| Message ID | 20191030131633.9356-1-sds@tycho.nsa.gov (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [RFC] selinux: add SELinux hooks for lockdown integrity and confidentiality | expand |
On 10/30/19 9:16 AM, Stephen Smalley wrote: > Add SELinux access control hooks for lockdown integrity and > confidentiality. This effectively mimics the current implementation of > lockdown (caveat noted below). If lockdown is enabled alongside SELinux, > then the lockdown access control will take precedence over the SELinux > lockdown implementation. > > Note that this SELinux implementation allows the integrity and > confidentiality reasons to be controlled independently from one another. > Thus, in an SELinux policy, one could allow integrity operations while > blocking confidentiality operations. NB This is intended to be the first of a series that will ultimately lead to finer-grained controls than just integrity and confidentiality, but wanted to get some feedback on it at this stage. Also anticipate greater controversy over exposing finer granularity since the lockdown reasons are free to change at any time, so this would be the baseline fallback position if finer grained controls are rejected. > > (original patch authored by an intern who wishes to remain anonymous; > I am signing off on his behalf) > > Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov> > --- > security/selinux/hooks.c | 22 ++++++++++++++++++++++ > security/selinux/include/classmap.h | 2 ++ > 2 files changed, 24 insertions(+) > > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c > index 36e531b91df2..6722c6b4ae74 100644 > --- a/security/selinux/hooks.c > +++ b/security/selinux/hooks.c > @@ -91,6 +91,7 @@ > #include <uapi/linux/mount.h> > #include <linux/fsnotify.h> > #include <linux/fanotify.h> > +#include <linux/security.h> > > #include "avc.h" > #include "objsec.h" > @@ -6799,6 +6800,25 @@ static void selinux_bpf_prog_free(struct bpf_prog_aux *aux) > } > #endif > > +static int selinux_lockdown(enum lockdown_reason what) > +{ > + u32 sid = current_sid(); > + > + if (what <= LOCKDOWN_INTEGRITY_MAX) > + return avc_has_perm(&selinux_state, > + sid, sid, > + SECCLASS_LOCKDOWN, LOCKDOWN__INTEGRITY, NULL); > + else if (what <= LOCKDOWN_CONFIDENTIALITY_MAX) > + return avc_has_perm(&selinux_state, > + sid, sid, > + SECCLASS_LOCKDOWN, LOCKDOWN__CONFIDENTIALITY, > + NULL); > + > + /* invalid reason */ > + pr_warn("SELinux: invalid lockdown reason\n"); > + return -EPERM; > +} > + > struct lsm_blob_sizes selinux_blob_sizes __lsm_ro_after_init = { > .lbs_cred = sizeof(struct task_security_struct), > .lbs_file = sizeof(struct file_security_struct), > @@ -7042,6 +7062,8 @@ static struct security_hook_list selinux_hooks[] __lsm_ro_after_init = { > LSM_HOOK_INIT(bpf_map_free_security, selinux_bpf_map_free), > LSM_HOOK_INIT(bpf_prog_free_security, selinux_bpf_prog_free), > #endif > + > + LSM_HOOK_INIT(locked_down, selinux_lockdown), > }; > > static __init int selinux_init(void) > diff --git a/security/selinux/include/classmap.h b/security/selinux/include/classmap.h > index 32e9b03be3dd..594c32febcd8 100644 > --- a/security/selinux/include/classmap.h > +++ b/security/selinux/include/classmap.h > @@ -244,6 +244,8 @@ struct security_class_mapping secclass_map[] = { > {"map_create", "map_read", "map_write", "prog_load", "prog_run"} }, > { "xdp_socket", > { COMMON_SOCK_PERMS, NULL } }, > + { "lockdown", > + { "integrity", "confidentiality", NULL } }, > { NULL } > }; > >
