Skip to contentSkip to conversation

Drivers: hv: vmbus: Fix variable assignments in hv_ringbuffer_read()

PATCH In review Andres Beltran 3 participants 5 messages Jul 24, 2020 – Jul 26, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Drivers: hv: vmbus: Fix variable assignments in hv_ringbuffer_read()

Assignments to buffer_actual_len and requestid happen before packetlen is checked to be within buflen. If this condition is true, hv_ringbuffer_read() returns with these variables already set to some value even though no data is actually read. This might create inconsistencies in any routine calling hv_ringbuffer_read(). Assign values to such pointers after the packetlen check.

---
 drivers/hv/ring_buffer.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/hv/ring_buffer.c +3 -2
index 356e22159e83..e277ce7372a4 100644
@@ -350,12 +350,13 @@ int hv_ringbuffer_read(struct vmbus_channel *channel,
 
 	offset = raw ? 0 : (desc->offset8 << 3);
 	packetlen = (desc->len8 << 3) - offset;
-	*buffer_actual_len = packetlen;
-	*requestid = desc->trans_id;
 
 	if (unlikely(packetlen > buflen))
 		return -ENOBUFS;
 
+	*buffer_actual_len = packetlen;
+	*requestid = desc->trans_id;
+
 	/* since ring is double mapped, only one copy is necessary */
 	memcpy(buffer, (const char *)desc + offset, packetlen);
 
Signature · 1 line
2.25.1
#1 lore raw
Quoted 23 lines · On Fri, 24 Jul 2020 09:46:06 -0700 "Andres Beltran" <[email protected]> wrote:

Assignments to buffer_actual_len and requestid happen before packetlen is checked to be within buflen. If this condition is true, hv_ringbuffer_read() returns with these variables already set to some value even though no data is actually read. This might create inconsistencies in any routine calling hv_ringbuffer_read(). Assign values to such pointers after the packetlen check.

---
 drivers/hv/ring_buffer.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/hv/ring_buffer.c +0 -1
index 356e22159e83..e277ce7372a4 100644
@@ -350,12 +350,13 @@ int hv_ringbuffer_read(struct vmbus_channel
 *channel,
 
 	offset = raw ? 0 : (desc->offset8 << 3);
 	packetlen = (desc->len8 << 3) - offset;
-	*buffer_actual_len = packetlen;

drivers/hv/ring_buffer.c int hv_ringbuffer_read(struct vmbus_channel · line 362

Quoted 10 lines
drivers/hv/ring_buffer.c +3 -1
-	*requestid = desc->trans_id;
 
 	if (unlikely(packetlen > buflen))
 		return -ENOBUFS;
 
+	*buffer_actual_len = packetlen;
+	*requestid = desc->trans_id;
+
 	/* since ring is double mapped, only one copy is necessary */
 	memcpy(buffer, (const char *)desc + offset, packetlen);

What is the rationale for this change, it may break other code.

A common API model in Windows world where this originated is to have a call where caller first
makes request and then if the requested buffer is not big enough the caller look at the actual length and allocate a bigger buffer.

Did you audit all the users of this API to make sure they aren't doing that.

maintainer
#2 lore raw
Quoted 53 lines

-----Original Message-----
From: Stephen Hemminger <[email protected]>
Sent: Friday, July 24, 2020 1:11 PM
To: Andres Beltran <[email protected]>
Cc: KY Srinivasan <[email protected]>; Haiyang Zhang <[email protected]>; Stephen Hemminger <[email protected]>; [email protected]; [email protected]; linux- [email protected]; Michael Kelley <[email protected]>; [email protected]; Saruhan Karademir <[email protected]>
Subject: Re: [PATCH] Drivers: hv: vmbus: Fix variable assignments in hv_ringbuffer_read()

Quoted 23 lines · On Fri, 24 Jul 2020 09:46:06 -0700 "Andres Beltran" <[email protected]> wrote:

Assignments to buffer_actual_len and requestid happen before packetlen is checked to be within buflen. If this condition is true, hv_ringbuffer_read() returns with these variables already set to some value even though no data is actually read. This might create inconsistencies in any routine calling hv_ringbuffer_read(). Assign values to such pointers after the packetlen check.

