mbox series

[RFC,0/2] dma-mapping: introduce helper for setting dma_parms

Message ID 20180911163050.28072-1-wsa+renesas@sang-engineering.com (mailing list archive)
Headers show
Series dma-mapping: introduce helper for setting dma_parms | expand

Message

Wolfram Sang Sept. 11, 2018, 4:30 p.m. UTC
Hi all,

commit 78c47830a5cb ("dma-debug: check scatterlist segments") triggers
for Renesas hardware I look after, so thanks for pointing out we should
have proper dma_parms for our DMA providers.

When trying to fix it, I became a bit puzzled about the life cycle of
the pointer to dma_parms. AFAIU most drivers leave the pointer dangling
on driver unbind. Check drivers/dma/bcm2835-dma.c, for example:

	od = devm_kzalloc(&pdev->dev, sizeof(*od), GFP_KERNEL);
	if (!od)
		return -ENOMEM;

	pdev->dev.dma_parms = &od->dma_parms;
	dma_set_max_seg_size(&pdev->dev, 0x3FFFFFFF);

And that's all about handling dma_parms. So, on unbind, the memory for
'od' gets freed and dma_params is a dangling pointer.

drivers/gpu/drm/exynos/exynos_drm_iommu.c seems to do it correctly:

static inline int configure_dma_max_seg_size(struct device *dev)
{
        if (!dev->dma_parms)
                dev->dma_parms = kzalloc(sizeof(*dev->dma_parms), GFP_KERNEL);
        if (!dev->dma_parms)
                return -ENOMEM;

        dma_set_max_seg_size(dev, DMA_BIT_MASK(32));
        return 0;
}

static inline void clear_dma_max_seg_size(struct device *dev)
{
        kfree(dev->dma_parms);
        dev->dma_parms = NULL;
}

But this seems error prone and quite some code to add for every DMA
provider. So, I wondered if we couldn't have a helper for that. After
some brainstorming, I favour a dmam_-type of function. It will ensure
the memory gets freed and the pointer cleared on unbind. And it should
be easy to use.

I attached an RFC which I tested on a Renesas R-Car H3 SoC with the internal
DMAC of the SD controller. A branch can be found here (still waiting for
buildbot results):

git://git.kernel.org/pub/scm/linux/kernel/git/wsa/linux.git renesas/sdhi/set_max_seg

I added the companion function dmam_free_dma_parms() for completeness
although there is no user yet. I'd be totally open to drop it until
someone needs it.

Please let me know what you think. If this is the right track, I'll be
willing to fix the dangling pointers with it, too.

Thanks and happy hacking,

   Wolfram



Wolfram Sang (2):
  dma-mapping: introduce helper for setting dma_parms
  mmc: sdhi: internal_dmac: set dma_parms

 drivers/mmc/host/renesas_sdhi_internal_dmac.c |  2 +
 include/linux/dma-mapping.h                   |  5 ++
 kernel/dma/mapping.c                          | 50 +++++++++++++++++++
 3 files changed, 57 insertions(+)

Comments

Robin Murphy Sept. 11, 2018, 5:24 p.m. UTC | #1
On 11/09/18 17:30, Wolfram Sang wrote:
> Hi all,
> 
> commit 78c47830a5cb ("dma-debug: check scatterlist segments") triggers
> for Renesas hardware I look after, so thanks for pointing out we should
> have proper dma_parms for our DMA providers.
> 
> When trying to fix it, I became a bit puzzled about the life cycle of
> the pointer to dma_parms. AFAIU most drivers leave the pointer dangling
> on driver unbind. Check drivers/dma/bcm2835-dma.c, for example:
> 
> 	od = devm_kzalloc(&pdev->dev, sizeof(*od), GFP_KERNEL);
> 	if (!od)
> 		return -ENOMEM;
> 
> 	pdev->dev.dma_parms = &od->dma_parms;
> 	dma_set_max_seg_size(&pdev->dev, 0x3FFFFFFF);
> 
> And that's all about handling dma_parms. So, on unbind, the memory for
> 'od' gets freed and dma_params is a dangling pointer.

