From: Jia-Ju Bai <[email protected]>
[ Upstream commit 6499a0db9b0f1e903d52f8244eacc1d4be00eea2 ]
The value av7110->debi_virt is stored in DMA memory, and it is assigned
to data, and thus data[0] can be modified at any time by malicious
hardware. In this case, "if (data[0] < 2)" can be passed, but then
data[0] can be changed into a large number, which may cause buffer
overflow when the code "av7110->ci_slot[data[0]]" is used.
To fix this possible bug, data[0] is assigned to a local variable, which
replaces the use of data[0].
Signed-off-by: Jia-Ju Bai <[email protected]>
Signed-off-by: Sean Young <[email protected]>
Signed-off-by: Mauro Carvalho Chehab <[email protected]>
Signed-off-by: Sasha Levin <[email protected]>
---
drivers/media/pci/ttpci/av7110.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/drivers/media/pci/ttpci/av7110.c b/drivers/media/pci/ttpci/av7110.c
index d6816effb8786..d02b5fd940c12 100644
--- a/drivers/media/pci/ttpci/av7110.c
+++ b/drivers/media/pci/ttpci/av7110.c
@@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
case DATA_CI_GET:
{
u8 *data = av7110->debi_virt;
+ u8 data_0 = data[0];
- if ((data[0] < 2) && data[2] == 0xff) {
+ if (data_0 < 2 && data[2] == 0xff) {
int flags = 0;
if (data[5] > 0)
flags |= CA_CI_MODULE_PRESENT;
if (data[5] > 5)
flags |= CA_CI_MODULE_READY;
- av7110->ci_slot[data[0]].flags = flags;
+ av7110->ci_slot[data_0].flags = flags;
} else
ci_get_data(&av7110->ci_rbuffer,
av7110->debi_virt,
--
2.25.1
Hi!
> The value av7110->debi_virt is stored in DMA memory, and it is assigned
> to data, and thus data[0] can be modified at any time by malicious
> hardware. In this case, "if (data[0] < 2)" can be passed, but then
> data[0] can be changed into a large number, which may cause buffer
> overflow when the code "av7110->ci_slot[data[0]]" is used.
>
> To fix this possible bug, data[0] is assigned to a local variable, which
> replaces the use of data[0].
I'm pretty sure hardware capable of manipulating memory can work
around any such checks, but...
> +++ b/drivers/media/pci/ttpci/av7110.c
> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
> case DATA_CI_GET:
> {
> u8 *data = av7110->debi_virt;
> + u8 data_0 = data[0];
>
> - if ((data[0] < 2) && data[2] == 0xff) {
> + if (data_0 < 2 && data[2] == 0xff) {
> int flags = 0;
> if (data[5] > 0)
> flags |= CA_CI_MODULE_PRESENT;
> if (data[5] > 5)
> flags |= CA_CI_MODULE_READY;
> - av7110->ci_slot[data[0]].flags = flags;
> + av7110->ci_slot[data_0].flags = flags;
This does not even do what it says. Compiler is still free to access
data[0] multiple times. It needs READ_ONCE() to be effective.
Best regards,
Pavel
--
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html
On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
> Hi!
>
> > The value av7110->debi_virt is stored in DMA memory, and it is assigned
> > to data, and thus data[0] can be modified at any time by malicious
> > hardware. In this case, "if (data[0] < 2)" can be passed, but then
> > data[0] can be changed into a large number, which may cause buffer
> > overflow when the code "av7110->ci_slot[data[0]]" is used.
> >
> > To fix this possible bug, data[0] is assigned to a local variable, which
> > replaces the use of data[0].
>
> I'm pretty sure hardware capable of manipulating memory can work
> around any such checks, but...
