libbpf-tools: add klockstat

[brho]

This is a port of BCC's klockstat. Differences from BCC:

• can specify a lock by ksym name, using -L
• tracks whichever task had the max time for acquire and hold, outputted
  when -s > 1 (otherwise it's cluttered).
• does not reset stats each interval by default. Can request with -R.

---

Usage: klockstat [-hRT] [-p PID] [-t TID] [-c FUNC] [-L LOCK] [-n NR_LOCKS]
[-s NR_STACKS] [-S SORT] [-d DURATION] [-i INTERVAL]

-p, --pid=PID Filter by process ID
-t, --tid=TID Filter by thread ID
-c, --caller=FUNC Filter by caller string prefix
-L, --lock=LOCK Filter by specific ksym lock name
-n, --locks=NR_LOCKS Number of locks to print
-s, --stacks=NR_STACKS Number of stack entries to print per lock
-S, --sort=SORT Sort by field:
acq_[max|total|count]
hld_[max|total|count]
-d, --duration=SECONDS Duration to trace
-i, --interval=SECONDS Print interval
-R, --reset Reset stats each interval
-T, --timestamp Print timestamp

-?, --help Give this help list
--usage Give a short usage message
-V, --version Print program version

Mandatory or optional arguments to long options are also mandatory or optional
for any corresponding short options.

Examples:
klockstat # trace system wide until ctrl-c
klockstat -d 5 # trace for 5 seconds
klockstat -i 5 # print stats every 5 seconds
klockstat -p 181 # trace process 181 only
klockstat -t 181 # trace thread 181 only
klockstat -c pipe_ # print only for lock callers with 'pipe_'
# prefix
klockstat -L cgroup_mutex # trace the cgroup_mutex lock only
klockstat -S acq_count # sort lock acquired results by acquire count
klockstat -S hld_total # sort lock held results by total held time
klockstat -S acq_count,hld_total # combination of above
klockstat -n 3 # display top 3 locks
klockstat -s 6 # display 6 stack entries per lock

---

Signed-off-by: Barret Rhoden brho@google.com

[davemarchevsky]

[buildbot, test this please]

[davemarchevsky]

[buildbot, ok to test]

[brho]

just checking if there's anything else needed on this PR.

[chenhengqi]

just checking if there's anything else needed on this PR.

Sorry, I missed the GitHub notification. Will take a look at this again this week.

[brho]

hi -

two updates on this in my recent change (9 Dec):

1. instead of tracking 'lock depth' as a lookup (which was interpreted as "task X's Nth lock acquired"), do lookups by task_id + lock_ptr, which is "Task X grabbing lock Y".

the biggest benefit is that we don't have "mismatched" lock/unlock pairs. so if a task grabs mutex A, then mutex B, it can unlock A before unlocking B. if you tracked by "depth", you'd just assume the first unlocked was the last to be locked.

an ancillary benefit is we don't have to worry about having unlock get paired with "ancient" locks, which means we don't need that "enabled" flag, or the bulk of comments justifying it.

finally, it cleans up the code a bit: one less MAP to maintain, since we no longer care about a task's "depth".

2. use atomics or other shared memory sync when accessing a lockstat structure. these are keyed by a callstack. although callstacks identify a line of code, e.g. vfs_read+0x12, they do not uniquely identify a lock. so you can have multiple threads working on the same struct in account(). i put the WRITE_ONCE/READ_ONCE in bits.bpf.h, but that could be moved elsewhere.

thanks!

[brho]

pushed fixes. thanks!

[brho]

just checking if this needs anything else. thanks.

[davemarchevsky]

Sorry for the merge delay here.

I would normally push back against the change to default behavior here, but I can see why it makes sense and am not an active user of klockstat so don't think it's a good idea to hold this up further.

Did a quick followup PR (#3796) adding "Differences from BCC" comment.