Limiting SECCOMP audit events

Post by Steve Grubb
Hello,
Over the last month, the amount of seccomp events in audit logs is
sky-rocketing. I have over a million events in the last 2 days. Most of this
is generated by firefox and qt webkit.
I am wondering if the audit package should ship a file for
/usr/lib/sysctl.d/60-auditd.conf
wherein it has
kernel.seccomp.actions_logged = kill_process kill_thread errno
Also, has anyone verified this sysctl is filtering audit events? Even with
the above, I have over a million events on a 4.14.3 kernel. Firefox alone is
generating over 50,000 events per hour.

I don't think you'd want to log errno -- AIUI, that's used regularly
by a lot of seccomp policy.

-Kees

--
Kees Cook
Pixel Security

Steve Grubb

2017-12-14 00:31:44 UTC

Post by Kees Cook

Post by Steve Grubb
Hello,
Over the last month, the amount of seccomp events in audit logs is
sky-rocketing. I have over a million events in the last 2 days. Most of
this is generated by firefox and qt webkit.
I am wondering if the audit package should ship a file for
/usr/lib/sysctl.d/60-auditd.conf
wherein it has
kernel.seccomp.actions_logged = kill_process kill_thread errno
Also, has anyone verified this sysctl is filtering audit events? Even with
the above, I have over a million events on a 4.14.3 kernel. Firefox alone
is generating over 50,000 events per hour.

I don't think you'd want to log errno -- AIUI, that's used regularly
by a lot of seccomp policy.

I'm not seeing any reporting errno. The ones reporting trap are coming in over
50,000 per hour. I don't think the filter is working.

[***@x2 ~]# cat /proc/sys/kernel/seccomp/actions_logged
kill_process kill_thread errno
[***@x2 ~]# date
Wed Dec 13 19:24:40 EST 2017
[***@x2 ~]# ausearch --start 19:24:40 -m seccomp --raw | aureport --event --
summary -i

Event Summary Report
======================
total type
======================
170 SECCOMP

In the time it took to type the command 170 seccomp events were recorded from
firefox.

[***@x2 ~]# ausearch --start 19:24:40 -m seccomp --just-one -i
----
node=x2 type=SECCOMP msg=audit(12/13/2017 19:24:56.454:199666) :
auid=sgrubb uid=sgrubb gid=sgrubb ses=3
subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 pid=3394
comm=Web Content exe=/usr/lib64/firefox/firefox sig=SIG0 arch=x86_64
syscall=stat compat=0 ip=0x7f909c828635 code=trap

^^ trap. With the sheer amount of events being recorded, I think it's necessary to
add a sysctl file to systems to suppress the logging. Especially when you consider
that systemd-journal also gratuitously grabs audit logs and sends them to rsyslog.

-Steve

Paul Moore

2017-12-14 01:43:38 UTC

Post by Kees Cook

Post by Steve Grubb
Hello,
Over the last month, the amount of seccomp events in audit logs is
sky-rocketing. I have over a million events in the last 2 days. Most of
this is generated by firefox and qt webkit.
I am wondering if the audit package should ship a file for
/usr/lib/sysctl.d/60-auditd.conf
wherein it has
kernel.seccomp.actions_logged = kill_process kill_thread errno
Also, has anyone verified this sysctl is filtering audit events? Even with
the above, I have over a million events on a 4.14.3 kernel. Firefox alone
is generating over 50,000 events per hour.

I don't think you'd want to log errno -- AIUI, that's used regularly
by a lot of seccomp policy.

I'm not seeing any reporting errno. The ones reporting trap are coming in
over 50,000 per hour. I don't think the filter is working.
kill_process kill_thread errno
Wed Dec 13 19:24:40 EST 2017
--summary -i
Event Summary Report
======================
total type
======================
170 SECCOMP
In the time it took to type the command 170 seccomp events were recorded
from firefox.
----
node=x2 type=SECCOMP msg=audit(12/13/2017 19:24:56.454:199666) : auid=sgrubb
uid=sgrubb gid=sgrubb ses=3
subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 pid=3394 comm=Web
Content exe=/usr/lib64/firefox/firefox sig=SIG0 arch=x86_64 syscall=stat
compat=0 ip=0x7f909c828635 code=trap
^^ trap. With the sheer amount of events being recorded, I think it's
necessary to add a sysctl file to systems to suppress the logging.
Especially when you consider that systemd-journal also gratuitously grabs
audit logs and sends them to rsyslog.

Looking at the kernel code, it looks like the actions_logged knob
isn't really intended to filter/drop seccomp events, but rather force
seccomp events to be loggged. Look at seccomp_log() to see what I
mean; there is still a call to audit_seccomp() at the end.

