Skip to contentSkip to conversation

perf tools: Annotate fixes, stdio progress indication, debuginfo-client in more places

PATCH Superseded Arnaldo Carvalho de Melo 2 participants 19 messages Sep 14 perf events

v5 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Arnaldo Carvalho de Melo

author maintainer
#0 lore raw

v5 cover perf tools: Annotate fixes, stdio progress indication, debuginfo-client in more places

Hi,

This came out of work on data type profiling, but none of them depends on that work and nothing in this series needs anything from it, so sending them separately.

The fixes:

- 'perf report -s type' spins forever, burning all of a CPU with no output, on the dwz compressed debug info of zlib-ng (libz.so.1):
die_collect_vars() saves the dwarf_dieoffset() of the type DIE, which is relative to the file that DIE lives in, the dwz common file for the types shared by more than one CU, and resolving one of those offsets in the main debug file does not fail, it parses whatever is at that offset there, in this case a typedef whose DW_AT_type refers to itself, which is what makes the "follow the typedefs and qualifiers until a pointer or an array type" loop in die_get_pointer_type() spin. The offset is now resolved in the file it was recorded as coming from, which the CU of the collected DIE tells exactly rather than having to be inferred from what is at the offset, and the type chases and the struct/union member nesting are bounded, so a debug info file broken in some other way makes perf give up on a type with a pr_debug instead of looking like it hung (patch 7);

- the data type browser prints, in its samples view (-n, annotate.show_nr_samples), a local variable initialized to zero and never updated, so every member shows up as having no samples while the period and percent columns for the same entry are filled in (patch 4);

- the 'perf data type profiling' shell test turns "this PMU cannot record these events" into a test failure: the script runs under 'set -e', so the bare 'perf mem record' aborts it through the EXIT trap, which reports a signal that never happened, instead of reaching the code right below meant to report the failure (patch 1).

The features:

- 'perf report --progress' prints the phase it is in and how far along it is to stderr: the TUI shows that with a progress bar, but in the stdio case the ui_progress updates that are already there are dropped on the floor. It is what makes a slow run tell itself apart from a stuck one (patch 5);

- the debuginfo for a DSO, and the kernel symbols, can now be fetched keyed by the build ID recorded in the perf.data file, using the debuginfod client, for the cases where they are not available locally under the name the DSO was opened with: a vmlinux for a kernel that since got upgraded, or a profile recorded on another machine. Both are off with --no-debuginfod, with core.debuginfod=false, and when the build-id cache is turned off (patches 2 and 3);

- 'perf mem record' asks for PERF_SAMPLE_CPU by default, as without it the cpu field is the (u32)-1 "no CPU info" sentinel and per-sample analysis cannot tell accesses from different cores apart from same-CPU traffic (patch 8), and uses the IBS swfilt software privilege filter, when the kernel exposes it, so that per-thread recording works where only system wide recording did (patch 9).

Which file a saved type DIE offset belongs to is settled with dwarf_cu_getdwarf(), new in elfutils 0.160, so the libdw feature test now probes for it and the message in Makefile.config says 0.160 where it said 0.157: 0.157 to 0.159 is from 2014 and has no such symbol, and those versions now lose dwarf support with that message instead of failing to link util/dwarf-aux.c.

Best regards,

- Arnaldo

What changed from v4 (f801de4e80289566):

PATCH 1/8, "perf test: Skip data_type_profiling when the PMU cannot record memory events" [Namhyung Kim review of PATCH v3 1/8]:

- Namhyung suggested adding a fallback to the 'swfilt' term for ibs_op instead of only skipping the test on a PMU that refuses per-thread memory events. The new PATCH 9 does that for the ibs_op memory events, using the swfilt software privilege filter the kernel exposes as a format term since v6.14, which makes the per-thread 'perf mem record' the test does work on AMD; the skip stays for kernels without it. The term is in the event name even for records that end up without exclude bits, e.g. plain system wide ones: 'perf record' adds those bits itself when the first open fails with EACCES, after the name was built, and the retry only succeeds with the term already in the name:

$ perf record -v -e ibs_op/ldlat=0/ -- true
kernel.perf_event_paranoid=2, trying to fall back to excluding kernel and hypervisor samples Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed. Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.

with no exclude bits the kernel discards nothing. The 'Test data symbol' shell test matches the event string, so its regex now accepts terms added after ldlat=150.

PATCH 2/8, "perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod" [sashiko-bot review of PATCH v4 2/8]:

- debuginfod__setup_urls_env() moved from util/debuginfo.c to util/util.c, next to perf_debuginfod_setup(), and its declaration from util/debuginfo.h to util/util.h: debuginfo.o is only built with CONFIG_LIBDW, so with libdebuginfod but without libdw the declaration was an empty stub and the setup was skipped exactly for build-id.c and probe-event.c, the debuginfod users that are built without libdw and that the commit message lists as reading DEBUGINFOD_URLS in every debuginfod_begin().

PATCH 8/8, "perf mem record: Request PERF_SAMPLE_CPU by default" [sashiko-bot review of PATCH v4 8/8]:

- The rec_argv array now reserves room for the fixed arguments __cmd_record() adds, up to eight with all the optional flags; a pre-existing under-allocation, the array was nine arguments per memory PMU plus the user arguments, and on a PMU with separate load and store events the four event arguments plus the fixed ones already filled it to the last slot, so this new --sample-cpu argument would write past it, e.g. with --phys-data --data-page-size --all-user.

Patches 1, 3, 4, 5, 6 and 7 are unchanged from v4: sashiko-bot found no issues in them.

What changed from v3 (2433e9f0bf86cfc3):

PATCH 2/8, "perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod" [sashiko-bot review of PATCH v3 2/8]:

- DEBUGINFOD_URLS is no longer set from the fetch path: setenv() is not thread safe and libdebuginfod reads it with getenv() in every debuginfod_begin(), which perf also does from build-id.c, probe-event.c and probe-finder.c, paths that don't go through the fetch lock, so a thread taking one of them could race with the setenv() a fetch was doing. This is now done from symbol__init(), before the threads are started, from the .urls files in /etc/debuginfod when it isn't set, and the pthread_once() and the call in debuginfod__fetch() are gone with it.

- The in-flight fetch wait was dead code and is gone: with debuginfod__fetch_lock held across the whole lookup and fetch, no second thread could ever see an entry being fetched, so the condition variable, the waiters and the publication dance could not run. debuginfo_lookup__new() became debuginfo_lookup__add(), called after a fetch that brought a file back: lookups remember only the successful fetches, one entry each, answered with a strdup() and re-fetched if that file goes away, and a request for a build ID being fetched waits for the lock and is answered from the misses list or from the entry, so a download still can't happen twice. debuginfod__missed_lock, always taken under the fetch lock, is gone with it.

PATCH 3/8, "perf symbol: Fall back to fetching the vmlinux by build ID" [sashiko-bot review of PATCH v3 3/8]:

- The check for a concurrent load, made after taking dso->lock back from the fetch, is now dso__has_symbols() and not dso__loaded():
dso__load() sets the latter even for an attempt that failed, so a thread whose fetch failed could make a thread whose fetch succeeded discard the vmlinux it just downloaded and carry on with an empty kernel dso, as loaded is set and nothing would try again.

PATCH 6/8, "perf scripts: Add perf-stuck, to tell where a running perf is stuck" [sashiko-bot review of PATCH v3 6/8]:

- The review asked whether the %s of a dwarf_diename() that returns NULL in perf-die-chain would abort the macro. Checked, not changed:
with GDB 17.2, printf prints "(null)" for a NULL char *, including for an inferior call returning one, and the macro carries on. The other %s in the file, dloc->ms->sym->name in perf-dso, can't be NULL
either: the only caller of find_data_type() dereferences ms->sym and ms->map to build the data_loc_info before that frame can exist, so no change there as well.

- Patches 1, 4, 5, 7 and 8 are unchanged from v3: sashiko-bot found no issues in them.

What changed from v2 (00cffc7df5abfa68):

PATCH 2/8, "perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod" [sashiko-bot review of PATCH v2 2/8]:

- The conditions to fetch, debuginfod being enabled, the build-id cache being on and the build ID not having been a miss already, are now looked at with the fetch lock held, as they are the state the fetch itself changes: deciding them outside the lock and fetching inside it let a second thread repeat the fetch the first one had just made, or repeat one that had just been turned off, and put the same build ID on the misses list twice.

- A build ID that is already being fetched is waited for, instead of fetched again: there was nothing that said "this one is in flight", so a second thread wanting it blocked on the fetch lock and then did its own complete fetch, which, while the first one is still running, is a second download of the same file, with a second client, a second progress line and a second turn at the terminal, for the same answer. It now sleeps on a condition variable and is woken with whatever the fetch settled: the file, when there is one, or the not available, when it found nothing or the user cancelled it, with no retry.

- And a fetch that brought a file back is remembered, so that the ones that need the same build ID later, e.g. annotating another symbol of the same DSO, are answered with a strdup() instead of another client:
the file stays in the debuginfod client cache, and its path is checked before being handed out, in case that cache gets cleaned from under us.

- The fetches are still one at a time, so a fetch for one build ID waits for a fetch for another one: the terminal mode, the signal dispositions, the progress line and the 's'/'d' keys are process global, and sharing them between concurrent fetches, refcounted, is left for later, noted in the code. It matters less than it sounds, now that the build IDs of a workload are answered from memory after the first pass over them.

PATCH 3/8, "perf symbol: Fall back to fetching the vmlinux by build ID" [sashiko-bot review of PATCH v2 3/8]:

- The fetch is no longer made with dso->lock held: dso__load() holds it across dso__load_kernel_sym(), and a fetch can block for a long time on the network, or on the terminal, waiting for the user, which would stall every other thread that needs the kernel dso. It is dropped just around the fetch, nothing of the dso is touched while it is, and the symbols are loaded with the lock held again, the same way dso__debuginfo() already does it for the debuginfo of a DSO. If another thread got the symbols for the dso in the meantime, the file that came back is not loaded a second time.

PATCH 5/8, "perf report: Add --progress option" [sashiko-bot review of PATCH v2 5/8]:

- A phase that starts when the progress stack is full is now just not shown, and its finish() is swallowed, instead of force completing the phase that encloses it to make room: that one still gets its own finish() later, which would then complete the phase above it, and so on, one premature completion per nesting level. Reproduced by building with STDIO_PROGRESS__MAX_DEPTH set to 1: v2 prints "Processing events... [100.0%]" while it is at 99.7%, before the nested phase even runs, v3 warns and keeps the outer phase to complete when it really does. With the 8 slots in the tree this cannot be hit, the phases perf has nest at most 2 deep, but the bookkeeping was wrong.

PATCH 6/8, "perf scripts: Add perf-stuck, to tell where a running perf is stuck" [sashiko-bot review of PATCH v2 6/8]:

- The perf-dso GDB macro read dloc->ms.map->dso->name and

    dloc->ms.sym->name, but struct data_loc_info.ms is a
    'struct map_symbol *', so GDB's C evaluator needs -> there: as it
    was, 'perf-stuck.sh -g' would have failed to evaluate it.

- The one pre-existing issue in PATCH 7/8, the unbounded member type chase in die_get_member_type() and the unbounded recursion in die_find_member(), is not addressed here, it went to the perf TODO list as item 180, as it is not on the path this series fixes.

Other changes since v2:

- PATCH 2/8: a refetch made because the cached file was cleaned from under us now updates the entry the same way the first fetch does, and an entry with no path and no error is fetched again instead of answered; as it was, the entry kept the first fetch's success with no path, so the later lookups it existed for returned success with a NULL path, which debuginfo__new_build_id() cannot open and the vmlinux fallback in PATCH 3 would take as a file that came back and dereference.

- PATCH 3/8: the commit message said the ignore_vmlinux_buildid that 'perf record' and 'perf probe' set internally is what keeps them away from the fallback, but 'perf probe' only sets it for the commands other than --list, --del and --add when given an offline vmlinux; the message now says that and that a 'perf probe' that does not set it, e.g. --funcs or --add without --vmlinux on a system with a restricted /proc/kallsyms, can have the kernel debuginfo fetched, like the other tools.

- PATCH 7/8: the kerneldoc of die_get_type_die() said the type would be looked up in the main file and then in the alt file, using the one that has a DIE with the saved tag, but the function deliberately resolves in only the file the offset was recorded as belonging to, with no fallback, and @from_alt was missing from the parameter list; it now describes that, and the die_collect_vars() and die_collect_global_vars() kerneldocs, which still said the type could be retrieved with dwarf_offdie() and the offset alone, now point at die_get_type_die().

- PATCH 7/8: the comment on the member->truncated flag said the browser reads it, but the browser picks between ';' and '{' from the children list, so a member cut by the nesting limit is drawn like a complete leaf; the comment now says the flag is for the JSON exporter added in a later series.

- PATCH 2/8: a search aborted by the user, with 's' or a signal, is now remembered for the rest of the session: the threads already waiting for that fetch were told there was nothing to share, but a request that arrived later started a new download of the same file, which the user may have skipped for being too big. debuginfod__misses now records such a cancellation as a cancellation, the 's' message says the build ID will not be fetched again in this session and the debug message tells a cancellation and a server miss apart. The lookup entry is also now allocated and published by a debuginfo_lookup__new() helper called before the fetch, instead of an allocation whose failure was only checked after it.

- Patches 1, 4 and 8 are unchanged from v2: sashiko-bot found no issues in them.

What changed from v1 (060dccedab1617dc):

PATCH 2/8, "perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod" [sashiko-bot review of PATCH 2/8]:

- Included <limits.h> for PATH_MAX, that was coming in through a glibc transitive include and fails to build with musl libc.