On Wed, Oct 30, 2019 at 11:29 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: > On 10/30/19 9:16 AM, Stephen Smalley wrote: > > Add SELinux access control hooks for lockdown integrity and > > confidentiality. This effectively mimics the current implementation of > > lockdown (caveat noted below). If lockdown is enabled alongside SELinux, > > then the lockdown access control will take precedence over the SELinux > > lockdown implementation. > > > > Note that this SELinux implementation allows the integrity and > > confidentiality reasons to be controlled independently from one another. > > Thus, in an SELinux policy, one could allow integrity operations while > > blocking confidentiality operations. > > NB This is intended to be the first of a series that will ultimately > lead to finer-grained controls than just integrity and confidentiality, > but wanted to get some feedback on it at this stage. Also anticipate > greater controversy over exposing finer granularity since the lockdown > reasons are free to change at any time, so this would be the baseline > fallback position if finer grained controls are rejected. > > > (original patch authored by an intern who wishes to remain anonymous; > > I am signing off on his behalf) I'm not going to comment on this on-list because IANAL, but it might be best to leave the comment above off of future postings as I think it somewhat confuses the principle behind the sign-off line. I understand you want to give credit where it is due, but without an explicit author's name/email I believe it is safer for you to assume that role. Put another way, the comment above makes me nervous about adding my own sign-off and merging it into the SELinux tree. > > Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov> > > --- > > security/selinux/hooks.c | 22 ++++++++++++++++++++++ > > security/selinux/include/classmap.h | 2 ++ > > 2 files changed, 24 insertions(+)
On Wed, Oct 30, 2019 at 9:16 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: > Add SELinux access control hooks for lockdown integrity and > confidentiality. This effectively mimics the current implementation of > lockdown (caveat noted below). If lockdown is enabled alongside SELinux, > then the lockdown access control will take precedence over the SELinux > lockdown implementation. > > Note that this SELinux implementation allows the integrity and > confidentiality reasons to be controlled independently from one another. > Thus, in an SELinux policy, one could allow integrity operations while > blocking confidentiality operations. > > (original patch authored by an intern who wishes to remain anonymous; > I am signing off on his behalf) > > Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov> > --- > security/selinux/hooks.c | 22 ++++++++++++++++++++++ > security/selinux/include/classmap.h | 2 ++ > 2 files changed, 24 insertions(+) > > diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c > index 36e531b91df2..6722c6b4ae74 100644 > --- a/security/selinux/hooks.c > +++ b/security/selinux/hooks.c > @@ -91,6 +91,7 @@ > #include <uapi/linux/mount.h> > #include <linux/fsnotify.h> > #include <linux/fanotify.h> > +#include <linux/security.h> > > #include "avc.h" > #include "objsec.h" > @@ -6799,6 +6800,25 @@ static void selinux_bpf_prog_free(struct bpf_prog_aux *aux) > } > #endif > > +static int selinux_lockdown(enum lockdown_reason what) > +{ > + u32 sid = current_sid(); > + > + if (what <= LOCKDOWN_INTEGRITY_MAX) > + return avc_has_perm(&selinux_state, > + sid, sid, > + SECCLASS_LOCKDOWN, LOCKDOWN__INTEGRITY, NULL); > + else if (what <= LOCKDOWN_CONFIDENTIALITY_MAX) > + return avc_has_perm(&selinux_state, > + sid, sid, > + SECCLASS_LOCKDOWN, LOCKDOWN__CONFIDENTIALITY, > + NULL); > + > + /* invalid reason */ > + pr_warn("SELinux: invalid lockdown reason\n"); > + return -EPERM; > +} I don't have any objections to adding a hook to control access to the lockdown functionality (I think it's a good idea), but I am a little nervous about the granularity of the control. Sticking with just an integrity and a confidentiality permission seems okay, but I worry about adding additional permissions until we have a better idea of how the lockdown functionality is adopted by developers and we see how the lockdown_reason evolves.
