apparmor: Enforce progressively tighter permissions for no_new_privs
Subsystems (2)
RFC PATCH 1/1 apparmor: Enforce progressively tighter permissions for no_new_privs
The current understanding of apparmor with respect to no_new_privs is at odds with how no_new_privs is implemented and understood by the rest of the kernel.
The documentation of no_new_privs states:
Quoted 3 lines
With ``no_new_privs`` set, ``execve()`` promises not to grant the privilege to do anything that could not have been done without the execve call.
And reading through the kernel except for apparmor that description matches what is implemented.
There are two major divergences of apparmor from this definition:
- proc_setattr enforces limitations when no_new_privs are set.
- the limitation is enforced from the apparent time when no_new_privs is set instead of guaranteeing that each execve has progressively more narrow permissions.
The code in apparmor that attempts to discover the apparmor label at the point where no_new_privs is set is not robust. The capture happens a long time after no_new_privs is set.
Capturing the label at the point where no_new_privs is set is practically impossible to implement robustly. Today the rule is struct cred can only be changed by it's current task. Today SECCOMP_FILTER_FLAG_TSYNC sets no_new_privs from another thread. A robust implementation would require changing something fundamental in how creds are managed for SECCOMP_FILTER_FLAG_TSYNC to be able to capture the cred at the point it is set.
Futhermore given the consistent documentation and how everything else implements no_new_privs, not having the permissions get progressively tighter is a footgun aimed at userspace. I fully expect it to break any security sensitive software that uses no_new_privs and was not deliberately designed and tested against apparmor.
Avoid the questionable and hard to fix implementation and the potential to confuse userspace by having no_new_privs enforce progressinvely tighter permissions.
---
I came accross this while examining the places cred_guard_mutex is used and trying to find a way to make those code paths less insane.
If it would be more pallatable I would not mind removing the task_no_new_privs test entirely from aa_change_hat and aa_change_profile as those are not part of exec, so arguably no_new_privs should not care about them at all.
Can we please get rid of the huge semantic wart and pain in the implementation?
security/apparmor/domain.c | 39 ++++---------------------------- security/apparmor/include/task.h | 4 ---- security/apparmor/task.c | 7 ------ 3 files changed, 4 insertions(+), 46 deletions(-)
security/apparmor/domain.c +4 -35
@@ -869,17 +869,6 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm) label =aa_get_newest_label (cred_label (bprm->cred)); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !ctx->nnp) - ctx->nnp =aa_get_label (label); -/* buffer freed below, name is pointer into buffer */ buffer =aa_get_buffer (false);if (!buffer) {
@@ -915,7 +904,7 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm)*/ if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) { error = -EPERM ; info ="no new privs" ; goto audit;
@@ -1158,16 +1147,6 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags) label =aa_get_newest_cred_label (cred); previous =aa_get_newest_label (ctx->previous); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (unconfined (label)) { info ="unconfined can not change_hat" ; error = -EPERM ;
@@ -1193,7 +1172,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;
@@ -1214,7 +1193,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (previous, ctx->nnp)) { + !aa_label_is_unconfined_subset (previous, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;
@@ -1303,16 +1282,6 @@ int aa_change_profile(const char *fqname, int flags) label =aa_get_current_label (); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (!fqname || !*fqname) {aa_put_label (label);AA_DEBUG ("no profile name" );
@@ -1409,7 +1378,7 @@ int aa_change_profile(const char *fqname, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;
security/apparmor/include/task.h +0 -4
@@ -17,13 +17,11 @@ static inline struct aa_task_ctx *task_ctx(struct task_struct *task)/* * struct aa_task_ctx - information for current task label change -* @nnp: snapshot of label at time of no_new_privs * @onexec: profile to transition to on next exec (MAY BE NULL) * @previous: profile the task may return to (MAY BE NULL) * @token: magic value the task must know for returning to @previous_profile */ struct aa_task_ctx { -struct aa_label *nnp;struct aa_label *onexec;struct aa_label *previous;u64 token;
