net: fix rte_vlan_insert with shared mbuf
PATCH 1/1 net: fix rte_vlan_insert with shared mbuf
If mbuf refcnt was > 1 then rte_vlan_insert() would incorrectly modify the original copy. Original code was expecting clone to make a copy (it doesn't). Better to let the caller deal with making a copy or setting up mbuf chain to allow for header to be added.
Also fix docbook comment about parameters (function takes pointer to mbuf).
--- lib/librte_net/rte_ether.h | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-)
lib/librte_net/rte_ether.h +4 -11
@@ -374,7 +374,7 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success * -EPERM: mbuf is is shared overwriting would be unsafe
@@ -385,16 +385,9 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) { -struct rte_mbuf *copy; - - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) -return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; - } +/* Can't directly insert header if mbuf is shared */ +if (rte_mbuf_refcnt_read (*m) >1 ) +return -EPERM ; oh = rte_pktmbuf_mtod(*m, struct ether_hdr *); nh = (struct ether_hdr *)
Signature · 1 line
2.17.1
Quoted 33 lines · Stephen Hemminger
If mbuf refcnt was > 1 then
rte_vlan_insert()would incorrectly modify the original copy. Original code was expecting clone to make a copy (it doesn't). Better to let the caller deal with making a copy or setting up mbuf chain to allow for header to be added.Also fix docbook comment about parameters (function takes pointer to mbuf).
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>--- lib/librte_net/rte_ether.h | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-)lib/librte_net/rte_ether.h +1
-3@@ -374,7 +374,7 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success * -EPERM: mbuf is is shared overwriting would be unsafe @@ -385,16 +385,9 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) {
lib/librte_net/rte_ether.h static inline int rte_vlan_insert(struct rte_mbuf **m) · line 389
Quoted 10 lines
lib/librte_net/rte_ether.h +2
-8-struct rte_mbuf *copy; - - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) -return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; - } +/* Can't directly insert header if mbuf is shared */ +if (rte_mbuf_refcnt_read (*m) >1 )
This check probably isn't sufficient. You need something more like:
if (rte_mbuf_refcnt_read(mbuf) > 1 ||
(RTE_MBUF_INDIRECT(mbuf) &&
rte_mbuf_refcnt_read(rte_mbuf_from_indirect(mbuf)) > 1)) {If this is a cloned packet, the refcnt will be 1. So you need to examine the parent mbuf to see if other clones exist.
Quoted 5 lines
lib/librte_net/rte_ether.h +1
-0+return -EPERM ; oh = rte_pktmbuf_mtod(*m, struct ether_hdr *); nh = (struct ether_hdr *)
Quoted 47 lines · On Tue, 26 Mar 2019 18:38:57 -0400 Chas Williams <[email protected]> wrote:
Quoted 33 lines · Stephen Hemminger
If mbuf refcnt was > 1 then
rte_vlan_insert()would incorrectly modify the original copy. Original code was expecting clone to make a copy (it doesn't). Better to let the caller deal with making a copy or setting up mbuf chain to allow for header to be added.Also fix docbook comment about parameters (function takes pointer to mbuf).
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>--- lib/librte_net/rte_ether.h | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-)lib/librte_net/rte_ether.h +1
-3@@ -374,7 +374,7 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success * -EPERM: mbuf is is shared overwriting would be unsafe @@ -385,16 +385,9 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) {lib/librte_net/rte_ether.h static inline int rte_vlan_insert(struct rte_mbuf **m) · line 389
Quoted 10 lines
lib/librte_net/rte_ether.h +2
-8-struct rte_mbuf *copy; - - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) -return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; - } +/* Can't directly insert header if mbuf is shared */ +if (rte_mbuf_refcnt_read (*m) >1 )This check probably isn't sufficient. You need something more like:
Quoted 7 lines
if (rte_mbuf_refcnt_read(mbuf) > 1 || (RTE_MBUF_INDIRECT(mbuf) && rte_mbuf_refcnt_read(rte_mbuf_from_indirect(mbuf)) > 1)) {If this is a cloned packet, the refcnt will be 1. So you need to examine the parent mbuf to see if other clones exist.