- debuginfod__setup_urls_env() now goes through pthread_once, so that the setenv() of DEBUGINFOD_URLS happens exactly once: setenv() is not thread safe and this is on the fetch path, which dso__debuginfo() deliberately takes outside dso__lock.

- The fetch now runs with debuginfod__fetch_lock held, so that the process global state it uses — the terminal settings, the SIGINT/ SIGTERM dispositions and the progress and cancellation state — is only ever touched by one fetch at a time. A second concurrent fetch would otherwise have taken the first one's raw mode as the state to restore, leaving the terminal broken when it was done, and would reset the cancellation state from under the fetch already in progress. Serializing also keeps two fetches from racing for the same keypresses and the same progress line, and a second fetch has nothing to gain from running in parallel with a first one reading the same kind of file off the same servers.

- An interrupt that lands after the fetch already succeeded is no longer swallowed: at that point the terminal and the signal dispositions are the original ones again, so the signal is re-raised, the same way the failure path does, instead of perf carrying on as if the user had not asked it to stop.

PATCH 6/8, "perf scripts: Add perf-stuck, to tell where a running perf is stuck" [sashiko-bot review of PATCH 6/8]:

- /proc/<pid>/stat is now parsed with the parenthesized command name removed first, so that a process with spaces in its name no longer shifts every field after it: watching a process named 'my sleep' used to report "sleep)" as its state and a garbage RSS.

- The CPU time is printed with awk's printf "%d" instead of relying on its default output format, that switches to scientific notation, which the bash arithmetic below cannot parse, once the sum of utime and stime goes past six digits, i.e. some 16 minutes of CPU at 100 Hz.

- The process can go away between the '-d /proc/<pid>' check and the read of /proc/<pid>/stat, so the read is now checked: a failed or empty one reports "process gone" instead of 'set -u' aborting the script on an unbound field, which is what would have hidden the very exit the script is meant to report.

- The command line printed when the script starts has its control characters stripped, so that a process started with escape sequences in its arguments, e.g. one replaying a log line, does not get them replayed on the terminal of whoever runs this.

- The two pre-existing issues sashiko-bot raised that are outside the scope of this series, the child entries and their hists arrays leaked by the data type browser teardown and the missing <stdio.h> and <stdlib.h> includes in ui/browsers/annotate-data.c, are recorded as items 178 and 179 of the perf TODO list for follow-up work.

- Patches 1, 3, 4, 5, 7 and 8 are unchanged from v1: sashiko-bot found no issues in them.

Arnaldo Carvalho de Melo (9):
perf test: Skip data_type_profiling when the PMU cannot record memory
events
perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod perf symbol: Fall back to fetching the vmlinux by build ID perf annotate-data: Show the sample count in the data-type browser perf report: Add --progress option
perf scripts: Add perf-stuck, to tell where a running perf is stuck perf annotate-data: Resolve type DIEs in the debug file they came from perf mem record: Request PERF_SAMPLE_CPU by default
perf mem record: Use the IBS swfilt filter when available

 tools/build/feature/test-libdw.c              |  15 +-
 tools/perf/Documentation/perf-annotate.txt    |  10 +
 tools/perf/Documentation/perf-config.txt      |  19 +
 tools/perf/Documentation/perf-mem.txt         |   4 +
 tools/perf/Documentation/perf-report.txt      |  28 ++
 tools/perf/Documentation/perf-top.txt         |  14 +
 tools/perf/Makefile.config                    |   2 +-
 tools/perf/arch/x86/util/mem-events.c         |  17 +-
 tools/perf/builtin-annotate.c                 |   2 +
 tools/perf/builtin-mem.c                      |  18 +-
 tools/perf/builtin-report.c                   |  17 +
 tools/perf/builtin-top.c                      |   6 +
 tools/perf/scripts/perf-stuck.gdb             | 104 +++++
 tools/perf/scripts/perf-stuck.sh              | 194 ++++++++++
 tools/perf/tests/shell/data_type_profiling.sh |  46 ++-
 tools/perf/tests/shell/test_data_symbol.sh    |   2 +-
 tools/perf/ui/Build                           |   1 +
 tools/perf/ui/browsers/annotate-data.c        |   2 +-
 tools/perf/ui/progress.h                      |   2 +
 tools/perf/ui/stdio/progress.c                | 185 +++++++++
 tools/perf/util/annotate-data.c               |  59 ++-
 tools/perf/util/annotate-data.h               |   3 +
 tools/perf/util/config.c                      |   3 +
 tools/perf/util/debuginfo.c                   | 523 ++++++++++++++++++++++++++
 tools/perf/util/debuginfo.h                   |  36 ++
 tools/perf/util/dso.c                         |  19 +
 tools/perf/util/dwarf-aux.c                   | 171 +++++++--
 tools/perf/util/dwarf-aux.h                   |  33 ++
 tools/perf/util/mem-events.c                  |  21 +-
 tools/perf/util/mem-events.h                  |   2 +
 tools/perf/util/ordered-events.c              |  17 +-
 tools/perf/util/session.c                     |  12 +-
 tools/perf/util/symbol.c                      |  83 +++-
 tools/perf/util/symbol_conf.h                 |   1 +
 tools/perf/util/util.c                        |  77 ++++
 tools/perf/util/util.h                        |   8 +
 36 files changed, 1691 insertions(+), 65 deletions(-)

base-commit: aa18964dd64511305de0711fed912054da6f5d18
v4-head: f801de4e80289566670e6b735198606b672e74e8
v3-head: 2433e9f0bf86cfc3d9966b467042563596c9ebdb
v2-head: 00cffc7df5abfa6850beb360deb830d8495163ce
v1-head: 060dccedab1617dcbf6e2be64ed97afe3f3fac63 --

Assisted-by LLM
author maintainer
#1 lore raw

PATCH v5 1/9 perf test: Skip data_type_profiling when the PMU cannot record memory events

From: Arnaldo Carvalho de Melo <[email protected]>

The test records with 'perf mem record' in per-thread mode and its only guard matches one specific message:

  perf mem record -o /dev/null -- true 2>&1 | \
        grep -q "failed: no PMU supports the memory events" && exit 2

A PMU that has memory events but refuses them per-thread falls through it, and AMD IBS does:

  $ perf mem record -o /dev/null -- true
  Error:
  Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed.
  Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.
  Error:
  Failure to open any events for recording.

What follows is not a skip either. The script runs under 'set -e', so the bare 'perf mem record' in test_basic_annotate() aborts it through the EXIT trap, which reports a signal that never happened and exits 1:

Basic Rust perf annotate test
Unexpected signal in test_basic_annotate

so 'perf test' turns "this PMU cannot record these events" into a test failure:

  $ perf test "data type profiling"
   87: perf data type profiling tests             : FAILED!

The 'if [ "x$?" != "x0" ]' right below the record was meant to report "[Failed: perf record]" and carry on with the other cases, but under 'set -e' it is dead code: the shell is gone before it is reached. The same goes for the 'perf annotate' call below it.

Skip when a trivial per-thread 'perf mem record' fails, and put both commands in the condition of an 'if', where 'set -e' leaves them alone, so that the
reporting works for the cases where only one workload cannot be recorded or annotated:

  $ perf test "data type profiling"
   87: perf data type profiling tests             : Skip

and, with just the new skip guard disabled on the same box, every case reports its failure instead of the first one aborting the run:

Basic Rust perf annotate test
Basic annotate [Failed: perf record]
Pipe Rust perf annotate test
Pipe annotate [Failed: perf record]
Basic C perf annotate test
Basic annotate [Failed: perf record]
Pipe C perf annotate test
Pipe annotate [Failed: perf record]

---
 tools/perf/tests/shell/data_type_profiling.sh | 46 +++++++++++++++----
 1 file changed, 36 insertions(+), 10 deletions(-)
tools/perf/tests/shell/data_type_profiling.sh +36 -10
index eca694600a0478d2..a916c410274aa888 100755
@@ -19,6 +19,15 @@ perfout=$(mktemp /tmp/__perf_test.perf.out.XXXXX)
 perf mem record -o /dev/null -- true  2>&1 | \
   		grep -q "failed: no PMU supports the memory events" && exit 2
 
+# Skip if per-thread mem record is not supported on this PMU (e.g. AMD IBS
+# needs system-wide '-a'): it is what the test records with below, and a
+# failing record must not be reported as a test failure.
+if ! perf mem record -o /dev/null -- true 2>/dev/null
+then
+  echo "Skip: cannot record memory events on this PMU"
+  exit 2
+fi
+
 cleanup() {
   rm -rf "${perfdata}" "${perfout}"
   rm -rf "${perfdata}".old
@@ -52,25 +61,42 @@ test_basic_annotate() {
     index=1 ;;
   esac
 
+  # Under 'set -e' a bare failing command aborts the script through the EXIT
+  # trap, so the commands that report a failure have to be the condition of
+  # an 'if' for that reporting to ever happen.
   if [ "x${mode}" == "xBasic" ]
   then
-    perf mem record -o "${perfdata}" ${testprogs[$index]} 2> /dev/null
+    if ! perf mem record -o "${perfdata}" ${testprogs[$index]} 2> /dev/null
+    then
+      echo "${mode} annotate [Failed: perf record]"
+      err=1
+      return
+    fi
   else
-    perf mem record -o - ${testprogs[$index]} 2> /dev/null > "${perfdata}"
-  fi
-  if [ "x$?" != "x0" ]
-  then
-    echo "${mode} annotate [Failed: perf record]"
-    err=1
-    return
+    if ! perf mem record -o - ${testprogs[$index]} 2> /dev/null > "${perfdata}"
+    then
+      echo "${mode} annotate [Failed: perf record]"
+      err=1
+      return
+    fi
   fi
 
   # Generate the annotated output file
   if [ "x${mode}" == "xBasic" ]
   then
-    perf annotate --code-with-type -i "${perfdata}" --stdio --percent-limit 1 2> /dev/null > "${perfout}"
+    if ! perf annotate --code-with-type -i "${perfdata}" --stdio --percent-limit 1 2> /dev/null > "${perfout}"
+    then
+      echo "${mode} annotate [Failed: perf annotate]"
+      err=1
+      return
+    fi
   else
-    perf annotate --code-with-type -i - --stdio 2> /dev/null --percent-limit 1 < "${perfdata}" > "${perfout}"
+    if ! perf annotate --code-with-type -i - --stdio 2> /dev/null --percent-limit 1 < "${perfdata}" > "${perfout}"
+    then
+      echo "${mode} annotate [Failed: perf annotate]"
+      err=1
+      return
+    fi
   fi
 
   # check if it has the target data type
Signature · 1 line
2.55.0
author maintainer
#3 lore raw

PATCH v5 2/9 perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

perf already uses debuginfod to fetch source files when annotating (via probe-finder.c) and 'perf probe' has open_from_debuginfod(), which queries debuginfo keyed by build ID when a module's debuginfo isn't found locally, largely the same thing this adds; eventually that one could be moved over to the new helper. For the analysis tools there was no way to obtain the debuginfo for a DSO in a profile when it isn't available locally under the name the DSO was opened with, for instance the vmlinux for the kernel a profile was recorded on when processing it on another machine, or after the kernel and its debuginfo package got upgraded in between.

Add debuginfo__find_build_id(), that uses the debuginfod client to locate a debuginfo file keyed by the build ID, checking its local cache first and then querying the servers in DEBUGINFOD_URLS, and debuginfo__new_build_id(), that opens the DWARF in the file it finds.

The debuginfod client fails when DEBUGINFOD_URLS isn't set even when what it wants is in its local cache, and the distro setup scripts that populate it from /etc/debuginfod don't reach cron jobs, systemd services and other environments that don't source the profile scripts, so also set it from the .urls files in /etc/debuginfod when not set. That has to happen before any thread that can call getenv() is started, as setenv() is not thread safe, and there are getenv()s outside the fetch
lock: libdebuginfod reads DEBUGINFOD_URLS in every debuginfod_begin(), which perf also does in build-id.c, probe-event.c and probe-finder.c, so do it from symbol__init(), on the single-threaded setup, and not lazily from the fetch path.

Querying servers, possibly third party ones, sends off-box the build IDs of the binaries being analysed and a fetch can take a while, so this is opt-out: on by default, off with --no-debuginfod, with core.debuginfod=false, per tool with report.debuginfod and top.debuginfod, and, since users that set buildid.dir to /dev/null (e.g. Linus) or otherwise turn the local build-id cache off clearly don't want fetched files stored on the box, off too in that case.

When a fetch is in progress in a terminal, stdio, the way it prints progress is how one gets out of it: 's' aborts the current fetch via the debuginfod client's progress callback protocol and remembers the build ID, so that the rest of the session doesn't ask for it again, the user may have skipped it for being too big; 'd' additionally disables debuginfod for the rest of the session and points at 'perf config core.debuginfod=false' to make that permanent -- rewriting the user's ~/.perfconfig from a keypress would silently drop its comments -- and SIGINT/SIGTERM are intercepted while the terminal is in raw mode, so that it is restored and the signal is re-raised when the user interrupts a fetch.

The fetches are serialized, as the terminal settings, the signal dispositions and the progress and cancellation state they use are process global, so two concurrent fetches would fight over them, and serializing also means that a request for a build ID that is being fetched doesn't start a second download of the same file: it waits for the fetch to finish and is answered from the entry the fetch published, keyed by the build ID, with the path of the file that came back, checked before being handed out in case the debuginfod client cache is cleaned from under us.

