The commit makes a claim without any kind of justification.
The claim is false, and the commit broke lsevtchn in dom0. It is also quite
obvious from XSM_TARGET that it has broken device model stubdoms too.
Whether to return information about a xen-owned evtchn is a matter of policy,
and it's not acceptable to short circuit the XSM on the matter.
This reverts commit f60ab5337f968e2f10c639ab59db7afb0fe4f7c3.
Fixes: f60ab5337f96 ("evtchn: refuse EVTCHNOP_status for Xen-bound event channels")
Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
---
CC: George Dunlap <George.Dunlap@citrix.com>
CC: Jan Beulich <JBeulich@suse.com>
CC: Stefano Stabellini <sstabellini@kernel.org>
CC: Julien Grall <julien@xen.org>
CC: Daniel Smith <dpsmith@apertussolutions.com>
---
xen/common/event_channel.c | 6 ------
1 file changed, 6 deletions(-)
diff --git a/xen/common/event_channel.c b/xen/common/event_channel.c
index 20f586cf5ecd..ae6c2f902645 100644
--- a/xen/common/event_channel.c
+++ b/xen/common/event_channel.c
@@ -1040,12 +1040,6 @@ int evtchn_status(evtchn_status_t *status)
read_lock(&d->event_lock);
- if ( consumer_is_xen(chn) )
- {
- rc = -EACCES;
- goto out;
- }
-
rc = xsm_evtchn_status(XSM_TARGET, d, chn);
if ( rc )
goto out;
base-commit: 7a09966e7b2823b70f6d56d0cf66c11124f4a3c1
--
2.30.2
On 4/2/24 13:06, Andrew Cooper wrote: > The commit makes a claim without any kind of justification. > > The claim is false, and the commit broke lsevtchn in dom0. It is also quite > obvious from XSM_TARGET that it has broken device model stubdoms too. > > Whether to return information about a xen-owned evtchn is a matter of policy, > and it's not acceptable to short circuit the XSM on the matter. > > This reverts commit f60ab5337f968e2f10c639ab59db7afb0fe4f7c3. > > Fixes: f60ab5337f96 ("evtchn: refuse EVTCHNOP_status for Xen-bound event channels") > Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com> > --- > CC: George Dunlap <George.Dunlap@citrix.com> > CC: Jan Beulich <JBeulich@suse.com> > CC: Stefano Stabellini <sstabellini@kernel.org> > CC: Julien Grall <julien@xen.org> > CC: Daniel Smith <dpsmith@apertussolutions.com> > --- > xen/common/event_channel.c | 6 ------ > 1 file changed, 6 deletions(-) > > diff --git a/xen/common/event_channel.c b/xen/common/event_channel.c > index 20f586cf5ecd..ae6c2f902645 100644 > --- a/xen/common/event_channel.c > +++ b/xen/common/event_channel.c > @@ -1040,12 +1040,6 @@ int evtchn_status(evtchn_status_t *status) > > read_lock(&d->event_lock); > > - if ( consumer_is_xen(chn) ) > - { > - rc = -EACCES; > - goto out; > - } > - > rc = xsm_evtchn_status(XSM_TARGET, d, chn); > if ( rc ) > goto out; > > base-commit: 7a09966e7b2823b70f6d56d0cf66c11124f4a3c1 Acked-by: Daniel P. Smith <dpsmith@apertussolutions.com>
On 02.04.2024 19:06, Andrew Cooper wrote: > The commit makes a claim without any kind of justification. Well, what does "have no business" leave open? > The claim is false, and the commit broke lsevtchn in dom0. Or alternatively lsevtchn was doing something that was never meant to work (from Xen's perspective). > It is also quite > obvious from XSM_TARGET that it has broken device model stubdoms too. Why would that be "obvious"? What business would a stubdom have to look at Xen's side of an evtchn? > Whether to return information about a xen-owned evtchn is a matter of policy, > and it's not acceptable to short circuit the XSM on the matter. I can certainly accept this as one possible view point. As in so many cases I'm afraid I dislike you putting it as if it was the only possible one. In summary: The supposed justification you claim is missing in the original change is imo also missing here then: What business would any entity in the system have to look at Xen's side of an event channel? Back at the time, 3 people agreed that it's "none". Jan