@@ -42,7 +40,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task);static inline void aa_free_task_ctx (struct aa_task_ctx *ctx) {if (ctx) { -aa_put_label (ctx->nnp);aa_put_label (ctx->previous);aa_put_label (ctx->onexec); }
@@ -57,7 +54,6 @@ static inline void aa_dup_task_ctx(struct aa_task_ctx *new, const struct aa_task_ctx *old) { *new = *old; -aa_get_label (new->nnp);aa_get_label (new->previous);aa_get_label (new->onexec); }
security/apparmor/task.c +0 -7
@@ -41,7 +41,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task)int aa_replace_current_label (struct aa_label *label) {struct aa_label *old =aa_current_raw_label (); -struct aa_task_ctx *ctx =task_ctx (current);struct cred *new;AA_BUG (!label);
@@ -56,12 +55,6 @@ int aa_replace_current_label(struct aa_label *label)if (!new)return -ENOMEM ; -if (ctx->nnp &&label_is_stale (ctx->nnp)) { -struct aa_label *tmp = ctx->nnp; - - ctx->nnp =aa_get_newest_label (tmp); -aa_put_label (tmp); - }if (unconfined (label) || (labels_ns (old) !=labels_ns (label)))/* * if switching to unconfined or a different label namespace
Signature · 3 lines
2.20.1 Eric
John Johansen
maintainerQuoted 5 lines · Eric W. Biederman
The current understanding of apparmor with respect to
no_new_privsis at odds with howno_new_privsis implemented and understood by the rest of the kernel.
Quoted 8 lines
The documentation of
no_new_privsstates:Quoted 3 lines
With ``no_new_privs`` set, ``execve()`` promises not to grant the privilege to do anything that could not have been done without the execve call.
And reading through the kernel except for apparmor that description matches what is implemented.
That is not correct.
commit 7b0d0b40cd78 ("selinux: Permit bounded transitions under
NO_NEW_PRIVS or NOSUID.")
Allows for bound transitions under selinux
and
commit af63f4193f9f selinux: Generalize support for NNP/nosuid SELinux
domain transitions
goes further and "Decouple NNP/nosuid from SELinux transitions".
Quoted 10 lines
There are two major divergences of apparmor from this definition:
-proc_setattrenforces limitations whenno_new_privsare set.
- the limitation is enforced from the apparent time whenno_new_privsis set instead of guaranteeing that each execve has progressively more narrow permissions.The code in apparmor that attempts to discover the apparmor label at the point where
no_new_privsis set is not robust. The capture happens a long time afterno_new_privsis set.
yes, but that shouldn't matter. As apparmor has not changed its label at any point between when no_new_privs was set and when the check is done. AppArmor is attempting to change it label, and if it finds NNP has been set we capture what the confinement was.
Quoted 3 lines
Capturing the label at the point where
no_new_privsis set is practically impossible to implement robustly. Today the rule is struct cred can only be changed by it's current task. Today
right, and apparmor only ever has the task update its own label.
Quoted 5 lines
SECCOMP_FILTER_FLAG_TSYNCsetsno_new_privsfrom another thread. A robust implementation would require changing something fundamental in how creds are managed forSECCOMP_FILTER_FLAG_TSYNCto be able to capture the cred at the point it is set.
I am open to supporting something like that.
Quoted 2 lines
Futhermore given the consistent documentation and how everything else implements
no_new_privs, not having the permissions get progressively
Again see above
Quoted 1 line
tighter is a footgun aimed at userspace. I fully expect it to break any
tighter is somewhat relative, nor is it only progressively tighter it is bounded against the snapshot of the label that was on the task.
Quoted 3 lines
security sensitive software that uses
no_new_privsand was not deliberately designed and tested against apparmor.
Currently the situation has become either an either or choice between the LSM and NNP. We are trying to walk a balance. Ideally apparmor would like to do something similar to selinux and decouple the label transition from NNP and nosuid via an internal capability, but we have not gone there yet.
Quoted 4 lines
Avoid the questionable and hard to fix implementation and the potential to confuse userspace by having
no_new_privsenforce progressinvely tighter permissions.
This would completely break several use cases.
