Hello,

There have been requests for more sophisticated perf event sample
filtering based on the sample data.  Recently the kernel added BPF
programs can access perf sample data and this is the userspace part
to enable such a filtering.

This still has some rough edges and needs more improvements.  But
I'd like to share the current work and get some feedback for the
directions and idea for further improvements.

v2 changes)
 * fix build error with the misc field  (Jiri)
 * add a destructor for filter expr  (Ian)
 * remove 'bpf:' prefix  (Arnaldo)
 * add '||' operator

The kernel changes are in the tip.git tree (perf/core branch) for now.
perf record has --filter option to set filters on the last specified
event in the command line.  It worked only for tracepoints and Intel
PT events so far.  This patchset extends it to have 'bpf:' prefix in
order to enable the general sample filters using BPF for any events.

A new filter expression parser was added (using flex/bison) to process
the filter string.  Right now, it only accepts very simple expressions
separated by comma.  I'd like to keep the filter expression as simple
as possible.

It requires samples satisfy all the filter expressions otherwise it'd
drop the sample.  IOW filter expressions are connected with logical AND
operations implicitly.

Essentially the BPF filter expression is:

  <term> <operator> <value> (("," | "||") <term> <operator> <value>)*

The <term> can be one of:
  ip, id, tid, pid, cpu, time, addr, period, txn, weight, phys_addr,
  code_pgsz, data_pgsz, weight1, weight2, weight3, ins_lat, retire_lat,
  p_stage_cyc, mem_op, mem_lvl, mem_snoop, mem_remote, mem_lock,
  mem_dtlb, mem_blk, mem_hops

The <operator> can be one of:
  ==, !=, >, >=, <, <=, &

The <value> can be one of:
  <number> (for any term)
  na, load, store, pfetch, exec (for mem_op)
  l1, l2, l3, l4, cxl, io, any_cache, lfb, ram, pmem (for mem_lvl)
  na, none, hit, miss, hitm, fwd, peer (for mem_snoop)
  remote (for mem_remote)
  na, locked (for mem_locked)
  na, l1_hit, l1_miss, l2_hit, l2_miss, any_hit, any_miss, walk, fault (for mem_dtlb)
  na, by_data, by_addr (for mem_blk)
  hops0, hops1, hops2, hops3 (for mem_hops)

I plan to improve it with range expressions like for ip or addr and it
should support symbols like the existing addr-filters.  Also cgroup
should understand and convert cgroup names to IDs.

Let's take a look at some examples.  The following is to profile a user
program on the command line.  When the frequency mode is used, it starts
with a very small period (i.e. 1) and adjust it on every interrupt (NMI)
to catch up the given frequency.

  $ ./perf record -- ./perf test -w noploop
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.263 MB perf.data (4006 samples) ]

  $ ./perf script -F pid,period,event,ip,sym | head
  36695          1 cycles:  ffffffffbab12ddd perf_event_exec
  36695          1 cycles:  ffffffffbab12ddd perf_event_exec
  36695          5 cycles:  ffffffffbab12ddd perf_event_exec
  36695         46 cycles:  ffffffffbab12de5 perf_event_exec
  36695       1163 cycles:  ffffffffba80a0eb x86_pmu_disable_all
  36695       1304 cycles:  ffffffffbaa19507 __hrtimer_get_next_event
  36695       8143 cycles:  ffffffffbaa186f9 __run_timers
  36695      69040 cycles:  ffffffffbaa0c393 rcu_segcblist_ready_cbs
  36695     355117 cycles:            4b0da4 noploop
  36695     321861 cycles:            4b0da4 noploop

If you want to skip the first few samples that have small periods, you
can do like this (note it requires root due to BPF).

  $ sudo ./perf record -e cycles --filter 'period > 10000' -- ./perf test -w noploop
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.262 MB perf.data (3990 samples) ]

  $ sudo ./perf script -F pid,period,event,ip,sym | head
  39524      58253 cycles:  ffffffffba97dac0 update_rq_clock
  39524     232657 cycles:            4b0da2 noploop
  39524     210981 cycles:            4b0da2 noploop
  39524     282882 cycles:            4b0da4 noploop
  39524     392180 cycles:            4b0da4 noploop
  39524     456058 cycles:            4b0da4 noploop
  39524     415196 cycles:            4b0da2 noploop
  39524     462721 cycles:            4b0da4 noploop
  39524     526272 cycles:            4b0da2 noploop
  39524     565569 cycles:            4b0da4 noploop

Maybe more useful example is when it deals with precise memory events.
On AMD processors with IBS, you can filter only memory load with L1
dTLB is missed like below.

