Re: [bug report] memory corruption panic caused by SG_IO ioctl()

[Douglas Gilbert]

On 2018-08-03 12:17 PM, Douglas Gilbert wrote:
> On 2018-08-03 11:47 AM, gaowanlong wrote:
> > Doug,
> >
> > On 2018-08-03 04:46 AM, Wanlong Gao wrote:
> > > Hi Martinand all folks,
> > >
> > >
> > > > Recently we find a kernel panic with memory corruption caused by SG_IO ioctl(),
> > > > and it can be easily reproduced by running following reproducer about
> > > > minutes,any idea?
> >
> > > Which kernel?
> >
> > We've tested with 4.17.11 and 4.18.rc7 and both reproduced.
> >
> > > And what are the underlying devices (e.g. does /dev/sg0 refer to a SATA disk,
> > > a real SCSI disk (SAS for example), USB mass storage, etc)?
> >
> > We tested in a qemu-kvm guest and the sg0 refer to a virtual SATA disk.
>
> Thanks for the prompt reply.
>
> The first test I am doing, and you can also do, is to replace the virtual
> SATA disk with a scsi_debug pseudo SCSI disk(s). This will tell us
> whether libata has a hand in this (as that was the case in a previous
> syzkaller report on the SG_IO ioctl()).
>
> > > Also can you get a copy of the kernel panic?
> >
> > Since the call traces are different every time it reproduced, that I didn't 
> > paste the
> > call trace or the vmcore, but this reproducer is very useful and I believe you 
> > can reproduce
> > it easily using the following code.
>
> Okay.
>
> As I write I'm running your reproducer with lk 4.18.0-rc6 against pseudo
> scsi_debug "disks". So far no problems (5 minutes) with no noise in syslog.

Ran for an hour before I stopped it. Before that I did a
  echo 1 > /sys/bus/pseudo/drivers/scsi_debug/opts

which causes a lot of noise in syslog. Then I could see every command was
being rejected with "LBA out of range". So I restarted scsi_debug with this:

  modprobe scsi_debug max_luns=8 sector_size=4096 virtual_gb=2000 ndelay=5000

To give 8 pseudo scsi disks of 2 TB size. Then it worked, this from syslog:
  sd 0:0:0:0: scsi_debug: tag=0x7e, cmd 08 f0 a8 77 d3 be 87 5d da 65 79 3f c7

That is certainly strange, a READ(6) [deprecated] with 13 bytes in the command!
But it doesn't seem to hurt scsi_debug. Still running 15 minutes later ...

Doug Gilbert