Quoted 171 lines
Fixes 9fcf78cca198 ("apparmor: update domain transitions that are subsets of confinement at nnp")Signed-off-by Eric W. Biederman <[email protected]>---
I came accross this while examining the places
cred_guard_mutexis used and trying to find a way to make those code paths less insane.If it would be more pallatable I would not mind removing the
task_no_new_privstest entirely fromaa_change_hatandaa_change_profileas those are not part of exec, so arguablyno_new_privsshould not care about them at all.Can we please get rid of the huge semantic wart and pain in the implementation?
security/apparmor/domain.c | 39 ++++---------------------------- security/apparmor/include/task.h | 4 ---- security/apparmor/task.c | 7 ------ 3 files changed, 4 insertions(+), 46 deletions(-)security/apparmor/domain.c +4
-35@@ -869,17 +869,6 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm) label =aa_get_newest_label (cred_label (bprm->cred)); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !ctx->nnp) - ctx->nnp =aa_get_label (label); -/* buffer freed below, name is pointer into buffer */ buffer =aa_get_buffer (false);if (!buffer) {@@ -915,7 +904,7 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm)*/ if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) { error = -EPERM ; info ="no new privs" ; goto audit;@@ -1158,16 +1147,6 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags) label =aa_get_newest_cred_label (cred); previous =aa_get_newest_label (ctx->previous); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (unconfined (label)) { info ="unconfined can not change_hat" ; error = -EPERM ;@@ -1193,7 +1172,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1214,7 +1193,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (previous, ctx->nnp)) { + !aa_label_is_unconfined_subset (previous, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1303,16 +1282,6 @@ int aa_change_profile(const char *fqname, int flags) label =aa_get_current_label (); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (!fqname || !*fqname) {aa_put_label (label);AA_DEBUG ("no profile name" );@@ -1409,7 +1378,7 @@ int aa_change_profile(const char *fqname, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;security/apparmor/include/task.h +0
-4@@ -17,13 +17,11 @@ static inline struct aa_task_ctx *task_ctx(struct task_struct *task)/* * struct aa_task_ctx - information for current task label change -* @nnp: snapshot of label at time of no_new_privs * @onexec: profile to transition to on next exec (MAY BE NULL) * @previous: profile the task may return to (MAY BE NULL) * @token: magic value the task must know for returning to @previous_profile */ struct aa_task_ctx { -struct aa_label *nnp;struct aa_label *onexec;struct aa_label *previous;u64 token;@@ -42,7 +40,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task);static inline void aa_free_task_ctx (struct aa_task_ctx *ctx) {if (ctx) { -aa_put_label (ctx->nnp);aa_put_label (ctx->previous);aa_put_label (ctx->onexec); }@@ -57,7 +54,6 @@ static inline void aa_dup_task_ctx(struct aa_task_ctx *new, const struct aa_task_ctx *old) { *new = *old; -aa_get_label (new->nnp);aa_get_label (new->previous);aa_get_label (new->onexec); }security/apparmor/task.c +0
-7@@ -41,7 +41,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task)int aa_replace_current_label (struct aa_label *label) {struct aa_label *old =aa_current_raw_label (); -struct aa_task_ctx *ctx =task_ctx (current);struct cred *new;AA_BUG (!label);@@ -56,12 +55,6 @@ int aa_replace_current_label(struct aa_label *label)if (!new)return -ENOMEM ; -if (ctx->nnp &&label_is_stale (ctx->nnp)) { -struct aa_label *tmp = ctx->nnp; - - ctx->nnp =aa_get_newest_label (tmp); -aa_put_label (tmp); - }if (unconfined (label) || (labels_ns (old) !=labels_ns (label)))/* * if switching to unconfined or a different label namespace
TL;DR ????selinux and apparmor ignore no_new_privs????
What?????
Quoted 15 lines · John Johansen
Quoted 5 lines · Eric W. Biederman
The current understanding of apparmor with respect to
no_new_privsis at odds with howno_new_privsis implemented and understood by the rest of the kernel.Quoted 8 lines
The documentation of
no_new_privsstates:Quoted 3 lines
With ``no_new_privs`` set, ``execve()`` promises not to grant the privilege to do anything that could not have been done without the execve call.
And reading through the kernel except for apparmor that description matches what is implemented.
Quoted 7 lines
That is not correct.
commit
7b0d0b40cd78("selinux: Permit bounded transitions under
NO_NEW_PRIVSor NOSUID.")Allows for bound transitions under selinux
and
As I understand a bound transition it is a transition to a state with a set of permissions that are a subset of what was previously held. Which is consistent with the mandate of no_new_privs.
