[PATCH 00/22] tracing: 'hist' triggers
From: Tom Zanussi
Date: Wed Feb 24 2016 - 13:40:59 EST
This is v14 of the 'hist triggers' patchset, which incorporates
suggestions from Steve, and drops the first 8 from v13.
Patch 1 can be merged into patch 8 from v13, 'tracing: Add lock-free
tracing_map'. Following the v13 suggestions, I made a couple changes
to the tracing map constants to clarify the field relationships, and
added a couple missing comments I noticed while at it.
Changes from v13:
- Changed tracing map defines to make the relationship between vals,
keys, and fields clear, and updated the code to use the new
- Added missing comments to a couple tracing map functions.
- Added 1 to the number of vals supported, to account for hitcount.
- Added parentheses around macro variables.
- Added "_FL" to hist_field_flags enums.
- Changed '!strncmp/strcmp' to 'strncmp/strcmp == 0' thoughout.
- Added '=' to attribute names.
- Added WARN_ON() where it made sense i.e. where field indices were
updated or used and it was't immediately clear from immdiate
context whether the index could possibly be invalid.
Changes from v12:
- Removed the unnecessary return val check and NULL assignment
pointed out by Namhyung, and did a small cleanup to make the error
handling in that function (create_hist_data()) more uniform, which
also included a small change to create_sort_keys().
- Removed the 'update cond flag' patch from the series since that's
now in tracing/for-next.
Changes from v11:
- Added tracing_map_lookup().
- Changed tracing_map_insert() to incorporate the hits/drops counter
updating. The changes allow clients to continue calling
tracing_map_insert() and update existing entries if desired, or
stop on the first drop, since the return value is still NULL in
- Changed the hist trigger code to continue calling
tracing_map_insert() rather than stopping on the first drop. The
idea is that there's no harm in continuing after the first drop -
if the user can't tolerate any drops, they'll still see a non-zero
drop count and can retry with a larger table, but if the user
doesn't care, the hit and drop counts are an accurate reflection
of how many events were tallied vs dropped.
- Added a new unreg_all() callback to event_command, and code to
call it when the trigger file is opened for truncate. Implemented
unreg_all() for hist triggers, which allows them to honor the
truncate flag while leaving existing triggers intact. This means
that '>>' should now be used for adding multiple hist triggers as
well as to clear/pause/cont an existing hist trigger.
- Included Namhyung's log2 patch.
- Fixed one of Masami's test cases which now requires using '>>' due
to the truncate changes.
- 'cont' and 'clear' no longer create a trigger if one doesn't exist.
- Rebased to latest trace/for-next.
- Various other smaller fixes and Documentation updates as noted by
Changes from v10:
- Rebased to latest trace/for-next.
- Added Masami's ftrace/hist triggers kselftest patches.
- Added Masami's Tested-by: to the series.
Changes from v9:
v10 adds some major new features - 'named' hist triggers and support
for multiple hist triggers on a single event - basically all of the
features and suggestions suggested by Masami, except for one: I
decided after all not to implement the 'pause/continue/clear' commands
on the hist file. I've gotten so used to the habit of calling up the
previous trigger-setting command in the command history and simply
appending a 'pause/cont/clear' command to it (and later removing it by
prepending a '!') that making the user change the command doesn't seem
as naturally usable. Masami, if you disagree, let me know and I'll
reconsider it but at this point I'd rather not make that change.
In addition, I think I've fixed all the other various smaller problems
mentioned in the previous review, please let me know if I missed any...
- Added support for 'named' hist triggers
- Added support for multiple hist triggers on a single trace event
- Added documentation and examples for the above new features
- Streamlined the README to only include the essentials for hist triggers
- Fixed the 'hex' modifier for values, which was previously ignored
- Replaced the kmalloc of large arrays with a simple page-based scheme
- Fixed the tracing_map_insert() case where jhash returns 0
- Fixed various other small problems noted along the way
Changes from v8:
Same as v8, but with the RFC patch [ftrace: Add function_hist tracer]
removed, and rebased to latest trace/for-next.
Changes from v7:
This version refactors the commits as suggested by Masami. There are
now more commits, but the result should be much more reviewable. The
ending code is the same as before, modulo a couple minor bug fixes I
discovered while refactoring and testing.
I've also reviewed and fixed a number of shortcomings and errors in
the comments, and have added a new discussion of the tracing_map data
structures after Steve mentioned he found them confusing and/or
Also, I kept Namhyung's string patch [tracing: Support string type key
properly] as submitted, but added a follow-on patch that refactors it
and fixes a problem I found with it that enabled static string keys to
contain random chars and therefore incorrect map insertions.
Changes from v6:
This version adds a new 'sym-offset' modifier as requested by Masami.
I implemented it as a modifier rather than using the trace option as
suggested, in part because I wanted to keep it all self-contained and
it seemed more consistent to just add it alongside the 'sym' modifier.
