selinux: allow kernfs symlinks to inherit parent directory context
PATCH 1/1 selinux: allow kernfs symlinks to inherit parent directory context
Currently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent fs root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability: kernfs_sovereign_symlinks
--- security/selinux/hooks.c | 5 ++++- security/selinux/include/security.h | 8 ++++++++ security/selinux/ss/services.c | 3 ++- 3 files changed, 14 insertions(+), 2 deletions(-)
security/selinux/hooks.c +4 -1
@@ -1475,7 +1475,10 @@ static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent/* Default to the fs superblock SID. */ sid = sbsec->sid; -if ((sbsec->flags &SE_SBGENFS ) && !S_ISLNK (inode->i_mode)) { +if (((sbsec->flags &SE_SBGENFS ) && + (!S_ISLNK (inode->i_mode))) || + (selinux_policycap_kernfs_sovereign_symlinks () && + (sbsec->flags &SE_SBGENFS_XATTR ))) {/* We must have a dentry to determine the label on * procfs inodes */ if (opt_dentry) {
security/selinux/include/security.h +8 -0
@@ -79,6 +79,7 @@ enum {POLICYDB_CAPABILITY_ALWAYSNETWORK ,POLICYDB_CAPABILITY_CGROUPSECLABEL ,POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION , +POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS , __POLICYDB_CAPABILITY_MAX };#define POLICYDB_CAPABILITY_MAX (__POLICYDB_CAPABILITY_MAX - 1)
@@ -209,6 +210,13 @@ static inline bool selinux_policycap_nnp_nosuid_transition(void)return state->policycap[POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION ]; } +static inline bool selinux_policycap_kernfs_sovereign_symlinks (void ) +{ +struct selinux_state *state = &selinux_state; + +return state->policycap[POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS ]; +} +int security_mls_enabled (struct selinux_state *state);int security_load_policy (struct selinux_state *state,void *data,size_t len);
security/selinux/ss/services.c +2 -1
@@ -73,7 +73,8 @@ const char *selinux_policycap_names[__POLICYDB_CAPABILITY_MAX] = {"extended_socket_class" ,"always_check_network" ,"cgroup_seclabel" , - "nnp_nosuid_transition" +"nnp_nosuid_transition" , + "kernfs_sovereign_symlinks" };static struct selinux_ss selinux_ss;
Signature · 1 line
2.25.0
Quoted 22 lines · Christian Göttsche
Currently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent fs root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability:kernfs_sovereign_symlinksSigned-off-by Christian Göttsche <[email protected]>--- security/selinux/hooks.c | 5 ++++- security/selinux/include/security.h | 8 ++++++++ security/selinux/ss/services.c | 3 ++- 3 files changed, 14 insertions(+), 2 deletions(-)security/selinux/hooks.c +0
-0
security/selinux/hooks.c static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent · line 1480
Quoted 7 lines
security/selinux/hooks.c +3
-1@@ -1475,7 +1475,10 @@ static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent/* Default to the fs superblock SID. */ sid = sbsec->sid; -if ((sbsec->flags &SE_SBGENFS ) && !S_ISLNK (inode->i_mode)) { + if (((sbsec->flags & SE_SBGENFS) && + (!S_ISLNK(inode->i_mode))) || + (selinux_policycap_kernfs_sovereign_symlinks() &&
Not fond of the name. 1) kernfs is a kernel implementation detail, shouldn't be exposed to policy; genfs is the policy construct 2) sovereign doesn't seem to fit the meaning of this capability; seclabel would be more appropriate.
security/selinux/hooks.c static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent · line 1481
Quoted 1 line
security/selinux/hooks.c +1
-0+ (sbsec->flags & SE_SBGENFS_XATTR))) {
Why limit this to SE_SBGENFS_XATTR filesystems? Why not just make the test:
if ((sbsec->flags & SE_SBGENFS) && (!S_ISLNK(inode->i_mode) ||
selinux_policycap_genfs_symlinkseclabel()))
or similar.
