On Mon, Jan 14, 2019 at 04:38:20PM +0000, Robin Murphy wrote: > On 14/01/2019 16:09, Thierry Reding wrote: > > On Mon, Jan 14, 2019 at 02:22:40PM +0100, Marek Szyprowski wrote: > > > This reverts commit 1874619a7df4b14b23b14877e705ae15325773e3. > > > > > > That patch broke IOMMU support for devices, which fails to probe for the > > > first time and use deferred probe approach. When non-NULL dma_ops is set > > > in arm_iommu_detach_device(), the given device later ignored by > > > arch_setup_dma_ops() and stays with non-IOMMU dma_ops. > > > > > > Reported-by: Tobias Jakobi > > > Fixes: 1874619a7df4 "ARM: dma-mapping: Set proper DMA ops in arm_iommu_detach_device()" > > > Signed-off-by: Marek Szyprowski > > > --- > > > arch/arm/mm/dma-mapping.c | 12 ++++++------ > > > 1 file changed, 6 insertions(+), 6 deletions(-) > > > > Can you point out exactly what drivers break because of this change? We > > need to find a solution that works for everyone. Reverting is only > > marginally useful because somebody will just end up wanting to revert > > the revert because a different driver is now broken. > > At first glance, it sounds like what we really want is for > arch_teardown_iommu_ops() to completely clear any ops that > arch_setup_dma_ops() installed - does the below suffice? > > Robin. > > ----->8----- > diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c > index f1e2922e447c..1e3e08a1c456 100644 > --- a/arch/arm/mm/dma-mapping.c > +++ b/arch/arm/mm/dma-mapping.c > @@ -2390,4 +2390,6 @@ void arch_teardown_dma_ops(struct device *dev) > return; > > arm_teardown_iommu_dma_ops(dev); > + /* Let arch_setup_dma_ops() start again from scratch upon re-probe */ > + set_dma_ops(dev, NULL); > } Seems to work fine on Tegra: Tested-by: Thierry Reding