Revert "xfrm: interface with if_id 0 should return error"
Subsystems (3)
PATCH 1/2 Revert "xfrm: interface with if_id 0 should return error"
From: Kai Lueke <[email protected]>
This reverts commit 8dce43919566f06e865f7e8949f5c10d8c2493f5 because it breaks userspace (e.g., Cilium is affected because it used id 0 for the dummy state https://github.com/cilium/cilium/pull/18789).
--- net/xfrm/xfrm_interface.c | 14 ++------------ 1 file changed, 2 insertions(+), 12 deletions(-)
net/xfrm/xfrm_interface.c +2 -12
@@ -637,16 +637,11 @@ static int xfrmi_newlink(struct net *src_net, struct net_device *dev, struct netlink_ext_ack *extack) {struct net *net =dev_net (dev); -struct xfrm_if_parms p = {}; +struct xfrm_if_parms p;struct xfrm_if *xi;int err;xfrmi_netlink_parms (data, &p); -if (!p.if_id) { -NL_SET_ERR_MSG (extack,"if_id must be non zero" ); -return -EINVAL ; - } - xi =xfrmi_locate (net, &p);if (xi)return -EEXIST ;
@@ -671,12 +666,7 @@ static int xfrmi_changelink(struct net_device *dev, struct nlattr *tb[], {struct xfrm_if *xi =netdev_priv (dev);struct net *net = xi->net; -struct xfrm_if_parms p = {}; - -if (!p.if_id) { -NL_SET_ERR_MSG (extack,"if_id must be non zero" ); -return -EINVAL ; - } +struct xfrm_if_parms p;xfrmi_netlink_parms (data, &p); xi =xfrmi_locate (net, &p);
Signature · 1 line
2.25.1
PATCH 2/2 Revert "xfrm: state and policy should fail if XFRMA_IF_ID 0"
From: Kai Lueke <[email protected]>
This reverts commit 68ac0f3810e76a853b5f7b90601a05c3048b8b54 because it breaks userspace (e.g., Cilium is affected because it used id 0 for the dummy state https://github.com/cilium/cilium/pull/18789).
--- net/xfrm/xfrm_user.c | 21 +++------------------ 1 file changed, 3 insertions(+), 18 deletions(-)
net/xfrm/xfrm_user.c +3 -18
@@ -630,13 +630,8 @@ static struct xfrm_state *xfrm_state_construct(struct net *net,xfrm_smark_init (attrs, &x->props.smark); -if (attrs[XFRMA_IF_ID ]) { +if (attrs[XFRMA_IF_ID ]) x->if_id =nla_get_u32 (attrs[XFRMA_IF_ID ]); -if (!x->if_id) { - err = -EINVAL ; - goto error; - } - } err =__xfrm_init_state (x, false, attrs[XFRMA_OFFLOAD_DEV ]);if (err)
@@ -1432,13 +1427,8 @@ static int xfrm_alloc_userspi(struct sk_buff *skb, struct nlmsghdr *nlh, mark =xfrm_mark_get (attrs, &m); -if (attrs[XFRMA_IF_ID ]) { +if (attrs[XFRMA_IF_ID ]) if_id =nla_get_u32 (attrs[XFRMA_IF_ID ]); -if (!if_id) { - err = -EINVAL ; - goto out_noput; - } - }if (p->info.seq) { x =xfrm_find_acq_byseq (net, mark, p->info.seq);
@@ -1751,13 +1741,8 @@ static struct xfrm_policy *xfrm_policy_construct(struct net *net, struct xfrm_usxfrm_mark_get (attrs, &xp->mark); -if (attrs[XFRMA_IF_ID ]) { +if (attrs[XFRMA_IF_ID ]) xp->if_id =nla_get_u32 (attrs[XFRMA_IF_ID ]); -if (!xp->if_id) { - err = -EINVAL ; - goto error; - } - }return xp; error:
Signature · 1 line
2.25.1
Hi Kai,
Quoted 8 lines · On Tue, Mar 1, 2022 at 3:15 PM <[email protected]> wrote:
From: Kai Lueke <[email protected]>
This reverts commit
8dce43919566f06e865f7e8949f5c10d8c2493f5because it breaks userspace (e.g., Cilium is affected because it used id 0 for the dummy state https://github.com/cilium/cilium/pull/18789).Signed-off-by Kai Lueke <[email protected]>
From the pull request you mentioned I understand the commit which affected Cilium userspace is 68ac0f3810e7
("xfrm: state and policy should fail if XFRMA_IF_ID 0").
Whereas 8dce43919566 ("xfrm: interface with if_id 0 should return error") involves xfrm interfaces which don't appear in the pull request.
In which case, why should that commit be reverted?
Eyal.
Hi,
Quoted 4 lines
Whereas
8dce43919566("xfrm: interface withif_id0 should return error") involves xfrm interfaces which don't appear in the pull request.In which case, why should that commit be reverted?