Quoted 4 lines
commit
af63f4193f9fselinux: Generalize support for NNP/nosuid SELinux
domain transitionsgoes further and "Decouple NNP/nosuid from SELinux transitions".
Yes. Looking at that commit I do see that selinux appears to be deliberately ignoring no_new_privs in specific cases.
WTF.
Quoted 20 lines
Quoted 10 lines
There are two major divergences of apparmor from this definition:
-proc_setattrenforces limitations whenno_new_privsare set.
- the limitation is enforced from the apparent time whenno_new_privsis set instead of guaranteeing that each execve has progressively more narrow permissions.The code in apparmor that attempts to discover the apparmor label at the point where
no_new_privsis set is not robust. The capture happens a long time afterno_new_privsis set.yes, but that shouldn't matter. As apparmor has not changed its label at any point between when
no_new_privswas set and when the check is done. AppArmor is attempting to change it label, and if it finds NNP has been set we capture what the confinement was.Quoted 3 lines
Capturing the label at the point where
no_new_privsis set is practically impossible to implement robustly. Today the rule is struct cred can only be changed by it's current task. Today
Quoted 8 lines
right, and apparmor only ever has the task update its own label.
Quoted 5 lines
SECCOMP_FILTER_FLAG_TSYNCsetsno_new_privsfrom another thread. A robust implementation would require changing something fundamental in how creds are managed forSECCOMP_FILTER_FLAG_TSYNCto be able to capture the cred at the point it is set.I am open to supporting something like that.
I can't see how it would be possible to be robust without completely changing the locking. Locking that right now in a simpler model we have not figured out how to make obviously correct.
Quoted 4 lines
Quoted 2 lines
Futhermore given the consistent documentation and how everything else implements
no_new_privs, not having the permissions get progressivelyAgain see above
Except where selinux deliberately ignores no_new_privs this is consitent.
Quoted 4 lines
Quoted 1 line
tighter is a footgun aimed at userspace. I fully expect it to break any
tighter is somewhat relative, nor is it only progressively tighter it is bounded against the snapshot of the label that was on the task.
Which is the BUG I am reporting. It should be progressingly tighter.
Quoted 9 lines
Quoted 3 lines
security sensitive software that uses
no_new_privsand was not deliberately designed and tested against apparmor.Currently the situation has become either an either or choice between the LSM and NNP. We are trying to walk a balance. Ideally apparmor would like to do something similar to selinux and decouple the label transition from NNP and nosuid via an internal capability, but we have not gone there yet.
Why do you need to escape no_new_privs. Why does anyone need to escape no_new_privs?
Quoted 6 lines
Quoted 4 lines
Avoid the questionable and hard to fix implementation and the potential to confuse userspace by having
no_new_privsenforce progressinvely tighter permissions.This would completely break several use cases.
Enforcing no_new_privs as documented would break userspace?
Isn't the opposite true that you are breaking people by not enforcing it?
Quoted 161 lines
Quoted 161 lines
Fixes 9fcf78cca198 ("apparmor: update domain transitions that are subsets of confinement at nnp")Signed-off-by Eric W. Biederman <[email protected]>---
I came accross this while examining the places
cred_guard_mutexis used and trying to find a way to make those code paths less insane.If it would be more pallatable I would not mind removing the
task_no_new_privstest entirely fromaa_change_hatandaa_change_profileas those are not part of exec, so arguablyno_new_privsshould not care about them at all.Can we please get rid of the huge semantic wart and pain in the implementation?
