Date:   Wed, 28 Aug 2019 12:35:57 +0200
From:   Jorge Ramirez <jorge.ramirez-ortiz@...aro.org>
To:     Srinivas Kandagatla <srinivas.kandagatla@...aro.org>
Cc:     Greg KH <gregkh@...uxfoundation.org>, arnd@...db.de,
        linux-arm-msm <linux-arm-msm@...r.kernel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Mayank Chopra <mak.chopra@...eaurora.org>
Subject: Re: [PATCH 4/5] misc: fastrpc: fix double refcounting on dmabuf

On 8/28/19 10:48, Srinivas Kandagatla wrote:
> 
> 
> On 28/08/2019 08:50, Jorge Ramirez wrote:
>> On 8/27/19 23:45, Srinivas Kandagatla wrote:
>>> On 23/08/2019 16:23, Jorge Ramirez-Ortiz wrote:
>>>> can you add me as a co-author to this patch please?
>>> No problem I can do that if you feel so!
>> yes please. thanks!
>>
>>>> since I spent about a day doing the analysis, sent you  a fix that
>>>> maintained the API used by the library and explained you how to
>>>> reproduce the issue I think it is just fair. > the  fact that the api
>>>> could be be modified and the fix be done a bit
>>>> differently- free dma buf ioctl removed- seems just a minor
>>>> implementation detail to me.
>>> No, that's not true, this is a clear fastrpc design issue.
>> IMO the ioctls defines the contract with userspace and the contract
>> establishes that userspace must call deallocate. the kernel wrongly
>> implemented to that contract since it doesn't handle the cases where
>> userspace can't send the release calls which leads to memory leaks. this
>> is what I meant by and implementation issue.
>>
>> if we had many fastrpc users, rolling out the design change that you
>> propose - removing an ioctl- would definitively have an impact. But
>> since that is not yet the case, there is not doubt that your patch makes
>> more sense.
> 
> Exactly before it make a way into other projects!
> 
>>
>> but my point was that there is not a huge gap in efforts between doing
>> one or the other.
> 
> Thats not the point, point is about right fix!

but I disagree with you about what 'right' means.

in this context, for you "right" meant potentially breaking some users
and implement the best possible kernel design.

for me, it meant continue to obey at the agreed ioctl interface to not
disturb the users.

but as I said, since there is not a significant pool of fastrpc users,
breaking backward compatibility with the fastrpc library is not
important hence why I agree that removing the ioctl was the better
choice (on this particular case).

> 
>>
>>> Userspace is already doing a refcount via mmap/unmap on that dmabuf fd,
>>> having an additional api adds another level of refcount which is totally
>>> redundant and is the root cause for this leak.
>> yes it is redundant but is not the root cause for this leak. the root
>> cause is that the driver doesnt handle the case where userspace didnt or
>> was not able to call release (and that is no more than adding allocated
>> buffers to a list and clean on exit)
> 
> I don't agree with you on that. We should not take an extra refcount in
> first place in driver.

of course taking an extra refcount is functionally pointless. But that
was a design decision that imposed something on the user. and the kernel
can certainly work with that 'silly' design decision by tracking the
memory in the driver.

is the right thing to do to keep less than ideal designs in the kernel
to support agreed user interface? IMO it depends on the use case

to my eyes the design was obviously wrong, I never questioned
that...that was very clear when I started tracing the code.

perhaps, rather than work around it, I should have considered that
removing the ioctl wouldnt be a big deal to anyone.

so I would have send two patches instead of the one I sent you

1. fix leak (keep track of allocated dma buffers and make sure
everything is released on exit)

2. remove unnecessary ioctl warning users.


> 
> let me explain it one more time!

cmon, I did understand it before we engaged in this discussion :)



> 
> dmabuf has to be mmaped in userspace app before it is used, and
> "Memory mappings that were created in the process shall be unmapped
> before the process is destroyed" so the refcount is taken care by
> mmap/unmap automatically.


I would like to leave the discussion here if that is ok with you (I
clearly understand your POV but I feel I am not doing a good job at
sharing my thoughts...we can do that offline if you want)


> 
> --srini
> 
> 
>>
> 

Powered by blists - more mailing lists