Correct me if I misunderstood this but reading the commit message it is explicitly labeled as a behavior change for userspace:
With this commit:
ip link add ipsec0 type xfrm dev lo if_id 0
Error: if_id must be non zero.
Changing behavior this way is from my understanding a regression because it breaks programs that happened to work before, even if they worked incorrect (cf. https://lwn.net/Articles/726021/ "The current process for Linux development says that kernel patches cannot break programs that rely on the ABI. That means a program that runs on the 4.0 kernel should be able to run on the 5.0 kernel, Levin said.").
Regards,
Kai
Hi Kai,
Quoted 14 lines · Kai Lüke
Hi,
Quoted 4 lines
Whereas
8dce43919566("xfrm: interface withif_id0 should return error") involves xfrm interfaces which don't appear in the pull request.In which case, why should that commit be reverted?
Correct me if I misunderstood this but reading the commit message it is explicitly labeled as a behavior change for userspace:
With this commit:
ip link add ipsec0 type xfrm dev loif_id0
Error:if_idmust be non zero.
Quoted 6 lines
Changing behavior this way is from my understanding a regression because it breaks programs that happened to work before, even if they worked incorrect (cf. https://lwn.net/Articles/726021/ "The current process for Linux development says that kernel patches cannot break programs that rely on the ABI. That means a program that runs on the 4.0 kernel should be able to run on the 5.0 kernel, Levin said.").
Well to some extent, but the point was that xfrm interfaces with if_id=0 were already broken, so returning an error to userspace in such case would be a better behavior.
So I'm not sure this is a regression but it's not up to me to decide these things.
Eyal
Quoted 24 lines · Eyal Birger
Hi Kai,
Quoted 14 lines · Kai Lüke
Hi,
Quoted 4 lines
Whereas
8dce43919566("xfrm: interface withif_id0 should return error") involves xfrm interfaces which don't appear in the pull request.In which case, why should that commit be reverted?
Correct me if I misunderstood this but reading the commit message it is explicitly labeled as a behavior change for userspace:
With this commit:
ip link add ipsec0 type xfrm dev loif_id0
Error:if_idmust be non zero.Quoted 6 lines
Changing behavior this way is from my understanding a regression because it breaks programs that happened to work before, even if they worked incorrect (cf. https://lwn.net/Articles/726021/ "The current process for Linux development says that kernel patches cannot break programs that rely on the ABI. That means a program that runs on the 4.0 kernel should be able to run on the 5.0 kernel, Levin said.").
Quoted 5 lines
Well to some extent, but the point was that xfrm interfaces with if_id=0 were already broken, so returning an error to userspace in such case would be a better behavior.
So I'm not sure this is a regression but it's not up to me to decide these things.
I agree with Eyal here. As far as Cilium is concerned, this is not causing any regression. Only the second commit, 68ac0f3810e7 ("xfrm:
state and policy should fail if XFRMA_IF_ID 0") causes issues in a previously-working setup in Cilium. We don't use xfrm interfaces.
Paul
Quoted 5 lines
I agree with Eyal here. As far as Cilium is concerned, this is not causing any regression. Only the second commit,
68ac0f3810e7("xfrm:
state and policy should fail ifXFRMA_IF_ID0") causes issues in a previously-working setup in Cilium. We don't use xfrm interfaces.
I see this as a very generic question of changing userspace behavior or not, regardless if we know how many users are affected, and from what I know there are similar cases in the kernel where the response was that breaking userspace is a no go - even if the intention was to be helpful by having early errors.
Greets,
Kai
Steffen Klassert
maintainerQuoted 10 lines · Kai Lüke
Quoted 5 lines
I agree with Eyal here. As far as Cilium is concerned, this is not causing any regression. Only the second commit,
68ac0f3810e7("xfrm:
state and policy should fail ifXFRMA_IF_ID0") causes issues in a previously-working setup in Cilium. We don't use xfrm interfaces.I see this as a very generic question of changing userspace behavior or not, regardless if we know how many users are affected, and from what I know there are similar cases in the kernel where the response was that breaking userspace is a no go - even if the intention was to be helpful by having early errors.
In general I agree that the userspace ABI has to be stable, but this never worked. We changed the behaviour from silently broken to notify userspace about a misconfiguration.
It is the question what is more annoying for the users. A bug that we can never fix, or changing a broken behaviour to something that tells you at least why it is not working.
In such a case we should gauge what's the better solution. Here I tend to keep it as it is.
Hi,
Quoted 10 lines
In general I agree that the userspace ABI has to be stable, but this never worked. We changed the behaviour from silently broken to notify userspace about a misconfiguration.
It is the question what is more annoying for the users. A bug that we can never fix, or changing a broken behaviour to something that tells you at least why it is not working.
In such a case we should gauge what's the better solution. Here I tend to keep it as it is.
alternatives are: docs to ensure the API is used the right way, maybe a dmesg log entry if wrong usage is detected, and filing bugs where the API is used wrong.
