[PATCH v5 00/23] perf trace: Fix BPF filtering and make tracing tests non-exclusive

[Ian Rogers <irogers@google.com>]

perf trace's BPF augmentation attaches to raw_syscalls:sys_enter and
raw_syscalls:sys_exit system wide, and used the program return value to
decide whether a syscall was interesting. Returning 0 from a
BPF_PROG_TYPE_TRACEPOINT program makes perf_trace_run_bpf_submit() drop
the event for every listener on that tracepoint, not just for the perf
trace that installed the program. Any concurrent perf trace, perf record
or ftrace session watching raw_syscalls therefore lost events, which is
one of the reasons so many of the perf trace and perf probe shell tests
had to be marked (exclusive) and run on their own.

Patches 1 to 6 are independent fixes to the code the rest of the series
goes on to rework or to rely on.

Patches 7 to 10 fix how an augmented argument is read back. They skip
the __data_loc internal tracepoint fields that syscalls:sys_enter_*
gained in 6.19 for syscalls only, bounds check every read of an
augmented payload, and stop the ring buffer padding at the end of a
sample being offered to the beautifiers as one. Patch 8 fixes the
SIGSEGV Arnaldo hit testing v4.

Patches 11 to 15 fix perf trace's filtering. They stop the return value
being used as a filter and do the filtering in BPF maps instead, stop
the sys_exit program array tail calling a sys_enter augmenter, and
replace the userspace PERF_RECORD_FORK/PERF_RECORD_EXIT bookkeeping with
BTF-typed raw tracepoint programs on sched_process_{fork,exit,exec}. A
task is then registered before its first syscall and evicted in
do_exit(), rather than whenever userspace next drains the ring buffer.

Patches 16 to 23 deal with the tests. Several collided with each other
through global state rather than through perf trace: fixed probe names,
clear_all_probes() disabling every tracepoint on the system, and perf
trace's hardcoded "probe:vfs_getname*" wildcard pinning probes belonging
to other tests. With those scoped to a pid they can drop (exclusive) and
run in parallel again.

Patches 1 and 2 fix pre-existing bugs that stand on their own and can be
applied ahead of the rest. Only patch 9 depends on them.

Tested on x86_64. The trace and probe tests pass under 'perf test -r3',
which runs the repeats concurrently. Every patch builds individually,
and the series also builds with BUILD_BPF_SKEL=0.

Changes since v4:
- New patch 1 fills in the size in the augmented argument header in the
  four augmenters that left it unset: sendto, perf_event_open,
  nanosleep and clock_nanosleep. They all build a payload for a fixed
  size type, and the beautifiers cast it directly rather than reading
  the length, so nothing has needed it so far. Patch 9 starts checking
  it. While there, report a failed bpf_probe_read_user() rather than
  claiming success for it: the payload length is set to zero and the
  error recorded in the header, so userspace falls back to printing the
  raw pointer instead of whatever the per-CPU scratch buffer last held.
- New patch 2 includes the augmented argument header in nanosleep's
  payload length. It used sizeof(augmented_args->args) where its three
  siblings use sizeof(u64) + sizeof(augmented_args->args), so the
  record was eight bytes short of the struct timespec it carries and
  tv_nsec was printed half from the ring buffer.
- Patch 7 only skips the internal tracepoint fields when it is setting
  up a syscall. syscall_arg_fmt__init_array() is also called for an
  ordinary tracepoint by evsel__init_tp_arg_scnprintf(), and
  trace__fprintf_tp_fields() walks that array in lockstep with the
  field list, so packing the entries shifted every formatter along by
  one from the field it belongs to. 'perf trace -e
  sched:sched_process_exec' printed its filename with the pid
  beautifier. The syscall path indexes the array packed and is
  unaffected, so the two callers now ask for what each of them needs.
- New patch 8 bounds checks an augmented argument before reading it.
  This is the SIGSEGV Arnaldo reported on v4. The check
  btf_struct_scnprintf() already had is lifted into a shared
  syscall_arg__augmented_args_valid() in beauty.h and used at every
  site that reads an augmented payload, in particular
  syscall_arg__scnprintf_buf(), which looped to a length taken from the
  payload with nothing bounding it and walked off the end of the 8K
  argbuf in bss.
- New patch 9 bounds the beautifiers that read a fixed size type out of
  a payload that may be shorter than it. perf_event_attr__fprintf()
  always reads a whole struct perf_event_attr, so the payload is copied
  into a zero padded local rather than read in place, and the AF_LOCAL
  socket address is printed with a length from the payload rather than
  as a string it need not be. Split out from patch 8 so the crash fix
  stays minimal.
- New patch 10 recognises the padding at the end of a sample for what
  it is. perf_sample_save_raw_data() rounds the raw size up to eight
  bytes and __output_skip() advances over the padding rather than
  zeroing it, so a 64 byte struct syscall_enter_args comes back as 68
  bytes and userspace took the four stale bytes for an augmented
  argument. A trailing run shorter than a struct augmented_arg cannot
  be one.
- Patch 13 notes that a descendant added to pids_filtered by
  sched_process_fork() can still have a sys_exit printed without a
  matching enter. The raw_syscalls:sys_exit evsel is still in the
  evlist and its tracepoint filter is set once, so it never learns of
  the descendants, and a tracepoint program can only suppress an event
  by vetoing it, which is what this series removes. It needs
  --filter-pids naming a process that then forks. Routing sys_exit
  through __augmented_syscalls__ the way sys_enter now is would close
  it.
- Patch 14 notes that a thread the target creates during the attach
  window is still missed by 'perf trace -t'. The re-enumeration reads
  /proc/<tid>/task/<tid>/children, which lists forked processes, and
  copy_process() makes a CLONE_THREAD child a sibling of its creator
  rather than a child of it. /proc does not record which task in a
  group created which, so such a sibling cannot be told apart from one
  -t was deliberately pointed away from.

