Skip to contentSkip to conversation

media: amphion: implement windsor encoder rpc interface

bug report Dan Carpenter 2 participants 4 messages Mar 10, 2022 Media (V4L/DVB)

lore
Share & tools
lore
#0 lore raw

Hello Ming Qian,

The patch d82977796c48: "media: amphion: implement windsor encoder rpc interface" from Feb 24, 2022, leads to the following Smatch static checker warning:

	drivers/media/platform/amphion/vpu_windsor.c:823 vpu_windsor_config_memory_resource()
	error: buffer overflow 'pool->enc_frames' 6 <= 7

drivers/media/platform/amphion/vpu_windsor.c
807 int vpu_windsor_config_memory_resource(struct vpu_shared_addr *shared,
808 u32 instance,
809 u32 type,
810 u32 index,
811 struct vpu_buffer *buf)
812 {
813 struct vpu_enc_mem_pool *pool;
814 struct vpu_enc_memory_resource *res;
815
816 if (instance >= VID_API_NUM_STREAMS)
^^^^^^^^^^^^^^^^^^^
This is 8.

    817                 return -EINVAL;
    818 
    819         pool = get_mem_pool(shared, instance);
    820 
    821         switch (type) {
    822         case MEM_RES_ENC:

--> 823 res = &pool->enc_frames[index];

This only has WINDSOR_MAX_SRC_FRAMES elements.

    824                 break;
    825         case MEM_RES_REF:
    826                 res = &pool->ref_frames[index];
    827                 break;
    828         case MEM_RES_ACT:
    829                 res = &pool->act_frame;
    830                 break;
    831         default:
    832                 return -EINVAL;
    833         }
    834 
    835         res->phys = buf->phys;
    836         res->virt = buf->phys - shared->boot_addr;
    837         res->size = buf->length;
    838 
    839         return 0;
    840 }

regards,
dan carpenter

#1 lore raw
Quoted 40 lines

-----Original Message-----
From: Dan Carpenter [mailto:[email protected]]
Sent: Thursday, March 10, 2022 2:57 PM
To: Ming Qian <[email protected]>
Cc: [email protected]
Subject: [EXT] [bug report] media: amphion: implement windsor encoder rpc interface

Caution: EXT Email

Hello Ming Qian,

The patch d82977796c48: "media: amphion: implement windsor encoder rpc interface" from Feb 24, 2022, leads to the following Smatch static checker warning:

drivers/media/platform/amphion/vpu_windsor.c:823
vpu_windsor_config_memory_resource()

        error: buffer overflow 'pool->enc_frames' 6 <= 7

drivers/media/platform/amphion/vpu_windsor.c
807 int vpu_windsor_config_memory_resource(struct vpu_shared_addr
*shared,

    808                                        u32 instance,
    809                                        u32 type,
    810                                        u32 index,
    811                                        struct vpu_buffer *buf)
    812 {
    813         struct vpu_enc_mem_pool *pool;
    814         struct vpu_enc_memory_resource *res;
    815
    816         if (instance >= VID_API_NUM_STREAMS)
                                ^^^^^^^^^^^^^^^^^^^ This is 8.

    817                 return -EINVAL;
    818
    819         pool = get_mem_pool(shared, instance);
    820
    821         switch (type) {
    822         case MEM_RES_ENC:
Quoted 3 lines

--> 823 res = &pool->enc_frames[index];

This only has WINDSOR_MAX_SRC_FRAMES elements.

Hi Dan,
I don't get the point, the instance and index is different, and one vpu core can support 8 instances (VID_API_NUM_STREAMS),
The enc_frame count of one instance won't exceed 6 (WINDSOR_MAX_SRC_FRAMES).

    Maybe I should add a check for the index like:

	If (index >= ARRAY_SIZE(pool->enc_frames))
		return -EINVAL;
Quoted 21 lines
    824                 break;
    825         case MEM_RES_REF:
    826                 res = &pool->ref_frames[index];
    827                 break;
    828         case MEM_RES_ACT:
    829                 res = &pool->act_frame;
    830                 break;
    831         default:
    832                 return -EINVAL;
    833         }
    834
    835         res->phys = buf->phys;
    836         res->virt = buf->phys - shared->boot_addr;
    837         res->size = buf->length;
    838
    839         return 0;
    840 }

regards,
dan carpenter

#2 lore raw
Quoted 22 lines · Ming Qian
Quoted 22 lines

drivers/media/platform/amphion/vpu_windsor.c
807 int vpu_windsor_config_memory_resource(struct vpu_shared_addr
*shared,

    808                                        u32 instance,
    809                                        u32 type,
    810                                        u32 index,
    811                                        struct vpu_buffer *buf)
    812 {
    813         struct vpu_enc_mem_pool *pool;
    814         struct vpu_enc_memory_resource *res;
    815
    816         if (instance >= VID_API_NUM_STREAMS)
                                ^^^^^^^^^^^^^^^^^^^ This is 8.