On 4/3/24 02:16, Jan Beulich wrote: > On 02.04.2024 19:06, Andrew Cooper wrote: >> The commit makes a claim without any kind of justification. > > Well, what does "have no business" leave open? Why does it not have any business? Why should a domain that creates an event channel not be able to inquire about its status? >> The claim is false, and the commit broke lsevtchn in dom0. > > Or alternatively lsevtchn was doing something that was never meant to work > (from Xen's perspective). Again, you have not said why this is a problem. What concern does it create? Does it open the door for access elevation, resource deprivation, or some other malicious behaviors? >> It is also quite >> obvious from XSM_TARGET that it has broken device model stubdoms too. > > Why would that be "obvious"? What business would a stubdom have to look at > Xen's side of an evtchn? Again, you have not expressed why it shouldn't be able to do so. >> Whether to return information about a xen-owned evtchn is a matter of policy, >> and it's not acceptable to short circuit the XSM on the matter. > > I can certainly accept this as one possible view point. As in so many cases > I'm afraid I dislike you putting it as if it was the only possible one. In fact, this commit is in violation of the XSM. It hard-codes a resource access check outside XSM, thus breaking the fine-grained access control of FLASK. > In summary: The supposed justification you claim is missing in the original > change is imo also missing here then: What business would any entity in the > system have to look at Xen's side of an event channel? Back at the time, 3 > people agreed that it's "none". As stated, you provided no reason or justification for "has no business" and by face value is an opinion that a few people agreed with. As for why, there could be a myriad number of reasons a domain may want to check the status of an interface it has with the hypervisor. From just logging its state for debug to throttling attempts at sending an event. So why, from a security/access control decision, does this access have to absolutely blocked, even from FLASK? v/r, dps
On 03.04.2024 13:10, Daniel P. Smith wrote: > On 4/3/24 02:16, Jan Beulich wrote: >> On 02.04.2024 19:06, Andrew Cooper wrote: >>> The commit makes a claim without any kind of justification. >> >> Well, what does "have no business" leave open? > > Why does it not have any business? Why should a domain that creates an > event channel not be able to inquire about its status? Event channels we talk about here are created via alloc_unbound_xen_event_channel(). IOW it's not any domain creating them. Once connected, the respective domain is of course fine to query its end of the channel. >>> The claim is false, and the commit broke lsevtchn in dom0. >> >> Or alternatively lsevtchn was doing something that was never meant to work >> (from Xen's perspective). > > Again, you have not said why this is a problem. What concern does it > create? Does it open the door for access elevation, resource > deprivation, or some other malicious behaviors? It exposes information that perhaps better wouldn't be exposed. Imo if Xen owned resource state is of interest, it would want exposing via hypfs. >>> It is also quite >>> obvious from XSM_TARGET that it has broken device model stubdoms too. >> >> Why would that be "obvious"? What business would a stubdom have to look at >> Xen's side of an evtchn? > > Again, you have not expressed why it shouldn't be able to do so. See above - not its resource, nor its guest's. >>> Whether to return information about a xen-owned evtchn is a matter of policy, >>> and it's not acceptable to short circuit the XSM on the matter. >> >> I can certainly accept this as one possible view point. As in so many cases >> I'm afraid I dislike you putting it as if it was the only possible one. > > In fact, this commit is in violation of the XSM. It hard-codes a > resource access check outside XSM, thus breaking the fine-grained access > control of FLASK. Perhaps; see below and see the question raised in the subsequent reply to the patch. >> In summary: The supposed justification you claim is missing in the original >> change is imo also missing here then: What business would any entity in the >> system have to look at Xen's side of an event channel? Back at the time, 3 >> people agreed that it's "none". > > As stated, you provided no reason or justification for "has no business" > and by face value is an opinion that a few people agreed with. As for > why, there could be a myriad number of reasons a domain may want to > check the status of an interface it has with the hypervisor. From just > logging its state for debug to throttling attempts at sending an event. > So why, from a security/access control decision, does this access have > to absolutely blocked, even from FLASK? I didn't say it absolutely needs to be blocked. I'm okay to become convinced otherwise. But in the description complaining about lack of reasons in the 3-4 year old change, just to then again not provide any reasons looks "interesting" to me. (And no, just to take that example, lsevtchn not working anymore on such channels is not on its own a reason. As indicated, it may well be that conceptually it was never supposed to be able to have access to this information. The latest not anymore when hypfs was introduced.) Jan