Changes since v3:
- New patch 3 returns -ENOMEM rather than -1 from evsel__set_filter(),
  evsel__append_filter(), evlist__set_tp_filter() and
  evlist__append_tp_filter(). An allocation is the only thing that can
  fail in any of them, and patch 7 goes on to print what they return,
  where -1 negated is EPERM and a failure to allocate would have been
  reported as "Operation not permitted".
- Patch 5 sets sc->args only once the arg_fmt array that is indexed
  alongside it has been allocated. Publishing the field first left a
  syscall with args set and arg_fmt NULL, and since sc->name is already
  set by then a later syscall__read_info() returns success without
  retrying the allocation, leaving syscall_arg_fmt__mask_val() to
  dereference NULL.
- Patch 7 adds the raw_syscalls tracepoints in trace__run() only if
  they are not in the evlist already, and stops ignoring the result of
  trace__add_syscall_newtp(). cmd_trace() adds them before it creates
  the bpf-output event, and on the paths where that creation fails they
  were added a second time, giving the session two enter and two exit
  evsels for the same pair of tracepoints.
- Patch 7 no longer abandons the session when a target does not fit in
  the pid map. bpf_map__update_elem() answers -E2BIG once max_entries
  keys are present, so a target with more threads than the map has room
  for took perf trace down with it, on exactly the large workloads
  where there is least else to reach for. The tasks that fit are added
  and a warning says how many did not.
- Patch 9 includes <sys/types.h> for the pid_t it uses rather than
  relying on the include chain to drag it in.
- Patch 9 no longer ends the session when the target has already
  exited. thread_map__new_by_pid() fails with ENOENT once
  /proc/<pid>/task is gone, which is the ordinary outcome of
  'perf trace -p' on a short lived process, so only ENOMEM is now
  propagated and everything else is logged and stepped over.
- Patch 9 reads /proc once per thread group instead of once per thread.
  'perf trace -p' names every thread of the target in the thread map
  and /proc/<tid>/task lists the whole group whichever thread it is
  asked through, so the enumeration was quadratic in the number of
  threads. On a 64 thread target it went from 65 directory reads and
  4097 entries to 2 and 65, for the same set of pids.
- New patch 10 removes from the BPF maps the tasks that died before
  userspace got round to writing them there. sched_process_exit() can
  only delete a pid that is already in the map, so a target that exits
  during startup leaves an entry behind for the rest of the session,
  and since pids are reused an unrelated task then gets traced, or
  silently filtered out, in its place. The map is walked with the
  cursor left on a task found alive, which is one the walk has decided
  to keep, so its own deletions can never leave the cursor on a key
  that htab_map_get_next_key() will answer by starting the walk over.