security/apparmor/domain.c | 39 ++++---------------------------- security/apparmor/include/task.h | 4 ---- security/apparmor/task.c | 7 ------ 3 files changed, 4 insertions(+), 46 deletions(-)security/apparmor/domain.c +4
-35@@ -869,17 +869,6 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm) label =aa_get_newest_label (cred_label (bprm->cred)); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !ctx->nnp) - ctx->nnp =aa_get_label (label); -/* buffer freed below, name is pointer into buffer */ buffer =aa_get_buffer (false);if (!buffer) {@@ -915,7 +904,7 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm)*/ if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) { error = -EPERM ; info ="no new privs" ; goto audit;@@ -1158,16 +1147,6 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags) label =aa_get_newest_cred_label (cred); previous =aa_get_newest_label (ctx->previous); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (unconfined (label)) { info ="unconfined can not change_hat" ; error = -EPERM ;@@ -1193,7 +1172,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1214,7 +1193,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (previous, ctx->nnp)) { + !aa_label_is_unconfined_subset (previous, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1303,16 +1282,6 @@ int aa_change_profile(const char *fqname, int flags) label =aa_get_current_label (); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (!fqname || !*fqname) {aa_put_label (label);AA_DEBUG ("no profile name" );@@ -1409,7 +1378,7 @@ int aa_change_profile(const char *fqname, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;security/apparmor/include/task.h +0
-4@@ -17,13 +17,11 @@ static inline struct aa_task_ctx *task_ctx(struct task_struct *task)/* * struct aa_task_ctx - information for current task label change -* @nnp: snapshot of label at time of no_new_privs * @onexec: profile to transition to on next exec (MAY BE NULL) * @previous: profile the task may return to (MAY BE NULL) * @token: magic value the task must know for returning to @previous_profile */ struct aa_task_ctx { -struct aa_label *nnp;struct aa_label *onexec;struct aa_label *previous;u64 token;@@ -42,7 +40,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task);static inline void aa_free_task_ctx (struct aa_task_ctx *ctx) {if (ctx) { -aa_put_label (ctx->nnp);aa_put_label (ctx->previous);aa_put_label (ctx->onexec); }@@ -57,7 +54,6 @@ static inline void aa_dup_task_ctx(struct aa_task_ctx *new, const struct aa_task_ctx *old) { *new = *old; -aa_get_label (new->nnp);aa_get_label (new->previous);aa_get_label (new->onexec); }security/apparmor/task.c +0
-1@@ -41,7 +41,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task)int aa_replace_current_label (struct aa_label *label) {struct aa_label *old =aa_current_raw_label (); -struct aa_task_ctx *ctx =task_ctx (current);struct cred *new;AA_BUG (!label);@@ -56,12 +55,6 @@ int aa_replace_current_label(struct aa_label *label)if (!new)return -ENOMEM ;
Quoted 10 lines
security/apparmor/task.c int aa_replace_current_label(struct aa_label *label) · line 60
Quoted 10 lines
security/apparmor/task.c +0
-6-if (ctx->nnp &&label_is_stale (ctx->nnp)) { -struct aa_label *tmp = ctx->nnp; - - ctx->nnp =aa_get_newest_label (tmp); -aa_put_label (tmp); - }if (unconfined (label) || (labels_ns (old) !=labels_ns (label)))/* * if switching to unconfined or a different label namespace
Eric
John Johansen
maintainerQuoted 5 lines · Eric W. Biederman
TL;DR ????selinux and apparmor ignore
no_new_privs????What?????
AppArmor does not ignore no_new_privs. Its mediation is bounded and it doesn't grant anything that wasn't allowed when NNP was set.
Quoted 287 lines
Quoted 15 lines · John Johansen
Quoted 5 lines · Eric W. Biederman
The current understanding of apparmor with respect to
no_new_privsis at odds with howno_new_privsis implemented and understood by the rest of the kernel.Quoted 8 lines
The documentation of
no_new_privsstates:Quoted 3 lines
With ``no_new_privs`` set, ``execve()`` promises not to grant the privilege to do anything that could not have been done without the execve call.
And reading through the kernel except for apparmor that description matches what is implemented.
Quoted 7 lines
That is not correct.
commit
7b0d0b40cd78("selinux: Permit bounded transitions under
NO_NEW_PRIVSor NOSUID.")Allows for bound transitions under selinux
andAs I understand a bound transition it is a transition to a state with a set of permissions that are a subset of what was previously held. Which is consistent with the mandate of
no_new_privs.Quoted 4 lines
commit
af63f4193f9fselinux: Generalize support for NNP/nosuid SELinux
domain transitionsgoes further and "Decouple NNP/nosuid from SELinux transitions".
Yes. Looking at that commit I do see that selinux appears to be deliberately ignoring
no_new_privsin specific cases.WTF.
Quoted 20 lines
Quoted 10 lines
There are two major divergences of apparmor from this definition:
-proc_setattrenforces limitations whenno_new_privsare set.