--
paul moore
www.paul-moore.com

Steve Grubb

2017-12-14 03:30:21 UTC

Post by Kees Cook

Post by Steve Grubb
Hello,
Over the last month, the amount of seccomp events in audit logs is
sky-rocketing. I have over a million events in the last 2 days. Most of
this is generated by firefox and qt webkit.
I am wondering if the audit package should ship a file for
/usr/lib/sysctl.d/60-auditd.conf
wherein it has
kernel.seccomp.actions_logged = kill_process kill_thread errno
Also, has anyone verified this sysctl is filtering audit events? Even with
the above, I have over a million events on a 4.14.3 kernel. Firefox alone
is generating over 50,000 events per hour.

I don't think you'd want to log errno -- AIUI, that's used regularly
by a lot of seccomp policy.

I'm not seeing any reporting errno. The ones reporting trap are coming in
over 50,000 per hour. I don't think the filter is working.
kill_process kill_thread errno
Wed Dec 13 19:24:40 EST 2017
--summary -i
Event Summary Report
======================
total type
======================
170 SECCOMP
In the time it took to type the command 170 seccomp events were recorded
from firefox.
----
auid=sgrubb uid=sgrubb gid=sgrubb ses=3
subj=unconfined_u:unconfined_r:unconfined_t:s0-s0:c0.c1023 pid=3394
comm=Web Content exe=/usr/lib64/firefox/firefox sig=SIG0 arch=x86_64
syscall=stat compat=0 ip=0x7f909c828635 code=trap
^^ trap. With the sheer amount of events being recorded, I think it's
necessary to add a sysctl file to systems to suppress the logging.
Especially when you consider that systemd-journal also gratuitously grabs
audit logs and sends them to rsyslog.

Looking at the kernel code, it looks like the actions_logged knob
isn't really intended to filter/drop seccomp events,

That's unfortunate. I thought this was a way to suppress generation of events. We
have a requirement that audit events be selective by the administrator. We need a
knob to drop some events. I guess, the only knob right now is the exclude filter.
That is probably too course.

Post by Paul Moore
but rather force seccomp events to be loggged. Look at seccomp_log() to see
what I mean; there is still a call to audit_seccomp() at the end.

Hmm. What do we do? I have more than a million seccomp events in 2 days. I'm
getting 14 events a second for seccomp just using my system. It's flooding
everything. I mean maybe the firefox and webkit people meant well using seccomp,
but how does a normal user get control back of their logs?

Thanks,
-Steve

Paul Moore

2017-12-14 12:42:26 UTC

...

Post by Paul Moore
Looking at the kernel code, it looks like the actions_logged knob
isn't really intended to filter/drop seccomp events,

That's unfortunate. I thought this was a way to suppress generation of
events. We have a requirement that audit events be selective by the
administrator. We need a knob to drop some events. I guess, the only knob
right now is the exclude filter. That is probably too course.

Post by Paul Moore
but rather force seccomp events to be loggged. Look at seccomp_log() to
see what I mean; there is still a call to audit_seccomp() at the end.

Hmm. What do we do?

I imagine we could put together a rather coarse grained action filter,
similar to what we have with "actions_logged" (maybe
"actions_silent"?), and perhaps add some additional audit filters for
seccomp for those who happen to have audit enabled. Both should be
relatively easy, the "actions_silent" field especially so.

--
paul moore
www.paul-moore.com

Steve Grubb

2017-12-14 15:29:38 UTC

Post by Paul Moore
Looking at the kernel code, it looks like the actions_logged knob
isn't really intended to filter/drop seccomp events,

That's unfortunate. I thought this was a way to suppress generation of
events. We have a requirement that audit events be selective by the
administrator. We need a knob to drop some events. I guess, the only knob
right now is the exclude filter. That is probably too course.

Post by Paul Moore
but rather force seccomp events to be loggged. Look at seccomp_log() to
see what I mean; there is still a call to audit_seccomp() at the end.

Hmm. What do we do?

OK. That would be helpful. This is eating up my log space. The biggest offenders
seem to be doing trap kind of events. I suppose if an errno was returned the
program would respond by erroring out. But since its a trap, I suspect something
looks around at data and then OK's it to proceed on which results in another trap.

-Steve

Tyler Hicks

2017-12-14 15:04:48 UTC

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

Post by Steve Grubb
Also, has anyone verified this sysctl is filtering audit events? Even
with the above, I have over a million events on a 4.14.3 kernel. Firefox
alone is generating over 50,000 events per hour.

Yes. I tested it a lot as the changes were being upstreamed and again
when I backported the changes to Ubuntu releases 17.10, 17.04, and 16.04
LTS. Those kernels have been released for a month and a half now and I
haven't heard of any issues.

I'll install a Fedora VM and see if I can determine what's happening.