Changes since v2:
- Rebased onto the current perf-tools-next.
- Patch 1 also includes <assert.h>, for the assert() in
  augmented_syscalls__create_bpf_output().
- New patch 2 makes evsel__put_and_free_priv() free the whole
  evsel_trace. It only zfree()d the struct, leaking the syscall_arg_fmt
  array hanging off it. No caller can reach that today, but patch 6
  adds one that discards a fully set up evsel.
- Patch 6 identifies the evsel to drop from the evlist by comparing
  against trace.syscalls.events.sys_enter instead of a strstr() for
  "syscalls:sys_enter". That substring also matches the per syscall
  syscalls:sys_enter_SYSCALL tracepoints, so a user asking for one of
  those by name would have had it removed from the evlist, and
  __augmented_syscalls__ described with its tracefs format rather than
  the raw tracepoint's.
- Patch 6 reports a failure to program the pid filters with the error
  that caused it. trace__run() sent everything trace__set_filter_pids()
  returned to out_error_mem, which prints "Not enough memory to run!".
  That was already a guess, and a wrong one now the function writes BPF
  maps too.
- Patch 7 moves the pid across in sched_process_exec() by deleting the
  old key before inserting the new one. The move is only a rename, but
  holding both keys at once needs a spare slot, and on a full map the
  insert failed with -E2BIG while the delete still succeeded, losing
  the task instead of moving it.
- Patch 7 no longer claims that a task forked during the attach window
  is still traced unaugmented. That was wrong: cmd_trace() removes the
  sys_enter evsel once the bpf-output event exists, so
  __augmented_syscalls__ is the only source of enter events and such a
  task is not reported at all. Describe what is really lost, why the
  tgid fallback is not kept as a safety net for it, note that only
  'perf trace -p' is exposed, and that closing it needs the target's
  descendants re-enumerated from /proc after the attach.
- Patch 7 drops the parent tgid test from sched_process_fork(), so a
  child inherits from the pid of the thread that called clone() and
  from nothing else. The lookups only ever match a task's own pid, and
  every thread of a -p target is enumerated from /proc/<pid>/task and
  inserted in its own right, so the tgid added no reach. What it did
  add was a child inheriting from a thread that is not traced itself,
  such as a sibling of the thread 'perf trace -t' selected.
- Patch 7 ends the session with the error that caused it when the BPF
  programs cannot be attached, rather than printing a bare errno left
  over from unwinding the attach. There is nothing to fall back to at
  that point, cmd_trace() has already built the evlist around
  __augmented_syscalls__ and dropped the sys_enter evsel, and the code
  being replaced did not fall back either: it ignored the result of
  augmented_raw_syscalls_bpf__attach() altogether.
- New patch 8 reads the target out of /proc again once the BPF programs
  are attached, so a task the target created while perf trace was
  starting up is traced rather than missed for the whole session. This
  is the gap patch 7 describes and left for later. It covers the
  target's new threads and, through task->children, anything it or they
  forked, to any depth. What is left is a task that made syscalls
  between sys_enter going live and being added to the map, which is
  momentary rather than lasting for the session.
- Patch 9 bails out if mktemp fails rather than carrying on with an
  empty $tmpdir. cd rejects the null directory, and the rmdir that was
  meant to undo the mktemp then failed to remove '' instead. Its
  cleanup() also no longer exits when it cannot cd out of the temporary
  directory, which would have skipped removing it. The removal takes an
  absolute path and does not need the cd to have succeeded.
- Patch 10 now narrows the disable in clear_all_probes() to the probes
  themselves rather than dropping it. Clearing kprobe_events or
  uprobe_events is all or nothing: dyn_events_release_all() returns
  -EBUSY without removing anything if it finds a probe that is still
  enabled, so simply removing the write could leave stale probes behind
  to collide with the next run. The set to disable is read from the
  kprobe_events and uprobe_events listings rather than assumed to be
  the groups perf uses, since a probe left enabled in another group,
  such as the default kprobes group used when kprobe_events is written
  directly, would abort the clear just the same.
- Patch 11 deletes the probes from an exit trap as well as on the way
  out. A pid scoped name is never seen again, so a run interrupted
  before cleanup_probe_vfs_getname() left its probes behind for good,
  a set per run, where the fixed name was at least found and reused by
  the next run.
- Patch 12 retries deleting the uprobe. Deletion writes uprobe_events
  just as addition does and can lose the same race with a concurrent
  test, and because the event name is now pid scoped a probe left
  behind is never overwritten by a later run. It also bails out if
  mktemp fails and quotes the path in the emptiness check, which
  unquoted would have tested the string "-s" and reported success.
- Patch 13 prints the tail of the output rather than the head. The
  pattern it matches only appears in the summary, which is printed
  after any trace output, so the head of the file is not the part that
  failed to match.

