured
>>>> eval: cannot open /etc/fstab: Device not configured
>>>> eval: swapon: Device not configured
>>>> Warning! No /etc/fstab: skipping disk checks
>>>> fstab: /etc/fstab:0: Device not configured
>>>>
>>>> Fatal trap 12: page fault while in kernel mode
>>>> cpuid = 1; apic id = 01
>>>> fault virtual address = 0x0
>>>> fault code = supervisor read, page not present
>>>> instruction pointer = 0x20:0xc0825fc4
>>>> stack pointer = 0x28:0xc5a088c8
>>>> frame pointer = 0x28:0xc5a08914
>>>> code segment = base 0x0, limit 0xfffff, type 0x1b
>>>> = DLP 0, pres 1, def32 1, gran 1
>>>> processor eflags = interrupt enabled, resume, IOPL = 0
>>>> current process = 91 (mount)
>>>> [ thread pid 91 tid 100056 ]
>>>> Stopped at g_access+0x24: mlvl 0(%ebx),%eax
>>>> db> where
>>>> Tracing pid 91 tid 100056 td 0xc84c42f0
>>>> g_access(c8481d34,0,1,1,0,…) at g_access+0x24/frame 0xc5a08914
>>>> ffs_mount(c8481d34,c0d78380,2,c5a08c00,c829ae6c,…) af ffs_mount+0xf74/frame 0xc5a08a34
>>>> vfs_donmount(c84c42f0,10000,0,c84cf200,c84cf200,…) at vfs_donmount+0x1423/frame 0xc5a08c24
>>>> sys_nmount(c84c42f0,c5a08ccc,c5a08cc4,1010006,c5a08d08,…) at sys_nmount+0x7f/frame 0xc5a08c48
>>>> syscall(c5a08d08) at syscall+0x443/frame 0xc508cfc
>>>> Xint0x80_syscall() at Xint0x80_syscall+0x21/frame 0xc5a08cfc
>>>> --- syscall (378, FreeBSD ELF32, sys_nmount), eip = 0x480d5feb, esp = 0xbfbfce1c, ebp = 0xbfbfd378 ---
>>>>
>>>> I'll fix my script to not do this, but it seems odd that fdisk -It can make the disk "go away".
>>>
>>> Please provide a full, unmodified copy of your script.
>>>
>>> What's confusing to me is that after your sed call (which I don't even
>>> understand, because it doesn't appear to be operating on anything except
>>> stdin/stdout, and we don't know what that is -- again, show the script),
>>> the kernel starts outputting indications that the root disk/filesystem
>>> or its related metadata disappeared:
>>>
>>>> vnode_pager_getpages: I/O read error
>>>> vm_fault: pager read error, pid 65 (fdisk)
>>>> pid 65 (fdisk), uid 0: exited on signal 11
>>>
>>> Except the kernel stack trace indicates something called sys_nmount(),
>>> which called vfs_donmount(), which called ffs_mount(), which calls
>>> g_access(). All of those scream to me "someone tried to mount
>>> something". fdisk does not do mounting.
>>
>> Right, which is why I copied the entire screen output -- it appears to me that the rc scripts had stumbled on until the kernel panicked.
>>
>>>
>>> fdisk also shouldn't be writing to LBA 0 (the MBR) if you used -I -t.
>>> I've been staring at fdisk.c for about 20 minutes now and I can't work
>>> out a situation where -I -t would cause the MBR to be rewritten
>>> actively.
>>>
>>> The only GEOM calls I see in fdisk.c that would get called are
>>> g_device_path(), g_open(), and g_close(). Actual device I/O uses read()
>>> and write() (only in write_s0() which shouldn't be called).
>>>
>>> Furthermore, GEOM has foot-shooting-prevention mechanisms in place (I'm
>>> talking about kern.geom.debugflags) to keep LBA 0 from being modified.
>>> Is your script setting that sysctl to 16/0x10 blindly? Ahem.
>>
>> No. The script is intended only to work for drives other than the one containing the boot partition.
>>
>>>
>>> It would also help if you could state exactly what 9.1-STABLE source
>>> you're using; if using svn provide revision (rXXXXXX), else provide
>>> uname -a output.
>>
>> rev 249788
>>
>>>
>>> Finally: I would suggest using gpart(8) instead going forward. This is
>>> a separate recommendation though; if somehow I'm overlooking something
>>> in fdisk.c where writes to LBA 0 really do happen, then that needs to
>>> get fixed. But gpart(8) is what you should use in general these days
>>> anyway.
>>>
>>
>> Seems like gpart was giving me some frustration with earlier versions of FreeBSD (7, I think) so I went with fdisk instead. Might work OK now...
>>
>> I have included the full script below.
>>
>> { snipping for brevity; for reference, see this url: }
>> { http://lists.freebsd.org/pipermail/freebsd-stable/2013-April/073234.html }
>
> Thanks for this.
>
> I could practically write a book on what's going on here. Rather than
> me spend hours of time reverse-engineering this, you're going to need to
> step up to the plate and see if you can figure out what exactly triggers
> the issue.
>
> I will give you this analysis about fdisk -I -t:
>
> When -I is specified, I_flag=1.
>
> When -t is specified, v_flag=1, and also v_flag=1.