The problem is that indirect buffers probably can't be overwritten because of lack of headroom.
Actually, why not push the problem into the pktmbuf_headroom logic?
Quoted 54 lines · Stephen Hemminger
Quoted 51 lines · On Tue, 26 Mar 2019 18:38:57 -0400 Chas Williams <[email protected]> wrote:
Quoted 33 lines · Stephen Hemminger
If mbuf refcnt was > 1 then
rte_vlan_insert()would incorrectly modify the original copy. Original code was expecting clone to make a copy (it doesn't). Better to let the caller deal with making a copy or setting up mbuf chain to allow for header to be added.Also fix docbook comment about parameters (function takes pointer to mbuf).
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>--- lib/librte_net/rte_ether.h | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-)lib/librte_net/rte_ether.h +1
-3@@ -374,7 +374,7 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success * -EPERM: mbuf is is shared overwriting would be unsafe @@ -385,16 +385,9 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) {lib/librte_net/rte_ether.h static inline int rte_vlan_insert(struct rte_mbuf **m) · line 389
Quoted 10 lines
lib/librte_net/rte_ether.h +2
-8-struct rte_mbuf *copy; - - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) -return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; - } +/* Can't directly insert header if mbuf is shared */ +if (rte_mbuf_refcnt_read (*m) >1 )This check probably isn't sufficient. You need something more like:
if (rte_mbuf_refcnt_read(mbuf) > 1 || (RTE_MBUF_INDIRECT(mbuf) && rte_mbuf_refcnt_read(rte_mbuf_from_indirect(mbuf)) > 1)) {
Quoted 8 lines
Quoted 3 lines
If this is a cloned packet, the refcnt will be 1. So you need to examine the parent mbuf to see if other clones exist.
The problem is that indirect buffers probably can't be overwritten because of lack of headroom.
Actually, why not push the problem into the
pktmbuf_headroomlogic?
That's not what the original code is checking and why it is checking. You should not modify the data of a packet that has other users. To check all the possible users, you need to check your refcnt and if a clone, check the parent to see if any other clones exist. If they do, you can't safely modify these packets. Yes, we have run into this bug. Yes, it was hard to find.
Someone local write a slightly different version of rte_vlan_insert that clones the packet and prepends an mbuf so you can safely insert the VLAN information. I will see about getting it submitted.
Quoted 61 lines · Chas Williams
Quoted 58 lines · Stephen Hemminger
Quoted 54 lines · On Tue, 26 Mar 2019 18:38:57 -0400 Chas Williams <[email protected]> wrote:
Quoted 33 lines · Stephen Hemminger
If mbuf refcnt was > 1 then
rte_vlan_insert()would incorrectly modify the original copy. Original code was expecting clone to make a copy (it doesn't). Better to let the caller deal with making a copy or setting up mbuf chain to allow for header to be added.Also fix docbook comment about parameters (function takes pointer to mbuf).
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>--- lib/librte_net/rte_ether.h | 15 ++++----------- 1 file changed, 4 insertions(+), 11 deletions(-)lib/librte_net/rte_ether.h +1
-3@@ -374,7 +374,7 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success * -EPERM: mbuf is is shared overwriting would be unsafe @@ -385,16 +385,9 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) {lib/librte_net/rte_ether.h static inline int rte_vlan_insert(struct rte_mbuf **m) · line 389
Quoted 10 lines
lib/librte_net/rte_ether.h +2
-8-struct rte_mbuf *copy; - - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) -return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; - } +/* Can't directly insert header if mbuf is shared */ +if (rte_mbuf_refcnt_read (*m) >1 )This check probably isn't sufficient. You need something more like:
if (rte_mbuf_refcnt_read(mbuf) > 1 || (RTE_MBUF_INDIRECT(mbuf) && rte_mbuf_refcnt_read(rte_mbuf_from_indirect(mbuf)) > 1)) {If this is a cloned packet, the refcnt will be 1. So you need to examine the parent mbuf to see if other clones exist.