Make dso__debuginfo() use debuginfo__new_build_id() as a fallback, keyed by the build ID recorded in the perf.data file, so that consumers such as the data type profiler can resolve the types of DSOs whose debuginfo can be fetched this way. Do the fetch outside dso__lock and remember the build IDs that were a miss and the ones whose search the user cancelled, so that consumers revisiting a set of DSOs repeatedly, such as the data type profiler on every hist entry DSO switch, don't pay server round trips per attempt and a cancelled download, maybe a file the user found too big, isn't restarted by the next request for the same build ID in the same session.

debuginfo__new_build_id() needs libdw to open the DWARF, so it lives in debuginfo.o, built only with CONFIG_LIBDW, while the libdebuginfod feature check is independent of NO_LIBDW; keep the new build ID prototypes under HAVE_LIBDW_SUPPORT too, with stubs otherwise, so that make NO_LIBDW=1 on a system that has the debuginfod client keeps linking.

---
 tools/perf/Documentation/perf-annotate.txt |  10 +
 tools/perf/Documentation/perf-config.txt   |  19 +
 tools/perf/Documentation/perf-report.txt   |  15 +
 tools/perf/Documentation/perf-top.txt      |  14 +
 tools/perf/builtin-annotate.c              |   2 +
 tools/perf/builtin-report.c                |   6 +
 tools/perf/builtin-top.c                   |   6 +
 tools/perf/util/config.c                   |   3 +
 tools/perf/util/debuginfo.c                | 523 +++++++++++++++++++++
 tools/perf/util/debuginfo.h                |  36 ++
 tools/perf/util/dso.c                      |  19 +
 tools/perf/util/symbol.c                   |  12 +
 tools/perf/util/symbol_conf.h              |   1 +
 tools/perf/util/util.c                     |  77 +++
 tools/perf/util/util.h                     |   8 +
 15 files changed, 751 insertions(+)
tools/perf/Documentation/perf-annotate.txt +10 -0
index 1a90b09a12d5abb1..25af1d166dc4c877 100644
@@ -58,6 +58,16 @@ OPTIONS
 --ignore-vmlinux::
 	Ignore vmlinux files.
 
+--debuginfod::
+--no-debuginfod::
+	Fetch debuginfo keyed by build ID from the debuginfod servers
+	configured in DEBUGINFOD_URLS, checking the local debuginfod
+	client cache first, when it is not available locally, on for
+	these commands by default.  See the --debuginfod option of
+	'perf report' for how to turn it off, including the 's' and
+	'd' keys that skip a fetch in progress while 'perf' is
+	waiting for it.
+
 --itrace::
 	Options for decoding instruction tracing data. The options are:
 
tools/perf/Documentation/perf-config.txt +19 -0
index 9b223f8928299945..688306abe847a0df 100644
@@ -216,6 +216,14 @@ core.*::
 	addr2line-timeout::
 		Sets a timeout (in milliseconds) for parsing 'addr2line'
 		output.  The default timeout is 5s.
+	debuginfod::
+		When set to 'false', disable fetching debuginfo keyed by
+		build ID from the debuginfod servers configured in
+		DEBUGINFOD_URLS.  It is on by default, can be overridden per
+		tool with the 'report.debuginfod' and 'top.debuginfod'
+		options and per invocation with --no-debuginfod; it is off
+		too when the local build-id cache is disabled, e.g.
+		'buildid.dir' set to /dev/null.
 
 tui.*, gtk.*::
 	Subcommands that can be configured here are 'top', 'report' and 'annotate'.
@@ -562,6 +570,14 @@ report.*::
 		This option can change default stat behavior with empty results.
 		If it's set true, 'perf report --stat' will not show 0 stats.
 
+	report.debuginfod::
+		Fetch debuginfo keyed by build ID from the debuginfod
+		servers configured in DEBUGINFOD_URLS, checking the local
+		debuginfod client cache first, when it is not available
+		locally.  On by default, set to 'false' to disable it for
+		'perf report', globally with 'core.debuginfod=false' or per
+		invocation with --no-debuginfod.
+
 top.*::
 	top.children::
 		Same as 'report.children'. So if it is enabled, the output of 'top'
@@ -569,6 +585,9 @@ top.*::
 		column by default.
 		The default is 'true'.
 
+	top.debuginfod::
+		Same as 'report.debuginfod', for 'perf top'.
+
 	top.call-graph::
 		This is identical to 'call-graph.record-mode', except it is
 		applicable only for 'top' subcommand. This option ONLY setup
tools/perf/Documentation/perf-report.txt +15 -0
index ae68ca402d0b4f0e..fed6af128ff07e4c 100644
@@ -366,6 +366,21 @@ OPTIONS
 --ignore-vmlinux::
 	Ignore vmlinux files.
 
+--debuginfod::
+--no-debuginfod::
+	Fetch debuginfo keyed by build ID from the debuginfod servers
+	configured in DEBUGINFOD_URLS, checking the local debuginfod
+	client cache first, when it is not available locally, on for
+	these commands by default.  It can be turned off per invocation
+	with --no-debuginfod, per tool with the "report.debuginfod"
+	config option or globally with "core.debuginfod" set to false.
+	While a fetch is in progress in the stdio interface, 's' skips
+	the current fetch and 'd' skips it and disables debuginfod for
+	the rest of the session, pointing at 'perf config' to make that
+	permanent.  It is disabled as well when the local build-id cache
+	is turned off, e.g. "buildid.dir" set to /dev/null, as that
+	asks for fetched files not to be kept on the box.
+
 --kallsyms=<file>::
         kallsyms pathname
 
tools/perf/Documentation/perf-top.txt +14 -0
index 2da2a16bbf260685..8bcb7b0ac4a2407a 100644
@@ -83,6 +83,20 @@ Default is to monitor all CPUS.
 --ignore-vmlinux::
 	Ignore vmlinux files.
 
+--debuginfod::
+--no-debuginfod::
+	Fetch debuginfo keyed by build ID from the debuginfod servers
+	configured in DEBUGINFOD_URLS, checking the local debuginfod
+	client cache first, when it is not available locally, on for
+	these commands by default.  Turn it off per invocation with
+	--no-debuginfod, with the "top.debuginfod" config option or
+	globally with "core.debuginfod" set to false.  While a fetch is
+	in progress in the stdio interface, 's' skips the current fetch
+	and 'd' skips it and disables debuginfod for the rest of the
+	session, pointing at 'perf config' to make that permanent.
+	Disabled as well when the build-id cache is off, e.g.
+	"buildid.dir" set to /dev/null.
+
 --kallsyms=<file>::
 	kallsyms pathname
 
tools/perf/builtin-annotate.c +2 -0
index 4638e6fdc39bb6b7..d14ae7d1c345cb79 100644
@@ -733,6 +733,8 @@ int cmd_annotate(int argc, const char **argv)
 	OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
                     "don't load vmlinux even if found"),
+	OPT_BOOLEAN(0, "debuginfod", &symbol_conf.debuginfod,
+                    "fetch debuginfo keyed by build ID from the debuginfod servers, on by default, use --no-debuginfod to turn off"),
 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 		   "file", "vmlinux pathname"),
 	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
tools/perf/builtin-report.c +6 -0
index b280ff9ff45e6485..4d3383d1daae2ed9 100644
@@ -134,6 +134,10 @@ static int report__config(const char *var, const char *value, void *cb)
 		symbol_conf.event_group = perf_config_bool(var, value);
 		return 0;
 	}
+	if (!strcmp(var, "report.debuginfod")) {
+		symbol_conf.debuginfod = perf_config_bool(var, value);
+		return 0;
+	}
 	if (!strcmp(var, "report.percent-limit")) {
 		double pcnt = strtof(value, NULL);
 
@@ -1346,6 +1350,8 @@ int cmd_report(int argc, const char **argv)
 		   "file", "vmlinux pathname"),
 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
                     "don't load vmlinux even if found"),
+	OPT_BOOLEAN(0, "debuginfod", &symbol_conf.debuginfod,
+                    "fetch debuginfo keyed by build ID from the debuginfod servers, on by default, use --no-debuginfod to turn off"),
 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 		   "file", "kallsyms pathname"),
 	OPT_BOOLEAN('f', "force", &symbol_conf.force, "don't complain, do it"),
tools/perf/builtin-top.c +6 -0
index c2562d49be46a9a1..aed45167d95005dd 100644
@@ -1436,6 +1436,10 @@ static int perf_top_config(const char *var, const char *value, void *cb __maybe_
 		symbol_conf.cumulate_callchain = perf_config_bool(var, value);
 		return 0;
 	}
+	if (!strcmp(var, "top.debuginfod")) {
+		symbol_conf.debuginfod = perf_config_bool(var, value);
+		return 0;
+	}
 
 	return 0;
 }
@@ -1508,6 +1512,8 @@ int cmd_top(int argc, const char **argv)
 		   "file", "vmlinux pathname"),
 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
 		    "don't load vmlinux even if found"),
+	OPT_BOOLEAN(0, "debuginfod", &symbol_conf.debuginfod,
+                    "fetch debuginfo keyed by build ID from the debuginfod servers, on by default, use --no-debuginfod to turn off"),
 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 		   "file", "kallsyms pathname"),
 	OPT_BOOLEAN('K', "hide_kernel_symbols", &top.hide_kernel_symbols,
tools/perf/util/config.c +3 -0
index b2972c35c1eca68c..31c6618d3b3daf22 100644
@@ -470,6 +470,9 @@ static int perf_default_core_config(const char *var, const char *value)
 	if (!strcmp(var, "core.addr2line-disable-warn"))
 		symbol_conf.addr2line_disable_warn = perf_config_bool(var, value);
 
+	if (!strcmp(var, "core.debuginfod"))
+		symbol_conf.debuginfod = perf_config_bool(var, value);
+
 	/* Add other config variables here. */
 	return 0;
 }
tools/perf/util/debuginfo.c +523 -0

546 lines not on this page · raw message

tools/perf/util/debuginfo.h +36 -0
index a52d69932815cd72..43b211a0dec174f5 100644
@@ -5,6 +5,8 @@
 #include <errno.h>
 #include <linux/compiler.h>
 
+struct build_id;
+
 #ifdef HAVE_LIBDW_SUPPORT
 
 #include "dwarf-aux.h"
@@ -54,6 +56,28 @@ static inline int debuginfo__get_text_offset(struct debuginfo *dbg __maybe_unuse
 #ifdef HAVE_DEBUGINFOD_SUPPORT
 int get_source_from_debuginfod(const char *raw_path, const char *sbuild_id,
 			       char **new_path);
+
+/*
+ * Finding a debuginfo file keyed by build ID uses the debuginfod client,
+ * but opening the DWARF in it needs libdw, i.e. these live in
+ * debuginfo.o, which is only built with CONFIG_LIBDW.
+ */
+#ifdef HAVE_LIBDW_SUPPORT
+int debuginfo__find_build_id(const struct build_id *bid, char **path);
+struct debuginfo *debuginfo__new_build_id(const struct build_id *bid);
+#else
+static inline int debuginfo__find_build_id(const struct build_id *bid __maybe_unused,
+					   char **path __maybe_unused)
+{
+	return -ENOTSUP;
+}
+
+static inline struct debuginfo *
+debuginfo__new_build_id(const struct build_id *bid __maybe_unused)
+{
+	return NULL;
+}
+#endif /* HAVE_LIBDW_SUPPORT */
 #else /* HAVE_DEBUGINFOD_SUPPORT */
 static inline int get_source_from_debuginfod(const char *raw_path __maybe_unused,
 					     const char *sbuild_id __maybe_unused,
@@ -61,6 +85,18 @@ static inline int get_source_from_debuginfod(const char *raw_path __maybe_unused
 {
 	return -ENOTSUP;
 }
+
+static inline int debuginfo__find_build_id(const struct build_id *bid __maybe_unused,
+					   char **path __maybe_unused)
+{
+	return -ENOTSUP;
+}
+
+static inline struct debuginfo *
+debuginfo__new_build_id(const struct build_id *bid __maybe_unused)
+{
+	return NULL;
+}
 #endif /* HAVE_DEBUGINFOD_SUPPORT */
 
 #endif /* _PERF_DEBUGINFO_H */
tools/perf/util/dso.c +19 -0
index 42bfe30a3b518e80..fe7c3b0265ba633a 100644
@@ -32,6 +32,7 @@
 #include "string2.h"
 #include "vdso.h"
 #include "annotate-data.h"
+#include "debuginfo.h"
 #include "libdw.h"
 
 static const char * const debuglink_paths[] = {
@@ -2073,5 +2074,23 @@ struct debuginfo *dso__debuginfo(struct dso *dso)
 
 	mutex_unlock(dso__lock(dso));
 	free(name);
+
+	/*
+	 * The debuginfo for a DSO in the profile may not be installed
+	 * locally, for instance the vmlinux for the kernel the profile was
+	 * recorded on when processing it on another machine, or after the
+	 * kernel and its debuginfo got upgraded in between.  Fall back to
+	 * fetching it keyed by the build ID recorded in the perf.data file,
+	 * using the debuginfod client, which checks its local cache first.
+	 *
+	 * Do it outside dso__lock, a fetch from a remote debuginfod server
+	 * can take a while and would otherwise block anything else using
+	 * this dso, and honour the opt-out, the user may have asked for
+	 * no debuginfod via --no-debuginfod, core.debuginfod=false or by
+	 * disabling the build-id cache.
+	 */
+	if (dinfo == NULL)
+		dinfo = debuginfo__new_build_id(dso__bid(dso));
+
 	return dinfo;
 }
tools/perf/util/symbol.c +12 -0
index 3587ad243159074f..1d70c43bb24ee279 100644
@@ -76,6 +76,8 @@ struct symbol_conf symbol_conf = {
 	.inline_name		= true,
 	.res_sample		= 0,
 	.addr2line_timeout_ms	= 5 * 1000,
+	/* Fetching debuginfo by build ID, off via --no-debuginfod, etc */
+	.debuginfod		= true,
 };
 
 struct map_list_node {
@@ -2534,6 +2536,16 @@ int symbol__init(struct perf_env *env)
 	if (symbol_conf.initialized)
 		return 0;
 