Tyler

Post by Steve Grubb

Thanks,
-Steve

Steve Grubb

2017-12-14 15:19:41 UTC

Post by Steve Grubb
Over the last month, the amount of seccomp events in audit logs is
sky-rocketing. I have over a million events in the last 2 days. Most of
this is generated by firefox and qt webkit.
I am wondering if the audit package should ship a file for
/usr/lib/sysctl.d/60-auditd.conf
wherein it has
kernel.seccomp.actions_logged = kill_process kill_thread errno

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Apparently I misunderstood this as a filter for audit events to be blocked.

Post by Tyler Hicks
I'll install a Fedora VM and see if I can determine what's happening.

I'm testing on F27 if that matters. It would appear to me that seccomp filtering is
becoming more used and is now dominating the events going into the audit logs.

# uptime
10:17:34 up 1:27, 1 user, load average: 0.23, 0.39, 0.53

[***@x2 ~]# ausearch --start today --raw | aureport --event --summary -i

Event Summary Report
======================
total type
======================
35561 SECCOMP
1041 SYSCALL
134 SERVICE_START
75 NETFILTER_CFG
46 SERVICE_STOP
46 CONFIG_CHANGE
24 AVC

All of the seccomp events are trap types.

-Steve

Tyler Hicks

2017-12-14 23:06:30 UTC

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.

I think the easiest description of the logic is in the commit message of
59f5cf44a38284eb9e76270c786fb6cc62ef8ac4:

if action == RET_ALLOW:
do not log
else if action == RET_KILL && RET_KILL in actions_logged:
log
else if action == RET_LOG && RET_LOG in actions_logged:
log
else if filter-requests-logging && action in actions_logged:
log
else if audit_enabled && process-is-being-audited:
log
else:
do not log

I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.

The reason is because I didn't get clear direction from the audit
folks about to do when audit is enabled and the process is being audited
and, therefore, I didn't feel comfortable rocking the boat. In that
situation, the decision to log is the same as it was in earlier kernels.
Specifically, you're hitting the last "else if" conditional in the
pseudocode above.

If you're happy with having the actions_logged sysctl control whether or
not to log seccomp actions taken for processes that are being audited,
then I think the following (untested) patch should do exactly what you
want. I imagine that you'd also want seccomp to emit audit events
whenever the value of the actions_logged sysctl is changed, which should
be pretty easy to do.

I hope this helps!

Tyler

diff --git a/include/linux/audit.h b/include/linux/audit.h
index af410d9..095b5dd 100644
--- a/include/linux/audit.h
+++ b/include/linux/audit.h
@@ -304,12 +304,6 @@ static inline void audit_inode_child(struct inode *parent,
}
void audit_core_dumps(long signr);

-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{
- if (audit_enabled && unlikely(!audit_dummy_context()))
- __audit_seccomp(syscall, signr, code);
-}
-
static inline void audit_ptrace(struct task_struct *t)
{
if (unlikely(!audit_dummy_context()))
@@ -502,8 +496,6 @@ static inline void audit_core_dumps(long signr)
{ }
static inline void __audit_seccomp(unsigned long syscall, long signr, int code)
{ }
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{ }
static inline int auditsc_get_stamp(struct audit_context *ctx,
struct timespec64 *t, unsigned int *serial)
{
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 5f0dfb2ab..914a707 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -590,12 +590,6 @@ static inline void seccomp_log(unsigned long syscall, long signr, u32 action,
*/
if (log)
return __audit_seccomp(syscall, signr, action);
-
- /*
- * Let the audit subsystem decide if the action should be audited based
- * on whether the current task itself is being audited.
- */
- return audit_seccomp(syscall, signr, action);
}

/*

Kees Cook

2017-12-14 23:16:18 UTC

Post by Tyler Hicks
The reason is because I didn't get clear direction from the audit
folks about to do when audit is enabled and the process is being audited
and, therefore, I didn't feel comfortable rocking the boat. In that
situation, the decision to log is the same as it was in earlier kernels.
Specifically, you're hitting the last "else if" conditional in the
pseudocode above.

Yeah, same for me: it's been entirely unclear what the desired
combination of audit vs seccomp should be. It seems like it should be
reporting everything when auditing a specific process, and then ...
something else? ... in the global context.