On 4/3/24 08:05, Jan Beulich wrote: > On 03.04.2024 13:10, Daniel P. Smith wrote: >> On 4/3/24 02:16, Jan Beulich wrote: >>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>> The commit makes a claim without any kind of justification. >>> >>> Well, what does "have no business" leave open? >> >> Why does it not have any business? Why should a domain that creates an >> event channel not be able to inquire about its status? > > Event channels we talk about here are created via > alloc_unbound_xen_event_channel(). IOW it's not any domain creating them. > Once connected, the respective domain is of course fine to query its end > of the channel. I would disagree, for instance alloc_unbound_xen_event_channel() is used in response to XEN_DOMCTL_vuart_op:XEN_DOMCTL_VUART_OP_INIT and XEN_DOMCTL_VM_EVENT_OP_PAGING:XEN_VM_EVENT_ENABLE, which are hypercalls by a domain and not something initiated by the hypervisor. >>>> The claim is false, and the commit broke lsevtchn in dom0. >>> >>> Or alternatively lsevtchn was doing something that was never meant to work >>> (from Xen's perspective). >> >> Again, you have not said why this is a problem. What concern does it >> create? Does it open the door for access elevation, resource >> deprivation, or some other malicious behaviors? > > It exposes information that perhaps better wouldn't be exposed. Imo if > Xen owned resource state is of interest, it would want exposing via > hypfs. You didn't answer why, just again expressed your opinion that it is not better exposed. And I would have to wholly disagree with the sentiment that hypfs exposure is the deciding factor what is or is not worth exposing. This thinking is completely orthogonal to FLASK and fine-grained access control. >>>> It is also quite >>>> obvious from XSM_TARGET that it has broken device model stubdoms too. >>> >>> Why would that be "obvious"? What business would a stubdom have to look at >>> Xen's side of an evtchn? >> >> Again, you have not expressed why it shouldn't be able to do so. > > See above - not its resource, nor its guest's. It is a resource provided to a domain that the domain can send/raise an event to and a backing domain that can bind to it, ie. the two parameters that must be passed to the allocation call. >>>> Whether to return information about a xen-owned evtchn is a matter of policy, >>>> and it's not acceptable to short circuit the XSM on the matter. >>> >>> I can certainly accept this as one possible view point. As in so many cases >>> I'm afraid I dislike you putting it as if it was the only possible one. >> >> In fact, this commit is in violation of the XSM. It hard-codes a >> resource access check outside XSM, thus breaking the fine-grained access >> control of FLASK. > > Perhaps; see below and see the question raised in the subsequent reply > to the patch. > >>> In summary: The supposed justification you claim is missing in the original >>> change is imo also missing here then: What business would any entity in the >>> system have to look at Xen's side of an event channel? Back at the time, 3 >>> people agreed that it's "none". >> >> As stated, you provided no reason or justification for "has no business" >> and by face value is an opinion that a few people agreed with. As for >> why, there could be a myriad number of reasons a domain may want to >> check the status of an interface it has with the hypervisor. From just >> logging its state for debug to throttling attempts at sending an event. >> So why, from a security/access control decision, does this access have >> to absolutely blocked, even from FLASK? > > I didn't say it absolutely needs to be blocked. I'm okay to become > convinced otherwise. But in the description complaining about lack of > reasons in the 3-4 year old change, just to then again not provide any > reasons looks "interesting" to me. (And no, just to take that example, > lsevtchn not working anymore on such channels is not on its own a > reason. As indicated, it may well be that conceptually it was never > supposed to be able to have access to this information. The latest not > anymore when hypfs was introduced.) This broke an existing behavior, whether that behavior is correct can always be questioned, does not justify leaving an incorrect implementation. And it is incorrect because as again you have not articulated why the lsevtchn behavior is wrong and thus whether this is the valid corrective action. v/r, dps