The chosen way led to having this change being introduced as part of an LTS kernel bugfix update, breaking user's clusters:
https://github.com/flatcar-linux/Flatcar/issues/626
Please rethink how you want to handle this, also for not making a precedent here so that this repeats.
Regards,
Kai (Flatcar Container Linux team)
Le 01/03/2022 à 17:44, Kai Lueke a écrit :
Quoted 9 lines
Hi,
Quoted 8 lines
In general I agree that the userspace ABI has to be stable, but this never worked. We changed the behaviour from silently broken to notify userspace about a misconfiguration.
It is the question what is more annoying for the users. A bug that we can never fix, or changing a broken behaviour to something that tells you at least why it is not working.
Quoted 6 lines
Quoted 2 lines
In such a case we should gauge what's the better solution. Here I tend to keep it as it is.
alternatives are: docs to ensure the API is used the right way, maybe a dmesg log entry if wrong usage is detected, and filing bugs where the API is used wrong.
I agree with that proposal (dmesg log). Breaking an existing script, even if it made something wrong is really painful. And maybe this broken xfrm interface was unused, so everything worked well before the patch and is now broken.
My two cents,
Nicolas
Jakub Kicinski
maintainerQuoted 10 lines · On Tue, 1 Mar 2022 17:10:01 +0100 Steffen Klassert wrote:
Quoted 5 lines
I see this as a very generic question of changing userspace behavior or not, regardless if we know how many users are affected, and from what I know there are similar cases in the kernel where the response was that breaking userspace is a no go - even if the intention was to be helpful by having early errors.
In general I agree that the userspace ABI has to be stable, but this never worked. We changed the behaviour from silently broken to notify userspace about a misconfiguration.
Quoted 6 lines
It is the question what is more annoying for the users. A bug that we can never fix, or changing a broken behaviour to something that tells you at least why it is not working.
In such a case we should gauge what's the better solution. Here I tend to keep it as it is.
Agreed. FWIW would be great if patch #2 started flowing towards Linus'es tree separately if the discussion on #1 is taking longer.
Hi,
Quoted 2 lines
Agreed. FWIW would be great if patch #2 started flowing towards Linus'es tree separately if the discussion on #1 is taking longer.
to preserve the initial goal of helping to uncover id 0 usage I think it would be best to have the revert be accompanied by a patch that instead creates a kernel log warning (or whatever).
Since I never did that I suggest to not wait for me. Also, feel free to do the revert yourself with a different commit message if mine didn't capture the things appropriately.
Regards,
Kai
Jakub Kicinski
maintainerQuoted 6 lines · On Wed, 2 Mar 2022 19:11:06 +0100 Kai Lueke wrote:
Quoted 2 lines
Agreed. FWIW would be great if patch #2 started flowing towards Linus'es tree separately if the discussion on #1 is taking longer.
to preserve the initial goal of helping to uncover id 0 usage I think it would be best to have the revert be accompanied by a patch that instead creates a kernel log warning (or whatever).
extack would be best, but that would mean a little bit of plumbing so more likely net-next material. Which would have to come after.
Quoted 3 lines
Since I never did that I suggest to not wait for me. Also, feel free to do the revert yourself with a different commit message if mine didn't capture the things appropriately.
TBH I'm not 100% clear on the nature of the regression. Does Cilium update the configuration later to make if_id be non-zero? Or the broken interface is not used but not being able to create it fails the whole configuration?
Steffen Klassert
maintainerQuoted 10 lines · Jakub Kicinski
Quoted 6 lines · On Wed, 2 Mar 2022 19:11:06 +0100 Kai Lueke wrote:
Quoted 2 lines
Agreed. FWIW would be great if patch #2 started flowing towards Linus'es tree separately if the discussion on #1 is taking longer.
to preserve the initial goal of helping to uncover id 0 usage I think it would be best to have the revert be accompanied by a patch that instead creates a kernel log warning (or whatever).
extack would be best, but that would mean a little bit of plumbing so more likely net-next material. Which would have to come after.
We need something that a user directly sees that the configuration is invalid. That the xfrm interface acts as a blackhole whith IF_ID 0 is one thing (revert #1).
The other thing is that IF_ID 0 for a policy/state internally means that these are not assigned to a xfrm interface. So a policy/state with IF_ID 0 will match flows that are not targeted to the xfrm interface. This means that confidential stuff could be sent to the wrong peer (revert #2).
Quoted 9 lines
Quoted 4 lines
Since I never did that I suggest to not wait for me. Also, feel free to do the revert yourself with a different commit message if mine didn't capture the things appropriately.
TBH I'm not 100% clear on the nature of the regression. Does Cilium update the configuration later to make
if_idbe non-zero? Or the broken interface is not used but not being able to create it fails the whole configuration?
As far I unerstood Pauls mail, only 68ac0f3810e7 ("xfrm: state and policy should fail if XFRMA_IF_ID 0") caused issues in Cilium.
So maybe just revert that one and document that IF_ID 0 on a policy/state is the same as configuring the policy/state without IF_ID.