Post by Tyler Hicks
If you're happy with having the actions_logged sysctl control whether or
not to log seccomp actions taken for processes that are being audited,
then I think the following (untested) patch should do exactly what you
want. I imagine that you'd also want seccomp to emit audit events
whenever the value of the actions_logged sysctl is changed, which should
be pretty easy to do.
I hope this helps!
Tyler
diff --git a/include/linux/audit.h b/include/linux/audit.h
index af410d9..095b5dd 100644
--- a/include/linux/audit.h
+++ b/include/linux/audit.h
@@ -304,12 +304,6 @@ static inline void audit_inode_child(struct inode *parent,
}
void audit_core_dumps(long signr);
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{
- if (audit_enabled && unlikely(!audit_dummy_context()))
- __audit_seccomp(syscall, signr, code);
-}
-
static inline void audit_ptrace(struct task_struct *t)
{
if (unlikely(!audit_dummy_context()))
@@ -502,8 +496,6 @@ static inline void audit_core_dumps(long signr)
{ }
static inline void __audit_seccomp(unsigned long syscall, long signr, int code)
{ }
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{ }
static inline int auditsc_get_stamp(struct audit_context *ctx,
struct timespec64 *t, unsigned int *serial)
{
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 5f0dfb2ab..914a707 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -590,12 +590,6 @@ static inline void seccomp_log(unsigned long syscall, long signr, u32 action,
*/
if (log)
return __audit_seccomp(syscall, signr, action);
-
- /*
- * Let the audit subsystem decide if the action should be audited based
- * on whether the current task itself is being audited.
- */
- return audit_seccomp(syscall, signr, action);
}
/*

If audit folks are happy with this, I am too. :)

-Kees

--
Kees Cook
Pixel Security

Paul Moore

2017-12-15 14:08:05 UTC

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.
I think the easiest description of the logic is in the commit message of
do not log
log
log
log
log
do not log
I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.
The reason is because I didn't get clear direction from the audit
folks about to do when audit is enabled and the process is being audited
and, therefore, I didn't feel comfortable rocking the boat. In that
situation, the decision to log is the same as it was in earlier kernels.
Specifically, you're hitting the last "else if" conditional in the
pseudocode above.
If you're happy with having the actions_logged sysctl control whether or
not to log seccomp actions taken for processes that are being audited,
then I think the following (untested) patch should do exactly what you
want. I imagine that you'd also want seccomp to emit audit events
whenever the value of the actions_logged sysctl is changed, which should
be pretty easy to do.
I hope this helps!
Tyler
diff --git a/include/linux/audit.h b/include/linux/audit.h
index af410d9..095b5dd 100644
--- a/include/linux/audit.h
+++ b/include/linux/audit.h
@@ -304,12 +304,6 @@ static inline void audit_inode_child(struct inode *parent,
}
void audit_core_dumps(long signr);
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{
- if (audit_enabled && unlikely(!audit_dummy_context()))
- __audit_seccomp(syscall, signr, code);
-}
-

Looks good to me but two things:

* Change the name of __audit_seccomp() to audit_seccomp() since we don't
have two functions anymore.

* Are we sure about removing the audit_enabled check? People got pretty
upset when it wasn't there in the past.

Post by Tyler Hicks
static inline void audit_ptrace(struct task_struct *t)
{
if (unlikely(!audit_dummy_context()))
@@ -502,8 +496,6 @@ static inline void audit_core_dumps(long signr)
{ }
static inline void __audit_seccomp(unsigned long syscall, long signr, int code)
{ }
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{ }
static inline int auditsc_get_stamp(struct audit_context *ctx,
struct timespec64 *t, unsigned int *serial)
{
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 5f0dfb2ab..914a707 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -590,12 +590,6 @@ static inline void seccomp_log(unsigned long syscall,
long signr, u32 action,
*/
if (log)
return __audit_seccomp(syscall, signr, action);
-
- /*
- * Let the audit subsystem decide if the action should be audited based
- * on whether the current task itself is being audited.
- */
- return audit_seccomp(syscall, signr, action);
}
/*
--
Linux-audit mailing list
https://www.redhat.com/mailman/listinfo/linux-audit

Tyler Hicks

2017-12-15 15:47:14 UTC

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.
I think the easiest description of the logic is in the commit message of
do not log
log
log
log
log
do not log
I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.
The reason is because I didn't get clear direction from the audit
folks about to do when audit is enabled and the process is being audited
and, therefore, I didn't feel comfortable rocking the boat. In that
situation, the decision to log is the same as it was in earlier kernels.
Specifically, you're hitting the last "else if" conditional in the
pseudocode above.
If you're happy with having the actions_logged sysctl control whether or
not to log seccomp actions taken for processes that are being audited,
then I think the following (untested) patch should do exactly what you
want. I imagine that you'd also want seccomp to emit audit events
whenever the value of the actions_logged sysctl is changed, which should
be pretty easy to do.
I hope this helps!
Tyler
diff --git a/include/linux/audit.h b/include/linux/audit.h
index af410d9..095b5dd 100644
--- a/include/linux/audit.h
+++ b/include/linux/audit.h
@@ -304,12 +304,6 @@ static inline void audit_inode_child(struct inode *parent,
}
void audit_core_dumps(long signr);
-static inline void audit_seccomp(unsigned long syscall, long signr, int code)
-{
- if (audit_enabled && unlikely(!audit_dummy_context()))
- __audit_seccomp(syscall, signr, code);
-}
-

* Change the name of __audit_seccomp() to audit_seccomp() since we don't
have two functions anymore.
* Are we sure about removing the audit_enabled check? People got pretty
upset when it wasn't there in the past.

