Re: Bad kfree of dma_parms in v5.7-rc5

[Ulf Hansson]

Marek, Tomi, Greg

On Wed, 20 May 2020 at 17:14, Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
>
> On Wed, 20 May 2020 at 15:28, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote:
> >
> > Hi Ulf,
> >
> > On 20.05.2020 15:12, Ulf Hansson wrote:
> > > + Greg
> > >
> > > On Wed, 20 May 2020 at 14:54, Marek Szyprowski <m.szyprowski@xxxxxxxxxxx> wrote:
> > >> On 20.05.2020 14:43, Tomi Valkeinen wrote:
> > >>> On 20/05/2020 12:22, Marek Szyprowski wrote:
> > >>>> On 20.05.2020 11:18, Tomi Valkeinen wrote:
> > >>>>> On 20/05/2020 12:13, Marek Szyprowski wrote:
> > >>>>>> On 20.05.2020 11:00, Tomi Valkeinen wrote:
> > >>>>>>> Commit 9495b7e92f716ab2bd6814fab5e97ab4a39adfdd ("driver core:
> > >>>>>>> platform: Initialize dma_parms for platform devices") v5.7-rc5 causes
> > >>>>>>> at least some v4l2 platform drivers to break when freeing resources.
> > >>>>>>>
> > >>>>>>> E.g. drivers/media/platform/ti-vpe/cal.c uses
> > >>>>>>> vb2_dma_contig_set_max_seg_size() and
> > >>>>>>> vb2_dma_contig_clear_max_seg_size() to manage the dma_params, and
> > >>>>>>> similar pattern is seen in other drivers too.
> > >>>>>>>
> > >>>>>>> After 9495b7e92f716ab2, vb2_dma_contig_set_max_seg_size() will not
> > >>>>>>> allocate anything, but vb2_dma_contig_clear_max_seg_size() will still
> > >>>>>>> kfree the dma_params.
> > >>>>>>>
> > >>>>>>> I'm not sure what's the proper fix here. A flag somewhere to indicate
> > >>>>>>> that vb2_dma_contig_set_max_seg_size() did allocate, and thus
> > >>>>>>> vb2_dma_contig_clear_max_seg_size() must free?
> > >>>>>>>
> > >>>>>>> Or drop the kzalloc and kfree totally, if dma_params is now supposed
> > >>>>>>> to always be there?
> > >>>>>> Thanks for reporting this issue!
> > >>>>>>
> > >>>>>> Once the mentioned commit has been merged, the code should assume that
> > >>>>>> the platform devices does have a struct dma_params allocated, so the
> > >>>>>> proper fix is to alloc dma_params only if the bus is not a platform
> > >>>>>> bus:
> > >>>>>>
> > >>>>>> if (!dev_is_platform(dev) && !dev->dma_parms) {
> > >>>>>>         dev->dma_parms = kzalloc(sizeof(*dev->dma_parms), GFP_KERNEL);
> > >>>>>>
> > >>>>>> same check for the free path.
> > >>>>> There is also "amba: Initialize dma_parms for amba devices". And the
> > >>>>> commit message says PCI devices do this too.
> > >>>>>
> > >>>>> Guessing this based on the device type doesn't sound like a good idea
> > >>>>> to me.
> > >>>> Indeed. Then replace the allocation with a simple check for NULL
> > >>>> dma_parms and return an error in such case. This should be enough for
> > >>>> v5.8. Later we can simply get rid of those helpers and inline setting
> > >>>> max segment size directly to the drivers.
> > > That seems like a good idea, in the long run.
> > >
> > >>> Is that valid either? Then we assume that dma_parms is always set up
> > >>> by someone else. That's true for platform devices and apparently some
> > >>> other devices, but is it true for all devices now?
> > >> # git grep vb2_dma_contig_set_max_seg_size | wc -l
> > >>
> > >> 18
> > >>
> > >> I've checked all clients of the vb2_dma_contig_set_max_seg_size
> > >> function. There are only 9 drivers, all of them are platform device
> > >> drivers. We don't care about off-tree users, so the proposed approach is
> > >> imho fine.
> > > Thanks for reporting and for looking into this. I apologize for the mess!
> > >
> > > There is one case, where the above solution could be a problem (unless
> > > I am wrong). That is, s5p_mfc_configure_2port_memory() that calls
> > > s5p_mfc_alloc_memdev(), which allocates/initializes an internal struct
> > > *device. Thus, this doesn't have the dev->dma_parms
> > > allocated/assigned.
> > Indeed, this one will fail.
> > > In other words, we would need to manage alloc/free for the
> > > dev->dma_parms to have a complete fix. Maybe in
> > > s5p_mfc_configure|unconfigure_2port_memory()!?
> > That would be the best place to allocate it.
> > > Additionally, I think reverting the offending commit, as discussed
> > > above, could cause even more issues, as it's even included for
> > > v5.6-stable kernels. I will go through all cases, more carefully this
> > > time, of how ->dma_parms is managed, to be sure there are no more
> > > conflicting cases.
> >
> > I've already posted a fix for ExynosDRM driver, which is also affected:
> > https://patchwork.kernel.org/patch/11559965/
>
> Alright, thanks for helping out!
>
> Please add a fixes/stable tag to it.
>
> Fixes: 9495b7e92f71 ("driver core: platform: Initialize dma_parms for
> platform devices")
> Cc: stable@xxxxxxxxxxxxxxx
>

FYI:

I have now double checked all cases where ->dma_params are being
allocated/freed. Besides those you (Marek/Tomi) you have found and
sent fixes for (many thanks!)  - I haven't found any additional cases
to worry about.

However, of course there are cleanups and removal of redundant code
that can be made, for some drivers/devices, which are based upon a
platform device. For example, some have their own "struct
device_dma_parameters", such as drivers/dma/dma-axi-dmac.c for
example. This is not a problem, but deserves to be cleaned up. I have
started to prepare patches for it.

Kind regards
Uffe