- the limitation is enforced from the apparent time whenno_new_privsis set instead of guaranteeing that each execve has progressively more narrow permissions.The code in apparmor that attempts to discover the apparmor label at the point where
no_new_privsis set is not robust. The capture happens a long time afterno_new_privsis set.yes, but that shouldn't matter. As apparmor has not changed its label at any point between when
no_new_privswas set and when the check is done. AppArmor is attempting to change it label, and if it finds NNP has been set we capture what the confinement was.Quoted 3 lines
Capturing the label at the point where
no_new_privsis set is practically impossible to implement robustly. Today the rule is struct cred can only be changed by it's current task. TodayQuoted 8 lines
right, and apparmor only ever has the task update its own label.
Quoted 5 lines
SECCOMP_FILTER_FLAG_TSYNCsetsno_new_privsfrom another thread. A robust implementation would require changing something fundamental in how creds are managed forSECCOMP_FILTER_FLAG_TSYNCto be able to capture the cred at the point it is set.I am open to supporting something like that.
I can't see how it would be possible to be robust without completely changing the locking. Locking that right now in a simpler model we have not figured out how to make obviously correct.
Quoted 4 lines
Quoted 2 lines
Futhermore given the consistent documentation and how everything else implements
no_new_privs, not having the permissions get progressivelyAgain see above
Except where selinux deliberately ignores
no_new_privsthis is consitent.Quoted 4 lines
Quoted 1 line
tighter is a footgun aimed at userspace. I fully expect it to break any
tighter is somewhat relative, nor is it only progressively tighter it is bounded against the snapshot of the label that was on the task.
Which is the BUG I am reporting. It should be progressingly tighter.
Quoted 9 lines
Quoted 3 lines
security sensitive software that uses
no_new_privsand was not deliberately designed and tested against apparmor.Currently the situation has become either an either or choice between the LSM and NNP. We are trying to walk a balance. Ideally apparmor would like to do something similar to selinux and decouple the label transition from NNP and nosuid via an internal capability, but we have not gone there yet.
Why do you need to escape
no_new_privs. Why does anyone need to escapeno_new_privs?Quoted 6 lines
Quoted 4 lines
Avoid the questionable and hard to fix implementation and the potential to confuse userspace by having
no_new_privsenforce progressinvely tighter permissions.This would completely break several use cases.
Enforcing
no_new_privsas documented would break userspace?Isn't the opposite true that you are breaking people by not enforcing it?
Quoted 161 lines
Quoted 161 lines
Fixes 9fcf78cca198 ("apparmor: update domain transitions that are subsets of confinement at nnp")Signed-off-by Eric W. Biederman <[email protected]>---
I came accross this while examining the places
cred_guard_mutexis used and trying to find a way to make those code paths less insane.If it would be more pallatable I would not mind removing the
task_no_new_privstest entirely fromaa_change_hatandaa_change_profileas those are not part of exec, so arguablyno_new_privsshould not care about them at all.Can we please get rid of the huge semantic wart and pain in the implementation?