Do you have any references to the complaints so that we can understand
them better? I remember being surprised by commit 96368701 adding the
audit_enabled check (my fault for not watching the list closer) and
having to revert it in Ubuntu with a distro patch.

After sleeping on it for a night, I'm now unsure if the patch I sent in
this thread is what you guys really want. I'll go back to talking in
pseudocode. This is what we have in 4.14:

if action == RET_ALLOW:
do not log
else if action == RET_KILL && RET_KILL in actions_logged:
log
else if action == RET_LOG && RET_LOG in actions_logged:
log
else if filter-requests-logging && action in actions_logged:
log
else if audit_enabled && process-is-being-audited:
log
else:
do not log

This is what the patch in this thread does:

--- a/seccomp-log.pseudo
+++ b/seccomp-log.pseudo
@@ -6,7 +6,5 @@
log
else if filter-requests-logging && action in actions_logged:
log
- else if audit_enabled && process-is-being-audited:
- log
else:
do not log

Instead of that change, now I'm wondering if this is what you really
want:

--- a/seccomp-log.pseudo
+++ b/seccomp-log.pseudo
@@ -6,7 +6,8 @@
log
else if filter-requests-logging && action in actions_logged:
log
- else if audit_enabled && process-is-being-audited:
+ else if audit_enabled && process-is-being-audited &&
+ action in actions_logged:
log
else:
do not log

After refactoring the 'action in actions_logged' check, it would leave
us with this:

if action == RET_ALLOW:
do not log
else if action not in actions_logged:
do not log
else if action == RET_KILL:
log
else if action == RET_LOG:
log
else if filter-requests-logging:
log
else if audit_enabled && process-is-being-audited:
log
else:
do not log

Tyler

Steve Grubb

2017-12-15 16:09:39 UTC

Post by Paul Moore
* Change the name of __audit_seccomp() to audit_seccomp() since we don't
have two functions anymore.
* Are we sure about removing the audit_enabled check? People got pretty
upset when it wasn't there in the past.

Yeah, this would let us drop the trap return. While errno can lead to a lot of
logging, in practice I just don't see them very often if ever.

-Steve

Post by Tyler Hicks
log
log
log
log
do not log

Paul Moore

2017-12-15 20:54:16 UTC

...

Post by Paul Moore
* Are we sure about removing the audit_enabled check? People got pretty
upset when it wasn't there in the past.

I'm somewhat limited at the moment (replying by phone) so I can't my find
it for you, but check the audit mailing list for the discussion.

Steve Grubb

2017-12-15 16:02:19 UTC

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.
I think the easiest description of the logic is in the commit message of
do not log
log
log
log
log
do not log
I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.

Yes, exactly. I have been experiencing large amounts of SECCOMP events
starting with qt webkit in kde and thought 4.14 would finally let me tame
those events. I have opened a couple bz asking developers if they really meant
to go live with a policy that is experiencing so many denials. But the
consensus is this is intended. (But I think they also have not actually tried
to use their audit logs.)

And here I thought you were also seeing large numbers of seccomp events and
were making a way to control what gets logged. In any event, I think we better
understand each other now. :-)

OK. Great. With developers starting to use the trap return value, audit logs
are getting swamped by benign events. We truly need a knob to eliminate the
noise from the signal.

Post by Tyler Hicks
I imagine that you'd also want seccomp to emit audit events whenever the
value of the actions_logged sysctl is changed, which should be pretty easy
to do.

Sure. If you want to add it, then it should be roughly like this:

struct tty_struct *tty;
const struct cred *cred;
struct audit_buffer *ab;
char comm[sizeof(current->comm)];

ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
if (unlikely(!ab))
return;

cred = current_cred();
tty = audit_get_tty(current);
audit_log_format(ab, "pid=%d uid=%u auid=%u tty=%s ses=%u",
task_tgid_nr(current),
from_kuid(&init_user_ns, cred->uid),
from_kuid(&init_user_ns,
audit_get_loginuid(current)),
tty ? tty_name(tty) : "(none)",
audit_get_sessionid(current));
audit_put_tty(tty);
audit_log_task_context(ab);
audit_log_format(ab, " comm=");
audit_log_untrustedstring(ab, get_task_comm(comm, current));
audit_log_d_path_exe(ab, current->mm);
audit_log_format(ab, "op=seccomp-logging");

<You can log the new value here if you wish - just can't have spaces in the
value. Numbers or mask is fine.>

audit_log_format(ab, " res=%u", res);

where res above is a 1 for success and 0 for failure. Failure is likely to be
due to not having the capability that allows setting the sysctl.