>
> > +++ b/drivers/media/pci/ttpci/av7110.c
> > @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
> > case DATA_CI_GET:
> > {
> > u8 *data = av7110->debi_virt;
> > + u8 data_0 = data[0];
> >
> > - if ((data[0] < 2) && data[2] == 0xff) {
> > + if (data_0 < 2 && data[2] == 0xff) {
> > int flags = 0;
> > if (data[5] > 0)
> > flags |= CA_CI_MODULE_PRESENT;
> > if (data[5] > 5)
> > flags |= CA_CI_MODULE_READY;
> > - av7110->ci_slot[data[0]].flags = flags;
> > + av7110->ci_slot[data_0].flags = flags;
>
> This does not even do what it says. Compiler is still free to access
> data[0] multiple times. It needs READ_ONCE() to be effective.
Yes, it seems quite dubious to me. If we *really* want to guard against
rogue hardware here, the whole DMA buffer should be copied. I don't
think it's worth it, a rogue PCI device can do much more harm.
--
Regards,
Laurent Pinchart
On Sat, Aug 29, 2020 at 08:16:00PM +0300, Laurent Pinchart wrote:
> On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
> > Hi!
> >
> > > The value av7110->debi_virt is stored in DMA memory, and it is assigned
> > > to data, and thus data[0] can be modified at any time by malicious
> > > hardware. In this case, "if (data[0] < 2)" can be passed, but then
> > > data[0] can be changed into a large number, which may cause buffer
> > > overflow when the code "av7110->ci_slot[data[0]]" is used.
> > >
> > > To fix this possible bug, data[0] is assigned to a local variable, which
> > > replaces the use of data[0].
> >
> > I'm pretty sure hardware capable of manipulating memory can work
> > around any such checks, but...
> >
> > > +++ b/drivers/media/pci/ttpci/av7110.c
> > > @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
> > > case DATA_CI_GET:
> > > {
> > > u8 *data = av7110->debi_virt;
> > > + u8 data_0 = data[0];
> > >
> > > - if ((data[0] < 2) && data[2] == 0xff) {
> > > + if (data_0 < 2 && data[2] == 0xff) {
> > > int flags = 0;
> > > if (data[5] > 0)
> > > flags |= CA_CI_MODULE_PRESENT;
> > > if (data[5] > 5)
> > > flags |= CA_CI_MODULE_READY;
> > > - av7110->ci_slot[data[0]].flags = flags;
> > > + av7110->ci_slot[data_0].flags = flags;
> >
> > This does not even do what it says. Compiler is still free to access
> > data[0] multiple times. It needs READ_ONCE() to be effective.
>
> Yes, it seems quite dubious to me. If we *really* want to guard against
> rogue hardware here, the whole DMA buffer should be copied. I don't
> think it's worth it, a rogue PCI device can do much more harm.
That is a good point. I'm not sure what the kernel could do to protect
against a malicious PCI device (that can do dma) so this patch is totally
pointless.
Thanks
Sean
On 2020/8/29 20:10, Pavel Machek wrote:
> Hi!
>
>> The value av7110->debi_virt is stored in DMA memory, and it is assigned
>> to data, and thus data[0] can be modified at any time by malicious
>> hardware. In this case, "if (data[0] < 2)" can be passed, but then
>> data[0] can be changed into a large number, which may cause buffer
>> overflow when the code "av7110->ci_slot[data[0]]" is used.
>>
>> To fix this possible bug, data[0] is assigned to a local variable, which
>> replaces the use of data[0].
> I'm pretty sure hardware capable of manipulating memory can work
> around any such checks, but...
>
>> +++ b/drivers/media/pci/ttpci/av7110.c
>> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
>> case DATA_CI_GET:
>> {
>> u8 *data = av7110->debi_virt;
>> + u8 data_0 = data[0];
>>
>> - if ((data[0] < 2) && data[2] == 0xff) {
>> + if (data_0 < 2 && data[2] == 0xff) {
>> int flags = 0;
>> if (data[5] > 0)
>> flags |= CA_CI_MODULE_PRESENT;
>> if (data[5] > 5)
>> flags |= CA_CI_MODULE_READY;
>> - av7110->ci_slot[data[0]].flags = flags;
>> + av7110->ci_slot[data_0].flags = flags;
> This does not even do what it says. Compiler is still free to access
> data[0] multiple times. It needs READ_ONCE() to be effective.
>
>
Thanks for this advice, I will submit a v2 patch soon.