Quoted 44 lines
security/selinux/hooks.c +0
-0/* We must have a dentry to determine the label on * procfs inodes */ if (opt_dentry) {security/selinux/include/security.h +8
-0@@ -79,6 +79,7 @@ enum {POLICYDB_CAPABILITY_ALWAYSNETWORK ,POLICYDB_CAPABILITY_CGROUPSECLABEL ,POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION , +POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS , __POLICYDB_CAPABILITY_MAX };#define POLICYDB_CAPABILITY_MAX (__POLICYDB_CAPABILITY_MAX - 1)@@ -209,6 +210,13 @@ static inline bool selinux_policycap_nnp_nosuid_transition(void)return state->policycap[POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION ]; } +static inline bool selinux_policycap_kernfs_sovereign_symlinks (void ) +{ +struct selinux_state *state = &selinux_state; + +return state->policycap[POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS ]; +} +int security_mls_enabled (struct selinux_state *state);int security_load_policy (struct selinux_state *state,void *data,size_t len);security/selinux/ss/services.c +2
-1@@ -73,7 +73,8 @@ const char *selinux_policycap_names[__POLICYDB_CAPABILITY_MAX] = {"extended_socket_class" ,"always_check_network" ,"cgroup_seclabel" , - "nnp_nosuid_transition" +"nnp_nosuid_transition" , + "kernfs_sovereign_symlinks" };static struct selinux_ss selinux_ss;
Am Fr., 24. Jan. 2020 um 19:53 Uhr schrieb Stephen Smalley <[email protected]>:
Quoted 27 lines · Christian Göttsche
Quoted 25 lines
Currently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent fs root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability:kernfs_sovereign_symlinksSigned-off-by Christian Göttsche <[email protected]>--- security/selinux/hooks.c | 5 ++++- security/selinux/include/security.h | 8 ++++++++ security/selinux/ss/services.c | 3 ++- 3 files changed, 14 insertions(+), 2 deletions(-)security/selinux/hooks.c +0
-0@@ -1475,7 +1475,10 @@ static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent/* Default to the fs superblock SID. */ sid = sbsec->sid;
Quoted 9 lines
security/selinux/hooks.c static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent · line 1480
Quoted 4 lines
security/selinux/hooks.c +3
-1-if ((sbsec->flags &SE_SBGENFS ) && !S_ISLNK (inode->i_mode)) { + if (((sbsec->flags & SE_SBGENFS) && + (!S_ISLNK(inode->i_mode))) || + (selinux_policycap_kernfs_sovereign_symlinks() &&Not fond of the name. 1) kernfs is a kernel implementation detail, shouldn't be exposed to policy; genfs is the policy construct 2) sovereign doesn't seem to fit the meaning of this capability; seclabel would be more appropriate.
Something like genfs_seclabel_symlinks?
Quoted 6 lines
security/selinux/hooks.c static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent · line 1481
Quoted 1 line
security/selinux/hooks.c +1
-0+ (sbsec->flags & SE_SBGENFS_XATTR))) {Why limit this to
SE_SBGENFS_XATTRfilesystems? Why not just make the test:if ((sbsec->flags & SE_SBGENFS) && (!S_ISLNK(inode->i_mode) ||
selinux_policycap_genfs_symlinkseclabel()))
or similar.
I somehow thought that this functionality is limited to filesystems with SE_SBGENFS_XATTR;
so I can expand the check to SE_SBGENFS.
Quoted 45 lines
Quoted 44 lines
security/selinux/hooks.c +0
-0/* We must have a dentry to determine the label on * procfs inodes */ if (opt_dentry) {security/selinux/include/security.h +8
-0@@ -79,6 +79,7 @@ enum {POLICYDB_CAPABILITY_ALWAYSNETWORK ,POLICYDB_CAPABILITY_CGROUPSECLABEL ,POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION , +POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS , __POLICYDB_CAPABILITY_MAX };#define POLICYDB_CAPABILITY_MAX (__POLICYDB_CAPABILITY_MAX - 1)@@ -209,6 +210,13 @@ static inline bool selinux_policycap_nnp_nosuid_transition(void)return state->policycap[POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION ]; } +static inline bool selinux_policycap_kernfs_sovereign_symlinks (void ) +{ +struct selinux_state *state = &selinux_state; + +return state->policycap[POLICYDB_CAPABILITY_KERNFS_SOVEREIGN_SYMLINKS ]; +} +int security_mls_enabled (struct selinux_state *state);int security_load_policy (struct selinux_state *state,void *data,size_t len);security/selinux/ss/services.c +2
-1@@ -73,7 +73,8 @@ const char *selinux_policycap_names[__POLICYDB_CAPABILITY_MAX] = {"extended_socket_class" ,"always_check_network" ,"cgroup_seclabel" , - "nnp_nosuid_transition" +"nnp_nosuid_transition" , + "kernfs_sovereign_symlinks" };static struct selinux_ss selinux_ss;
Quoted 16 lines · Christian Göttsche
Am Fr., 24. Jan. 2020 um 19:53 Uhr schrieb Stephen Smalley <[email protected]>:
Quoted 14 lines · Christian Göttsche
Quoted 12 lines
Currently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent fs root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability:kernfs_sovereign_symlinksSigned-off-by Christian Göttsche <[email protected]>---
Quoted 6 lines
Quoted 4 lines
Not fond of the name. 1) kernfs is a kernel implementation detail, shouldn't be exposed to policy; genfs is the policy construct 2) sovereign doesn't seem to fit the meaning of this capability; seclabel would be more appropriate.