Post by Tyler Hicks
I hope this helps!

Thanks!

-Steve

Post by Tyler Hicks
diff --git a/include/linux/audit.h b/include/linux/audit.h
index af410d9..095b5dd 100644
--- a/include/linux/audit.h
+++ b/include/linux/audit.h
@@ -304,12 +304,6 @@ static inline void audit_inode_child(struct inode
*parent, }
void audit_core_dumps(long signr);
-static inline void audit_seccomp(unsigned long syscall, long signr, int
code) -{
- if (audit_enabled && unlikely(!audit_dummy_context()))
- __audit_seccomp(syscall, signr, code);
-}
-
static inline void audit_ptrace(struct task_struct *t)
{
if (unlikely(!audit_dummy_context()))
@@ -502,8 +496,6 @@ static inline void audit_core_dumps(long signr)
{ }
static inline void __audit_seccomp(unsigned long syscall, long signr, int
code) { }
-static inline void audit_seccomp(unsigned long syscall, long signr, int
code) -{ }
static inline int auditsc_get_stamp(struct audit_context *ctx,
struct timespec64 *t, unsigned int *serial)
{
diff --git a/kernel/seccomp.c b/kernel/seccomp.c
index 5f0dfb2ab..914a707 100644
--- a/kernel/seccomp.c
+++ b/kernel/seccomp.c
@@ -590,12 +590,6 @@ static inline void seccomp_log(unsigned long syscall,
long signr, u32 action, */
if (log)
return __audit_seccomp(syscall, signr, action);
-
- /*
- * Let the audit subsystem decide if the action should be audited based
- * on whether the current task itself is being audited.
- */
- return audit_seccomp(syscall, signr, action);
}
/*

Steve Grubb

2018-01-02 20:03:32 UTC

Hello,

I know people have been busy with the holidays and things...but I just wanted
to mention I'm still seeing 100's of thousands of seccomp events hitting the
audit logs every day.

# ausearch --start today -m seccomp --raw | aureport -x --summary

Executable Summary Report
=================================
total file
=================================
209843 /usr/lib64/firefox/firefox
2196 /usr/lib64/qt5/libexec/QtWebEngineProcess

Has anyone looked at it beyond pseudo code?

-Steve

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.
I think the easiest description of the logic is in the commit message of
do not log
log
log
log
log
do not log
I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.

Yes, exactly. I have been experiencing large amounts of SECCOMP events
starting with qt webkit in kde and thought 4.14 would finally let me tame
those events. I have opened a couple bz asking developers if they really
meant to go live with a policy that is experiencing so many denials. But
the consensus is this is intended. (But I think they also have not
actually tried to use their audit logs.)

And here I thought you were also seeing large numbers of seccomp events and
were making a way to control what gets logged. In any event, I think we
better understand each other now. :-)

OK. Great. With developers starting to use the trap return value, audit
logs are getting swamped by benign events. We truly need a knob to
eliminate the noise from the signal.

Post by Tyler Hicks
I imagine that you'd also want seccomp to emit audit events whenever the
value of the actions_logged sysctl is changed, which should be pretty easy
to do.

struct tty_struct *tty;
const struct cred *cred;
struct audit_buffer *ab;
char comm[sizeof(current->comm)];
ab = audit_log_start(NULL, GFP_KERNEL, AUDIT_CONFIG_CHANGE);
if (unlikely(!ab))
return;
cred = current_cred();
tty = audit_get_tty(current);
audit_log_format(ab, "pid=%d uid=%u auid=%u tty=%s ses=%u",
task_tgid_nr(current),
from_kuid(&init_user_ns, cred->uid),
from_kuid(&init_user_ns,
audit_get_loginuid(current)),
tty ? tty_name(tty) : "(none)",
audit_get_sessionid(current));
audit_put_tty(tty);
audit_log_task_context(ab);
audit_log_format(ab, " comm=");
audit_log_untrustedstring(ab, get_task_comm(comm, current));
audit_log_d_path_exe(ab, current->mm);
audit_log_format(ab, "op=seccomp-logging");
<You can log the new value here if you wish - just can't have spaces in the
value. Numbers or mask is fine.>
audit_log_format(ab, " res=%u", res);
where res above is a 1 for success and 0 for failure. Failure is likely to
be due to not having the capability that allows setting the sysctl.

Post by Tyler Hicks
I hope this helps!

