| Message ID | 148517353630.18128.12799811817831597628.sendpatchset@little-apple (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Magnus, On 23/01/17 12:12, Magnus Damm wrote: > From: Magnus Damm <damm+renesas@opensource.se> > > Consider failure of iommu_get_domain_for_dev() as non-critical and > get rid of the warning printout. This allows IOMMU properties to be > included in the DTB even though the kernel is configured with > CONFIG_IOMMU_API=n or in case a particular IOMMU driver refuses to > enable IOMMU support for a certain slave device and returns error > from the ->add_device() callback. > > This is only a cosmetic change that removes console warning printouts. The warning is there for a reason - at this point, we *expected* the device to be using an IOMMU for DMA, so a failure is significant. Rather than masking genuine failures in other cases because your case deliberately breaks that expectation, simply change the expectation - i.e. rather than letting of_xlate() succeed then failing add_device() later, reject the of_xlate() call up-front such that the DMA layer never gets told about the IOMMU in the first place. Robin. > Signed-off-by: Magnus Damm <damm+renesas@opensource.se> > --- > > arch/arm64/mm/dma-mapping.c | 10 +++++++++- > 1 file changed, 9 insertions(+), 1 deletion(-) > > --- 0001/arch/arm64/mm/dma-mapping.c > +++ work/arch/arm64/mm/dma-mapping.c 2017-01-23 20:54:40.060607110 +0900 > @@ -827,11 +827,19 @@ static bool do_iommu_attach(struct devic > struct iommu_domain *domain = iommu_get_domain_for_dev(dev); > > /* > + * In case IOMMU support is excluded from the kernel or if the device > + * is not hooked up to any IOMMU group then be silent and keep the > + * old dma_ops. > + */ > + if (!domain) > + return false; > + > + /* > * If the IOMMU driver has the DMA domain support that we require, > * then the IOMMU core will have already configured a group for this > * device, and allocated the default domain for that group. > */ > - if (!domain || iommu_dma_init_domain(domain, dma_base, size, dev)) { > + if (iommu_dma_init_domain(domain, dma_base, size, dev)) { > pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n", > dev_name(dev)); > return false; > _______________________________________________ > iommu mailing list > iommu@lists.linux-foundation.org > https://lists.linuxfoundation.org/mailman/listinfo/iommu >
Hi Robin, >> Consider failure of iommu_get_domain_for_dev() as non-critical and >> get rid of the warning printout. This allows IOMMU properties to be >> included in the DTB even though the kernel is configured with >> CONFIG_IOMMU_API=n or in case a particular IOMMU driver refuses to >> enable IOMMU support for a certain slave device and returns error >> from the ->add_device() callback. >> >> This is only a cosmetic change that removes console warning printouts. > >The warning is there for a reason - at this point, we *expected* the >device to be using an IOMMU for DMA, so a failure is significant. Rather >than masking genuine failures in other cases because your case >deliberately breaks that expectation, simply change the expectation - >i.e. rather than letting of_xlate() succeed then failing add_device() >later, reject the of_xlate() call up-front such that the DMA layer never >gets told about the IOMMU in the first place. > >Robin. > With the iommu probe deferral patches, this behavior would change where the arch_setup_dma_ops would never be called if there is an error from xlate or add_device. But also the error value from xlate/add_device is returned back and the probe of the device would fail for any error. So if there can be cases like above, where the xlate/add_device callbacks can return error for specific reasons, should only EPROBE_DEFER be considered and rest of the errors be filtered out with a WARN probably ? Regards, Sricharan >> Signed-off-by: Magnus Damm <damm+renesas@opensource.se> >> --- >> >> arch/arm64/mm/dma-mapping.c | 10 +++++++++- >> 1 file changed, 9 insertions(+), 1 deletion(-) >> >> --- 0001/arch/arm64/mm/dma-mapping.c >> +++ work/arch/arm64/mm/dma-mapping.c 2017-01-23 20:54:40.060607110 +0900 >> @@ -827,11 +827,19 @@ static bool do_iommu_attach(struct devic >> struct iommu_domain *domain = iommu_get_domain_for_dev(dev); >> >> /* >> + * In case IOMMU support is excluded from the kernel or if the device >> + * is not hooked up to any IOMMU group then be silent and keep the >> + * old dma_ops. >> + */ >> + if (!domain) >> + return false; >> + >> + /* >> * If the IOMMU driver has the DMA domain support that we require, >> * then the IOMMU core will have already configured a group for this >> * device, and allocated the default domain for that group. >> */ >> - if (!domain || iommu_dma_init_domain(domain, dma_base, size, dev)) { >> + if (iommu_dma_init_domain(domain, dma_base, size, dev)) { >> pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n", >> dev_name(dev)); >> return false; >> _______________________________________________ >> iommu mailing list >> iommu@lists.linux-foundation.org >> https://lists.linuxfoundation.org/mailman/listinfo/iommu >> > >_______________________________________________ >iommu mailing list >iommu@lists.linux-foundation.org >https://lists.linuxfoundation.org/mailman/listinfo/iommu
Hi Robin, On Mon, Jan 23, 2017 at 9:34 PM, Robin Murphy <robin.murphy@arm.com> wrote: > Hi Magnus, > > On 23/01/17 12:12, Magnus Damm wrote: >> From: Magnus Damm <damm+renesas@opensource.se> >> >> Consider failure of iommu_get_domain_for_dev() as non-critical and >> get rid of the warning printout. This allows IOMMU properties to be >> included in the DTB even though the kernel is configured with >> CONFIG_IOMMU_API=n or in case a particular IOMMU driver refuses to >> enable IOMMU support for a certain slave device and returns error >> from the ->add_device() callback. >> >> This is only a cosmetic change that removes console warning printouts. > > The warning is there for a reason - at this point, we *expected* the > device to be using an IOMMU for DMA, so a failure is significant. Rather > than masking genuine failures in other cases because your case > deliberately breaks that expectation, simply change the expectation - > i.e. rather than letting of_xlate() succeed then failing add_device() > later, reject the of_xlate() call up-front such that the DMA layer never > gets told about the IOMMU in the first place. Thanks for pointing me in the right direction! I will try to handle this in xlate() instead. Cheers, / magnus
--- 0001/arch/arm64/mm/dma-mapping.c +++ work/arch/arm64/mm/dma-mapping.c 2017-01-23 20:54:40.060607110 +0900 @@ -827,11 +827,19 @@ static bool do_iommu_attach(struct devic struct iommu_domain *domain = iommu_get_domain_for_dev(dev); /* + * In case IOMMU support is excluded from the kernel or if the device + * is not hooked up to any IOMMU group then be silent and keep the + * old dma_ops. + */ + if (!domain) + return false; + + /* * If the IOMMU driver has the DMA domain support that we require, * then the IOMMU core will have already configured a group for this * device, and allocated the default domain for that group. */ - if (!domain || iommu_dma_init_domain(domain, dma_base, size, dev)) { + if (iommu_dma_init_domain(domain, dma_base, size, dev)) { pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n", dev_name(dev)); return false;