On Wed, Feb 16, 2022 at 12:59:31AM +0100, Rob Herring wrote:
> On Tue, Feb 15, 2022 at 4:58 PM Rob Herring <robh@xxxxxxxxxx> wrote:
> >
> > On Mon, 14 Feb 2022 16:32:08 +0100, Mårten Lindahl wrote:
> > > When unbinding/binding a driver with DMA mapped memory, the DMA map is
> > > not freed when the driver is reloaded. This leads to a memory leak when
> > > the DMA map is overwritten when reprobing the driver.
> > >
> > > This can be reproduced with a platform driver having a dma-range:
> > >
> > > dummy {
> > > ...
> > > #address-cells = <0x2>;
> > > #size-cells = <0x2>;
> > > ranges;
> > > dma-ranges = <...>;
> > > ...
> > > };
> > >
> > > and then unbinding/binding it:
> > >
> > > ~# echo soc:dummy >/sys/bus/platform/drivers/<driver>/unbind
> > >
> > > DMA map object 0xffffff800b0ae540 still being held by &pdev->dev
> > >
> > > ~# echo soc:dummy >/sys/bus/platform/drivers/<driver>/bind
> > > ~# echo scan > /sys/kernel/debug/kmemleak
> > > ~# cat /sys/kernel/debug/kmemleak
> > > unreferenced object 0xffffff800b0ae540 (size 64):
> > > comm "sh", pid 833, jiffies 4295174550 (age 2535.352s)
> > > hex dump (first 32 bytes):
> > > 00 00 00 80 00 00 00 00 00 00 00 00 00 00 00 00 ................
> > > 00 00 00 80 00 00 00 00 00 00 00 80 00 00 00 00 ................
> > > backtrace:
> > > [<ffffffefd1694708>] create_object.isra.0+0x108/0x344
> > > [<ffffffefd1d1a850>] kmemleak_alloc+0x8c/0xd0
> > > [<ffffffefd167e2d0>] __kmalloc+0x440/0x6f0
> > > [<ffffffefd1a960a4>] of_dma_get_range+0x124/0x220
> > > [<ffffffefd1a8ce90>] of_dma_configure_id+0x40/0x2d0
> > > [<ffffffefd198b68c>] platform_dma_configure+0x5c/0xa4
> > > [<ffffffefd198846c>] really_probe+0x8c/0x514
> > > [<ffffffefd1988990>] __driver_probe_device+0x9c/0x19c
> > > [<ffffffefd1988cd8>] device_driver_attach+0x54/0xbc
> > > [<ffffffefd1986634>] bind_store+0xc4/0x120
> > > [<ffffffefd19856e0>] drv_attr_store+0x30/0x44
> > > [<ffffffefd173c9b0>] sysfs_kf_write+0x50/0x60
> > > [<ffffffefd173c1c4>] kernfs_fop_write_iter+0x124/0x1b4
> > > [<ffffffefd16a013c>] new_sync_write+0xdc/0x160
> > > [<ffffffefd16a256c>] vfs_write+0x23c/0x2a0
> > > [<ffffffefd16a2758>] ksys_write+0x64/0xec
> > >
> > > Don't get a new dma_range_map if there already is one. Check for an
> > > existing map and reuse it, or else get the map as before. This will
> > > prevent overwriting the old map which is still valid.
> > >
> > > Signed-off-by: Mårten Lindahl <marten.lindahl@xxxxxxxx>
> > > ---
> > >
> > > v2:
> > > - Reuse range map instead of getting a new and freeing the old.
> > >
> > > drivers/of/device.c | 9 +++++++--
> > > 1 file changed, 7 insertions(+), 2 deletions(-)
> > >
> >
> > Applied, thanks!
Hi Rob!
>
> Ummm, a bit too quick here. I was looking at the history to add a
> Fixes tag. Looking at commit 89c7cb1608ac ("of/device: Update
> dma_range_map only when dev has valid dma-ranges"), I think this
> change may cause a change in behavior as 'ret' is used as a 'we have a
> valid dma-ranges' flag. So if dma_range_map was set, but not from
> dma-ranges, the behavior will be different.
I really appreciate your second review on this. It is not very obvious
how the ret flag is used in that case, but I agree it may change
expected behavior.
>
> Instead, I think this needs to be something like the following patch.
> Really we have 3 occurrences of the same clean-up in dd.c and that
> should be refactored. But the below is probably better for a backport
> and refactoring should come after.
Actually, when I found the leak I saw two places to free the
dma_range_map, the first where I put it in patch1, and the second in
__device_release_driver. Now, as you also suggest, it seems the better
place to do it is in __device_release_driver. But I must admit I missed
the place in really_probe, so again, your eyes on this are very helpful.
I will make a new patch on dd instead. Thanks!
Kind regards
Mårten
>
> diff --git a/drivers/base/dd.c b/drivers/base/dd.c
> index 9eaaff2f556c..f0a508fd7913 100644
> --- a/drivers/base/dd.c
> +++ b/drivers/base/dd.c
> @@ -629,6 +632,9 @@ static int really_probe(struct device *dev, struct
> device_driver *drv)
> drv->remove(dev);
>
> devres_release_all(dev);
> + arch_teardown_dma_ops(dev);
> + kfree(dev->dma_range_map);
> + dev->dma_range_map = NULL;
> driver_sysfs_remove(dev);
> dev->driver = NULL;
> dev_set_drvdata(dev, NULL);
> @@ -1209,6 +1215,8 @@ static void __device_release_driver(struct
> device *dev, struct device *parent)
>
> devres_release_all(dev);
> arch_teardown_dma_ops(dev);
> + kfree(dev->dma_range_map);
> + dev->dma_range_map = NULL;
> dev->driver = NULL;
> dev_set_drvdata(dev, NULL);
> if (dev->pm_domain && dev->pm_domain->dismiss)
