Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 7:46 PM, Steven Rostedt wrote: > On Fri, 14 Aug 2020 19:22:13 +0200 > peter enderborg wrote: > >> On 8/14/20 7:08 PM, Stephen Smalley wrote: >>> On Fri, Aug 14, 2020 at 1:07 PM peter enderborg >>> wrote: On 8/14/20 6:51 PM, Stephen Smalley wrote: > On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen > wrote: >> On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley >> wrote: >>> An explanation here of how one might go about decoding audited and >>> tclass would be helpful to users (even better would be a script to do it >>> for them). Again, I know how to do that but not everyone using >>> perf/ftrace will. >> What about something along those lines: >> >> The tclass value can be mapped to a class by searching >> security/selinux/flask.h. The audited value is a bit field of the >> permissions described in security/selinux/av_permissions.h for the >> corresponding class. > Sure, I guess that works. Would be nice if we just included the class > and permission name(s) in the event itself but I guess you viewed that > as too heavyweight? The class name is added in part 2. Im not sure how a proper format for permission would look like in trace terms. It is a list, right? >>> Yes. See avc_audit_pre_callback() for example code to log the permission >>> names. >> I wrote about that on some of the previous sets. The problem is that trace >> format is quite fixed. So it is lists are not >> that easy to handle if you want to filter in them. You can have a trace >> event for each of them. You can also add >> additional trace event "selinux_audied_permission" for each permission. With >> that you can filter out tclass or permissions. >> >> But the basic thing we would like at the moment is a event that we can debug >> in user space. > We have a trace_seq p helper, that lets you create strings in > TP_printk(). I should document this more. Thus you can do: > > extern const char *audit_perm_to_name(struct trace_seq *p, u16 class, u32 > audited); > #define __perm_to_name(p, class, audited) audit_perm_to_name(p, class, > audited) > > TP_printk("tclass=%u audited=%x (%s)", > __entry->tclass, > __entry->audited, > __perm_to_name(__entry->tclass, __entry->audited)) > > > const char *audit_perm_to_name(struct trace_seq *p, u16 tclass, u32 av) > { > const char *ret = trace_seq_buffer_ptr(p); > int i, perm; > > ( some check for tclass integrity here) > > perms = secclass_map[tclass-1].perms; > > i = 0; > perm = 1; > while (i < (sizeof(av) * 8)) { > if ((perm & av) && perms[i]) { > trace_seq_printf(p, " %s", perms[i]); > av &= ~perm; > } > i++; > perm <<= 1; > } > > return ret; > } > > Note, this wont work for perf and trace-cmd as it wouldn't know how to > parse it, but if the tclass perms are stable, you could create a plugin > to libtraceevent that can do the above as well. > > -- Steve That works fine. I will do this as third patch in our patch-set. But I think we also should export the permission-map somewhere. I don’t think there is any good place for it in tracefs. So selinuxfs or debugfs might do? And I think it is more useful to print what is denied than what is audited but that does not match the trace event name.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 7:46 PM, Steven Rostedt wrote: > On Fri, 14 Aug 2020 19:22:13 +0200 > peter enderborg wrote: > >> On 8/14/20 7:08 PM, Stephen Smalley wrote: >>> On Fri, Aug 14, 2020 at 1:07 PM peter enderborg >>> wrote: On 8/14/20 6:51 PM, Stephen Smalley wrote: > On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen > wrote: >> On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley >> wrote: >>> An explanation here of how one might go about decoding audited and >>> tclass would be helpful to users (even better would be a script to do it >>> for them). Again, I know how to do that but not everyone using >>> perf/ftrace will. >> What about something along those lines: >> >> The tclass value can be mapped to a class by searching >> security/selinux/flask.h. The audited value is a bit field of the >> permissions described in security/selinux/av_permissions.h for the >> corresponding class. > Sure, I guess that works. Would be nice if we just included the class > and permission name(s) in the event itself but I guess you viewed that > as too heavyweight? The class name is added in part 2. Im not sure how a proper format for permission would look like in trace terms. It is a list, right? >>> Yes. See avc_audit_pre_callback() for example code to log the permission >>> names. >> I wrote about that on some of the previous sets. The problem is that trace >> format is quite fixed. So it is lists are not >> that easy to handle if you want to filter in them. You can have a trace >> event for each of them. You can also add >> additional trace event "selinux_audied_permission" for each permission. With >> that you can filter out tclass or permissions. >> >> But the basic thing we would like at the moment is a event that we can debug >> in user space. > We have a trace_seq p helper, that lets you create strings in > TP_printk(). I should document this more. Thus you can do: > > extern const char *audit_perm_to_name(struct trace_seq *p, u16 class, u32 > audited); > #define __perm_to_name(p, class, audited) audit_perm_to_name(p, class, > audited) > > TP_printk("tclass=%u audited=%x (%s)", > __entry->tclass, > __entry->audited, > __perm_to_name(__entry->tclass, __entry->audited)) > > > const char *audit_perm_to_name(struct trace_seq *p, u16 tclass, u32 av) > { > const char *ret = trace_seq_buffer_ptr(p); > int i, perm; > > ( some check for tclass integrity here) > > perms = secclass_map[tclass-1].perms; > > i = 0; > perm = 1; > while (i < (sizeof(av) * 8)) { > if ((perm & av) && perms[i]) { > trace_seq_printf(p, " %s", perms[i]); > av &= ~perm; > } > i++; > perm <<= 1; > } > > return ret; > } > > Note, this wont work for perf and trace-cmd as it wouldn't know how to > parse it, but if the tclass perms are stable, you could create a plugin > to libtraceevent that can do the above as well. > > -- Steve Something like: while (i < (sizeof(av) * 8)) { if ((perm & av) && perms[i]) { if (!(perm & avdenied)) trace_seq_printf(p, " %s", perms[i]); else trace_seq_printf(p, " !%s", perms[i]); av &= ~perm; And you get information about denied too.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Fri, 14 Aug 2020 20:50:47 +0200 peter enderborg wrote: > On 8/14/20 8:30 PM, Steven Rostedt wrote: > > On Fri, 14 Aug 2020 20:06:34 +0200 > > peter enderborg wrote: > > > >> Im find with that, but then you can not do filtering? I would be > >> pretty neat with a filter saying tclass=file permission=write. > >> > > Well, if the mapping is stable, you could do: > > > > (tclass == 6) && (audited & 0x4) > > It does not happen to exist a hook for translate strings to numeric values > when inserting filter? > How would you imagine such a hook existing? Something that would be specific to each trace event class, where you can register at boot up a mapping of names to values? Or a function that would translate it? -- Steve
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 8:30 PM, Steven Rostedt wrote: > On Fri, 14 Aug 2020 20:06:34 +0200 > peter enderborg wrote: > >> Im find with that, but then you can not do filtering? I would be >> pretty neat with a filter saying tclass=file permission=write. >> > Well, if the mapping is stable, you could do: > > (tclass == 6) && (audited & 0x4) It does not happen to exist a hook for translate strings to numeric values when inserting filter? > -- Steve
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Fri, 14 Aug 2020 20:06:34 +0200 peter enderborg wrote: > Im find with that, but then you can not do filtering? I would be > pretty neat with a filter saying tclass=file permission=write. > Well, if the mapping is stable, you could do: (tclass == 6) && (audited & 0x4) -- Steve
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 7:46 PM, Steven Rostedt wrote: > On Fri, 14 Aug 2020 19:22:13 +0200 > peter enderborg wrote: > >> On 8/14/20 7:08 PM, Stephen Smalley wrote: >>> On Fri, Aug 14, 2020 at 1:07 PM peter enderborg >>> wrote: On 8/14/20 6:51 PM, Stephen Smalley wrote: > On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen > wrote: >> On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley >> wrote: >>> An explanation here of how one might go about decoding audited and >>> tclass would be helpful to users (even better would be a script to do it >>> for them). Again, I know how to do that but not everyone using >>> perf/ftrace will. >> What about something along those lines: >> >> The tclass value can be mapped to a class by searching >> security/selinux/flask.h. The audited value is a bit field of the >> permissions described in security/selinux/av_permissions.h for the >> corresponding class. > Sure, I guess that works. Would be nice if we just included the class > and permission name(s) in the event itself but I guess you viewed that > as too heavyweight? The class name is added in part 2. Im not sure how a proper format for permission would look like in trace terms. It is a list, right? >>> Yes. See avc_audit_pre_callback() for example code to log the permission >>> names. >> I wrote about that on some of the previous sets. The problem is that trace >> format is quite fixed. So it is lists are not >> that easy to handle if you want to filter in them. You can have a trace >> event for each of them. You can also add >> additional trace event "selinux_audied_permission" for each permission. With >> that you can filter out tclass or permissions. >> >> But the basic thing we would like at the moment is a event that we can debug >> in user space. > We have a trace_seq p helper, that lets you create strings in > TP_printk(). I should document this more. Thus you can do: > > extern const char *audit_perm_to_name(struct trace_seq *p, u16 class, u32 > audited); > #define __perm_to_name(p, class, audited) audit_perm_to_name(p, class, > audited) > > TP_printk("tclass=%u audited=%x (%s)", > __entry->tclass, > __entry->audited, > __perm_to_name(__entry->tclass, __entry->audited)) > > > const char *audit_perm_to_name(struct trace_seq *p, u16 tclass, u32 av) > { > const char *ret = trace_seq_buffer_ptr(p); > int i, perm; > > ( some check for tclass integrity here) > > perms = secclass_map[tclass-1].perms; > > i = 0; > perm = 1; > while (i < (sizeof(av) * 8)) { > if ((perm & av) && perms[i]) { > trace_seq_printf(p, " %s", perms[i]); > av &= ~perm; > } > i++; > perm <<= 1; > } > > return ret; > } > > Note, this wont work for perf and trace-cmd as it wouldn't know how to > parse it, but if the tclass perms are stable, you could create a plugin > to libtraceevent that can do the above as well. > > -- Steve Im find with that, but then you can not do filtering? I would be pretty neat with a filter saying tclass=file permission=write.