>
> Function open_disk(), when fdisk is used with the -I option, will call
> g_open() with the read-write flag set to 1. Whether or not this
> succeeds I don't know (and if it fails, but only with EPERM, then it
> retries in read-only mode silently). The -I flag correlates with the
> I_flag variable (do not confuse this with i_flag):
>
> 726 static int
> 727 open_disk(int flag)
> 728 {
> 729 int rwmode;
> 730
> 731 /* Write mode if one of these flags are set. */
> 732 rwmode = (a_flag || I_flag || B_flag || flag);
> 733 fd = g_open(disk, rwmode);
> 734 /* If the mode fails, try read-only if we didn't. */
> 735 if (fd == -1 && errno == EPERM && rwmode)
> 736 fd = g_open(disk, 0);
> 737 if (fd == -1 && errno == ENXIO)
> 738 return -2;
> 739 if (fd == -1) {
> 740 warnx("can't open device %s", disk);
> 741 return -1;
> 742 }
> 743 if (get_params() == -1) {
> 744 warnx("can't get disk parameters on %s", disk);
> 745 return -1;
> 746 }
> 747 return fd;
> 748 }
>
> Variable fd is global.
>
> After this call to open_disk(), read_disk() is used, but that's only
> doing read operations on fd.
>
> After this, the if (I_flag) code gets run. This calls read_s0(),
> reset_boot() (sounds ominous but isn't), and dos().
>
> read_s0() does not issue any write I/O to fd, or call any functions that
> issue write I/O.
>
> reset_boot() just resets the in-memory-copy of the partition table.
> It does not modify anything on disk.
>
> dos() does not do any I/O at all.
>
> At this point, if v_flag is set (which it is), print_s0() gets run.
>
> print_s0() calls print_params(), which simply prints out the
> in-memory-copy of C/H/S from the disk label and so on. No file I/O is
> done. Once that's done, it calls print_part() on each partition,
> which just outputs all the details -- again, no file I/O is done.
>
> Finally, at this stage, if t_flag ISN'T set, then write_s0() gets run.
> In this case write_s0() does not get called because t_flag=1. FYI,
> write_s0() is what does the actual write I/O to LBA 0/MBR. After that,
> exit(0) is called.
>
> So even though -I -t calls g_open() with the read-write flag set, I
> don't see anything that indicates writing to LBA 0/MBR happens.
>
> So I do not see how fdisk -I -t could cause this situation.
>
> fdisk -v, maybe, but again, you'll need to do the testing.
Thanks for the analysis.
> Now I have a question for you: how did you manage to get this output?
>
>>>> + fdisk -It ada0
>>>> + /rescue/sed -En 's,.*start ([0-9]+).*size ([0-9]+).*,\1 + \2,p'
>
> Because this looks like /bin/sh -x output, but I need to know if that's
> the case or not.
Yes, I inserted "set -x" at the top of the function to get this output.
> /bin/sh -x claims to echo commands to stderr ***before*** they're
> executed.
>
> So I'm then left wondering why we don't see output that equates to the
> equivalent of this line:
>
> eval $(fdisk -v $DISK | $SED -En 's,.*start ([0-9]+).*size ([0-9]+).*,curroff=\1 currsize=\2,p')
>
> Instead, we start seeing this:
>
>>>> vnode_pager_getpages: I/O read error
>>>> vm_fault: pager read error, pid 65 (fdisk)
>>>> pid 65 (fdisk), uid 0: exited on signal 11
>>>> eval: arithmetic expression: expecting primary: ""
>>> Entropy harvesting: point_to_pointeval: date: Device not configured
>>> eval: df: Device not configured
>>> eval: dmesg: Device not configured
>>> cat: /bin/ls: Device not configured
>
> Your script has only 1 eval statement (and eval is very very dangerous.
> I cannot stress this enough. If you ever think you need eval in shell
> scripts, you probably don't.)
>
> Your script does not call df, dmesg, date, or /bin/ls. So why are these
> mentioned? And "Entropy harvesting" comes from dmesg/the kernel message
> buffer too, how is that ending up there?
Hmm, maybe the quoting inside the $() is not working the way I expected. Perhaps the * symbols are being expanded as wildcards by the shell as a result of the eval.
> Possibly the eval: error line only gets output by sh ***after*** all the
> preceding [broken] stuff gets run.
>
> But I'm also confused, because there isn't anything arithmetic-oriented
> in your eval line, so why is it talking about arithmetic expressions?
> You don't use expr either, so the only math operation comes BEFORE all
> of that, specifically here:
>
> physsize=$(($(fdisk -It $DISK | $SED -En 's,.*start ([0-9]+).*size ([0-9]+).*,\1 + \2,p')))
>
> My gut feeling here is that something "unexpected" happened and your
> script went totally haywire as a result (probably some unexpected output
> that got turned into something you didn't expect). My favourite is
> seeing asterisk/wildcards expanded to pull in all the filenames in $cwd.
>
> I'm sorry to tell you, but there is a point when writing shell scripts
> becomes unreliable/unmanageable/results in too much risk, and is time to
> consider writing such things in an actual programming language
> (preferably one without reliance on CLI tools, but real APIs). I know
> you don't need to hear that right now, but it's true.
There are few tools available at the time this code needs to run -- note the use of sed from /rescue as an example. The only alternative would probably be C code. Still seems odd that userland code could crash the kernel this way.
> See if you can work out exactly what line begins causing problems for
> you. My guess is that it's the result of fdisk segfaulting, but I'm
> honestly not sure because the above output doesn't make entire sense.
>
> Let us know what you determine/find out.
Thanks for your time, Jeremy.
Guy
More information about the freebsd-stable
mailing list