security/apparmor/domain.c | 39 ++++---------------------------- security/apparmor/include/task.h | 4 ---- security/apparmor/task.c | 7 ------ 3 files changed, 4 insertions(+), 46 deletions(-)security/apparmor/domain.c +4
-35@@ -869,17 +869,6 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm) label =aa_get_newest_label (cred_label (bprm->cred)); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !ctx->nnp) - ctx->nnp =aa_get_label (label); -/* buffer freed below, name is pointer into buffer */ buffer =aa_get_buffer (false);if (!buffer) {@@ -915,7 +904,7 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm)*/ if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) { error = -EPERM ; info ="no new privs" ; goto audit;@@ -1158,16 +1147,6 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags) label =aa_get_newest_cred_label (cred); previous =aa_get_newest_label (ctx->previous); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (unconfined (label)) { info ="unconfined can not change_hat" ; error = -EPERM ;@@ -1193,7 +1172,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1214,7 +1193,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (previous, ctx->nnp)) { + !aa_label_is_unconfined_subset (previous, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1303,16 +1282,6 @@ int aa_change_profile(const char *fqname, int flags) label =aa_get_current_label (); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (!fqname || !*fqname) {aa_put_label (label);AA_DEBUG ("no profile name" );@@ -1409,7 +1378,7 @@ int aa_change_profile(const char *fqname, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;security/apparmor/include/task.h +0
-4@@ -17,13 +17,11 @@ static inline struct aa_task_ctx *task_ctx(struct task_struct *task)/* * struct aa_task_ctx - information for current task label change -* @nnp: snapshot of label at time of no_new_privs * @onexec: profile to transition to on next exec (MAY BE NULL) * @previous: profile the task may return to (MAY BE NULL) * @token: magic value the task must know for returning to @previous_profile */ struct aa_task_ctx { -struct aa_label *nnp;struct aa_label *onexec;struct aa_label *previous;u64 token;@@ -42,7 +40,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task);static inline void aa_free_task_ctx (struct aa_task_ctx *ctx) {if (ctx) { -aa_put_label (ctx->nnp);aa_put_label (ctx->previous);aa_put_label (ctx->onexec); }@@ -57,7 +54,6 @@ static inline void aa_dup_task_ctx(struct aa_task_ctx *new, const struct aa_task_ctx *old) { *new = *old; -aa_get_label (new->nnp);aa_get_label (new->previous);aa_get_label (new->onexec); }security/apparmor/task.c +0
-1@@ -41,7 +41,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task)int aa_replace_current_label (struct aa_label *label) {struct aa_label *old =aa_current_raw_label (); -struct aa_task_ctx *ctx =task_ctx (current);struct cred *new;AA_BUG (!label);@@ -56,12 +55,6 @@ int aa_replace_current_label(struct aa_label *label)if (!new)return -ENOMEM ;Quoted 10 lines
security/apparmor/task.c int aa_replace_current_label(struct aa_label *label) · line 60
Quoted 10 lines
security/apparmor/task.c +0
-6-if (ctx->nnp &&label_is_stale (ctx->nnp)) { -struct aa_label *tmp = ctx->nnp; - - ctx->nnp =aa_get_newest_label (tmp); -aa_put_label (tmp); - }if (unconfined (label) || (labels_ns (old) !=labels_ns (label)))/* * if switching to unconfined or a different label namespace Eric
This should now Cc the correct email address for James Morris.
Quoted 210 lines · [email protected] (Eric W. Biederman) writes:
The current understanding of apparmor with respect to
no_new_privsis at odds with howno_new_privsis implemented and understood by the rest of the kernel.The documentation of
no_new_privsstates:Quoted 3 lines
With ``no_new_privs`` set, ``execve()`` promises not to grant the privilege to do anything that could not have been done without the execve call.
And reading through the kernel except for apparmor that description matches what is implemented.
There are two major divergences of apparmor from this definition:
-proc_setattrenforces limitations whenno_new_privsare set.
- the limitation is enforced from the apparent time whenno_new_privsis set instead of guaranteeing that each execve has progressively more narrow permissions.The code in apparmor that attempts to discover the apparmor label at the point where
no_new_privsis set is not robust. The capture happens a long time afterno_new_privsis set.Capturing the label at the point where
no_new_privsis set is practically impossible to implement robustly. Today the rule is struct cred can only be changed by it's current task. TodaySECCOMP_FILTER_FLAG_TSYNCsetsno_new_privsfrom another thread. A robust implementation would require changing something fundamental in how creds are managed forSECCOMP_FILTER_FLAG_TSYNCto be able to capture the cred at the point it is set.Futhermore given the consistent documentation and how everything else implements
no_new_privs, not having the permissions get progressively tighter is a footgun aimed at userspace. I fully expect it to break any security sensitive software that usesno_new_privsand was not deliberately designed and tested against apparmor.Avoid the questionable and hard to fix implementation and the potential to confuse userspace by having
no_new_privsenforce progressinvely tighter permissions.Fixes 9fcf78cca198 ("apparmor: update domain transitions that are subsets of confinement at nnp")Signed-off-by Eric W. Biederman <[email protected]>---
I came accross this while examining the places
cred_guard_mutexis used and trying to find a way to make those code paths less insane.If it would be more pallatable I would not mind removing the
task_no_new_privstest entirely fromaa_change_hatandaa_change_profileas those are not part of exec, so arguablyno_new_privsshould not care about them at all.Can we please get rid of the huge semantic wart and pain in the implementation?