Best wishes,
Jia-Ju Bai
On 2020/8/30 1:16, Laurent Pinchart wrote:
> On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
>> Hi!
>>
>>> The value av7110->debi_virt is stored in DMA memory, and it is assigned
>>> to data, and thus data[0] can be modified at any time by malicious
>>> hardware. In this case, "if (data[0] < 2)" can be passed, but then
>>> data[0] can be changed into a large number, which may cause buffer
>>> overflow when the code "av7110->ci_slot[data[0]]" is used.
>>>
>>> To fix this possible bug, data[0] is assigned to a local variable, which
>>> replaces the use of data[0].
>> I'm pretty sure hardware capable of manipulating memory can work
>> around any such checks, but...
>>
>>> +++ b/drivers/media/pci/ttpci/av7110.c
>>> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
>>> case DATA_CI_GET:
>>> {
>>> u8 *data = av7110->debi_virt;
>>> + u8 data_0 = data[0];
>>>
>>> - if ((data[0] < 2) && data[2] == 0xff) {
>>> + if (data_0 < 2 && data[2] == 0xff) {
>>> int flags = 0;
>>> if (data[5] > 0)
>>> flags |= CA_CI_MODULE_PRESENT;
>>> if (data[5] > 5)
>>> flags |= CA_CI_MODULE_READY;
>>> - av7110->ci_slot[data[0]].flags = flags;
>>> + av7110->ci_slot[data_0].flags = flags;
>> This does not even do what it says. Compiler is still free to access
>> data[0] multiple times. It needs READ_ONCE() to be effective.
> Yes, it seems quite dubious to me. If we *really* want to guard against
> rogue hardware here, the whole DMA buffer should be copied. I don't
> think it's worth it, a rogue PCI device can do much more harm.
>
From the original driver code, data[0] is considered to be bad and thus
it should be checked, because the content of the DMA buffer may be
problematic.
Based on this consideration, data[0] can be also modified to bypass the
check, and thus its value should be copied to a local variable for the
check and use.
I agree with Pavel that the compiler optimization may drop the copying
operation, and thus READ_ONCE() should be used here.
I will submit a v2 patch soon.
Best wishes,
Jia-Ju Bai
Hi Jia-Ju,
On Sun, Aug 30, 2020 at 03:33:11PM +0800, Jia-Ju Bai wrote:
> On 2020/8/30 1:16, Laurent Pinchart wrote:
> > On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
> >> Hi!
> >>
> >>> The value av7110->debi_virt is stored in DMA memory, and it is assigned
> >>> to data, and thus data[0] can be modified at any time by malicious
> >>> hardware. In this case, "if (data[0] < 2)" can be passed, but then
> >>> data[0] can be changed into a large number, which may cause buffer
> >>> overflow when the code "av7110->ci_slot[data[0]]" is used.
> >>>
> >>> To fix this possible bug, data[0] is assigned to a local variable, which
> >>> replaces the use of data[0].
> >> I'm pretty sure hardware capable of manipulating memory can work
> >> around any such checks, but...
> >>
> >>> +++ b/drivers/media/pci/ttpci/av7110.c
> >>> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
> >>> case DATA_CI_GET:
> >>> {
> >>> u8 *data = av7110->debi_virt;
> >>> + u8 data_0 = data[0];
> >>>
> >>> - if ((data[0] < 2) && data[2] == 0xff) {
> >>> + if (data_0 < 2 && data[2] == 0xff) {
> >>> int flags = 0;
> >>> if (data[5] > 0)
> >>> flags |= CA_CI_MODULE_PRESENT;
> >>> if (data[5] > 5)
> >>> flags |= CA_CI_MODULE_READY;
> >>> - av7110->ci_slot[data[0]].flags = flags;
> >>> + av7110->ci_slot[data_0].flags = flags;
> >>
> >> This does not even do what it says. Compiler is still free to access
> >> data[0] multiple times. It needs READ_ONCE() to be effective.
> >
> > Yes, it seems quite dubious to me. If we *really* want to guard against
> > rogue hardware here, the whole DMA buffer should be copied. I don't
> > think it's worth it, a rogue PCI device can do much more harm.
