Re: [PATCH][venus-for-next-v5.14] media: venus: hfi_cmds: Fix packet size calculation

From: Gustavo A. R. Silva
Date: Wed Jun 02 2021 - 18:02:17 EST




On 6/2/21 15:34, Gustavo A. R. Silva wrote:
>
>
> On 6/2/21 14:55, Kees Cook wrote:
>> On Tue, Jun 01, 2021 at 01:46:16PM -0500, Gustavo A. R. Silva wrote:
>>> Now that a one-element array was replaced with a flexible-array member
>>> in struct hfi_sys_set_property_pkt, use the struct_size() helper to
>>> correctly calculate the packet size.
>>>
>>> Fixes: 701e10b3fd9f ("media: venus: hfi_cmds.h: Replace one-element array with flexible-array member")
>>> Signed-off-by: Gustavo A. R. Silva <gustavoars@xxxxxxxxxx>
>>> ---
>>> BTW... it seems that a similar problem is present in
>>> https://lore.kernel.org/linux-hardening/20210211001044.GA69612@embeddedor/
>>> and that is what is causing the regression. I will send v2 of that
>>> patch, shortly. Thanks.
>>>
>>> drivers/media/platform/qcom/venus/hfi_cmds.c | 8 ++++----
>>> 1 file changed, 4 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/media/platform/qcom/venus/hfi_cmds.c b/drivers/media/platform/qcom/venus/hfi_cmds.c
>>> index 11a8347e5f5c..c86279e5d6e8 100644
>>> --- a/drivers/media/platform/qcom/venus/hfi_cmds.c
>>> +++ b/drivers/media/platform/qcom/venus/hfi_cmds.c
>>> @@ -27,7 +27,7 @@ void pkt_sys_idle_indicator(struct hfi_sys_set_property_pkt *pkt, u32 enable)
>>> {
>>> struct hfi_enable *hfi = (struct hfi_enable *)&pkt->data[1];
>>>
>>> - pkt->hdr.size = sizeof(*pkt) + sizeof(*hfi) + sizeof(u32);
>>> + pkt->hdr.size = struct_size(pkt, data, 2) + sizeof(*hfi);
>>
>> I think this should be "1" not "2".
>>
>> (i.e. there is a single "data" item, followed by an entire *hfi (which
>> starts immediate after data[0]).
>
> Yeah; I see your point. Here I just wanted to preserve the exact same size
> as the original code, which turns out has a "benign" off-by-one issue.

And the reason why I say it's a "benign" off-by-one (which is actually off-by-four-bytes)
issue is because of this piece of code:

drivers/media/platform/qcom/venus/hfi_venus.c:
864 static int venus_sys_set_idle_message(struct venus_hfi_device *hdev,
865 bool enable)
866 {
867 struct hfi_sys_set_property_pkt *pkt;
868 u8 packet[IFACEQ_VAR_SMALL_PKT_SIZE];
869 int ret;
870
871 if (!enable)
872 return 0;
873
874 pkt = (struct hfi_sys_set_property_pkt *)packet;
875
876 pkt_sys_idle_indicator(pkt, enable);
877
878 ret = venus_iface_cmdq_write(hdev, pkt, false);
879 if (ret)
880 return ret;
881
882 return 0;
883 }

IFACEQ_VAR_SMALL_PKT_SIZE at line 868 is of size 100, which is greater than

sizeof(*pkt) + sizeof(*hfi) + sizeof(u32); in the original code:

drivers/media/platform/qcom/venus/hfi_cmds.c:
26 void pkt_sys_idle_indicator(struct hfi_sys_set_property_pkt *pkt, u32 enable)
27 {
28 struct hfi_enable *hfi = (struct hfi_enable *)&pkt->data[1];
29
30 pkt->hdr.size = sizeof(*pkt) + sizeof(*hfi) + sizeof(u32);
31 pkt->hdr.pkt_type = HFI_CMD_SYS_SET_PROPERTY;
32 pkt->num_properties = 1;
33 pkt->data[0] = HFI_PROPERTY_SYS_IDLE_INDICATOR;
34 hfi->enable = enable;
35 }

--
Gustavo

> I'll fix it up and respin.
>
> Thanks!
> --
> Gustavo