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Fri, 14 Aug 2020 19:22:13 +0200 peter enderborg wrote: > On 8/14/20 7:08 PM, Stephen Smalley wrote: > > On Fri, Aug 14, 2020 at 1:07 PM peter enderborg > > wrote: > >> On 8/14/20 6:51 PM, Stephen Smalley wrote: > >>> On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen > >>> wrote: > On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley > wrote: > > An explanation here of how one might go about decoding audited and > > tclass would be helpful to users (even better would be a script to do it > > for them). Again, I know how to do that but not everyone using > > perf/ftrace will. > What about something along those lines: > > The tclass value can be mapped to a class by searching > security/selinux/flask.h. The audited value is a bit field of the > permissions described in security/selinux/av_permissions.h for the > corresponding class. > >>> Sure, I guess that works. Would be nice if we just included the class > >>> and permission name(s) in the event itself but I guess you viewed that > >>> as too heavyweight? > >> The class name is added in part 2. Im not sure how a proper format for > >> permission > >> would look like in trace terms. It is a list, right? > > Yes. See avc_audit_pre_callback() for example code to log the permission > > names. > > I wrote about that on some of the previous sets. The problem is that trace > format is quite fixed. So it is lists are not > that easy to handle if you want to filter in them. You can have a trace event > for each of them. You can also add > additional trace event "selinux_audied_permission" for each permission. With > that you can filter out tclass or permissions. > > But the basic thing we would like at the moment is a event that we can debug > in user space. We have a trace_seq p helper, that lets you create strings in TP_printk(). I should document this more. Thus you can do: extern const char *audit_perm_to_name(struct trace_seq *p, u16 class, u32 audited); #define __perm_to_name(p, class, audited) audit_perm_to_name(p, class, audited) TP_printk("tclass=%u audited=%x (%s)", __entry->tclass, __entry->audited, __perm_to_name(__entry->tclass, __entry->audited)) const char *audit_perm_to_name(struct trace_seq *p, u16 tclass, u32 av) { const char *ret = trace_seq_buffer_ptr(p); int i, perm; ( some check for tclass integrity here) perms = secclass_map[tclass-1].perms; i = 0; perm = 1; while (i < (sizeof(av) * 8)) { if ((perm & av) && perms[i]) { trace_seq_printf(p, " %s", perms[i]); av &= ~perm; } i++; perm <<= 1; } return ret; } Note, this wont work for perf and trace-cmd as it wouldn't know how to parse it, but if the tclass perms are stable, you could create a plugin to libtraceevent that can do the above as well. -- Steve
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 7:08 PM, Stephen Smalley wrote: > On Fri, Aug 14, 2020 at 1:07 PM peter enderborg > wrote: >> On 8/14/20 6:51 PM, Stephen Smalley wrote: >>> On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen wrote: On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley wrote: > An explanation here of how one might go about decoding audited and > tclass would be helpful to users (even better would be a script to do it > for them). Again, I know how to do that but not everyone using > perf/ftrace will. What about something along those lines: The tclass value can be mapped to a class by searching security/selinux/flask.h. The audited value is a bit field of the permissions described in security/selinux/av_permissions.h for the corresponding class. >>> Sure, I guess that works. Would be nice if we just included the class >>> and permission name(s) in the event itself but I guess you viewed that >>> as too heavyweight? >> The class name is added in part 2. Im not sure how a proper format for >> permission >> would look like in trace terms. It is a list, right? > Yes. See avc_audit_pre_callback() for example code to log the permission > names. I wrote about that on some of the previous sets. The problem is that trace format is quite fixed. So it is lists are not that easy to handle if you want to filter in them. You can have a trace event for each of them. You can also add additional trace event "selinux_audied_permission" for each permission. With that you can filter out tclass or permissions. But the basic thing we would like at the moment is a event that we can debug in user space.