Quoted 7 lines
Quoted 4 lines
The problem is that indirect buffers probably can't be overwritten because of lack of headroom.
Actually, why not push the problem into the
pktmbuf_headroomlogic?That's not what the original code is checking and why it is checking. You should not modify the data of a packet that has other users.
+1, commit log mentions from the same problem I think.
Quoted 9 lines
To check
all the possible users, you need to check your refcnt and if a clone, check the parent to see if any other clones exist. If they do, you can't safely modify these packets. Yes, we have run into this bug. Yes, it was hard to find.Someone local write a slightly different version of
rte_vlan_insertthat clones the packet and prepends an mbuf so you can safely insert the VLAN information. I will see about getting it submitted.
Thanks for it, it looks like duplicating the mbuf was the original intention of the code.
But if we can't get updated 'rte_vlan_insert' timely, I suggest getting this patch with extended checks that you suggested. So can it be possible to make new version of this patch in any case?
RFC v2 1/1 net: fix rte_vlan_insert with shared mbuf
If mbuf is shared then rte_vlan_insert() would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.
---
v2 - compile tested only, do copy/clone.
lib/librte_net/rte_ether.h | 37 ++++++++++++++++++++++++++++--------- 1 file changed, 28 insertions(+), 9 deletions(-)
lib/librte_net/rte_ether.h +28 -9
@@ -374,10 +374,10 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success -* -EPERM: mbuf is is shared overwriting would be unsafe +* -ENOMEM: could not allocate mbuf for header * -ENOSPC: not enough headroom in mbuf */ static inline int rte_vlan_insert (struct rte_mbuf **m)
@@ -385,15 +385,34 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) { -struct rte_mbuf *copy; +/* Can't safely directly insert header if mbuf is shared or indirect */ +if (!RTE_MBUF_DIRECT (*m) ||rte_mbuf_refcnt_read (*m) >1 ) { +struct rte_mempool *mp = (*m)->pool; +struct rte_mbuf *md, *mh; + + mh =rte_pktmbuf_alloc (mp); +if (unlikely (mh ==NULL )) +return -ENOMEM ; + + mh->tx_offload = (*m)->tx_offload; + mh->vlan_tci = (*m)->vlan_tci; + mh->vlan_tci_outer = (*m)->vlan_tci_outer; + mh->port = (*m)->port; + mh->ol_flags = (*m)->ol_flags; + mh->packet_type = (*m)->packet_type; - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) + md =rte_pktmbuf_clone (*m, mp); +if (unlikely (md ==NULL )) { +rte_pktmbuf_free (mh);return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; + } + + mh->next = md; + mh->nb_segs = md->nb_segs +1 ; +memcpy (rte_pktmbuf_append (mh,ETHER_HDR_LEN ), + rte_pktmbuf_mtod(md, void *), ETHER_HDR_LEN); +rte_pktmbuf_adj (md,ETHER_HDR_LEN ); + *m = mh; } oh = rte_pktmbuf_mtod(*m, struct ether_hdr *);
Signature · 1 line
2.17.1
Unfortunately, I think the complete fix is more complicated than this. Drivers that use rte_vlan_insert don't anticipate that the mbuf might change and that (hardware) transmit can fail.
They make a copy of the mbuf pointer from the incoming transmit list and don't update the original if rte_vlan_insert creates a new mbuf. If transmit fails, the application needs to be the given the new mbuf for re-transmission or freeing.