On 10/31/19 5:59 AM, Paul Moore wrote: > On Wed, Oct 30, 2019 at 9:16 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: >> Add SELinux access control hooks for lockdown integrity and >> confidentiality. This effectively mimics the current implementation of >> lockdown (caveat noted below). If lockdown is enabled alongside SELinux, >> then the lockdown access control will take precedence over the SELinux >> lockdown implementation. >> >> Note that this SELinux implementation allows the integrity and >> confidentiality reasons to be controlled independently from one another. >> Thus, in an SELinux policy, one could allow integrity operations while >> blocking confidentiality operations. >> >> (original patch authored by an intern who wishes to remain anonymous; >> I am signing off on his behalf) >> >> Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov> >> --- >> security/selinux/hooks.c | 22 ++++++++++++++++++++++ >> security/selinux/include/classmap.h | 2 ++ >> 2 files changed, 24 insertions(+) >> >> diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c >> index 36e531b91df2..6722c6b4ae74 100644 >> --- a/security/selinux/hooks.c >> +++ b/security/selinux/hooks.c >> @@ -91,6 +91,7 @@ >> #include <uapi/linux/mount.h> >> #include <linux/fsnotify.h> >> #include <linux/fanotify.h> >> +#include <linux/security.h> >> >> #include "avc.h" >> #include "objsec.h" >> @@ -6799,6 +6800,25 @@ static void selinux_bpf_prog_free(struct bpf_prog_aux *aux) >> } >> #endif >> >> +static int selinux_lockdown(enum lockdown_reason what) >> +{ >> + u32 sid = current_sid(); >> + >> + if (what <= LOCKDOWN_INTEGRITY_MAX) >> + return avc_has_perm(&selinux_state, >> + sid, sid, >> + SECCLASS_LOCKDOWN, LOCKDOWN__INTEGRITY, NULL); >> + else if (what <= LOCKDOWN_CONFIDENTIALITY_MAX) >> + return avc_has_perm(&selinux_state, >> + sid, sid, >> + SECCLASS_LOCKDOWN, LOCKDOWN__CONFIDENTIALITY, >> + NULL); >> + >> + /* invalid reason */ >> + pr_warn("SELinux: invalid lockdown reason\n"); >> + return -EPERM; >> +} > > I don't have any objections to adding a hook to control access to the > lockdown functionality (I think it's a good idea), but I am a little > nervous about the granularity of the control. Sticking with just an > integrity and a confidentiality permission seems okay, but I worry > about adding additional permissions until we have a better idea of how > the lockdown functionality is adopted by developers and we see how the > lockdown_reason evolves. Ok, so let's discuss what if anything else is needed for a final non-RFC version of this patch. One thing that I wondered about was whether we ought to include the reason information in the audit record as supplemental data via a new LSM_AUDIT_DATA_* type for help in policy debugging / development. However, the lockdown_reasons[] string array is presently private to the lockdown module so we would have to export that or replicate it for creating a string representation of the reason if we were to do so. That would expose the reasons in terms of audit data but not as a basis for the permission check. Note that the lockdown module logs these reason strings via pr_notice() when it denies access, so it appears that exposing the strings as part of audit data would not introduce any extra kernel stable ABI guarantees? I also wasn't sure about the pr_warn() above. If we reach it, it is effectively a kernel bug. We could mirror what the lockdown module does in lockdown_is_locked_down(), i.e. use WARN() instead. Of course, the SELinux hook won't even be reached in this case if the lockdown module is enabled, but the lockdown module could be disabled so I guess we need to check it too. If we take the lockdown class with just integrity and confidentiality permissions now and later introduce finer granularity, we'll presumably need a policy capability to select whether the coarse-grained or fine-grained permissions are used.