  $ sudo ./perf record -ad -e ibs_op//p \
  > --filter 'mem_op == load, mem_dtlb > l1_hit' sleep 1
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 1.338 MB perf.data (15 samples) ]

  $ sudo ./perf script -F data_src | head
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          49080142 |OP LOAD|LVL L1 hit|SNP N/A|TLB L2 hit|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          51088842 |OP LOAD|LVL L3 or Remote Cache (1 hop) hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
          49080442 |OP LOAD|LVL L2 hit|SNP N/A|TLB L2 hit|LCK N/A|BLK  N/A
          51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A

You can also check the number of dropped samples in LOST_SAMPLES events
using perf report --stat command.

  $ sudo ./perf report --stat

  Aggregated stats:
             TOTAL events:      16066
              MMAP events:         22  ( 0.1%)
              COMM events:       4166  (25.9%)
              EXIT events:          1  ( 0.0%)
          THROTTLE events:        816  ( 5.1%)
        UNTHROTTLE events:        613  ( 3.8%)
              FORK events:       4165  (25.9%)
            SAMPLE events:         15  ( 0.1%)
             MMAP2 events:       6133  (38.2%)
      LOST_SAMPLES events:          1  ( 0.0%)
           KSYMBOL events:         69  ( 0.4%)
         BPF_EVENT events:         57  ( 0.4%)
    FINISHED_ROUND events:          3  ( 0.0%)
          ID_INDEX events:          1  ( 0.0%)
        THREAD_MAP events:          1  ( 0.0%)
           CPU_MAP events:          1  ( 0.0%)
         TIME_CONV events:          1  ( 0.0%)
     FINISHED_INIT events:          1  ( 0.0%)
  ibs_op//p stats:
            SAMPLE events:         15
      LOST_SAMPLES events:       3991

Note that the total aggregated stats show 1 LOST_SAMPLES event but
per event stats show 3991 events because it's the actual number of
dropped samples while the aggregated stats has the number of record.
Maybe we need to change the per-event stats to 'LOST_SAMPLES count'
to avoid the confusion.

The code is available at 'perf/bpf-filter-v2' branch in my tree.

  git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git

Again, you need tip/perf/core kernel for this to work.
Any feedback is welcome.

Thanks,
Namhyung

Namhyung Kim (8):
  perf bpf filter: Introduce basic BPF filter expression
  perf bpf filter: Implement event sample filtering
  perf record: Add BPF event filter support
  perf record: Record dropped sample count
  perf bpf filter: Add 'pid' sample data support
  perf bpf filter: Add more weight sample data support
  perf bpf filter: Add data_src sample data support
  perf bpf filter: Add logical OR operator

 tools/lib/perf/include/perf/event.h          |   2 +
 tools/perf/Documentation/perf-record.txt     |  15 +-
 tools/perf/Makefile.perf                     |   2 +-
 tools/perf/builtin-record.c                  |  46 ++++--
 tools/perf/util/Build                        |  16 ++
 tools/perf/util/bpf-filter.c                 | 142 ++++++++++++++++
 tools/perf/util/bpf-filter.h                 |  49 ++++++
 tools/perf/util/bpf-filter.l                 | 147 +++++++++++++++++
 tools/perf/util/bpf-filter.y                 |  77 +++++++++
 tools/perf/util/bpf_counter.c                |   3 +-
 tools/perf/util/bpf_skel/sample-filter.h     |  27 +++
 tools/perf/util/bpf_skel/sample_filter.bpf.c | 164 +++++++++++++++++++
 tools/perf/util/evsel.c                      |   2 +
 tools/perf/util/evsel.h                      |   7 +-
 tools/perf/util/parse-events.c               |   8 +-
 tools/perf/util/session.c                    |   3 +-
 16 files changed, 682 insertions(+), 28 deletions(-)
 create mode 100644 tools/perf/util/bpf-filter.c
 create mode 100644 tools/perf/util/bpf-filter.h
 create mode 100644 tools/perf/util/bpf-filter.l
 create mode 100644 tools/perf/util/bpf-filter.y
 create mode 100644 tools/perf/util/bpf_skel/sample-filter.h
 create mode 100644 tools/perf/util/bpf_skel/sample_filter.bpf.c


base-commit: f9fa0778ee7349a9aa3d2ea10e9f2ab843a0b44e
-- 
2.39.2.637.g21b0678d19-goog

