Re: [syzbot] KASAN: use-after-free Read in kernfs_next_descendant_post (2)

[Dmitry Torokhov <dmitry.torokhov@gmail.com>]

On Mon, Nov 14, 2022 at 09:34:16AM -0800, Luis Chamberlain wrote:
> On Mon, Oct 31, 2022 at 12:53:00PM -1000, Tejun Heo wrote:
> > (cc'ing Luis for firmware loader and quoting the whole body)
> > 
> > On Sat, Oct 22, 2022 at 06:52:28AM +0800, Hillf Danton wrote:
> > > On 20 Oct 2022 00:15:40 -0700
> > > > syzbot has found a reproducer for the following issue on:
> > > > 
> > > > HEAD commit:    55be6084c8e0 Merge tag 'timers-core-2022-10-05' of git://g..
> > > > git tree:       upstream
> > > > console+strace: https://syzkaller.appspot.com/x/log.txt?x=1449d53c880000
> > > > kernel config:  https://syzkaller.appspot.com/x/.config?x=df75278aabf0681a
> > > > dashboard link: https://syzkaller.appspot.com/bug?extid=6bc35f3913193fe7f0d3
> > > > compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.2
> > > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=14e01c72880000
> > > > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=1128908c880000
> > > 
> > > See if the change to ueagle driver alone can survive syzbot test.
> > > 
> > > #syz test https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git  aae703b02f92
> > > 
> > > --- a/include/linux/netdevice.h
> > > +++ b/include/linux/netdevice.h
> > > @@ -3663,8 +3663,9 @@ static inline bool netif_attr_test_online(unsigned long j,
> > >  static inline unsigned int netif_attrmask_next(int n, const unsigned long *srcp,
> > >  					       unsigned int nr_bits)
> > >  {
> > > -	/* n is a prior cpu */
> > > -	cpu_max_bits_warn(n + 1, nr_bits);
> > > +	/* -1 is a legal arg here. */
> > > +	if (n != -1)
> > > +		cpu_max_bits_warn(n, nr_bits);
> > >  
> > >  	if (srcp)
> > >  		return find_next_bit(srcp, nr_bits, n + 1);
> > > @@ -3685,8 +3686,9 @@ static inline int netif_attrmask_next_and(int n, const unsigned long *src1p,
> > >  					  const unsigned long *src2p,
> > >  					  unsigned int nr_bits)
> > >  {
> > > -	/* n is a prior cpu */
> > > -	cpu_max_bits_warn(n + 1, nr_bits);
> > > +	/* -1 is a legal arg here. */
> > > +	if (n != -1)
> > > +		cpu_max_bits_warn(n, nr_bits);
> > >  
> > >  	if (src1p && src2p)
> > >  		return find_next_and_bit(src1p, src2p, nr_bits, n + 1);
> > > --- a/drivers/usb/atm/ueagle-atm.c
> > > +++ b/drivers/usb/atm/ueagle-atm.c
> > > @@ -597,9 +597,8 @@ static int uea_send_modem_cmd(struct usb
> > >  }
> > >  
> > >  static void uea_upload_pre_firmware(const struct firmware *fw_entry,
> > > -								void *context)
> > > +					struct usb_device *usb)
> > >  {
> > > -	struct usb_device *usb = context;
> > >  	const u8 *pfw;
> > >  	u8 value;
> > >  	u32 crc = 0;
> > > @@ -679,6 +678,7 @@ static int uea_load_firmware(struct usb_
> > >  {
> > >  	int ret;
> > >  	char *fw_name = EAGLE_FIRMWARE;
> > > +	const struct firmware *fw;
> > >  
> > >  	uea_enters(usb);
> > >  	uea_info(usb, "pre-firmware device, uploading firmware\n");
> > > @@ -701,13 +701,13 @@ static int uea_load_firmware(struct usb_
> > >  		break;
> > >  	}
> > >  
> > > -	ret = request_firmware_nowait(THIS_MODULE, 1, fw_name, &usb->dev,
> > > -					GFP_KERNEL, usb,
> > > -					uea_upload_pre_firmware);
> > > +	ret = request_firmware(&fw, fw_name, &usb->dev);
> > >  	if (ret)
> > >  		uea_err(usb, "firmware %s is not available\n", fw_name);
> > > -	else
> > > +	else {
> > >  		uea_info(usb, "loading firmware %s\n", fw_name);
> > > +		uea_upload_pre_firmware(fw, usb);
> > > +	}
> > >  
> > >  	uea_leaves(usb);
> > >  	return ret;
> > 
> > So, the problem is that while request_firmware_nowait() inc's the ref on the
> > device, if the device gets removed later, having a ref isn't sufficient for
> > adding stuff to the device. A relatively easy solution would be putting
> > these firmware load work items into its own workqueue and flushing it on
> > device removal path. Luis, what do you think?
> 
> Since we *can* remove a device after we get a module reference and
> since fw_cache_is_setup() tries to use the device before get_device()
> (even though this is not the issue reported), I think perhaps the fix
> below may be generic and best.

I do not see how moving the point where we acquire device refcount
around fixes anything. Caller of request_firmware_nowait() is supposed
to have a valid reference to device object and it is supposed to stay
valid for the entire duration of request_firmware_nowait(). Grabbing
and extra reference only matters if the device (or other refcounted
structure) is being passed to another thread of execution.

I think what Tejun is saying is the only way to fix this. Similarly to
work struct, where users are supposed to call cancel_work_sync() during
teardown, users of request_firmware_nowait() need to wait for it to
complete before continuing with tearing down the instance. See for
example ims-pcu driver where it tries to request firmware asynchronously
when it finds the device in bootloader mode, and is waiting for it
completion when handling device disconnect:

https://elixir.bootlin.com/linux/v6.1-rc3/source/drivers/input/misc/ims-pcu.c#L1978

Thanks.

-- 
Dmitry