+	/*
+	 * Have DEBUGINFOD_URLS set before the threads that can take the
+	 * debuginfod fetch path, and that getenv() it, e.g. in the every
+	 * debuginfod_begin() of the other debuginfod client users, are
+	 * started: set it up here, from the .urls files in /etc/debuginfod
+	 * when it isn't set, instead of doing it lazily from the fetch
+	 * path itself, setenv() not being thread safe.
+	 */
+	debuginfod__setup_urls_env();
+
 	symbol_conf.priv_size = PERF_ALIGN(symbol_conf.priv_size, sizeof(u64));
 
 	symbol__elf_init();
tools/perf/util/symbol_conf.h +1 -0
index 71f60081a85bb18d..a56b1d2d843b9ff1 100644
@@ -45,6 +45,7 @@ struct symbol_conf {
 			force,
 			ignore_vmlinux,
 			ignore_vmlinux_buildid,
+			debuginfod,
 			show_kernel_path,
 			use_modules,
 			allow_aliases,
tools/perf/util/util.c +77 -0

83 lines not on this page · raw message

tools/perf/util/util.h +8 -0
index f8f1ff603c728086..1f763f9960bd0756 100644
@@ -82,6 +82,14 @@ struct perf_debuginfod {
 };
 void perf_debuginfod_setup(struct perf_debuginfod *di);
 
+#ifdef HAVE_DEBUGINFOD_SUPPORT
+void debuginfod__setup_urls_env(void);
+#else
+static inline void debuginfod__setup_urls_env(void)
+{
+}
+#endif
+
 const char *perf_basename(const char *path);
 
 char *filename_with_chroot(int pid, const char *filename);
Signature · 1 line
2.55.0
author maintainer
#5 lore raw

PATCH v5 3/9 perf symbol: Fall back to fetching the vmlinux by build ID

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

A profile recorded on a kernel that is no longer installed, be it because the machine was rebooted into a new kernel or because the profile is being processed on another machine, can't have its kernel symbols resolved: /proc/kallsyms matches the running kernel, not the one in the profile, and is refused when restricted, e.g. with kernel.perf_event_paranoid > 1, while the build-id cache may carry just a kallsyms copy with zeroed addresses. With no kernel symbols the kernel samples can't be annotated, so they all end up in the '(unknown)' data type and the kernel is missing from the data type profile JSON "dsos" entry.

As a last resort, when the kernel symbols can't be found locally and the user didn't specify a kallsyms file, fetch the vmlinux keyed by the kernel build ID recorded in the perf.data file using the debuginfod client and use it for symbols, which also provides the DWARF debuginfo needed for data type profiling.

The fetch is made with dso->lock dropped, the same way dso__debuginfo() does it for the debuginfo of a DSO, and the fetched file is loaded with the lock held again, only if another thread didn't get the symbols in the meantime, checked with dso__has_symbols(): dso__loaded() is set by dso__load() even for an attempt that failed, and would make a successful fetch throw the file away.

Like the other vmlinux sources this honors --ignore-vmlinux and --ignore-vmlinux_buildid: a fetched vmlinux is still a vmlinux, and the fetch is keyed by the build ID, so both flags skip it. 'perf record' sets ignore_vmlinux_buildid internally, and so does 'perf probe' for the commands other than --list, --del and --add when given an offline vmlinux, keeping those away from the fetch as well. A 'perf probe' that doesn't set it, e.g. --funcs or --add without --vmlinux on a system with a restricted /proc/kallsyms, can have the kernel debuginfo fetched, like the other tools. The fetch itself follows the opt-out controls added in the previous patch, --no-debuginfod, core.debuginfod=false and the build-id-cache-is-off case, and can be skipped with the 's'/'d' keys while it runs. Build IDs that were a miss are remembered, not queried again on every dso__load() retry.

With a profile recorded on a system running kernel 7.1.10, later processed after it was upgraded to 7.1.13, 'perf report -s type ' went from having all 24.91% of the kernel samples as '(unknown)' data types to resolving struct task_struct, struct rq, struct tty_struct, struct qspinlock, etc, with the kernel DSO, keyed by its build ID, appearing in the data type profile JSON.

---
 tools/perf/util/symbol.c | 71 +++++++++++++++++++++++++++++++++++++++-
 1 file changed, 70 insertions(+), 1 deletion(-)
tools/perf/util/symbol.c +70 -1
index 1d70c43bb24ee279..ccb6217119416dec 100644
@@ -20,6 +20,7 @@
 #include "cap.h"
 #include "cpumap.h"
 #include "debug.h"
+#include "debuginfo.h"
 #include "demangle-cxx.h"
 #include "demangle-java.h"
 #include "demangle-ocaml.h"
@@ -2191,12 +2192,33 @@ static char *dso__find_kallsyms(struct dso *dso, struct map *map)
 	return strdup(path);
 }
 
+/*
+ * Last resort when the symbols for the kernel the profile was recorded
+ * on can't be found locally: fetch the vmlinux keyed by the build ID
+ * recorded in the perf.data file using the debuginfod client, which
+ * checks its local cache first, e.g. when processing the profile on
+ * another machine or after the kernel and its debuginfo package got
+ * upgraded in between.
+ */
+/*
+ * The fetch itself, that dso__load_kernel_sym() calls with dso->lock
+ * dropped, see the comment there.
+ */
+static int dso__fetch_vmlinux_build_id(struct dso *dso, char **path)
+{
+	if (!dso__has_build_id(dso))
+		return -1;
+
+	return debuginfo__find_build_id(dso__bid(dso), path);
+}
+
 static int dso__load_kernel_sym(struct dso *dso, struct map *map)
 {
 	int err;
 	const char *kallsyms_filename = NULL;
 	char *kallsyms_allocated_filename = NULL;
 	char *filename = NULL;
+	bool user_kallsyms = false;
 
 	/*
 	 * Step 1: if the user specified a kallsyms or vmlinux filename, use
@@ -2215,6 +2237,7 @@ static int dso__load_kernel_sym(struct dso *dso, struct map *map)
 	 */
 	if (symbol_conf.kallsyms_name != NULL) {
 		kallsyms_filename = symbol_conf.kallsyms_name;
+		user_kallsyms = true;
 		goto do_kallsyms;
 	}
 
@@ -2257,7 +2280,53 @@ static int dso__load_kernel_sym(struct dso *dso, struct map *map)
 		pr_debug("Using %s for symbols\n", kallsyms_filename);
 	free(kallsyms_allocated_filename);
 
-	if (err > 0 && !dso__is_kcore(dso)) {
+	/*
+	 * The kallsyms may be unavailable or restricted, e.g.
+	 * /proc/kallsyms with kernel.perf_event_paranoid > 1, try to fetch
+	 * the vmlinux keyed by the build ID using debuginfod as a last
+	 * resort, honoring --ignore-vmlinux and --ignore-vmlinux_buildid
+	 * like the other vmlinux sources above.
+	 */
+	if (err <= 0 && !user_kallsyms &&
+	    !symbol_conf.ignore_vmlinux &&
+	    !symbol_conf.ignore_vmlinux_buildid) {
+		char *fetched_path = NULL;
+
+		/*
+		 * dso__load() holds dso->lock while it calls us, and the
+		 * fetch below can take a long time, blocked on the network
+		 * or on the terminal, waiting for the user: do it with the
+		 * lock dropped, as dso__debuginfo() does for the debuginfo
+		 * of a DSO, so that the threads that need this dso don't get
+		 * stuck behind a server round trip.  Nothing of the dso is
+		 * touched by the fetch, the symbols are loaded with the lock
+		 * held again, and only if the fetch brought a file back.
+		 */
+		mutex_unlock(dso__lock(dso));
+		err = dso__fetch_vmlinux_build_id(dso, &fetched_path);
+		mutex_lock(dso__lock(dso));
+
+		if (err) {
+			zfree(&fetched_path);
+		} else if (dso__has_symbols(dso)) {
+			/*
+			 * Somebody else got the symbols for this dso while
+			 * the lock was dropped for the fetch, use those
+			 * instead of loading the file that came back a
+			 * second time.  dso__has_symbols() and not
+			 * dso__loaded(): dso__load() sets the latter even
+			 * when it fails, so an attempt that failed would
+			 * have us throw a file that came back away.
+			 */
+			pr_debug("%s got its symbols while its vmlinux was being fetched, using them\n",
+				 dso__name(dso));
+			zfree(&fetched_path);
+			err = 1;
+		} else {
+			/* Takes ownership of 'fetched_path' even when it fails */
+			err = dso__load_vmlinux(dso, map, fetched_path, true);
+		}
+	} else if (err > 0 && !dso__is_kcore(dso)) {
 		struct maps *kmaps = map__kmaps(map);
 
 		dso__set_binary_type(dso, DSO_BINARY_TYPE__KALLSYMS);
Signature · 1 line
2.55.0
author maintainer
#7 lore raw

PATCH v5 4/9 perf annotate-data: Show the sample count in the data-type browser

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

The data-type browser has a samples view, selected with -n (or with annotate.show_nr_samples), in which browser__write_overhead() prints a local nr_samples variable that is initialized to zero and never updated, so every member is listed as having no samples while the period and percent columns for the same entry are filled in.

Print the histogram entry's own count instead.

This predates the load/store counter split, so fix it ahead of that
patch: the split then only has to adapt a line that is already correct, and this fix can be picked on its own.

Fixes d001c7a7f4736743 ("perf annotate-data: Add hist_entry__annotate_data_tui()")
Assisted-by LLM
---
 tools/perf/ui/browsers/annotate-data.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
tools/perf/ui/browsers/annotate-data.c +1 -1
index aa8c89fe2e82c1c5..1080ed1a40d2609a 100644
@@ -370,7 +370,7 @@ static void browser__write_overhead(struct ui_browser *uib,
 	u64 period = hist->period;
 	double percent = total->period ? (100.0 * period / total->period) : 0;
 	bool current = ui_browser__is_current_entry(uib, row);
-	int nr_samples = 0;
+	int nr_samples = hist->nr_samples;
 
 	ui_browser__set_percent_color(uib, percent, current);
 
Signature · 1 line
2.55.0
author maintainer
#9 lore raw

PATCH v5 5/9 perf report: Add --progress option

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

Processing a large data type profiling session, e.g. an AMD IBS one, can take a long time and, when using the stdio output, 'perf report' gives no feedback about which phase it is in nor about how far along it is: the TUI has a progress bar for that, but in the stdio case the ui_progress updates, that are already there, are dropped on the floor.

Add a --progress option that installs a stdio backend for ui_progress, ui/stdio/progress.c, printing the phase title, the percentage done and the current/total counts to stderr:

Processing events... [ 42.3%] 317M / 746M Merging related events... [ 61.0%] 309026 / 506686 Sorting events for output... [ 98.2%] 14132 / 14387

The first one is the perf.data file size based progress already kept while reading events, sized with the same unit_number__scnprintf() used by the TUI progress bar title; the other two are the hist entry based ones for the hist entry merging (collapse) and output sorting phases, that print raw counts.

Steps are 1% of the phase total, the default (total / 16) is tuned for the character cell based TUI bar, and the last update normally stops short of the total, so finishing a phase prints it as complete.

When stderr is a tty the line is updated in place, with each phase getting a line of its own, otherwise one line is printed per update, so that redirecting stderr to a file leaves a readable log of the phases.

Phases can be nested, e.g. the ordered events flushes that take place while the "Processing events..." phase is still in progress, so the backend keeps track of the ones started so far to be able to complete the right one when a phase finishes, as ui_progress__finish() gets no arguments.

That bookkeeping requires ui_progress__init() and ui_progress__finish() to be paired, and a few callers had paths returning early without the
finish: do_flush() in ordered-events.c on session_done() and on deliver() errors, and the ENOMEM paths right after the init in __perf_session__process_pipe_events() and __perf_session__process_dir_events().

Fix those, with a backend that kept the stale entries it would print through dangling pointers to returned stack frames, and have the update() method ignore anything that doesn't match the innermost running phase, printing nothing is better than printing another phase's numbers or reading the stack array with index -1.