That's the typical case - the dma_parms structure is simply embedded in 
some other private data which tends to have the appropriate lifetime 
anyway. I can't see that the dangling pointer is an issue when it's set 
unconditionally on probe and valid until remove, because anyone 
dereferencing dev->dma_parms when dev has no driver bound is doing 
something very very wrong anyway. I suppose we could zero it in 
__device_release_driver() for good measure though - shame we've found 
something dma_deconfigure() could have been useful for just after we 
killed it ;)

> 
> drivers/gpu/drm/exynos/exynos_drm_iommu.c seems to do it correctly:

...but could be even simpler if it was just included in 
exynos_drm_private. Realistically, I struggle to imagine a driver which 
would need to set dma_parms that didn't already have some other private 
state into which it could fit.

Note that ultimately we'd like to have a single structure to hold all 
the masks and other gubbins (per the very original intent of dma_parms), 
so there's a fair chance of this getting fundamentally rejigged at some 
point anyway.

Robin.

> static inline int configure_dma_max_seg_size(struct device *dev)
> {
>          if (!dev->dma_parms)
>                  dev->dma_parms = kzalloc(sizeof(*dev->dma_parms), GFP_KERNEL);
>          if (!dev->dma_parms)
>                  return -ENOMEM;
> 
>          dma_set_max_seg_size(dev, DMA_BIT_MASK(32));
>          return 0;
> }
> 
> static inline void clear_dma_max_seg_size(struct device *dev)
> {
>          kfree(dev->dma_parms);
>          dev->dma_parms = NULL;
> }
> 
> But this seems error prone and quite some code to add for every DMA
> provider. So, I wondered if we couldn't have a helper for that. After
> some brainstorming, I favour a dmam_-type of function. It will ensure
> the memory gets freed and the pointer cleared on unbind. And it should
> be easy to use.
> 
> I attached an RFC which I tested on a Renesas R-Car H3 SoC with the internal
> DMAC of the SD controller. A branch can be found here (still waiting for
> buildbot results):
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/wsa/linux.git renesas/sdhi/set_max_seg
> 
> I added the companion function dmam_free_dma_parms() for completeness
> although there is no user yet. I'd be totally open to drop it until
> someone needs it.
> 
> Please let me know what you think. If this is the right track, I'll be
> willing to fix the dangling pointers with it, too.
> 
> Thanks and happy hacking,
> 
>     Wolfram
> 
> 
> 
> Wolfram Sang (2):
>    dma-mapping: introduce helper for setting dma_parms
>    mmc: sdhi: internal_dmac: set dma_parms
> 
>   drivers/mmc/host/renesas_sdhi_internal_dmac.c |  2 +
>   include/linux/dma-mapping.h                   |  5 ++
>   kernel/dma/mapping.c                          | 50 +++++++++++++++++++
>   3 files changed, 57 insertions(+)
>
Wolfram Sang Sept. 11, 2018, 11:39 p.m. UTC | #2
Hi Robin,

> > 	od = devm_kzalloc(&pdev->dev, sizeof(*od), GFP_KERNEL);
> > 	if (!od)
> > 		return -ENOMEM;
> > 
> > 	pdev->dev.dma_parms = &od->dma_parms;
> > 	dma_set_max_seg_size(&pdev->dev, 0x3FFFFFFF);
> > 
> > And that's all about handling dma_parms. So, on unbind, the memory for
> > 'od' gets freed and dma_params is a dangling pointer.
> 
> That's the typical case - the dma_parms structure is simply embedded in some
> other private data which tends to have the appropriate lifetime anyway. I
> can't see that the dangling pointer is an issue when it's set
> unconditionally on probe and valid until remove, because anyone
> dereferencing dev->dma_parms when dev has no driver bound is doing something
> very very wrong anyway. I suppose we could zero it in
> __device_release_driver() for good measure though - shame we've found
> something dma_deconfigure() could have been useful for just after we killed
> it ;)

I see. Yes, I was aware that the misuse of this dangling pointer is
somewhat academical. Yet, it was easy to fix and clearing this pointer
is good programming practice, I'd say. I agree that clearing the pointer
in __device_release_driver is a good option, too, if documentation about
its expected life cycle (== get's cleared on unbind) is clear about
that. Probably that life cycle confusion led to the more complicated
code in the exynos_drm driver. I will look into all of that tomorrow.

> Note that ultimately we'd like to have a single structure to hold all the
> masks and other gubbins (per the very original intent of dma_parms), so

I was wondering about that, yes.

> there's a fair chance of this getting fundamentally rejigged at some point
> anyway.

Makes sense. Yet, as this change is gonna be small, I think it's still
nice to have.

Thanks for the input!

   Wolfram