Thanks!
-Steve

--
Linux-audit mailing list
https://www.redhat.com/mailman/listinfo/linux-audit

Tyler Hicks

2018-01-03 02:52:55 UTC

Post by Steve Grubb
Hello,
I know people have been busy with the holidays and things...but I just wanted
to mention I'm still seeing 100's of thousands of seccomp events hitting the
audit logs every day.
# ausearch --start today -m seccomp --raw | aureport -x --summary
Executable Summary Report
=================================
total file
=================================
209843 /usr/lib64/firefox/firefox
2196 /usr/lib64/qt5/libexec/QtWebEngineProcess
Has anyone looked at it beyond pseudo code?

I started to throw together a quick couple of patches prior to the
holidays but didn't finish. Things aren't looking good for the next few
weeks for me so someone else should take over if it is important for 4.16.

Tyler

Post by Steve Grubb
-Steve

I agree with Kees here. IMO, you only want "kill_process kill_thread"
which is the default.

The default appears to be all of the types of events without setting
kernel.seccomp.actions_logged.

Ah, right. I didn't correctly remember the final implementation details.
The default sysctl setting is to allow all actions except for RET_ALLOW
to be logged.
I think the easiest description of the logic is in the commit message of
do not log
log
log
log
log
do not log
I think I originally misunderstood your first email in this thread. I
thought you were saying that you were experiencing more seccomp audit
events in 4.14 versus 4.13 and that you felt a regression had been
introduced. After rereading, I think you're asking why you're getting
seccomp RET_TRAP actions logged even though "trap" isn't in the
actions_logged sysctl.

Yes, exactly. I have been experiencing large amounts of SECCOMP events
starting with qt webkit in kde and thought 4.14 would finally let me tame
those events. I have opened a couple bz asking developers if they really
meant to go live with a policy that is experiencing so many denials. But
the consensus is this is intended. (But I think they also have not
actually tried to use their audit logs.)

And here I thought you were also seeing large numbers of seccomp events and
were making a way to control what gets logged. In any event, I think we
better understand each other now. :-)

OK. Great. With developers starting to use the trap return value, audit
logs are getting swamped by benign events. We truly need a knob to
eliminate the noise from the signal.

Post by Tyler Hicks
I imagine that you'd also want seccomp to emit audit events whenever the
value of the actions_logged sysctl is changed, which should be pretty easy
to do.

Post by Tyler Hicks
I hope this helps!

Thanks!
-Steve

based

Post by Tyler Hicks
- * on whether the current task itself is being audited.
- */
- return audit_seccomp(syscall, signr, action);
}
/*

--
Linux-audit mailing list
https://www.redhat.com/mailman/listinfo/linux-audit

Paul Moore

2018-01-03 14:25:12 UTC

This is also on my todo list, but it sits behind fixing one last
libseccomp bug and getting a new release out. I made some good
progress on the libseccomp bug right before the holiday, but I think
there is still a days worth of work left before it is ready to be
merged. I'm also traveling for the next week so I doubt I'll have any
serious time to devote to the kernel patch(es).

I can't remember what Tyler's last thought was on the logic, but I
imagine I'll just wait until I see some patches to review/merge, or I
can go back in the thread if I happen to have time before anyone else.

Also, to set expectations, since we are currently at -rc6, this is
likely going to need to wait until 4.17 at the earliest as I generally
don't like merging new functionality in the last week or two before
the merge window.

Also (part two), we should add a test case to the audit-testsuite for
any new knobs that affect the SECCOMP records.

--
paul moore
www.paul-moore.com

Steve Grubb

2018-04-17 22:54:28 UTC

Hello,

Ping? SECCOMP events are still flooding the system. Can we do something
hackish to turn this off until a better solution can be created?

Thanks,
-Steve

Post by Steve Grubb
Hello,
I know people have been busy with the holidays and things...but I just
wanted to mention I'm still seeing 100's of thousands of seccomp events
hitting the audit logs every day.
# ausearch --start today -m seccomp --raw | aureport -x --summary
Executable Summary Report
=================================
total file
=================================
209843 /usr/lib64/firefox/firefox
2196 /usr/lib64/qt5/libexec/QtWebEngineProcess
Has anyone looked at it beyond pseudo code?

This is also on my todo list, but it sits behind fixing one last
libseccomp bug and getting a new release out. I made some good
progress on the libseccomp bug right before the holiday, but I think
there is still a days worth of work left before it is ready to be
merged. I'm also traveling for the next week so I doubt I'll have any
serious time to devote to the kernel patch(es).
I can't remember what Tyler's last thought was on the logic, but I
imagine I'll just wait until I see some patches to review/merge, or I
can go back in the thread if I happen to have time before anyone else.
Also, to set expectations, since we are currently at -rc6, this is
likely going to need to wait until 4.17 at the earliest as I generally
don't like merging new functionality in the last week or two before
the merge window.
Also (part two), we should add a test case to the audit-testsuite for
any new knobs that affect the SECCOMP records.