Committer testing:

⬢ [acme@toolbx perf-tools-next]$ perf record -o perf.data.small -- sleep 0.2 [ perf record: Woken up 2 times to write data ]
[ perf record: Captured and wrote 0.002 MB perf.data.small (7 samples) ] ⬢ [acme@toolbx perf-tools-next]$ perf report --progress -i perf.data.small > /dev/null Processing events... [ 0.0%] 0B / 2K
Processing events... [ 99.6%] 2040B / 2K
Processing time ordered events... [100.0%] 13 / 13
Processing events... [100.0%] 2K / 2K
Sorting events for output... [100.0%] 5 / 5
⬢ [acme@toolbx perf-tools-next]$

Nothing goes to stderr without --progress:

  ⬢ [acme@toolbx perf-tools-next]$ perf report --stdio -i perf.data.small > /dev/null 2> stderr.txt
  ⬢ [acme@toolbx perf-tools-next]$ wc -c stderr.txt
  0 stderr.txt
  ⬢ [acme@toolbx perf-tools-next]$

With a 321 MB perf.data (perf mem record, i7-14700K):

  ⬢ [acme@toolbx perf-tools-next]$ perf report --progress -i perf.data.i7 > /dev/null
  Processing events... [  0.0%] 0B / 306M
  [...]
  Sorting events for output... [100.0%] 29996 / 29996
  ⬢ [acme@toolbx perf-tools-next]$

Which is what prompted this: an AMD IBS data type profile session hangs in the DWARF type resolution done when merging hist entries, with --progress it now is visible that it is not the event loading that is stuck, but the merging, at a specific hist entry:

  ⬢ [acme@toolbx perf-tools-next]$ perf report --progress -s type -i perf.data.ibs > /dev/null
  Processing events... [  0.0%] 0B / 746M
  [...]
  Processing events... [100.0%] 746M / 746M
  Merging related events... [ 58.0%] 293828 / 506686
  Merging related events... [ 59.0%] 298894 / 506686
  Merging related events... [ 60.0%] 303960 / 506686
  Merging related events... [ 61.0%] 309026 / 506686

  ⬢ [acme@toolbx perf-tools-next]$ gdb -p $(pidof perf) -batch -ex 'bt 6' -ex detach
  #0  0x0000000000773f31 in die_get_pointer_type ()
  #1  0x000000000077b817 in find_data_type ()
  #2  0x0000000000636f20 in __hist_entry__get_data_type ()
  #3  0x0000000000639c65 in hist_entry__get_data_type ()
  #4  0x00000000006e57a8 in sort__type_collapse ()
  #5  0x00000000006ef0ee in hists__collapse_resort ()
  ⬢ [acme@toolbx perf-tools-next]$

The report related entries in 'perf test' pass:

⬢ [acme@toolbx perf-tools-next]$ perf test 17 27 30 31 85 88

  17: Match and link multiple hists                              : Ok
  27: Filter hist entries                                        : Ok
  30: Sort output of hist entries                                : Ok
  31: Cumulate child hist entries                                : Ok

85: Test that perf report includes file offsets and event type names in diagnostic messages. : Skip 88: Test that perf report handles truncated perf.data gracefully (no crash, no segfault — clean error exit).: Ok ⬢ [acme@toolbx perf-tools-next]$

---
 tools/perf/Documentation/perf-report.txt |  13 ++
 tools/perf/builtin-report.c              |  11 ++
 tools/perf/ui/Build                      |   1 +
 tools/perf/ui/progress.h                 |   2 +
 tools/perf/ui/stdio/progress.c           | 185 +++++++++++++++++++++++
 tools/perf/util/ordered-events.c         |  17 ++-
 tools/perf/util/session.c                |  12 +-
 7 files changed, 234 insertions(+), 7 deletions(-)
 create mode 100644 tools/perf/ui/stdio/progress.c
tools/perf/Documentation/perf-report.txt +13 -0
index fed6af128ff07e4c..2db4b069546130f0 100644
@@ -29,6 +29,19 @@ OPTIONS
 --quiet::
 	Do not show any warnings or messages.  (Suppress -v)
 
+--progress::
+	Show progress for each of the processing phases, printing the
+	percentage done and the current/total counts: the first phase
+	counts the bytes of the perf.data file processed so far, the
+	merge and sort phases count hist entries, e.g.:
+
+	Processing events... [ 42.3%] 4G / 10G
+	Merging related events... [  7.1%] 1024 / 14387
+	Sorting events for output... [ 98.2%] 14132 / 14387
+
+	It is a no-op when using the TUI or GTK browsers, that already
+	present progress information.
+
 -n::
 --show-nr-samples::
 	Show the number of samples for each symbol
tools/perf/builtin-report.c +11 -0
index 4d3383d1daae2ed9..fe59a429b9d6eef9 100644
@@ -87,6 +87,7 @@ struct report {
 	bool			use_gtk;
 #endif
 	bool			use_stdio;
+	bool			progress;
 	bool			show_full_info;
 	bool			show_threads;
 	bool			inverted_callchain;
@@ -1373,6 +1374,8 @@ int cmd_report(int argc, const char **argv)
 		    "Use the stdio interface"),
 	OPT_BOOLEAN(0, "weights", &symbol_conf.annotate_weight,
 			"Show or hide weight columns in annotation. Default show if non-zero."),
+	OPT_BOOLEAN(0, "progress", &report.progress,
+		    "Show progress while processing the perf.data file"),
 	OPT_BOOLEAN(0, "header", &report.header, "Show data header."),
 	OPT_BOOLEAN(0, "header-only", &report.header_only,
 		    "Show only data header."),
@@ -1765,6 +1768,14 @@ int cmd_report(int argc, const char **argv)
 	else
 		use_browser = 0;
 
+	/*
+	 * The TUI/GTK browsers already show progress, this is for the stdio
+	 * case, where we print the percentage of the perf.data file that was
+	 * processed so far, for each of the processing phases.
+	 */
+	if (report.progress && use_browser == 0)
+		stdio_progress__init();
+
 	if (report.data_type && use_browser == 1) {
 		symbol_conf.annotate_data_member = true;
 		symbol_conf.annotate_data_sample = true;
tools/perf/ui/Build +1 -0
index 6005f813c9e3990c..a7b1740d51c80f23 100644
@@ -4,6 +4,7 @@ perf-ui-y += progress.o
 perf-ui-y += util.o
 perf-ui-y += hist.o
 perf-ui-y += stdio/hist.o
+perf-ui-y += stdio/progress.o
 
 CFLAGS_setup.o += -DLIBDIR="BUILD_STR($(LIBDIR))"
 
tools/perf/ui/progress.h +2 -0
index 4f52c37b2f099a82..03f1a8bb260ba076 100644
@@ -23,6 +23,8 @@ void __ui_progress__init(struct ui_progress *p, u64 total,
 
 void ui_progress__update(struct ui_progress *p, u64 adv);
 
+void stdio_progress__init(void);
+
 struct ui_progress_ops {
 	void (*init)(struct ui_progress *p);
 	void (*update)(struct ui_progress *p);
tools/perf/ui/stdio/progress.c new file +185 -0

185 lines not on this page · raw message

tools/perf/util/ordered-events.c +13 -4
index a5857f9f5af2d3de..4063403d4978b45e 100644
@@ -237,14 +237,16 @@ static int do_flush(struct ordered_events *oe, bool show_progress)
 		ui_progress__init(&prog, oe->nr_events, "Processing time ordered events...");
 
 	list_for_each_entry_safe(iter, tmp, head, list) {
-		if (session_done())
-			return 0;
+		if (session_done()) {
+			ret = 0;
+			goto out_progress;
+		}
 
 		if (iter->timestamp > limit)
 			break;
 		ret = oe->deliver(oe, iter);
 		if (ret < 0)
-			return ret;
+			goto out_progress;
 
 		ordered_events__delete(oe, iter);
 		oe->last_flush = iter->timestamp;
@@ -258,10 +260,17 @@ static int do_flush(struct ordered_events *oe, bool show_progress)
 	else if (last_ts <= limit)
 		oe->last = list_entry(head->prev, struct ordered_event, list);
 
+	ret = 0;
+out_progress:
+	/*
+	 * Always pair ui_progress__init() with ui_progress__finish(), the
+	 * stdio progress backend tracks phases on a stack and an early
+	 * return that skipped the finish would leave the dead 'prog' on it.
+	 */
 	if (show_progress)
 		ui_progress__finish();
 
-	return 0;
+	return ret;
 }
 
 static int __ordered_events__flush(struct ordered_events *oe, enum oe_flush how,
tools/perf/util/session.c +9 -3
index 3237870a1a34b62c..85166042e899aada 100644
@@ -3146,8 +3146,12 @@ static int __perf_session__process_pipe_events(struct perf_session *session)
 	cur_size = sizeof(union perf_event);
 
 	buf = malloc(cur_size);
-	if (!buf)
-		return -errno;
+	if (!buf) {
+		err = -errno;
+		if (update_prog)
+			ui_progress__finish();
+		return err;
+	}
 	ordered_events__set_copy_on_queue(oe, true);
 more:
 	event = buf;
@@ -3646,8 +3650,10 @@ static int __perf_session__process_dir_events(struct perf_session *session)
 	}
 
 	rd = calloc(nr_readers, sizeof(struct reader));
-	if (!rd)
+	if (!rd) {
+		ui_progress__finish();
 		return -ENOMEM;
+	}
 
 	rd[0] = (struct reader) {
 		.fd		 = perf_data__fd(session->data),
Signature · 1 line
2.55.0
author maintainer
#11 lore raw

PATCH v5 6/9 perf scripts: Add perf-stuck, to tell where a running perf is stuck

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

A perf that takes forever is hard to tell apart from one that is stuck in a loop, and when it is stuck there is no way to know where without attaching gdb to it and looking around, which is what this does, from the outside, sampling /proc/<pid> at a fixed interval:

  ⬢ [acme@toolbx perf-tools-next]$ tools/perf/scripts/perf-stuck.sh -i 15 -n 4 -l ibs.log $(pgrep -x perf)
  watching 2297313 (perf report --progress -s type -i perf.data.ibs) every 15s
  09:39:34 state=R cpu=+0 (0.00s) rss=664295kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB  Merging related events... [ 39.0%] 197574 / 506686
  09:39:49 state=R cpu=+1496 (14.96s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=1 Merging related events... [ 39.0%] 197574 / 506686
  09:40:04 state=R cpu=+1497 (14.97s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=2 Merging related events... [ 39.0%] 197574 / 506686
  09:40:19 state=R cpu=+1497 (14.97s) rss=502250kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=0 Merging related events... [ 61.0%] 309026 / 506686

The CPU time used grows by a whole interval on every sample while the [stack] mapping, that would be moving down if this was recursion, stays put, so that one is spinning, and the last line of the progress log of 'perf report --progress' tells in which phase.

With -g it runs gdb when no progress is made for two consecutive samples, using the perf-stuck.gdb that sits next to it, which adds the perf-die-chain command used by the next patch, to print the DIE chain a DWARF type chasing loop is walking when the perf being watched is stuck in one of those.

It is a prototype: this wants to become a first class 'perf stuck' command, sampling a running process from inside perf, instead of this shell script poking at /proc and shelling out to gdb.

Example:

⬢ [acme@toolbx perf-tools-next]$ tools/perf/scripts/perf-stuck.sh -i 1 -n 3 -g $(pgrep -x sleep) watching 2342826 (sleep 45 ) every 1s
11:58:36 state=S cpu=+0 (0.00s) rss=492kB stack=7ffe21f89000-7ffe21faa000 size=132kB stuck=2 (no progress log) ... gdb output of 2342826 in /tmp/perf-stuck-gdb.Uz9i5p
⬢ [acme@toolbx perf-tools-next]$ tail -4 /tmp/perf-stuck-gdb.Uz9i5p
#3 0x0000555bc6ddc28f in main ()
not in a DWARF type chaser, try: bt
not in find_data_type()
[Inferior 1 (process 2342826) detached]
⬢ [acme@toolbx perf-tools-next]$

---
 tools/perf/scripts/perf-stuck.gdb | 104 ++++++++++++++++
 tools/perf/scripts/perf-stuck.sh  | 194 ++++++++++++++++++++++++++++++
 2 files changed, 298 insertions(+)
 create mode 100644 tools/perf/scripts/perf-stuck.gdb
 create mode 100755 tools/perf/scripts/perf-stuck.sh
tools/perf/scripts/perf-stuck.gdb new file +104 -0
new file mode 100644
index 0000000000000000..53e9019828b6ef62
@@ -0,0 +1,104 @@
+# SPDX-License-Identifier: GPL-2.0
+#
+# gdb commands for a perf that is stuck, used by perf-stuck.sh -g and usable
+# directly:
+#
+#   gdb -p $(pgrep -x perf) -batch -x perf-stuck.gdb -ex bt
+#
+# PROTOTYPE: part of the perf-stuck.sh stopgap, see the note at the start of
+# that script: this wants to move into a first class 'perf stuck' command,
+# which would print these DIE chains by itself, without gdb.
+#
+# The settings are the ones that keep a batch attach from stopping to ask
+# questions (debuginfod, pagination) and that make the output readable.
+#
+# The commands are for the DWARF type chasers in util/dwarf-aux.c, the
+# functions a data type profiling 'perf report -s type' spins in when a
+# debug info file has a type chain that got into a cycle:
+#
+#   perf-die-chain <function> <die variable> [iterations]
+#   perf-die-chain-all [iterations]
+#   perf-dso
+#
+# For each iteration of the chasing loop they print the DIE address, the
+# CU it came from, its offset in the debug file, its tag and its name: a
+# cycle shows up as the same handful of (addr, cu) pairs repeating, and a
+# CU that changes from one iteration to the next means the chase is
+# hopping between a debug file and its dwz common file.
+
+set pagination off
+set confirm off
+set debuginfod enabled off
+set print pretty on
+set height 0
+set width 0
+
+define perf-die-chain
+  if $argc < 2
+    printf "usage: perf-die-chain <function> <die variable> [iterations]\n"
+  else
+    frame function $arg0
+    if $argc == 3
+      set $perf_die_chain_n = $arg2
+    else
+      set $perf_die_chain_n = 10
+    end
+    set $perf_die_chain_head = $pc
+    set $perf_die_chain_i = 0
+    while $perf_die_chain_i < $perf_die_chain_n
+      printf "chain[%d] die=%p addr=%p cu=%p off=0x%lx tag=%d name=%s\n", $perf_die_chain_i, $arg1, $arg1->addr, $arg1->cu, ((Dwarf_Off) dwarf_dieoffset($arg1)), ((int) dwarf_tag($arg1)), ((char *) dwarf_diename($arg1))
+      until *$perf_die_chain_head
+      set $perf_die_chain_i = $perf_die_chain_i + 1
+    end
+  end
+end
+
+document perf-die-chain
+Print the DIE chain being walked by a DWARF type chasing loop.
+usage: perf-die-chain <function> <die variable> [iterations]
+  perf-die-chain die_get_pointer_type type_die
+  perf-die-chain __die_get_real_type vr_die
+  perf-die-chain die_get_real_type vr_die
+end
+
+define perf-die-chain-all
+  if $argc == 0
+    set $perf_die_chain_n = 10
+  else
+    set $perf_die_chain_n = $arg0
+  end
+  if $_any_caller_is("die_get_pointer_type", 20)
+    printf "stuck in die_get_pointer_type():\n"
+    perf-die-chain die_get_pointer_type type_die $perf_die_chain_n
+  else
+    if $_any_caller_is("__die_get_real_type", 20)
+      printf "stuck in __die_get_real_type():\n"
+      perf-die-chain __die_get_real_type vr_die $perf_die_chain_n
+    else
+      if $_any_caller_is("die_get_real_type", 20)
+        printf "stuck in die_get_real_type():\n"
+        perf-die-chain die_get_real_type vr_die $perf_die_chain_n
+      else
+        printf "not in a DWARF type chaser, try: bt\n"
+      end
+    end
+  end
+end
+
+document perf-die-chain-all
+Find which DWARF type chaser the process is in and print the DIE chain.
+usage: perf-die-chain-all [iterations]
+end
+
+define perf-dso
+  if $_any_caller_is("find_data_type", 20)
+    frame function find_data_type
+    printf "dso=%s ip=0x%lx sym=%s\n", dloc->ms->map->dso->name, dloc->ip, dloc->ms->sym->name
+  else
+    printf "not in find_data_type()\n"
+  end
+end
+
+document perf-dso
+Print the dso, ip and symbol of the data location being resolved.
+end
tools/perf/scripts/perf-stuck.sh new file +194 -0

