| Message ID | 1513696821-32291-1-git-send-email-tomasz.nowicki@caviumnetworks.com (mailing list archive) |
|---|---|
| State | New, archived |
| Delegated to: | Bjorn Helgaas |
| Headers | show |
On Tue, 19 Dec 2017 16:20:21 +0100 Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com> wrote: > While iterating over DMA aliases for a PCI device, for some rare cases > (i.e. PCIe-to-PCI/X bridges) we may get exactly the same ID as initial child > device. In turn, the same ID may get registered for a device multiple times. > Eventually IOMMU driver may try to configure the same ID within domain > multiple times too which for some IOMMU drivers is illegal and causes kernel > panic. > > Rule out ID duplication prior to device ID array registration. > > CC: stable@vger.kernel.org # v4.14+ You've identified a release, is there a specific commit this fixes? > Signed-off-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com> > --- > drivers/iommu/iommu.c | 28 ++++++++++++++++++++++++++++ > 1 file changed, 28 insertions(+) > > diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c > index 3de5c0b..9b2c138 100644 > --- a/drivers/iommu/iommu.c > +++ b/drivers/iommu/iommu.c > @@ -1945,6 +1945,31 @@ void iommu_fwspec_free(struct device *dev) > } > EXPORT_SYMBOL_GPL(iommu_fwspec_free); > > +static void iommu_fwspec_remove_ids_dup(struct device *dev, u32 *ids, > + int *num_ids) > +{ > + struct iommu_fwspec *fwspec = dev->iommu_fwspec; > + int i, j, k, valid_ids = *num_ids; > + > + for (i = 0; i < valid_ids; i++) { > + for (j = 0; j < fwspec->num_ids; j++) { > + if (ids[i] != fwspec->ids[j]) > + continue; > + > + dev_info(dev, "found 0x%x ID duplication, skipped\n", > + ids[i]); > + > + for (k = i + 1; k < valid_ids; k++) > + ids[k - 1] = ids[k]; Use memmove()? > + > + valid_ids--; > + break; At this point ids[i] is not the ids[i] that we tested for dupes, it's what was ids[i + 1], but we're going to i++ on the next iteration and we therefore never test that entry. > + } > + } > + > + *num_ids = valid_ids; > +} > + > int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids) > { > struct iommu_fwspec *fwspec = dev->iommu_fwspec; > @@ -1954,6 +1979,9 @@ int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids) > if (!fwspec) > return -EINVAL; > > + /* Rule out IDs already registered */ > + iommu_fwspec_remove_ids_dup(dev, ids, &num_ids); > + > size = offsetof(struct iommu_fwspec, ids[fwspec->num_ids + num_ids]); > if (size > sizeof(*fwspec)) { > fwspec = krealloc(dev->iommu_fwspec, size, GFP_KERNEL);
On 19.12.2017 17:37, Alex Williamson wrote: > On Tue, 19 Dec 2017 16:20:21 +0100 > Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com> wrote: > >> While iterating over DMA aliases for a PCI device, for some rare cases >> (i.e. PCIe-to-PCI/X bridges) we may get exactly the same ID as initial child >> device. In turn, the same ID may get registered for a device multiple times. >> Eventually IOMMU driver may try to configure the same ID within domain >> multiple times too which for some IOMMU drivers is illegal and causes kernel >> panic. >> >> Rule out ID duplication prior to device ID array registration. >> >> CC: stable@vger.kernel.org # v4.14+ > > You've identified a release, is there a specific commit this fixes? Yes, it was triggered by converting drm_pci_init() to pci_register_driver() in ast_drv.c Fixes: 10631d724def ("drm/pci: Deprecate drm_pci_init/exit completely ") > >> Signed-off-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com> >> --- >> drivers/iommu/iommu.c | 28 ++++++++++++++++++++++++++++ >> 1 file changed, 28 insertions(+) >> >> diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c >> index 3de5c0b..9b2c138 100644 >> --- a/drivers/iommu/iommu.c >> +++ b/drivers/iommu/iommu.c >> @@ -1945,6 +1945,31 @@ void iommu_fwspec_free(struct device *dev) >> } >> EXPORT_SYMBOL_GPL(iommu_fwspec_free); >> >> +static void iommu_fwspec_remove_ids_dup(struct device *dev, u32 *ids, >> + int *num_ids) >> +{ >> + struct iommu_fwspec *fwspec = dev->iommu_fwspec; >> + int i, j, k, valid_ids = *num_ids; >> + >> + for (i = 0; i < valid_ids; i++) { >> + for (j = 0; j < fwspec->num_ids; j++) { >> + if (ids[i] != fwspec->ids[j]) >> + continue; >> + >> + dev_info(dev, "found 0x%x ID duplication, skipped\n", >> + ids[i]); >> + >> + for (k = i + 1; k < valid_ids; k++) >> + ids[k - 1] = ids[k]; > > Use memmove()? Right. > >> + >> + valid_ids--; >> + break; > > At this point ids[i] is not the ids[i] that we tested for dupes, it's > what was ids[i + 1], but we're going to i++ on the next iteration and > we therefore never test that entry. Good point. Now the fundamental question is where we should put the patch, here or in SMMUv3 driver as per Robin suggestion. Thanks, Tomasz
diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c index 3de5c0b..9b2c138 100644 --- a/drivers/iommu/iommu.c +++ b/drivers/iommu/iommu.c @@ -1945,6 +1945,31 @@ void iommu_fwspec_free(struct device *dev) } EXPORT_SYMBOL_GPL(iommu_fwspec_free); +static void iommu_fwspec_remove_ids_dup(struct device *dev, u32 *ids, + int *num_ids) +{ + struct iommu_fwspec *fwspec = dev->iommu_fwspec; + int i, j, k, valid_ids = *num_ids; + + for (i = 0; i < valid_ids; i++) { + for (j = 0; j < fwspec->num_ids; j++) { + if (ids[i] != fwspec->ids[j]) + continue; + + dev_info(dev, "found 0x%x ID duplication, skipped\n", + ids[i]); + + for (k = i + 1; k < valid_ids; k++) + ids[k - 1] = ids[k]; + + valid_ids--; + break; + } + } + + *num_ids = valid_ids; +} + int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids) { struct iommu_fwspec *fwspec = dev->iommu_fwspec; @@ -1954,6 +1979,9 @@ int iommu_fwspec_add_ids(struct device *dev, u32 *ids, int num_ids) if (!fwspec) return -EINVAL; + /* Rule out IDs already registered */ + iommu_fwspec_remove_ids_dup(dev, ids, &num_ids); + size = offsetof(struct iommu_fwspec, ids[fwspec->num_ids + num_ids]); if (size > sizeof(*fwspec)) { fwspec = krealloc(dev->iommu_fwspec, size, GFP_KERNEL);
While iterating over DMA aliases for a PCI device, for some rare cases (i.e. PCIe-to-PCI/X bridges) we may get exactly the same ID as initial child device. In turn, the same ID may get registered for a device multiple times. Eventually IOMMU driver may try to configure the same ID within domain multiple times too which for some IOMMU drivers is illegal and causes kernel panic. Rule out ID duplication prior to device ID array registration. CC: stable@vger.kernel.org # v4.14+ Signed-off-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com> --- drivers/iommu/iommu.c | 28 ++++++++++++++++++++++++++++ 1 file changed, 28 insertions(+)