Quoted 73 lines · Stephen Hemminger
If mbuf is shared then
rte_vlan_insert()would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>---
v2 - compile tested only, do copy/clone.lib/librte_net/rte_ether.h | 37 ++++++++++++++++++++++++++++--------- 1 file changed, 28 insertions(+), 9 deletions(-)lib/librte_net/rte_ether.h +28
-9@@ -374,10 +374,10 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success -* -EPERM: mbuf is is shared overwriting would be unsafe +* -ENOMEM: could not allocate mbuf for header * -ENOSPC: not enough headroom in mbuf */ static inline int rte_vlan_insert (struct rte_mbuf **m)@@ -385,15 +385,34 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) { -struct rte_mbuf *copy; +/* Can't safely directly insert header if mbuf is shared or indirect */ +if (!RTE_MBUF_DIRECT (*m) ||rte_mbuf_refcnt_read (*m) >1 ) { +struct rte_mempool *mp = (*m)->pool; +struct rte_mbuf *md, *mh; + + mh =rte_pktmbuf_alloc (mp); +if (unlikely (mh ==NULL )) +return -ENOMEM ; + + mh->tx_offload = (*m)->tx_offload; + mh->vlan_tci = (*m)->vlan_tci; + mh->vlan_tci_outer = (*m)->vlan_tci_outer; + mh->port = (*m)->port; + mh->ol_flags = (*m)->ol_flags; + mh->packet_type = (*m)->packet_type; - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) + md =rte_pktmbuf_clone (*m, mp); +if (unlikely (md ==NULL )) { +rte_pktmbuf_free (mh);return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; + } + + mh->next = md; + mh->nb_segs = md->nb_segs +1 ; +memcpy (rte_pktmbuf_append (mh,ETHER_HDR_LEN ), + rte_pktmbuf_mtod(md, void *), ETHER_HDR_LEN); +rte_pktmbuf_adj (md,ETHER_HDR_LEN ); + *m = mh; } oh = rte_pktmbuf_mtod(*m, struct ether_hdr *);
Quoted 4 lines · On Sat, 30 Mar 2019 08:41:33 -0400 Chas Williams <[email protected]> wrote:
Unfortunately, I think the complete fix is more complicated than this. Drivers that use
rte_vlan_insertdon't anticipate that the mbuf might change and that (hardware) transmit can fail.
Quoted 13 lines
They make a copy of the mbuf pointer from the incoming transmit list and don't update the original if
rte_vlan_insertcreates a new mbuf. If transmit fails, the application needs to be the given the new mbuf for re-transmission or freeing.Quoted 7 lines · Stephen Hemminger
If mbuf is shared then
rte_vlan_insert()would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>
The virtio driver is buggy, it saves original mbuf and doesn't handle rewrite. Dpaa2 is fine, should never happen since it checks refcnt etc. Netvsc PMD is using this on rx path and is safe.AF_packet PMD should work fine as well.
So virtio is the one that needs fixing
Quoted 18 lines · Stephen Hemminger
Quoted 15 lines · On Sat, 30 Mar 2019 08:41:33 -0400 Chas Williams <[email protected]> wrote:
Unfortunately, I think the complete fix is more complicated than this. Drivers that use
rte_vlan_insertdon't anticipate that the mbuf might change and that (hardware) transmit can fail.They make a copy of the mbuf pointer from the incoming transmit list and don't update the original if
rte_vlan_insertcreates a new mbuf. If transmit fails, the application needs to be the given the new mbuf for re-transmission or freeing.Quoted 5 lines · Stephen Hemminger
If mbuf is shared then
rte_vlan_insert()would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.
Quoted 7 lines
Quoted 2 lines
Quoted 2 lines
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>The virtio driver is buggy, it saves original mbuf and doesn't handle rewrite. Dpaa2 is fine, should never happen since it checks refcnt etc. Netvsc PMD is using this on rx path and is safe.
AF_packetPMD should work fine as well.
af_packet is broken as well. It needs to defer rte_vlan_insert until after it gets the next incoming frame. Otherwise the break can exit the loop early.