Something like
genfs_seclabel_symlinks?
Works for me.
Quoted 11 lines
Quoted 7 lines
Quoted 2 lines
+ (
sbsec->flags&SE_SBGENFS_XATTR))) {Why limit this to
SE_SBGENFS_XATTRfilesystems? Why not just make the test:if ((sbsec->flags & SE_SBGENFS) && (!S_ISLNK(inode->i_mode) ||
selinux_policycap_genfs_symlinkseclabel()))
or similar.I somehow thought that this functionality is limited to filesystems with
SE_SBGENFS_XATTR;
so I can expand the check toSE_SBGENFS.
I could be wrong but I don't see why it would need to be limited in that way.
PATCH v2 1/1 selinux: allow kernfs symlinks to inherit parent directory context
In reply to Christian GöttscheCurrently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent filesystem root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability: genfs_seclabel_symlinks
---
v2:
incorporate feedback from Stephen Smalley
- changed polcap name
- extended affected filesystems from SE_SBGENFS_XATTR to SE_SBGENFS
security/selinux/hooks.c | 4 +++- security/selinux/include/security.h | 8 ++++++++ security/selinux/ss/services.c | 3 ++- 3 files changed, 13 insertions(+), 2 deletions(-)
security/selinux/hooks.c +3 -1
@@ -1475,7 +1475,9 @@ static int inode_doinit_with_dentry(struct inode *inode, struct dentry *opt_dent/* Default to the fs superblock SID. */ sid = sbsec->sid; -if ((sbsec->flags &SE_SBGENFS ) && !S_ISLNK (inode->i_mode)) { +if ((sbsec->flags &SE_SBGENFS ) && + (!S_ISLNK (inode->i_mode) || +selinux_policycap_genfs_seclabel_symlinks ())) {/* We must have a dentry to determine the label on * procfs inodes */ if (opt_dentry) {
security/selinux/include/security.h +8 -0
@@ -79,6 +79,7 @@ enum {POLICYDB_CAPABILITY_ALWAYSNETWORK ,POLICYDB_CAPABILITY_CGROUPSECLABEL ,POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION , +POLICYDB_CAPABILITY_GENFS_SECLABEL_SYMLINKS , __POLICYDB_CAPABILITY_MAX };#define POLICYDB_CAPABILITY_MAX (__POLICYDB_CAPABILITY_MAX - 1)
@@ -209,6 +210,13 @@ static inline bool selinux_policycap_nnp_nosuid_transition(void)return state->policycap[POLICYDB_CAPABILITY_NNP_NOSUID_TRANSITION ]; } +static inline bool selinux_policycap_genfs_seclabel_symlinks (void ) +{ +struct selinux_state *state = &selinux_state; + +return state->policycap[POLICYDB_CAPABILITY_GENFS_SECLABEL_SYMLINKS ]; +} +int security_mls_enabled (struct selinux_state *state);int security_load_policy (struct selinux_state *state,void *data,size_t len);
security/selinux/ss/services.c +2 -1
@@ -73,7 +73,8 @@ const char *selinux_policycap_names[__POLICYDB_CAPABILITY_MAX] = {"extended_socket_class" ,"always_check_network" ,"cgroup_seclabel" , - "nnp_nosuid_transition" +"nnp_nosuid_transition" , + "genfs_seclabel_symlinks" };static struct selinux_ss selinux_ss;
Signature · 1 line
2.25.0
Quoted 11 lines · Christian Göttsche
Currently symlinks on kernel filesystems, like sysfs, are labeled on creation with the parent filesystem root sid.
Allow symlinks to inherit the parent directory context, so fine-grained kernfs labeling can be applied to symlinks too and checking contexts doesn't complain about them.
For backward-compatibility this behavior is contained in a new policy
capability:genfs_seclabel_symlinksSigned-off-by Christian Göttsche <[email protected]>
(pruned cc list to omit non-selinux kernel maintainers since this change is entirely self-contained to selinux; I'm guessing you blindly took the results of scripts/get_maintainer.pl which tends to over-approximate - I only use it as a guide/hint and rarely use the full list it provides.)