Also, hist triggers arent't really a tracer and therefore don't
directly tie into the option update/callback mechanism so making use
of it isn't as simple as a normal tracer.
I also changed the sort key specification to be stricter and signal an
error if the specified sort key wasn't found (rather than defaulting
to hitcount in those cases), also suggested by Masami. Thanks,
Masami, for your input!
Also updated the Documentation and tracing/README to reflect the
Changes from v5:
This version adds support for compound keys, along with the related
ability to sort using primary and secondary keys. This was mentioned
in previous versions as the last important piece that remained
unimplemented, and is now implemented. (I didn't have time to get to
the couple of enhancements suggested by Masami, but I expect to be
able to add those later on top of these.)
Because we now support compound keys and it's not immediately clear in
the output exactly which fields correspond to keys, the key(s),
compound or not, are now enclosed by curly braces.
The Documentation and README have been updated to reflect the changes,
and several new examples have been added to illustrate how to use
Also, the code was updated to work with the new ftrace_event_file,
etc, renaming in tracing/for-next.
Changes from v4:
This version addresses some problems and suggestions made by Daniel
Wagner - a lot of the code was reworked to get rid of the distinction
between keys and values, and as a result, both keys and values can be
used as sort keys. As suggested, it also allows 'val=' to be absent
in a trigger command - if no 'val' is specified, hitcount is assumed
and automatically used as the only val.
The map code was also separated out into a separate file,
tracing_map.c, allowing it to be reused. It also adds a second tracer
called function_hist that actually does reuse the code, as an RFC
Patch 01/10 [tracing: Update cond flag when enabling or disabling..]
is a fix for a problem noticed by Daniel and that fixes a problem in
existing trigger code and should be applied regardless of whether the
rest of the patchset is merged.
As mentioned, patch 10/10 is an RFC patch implementing a new tracer
based on the function tracer code. It's a fun little tool and is
useful for a specific problem I'm working on (and is also a nice test
of the tracing_map code), but is an RFC because first, I'm not sure it
would really be of general interest and secondly, it's POC-level
quality and I'd need to spend more time fixing it up to make it
upstreamable, but I don't want to waste my time if not.
There are a couple of important bits of functionality that were
present in v1 but not yet reimplemented in v5.
The first is support for compound keys. Currently, maps can only be
keyed on a single event field, whereas in v1 they could be keyed on
multiple keys. With support for compound keys, you can create much
more interesting output, such as for example per-pid lists of
syscalls or read counts e.g.:
# echo 'hist:keys=common_pid.execname,id.syscall:vals=hitcount' > \
# cat /sys/kernel/debug/tracing/events/raw_syscalls/sys_enter/hist
key: common_pid:bash, id:sys_write vals: count:69
key: common_pid:bash, id:sys_rt_sigprocmask vals: count:218
key: common_pid:update-notifier, id:sys_poll vals: count:37
key: common_pid:update-notifier, id:sys_recvfrom vals: count:118
key: common_pid:deja-dup-monito, id:sys_sendto vals: count:1
key: common_pid:deja-dup-monito, id:sys_read vals: count:4
key: common_pid:deja-dup-monito, id:sys_poll vals: count:8
key: common_pid:deja-dup-monito, id:sys_recvmsg vals: count:8
key: common_pid:deja-dup-monito, id:sys_getegid vals: count:8
key: common_pid:emacs, id:sys_fsync vals: count:1
key: common_pid:emacs, id:sys_open vals: count:1
key: common_pid:emacs, id:sys_symlink vals: count:2
key: common_pid:emacs, id:sys_poll vals: count:23
key: common_pid:emacs, id:sys_select vals: count:23
key: common_pid:emacs, id:unknown_syscall vals: count:34
key: common_pid:emacs, id:sys_ioctl vals: count:60
key: common_pid:emacs, id:sys_rt_sigprocmask vals: count:116
key: common_pid:cat, id:sys_munmap vals: count:1
key: common_pid:cat, id:sys_fadvise64 vals: count:1
Related to that is support for sorting on multiple fields. Currently,
you can sort using only a primary key. Being able to sort on multiple
or at least a secondary key is indispensible for seeing trends when
displaying multiple values.
Changes from v3:
v4 fixes the race in tracing_map_insert() noted in v3, where
map.val.key could be checked even if map.val wasn't yet set. The
simple fix for that in tracing_map_insert() introduces the possibility
of duplicates in the map, which though rare, need to be accounted for
in the output. To address that, duplicate-merging code was added to
the map-printing code.
It was also pointed out that it didn't seem correct to include
module.h, but the fix for that has deeper roots and is being addressed
by a separate patchset; for now we need to continue including
module.h, though prompted by that I did some other header include
The functionality remains the same as v2, but this version no longer
tries to export and use bpf_maps, and more importantly removes the
associated GFP_NOTRACE/trace event hacks and kmem macros required to
work around the bpf_map implementation.