On 03.04.2024 15:27, Daniel P. Smith wrote: > On 4/3/24 08:05, Jan Beulich wrote: >> On 03.04.2024 13:10, Daniel P. Smith wrote: >>> On 4/3/24 02:16, Jan Beulich wrote: >>>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>>> It is also quite >>>>> obvious from XSM_TARGET that it has broken device model stubdoms too. >>>> >>>> Why would that be "obvious"? What business would a stubdom have to look at >>>> Xen's side of an evtchn? >>> >>> Again, you have not expressed why it shouldn't be able to do so. >> >> See above - not its resource, nor its guest's. > > It is a resource provided to a domain that the domain can send/raise an > event to and a backing domain that can bind to it, ie. the two > parameters that must be passed to the allocation call. Before writing this particular part of your reply, did you look as evtchn_send()? Sending on such ports is similarly denied without involving XSM. For a good reason, stated in the accompanying comment. It is therefore simply inconsistent to allow any kind of other operation on such ports. Hence the patch that Andrew now deems needs reverting. In fact I view these ports living in the guest's event channel space as similarly inappropriate as the ioreq pages - until a few years back - living in the guest's GFN space. Jan
On 03.04.2024 15:27, Daniel P. Smith wrote: > On 4/3/24 08:05, Jan Beulich wrote: >> On 03.04.2024 13:10, Daniel P. Smith wrote: >>> On 4/3/24 02:16, Jan Beulich wrote: >>>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>>> The commit makes a claim without any kind of justification. >>>> >>>> Well, what does "have no business" leave open? >>> >>> Why does it not have any business? Why should a domain that creates an >>> event channel not be able to inquire about its status? >> >> Event channels we talk about here are created via >> alloc_unbound_xen_event_channel(). IOW it's not any domain creating them. >> Once connected, the respective domain is of course fine to query its end >> of the channel. > > I would disagree, for instance alloc_unbound_xen_event_channel() is used > in response to XEN_DOMCTL_vuart_op:XEN_DOMCTL_VUART_OP_INIT and > XEN_DOMCTL_VM_EVENT_OP_PAGING:XEN_VM_EVENT_ENABLE, which are hypercalls > by a domain and not something initiated by the hypervisor. Those ports, aiui, aren't supposed to be used by the caller for other than connecting an inter-domain port to the other side. >>>>> The claim is false, and the commit broke lsevtchn in dom0. >>>> >>>> Or alternatively lsevtchn was doing something that was never meant to work >>>> (from Xen's perspective). >>> >>> Again, you have not said why this is a problem. What concern does it >>> create? Does it open the door for access elevation, resource >>> deprivation, or some other malicious behaviors? >> >> It exposes information that perhaps better wouldn't be exposed. Imo if >> Xen owned resource state is of interest, it would want exposing via >> hypfs. > > You didn't answer why, just again expressed your opinion that it is not > better exposed. I'm sorry, but "better wouldn't be exposed" includes the "why" part already imo: Information should simply not be exposed unduly. For every bit of exposed information, there ought to be a reason (and then the right vehicle used for exposure). >>>>> It is also quite >>>>> obvious from XSM_TARGET that it has broken device model stubdoms too. >>>> >>>> Why would that be "obvious"? What business would a stubdom have to look at >>>> Xen's side of an evtchn? >>> >>> Again, you have not expressed why it shouldn't be able to do so. >> >> See above - not its resource, nor its guest's. > > It is a resource provided to a domain that the domain can send/raise an > event to and a backing domain that can bind to it, ie. the two > parameters that must be passed to the allocation call. I don't think so: As per above (my understanding may be wrong), it's only the other side of the connection which is available for use by a domain. Over night I was pretty close to admitting a mistake there, but upon re- checking of the sources I could only find this view of mine supported. Which doesn't mean I'm viewing things correctly; please point out my mistake if there is any. >>>>> Whether to return information about a xen-owned