This looks fine to me code-wise. Have you tried enabling this new policy capability in policy and seeing the effects of it? I remember a problem in the way-back time that motivated adding the S_ISLNK() exception for proc. IIRC, the issue was that policies specified "genfscon proc /net system_u:object_r:proc_net_t:s0" to label everything under /proc/net with proc_net_t by default, and when /proc/net was changed to be a symlink to /proc/self/net as part of the network namespaces work, this caused the symlink to be labeled proc_net_t, but since previously there were no symlinks labeled proc_net_t, no processes were allowed to read/follow the symlink and existing systems broke. Exempting symlinks caused the /proc/net symlink to be labeled proc_t instead, which was already widely allowed. To avoid this problem when enabling this capability, you will either need to allow proc_net_t:lnk_file read widely or you will need to change the genfscon statement to avoid labeling the symlink (there is an optional file mode field in genfscon statements similar to that of file_contexts, e.g.
genfscon proc /net -d system_u:object_r:proc_net_t:s0
genfscon proc /net -- system_u:object_r:proc_net_t:s0
would only label directories and regular files with proc_net_t.
[...]
Am Mi., 29. Jan. 2020 um 14:34 Uhr schrieb Stephen Smalley <[email protected]>:
Quoted 11 lines
This looks fine to me code-wise. Have you tried enabling this new policy capability in policy and seeing the effects of it? I remember a problem in the way-back time that motivated adding the
S_ISLNK()exception for proc. IIRC, the issue was that policies specified "genfscon proc /net system_u:object_r:proc_net_t:s0" to label everything under/proc/netwithproc_net_tby default, and when/proc/netwas changed to be a symlink to/proc/self/netas part of the network namespaces work, this caused the symlink to be labeledproc_net_t, but since previously there were no symlinks labeledproc_net_t, no processes were allowed to read/follow the symlink and existing systems broke. Exempting symlinks caused the/proc/netsymlink to be labeledproc_t
Quoted 8 lines
instead, which was already widely allowed. To avoid this problem when enabling this capability, you will either need to allow proc_net_t:lnk_file read widely or you will need to change the genfscon statement to avoid labeling the symlink (there is an optional file mode field in genfscon statements similar to that of
file_contexts, e.g.
genfscon proc /net -d system_u:object_r:proc_net_t:s0 genfscon proc /net -- system_u:object_r:proc_net_t:s0
would only label directories and regular files withproc_net_t.
I tested it with a refpolicy alike policy, where kernel_read_network_state() allows access to symlinks [1]. Just systemd pid 1 wants to read the symlink only without reading any files, so I added kernel_search_network_state(systemd_t).
Paul Moore
maintainerQuoted 21 lines · Christian Göttsche
Am Mi., 29. Jan. 2020 um 14:34 Uhr schrieb Stephen Smalley <[email protected]>:
Quoted 11 lines
This looks fine to me code-wise. Have you tried enabling this new policy capability in policy and seeing the effects of it? I remember a problem in the way-back time that motivated adding the
S_ISLNK()exception for proc. IIRC, the issue was that policies specified "genfscon proc /net system_u:object_r:proc_net_t:s0" to label everything under/proc/netwithproc_net_tby default, and when/proc/netwas changed to be a symlink to/proc/self/netas part of the network namespaces work, this caused the symlink to be labeledproc_net_t, but since previously there were no symlinks labeledproc_net_t, no processes were allowed to read/follow the symlink and existing systems broke. Exempting symlinks caused the/proc/netsymlink to be labeledproc_tQuoted 8 lines
instead, which was already widely allowed. To avoid this problem when enabling this capability, you will either need to allow proc_net_t:lnk_file read widely or you will need to change the genfscon statement to avoid labeling the symlink (there is an optional file mode field in genfscon statements similar to that of
file_contexts, e.g.
genfscon proc /net -d system_u:object_r:proc_net_t:s0 genfscon proc /net -- system_u:object_r:proc_net_t:s0
would only label directories and regular files withproc_net_t.
Quoted 6 lines
I tested it with a refpolicy alike policy, where
kernel_read_network_state()allows access to symlinks [1]. Just systemd pid 1 wants to read the symlink only without reading any files, so I addedkernel_search_network_state(systemd_t).
Thanks. FWIW, I think the netns procfs case is a little different simply because it was a change in the core kernel behavior and not a change in the access controls. While we have tricks, e.g. policy capabilities, to help with migrating users/admins to new access control changes, we don't always have the same abilities when the core kernel changes.
One could make the argument that the netns procfs change was a "break userspace" type of change, but I doubt you would have gotten far with that because well ... containers.
Anyway, I've gone ahead and queued the patch up in selinux/next, you should see it in the tree once the merge window closes.
Signature · 2 lines
paul moore www.paul-moore.com