Changes since v1:
- New patch 1 includes <sys/types.h> and <string.h> for the pid_t and
  strcmp() uses that were relying on the include chain happening to
  drag them in, which does not hold on libcs such as musl.
- Patch 6 no longer returns success when the event qualifier filter
  string fails to allocate. err now defaults to 0 because either
  tracepoint may legitimately be absent, so the ENOMEM path has to set
  the error itself rather than rely on that default. It also includes
  <stdbool.h> for the bool parameters it adds to trace_augment.h.
- Patch 9 removes the temporary directory if the cd into it fails.
  That happens before the cleanup trap is installed, so the directory
  would otherwise be left behind in /tmp.

Ian Rogers (23):
  perf trace: Set the augmented arg header in the augmenters that omit
    it
  perf trace: Include the augmented arg header in nanosleep's payload
    length
  perf trace: Include the headers declaring pid_t, strcmp and assert
  perf trace: Free the whole evsel_trace in evsel__put_and_free_priv
  perf evsel: Report an allocation failure as ENOMEM when setting
    filters
  perf trace: Start BPF summary before starting workload
  perf trace: Skip internal tracepoint fields in formatting and beauty
    map
  perf trace: Bounds check augmented arguments before reading them
  perf trace: Bound the fixed size augmented argument beautifiers
  perf trace: Do not read sample padding as an augmented argument
  perf trace: Do not set unaugmented BPF program on sys_exit map
  perf trace: Filter events in BPF and avoid tracepoint vetoes
  perf trace: Handle fork and exit directly in BPF filter maps
  perf trace: Enumerate the target again once BPF is attached
  perf trace: Drop targets that died before they were filtered
  perf test test_task_analyzer: Isolate in temporary directory and make
    non-exclusive
  perf test common: Only disable probes in clear_all_probes
  perf test probe_vfs_getname: Scope probe name to PID and make
    non-exclusive
  perf test record+probe_libc_inet_pton: Scope event to PID, add
    retries, and make non-exclusive
  perf test trace_summary: Improve error diagnostics
  perf test trace_btf_general: Drop --max-events=1 and make
    non-exclusive
  perf test trace_summary: Make non-exclusive
  perf test uprobe_from_different_cu: Scope probe name to PID

 tools/perf/Documentation/perf-trace.txt       |   5 +
 tools/perf/builtin-trace.c                    | 863 +++++++++++++++---
 tools/perf/tests/shell/common/init.sh         |  33 +-
 .../perf/tests/shell/lib/probe_vfs_getname.sh |  51 +-
 tools/perf/tests/shell/probe_vfs_getname.sh   |   3 +-
 .../shell/record+probe_libc_inet_pton.sh      | 107 ++-
 .../shell/record+script_probe_vfs_getname.sh  |  18 +-
 tools/perf/tests/shell/test_task_analyzer.sh  |  20 +-
 .../shell/test_uprobe_from_different_cu.sh    |  11 +-
 .../tests/shell/trace+probe_vfs_getname.sh    |   9 +
 tools/perf/tests/shell/trace_btf_general.sh   |   8 +-
 tools/perf/tests/shell/trace_summary.sh       |  16 +-
 tools/perf/trace/beauty/beauty.h              |  43 +
 tools/perf/trace/beauty/perf_event_open.c     |  38 +-
 tools/perf/trace/beauty/sockaddr.c            |  55 +-
 tools/perf/trace/beauty/timespec.c            |   2 +-
 .../bpf_skel/augmented_raw_syscalls.bpf.c     | 407 ++++++++-
 tools/perf/util/bpf_trace_augment.c           | 288 +++++-
 tools/perf/util/evlist.c                      |  12 +-
 tools/perf/util/evsel.c                       |   4 +-
 tools/perf/util/trace_augment.h               |  33 +-
 21 files changed, 1801 insertions(+), 225 deletions(-)


base-commit: 9db4e9d6cc7c7d318fb6d6bb73abe3fed27385a8
-- 
2.56.0.rc1.315.gc6ed9934b7-goog