>
> From the original driver code, data[0] is considered to be bad and thus
> it should be checked, because the content of the DMA buffer may be
> problematic.
>
> Based on this consideration, data[0] can be also modified to bypass the
> check, and thus its value should be copied to a local variable for the
> check and use.
What makes you think the hardware would do that ?
> I agree with Pavel that the compiler optimization may drop the copying
> operation, and thus READ_ONCE() should be used here.
> I will submit a v2 patch soon.
--
Regards,
Laurent Pinchart
On 2020/8/31 6:25, Laurent Pinchart wrote:
> Hi Jia-Ju,
>
> On Sun, Aug 30, 2020 at 03:33:11PM +0800, Jia-Ju Bai wrote:
>> On 2020/8/30 1:16, Laurent Pinchart wrote:
>>> On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
>>>> Hi!
>>>>
>>>>> The value av7110->debi_virt is stored in DMA memory, and it is assigned
>>>>> to data, and thus data[0] can be modified at any time by malicious
>>>>> hardware. In this case, "if (data[0] < 2)" can be passed, but then
>>>>> data[0] can be changed into a large number, which may cause buffer
>>>>> overflow when the code "av7110->ci_slot[data[0]]" is used.
>>>>>
>>>>> To fix this possible bug, data[0] is assigned to a local variable, which
>>>>> replaces the use of data[0].
>>>> I'm pretty sure hardware capable of manipulating memory can work
>>>> around any such checks, but...
>>>>
>>>>> +++ b/drivers/media/pci/ttpci/av7110.c
>>>>> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
>>>>> case DATA_CI_GET:
>>>>> {
>>>>> u8 *data = av7110->debi_virt;
>>>>> + u8 data_0 = data[0];
>>>>>
>>>>> - if ((data[0] < 2) && data[2] == 0xff) {
>>>>> + if (data_0 < 2 && data[2] == 0xff) {
>>>>> int flags = 0;
>>>>> if (data[5] > 0)
>>>>> flags |= CA_CI_MODULE_PRESENT;
>>>>> if (data[5] > 5)
>>>>> flags |= CA_CI_MODULE_READY;
>>>>> - av7110->ci_slot[data[0]].flags = flags;
>>>>> + av7110->ci_slot[data_0].flags = flags;
>>>> This does not even do what it says. Compiler is still free to access
>>>> data[0] multiple times. It needs READ_ONCE() to be effective.
>>> Yes, it seems quite dubious to me. If we *really* want to guard against
>>> rogue hardware here, the whole DMA buffer should be copied. I don't
>>> think it's worth it, a rogue PCI device can do much more harm.
>> From the original driver code, data[0] is considered to be bad and thus
>> it should be checked, because the content of the DMA buffer may be
>> problematic.
>>
>> Based on this consideration, data[0] can be also modified to bypass the
>> check, and thus its value should be copied to a local variable for the
>> check and use.
> What makes you think the hardware would do that ?
>
Several recent papers show that the bad values from malicious or
problematic hardware can cause security problems:
[NDSS'19] PeriScope: An Effective Probing and Fuzzing Framework for the
Hardware-OS Boundary
[NDSS'19] Thunderclap: Exploring Vulnerabilities in Operating System
IOMMU Protection via DMA from Untrustworthy Peripherals
[USENIX Security'20] USBFuzz: A Framework for Fuzzing USB Drivers by
Device Emulation
In this case, the values from DMA can be bad, and the driver should
carefully check these values to avoid security problems.
IOMMU is an effective method to prevent the hardware from accessing
arbitrary memory address via DMA, but it does not check whether the
values from DMA are safe.
I find that some drivers (including the av7110 driver) check (or try to
check) the values from DMA, and thus I think these drivers have
considered such security problems.
However, some of these checks are not rigorous, so that they can be
bypassed in some cases. The problem that I reported is such an example.