Paul Moore

2018-04-18 01:57:10 UTC

Post by Steve Grubb
Hello,
Ping? SECCOMP events are still flooding the system. Can we do something
hackish to turn this off until a better solution can be created?

Pong?

The only workarounds I can think of would be to disable audit or
create a filter rule excluding auditing for the noisy process. I've
never tried the latter, but I'm pretty sure it would work.

Post by Steve Grubb
Hello,
I know people have been busy with the holidays and things...but I just
wanted to mention I'm still seeing 100's of thousands of seccomp events
hitting the audit logs every day.
# ausearch --start today -m seccomp --raw | aureport -x --summary
Executable Summary Report
=================================
total file
=================================
209843 /usr/lib64/firefox/firefox
2196 /usr/lib64/qt5/libexec/QtWebEngineProcess
Has anyone looked at it beyond pseudo code?

This is also on my todo list, but it sits behind fixing one last
libseccomp bug and getting a new release out. I made some good
progress on the libseccomp bug right before the holiday, but I think
there is still a days worth of work left before it is ready to be
merged. I'm also traveling for the next week so I doubt I'll have any
serious time to devote to the kernel patch(es).
I can't remember what Tyler's last thought was on the logic, but I
imagine I'll just wait until I see some patches to review/merge, or I
can go back in the thread if I happen to have time before anyone else.
Also, to set expectations, since we are currently at -rc6, this is
likely going to need to wait until 4.17 at the earliest as I generally
don't like merging new functionality in the last week or two before
the merge window.
Also (part two), we should add a test case to the audit-testsuite for
any new knobs that affect the SECCOMP records.

--
paul moore
www.paul-moore.com

Tyler Hicks

2018-04-25 00:00:46 UTC

Post by Steve Grubb
Hello,
Ping? SECCOMP events are still flooding the system. Can we do something
hackish to turn this off until a better solution can be created?

Pong?
The only workarounds I can think of would be to disable audit or
create a filter rule excluding auditing for the noisy process. I've
never tried the latter, but I'm pretty sure it would work.

I've pushed two branches which have slightly different behaviors. They
only differ by the last patch in each branch. The tree is here:

https://git.kernel.org/pub/scm/linux/kernel/git/tyhicks/linux.git/

1) seccomp-auditing/option-1-consistent-behavior
This branch removes all special casing around audited processes. The
decision on whether or not to audit an action no longer considers
whether or not the process is being audited. RET_TRAP, RET_TRACE,
and RET_ERRNO actions will only be logged if the application loads
the filter with the SECCOMP_FILTER_FLAG_LOG bit set. The admin has
the final say via the kernel.seccomp.actions_logged sysctl.

2) seccomp-auditing/option-2-honor-sysctl
This branch continues to special case audited processes. The decision
to log RET_TRAP, RET_TRACE, and RET_ERRNO actions depends on if the
SECCOMP_FILTER_FLAG_LOG bit being set OR if the process is being
audited. The admin has the final say via the
kernel.seccomp.actions_logged sysctl.

I prefer option #1. Play with both implementations and let me know what
works best for your requirements. Both branches share the same
underlying patches for emitting audit records on writes to the
kernel.seccomp.actions_logged sysctl.

Tyler

Post by Steve Grubb
Hello,
I know people have been busy with the holidays and things...but I just
wanted to mention I'm still seeing 100's of thousands of seccomp events
hitting the audit logs every day.
# ausearch --start today -m seccomp --raw | aureport -x --summary
Executable Summary Report
=================================
total file
=================================
209843 /usr/lib64/firefox/firefox
2196 /usr/lib64/qt5/libexec/QtWebEngineProcess
Has anyone looked at it beyond pseudo code?

This is also on my todo list, but it sits behind fixing one last
libseccomp bug and getting a new release out. I made some good
progress on the libseccomp bug right before the holiday, but I think
there is still a days worth of work left before it is ready to be
merged. I'm also traveling for the next week so I doubt I'll have any
serious time to devote to the kernel patch(es).
I can't remember what Tyler's last thought was on the logic, but I
imagine I'll just wait until I see some patches to review/merge, or I
can go back in the thread if I happen to have time before anyone else.
Also, to set expectations, since we are currently at -rc6, this is
likely going to need to wait until 4.17 at the earliest as I generally
don't like merging new functionality in the last week or two before
the merge window.
Also (part two), we should add a test case to the audit-testsuite for
any new knobs that affect the SECCOMP records.

Paul Moore

2018-04-26 14:41:24 UTC