194 lines not on this page · raw message

Signature · 1 line
2.55.0
author maintainer
#13 lore raw

PATCH v5 7/9 perf annotate-data: Resolve type DIEs in the debug file they came from

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

A 'perf report -s type' on a 783 MB AMD IBS data type profiling session hangs, burning all of a CPU and producing no output:

  ⬢ [acme@toolbx perf-tools-next]$ perf report --progress -s type -i perf.data.ibs > /dev/null
  Processing events... [100.0%] 746M / 746M
  Merging related events... [ 58.0%] 293828 / 506686
  Merging related events... [ 59.0%] 298894 / 506686
  Merging related events... [ 60.0%] 303960 / 506686
  Merging related events... [ 61.0%] 309026 / 506686
  [ ... nothing else, ever ... ]

It is a spin and not a slow path: with perf-stuck the CPU time used grows by a whole interval on every sample, while the [stack] mapping, which would be moving down if this was recursion, stays put:

  ⬢ [acme@toolbx perf-tools-next]$ tools/perf/scripts/perf-stuck.sh -i 15 -n 4 -l ibs.log $(pgrep -x perf)
  watching 2297313 (perf report --progress -s type -i perf.data.ibs) every 15s
  09:39:34 state=R cpu=+0 (0.00s) rss=664295kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB  Merging related events... [ 39.0%] 197574 / 506686
  09:39:49 state=R cpu=+1496 (14.96s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=1 Merging related events... [ 39.0%] 197574 / 506686
  09:40:04 state=R cpu=+1497 (14.97s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=2 Merging related events... [ 39.0%] 197574 / 506686
  09:40:19 state=R cpu=+1497 (14.97s) rss=502250kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=0 Merging related events... [ 61.0%] 309026 / 506686

Attaching gdb and stepping the loop, printing dwarf_dieoffset() and dwarf_tag() for the DIE being chased on each trip round it, shows the chase never moving, on a typedef that refers to itself:

  ⬢ [acme@toolbx perf-tools-next]$ gdb -p $(pgrep -x perf) -batch -x tools/perf/scripts/perf-stuck.gdb -ex 'perf-die-chain die_get_pointer_type type_die 8' -ex perf-dso -ex detach
  stuck in die_get_pointer_type():
  #3  0x0000000000772421 in die_get_pointer_type (type_die=0x7ffc3ae4b710, type_die@entry=0x7ffc3ae4b6f0, die_mem=die_mem@entry=0x7ffc3ae4b710) at util/dwarf-aux.c:327
  327			type_die = die_get_type(type_die, die_mem);
  chain[0] die=0x7ffc3ae4b710 addr=0x7ff6811591ef cu=0x44ef0878 off=0x1f tag=22 name=(null)
  chain[1] die=0x7ffc3ae4b710 addr=0x7ff6811591ef cu=0x44ef0878 off=0x1f tag=22 name=(null)
  [ ... the very same DIE, forever ... ]
  dso=/usr/lib64/libz.so.1.3.1.zlib-ng ip=0xe2e sym=build_tree

The DIE is at offset 0x1f of the debug info of libz.so.1, which is zlib-ng, and is one of the dwz compressed ones: the type DIEs shared by more than one CU live in the common file, where 0x1f is a perfectly good DW_TAG_base_type:

⬢ [acme@toolbx perf-tools-next]$ readelf --debug-dump=info /usr/lib/debug/.dwz/zlib-ng-2.3.3-3.fc44.x86_64 | sed -n '/Compilation Unit @ offset 0:/,/Compilation Unit @ offset 0x5f:/p'
Compilation Unit @ offset 0:
Length: 0x5b (32-bit)
Version: 5
Unit Type: DW_UT_partial (3)
<0><c>: Abbrev Number: 1 (DW_TAG_partial_unit)
[...]
<1><1f>: Abbrev Number: 62 (DW_TAG_base_type)

      <20>   DW_AT_byte_size   : 4
      <21>   DW_AT_encoding    : 5	(signed)
      <22>   DW_AT_name        : int

while in the main debug file that same offset is not a DIE at all, it is the start of another unit's header:

⬢ [acme@toolbx perf-tools-next]$ readelf --debug-dump=info /usr/lib/debug/usr/lib64/libz.so.1.3.1.zlib-ng-2.3.3-3.fc44.x86_64.debug | grep 'Compilation Unit @' | head -2
Compilation Unit @ offset 0:
Compilation Unit @ offset 0x1f:

die_collect_vars() saves the dwarf_dieoffset() of the type DIE, which is relative to the file that DIE lives in, and update_var_state() then hands that offset to dwarf_offdie() together with the main debug file, where dwarf_offdie() parses whatever is there as a DIE, in this case the one typedef that refers to itself, which is what makes the "follow the typedefs and qualifiers until a pointer or an array type" loop in die_get_pointer_type() spin.

So record, next to the offset, whether the type DIE was in the file the variable DIE came from or in the dwz common one, and resolve the offset in the file it came from, which is what the new die_get_type_die() does.

Which of the two that is does not have to be guessed from what is at the
offset: dwz encodes the references into its common file as DW_FORM_GNU_ref_alt, so elfutils resolves them into the alt Dwarf and the CU of the resulting type DIE belongs to that other file, so comparing the Dwarf each of the two CUs belongs to, die_same_file(), settles it exactly, for any number of hops from the variable DIE.

Then bound the chases themselves: no sane chain of typedefs and qualifiers is 32 DIEs long, no sane nesting for the struct and union members that __add_member_cb() follows recursively is 8 deep, and the same recursion bound covers the type names die_get_typename_from_type() builds by following pointers and arrays, so that a debug info file broken in some other way makes perf give up on a type, telling about it with pr_debug, visible with -v, instead of looking like it hung. The member nesting bound is reported with pr_debug_dtp (visible with -vvv or -D type-profile).

Cycles can't occur in member trees from valid DWARF, embedded members can't be recursive in C, so the nesting bound only ever bites legitimate depth and the member where the recursion was cut is marked 'truncated', which the JSON exporter added in a subsequent series reports to its consumers, so that they can tell a truncated tree from one that really ends there.

die_get_type_die() then has no fallback to the other file: resolving an alt file offset in the main file is the misparse above, so it resolves in the file the offset was recorded as belonging to, and gives up on the type when the offset does not resolve there. The tag the type DIE had is kept as a sanity check, a mismatch means the debug info changed under perf or is broken in yet another way, and the bounds on the type chasers remain the backstop: a debug info file broken in some other way still makes perf give up on a type with a pr_debug instead of hanging.

Testing:

Before, on the 783 MB AMD IBS session, killed after some 8 minutes stuck at 61%:

  ⬢ [acme@toolbx perf-tools-next]$ perf report --progress -s type -i perf.data.ibs > /dev/null
  Processing events... [100.0%] 746M / 746M
  Merging related events... [ 61.0%] 309026 / 506686
  ⬢ [acme@toolbx perf-tools-next]$

After, the whole session is processed, and the zlib-ng types, from the build_tree() hist entry that used to hang it, show up:

⬢ [acme@toolbx perf-tools-next]$ perf report --progress -s type -i perf.data.ibs > ibs.out Processing events... [100.0%] 746M / 746M
Merging related events... [100.0%] 506686 / 506686
Sorting events for output... [100.0%] 1428 / 1428
⬢ [acme@toolbx perf-tools-next]$ grep -E "deflate_state|inflate_state|internal_state" ibs.out

      0.00%  deflate_state
      0.00%  deflate_state*
      0.00%  struct inflate_state
      0.00%  struct inflate_state*
      0.00%  struct internal_state
  ⬢ [acme@toolbx perf-tools-next]$ perf test 17 27 30 31 85 88
  17: Match and link multiple hists                              : Ok
  27: Filter hist entries                                        : Ok
  30: Sort output of hist entries                                : Ok
  31: Cumulate child hist entries                                : Ok

85: Test that perf report includes file offsets and event type names in diagnostic messages. : Ok 88: Test that perf report handles truncated perf.data gracefully (no crash, no segfault — clean error exit).: Skip ⬢ [acme@toolbx perf-tools-next]$

Requiring elfutils 0.160 for this: dwarf_cu_getdwarf(), the function that tells which Dwarf a CU belongs to, first appeared in 0.160 ("libdw:
New functions dwarf_cu_getdwarf, dwarf_cu_die", elfutils NEWS), so the libdw feature test now probes for it, in tools/build/feature/test-libdw.c, and Makefile.config says 0.160 where it said 0.157. The probe takes the address instead of calling it, as that is all that is needed to make a 0.157-0.159 elfutils, from 2014 and without the symbol, disable dwarf support with the existing message rather than fail to link dwarf-aux.c.

Fixes 06b2ce75386df04b ("perf annotate-data: Maintain variable type info")
Fixes 55ee3d005d62279d ("perf annotate-data: Add a cache for global variable types")
Assisted-by LLM
---
 tools/build/feature/test-libdw.c |  15 ++-
 tools/perf/Makefile.config       |   2 +-
 tools/perf/util/annotate-data.c  |  59 ++++++++---
 tools/perf/util/annotate-data.h  |   3 +
 tools/perf/util/dwarf-aux.c      | 171 +++++++++++++++++++++++++++----
 tools/perf/util/dwarf-aux.h      |  33 ++++++
 6 files changed, 247 insertions(+), 36 deletions(-)
tools/build/feature/test-libdw.c +14 -1
index aabd63ca76b4d7e6..23e1ba6ff3466b9f 100644
@@ -49,8 +49,21 @@ int test_elfutils(void)
 	return 0;
 }
 
+/*
+ * elfutils 0.160 and later: used to tell which debug file a DIE lives in,
+ * the dwz alt file or the main one, see die_same_file() in
+ * tools/perf/util/dwarf-aux.c.  Only the symbol is needed, so take its
+ * address instead of calling it.
+ */
+int test_libdw_cu_getdwarf(void)
+{
+	void *sym = (void *)dwarf_cu_getdwarf;
+
+	return sym == NULL;
+}
+
 int main(void)
 {
 	return test_libdw() + test_libdw_unwind() + test_libdw_getlocations() +
-	       test_libdw_getcfi() + test_elfutils();
+	       test_libdw_getcfi() + test_libdw_cu_getdwarf() + test_elfutils();
 }
tools/perf/Makefile.config +1 -1
index 4d5993da9f94579f..fa78f50db60179f3 100644
@@ -470,7 +470,7 @@ else
   else
     ifneq ($(feature-libdw), 1)
       ifndef NO_LIBDW
-        $(warning No libdw.h found or old libdw.h found or elfutils is older than 0.157, disables dwarf support. Please install new elfutils-devel/libdw-dev)
+        $(warning No libdw.h found or old libdw.h found or elfutils is older than 0.160, disables dwarf support. Please install new elfutils-devel/libdw-dev)
         NO_LIBDW := 1
       endif
     endif # Dwarf support
tools/perf/util/annotate-data.c +47 -12

133 lines not on this page · raw message

tools/perf/util/annotate-data.h +3 -0
index c26130744260955f..d85866e83fbda69a 100644
@@ -57,6 +57,9 @@ struct annotated_member {
 	char *var_name;
 	int offset;
 	int size;
+	unsigned int depth;
+	/* Children not expanded because the nesting limit was reached */
+	bool truncated;
 };
 
 /**
tools/perf/util/dwarf-aux.c +149 -22

266 lines not on this page · raw message

tools/perf/util/dwarf-aux.h +33 -0
index 161f0bf980b6ee6a..149e0cbf63ccc82a 100644
@@ -152,6 +152,8 @@ int die_get_scopes(Dwarf_Die *cu_die, Dwarf_Addr pc, Dwarf_Die **scopes);
 struct die_var_type {
 	struct die_var_type *next;
 	u64 die_off;
+	int die_tag;
+	bool from_alt;	/* die_off is relative to the alt (dwz) file */
 	u64 addr;
 	u64 end;        /* end address of location range */
 	int reg;
@@ -183,6 +185,37 @@ Dwarf_Die *die_find_variable_by_addr(Dwarf_Die *sc_die, Dwarf_Addr addr,
 /* Save all variables and parameters in this scope */
 void die_collect_vars(Dwarf_Die *sc_die, struct die_var_type **var_types);
 
