[PATCH] misc: nsm: bound the device-reported response length

[PATCH] misc: nsm: bound the device-reported response length

[Bryam Vargas]

nsm_sendrecv_msg_locked() stores the virtqueue used-ring length reported
by the NSM device into msg->resp.len without bounding it to the response
buffer. A malicious or buggy backend can report a length larger than the
response buffer; parse_resp_raw() then copies that many bytes out of the
fixed buffer to user space, disclosing adjacent kernel heap (an
out-of-bounds read). The request path already floors its length in
fill_req_raw(); the response path lacks the symmetric check.

Clamp the stored length to the size of the response buffer. Well-behaved
devices report no more than the posted buffer size, so conforming traffic
is unaffected.

Fixes: b9873755a6c8 ("misc: Add Nitro Secure Module driver")
Cc: stable@vger.kernel.org
Signed-off-by: Bryam Vargas <hexlabsecurity@proton.me>
---
 drivers/misc/nsm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/misc/nsm.c b/drivers/misc/nsm.c
index ef7b32742340..f759fbba049a 100644
--- a/drivers/misc/nsm.c
+++ b/drivers/misc/nsm.c
@@ -243,7 +243,7 @@ static int nsm_sendrecv_msg_locked(struct nsm *nsm)
 		goto cleanup;
 	}
 
-	msg->resp.len = len;
+	msg->resp.len = min_t(unsigned int, len, sizeof(msg->resp.data));
 
 	rc = 0;
 

---
base-commit: 1a3746ccbb0a97bed3c06ccde6b880013b1dddc1
change-id: 20260620-b4-disp-a54b7dd6-4cbda73d614b

Best regards,
-- 
Bryam Vargas <hexlabsecurity@proton.me>

[PATCH] misc: nsm: bound the device-reported response length

[Bryam Vargas via B4 Relay]

From: Bryam Vargas <hexlabsecurity@proton.me>

nsm_sendrecv_msg_locked() stores the virtqueue used-ring length reported
by the NSM device into msg->resp.len without bounding it to the response
buffer. A malicious or buggy backend can report a length larger than the
response buffer; parse_resp_raw() then copies that many bytes out of the
fixed buffer to user space, disclosing adjacent kernel heap (an
out-of-bounds read). The request path already floors its length in
fill_req_raw(); the response path lacks the symmetric check.

Clamp the stored length to the size of the response buffer. Well-behaved
devices report no more than the posted buffer size, so conforming traffic
is unaffected.

Fixes: b9873755a6c8 ("misc: Add Nitro Secure Module driver")
Cc: stable@vger.kernel.org
Signed-off-by: Bryam Vargas <hexlabsecurity@proton.me>
---
 drivers/misc/nsm.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/misc/nsm.c b/drivers/misc/nsm.c
index ef7b32742340..f759fbba049a 100644
--- a/drivers/misc/nsm.c
+++ b/drivers/misc/nsm.c
@@ -243,7 +243,7 @@ static int nsm_sendrecv_msg_locked(struct nsm *nsm)
 		goto cleanup;
 	}
 
-	msg->resp.len = len;
+	msg->resp.len = min_t(unsigned int, len, sizeof(msg->resp.data));
 
 	rc = 0;
 

---
base-commit: 1a3746ccbb0a97bed3c06ccde6b880013b1dddc1
change-id: 20260620-b4-disp-a54b7dd6-4cbda73d614b

Best regards,
-- 
Bryam Vargas <hexlabsecurity@proton.me>

Re: [PATCH] misc: nsm: bound the device-reported response length

[Greg Kroah-Hartman]

On Sat, Jun 20, 2026 at 09:42:11PM -0500, Bryam Vargas via B4 Relay wrote:
> From: Bryam Vargas <hexlabsecurity@proton.me>
> 
> nsm_sendrecv_msg_locked() stores the virtqueue used-ring length reported
> by the NSM device into msg->resp.len without bounding it to the response
> buffer. A malicious or buggy backend can report a length larger than the
> response buffer; parse_resp_raw() then copies that many bytes out of the
> fixed buffer to user space, disclosing adjacent kernel heap (an
> out-of-bounds read). The request path already floors its length in
> fill_req_raw(); the response path lacks the symmetric check.
> 
> Clamp the stored length to the size of the response buffer. Well-behaved
> devices report no more than the posted buffer size, so conforming traffic
> is unaffected.