evtchn is a matter of policy, >>>>> and it's not acceptable to short circuit the XSM on the matter. >>>> >>>> I can certainly accept this as one possible view point. As in so many cases >>>> I'm afraid I dislike you putting it as if it was the only possible one. >>> >>> In fact, this commit is in violation of the XSM. It hard-codes a >>> resource access check outside XSM, thus breaking the fine-grained access >>> control of FLASK. >> >> Perhaps; see below and see the question raised in the subsequent reply >> to the patch. >> >>>> In summary: The supposed justification you claim is missing in the original >>>> change is imo also missing here then: What business would any entity in the >>>> system have to look at Xen's side of an event channel? Back at the time, 3 >>>> people agreed that it's "none". >>> >>> As stated, you provided no reason or justification for "has no business" >>> and by face value is an opinion that a few people agreed with. As for >>> why, there could be a myriad number of reasons a domain may want to >>> check the status of an interface it has with the hypervisor. From just >>> logging its state for debug to throttling attempts at sending an event. >>> So why, from a security/access control decision, does this access have >>> to absolutely blocked, even from FLASK? >> >> I didn't say it absolutely needs to be blocked. I'm okay to become >> convinced otherwise. But in the description complaining about lack of >> reasons in the 3-4 year old change, just to then again not provide any >> reasons looks "interesting" to me. (And no, just to take that example, >> lsevtchn not working anymore on such channels is not on its own a >> reason. As indicated, it may well be that conceptually it was never >> supposed to be able to have access to this information. The latest not >> anymore when hypfs was introduced.) > > This broke an existing behavior, whether that behavior is correct can > always be questioned, does not justify leaving an incorrect > implementation. And it is incorrect because as again you have not > articulated why the lsevtchn behavior is wrong and thus whether this is > the valid corrective action. Again - if lsevtchn is supposed to be able to access Xen-internal resources, _that_ is what needs justifying. Otherwise my take is that it is supposed to only access domain resources. Jan
On 03.04.2024 08:16, Jan Beulich wrote: > On 02.04.2024 19:06, Andrew Cooper wrote: >> Whether to return information about a xen-owned evtchn is a matter of policy, >> and it's not acceptable to short circuit the XSM on the matter. > > I can certainly accept this as one possible view point. As in so many cases > I'm afraid I dislike you putting it as if it was the only possible one. Further to this: Is there even a way to express the same denial in XSM? alloc_unbound_xen_event_channel() doesn't specifically "mark" such a channel, and (yes, it could in principle be open-coded in Flask code) consumer_is_xen() is private to event_channel.c. I also dare to question whether in SILO mode status information like this should be available. Jan
On 4/3/24 02:52, Jan Beulich wrote: > On 03.04.2024 08:16, Jan Beulich wrote: >> On 02.04.2024 19:06, Andrew Cooper wrote: >>> Whether to return information about a xen-owned evtchn is a matter of policy, >>> and it's not acceptable to short circuit the XSM on the matter. >> >> I can certainly accept this as one possible view point. As in so many cases >> I'm afraid I dislike you putting it as if it was the only possible one. > > Further to this: Is there even a way to express the same denial in XSM? > alloc_unbound_xen_event_channel() doesn't specifically "mark" such a > channel, and (yes, it could in principle be open-coded in Flask code) > consumer_is_xen() is private to event_channel.c. I also dare to question > whether in SILO mode status information like this should be available. To build on the previous response: if the natural failure return value is -EACCESS in response to a domain resource access attempt, then the probability is extremely high that it should be implemented under a XSM hook and not hard-coded into the resource logic. v/r, dps