+/*
+ * Get the type DIE saved by die_collect_vars()/die_collect_global_vars().
+ *
+ * The offsets those save are the dwarf_dieoffset() of the type DIE, which is
+ * relative to the debug file that DIE lives in: the dwz common file, the alt
+ * file in libdw terms, for the types shared by more than one CU, the main
+ * file for the rest.  Resolving an alt file offset in the main file does not
+ * fail: dwarf_offdie() parses whatever is at that offset there, and an offset
+ * that is a CU header in the main file reads back as a typedef whose
+ * DW_AT_type refers to itself, which is what hung 'perf report -s type' on
+ * the dwz compressed debug info of zlib-ng (libz.so.1).
+ *
+ * So @from_alt, recorded when the offset was saved, says which of the two
+ * files to resolve it in.  It is not inferred from the DIE contents: dwz
+ * encodes references into the common file as DW_FORM_GNU_ref_alt, so
+ * elfutils resolves them into the alt Dwarf and the CU of the type DIE then
+ * belongs to that other file, which die_same_file() compares exactly.
+ *
+ * There is deliberately no fallback to the other file when the offset does
+ * not resolve: that fallback is the misparse above.
+ *
+ * @die_tag is then only a sanity check: the offset is of a DIE that had this
+ * tag when it was saved, so a mismatch means the debug info changed under us,
+ * or is broken, and giving up on the type is the right answer.
+ */
+Dwarf_Die *die_get_type_die(Dwarf *dbg, u64 die_off, int die_tag, bool from_alt,
+			    Dwarf_Die *die_mem);
+
+/* Whether two DIEs live in the same debug file */
+bool die_same_file(Dwarf_Die *die_a, Dwarf_Die *die_b);
+
 /* Save all global variables in this CU */
 void die_collect_global_vars(Dwarf_Die *cu_die, struct die_var_type **var_types);
 
Signature · 1 line
2.55.0
author maintainer
#15 lore raw

PATCH v5 8/9 perf mem record: Request PERF_SAMPLE_CPU by default

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

The data-type profiling per-sample stream keys cross-CPU contention on sample->cpu: without PERF_SAMPLE_CPU the cpu field is the (u32)-1 "no CPU info" sentinel, documented as such in perf_session__deliver_event(), and same-instance reads and writes from different cores are indistinguishable from same-CPU traffic, so pahole's false-sharing detector cannot tell them apart.

builtin-record.c already defines --sample-cpu and 'perf mem record' forwards unknown options to the record parser, so passing it explicitly works today; make it the default, next to the -d (addr) and -W (weight) the command already requests, documenting it in perf-mem(1).

The rec_argv array is sized as nine arguments per memory PMU plus the user arguments, not counting the arguments __cmd_record() adds itself, up to eight with all the optional flags. On PMUs with separate load and store events the four event arguments plus the fixed ones already filled the array to its last slot, so this new argument would write past it; reserve space for the fixed arguments explicitly.

---
 tools/perf/Documentation/perf-mem.txt |  4 ++++
 tools/perf/builtin-mem.c              | 18 ++++++++++++++++--
 2 files changed, 20 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-mem.txt +4 -0
index 4d164836d0943119..fe51c5e3333dc4a0 100644
@@ -14,6 +14,10 @@ DESCRIPTION
 -----------
 "perf mem record" runs a command and gathers memory operation data
 from it, into perf.data. Perf record options are accepted and are passed through.
+It also requests the address (-d), the weight (-W, where supported) and the
+CPU id (--sample-cpu) of every sampled access by default; the CPU id is what
+lets per-sample analysis tell reads and writes to the same data from
+different cores apart from same-CPU traffic.
 
 "perf mem report" displays the result. It invokes perf report with the
 right set of options to display a memory access profile. By default, loads
tools/perf/builtin-mem.c +16 -2
index 6101a26b3a781e69..6f38cda1a45ada16 100644
@@ -99,8 +99,13 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem,
 	argc = parse_options(argc, argv, options, record_usage,
 			     PARSE_OPT_KEEP_UNKNOWN);
 
-	/* Max number of arguments multiplied by number of PMUs that can support them. */
-	rec_argc = argc + 9 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1);
+	/*
+	 * Max number of arguments multiplied by number of PMUs that can
+	 * support them, plus the arguments added directly below, at most:
+	 * "record", "-W", "-d", "--sample-cpu", "--phys-data",
+	 * "--data-page-size", "--all-user" and "--all-kernel".
+	 */
+	rec_argc = argc + 8 + 9 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1);
 
 	if (mem->cpu_list)
 		rec_argc += 2;
@@ -135,6 +140,15 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem,
 
 	rec_argv[i++] = "-d";
 
+	/*
+	 * The data-type profiling per-sample stream keys cross-CPU
+	 * contention on sample->cpu (PERF_SAMPLE_CPU); without it the cpu
+	 * field is the (u32)-1 'no CPU info' sentinel and same-instance
+	 * reads and writes from different cores are indistinguishable
+	 * from same-CPU traffic.
+	 */
+	rec_argv[i++] = "--sample-cpu";
+
 	if (mem->phys_addr)
 		rec_argv[i++] = "--phys-data";
 
Signature · 1 line
2.55.0
author maintainer
#17 lore raw

PATCH v5 9/9 perf mem record: Use the IBS swfilt filter when available

In reply to Arnaldo Carvalho de Melo

From: Arnaldo Carvalho de Melo <[email protected]>

IBS events with exclude_{user,kernel} bits, as used in per-thread mode when kernel samples are not allowed, are rejected by the kernel on hardware without the privilege filter, so a per-thread 'perf mem record' fails:

  $ perf mem record -o /dev/null -- true
  Error:
  Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed.
  Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.

Kernel v6.14 added swfilt, a software privilege filter that makes those events usable per-thread, exposed as the 'swfilt' format term (d29e744c71673a71 "perf/x86: Relax privilege filter restriction on AMD IBS"):

  $ perf record -e ibs_op/ldlat=0,swfilt=1/ -- true

Give the ibs_op memory events a name variant with the swfilt term, used when the PMU exposes it as a format term, keeping the names that need system wide mode otherwise.

The variant is used even for records that end up without exclude bits, e.g. a plain system wide one: perf record adds those bits itself when the first open fails with EACCES on an unprivileged setup, after the event name has been built, so the term has to be in it for that retry:

  $ perf record -v -e ibs_op/ldlat=0/ -- true
  kernel.perf_event_paranoid=2, trying to fall back to excluding kernel and hypervisor  samples
  Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed.
  Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.

With no exclude bits the kernel discards nothing.

This makes the per-thread 'perf mem record' that the data type profiling shell test does work: the skip added by patch 1 is driven by that record failing, so it now runs on AMD kernels with swfilt, and keeps skipping on kernels without it. The 'Test data symbol' shell test matches the exact ibs_op event string, so its regex now accepts terms added after ldlat=150.

---
 tools/perf/arch/x86/util/mem-events.c      | 17 ++++++++++++++---
 tools/perf/tests/shell/test_data_symbol.sh |  2 +-
 tools/perf/util/mem-events.c               | 21 +++++++++++++++++----
 tools/perf/util/mem-events.h               |  2 ++
 4 files changed, 34 insertions(+), 8 deletions(-)
tools/perf/arch/x86/util/mem-events.c +14 -3
index b38f519020ff8c6f..034053dc762101dc 100644
@@ -7,7 +7,10 @@
 
 #define MEM_LOADS_AUX		0x8203
 
-#define E(t, n, s, l, a) { .tag = t, .name = n, .event_name = s, .ldlat = l, .aux_event = a }
+#define E_INIT(t, n, s, l, a, sf) {					\
+	.tag = t, .name = n, .event_name = s, .swfilt_name = sf,	\
+	.ldlat = l, .aux_event = a }
+#define E(t, n, s, l, a) E_INIT(t, n, s, l, a, NULL)
 
 struct perf_mem_event perf_mem_events_intel[PERF_MEM_EVENTS__MAX] = {
 	E("ldlat-loads",	"%s/mem-loads,ldlat=%u/P",	"mem-loads",	true,	0),
@@ -21,14 +24,22 @@ struct perf_mem_event perf_mem_events_intel_aux[PERF_MEM_EVENTS__MAX] = {
 	E(NULL,			NULL,				NULL,		false,	0),
 };
 
+/*
+ * IBS events with exclude_{user,kernel} bits set, as used by perf to
+ * record per-thread when kernel samples are not allowed, are rejected
+ * by the kernel on hardware without the privilege filter unless the
+ * swfilt software filter is used, so these events carry a variant of
+ * their names with the swfilt term, used by perf_pmu__mem_events_name()
+ * when the kernel exposes the term.
+ */
 struct perf_mem_event perf_mem_events_amd[PERF_MEM_EVENTS__MAX] = {
 	E(NULL,		NULL,		NULL,	false,	0),
 	E(NULL,		NULL,		NULL,	false,	0),
-	E("mem-ldst",	"%s//",		NULL,	false,	0),
+	E_INIT("mem-ldst",	"%s//",		NULL,	false,	0, "%s/swfilt=1/"),
 };
 
 struct perf_mem_event perf_mem_events_amd_ldlat[PERF_MEM_EVENTS__MAX] = {
 	E(NULL,		NULL,		NULL,	false,	0),
 	E(NULL,		NULL,		NULL,	false,	0),
-	E("mem-ldst",	"%s/ldlat=%u/",	NULL,	true,	0),
+	E_INIT("mem-ldst",	"%s/ldlat=%u/",	NULL,	true,	0, "%s/ldlat=%u,swfilt=1/"),
 };
tools/perf/tests/shell/test_data_symbol.sh +1 -1
index d61b5659a46d9a77..46362267bf28a5e9 100755
@@ -73,7 +73,7 @@ if (($is_amd >= 1)); then
 		fi
 
 		mem_events="$(perf mem record -v --ldlat=150 -e list 2>&1)"
-		if ! [[ "$mem_events" =~ ^mem-ldst.*ibs_op/ldlat=150/.*available ]]; then
+		if ! [[ "$mem_events" =~ ^mem-ldst.*ibs_op/ldlat=150[,/].*available ]]; then
 			echo "ERROR: --ldlat not honored?"
 			exit 1
 		fi
tools/perf/util/mem-events.c +17 -4
index 0b49fce251fcc184..8f74cd085e500231 100644
@@ -82,6 +82,7 @@ static const char *perf_pmu__mem_events_name(struct perf_pmu *pmu, int i,
 					     char *buf, size_t buf_size)
 {
 	struct perf_mem_event *e;
+	const char *name;
 
 	if (i >= PERF_MEM_EVENTS__MAX || !pmu)
 		return NULL;
@@ -90,24 +91,36 @@ static const char *perf_pmu__mem_events_name(struct perf_pmu *pmu, int i,
 	if (!e || !e->name)
 		return NULL;
 
+	/*
+	 * Use the swfilt variant of the name when the PMU exposes the term.
+	 * It is not conditional on the event already having exclude bits:
+	 * perf record adds those bits itself when the first open fails with
+	 * EACCES on an unprivileged setup, after this name has been built,
+	 * and that retry only succeeds with the term in the name.  With no
+	 * exclude bits the kernel doesn't discard anything.
+	 */
+	name = e->name;
+	if (e->swfilt_name && perf_pmu__has_format(pmu, "swfilt"))
+		name = e->swfilt_name;
+
 	if (i == PERF_MEM_EVENTS__LOAD || i == PERF_MEM_EVENTS__LOAD_STORE) {
 		if (e->ldlat) {
 			if (!e->aux_event) {
 				/* ARM and Most of Intel */
 				scnprintf(buf, buf_size,
-					  e->name, pmu->name,
+					  name, pmu->name,
 					  perf_mem_events__loads_ldlat);
 			} else {
 				/* Intel with mem-loads-aux event */
 				scnprintf(buf, buf_size,
-					  e->name, pmu->name, pmu->name,
+					  name, pmu->name, pmu->name,
 					  perf_mem_events__loads_ldlat);
 			}
 		} else {
 			if (!e->aux_event) {
 				/* AMD and POWER */
 				scnprintf(buf, buf_size,
-					  e->name, pmu->name);
+					  name, pmu->name);
 			} else {
 				return NULL;
 			}
@@ -117,7 +130,7 @@ static const char *perf_pmu__mem_events_name(struct perf_pmu *pmu, int i,
 
 	if (i == PERF_MEM_EVENTS__STORE) {
 		scnprintf(buf, buf_size,
-			  e->name, pmu->name);
+			  name, pmu->name);
 		return buf;
 	}
 
tools/perf/util/mem-events.h +2 -0
index 5b98076904b0b689..41f628fad10709b9 100644
@@ -11,6 +11,8 @@ struct perf_mem_event {
 	u32		aux_event;
 	const char	*tag;
 	const char	*name;
+	/* Name with the swfilt software privilege filter, when supported. */
+	const char	*swfilt_name;
 	const char	*event_name;
 };
 
Signature · 1 line
2.55.0
#2 lore raw

Re: [PATCH 1/9] perf test: Skip data_type_profiling when the PMU cannot record memory events

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 18 lines

From: Arnaldo Carvalho de Melo <[email protected]>

The test records with 'perf mem record' in per-thread mode and its only guard matches one specific message:

  perf mem record -o /dev/null -- true 2>&1 | \
        grep -q "failed: no PMU supports the memory events" && exit 2

A PMU that has memory events but refuses them per-thread falls through it, and AMD IBS does:

  $ perf mem record -o /dev/null -- true
  Error:
  Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed.
  Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.
[ ... ]
  Pipe annotate [Failed: perf record]
Quoted 3 lines

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
#4 lore raw

Re: [PATCH 2/9] perf debuginfo: Fetch debuginfo keyed by build ID using debuginfod

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 13 lines

From: Arnaldo Carvalho de Melo <[email protected]>

perf already uses debuginfod to fetch source files when annotating (via probe-finder.c) and 'perf probe' has open_from_debuginfod(), which queries debuginfo keyed by build ID when a module's debuginfo isn't found locally, largely the same thing this adds; eventually that one could be moved over to the new helper. For the analysis tools there was no way to obtain the debuginfo for a DSO in a profile when it isn't available locally under the name the DSO was opened with, for instance the vmlinux for the kernel a profile was recorded on when processing it on another machine, or after the kernel and its debuginfo package got upgraded in between.

