Date:   Thu, 6 May 2021 15:08:08 +0200
From:   Peter Rosin <peda@...ntia.se>
To:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>
Cc:     linux-kernel@...r.kernel.org,
        Atul Gopinathan <atulgopinathan@...il.com>,
        Jens Axboe <axboe@...nel.dk>, stable <stable@...r.kernel.org>
Subject: Re: [PATCH 27/69] cdrom: gdrom: deallocate struct gdrom_unit fields
 in remove_gdrom

Hi!

On 2021-05-06 12:24, Greg Kroah-Hartman wrote:
> On Mon, May 03, 2021 at 04:13:18PM +0200, Peter Rosin wrote:
>> Hi!
>>
>> On 2021-05-03 13:56, Greg Kroah-Hartman wrote:
>>> From: Atul Gopinathan <atulgopinathan@...il.com>
>>>
>>> The fields, "toc" and "cd_info", of "struct gdrom_unit gd" are allocated
>>> in "probe_gdrom()". Prevent a memory leak by making sure "gd.cd_info" is
>>> deallocated in the "remove_gdrom()" function.
>>>
>>> Also prevent double free of the field "gd.toc" by moving it from the
>>> module's exit function to "remove_gdrom()". This is because, in
>>> "probe_gdrom()", the function makes sure to deallocate "gd.toc" in case
>>> of any errors, so the exit function invoked later would again free
>>> "gd.toc".
>>>
>>> The patch also maintains consistency by deallocating the above mentioned
>>> fields in "remove_gdrom()" along with another memory allocated field
>>> "gd.disk".
>>>
>>> Suggested-by: Jens Axboe <axboe@...nel.dk>
>>> Cc: Peter Rosin <peda@...ntia.se>
>>> Cc: stable <stable@...r.kernel.org>
>>> Signed-off-by: Atul Gopinathan <atulgopinathan@...il.com>
>>> Signed-off-by: Greg Kroah-Hartman <gregkh@...uxfoundation.org>
>>> ---
>>>  drivers/cdrom/gdrom.c | 3 ++-
>>>  1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/cdrom/gdrom.c b/drivers/cdrom/gdrom.c
>>> index 7f681320c7d3..6c4f6139f853 100644
>>> --- a/drivers/cdrom/gdrom.c
>>> +++ b/drivers/cdrom/gdrom.c
>>> @@ -830,6 +830,8 @@ static int remove_gdrom(struct platform_device *devptr)
>>>  	if (gdrom_major)
>>>  		unregister_blkdev(gdrom_major, GDROM_DEV_NAME);
>>>  	unregister_cdrom(gd.cd_info);
>>> +	kfree(gd.cd_info);
>>> +	kfree(gd.toc);
>>>  
>>>  	return 0;
>>>  }
>>> @@ -861,7 +863,6 @@ static void __exit exit_gdrom(void)
>>>  {
>>>  	platform_device_unregister(pd);
>>>  	platform_driver_unregister(&gdrom_driver);
>>> -	kfree(gd.toc);
>>>  }
>>>  
>>>  module_init(init_gdrom);
>>>
>>
>> I worry about the gd.toc = NULL; statement in init_gdrom(). It sets off
>> all kinds of warnings with me. It looks completely bogus, but the fact
>> that it's there at all makes me go hmmmm.
> 
> Yeah, that's bogus.
> 
>> probe_gdrom_setupcd() will arrange for gdrom_ops to be used, including
>> .get_last_session pointing to gdrom_get_last_session() 
>>
>> gdrom_get_last_session() will use gd.toc, if it is non-NULL.
>>
>> The above will all be registered externally to the driver with the call
>> to register_cdrom() in probe_gdrom(), before a possible stale gd.toc is
>> overwritten with a new one at the end of probe_gdrom().
> 
> But can that really happen given that it hasn't ever happened before in
> a real system?  :)
> 
>> Side note, .get_last_session is an interesting name in this context, but
>> I have no idea if it might be called in the "bad" window (but relying on
>> that to not be the case would be ... subtle).
>>
>> So, by simply freeing gd.toc in remove_gdrom() without also setting
>> it to NULL, it looks like a potential use after free of gd.toc is
>> introduced, replacing a potential leak. Not good.
> 
> So should we set it to NULL after freeing it?  Is that really going to
> help here given that the probe failed?  Nothing can use it after
> remove_gdrom() is called because unregiser_* is called already.
> 
> I don't see the race here, sorry.
> 
>> The same is not true for gd.cd_info as far as I can tell, but it's a bit
>> subtle. gdrom_probe() calls gdrom_execute_diagnostics() before the stale
>> gd.cd_info is overwritten, and gdrom_execute_diagnostic() passes the
>> stale pointer to gdrom_hardreset(), which luckily doesn't use it. But
>> this is - as hinted - a bit too subtle for me. I would prefer to have
>> remove_gdrom() also clear out the gd.cd_info pointer.
> 
> Ok, but again, how can that be used after remove_gdrom() is called?
> 
>> In addition to adding these clears of gd.toc and gd.cd_info to
>> remove_gdrom(), they also need to be cleared in case probe fails.
>>
>> Or instead, maybe add a big fat
>> 	memset(&gd, 0, sizeof(gd));
>> at the top of probe?
> 
> Really, that's what is happening today as there is only 1 device here,
> and the whole structure was zeroed out already.  So that would be a
> no-op.
> 
>> Or maybe the struct gdrom_unit should simply be kzalloc:ed? But that
>> triggers some . to -> churn...
> 
> Yes, ideally that would be the correct change, but given that you can
> only have 1 device in the system at a time of this type, it's not going
> to make much difference at all here.
> 
>> Anyway, the patch as proposed gets a NACK from me.
> 
> Why?  It fixes the obvious memory leak, right?  Worst case you are
> saying we should also set to NULL these pointers, but I can not see how
> they are accessed as we have already torn everything down.

I'm thinking this:

1. init_gdrom() is called. gd.toc is NULL and is bogusly re-set to NULL.
2. probe_gdrom() is called and succeeds. gd.toc is allocted.
3. device is used, etc etc, whatever
4. remove_gdrom() is called. gd.toc is freed (but not set to NULL).
5. probe_gdrom() is called again. Boom.

In 5, gd.toc is not NULL, and is pointing to whatever. It is
potentially used by probe_gdrom() before it is (re-)allocated.

I suppose the above can only happen if the module is compiled in.

Without this patch, we are "safe" because gd.toc still points to the old
thing which is leaked once a new gd.toc is allocated by the second probe.

Cheers,
Peter

Powered by blists - more mailing lists