On Thu, Oct 31, 2019 at 10:01 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: > On 10/31/19 5:59 AM, Paul Moore wrote: > > On Wed, Oct 30, 2019 at 9:16 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: > >> Add SELinux access control hooks for lockdown integrity and > >> confidentiality. This effectively mimics the current implementation of > >> lockdown (caveat noted below). If lockdown is enabled alongside SELinux, > >> then the lockdown access control will take precedence over the SELinux > >> lockdown implementation. > >> > >> Note that this SELinux implementation allows the integrity and > >> confidentiality reasons to be controlled independently from one another. > >> Thus, in an SELinux policy, one could allow integrity operations while > >> blocking confidentiality operations. ... > > I don't have any objections to adding a hook to control access to the > > lockdown functionality (I think it's a good idea), but I am a little > > nervous about the granularity of the control. Sticking with just an > > integrity and a confidentiality permission seems okay, but I worry > > about adding additional permissions until we have a better idea of how > > the lockdown functionality is adopted by developers and we see how the > > lockdown_reason evolves. > > Ok, so let's discuss what if anything else is needed for a final non-RFC > version of this patch. > > One thing that I wondered about was whether we ought to include the > reason information in the audit record as supplemental data via a new > LSM_AUDIT_DATA_* type for help in policy debugging / development. > However, the lockdown_reasons[] string array is presently private to the > lockdown module so we would have to export that or replicate it for > creating a string representation of the reason if we were to do so. > That would expose the reasons in terms of audit data but not as a basis > for the permission check. Note that the lockdown module logs these > reason strings via pr_notice() when it denies access, so it appears that > exposing the strings as part of audit data would not introduce any extra > kernel stable ABI guarantees? That is an interesting question: do we consider dmesg output to be part of the stable kernel API? My hunch would be "no", as I've seen things change quite a bit there over the years, but IANL (I Am Not Linus). However, that said, logging a reason string via audit seems like a good idea (especially since there is presently a many-to-one mapping between reasons and the SELinux permission). Further, while the audit field name is part of the kernel API, the value is much more open. > I also wasn't sure about the pr_warn() above. If we reach it, it is > effectively a kernel bug. We could mirror what the lockdown module does > in lockdown_is_locked_down(), i.e. use WARN() instead. Of course, the > SELinux hook won't even be reached in this case if the lockdown module > is enabled, but the lockdown module could be disabled so I guess we need > to check it too. Since this seems security relevant, I wonder if we should be using SELINUX_ERR? > If we take the lockdown class with just integrity and confidentiality > permissions now and later introduce finer granularity, we'll presumably > need a policy capability to select whether the coarse-grained or > fine-grained permissions are used. True, although I'm not overly worried about the need to use policy capabilities; I'm more worried about jumping into too fine a granularity before we see how this will be used.
On 11/7/19 12:48 PM, Paul Moore wrote: > On Thu, Oct 31, 2019 at 10:01 AM Stephen Smalley <sds@tycho.nsa.gov> wrote: > That is an interesting question: do we consider dmesg output to be > part of the stable kernel API? My hunch would be "no", as I've seen > things change quite a bit there over the years, but IANL (I Am Not > Linus). However, that said, logging a reason string via audit seems > like a good idea (especially since there is presently a many-to-one > mapping between reasons and the SELinux permission). Further, while > the audit field name is part of the kernel API, the value is much more > open. Ok, any preferences on the audit field name or should we just create one and cc linux-audit on the next RFC? lockdown_reason=? >> I also wasn't sure about the pr_warn() above. If we reach it, it is >> effectively a kernel bug. We could mirror what the lockdown module does >> in lockdown_is_locked_down(), i.e. use WARN() instead. Of course, the >> SELinux hook won't even be reached in this case if the lockdown module >> is enabled, but the lockdown module could be disabled so I guess we need >> to check it too. > > Since this seems security relevant, I wonder if we should be using SELINUX_ERR? The benefit of a WARN() is that it will give us a stack trace showing the offending caller in the kernel, which would be useful since it would be a buggy caller passing an invalid lockdown reason (LOCKDOWN_NONE or >= LOCKDOWN_CONFIDENTIALITY_MAX). pr_warn() or audit_log() won't give us that info. We could do both of course.