On 03.04.2024 13:50, Daniel P. Smith wrote: > On 4/3/24 02:52, Jan Beulich wrote: >> On 03.04.2024 08:16, Jan Beulich wrote: >>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>> Whether to return information about a xen-owned evtchn is a matter of policy, >>>> and it's not acceptable to short circuit the XSM on the matter. >>> >>> I can certainly accept this as one possible view point. As in so many cases >>> I'm afraid I dislike you putting it as if it was the only possible one. >> >> Further to this: Is there even a way to express the same denial in XSM? >> alloc_unbound_xen_event_channel() doesn't specifically "mark" such a >> channel, and (yes, it could in principle be open-coded in Flask code) >> consumer_is_xen() is private to event_channel.c. I also dare to question >> whether in SILO mode status information like this should be available. > > To build on the previous response: if the natural failure return value > is -EACCESS in response to a domain resource access attempt, then the > probability is extremely high that it should be implemented under a XSM > hook and not hard-coded into the resource logic. Possibly. But first of all - could you answer the earlier question I raised? Jan
On 4/3/24 07:54, Jan Beulich wrote: > On 03.04.2024 13:50, Daniel P. Smith wrote: >> On 4/3/24 02:52, Jan Beulich wrote: >>> On 03.04.2024 08:16, Jan Beulich wrote: >>>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>>> Whether to return information about a xen-owned evtchn is a matter of policy, >>>>> and it's not acceptable to short circuit the XSM on the matter. >>>> >>>> I can certainly accept this as one possible view point. As in so many cases >>>> I'm afraid I dislike you putting it as if it was the only possible one. >>> >>> Further to this: Is there even a way to express the same denial in XSM? >>> alloc_unbound_xen_event_channel() doesn't specifically "mark" such a >>> channel, and (yes, it could in principle be open-coded in Flask code) >>> consumer_is_xen() is private to event_channel.c. I also dare to question >>> whether in SILO mode status information like this should be available. >> >> To build on the previous response: if the natural failure return value >> is -EACCESS in response to a domain resource access attempt, then the >> probability is extremely high that it should be implemented under a XSM >> hook and not hard-coded into the resource logic. > > Possibly. But first of all - could you answer the earlier question I raised? Don't need to, this change subverts/violates the access control framework. If the desire is to make this access decision for the default/dummy policy, then codify it there. Otherwise I will be ack'ing this change since it is access control and falls under the purview of XSM. v/r, dps
On 03.04.2024 15:31, Daniel P. Smith wrote: > On 4/3/24 07:54, Jan Beulich wrote: >> On 03.04.2024 13:50, Daniel P. Smith wrote: >>> On 4/3/24 02:52, Jan Beulich wrote: >>>> On 03.04.2024 08:16, Jan Beulich wrote: >>>>> On 02.04.2024 19:06, Andrew Cooper wrote: >>>>>> Whether to return information about a xen-owned evtchn is a matter of policy, >>>>>> and it's not acceptable to short circuit the XSM on the matter. >>>>> >>>>> I can certainly accept this as one possible view point. As in so many cases >>>>> I'm afraid I dislike you putting it as if it was the only possible one. >>>> >>>> Further to this: Is there even a way to express the same denial in XSM? >>>> alloc_unbound_xen_event_channel() doesn't specifically "mark" such a >>>> channel, and (yes, it could in principle be open-coded in Flask code) >>>> consumer_is_xen() is private to event_channel.c. I also dare to question >>>> whether in SILO mode status information like this should be available. >>> >>> To build on the previous response: if the natural failure return value >>> is -EACCESS in response to a domain resource access attempt, then the >>> probability is extremely high that it should be implemented under a XSM >>> hook and not hard-coded into the resource logic. >> >> Possibly. But first of all - could you answer the earlier question I raised? > > Don't need to, this change subverts/violates the access control > framework. If the desire is to make this access decision for the > default/dummy policy, then codify it there. Otherwise I will be ack'ing > this change since it is access control and falls under the purview of XSM. If Xen internals like this are to be exposable (and controlled by XSM), why would other Xen internals not similarly be (optionally) exposed? Further, since above you referred to EACCES being what XSM is supposed to control: xsm_default_action() used EPERM, and (presumably; too long ago) EACCES was chosen here precisely to make it not look like an XSM surrogate. Jan
© 2016 - 2024 Red Hat, Inc.