Em Sat, Feb 18, 2023 at 10:13:21PM -0800, Namhyung Kim escreveu:
> Hello,
> 
> There have been requests for more sophisticated perf event sample
> filtering based on the sample data.  Recently the kernel added BPF
> programs can access perf sample data and this is the userspace part
> to enable such a filtering.
> 
> This still has some rough edges and needs more improvements.  But
> I'd like to share the current work and get some feedback for the
> directions and idea for further improvements.
> 
> v2 changes)
>  * fix build error with the misc field  (Jiri)
>  * add a destructor for filter expr  (Ian)
>  * remove 'bpf:' prefix  (Arnaldo)
>  * add '||' operator
> 
> The kernel changes are in the tip.git tree (perf/core branch) for now.
> perf record has --filter option to set filters on the last specified
> event in the command line.  It worked only for tracepoints and Intel
> PT events so far.  This patchset extends it to have 'bpf:' prefix in
> order to enable the general sample filters using BPF for any events.
> 
> A new filter expression parser was added (using flex/bison) to process
> the filter string.  Right now, it only accepts very simple expressions
> separated by comma.  I'd like to keep the filter expression as simple
> as possible.
> 
> It requires samples satisfy all the filter expressions otherwise it'd
> drop the sample.  IOW filter expressions are connected with logical AND
> operations implicitly.
> 
> Essentially the BPF filter expression is:
> 
>   <term> <operator> <value> (("," | "||") <term> <operator> <value>)*

So "," means "&&" ?

I understand that its less characters, but can't we use the well
established '&&' instead? :-)