On Thu, Nov 7, 2019 at 1:07 PM Stephen Smalley <sds@tycho.nsa.gov> wrote: > On 11/7/19 12:48 PM, Paul Moore wrote: > > On Thu, Oct 31, 2019 at 10:01 AM Stephen Smalley <sds@tycho.nsa.gov> > wrote: > > That is an interesting question: do we consider dmesg output to be > > part of the stable kernel API? My hunch would be "no", as I've seen > > things change quite a bit there over the years, but IANL (I Am Not > > Linus). However, that said, logging a reason string via audit seems > > like a good idea (especially since there is presently a many-to-one > > mapping between reasons and the SELinux permission). Further, while > > the audit field name is part of the kernel API, the value is much more > > open. > > Ok, any preferences on the audit field name or should we just create one > and cc linux-audit on the next RFC? lockdown_reason=? Definitely CC linux-audit as I expect Steve will want to have his say. FWIW, "lockdown_reason" seems reasonable to me. > >> I also wasn't sure about the pr_warn() above. If we reach it, it is > >> effectively a kernel bug. We could mirror what the lockdown module does > >> in lockdown_is_locked_down(), i.e. use WARN() instead. Of course, the > >> SELinux hook won't even be reached in this case if the lockdown module > >> is enabled, but the lockdown module could be disabled so I guess we need > >> to check it too. > > > > Since this seems security relevant, I wonder if we should be using SELINUX_ERR? > > The benefit of a WARN() is that it will give us a stack trace showing > the offending caller in the kernel, which would be useful since it would > be a buggy caller passing an invalid lockdown reason (LOCKDOWN_NONE or > >= LOCKDOWN_CONFIDENTIALITY_MAX). pr_warn() or audit_log() won't give > us that info. We could do both of course. It's a balance between development needs and freaking out administrators (although perhaps rightly so). I also worry a bit that WARN can be disabled at build time so having something like SELINUX_ERR could be a good fallback, assuming we did both.
diff --git a/security/selinux/hooks.c b/security/selinux/hooks.c index 36e531b91df2..6722c6b4ae74 100644 --- a/security/selinux/hooks.c +++ b/security/selinux/hooks.c @@ -91,6 +91,7 @@ #include <uapi/linux/mount.h> #include <linux/fsnotify.h> #include <linux/fanotify.h> +#include <linux/security.h> #include "avc.h" #include "objsec.h" @@ -6799,6 +6800,25 @@ static void selinux_bpf_prog_free(struct bpf_prog_aux *aux) } #endif +static int selinux_lockdown(enum lockdown_reason what) +{ + u32 sid = current_sid(); + + if (what <= LOCKDOWN_INTEGRITY_MAX) + return avc_has_perm(&selinux_state, + sid, sid, + SECCLASS_LOCKDOWN, LOCKDOWN__INTEGRITY, NULL); + else if (what <= LOCKDOWN_CONFIDENTIALITY_MAX) + return avc_has_perm(&selinux_state, + sid, sid, + SECCLASS_LOCKDOWN, LOCKDOWN__CONFIDENTIALITY, + NULL); + + /* invalid reason */ + pr_warn("SELinux: invalid lockdown reason\n"); + return -EPERM; +} + struct lsm_blob_sizes selinux_blob_sizes __lsm_ro_after_init = { .lbs_cred = sizeof(struct task_security_struct), .lbs_file = sizeof(struct file_security_struct), @@ -7042,6 +7062,8 @@ static struct security_hook_list selinux_hooks[] __lsm_ro_after_init = { LSM_HOOK_INIT(bpf_map_free_security, selinux_bpf_map_free), LSM_HOOK_INIT(bpf_prog_free_security, selinux_bpf_prog_free), #endif + + LSM_HOOK_INIT(locked_down, selinux_lockdown), }; static __init int selinux_init(void) diff --git a/security/selinux/include/classmap.h b/security/selinux/include/classmap.h index 32e9b03be3dd..594c32febcd8 100644 --- a/security/selinux/include/classmap.h +++ b/security/selinux/include/classmap.h @@ -244,6 +244,8 @@ struct security_class_mapping secclass_map[] = { {"map_create", "map_read", "map_write", "prog_load", "prog_run"} }, { "xdp_socket", { COMMON_SOCK_PERMS, NULL } }, + { "lockdown", + { "integrity", "confidentiality", NULL } }, { NULL } };
Add SELinux access control hooks for lockdown integrity and confidentiality. This effectively mimics the current implementation of lockdown (caveat noted below). If lockdown is enabled alongside SELinux, then the lockdown access control will take precedence over the SELinux lockdown implementation. Note that this SELinux implementation allows the integrity and confidentiality reasons to be controlled independently from one another. Thus, in an SELinux policy, one could allow integrity operations while blocking confidentiality operations. (original patch authored by an intern who wishes to remain anonymous; I am signing off on his behalf) Signed-off-by: Stephen Smalley <sds@tycho.nsa.gov> --- security/selinux/hooks.c | 22 ++++++++++++++++++++++ security/selinux/include/classmap.h | 2 ++ 2 files changed, 24 insertions(+)