Quoted 1 line
So virtio is the one that needs fixing
I agree with this.
Quoted 4 lines · Stephen Hemminger
If mbuf is shared then
rte_vlan_insert()would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.
Hi Stephen, Chas,
Since this patch is still an RFC, and we are after RC1, I think it is better to consider this patch for next release.
But for this release, what do you think having a patch to return error on 'rte_vlan_insert()' for shared mbufs?
Quoted 69 lines
Fixes c974021a5949 ("ether: add soft vlan encap/decap")Signed-off-by Stephen Hemminger <[email protected]>---
v2 - compile tested only, do copy/clone.lib/librte_net/rte_ether.h | 37 ++++++++++++++++++++++++++++--------- 1 file changed, 28 insertions(+), 9 deletions(-)lib/librte_net/rte_ether.h +28
-9@@ -374,10 +374,10 @@ static inline int rte_vlan_strip(struct rte_mbuf *m)* Software version of VLAN unstripping * * @param m -* The packet mbuf. +* Pointer to the packet mbuf. * @return * - 0: On success -* -EPERM: mbuf is is shared overwriting would be unsafe +* -ENOMEM: could not allocate mbuf for header * -ENOSPC: not enough headroom in mbuf */ static inline int rte_vlan_insert (struct rte_mbuf **m)@@ -385,15 +385,34 @@ static inline int rte_vlan_insert(struct rte_mbuf **m)struct ether_hdr *oh, *nh;struct vlan_hdr *vh; -/* Can't insert header if mbuf is shared */ -if (rte_mbuf_refcnt_read (*m) >1 ) { -struct rte_mbuf *copy; +/* Can't safely directly insert header if mbuf is shared or indirect */ +if (!RTE_MBUF_DIRECT (*m) ||rte_mbuf_refcnt_read (*m) >1 ) { +struct rte_mempool *mp = (*m)->pool; +struct rte_mbuf *md, *mh; + + mh =rte_pktmbuf_alloc (mp); +if (unlikely (mh ==NULL )) +return -ENOMEM ; + + mh->tx_offload = (*m)->tx_offload; + mh->vlan_tci = (*m)->vlan_tci; + mh->vlan_tci_outer = (*m)->vlan_tci_outer; + mh->port = (*m)->port; + mh->ol_flags = (*m)->ol_flags; + mh->packet_type = (*m)->packet_type; - copy =rte_pktmbuf_clone (*m, (*m)->pool); -if (unlikely (copy ==NULL )) + md =rte_pktmbuf_clone (*m, mp); +if (unlikely (md ==NULL )) { +rte_pktmbuf_free (mh);return -ENOMEM ; -rte_pktmbuf_free (*m); - *m = copy; + } + + mh->next = md; + mh->nb_segs = md->nb_segs +1 ; +memcpy (rte_pktmbuf_append (mh,ETHER_HDR_LEN ), + rte_pktmbuf_mtod(md, void *), ETHER_HDR_LEN); +rte_pktmbuf_adj (md,ETHER_HDR_LEN ); + *m = mh; } oh = rte_pktmbuf_mtod(*m, struct ether_hdr *);
Quoted 6 lines · Ferruh Yigit
Quoted 4 lines · Stephen Hemminger
If mbuf is shared then
rte_vlan_insert()would clobber the original Ethernet header. The changed version handles this by getting an mbuf that will hold the new Ethernet and VLAN header followed by another mbuf (cloned) for the data.
Quoted 7 lines
Hi Stephen, Chas,
Since this patch is still an RFC, and we are after RC1, I think it is better to consider this patch for next release.
But for this release, what do you think having a patch to return error on '
rte_vlan_insert()' for shared mbufs?
nack for this patch
The other approach to not support the shared mbuf merged:
https://git.dpdk.org/dpdk/commit/?id=15a74163b12