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Fri, Aug 14, 2020 at 1:07 PM peter enderborg wrote: > > On 8/14/20 6:51 PM, Stephen Smalley wrote: > > On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen wrote: > >> On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley > >> wrote: > >>> An explanation here of how one might go about decoding audited and > >>> tclass would be helpful to users (even better would be a script to do it > >>> for them). Again, I know how to do that but not everyone using > >>> perf/ftrace will. > >> What about something along those lines: > >> > >> The tclass value can be mapped to a class by searching > >> security/selinux/flask.h. The audited value is a bit field of the > >> permissions described in security/selinux/av_permissions.h for the > >> corresponding class. > > Sure, I guess that works. Would be nice if we just included the class > > and permission name(s) in the event itself but I guess you viewed that > > as too heavyweight? > > The class name is added in part 2. Im not sure how a proper format for > permission > would look like in trace terms. It is a list, right? Yes. See avc_audit_pre_callback() for example code to log the permission names.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/14/20 6:51 PM, Stephen Smalley wrote: > On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen wrote: >> On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley >> wrote: >>> An explanation here of how one might go about decoding audited and >>> tclass would be helpful to users (even better would be a script to do it >>> for them). Again, I know how to do that but not everyone using >>> perf/ftrace will. >> What about something along those lines: >> >> The tclass value can be mapped to a class by searching >> security/selinux/flask.h. The audited value is a bit field of the >> permissions described in security/selinux/av_permissions.h for the >> corresponding class. > Sure, I guess that works. Would be nice if we just included the class > and permission name(s) in the event itself but I guess you viewed that > as too heavyweight? The class name is added in part 2. Im not sure how a proper format for permission would look like in trace terms. It is a list, right?
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Fri, Aug 14, 2020 at 9:05 AM Thiébaud Weksteen wrote: > > On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley > wrote: > > > > An explanation here of how one might go about decoding audited and > > tclass would be helpful to users (even better would be a script to do it > > for them). Again, I know how to do that but not everyone using > > perf/ftrace will. > > What about something along those lines: > > The tclass value can be mapped to a class by searching > security/selinux/flask.h. The audited value is a bit field of the > permissions described in security/selinux/av_permissions.h for the > corresponding class. Sure, I guess that works. Would be nice if we just included the class and permission name(s) in the event itself but I guess you viewed that as too heavyweight?
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On Thu, Aug 13, 2020 at 5:41 PM Stephen Smalley wrote: > > An explanation here of how one might go about decoding audited and > tclass would be helpful to users (even better would be a script to do it > for them). Again, I know how to do that but not everyone using > perf/ftrace will. What about something along those lines: The tclass value can be mapped to a class by searching security/selinux/flask.h. The audited value is a bit field of the permissions described in security/selinux/av_permissions.h for the corresponding class.
Re: [PATCH v2 1/2] selinux: add tracepoint on denials
On 8/13/20 10:48 AM, Thiébaud Weksteen wrote: The audit data currently captures which process and which target is responsible for a denial. There is no data on where exactly in the process that call occurred. Debugging can be made easier by being able to reconstruct the unified kernel and userland stack traces [1]. Add a tracepoint on the SELinux denials which can then be used by userland (i.e. perf). Although this patch could manually be added by each OS developer to trouble shoot a denial, adding it to the kernel streamlines the developers workflow. It is possible to use perf for monitoring the event: # perf record -e avc:selinux_audited -g -a ^C # perf report -g [...] 6.40% 6.40% audited=80 tclass=4 | __libc_start_main | |--4.60%--__GI___ioctl | entry_SYSCALL_64 | do_syscall_64 | __x64_sys_ioctl | ksys_ioctl | binder_ioctl | binder_set_nice | can_nice | capable | security_capable | cred_has_capability.isra.0 | slow_avc_audit | common_lsm_audit | avc_audit_post_callback | avc_audit_post_callback | It is also possible to use the ftrace interface: # echo 1 > /sys/kernel/debug/tracing/events/avc/selinux_audited/enable # cat /sys/kernel/debug/tracing/trace tracer: nop entries-in-buffer/entries-written: 1/1 #P:8 [...] dmesg-3624 [001] 13072.325358: selinux_denied: audited=80 tclass=4 An explanation here of how one might go about decoding audited and tclass would be helpful to users (even better would be a script to do it for them). Again, I know how to do that but not everyone using perf/ftrace will.