The tracing_map functionality is instead built on top of a simple
lock-free map algorithm originated by Dr. Cliff Click (see references
in the code for more details), which though too restrictive to be
general-purpose in its current form, functions nicely as a
special-purpose tracing map.
v3 also moves the hist triggers code into a separate file and puts it
all behind a new config option, CONFIG_HIST_TRIGGERS. It also merges
in the sorting code rather than keeping it as a separate patch.
This patchset also includes a couple other new and related triggers,
enable_hist and disable_hist, very similar to the existing
enable_event/disable_event triggers used to automatically enable and
disable events based on a triggering condition, but in this case
allowing hist triggers to be enabled and disabled in the same way.
- Added an insert check for val before checking the key associated with val
- Added code to merge possible duplicates in the map
Changes from v2:
- reimplemented tracing_map, replacing bpf_map with nmi-safe/lock-free map
- removed GPF_NOTRACE, kmalloc/free macros and event hacks needed by bpf_maps
- moved hist triggers from trace_events_trigger.c to trace_events_hist.c
- added CONFIG_HIST_TRIGGERS config option
- consolidated sorting code with main patch
Changes from v1:
- completely rewritten on top of tracing_map (renamed and exported bpf_map)
- added map clearing and client ops to tracing_map
- changed the name from 'hash' triggers to 'hist' triggers
- added new trigger 'pause' feature
- added new enable_hist and disable_hist triggers
- added usage for hist/enable_hist/disable hist to tracing/README
- moved examples into Documentation/trace/event.txt
- added ___GFP_NOTRACE, kmalloc/kfree macros, and conditional kmem tracepoints
The following changes since commit c230d332a81237cfef69ebc6060cf597889ede77:
tracing: Add lock-free tracing_map (2016-02-20 17:12:59 -0600)
are available in the git repository at:
Masami Hiramatsu (2):
kselftests/ftrace : Add event trigger testcases
kselftests/ftrace: Add hist trigger testcases
Namhyung Kim (2):
tracing: Support string type key properly
tracing: Add hist trigger 'log2' modifier
Tom Zanussi (18):
tracing: Update some tracing_map constants and comments
tracing: Add 'hist' event trigger command
tracing: Add hist trigger support for multiple values ('vals=' param)
tracing: Add hist trigger support for compound keys
tracing: Add hist trigger support for user-defined sorting ('sort='
tracing: Add hist trigger support for pausing and continuing a trace
tracing: Add hist trigger support for clearing a trace
tracing: Add hist trigger 'hex' modifier for displaying numeric fields
tracing: Add hist trigger 'sym' and 'sym-offset' modifiers
tracing: Add hist trigger 'execname' modifier
tracing: Add hist trigger 'syscall' modifier
tracing: Add hist trigger support for stacktraces as keys
tracing: Remove restriction on string position in hist trigger keys
tracing: Add enable_hist/disable_hist triggers
tracing: Add 'hist' trigger Documentation
tracing: Add support for multiple hist triggers per event
tracing: Add support for named triggers
tracing: Add support for named hist triggers
Documentation/trace/events.txt | 1555 +++++++++++++++++
include/linux/trace_events.h | 2 +
kernel/trace/Kconfig | 14 +
kernel/trace/Makefile | 1 +
kernel/trace/trace.c | 57 +
kernel/trace/trace.h | 52 +
kernel/trace/trace_events.c | 4 +
kernel/trace/trace_events_hist.c | 1746 ++++++++++++++++++++
kernel/trace/trace_events_trigger.c | 215 ++-
kernel/trace/tracing_map.c | 4 +-
kernel/trace/tracing_map.h | 5 +-
tools/testing/selftests/ftrace/test.d/functions | 9 +
.../ftrace/test.d/trigger/trigger-eventonoff.tc | 64 +
.../ftrace/test.d/trigger/trigger-filter.tc | 59 +
.../ftrace/test.d/trigger/trigger-hist-mod.tc | 65 +
.../ftrace/test.d/trigger/trigger-hist.tc | 83 +
.../ftrace/test.d/trigger/trigger-multihist.tc | 73 +
.../ftrace/test.d/trigger/trigger-snapshot.tc | 56 +
.../ftrace/test.d/trigger/trigger-stacktrace.tc | 53 +
.../ftrace/test.d/trigger/trigger-traceonoff.tc | 58 +
20 files changed, 4140 insertions(+), 35 deletions(-)
create mode 100644 kernel/trace/trace_events_hist.c
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-eventonoff.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-filter.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-hist-mod.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-hist.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-multihist.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-snapshot.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-stacktrace.tc
create mode 100644 tools/testing/selftests/ftrace/test.d/trigger/trigger-traceonoff.tc