security/apparmor/domain.c | 39 ++++---------------------------- security/apparmor/include/task.h | 4 ---- security/apparmor/task.c | 7 ------ 3 files changed, 4 insertions(+), 46 deletions(-)security/apparmor/domain.c +4
-35@@ -869,17 +869,6 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm) label =aa_get_newest_label (cred_label (bprm->cred)); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !ctx->nnp) - ctx->nnp =aa_get_label (label); -/* buffer freed below, name is pointer into buffer */ buffer =aa_get_buffer (false);if (!buffer) {@@ -915,7 +904,7 @@ int apparmor_bprm_creds_for_exec(struct linux_binprm *bprm)*/ if ((bprm->unsafe &LSM_UNSAFE_NO_NEW_PRIVS ) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) { error = -EPERM ; info ="no new privs" ; goto audit;@@ -1158,16 +1147,6 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags) label =aa_get_newest_cred_label (cred); previous =aa_get_newest_label (ctx->previous); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (unconfined (label)) { info ="unconfined can not change_hat" ; error = -EPERM ;@@ -1193,7 +1172,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1214,7 +1193,7 @@ int aa_change_hat(const char *hats[], int count, u64 token, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (previous, ctx->nnp)) { + !aa_label_is_unconfined_subset (previous, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;@@ -1303,16 +1282,6 @@ int aa_change_profile(const char *fqname, int flags) label =aa_get_current_label (); -/* -* Detect no new privs being set, and store the label it -* occurred under. Ideally this would happen when nnp -* is set but there isn't a good way to do that yet. -* -* Testing for unconfined must be done before the subset test -*/ -if (task_no_new_privs (current) && !unconfined (label) && !ctx->nnp) - ctx->nnp =aa_get_label (label); -if (!fqname || !*fqname) {aa_put_label (label);AA_DEBUG ("no profile name" );@@ -1409,7 +1378,7 @@ int aa_change_profile(const char *fqname, int flags)* reduce restrictions. */ if (task_no_new_privs (current) && !unconfined (label) && - !aa_label_is_unconfined_subset (new, ctx->nnp)) { + !aa_label_is_unconfined_subset (new, label)) {/* not an apparmor denial per se, so don't log it */ AA_DEBUG ("no_new_privs - change_hat denied" ); error = -EPERM ;security/apparmor/include/task.h +0
-4@@ -17,13 +17,11 @@ static inline struct aa_task_ctx *task_ctx(struct task_struct *task)/* * struct aa_task_ctx - information for current task label change -* @nnp: snapshot of label at time of no_new_privs * @onexec: profile to transition to on next exec (MAY BE NULL) * @previous: profile the task may return to (MAY BE NULL) * @token: magic value the task must know for returning to @previous_profile */ struct aa_task_ctx { -struct aa_label *nnp;struct aa_label *onexec;struct aa_label *previous;u64 token;@@ -42,7 +40,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task);static inline void aa_free_task_ctx (struct aa_task_ctx *ctx) {if (ctx) { -aa_put_label (ctx->nnp);aa_put_label (ctx->previous);aa_put_label (ctx->onexec); }@@ -57,7 +54,6 @@ static inline void aa_dup_task_ctx(struct aa_task_ctx *new, const struct aa_task_ctx *old) { *new = *old; -aa_get_label (new->nnp);aa_get_label (new->previous);aa_get_label (new->onexec); }security/apparmor/task.c +0
-7@@ -41,7 +41,6 @@ struct aa_label *aa_get_task_label(struct task_struct *task)int aa_replace_current_label (struct aa_label *label) {struct aa_label *old =aa_current_raw_label (); -struct aa_task_ctx *ctx =task_ctx (current);struct cred *new;AA_BUG (!label);@@ -56,12 +55,6 @@ int aa_replace_current_label(struct aa_label *label)if (!new)return -ENOMEM ; -if (ctx->nnp &&label_is_stale (ctx->nnp)) { -struct aa_label *tmp = ctx->nnp; - - ctx->nnp =aa_get_newest_label (tmp); -aa_put_label (tmp); - }if (unconfined (label) || (labels_ns (old) !=labels_ns (label)))/* * if switching to unconfined or a different label namespace