Is this really the only place where a "buggy" device that is bound to
the driver can cause any problems?  Shouldn't the driver only be bound
to trusted devices to start with?

thanks,

greg k-h

Re: [PATCH] misc: nsm: bound the device-reported response length

[Graf (AWS), Alexander]

On 21.06.26 07:38, Greg Kroah-Hartman wrote:
> On Sat, Jun 20, 2026 at 09:42:11PM -0500, Bryam Vargas via B4 Relay wrote:
>> From: Bryam Vargas <hexlabsecurity@proton.me>
>>
>> nsm_sendrecv_msg_locked() stores the virtqueue used-ring length reported
>> by the NSM device into msg->resp.len without bounding it to the response
>> buffer. A malicious or buggy backend can report a length larger than the
>> response buffer; parse_resp_raw() then copies that many bytes out of the
>> fixed buffer to user space, disclosing adjacent kernel heap (an
>> out-of-bounds read). The request path already floors its length in
>> fill_req_raw(); the response path lacks the symmetric check.
>>
>> Clamp the stored length to the size of the response buffer. Well-behaved
>> devices report no more than the posted buffer size, so conforming traffic
>> is unaffected.
> Is this really the only place where a "buggy" device that is bound to
> the driver can cause any problems?  Shouldn't the driver only be bound
> to trusted devices to start with?


It's close to impossible to understand whether the backing device is 
"trusted" or not from a VM's point of view. Even all the fancy "trusted 
PCIe" logic is about as secure as "secure HDMI": not at all. The best 
thing a VM can do is trust specific drivers IMHO to reduce its TCB. I 
think as a stick in the ground saying "all virtio drivers are trusted" 
makes a lot of sense given that a lot of the virtualized world runs on 
virtio.

Bryam, under that premise, could you please also look at other parts of 
the nsm driver that could potentially break the premise of a malicious 
host? A quick AI scan revealed that the cbor_short_size switch has no 
default branch, which leads to uninitialized array_len. Since the report 
looks heavily AI influenced, maybe yours actually finds more?

For this patch:

Reviewed-by: Alexander Graf <graf@amazon.com>


Thanks,

Alex



Amazon Web Services Development Center Germany GmbH
Tamara-Danz-Str. 13
10243 Berlin
Geschaeftsfuehrung: Christof Hellmis, Andreas Stieger
Eingetragen am Amtsgericht Charlottenburg unter HRB 257764 B
Sitz: Berlin
Ust-ID: DE 365 538 597

Re: [PATCH] misc: nsm: bound the device-reported response length

[Bryam Vargas]

On 2026-06-21, Alexander Graf wrote:
> I think as a stick in the ground saying "all virtio drivers are trusted"
> makes a lot of sense given that a lot of the virtualized world runs on
> virtio.

Agreed -- a guest can't attest its backend, so the practical TCB line is the
driver, not the device. I'll treat the NSM transport as untrusted and sweep the
rest of nsm.c on that basis.

> A quick AI scan revealed that the cbor_short_size switch has no default
> branch, which leads to uninitialized array_len.

Confirmed, and it's an uninitialized read rather than an OOB. The switch covers
23..27 with no default, so a device picking 0..22 or 28..31 matches no case and
array_len reaches the two trailing checks unset. Those clamp the result to
<= resp_len - array_offset and <= INT_MAX, so the memcpy in
parse_resp_get_random() stays inside resp->data. What's left is the
uninitialized length plus a plain decode bug: a short array of 0..22 elements is
read from garbage instead of cbor_short_size.

The existing case already handles 23 correctly, so the fix widens that label to
the whole short-form range (same body, no new case value -> no duplicate) and
adds a rejecting default:

-       case CBOR_SHORT_SIZE_MAX_VALUE:        /* short encoding */
+       case 0 ... CBOR_SHORT_SIZE_MAX_VALUE:  /* short form: len IS the value */
                array_len = cbor_short_size;
                break;
        ... U8/U16/U32/U64 cases unchanged ...
+       default:                               /* 28..31 reserved/indefinite */
+               return -EFAULT;

By inspection so far. I'll send that as its own patch this week once it's
through KMSAN here, and the wider nsm.c pass before -rc3.

Thanks for the review,
Bryam