    817                 return -EINVAL;
    818
    819         pool = get_mem_pool(shared, instance);
    820
    821         switch (type) {
    822         case MEM_RES_ENC:

--> 823 res = &pool->enc_frames[index];

Quoted 6 lines
Quoted 1 line

This only has WINDSOR_MAX_SRC_FRAMES elements.

Hi Dan,
I don't get the point, the instance and index is different, and
one vpu core can support 8 instances (VID_API_NUM_STREAMS),
The enc_frame count of one instance won't exceed 6 (WINDSOR_MAX_SRC_FRAMES).

Hi Ming,

I appologize. I mis-understood what Smatch was saying and mis-read the code as well. I got "instance" and "index" mixed up as you say.

However, when I re-reviewed the code now it looks like Smatch is correct and there is a potential buffer overflow. The "index" variable comes from vpu_iface_unpack_msg_data() so I do not think it can be trusted. We need to have an upper bounds.

There are 3 upper bounds checks in venc_request_mem_resource() but they only check that "index" is in the 0-7 range.

	if (enc_frame_num > ARRAY_SIZE(venc->enc)) {
	if (ref_frame_num > ARRAY_SIZE(venc->ref)) {
	if (act_frame_num > ARRAY_SIZE(venc->act)) {

These ->enc, ->ref and ->act arrays all have 8 elements.

However, as noted by Smatch the pool->enc_frames[] array only has 6 elements and the pool->ref_frames[] array only has 3 elements. So we need to add bounds checking before both accesses.

Quoted 5 lines
    Maybe I should add a check for the index like:

	If (index >= ARRAY_SIZE(pool->enc_frames))
		return -EINVAL;

Yes, please. Or maybe we need to make the arrays larger to match the arrays in venc_request_mem_resource()?

Quoted 4 lines
Quoted 4 lines
    824                 break;
    825         case MEM_RES_REF:
    826                 res = &pool->ref_frames[index];

And here as well.

regards,
dan carpenter

#3 lore raw
Quoted 54 lines · Ming Qian
Quoted 22 lines
Quoted 22 lines

drivers/media/platform/amphion/vpu_windsor.c
807 int vpu_windsor_config_memory_resource(struct
vpu_shared_addr *shared,

    808                                        u32 instance,
    809                                        u32 type,
    810                                        u32 index,
    811                                        struct vpu_buffer *buf)
    812 {
    813         struct vpu_enc_mem_pool *pool;
    814         struct vpu_enc_memory_resource *res;
    815
    816         if (instance >= VID_API_NUM_STREAMS)
                                ^^^^^^^^^^^^^^^^^^^ This is 8.

    817                 return -EINVAL;
    818
    819         pool = get_mem_pool(shared, instance);
    820
    821         switch (type) {
    822         case MEM_RES_ENC:

--> 823 res = &pool->enc_frames[index];

Quoted 6 lines
Quoted 1 line

This only has WINDSOR_MAX_SRC_FRAMES elements.

Hi Dan,
I don't get the point, the instance and index is different, and
one vpu core can support 8 instances (VID_API_NUM_STREAMS), The enc_frame count of one instance won't exceed 6

(WINDSOR_MAX_SRC_FRAMES).

Hi Ming,

I appologize. I mis-understood what Smatch was saying and mis-read the code as well. I got "instance" and "index" mixed up as you say.

However, when I re-reviewed the code now it looks like Smatch is correct and there is a potential buffer overflow. The "index" variable comes from vpu_iface_unpack_msg_data() so I do not think it can be trusted. We need to have an upper bounds.

There are 3 upper bounds checks in venc_request_mem_resource() but they only check that "index" is in the 0-7 range.

        if (enc_frame_num > ARRAY_SIZE(venc->enc)) {
        if (ref_frame_num > ARRAY_SIZE(venc->ref)) {
        if (act_frame_num > ARRAY_SIZE(venc->act)) {

These ->enc, ->ref and ->act arrays all have 8 elements.

However, as noted by Smatch the pool->enc_frames[] array only has 6 elements and the pool->ref_frames[] array only has 3 elements. So we need to add bounds checking before both accesses.

Quoted 9 lines
Quoted 5 lines
    Maybe I should add a check for the index like:

      If (index >= ARRAY_SIZE(pool->enc_frames))
              return -EINVAL;

Yes, please. Or maybe we need to make the arrays larger to match the arrays in venc_request_mem_resource()?

Hi Dan,
Yes, you're right, I have also noticed that.
I will make a patch that add the checking code.
Here the array size is defined in the firmware, so I think I should not change them.

Thanks for your review.

Ming

Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
    824                 break;
    825         case MEM_RES_REF:
    826                 res = &pool->ref_frames[index];

And here as well.

Yes, I'll check it too.

Quoted 4 lines

regards,
dan carpenter