- Arnaldo
 
> The <term> can be one of:
>   ip, id, tid, pid, cpu, time, addr, period, txn, weight, phys_addr,
>   code_pgsz, data_pgsz, weight1, weight2, weight3, ins_lat, retire_lat,
>   p_stage_cyc, mem_op, mem_lvl, mem_snoop, mem_remote, mem_lock,
>   mem_dtlb, mem_blk, mem_hops
> 
> The <operator> can be one of:
>   ==, !=, >, >=, <, <=, &
> 
> The <value> can be one of:
>   <number> (for any term)
>   na, load, store, pfetch, exec (for mem_op)
>   l1, l2, l3, l4, cxl, io, any_cache, lfb, ram, pmem (for mem_lvl)
>   na, none, hit, miss, hitm, fwd, peer (for mem_snoop)
>   remote (for mem_remote)
>   na, locked (for mem_locked)
>   na, l1_hit, l1_miss, l2_hit, l2_miss, any_hit, any_miss, walk, fault (for mem_dtlb)>   na, by_data, by_addr (for mem_blk)
>   hops0, hops1, hops2, hops3 (for mem_hops)
> 
> I plan to improve it with range expressions like for ip or addr and it
> should support symbols like the existing addr-filters.  Also cgroup
> should understand and convert cgroup names to IDs.
> 
> Let's take a look at some examples.  The following is to profile a user
> program on the command line.  When the frequency mode is used, it starts
> with a very small period (i.e. 1) and adjust it on every interrupt (NMI)
> to catch up the given frequency.
> 
>   $ ./perf record -- ./perf test -w noploop
>   [ perf record: Woken up 1 times to write data ]
>   [ perf record: Captured and wrote 0.263 MB perf.data (4006 samples) ]
> 
>   $ ./perf script -F pid,period,event,ip,sym | head
>   36695          1 cycles:  ffffffffbab12ddd perf_event_exec
>   36695          1 cycles:  ffffffffbab12ddd perf_event_exec
>   36695          5 cycles:  ffffffffbab12ddd perf_event_exec
>   36695         46 cycles:  ffffffffbab12de5 perf_event_exec
>   36695       1163 cycles:  ffffffffba80a0eb x86_pmu_disable_all
>   36695       1304 cycles:  ffffffffbaa19507 __hrtimer_get_next_event
>   36695       8143 cycles:  ffffffffbaa186f9 __run_timers
>   36695      69040 cycles:  ffffffffbaa0c393 rcu_segcblist_ready_cbs
>   36695     355117 cycles:            4b0da4 noploop
>   36695     321861 cycles:            4b0da4 noploop
> 
> If you want to skip the first few samples that have small periods, you
> can do like this (note it requires root due to BPF).
> 
>   $ sudo ./perf record -e cycles --filter 'period > 10000' -- ./perf test -w noploop
>   [ perf record: Woken up 1 times to write data ]
>   [ perf record: Captured and wrote 0.262 MB perf.data (3990 samples) ]
> 
>   $ sudo ./perf script -F pid,period,event,ip,sym | head
>   39524      58253 cycles:  ffffffffba97dac0 update_rq_clock
>   39524     232657 cycles:            4b0da2 noploop
>   39524     210981 cycles:            4b0da2 noploop
>   39524     282882 cycles:            4b0da4 noploop
>   39524     392180 cycles:            4b0da4 noploop
>   39524     456058 cycles:            4b0da4 noploop
>   39524     415196 cycles:            4b0da2 noploop
>   39524     462721 cycles:            4b0da4 noploop
>   39524     526272 cycles:            4b0da2 noploop
>   39524     565569 cycles:            4b0da4 noploop
> 
> Maybe more useful example is when it deals with precise memory events.
> On AMD processors with IBS, you can filter only memory load with L1
> dTLB is missed like below.
> 
>   $ sudo ./perf record -ad -e ibs_op//p \
>   > --filter 'mem_op == load, mem_dtlb > l1_hit' sleep 1
>   [ perf record: Woken up 1 times to write data ]
>   [ perf record: Captured and wrote 1.338 MB perf.data (15 samples) ]
> 
>   $ sudo ./perf script -F data_src | head
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           49080142 |OP LOAD|LVL L1 hit|SNP N/A|TLB L2 hit|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           51088842 |OP LOAD|LVL L3 or Remote Cache (1 hop) hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
>           49080442 |OP LOAD|LVL L2 hit|SNP N/A|TLB L2 hit|LCK N/A|BLK  N/A
>           51080242 |OP LOAD|LVL LFB/MAB hit|SNP N/A|TLB L2 miss|LCK N/A|BLK  N/A
> 
> You can also check the number of dropped samples in LOST_SAMPLES events
> using perf report --stat command.
> 
>   $ sudo ./perf report --stat
> 
>   Aggregated stats:
>              TOTAL events:      16066
>               MMAP events:         22  ( 0.1%)
>               COMM events:       4166  (25.9%)
>               EXIT events:          1  ( 0.0%)
>           THROTTLE events:        816  ( 5.1%)
>         UNTHROTTLE events:        613  ( 3.8%)
>               FORK events:       4165  (25.9%)
>             SAMPLE events:         15  ( 0.1%)
>              MMAP2 events:       6133  (38.2%)
>       LOST_SAMPLES events:          1  ( 0.0%)
>            KSYMBOL events:         69  ( 0.4%)
>          BPF_EVENT events:         57  ( 0.4%)
>     FINISHED_ROUND events:          3  ( 0.0%)
>           ID_INDEX events:          1  ( 0.0%)
>         THREAD_MAP events:          1  ( 0.0%)
>            CPU_MAP events:          1  ( 0.0%)
>          TIME_CONV events:          1  ( 0.0%)
>      FINISHED_INIT events:          1  ( 0.0%)
>   ibs_op//p stats:
>             SAMPLE events:         15
>       LOST_SAMPLES events:       3991
> 
> Note that the total aggregated stats show 1 LOST_SAMPLES event but
> per event stats show 3991 events because it's the actual number of
> dropped samples while the aggregated stats has the number of record.
> Maybe we need to change the per-event stats to 'LOST_SAMPLES count'
> to avoid the confusion.
> 
> The code is available at 'perf/bpf-filter-v2' branch in my tree.
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/namhyung/linux-perf.git
> 
> Again, you need tip/perf/core kernel for this to work.
> Any feedback is welcome.
> 
> Thanks,
> Namhyung
> 
> Namhyung Kim (8):
>   perf bpf filter: Introduce basic BPF filter expression
>   perf bpf filter: Implement event sample filtering
>   perf record: Add BPF event filter support
>   perf record: Record dropped sample count
>   perf bpf filter: Add 'pid' sample data support
>   perf bpf filter: Add more weight sample data support
>   perf bpf filter: Add data_src sample data support
>   perf bpf filter: Add logical OR operator
> 
>  tools/lib/perf/include/perf/event.h          |   2 +
>  tools/perf/Documentation/perf-record.txt     |  15 +-
>  tools/perf/Makefile.perf                     |   2 +-
>  tools/perf/builtin-record.c                  |  46 ++++--
>  tools/perf/util/Build                        |  16 ++
>  tools/perf/util/bpf-filter.c                 | 142 ++++++++++++++++
>  tools/perf/util/bpf-filter.h                 |  49 ++++++
>  tools/perf/util/bpf-filter.l                 | 147 +++++++++++++++++
>  tools/perf/util/bpf-filter.y                 |  77 +++++++++
>  tools/perf/util/bpf_counter.c                |   3 +-
>  tools/perf/util/bpf_skel/sample-filter.h     |  27 +++
>  tools/perf/util/bpf_skel/sample_filter.bpf.c | 164 +++++++++++++++++++
>  tools/perf/util/evsel.c                      |   2 +
>  tools/perf/util/evsel.h                      |   7 +-
>  tools/perf/util/parse-events.c               |   8 +-
>  tools/perf/util/session.c                    |   3 +-
>  16 files changed, 682 insertions(+), 28 deletions(-)
>  create mode 100644 tools/perf/util/bpf-filter.c
>  create mode 100644 tools/perf/util/bpf-filter.h
>  create mode 100644 tools/perf/util/bpf-filter.l
>  create mode 100644 tools/perf/util/bpf-filter.y
>  create mode 100644 tools/perf/util/bpf_skel/sample-filter.h
>  create mode 100644 tools/perf/util/bpf_skel/sample_filter.bpf.c
> 
> 
> base-commit: f9fa0778ee7349a9aa3d2ea10e9f2ab843a0b44e
> -- 
> 2.39.2.637.g21b0678d19-goog
> 