Best wishes,
Jia-Ju Bai
Hi Jia-Ju,
On Mon, Aug 31, 2020 at 09:45:14PM +0800, Jia-Ju Bai wrote:
> On 2020/8/31 6:25, Laurent Pinchart wrote:
> > On Sun, Aug 30, 2020 at 03:33:11PM +0800, Jia-Ju Bai wrote:
> >> On 2020/8/30 1:16, Laurent Pinchart wrote:
> >>> On Sat, Aug 29, 2020 at 02:10:20PM +0200, Pavel Machek wrote:
> >>>> Hi!
> >>>>
> >>>>> The value av7110->debi_virt is stored in DMA memory, and it is assigned
> >>>>> to data, and thus data[0] can be modified at any time by malicious
> >>>>> hardware. In this case, "if (data[0] < 2)" can be passed, but then
> >>>>> data[0] can be changed into a large number, which may cause buffer
> >>>>> overflow when the code "av7110->ci_slot[data[0]]" is used.
> >>>>>
> >>>>> To fix this possible bug, data[0] is assigned to a local variable, which
> >>>>> replaces the use of data[0].
> >>>>
> >>>> I'm pretty sure hardware capable of manipulating memory can work
> >>>> around any such checks, but...
> >>>>
> >>>>> +++ b/drivers/media/pci/ttpci/av7110.c
> >>>>> @@ -424,14 +424,15 @@ static void debiirq(unsigned long cookie)
> >>>>> case DATA_CI_GET:
> >>>>> {
> >>>>> u8 *data = av7110->debi_virt;
> >>>>> + u8 data_0 = data[0];
> >>>>>
> >>>>> - if ((data[0] < 2) && data[2] == 0xff) {
> >>>>> + if (data_0 < 2 && data[2] == 0xff) {
> >>>>> int flags = 0;
> >>>>> if (data[5] > 0)
> >>>>> flags |= CA_CI_MODULE_PRESENT;
> >>>>> if (data[5] > 5)
> >>>>> flags |= CA_CI_MODULE_READY;
> >>>>> - av7110->ci_slot[data[0]].flags = flags;
> >>>>> + av7110->ci_slot[data_0].flags = flags;
> >>>>
> >>>> This does not even do what it says. Compiler is still free to access
> >>>> data[0] multiple times. It needs READ_ONCE() to be effective.
> >>>
> >>> Yes, it seems quite dubious to me. If we *really* want to guard against
> >>> rogue hardware here, the whole DMA buffer should be copied. I don't
> >>> think it's worth it, a rogue PCI device can do much more harm.
> >>
> >> From the original driver code, data[0] is considered to be bad and thus
> >> it should be checked, because the content of the DMA buffer may be
> >> problematic.
> >>
> >> Based on this consideration, data[0] can be also modified to bypass the
> >> check, and thus its value should be copied to a local variable for the
> >> check and use.
> >
> > What makes you think the hardware would do that ?
>
> Several recent papers show that the bad values from malicious or
> problematic hardware can cause security problems:
> [NDSS'19] PeriScope: An Effective Probing and Fuzzing Framework for the
> Hardware-OS Boundary
> [NDSS'19] Thunderclap: Exploring Vulnerabilities in Operating System
> IOMMU Protection via DMA from Untrustworthy Peripherals
> [USENIX Security'20] USBFuzz: A Framework for Fuzzing USB Drivers by
> Device Emulation
>
> In this case, the values from DMA can be bad, and the driver should
> carefully check these values to avoid security problems.
> IOMMU is an effective method to prevent the hardware from accessing
> arbitrary memory address via DMA, but it does not check whether the
> values from DMA are safe.
>
> I find that some drivers (including the av7110 driver) check (or try to
> check) the values from DMA, and thus I think these drivers have
> considered such security problems.
> However, some of these checks are not rigorous, so that they can be
> bypassed in some cases. The problem that I reported is such an example.
The AV7110 is an old chip, I'm not even sure if it can be used with a
modern system that supports IOMMUs for PCI devices. Without that, it's
game over anyway. Before trying to address the issue of a malicious
AV7110 playing with DMA and CPU races, I would ensure that it's worth
it.
--
Regards,
Laurent Pinchart