We are expending the default DMA segmentation boundary to its
possible maximum value (ULONG_MAX) to indicate that a device
doesn't specify a boundary limit. So all dma_get_seg_boundary
callers should take a precaution with the return values since
it would easily get overflowed.
I scanned the entire kernel tree for all the existing callers
and found that most of callers may get overflowed in two ways:
either "+ 1" or passing it to ALIGN() that does "+ mask".
According to kernel defines:
#define ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
#define ALIGN(x, a) ALIGN_MASK(x, (typeof(x))(a) - 1)
We can simplify the logic here:
ALIGN(boundary + 1, 1 << shift) >> shift
= ALIGN_MASK(b + 1, (1 << s) - 1) >> s
= {[b + 1 + (1 << s) - 1] & ~[(1 << s) - 1]} >> s
= [b + 1 + (1 << s) - 1] >> s
= [b + (1 << s)] >> s
= (b >> s) + 1
So this series of patches fix the potential overflow with this
overflow-free shortcut.
As I don't think that I have these platforms, marking RFT.
Thanks
Nic
Nicolin Chen (7):
powerpc/iommu: Avoid overflow at boundary_size
alpha: Avoid overflow at boundary_size
ia64/sba_iommu: Avoid overflow at boundary_size
s390/pci_dma: Avoid overflow at boundary_size
sparc: Avoid overflow at boundary_size
x86/amd_gart: Avoid overflow at boundary_size
parisc: Avoid overflow at boundary_size
arch/alpha/kernel/pci_iommu.c | 10 ++++------
arch/ia64/hp/common/sba_iommu.c | 4 ++--
arch/powerpc/kernel/iommu.c | 11 +++++------
arch/s390/pci/pci_dma.c | 4 ++--
arch/sparc/kernel/iommu-common.c | 9 +++------
arch/sparc/kernel/iommu.c | 4 ++--
arch/sparc/kernel/pci_sun4v.c | 4 ++--
arch/x86/kernel/amd_gart_64.c | 4 ++--
drivers/parisc/ccio-dma.c | 4 ++--
drivers/parisc/sba_iommu.c | 4 ++--
10 files changed, 26 insertions(+), 32 deletions(-)
--
2.17.1
On 8/21/20 1:19 AM, Nicolin Chen wrote:
We are expending the default DMA segmentation boundary to its
possible maximum value (ULONG_MAX) to indicate that a device
doesn't specify a boundary limit. So all dma_get_seg_boundary
callers should take a precaution with the return values since
it would easily get overflowed.
I scanned the entire kernel tree for all the existing callers
and found that most of callers may get overflowed in two ways:
either "+ 1" or passing it to ALIGN() that does "+ mask".
According to kernel defines:
#define ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
#define ALIGN(x, a) ALIGN_MASK(x, (typeof(x))(a) - 1)
We can simplify the logic here:
ALIGN(boundary + 1, 1 << shift) >> shift
= ALIGN_MASK(b + 1, (1 << s) - 1) >> s
= {[b + 1 + (1 << s) - 1] & ~[(1 << s) - 1]} >> s
= [b + 1 + (1 << s) - 1] >> s
= [b + (1 << s)] >> s
= (b >> s) + 1
So this series of patches fix the potential overflow with this
overflow-free shortcut.
Hi Nicolin,
haven't seen any other feedback from other maintainers,
so I guess you will resend this?
On first glance it seems to make sense.
I'm a little confused why it is only a "potential overflow"
while this part
"We are expending the default DMA segmentation boundary to its
possible maximum value (ULONG_MAX) to indicate that a device
doesn't specify a boundary limit"
sounds to me like ULONG_MAX is actually used, does that
mean there are currently no devices which do not specify a
boundary limit?
As I don't think that I have these platforms, marking RFT.
Thanks
Nic
Nicolin Chen (7):
powerpc/iommu: Avoid overflow at boundary_size
alpha: Avoid overflow at boundary_size
ia64/sba_iommu: Avoid overflow at boundary_size
s390/pci_dma: Avoid overflow at boundary_size
sparc: Avoid overflow at boundary_size
x86/amd_gart: Avoid overflow at boundary_size
parisc: Avoid overflow at boundary_size
arch/alpha/kernel/pci_iommu.c | 10 ++++------
arch/ia64/hp/common/sba_iommu.c | 4 ++--
arch/powerpc/kernel/iommu.c | 11 +++++------
arch/s390/pci/pci_dma.c | 4 ++--
arch/sparc/kernel/iommu-common.c | 9 +++------
arch/sparc/kernel/iommu.c | 4 ++--
arch/sparc/kernel/pci_sun4v.c | 4 ++--
arch/x86/kernel/amd_gart_64.c | 4 ++--
drivers/parisc/ccio-dma.c | 4 ++--
drivers/parisc/sba_iommu.c | 4 ++--
10 files changed, 26 insertions(+), 32 deletions(-)
Hi Niklas,
On Tue, Aug 25, 2020 at 12:16:27PM +0200, Niklas Schnelle wrote:
On 8/21/20 1:19 AM, Nicolin Chen wrote:
quoted
We are expending the default DMA segmentation boundary to its
possible maximum value (ULONG_MAX) to indicate that a device
doesn't specify a boundary limit. So all dma_get_seg_boundary
callers should take a precaution with the return values since
it would easily get overflowed.
I scanned the entire kernel tree for all the existing callers
and found that most of callers may get overflowed in two ways:
either "+ 1" or passing it to ALIGN() that does "+ mask".
According to kernel defines:
#define ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))
#define ALIGN(x, a) ALIGN_MASK(x, (typeof(x))(a) - 1)
We can simplify the logic here:
ALIGN(boundary + 1, 1 << shift) >> shift
= ALIGN_MASK(b + 1, (1 << s) - 1) >> s
= {[b + 1 + (1 << s) - 1] & ~[(1 << s) - 1]} >> s
= [b + 1 + (1 << s) - 1] >> s
= [b + (1 << s)] >> s
= (b >> s) + 1
So this series of patches fix the potential overflow with this
overflow-free shortcut.
haven't seen any other feedback from other maintainers,
I am wondering this too...whether I sent correctly or not.
so I guess you will resend this?
Do I need to? Though I won't mind doing so if it's necessary..
On first glance it seems to make sense.
I'm a little confused why it is only a "potential overflow"
while this part
"We are expending the default DMA segmentation boundary to its
possible maximum value (ULONG_MAX) to indicate that a device
doesn't specify a boundary limit"
sounds to me like ULONG_MAX is actually used, does that
mean there are currently no devices which do not specify a
boundary limit?
Sorry for the confusion. We actually applied ULONG_MAX change
last week but reverted it right after, due to a bug report at
one of these "potential" overflows. So at this moment the top
of the tree doesn't set default boundary to ULONG_MAX yet.
Thanks
Nic