-- 

- Arnaldo

On Thu, Feb 23, 2023 at 6:47 AM Arnaldo Carvalho de Melo
<acme@kernel.org> wrote:
>
> Em Sat, Feb 18, 2023 at 10:13:21PM -0800, Namhyung Kim escreveu:
> > Hello,
> >
> > There have been requests for more sophisticated perf event sample
> > filtering based on the sample data.  Recently the kernel added BPF
> > programs can access perf sample data and this is the userspace part
> > to enable such a filtering.
> >
> > This still has some rough edges and needs more improvements.  But
> > I'd like to share the current work and get some feedback for the
> > directions and idea for further improvements.
> >
> > v2 changes)
> >  * fix build error with the misc field  (Jiri)
> >  * add a destructor for filter expr  (Ian)
> >  * remove 'bpf:' prefix  (Arnaldo)
> >  * add '||' operator
> >
> > The kernel changes are in the tip.git tree (perf/core branch) for now.
> > perf record has --filter option to set filters on the last specified
> > event in the command line.  It worked only for tracepoints and Intel
> > PT events so far.  This patchset extends it to have 'bpf:' prefix in
> > order to enable the general sample filters using BPF for any events.
> >
> > A new filter expression parser was added (using flex/bison) to process
> > the filter string.  Right now, it only accepts very simple expressions
> > separated by comma.  I'd like to keep the filter expression as simple
> > as possible.
> >
> > It requires samples satisfy all the filter expressions otherwise it'd
> > drop the sample.  IOW filter expressions are connected with logical AND
> > operations implicitly.
> >
> > Essentially the BPF filter expression is:
> >
> >   <term> <operator> <value> (("," | "||") <term> <operator> <value>)*
>
> So "," means "&&" ?

Right. :)

>
> I understand that its less characters, but can't we use the well
> established '&&' instead? :-)

It's doable but I was a bit afraid of the complexity or confusion
it might bring up.  It should handle expressions like 'A && B || C'
then we need to be clear on the precedence of them.  Currently
it would behave like 'A && (B || C)' but users might expect
'(A && B) || C'.

Then we might want to add the parentheses explicitly to adjust
the operator priority.  That could result in complex expressions
possibly with nested parentheses.  I'm not sure if we really want
this.  Considering the restrictions in BPF, I think we should keep
the expressions as simple as possible.

Thanks,
Namhyung