On Tue, 10 Feb 2026 14:14:15 +0900
Masami Hiramatsu (Google) <[email protected]> wrote:
> Hmm, OK. Now I found how sysfs handles it.
>
> /*
> * For regular files, if the opener has CAP_DAC_OVERRIDE, open(2)
> * succeeds regardless of the RW permissions. sysfs had an extra
> * layer of enforcement where open(2) fails with -EACCES regardless
> * of CAP_DAC_OVERRIDE if the permission doesn't have the
> * respective read or write access at all (none of S_IRUGO or
> * S_IWUGO) or the respective operation isn't implemented. The
> * following flag enables that behavior.
> */
> KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK = 0x0002,
>
> So for the similar reason, I will make tracefs to check the permission
> even if CAP_DAC_OVERRIDE is set. (But this check should be done in general,
> instead of each open() operation)
>
I don't believe this is the same. This is about an instance being truly
read only. The instance is special, not the files. Note, permissions can
be changed by root too. After applying your patches, I did the following:
~# cd /sys/kernel/tracing/instances/backup/
~# ls -l current_tracer
-r--r----- 1 root root 0 Feb 11 10:29 current_tracer
~# cat current_tracer
nop
~# cat trace
# tracer: nop
#
# entries-in-buffer/entries-written: 0/0 #P:8
#
# _-----=> irqs-off/BH-disabled
# / _----=> need-resched
# | / _---=> hardirq/softirq
# || / _--=> preempt-depth
# ||| / _-=> migrate-disable
# |||| / delay
# TASK-PID CPU# ||||| TIMESTAMP FUNCTION
# | | | ||||| | |
~# chmod 664 current_tracer
~# ls -l current_tracer
-rw-rw-r-- 1 root root 0 Feb 11 10:29 current_tracer
~# echo function > current_tracer
~# cat current_tracer
function
~# cat trace
# tracer: function
#
# entries-in-buffer/entries-written: 1750306/2076556 #P:8
#
# _-----=> irqs-off/BH-disabled
# / _----=> need-resched
# | / _---=> hardirq/softirq
# || / _--=> preempt-depth
# ||| / _-=> migrate-disable
# |||| / delay
# TASK-PID CPU# ||||| TIMESTAMP FUNCTION
# | | | ||||| | |
bash-1056 [001] ..... 231.448852: mutex_unlock
<-tracing_set_tracer
<idle>-0 [002] ...1. 231.448853: arch_cpu_idle_exit <-do_idle
##### CPU 7 buffer started ####
<idle>-0 [007] ...1. 231.448853: arch_cpu_idle_exit <-do_idle
bash-1056 [001] ..... 231.448854: __mutex_unlock_slowpath
<-tracing_set_tracer
<idle>-0 [002] d..1. 231.448855: arch_cpu_idle_enter <-do_idle
<idle>-0 [007] d..1. 231.448855: arch_cpu_idle_enter <-do_idle
<idle>-0 [007] d..1. 231.448855: tsc_verify_tsc_adjust
<-arch_cpu_idle_enter
<idle>-0 [002] d..1. 231.448855: tsc_verify_tsc_adjust
<-arch_cpu_idle_enter
bash-1056 [001] d.... 231.448856:
fpregs_assert_state_consistent <-arch_exit_to_user_mode_prepare
<idle>-0 [007] d..1. 231.448856: local_touch_nmi <-do_idle
<idle>-0 [002] d..1. 231.448856: local_touch_nmi <-do_idle
bash-1056 [001] d.... 231.448856: switch_fpu_return
<-arch_exit_to_user_mode_prepare
<idle>-0 [007] d..1. 231.448856:
rcu_nocb_flush_deferred_wakeup <-do_idle
<idle>-0 [002] d..1. 231.448856:
rcu_nocb_flush_deferred_wakeup <-do_idle
<idle>-0 [007] d..1. 231.448857: cpuidle_get_cpu_driver
<-do_idle
<idle>-0 [002] d..1. 231.448857: cpuidle_get_cpu_driver
<-do_idle
<idle>-0 [007] d..1. 231.448857: cpuidle_not_available
<-do_idle
[..]
Not too read only!
I change permissions all the time for tracefs files, so I don't want
that changed.
This is not the same as sysfs. Let's keep it simple. Have all the open
callers that can do writes return error -EACCES if a file allows writes
and is open for write, but is part of the read only instance.
-- Steve