---
 drivers/hv/ring_buffer.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/hv/ring_buffer.c +0 -1
index 356e22159e83..e277ce7372a4 100644
@@ -350,12 +350,13 @@ int hv_ringbuffer_read(struct vmbus_channel
 *channel,
 
 	offset = raw ? 0 : (desc->offset8 << 3);
 	packetlen = (desc->len8 << 3) - offset;
-	*buffer_actual_len = packetlen;

drivers/hv/ring_buffer.c int hv_ringbuffer_read(struct vmbus_channel · line 362

Quoted 10 lines
drivers/hv/ring_buffer.c +3 -1
-	*requestid = desc->trans_id;
 
 	if (unlikely(packetlen > buflen))
 		return -ENOBUFS;
 
+	*buffer_actual_len = packetlen;
+	*requestid = desc->trans_id;
+
 	/* since ring is double mapped, only one copy is necessary */
 	memcpy(buffer, (const char *)desc + offset, packetlen);

What is the rationale for this change, it may break other code.

Quoted 6 lines

A common API model in Windows world where this originated is to have a call where caller first
makes request and then if the requested buffer is not big enough the caller look at the actual length and allocate a bigger buffer.

Did you audit all the users of this API to make sure they aren't doing that.

I took a quick look. I saw a case which treats the ringbuffer_read as successful if the buffer_actual_len > 0. So if hv_ringbuffer_read() should not be fixed this way, then all callers like balloon_onchannelcallback() need a fix.

1476 static void balloon_onchannelcallback(void *context) 1477 {

1478         struct hv_device *dev = context;
1479         u32 recvlen;
1480         u64 requestid;
1481         struct dm_message *dm_msg;
1482         struct dm_header *dm_hdr;
1483         struct hv_dynmem_device *dm = hv_get_drvdata(dev);
1484         struct dm_balloon *bal_msg;
1485         struct dm_hot_add *ha_msg;
1486         union dm_mem_page_range *ha_pg_range;
1487         union dm_mem_page_range *ha_region;

1488
1489 memset(recv_buffer, 0, sizeof(recv_buffer));
1490 vmbus_recvpacket(dev->channel, recv_buffer,
1491 HV_HYP_PAGE_SIZE, &recvlen, &requestid);
1492
1493 if (recvlen > 0) {
1494 dm_msg = (struct dm_message *)recv_buffer;
1495 dm_hdr = &dm_msg->hdr;

Thanks,
- Haiyang

#3 lore raw
In reply to Stephen Hemminger
Quoted 9 lines · Stephen Hemminger

What is the rationale for this change, it may break other code.

A common API model in Windows world where this originated is to have a call where caller first
makes request and then if the requested buffer is not big enough the caller look at the actual length and allocate a bigger buffer.

Did you audit all the users of this API to make sure they aren't doing that.

The rationale for the change was to solve instances like the one @Haiyang Zhang pointed out, especially in hv_utils, which needs additional hardening. Unfortunately, there is an instance in hv_pci_onchannelcallback() that does what you just described. Thus, the fix will have to be made to all the callers of vmbus_recvpacket() and vmbus_recvpacket_raw() to make sure they check the return value, which most callers are not doing now. Thanks for pointing out this behavior. I was not aware that the length can be checked by callers to allocate a bigger buffer.

maintainer
#4 lore raw
Quoted 27 lines

-----Original Message-----
From: Andres Beltran <[email protected]>
Sent: Friday, July 24, 2020 7:04 PM
To: Stephen Hemminger <[email protected]>
Cc: KY Srinivasan <[email protected]>; Haiyang Zhang <[email protected]>; Stephen Hemminger <[email protected]>; Wei Liu <[email protected]>; [email protected]; linux- [email protected]; Michael Kelley <[email protected]>; Andrea Parri <[email protected]>; Saruhan Karademir <[email protected]>
Subject: Re: [PATCH] Drivers: hv: vmbus: Fix variable assignments in hv_ringbuffer_read()

Quoted 9 lines · Stephen Hemminger

What is the rationale for this change, it may break other code.

A common API model in Windows world where this originated is to have a call where caller first
makes request and then if the requested buffer is not big enough the caller look at the actual length and allocate a bigger buffer.

Did you audit all the users of this API to make sure they aren't doing that.

Quoted 9 lines

The rationale for the change was to solve instances like the one @Haiyang Zhang pointed out, especially in hv_utils, which needs additional hardening. Unfortunately, there is an instance in hv_pci_onchannelcallback() that does what you just described. Thus, the fix will have to be made to all the callers of vmbus_recvpacket() and vmbus_recvpacket_raw() to make sure they check the return value, which most callers are not doing now. Thanks for pointing out this behavior. I was not aware that the length can be checked by callers to allocate a bigger buffer.

To prevent future coding error, please add code comments for hv_ringbuffer_read() to indicate that the buffer_actual_len may be nonzero when the function fails, and should not be used to determine if the function succeeds or not.

Thanks,
- Haiyang