Quoted 8 lines

Add debuginfo__find_build_id(), that uses the debuginfod client to locate a debuginfo file keyed by the build ID, checking its local [ ... ]
that make NO_LIBDW=1 on a system that has the debuginfod client keeps linking.

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=2
#6 lore raw

Re: [PATCH 3/9] perf symbol: Fall back to fetching the vmlinux by build ID

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 13 lines

From: Arnaldo Carvalho de Melo <[email protected]>

A profile recorded on a kernel that is no longer installed, be it because the machine was rebooted into a new kernel or because the profile is being processed on another machine, can't have its kernel symbols resolved: /proc/kallsyms matches the running kernel, not the one in the profile, and is refused when restricted, e.g. with kernel.perf_event_paranoid > 1, while the build-id cache may carry just a kallsyms copy with zeroed addresses. With no kernel symbols the kernel samples can't be annotated, so they all end up in the '(unknown)' data type and the kernel is missing from the data type profile JSON "dsos" entry.

Quoted 8 lines

As a last resort, when the kernel symbols can't be found locally and the user didn't specify a kallsyms file, fetch the vmlinux keyed by [ ... ]
struct tty_struct, struct qspinlock, etc, with the kernel DSO, keyed by its build ID, appearing in the data type profile JSON.

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=3
#8 lore raw

Re: [PATCH 4/9] perf annotate-data: Show the sample count in the data-type browser

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 10 lines

From: Arnaldo Carvalho de Melo <[email protected]>

The data-type browser has a samples view, selected with -n (or with annotate.show_nr_samples), in which browser__write_overhead() prints a local nr_samples variable that is initialized to zero and never updated, so every member is listed as having no samples while the period and percent columns for the same entry are filled in.

Print the histogram entry's own count instead.

Quoted 7 lines

This predates the load/store counter split, so fix it ahead of that
patch: the split then only has to adapt a line that is already correct, and this fix can be picked on its own.

Fixes d001c7a7f4736743 ("perf annotate-data: Add hist_entry__annotate_data_tui()")
Assisted-by LLM

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=4
#10 lore raw

Re: [PATCH 5/9] perf report: Add --progress option

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 19 lines

From: Arnaldo Carvalho de Melo <[email protected]>

Processing a large data type profiling session, e.g. an AMD IBS one, can take a long time and, when using the stdio output, 'perf report' gives no feedback about which phase it is in nor about how far along it is: the TUI has a progress bar for that, but in the stdio case the ui_progress updates, that are already there, are dropped on the floor.

Add a --progress option that installs a stdio backend for ui_progress, ui/stdio/progress.c, printing the phase title, the percentage done and the current/total counts to stderr:

Processing events... [ 42.3%] 317M / 746M
Merging related events... [ 61.0%] 309026 / 506686
Sorting events for output... [ 98.2%] 14132 / 14387
[ ... ]
88: Test that perf report handles truncated perf.data gracefully (no crash, no segfault — clean error exit).: Ok ⬢ [acme@toolbx perf-tools-next]$

Quoted 2 lines

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=5
#12 lore raw

Re: [PATCH 6/9] perf scripts: Add perf-stuck, to tell where a running perf is stuck

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 14 lines

From: Arnaldo Carvalho de Melo <[email protected]>

A perf that takes forever is hard to tell apart from one that is stuck in a loop, and when it is stuck there is no way to know where without attaching gdb to it and looking around, which is what this does, from the outside, sampling /proc/<pid> at a fixed interval:

  ⬢ [acme@toolbx perf-tools-next]$ tools/perf/scripts/perf-stuck.sh -i 15 -n 4 -l ibs.log $(pgrep -x perf)
  watching 2297313 (perf report --progress -s type -i perf.data.ibs) every 15s
  09:39:34 state=R cpu=+0 (0.00s) rss=664295kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB  Merging related events... [ 39.0%] 197574 / 506686
  09:39:49 state=R cpu=+1496 (14.96s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=1 Merging related events... [ 39.0%] 197574 / 506686
  09:40:04 state=R cpu=+1497 (14.97s) rss=665383kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=2 Merging related events... [ 39.0%] 197574 / 506686
  09:40:19 state=R cpu=+1497 (14.97s) rss=502250kB stack=7ffc3ae31000-7ffc3ae52000 size=132kB stuck=0 Merging related events... [ 61.0%] 309026 / 506686
Quoted 7 lines

The CPU time used grows by a whole interval on every sample while the [ ... ]
[Inferior 1 (process 2342826) detached]
⬢ [acme@toolbx perf-tools-next]$

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=6
#14 lore raw

Re: [PATCH 7/9] perf annotate-data: Resolve type DIEs in the debug file they came from

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:

Pre-existing issues:
- [High] Unbounded recursive calls in __die_find_member_cb when processing anonymous structures can lead to stack exhaustion.
- [High] Missing loop iteration bound in die_get_member_type allows an infinite loop on cyclic DWARF structure references. --

commit ea0e3dfe9a708661f7cdcfe58f9fbe047a513329
Author: Arnaldo Carvalho de Melo <[email protected]>

perf annotate-data: Resolve type DIEs in the debug file they came from

A 'perf report -s type' on a 783 MB AMD IBS data type profiling session hangs, burning all of a CPU and producing no output. This resolves the type DIEs in the debug file they came from to prevent hanging on self-referencing typedefs and qualifiers. It also adds depth and recursion bounds to type chasers.

Quoted 4 lines
tools/perf/util/annotate-data.c +0 -0
index 4e4c587640823..06d27868887bd 100644

[ ... ]

Quoted 5 lines
tools/perf/util/annotate-data.c +1 -0
@@ -221,6 +221,15 @@ static bool data_type_less(struct rb_node *node_a, const struct rb_node *node_b)
 	return strcmp(a->self.type_name, b->self.type_name) < 0;
 }
 
+/*

tools/perf/util/annotate-data.c static bool data_type_less(struct rb_node *node_a, const struct rb_node *node_b) · line 234

Quoted 10 lines
tools/perf/util/annotate-data.c +8 -0
+ * Members of struct/union members are added recursively, and the same DIE
+ * that is not what it looks like, the one that makes the type chasers in
+ * util/dwarf-aux.c spin, can make a member's type point back at one of its
+ * own ancestors, recursing until the stack is gone.  Nothing usable comes
+ * out of nesting members this deep anyway.
+ */
+#define MAX_MEMBER_DEPTH 8
+
 /* Recursively add new members for struct/union */
 static int __add_member_cb(Dwarf_Die *die, void *arg)

[Severity: High]
This is a pre-existing issue, but while we are adding MAX_MEMBER_DEPTH to prevent stack exhaustion here, what about __die_find_member_cb() in tools/perf/util/dwarf-aux.c?

If we process a corrupted DWARF file with an anonymous structure member that forms a cycle, couldn't this initiate an unbounded recursive search down the anonymous structure member chain and also exhaust the stack?

tools/perf/util/dwarf-aux.c:__die_find_member_cb() {
        ...
        else if (!dwarf_diename(die_mem)) {     /* Unnamed structure */
                Dwarf_Die type_die, tmp_die;
                if (die_get_type(die_mem, &type_die) &&
                    die_find_member(&type_die, name, &tmp_die))
                        return DIE_FIND_CB_END;
        }
        ...
}
Quoted 4 lines
tools/perf/util/dwarf-aux.c +0 -0
index d7160f87ac7d7..7acb431fd34a8 100644

[ ... ]

Quoted 13 lines
tools/perf/util/dwarf-aux.c +9 -0
@@ -266,16 +266,35 @@ Dwarf_Die *die_get_type(Dwarf_Die *vr_die, Dwarf_Die *die_mem)
 		return NULL;
 }
 
+/*
+ * The chases below cross typedefs and qualifiers to get to the type that
+ * is actually meant, and a DIE that is not what it looks like, e.g. one
+ * parsed at an offset that is not the start of a DIE in the file it was
+ * resolved in, can have a DW_AT_type that refers back to itself, which
+ * makes them spin forever: 'perf report -s type' did exactly that on the
+ * dwz compressed debug info of zlib-ng (libz.so.1), burning all of a CPU
+ * with no output while resolving a hist entry in build_tree().
+ *

tools/perf/util/dwarf-aux.c Dwarf_Die *die_get_type(Dwarf_Die *vr_die, Dwarf_Die *die_mem) · line 286

Quoted 10 lines
tools/perf/util/dwarf-aux.c +6 -1
+ * No sane chain is this long, so give up instead of hanging, telling about
+ * it so that the broken debug info can be looked at.
+ */
+#define MAX_TYPE_CHASE 32
+
 /* Get a type die, but skip qualifiers */
 Dwarf_Die *__die_get_real_type(Dwarf_Die *vr_die, Dwarf_Die *die_mem)
 {
-	int tag;
+	int tag, chase = 0;

[Severity: High]
This isn't a bug introduced by this patch, but since we are preventing infinite spinning on cyclic DWARF types here, is there a similar missing loop iteration bound in die_get_member_type() in tools/perf/util/dwarf-aux.c?

If a crafted DWARF file contains a cyclic struct/union definition (e.g., an anonymous embedded struct whose type points back to itself) located at offset 0, would this traverse downwards infinitely and hang the tool since offset and tag remain unchanged?

tools/perf/util/dwarf-aux.c:die_get_member_type() {
        ...
        while (tag == DW_TAG_structure_type || tag == DW_TAG_union_type) {
                member = die_find_child(&mb_type, __die_find_member_offset_cb,
                                        (void *)(long)offset, die_mem);
                if (member == NULL)
                        return NULL;

                if (die_get_real_type(member, &mb_type) == NULL)
                        return NULL;

                tag = dwarf_tag(&mb_type);
        ...
}
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=7
#16 lore raw

Re: [PATCH 8/9] perf mem record: Request PERF_SAMPLE_CPU by default

In reply to Arnaldo Carvalho de Melo
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Heap buffer overflow when constructing rec_argv in builtin-c2c.c for systems with a single memory PMU and multiple memory events. --

commit 6b306f5272233833a7987c9e447a8e3a91ccaf75
Author: Arnaldo Carvalho de Melo <[email protected]>

perf mem record: Request PERF_SAMPLE_CPU by default

This commit enables requesting PERF_SAMPLE_CPU by default for perf mem record. It allows the false-sharing detector to distinguish between reads and writes from different cores and same-CPU traffic by providing a valid cpu field. Additionally, it explicitly increases the rec_argv array allocation size to prevent buffer overflows when adding these extra arguments.

[ ... ]

Quoted 7 lines
tools/perf/builtin-mem.c +0 -0
index 6101a26b3a781..6f38cda1a45ad 100644
@@ -99,8 +99,13 @@ static int __cmd_record(int argc, const char **argv, struct perf_mem *mem,
 	argc = parse_options(argc, argv, options, record_usage,
 			     PARSE_OPT_KEEP_UNKNOWN);

tools/perf/builtin-mem.c static int __cmd_record(int argc, const char **argv, struct perf_mem *mem, · line 108

Quoted 10 lines
tools/perf/builtin-mem.c +7 -2
 
-	/* Max number of arguments multiplied by number of PMUs that can support them. */
-	rec_argc = argc + 9 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1);
+	/*
+	 * Max number of arguments multiplied by number of PMUs that can
+	 * support them, plus the arguments added directly below, at most:
+	 * "record", "-W", "-d", "--sample-cpu", "--phys-data",
+	 * "--data-page-size", "--all-user" and "--all-kernel".
+	 */
+	rec_argc = argc + 8 + 9 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1);

[Severity: High]
This isn't a bug introduced by this patch, but since this explicitly fixes an allocation under-sizing here, should a similar fix be applied to the copy-pasted code in builtin-c2c.c?

In tools/perf/builtin-c2c.c:perf_c2c__record():

        /* Max number of arguments multiplied by number of PMUs that can support them. */
        rec_argc = argc + 11 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1);

        rec_argv = calloc(rec_argc + 1, sizeof(char *));

        ...

        for (j = 0; j < argc; j++, i++)
                rec_argv[i] = argv[j];

If a user runs perf c2c record -e spe-load,spe-store,spe-ldst -u -k on a system with a single memory PMU, rec_argc scales as argc + 11 * 1. However, 5 fixed flags are added, followed by 6 event flags (2 for each of the 3 events), and then 2 optional flags (-u, -k).

This totals 13 flags added, which would write 2 pointers past the allocated bounds in the final loop. Since this also overwrites calloc's implicit NULL terminator, cmd_record will then read out-of-bounds.

Can this heap buffer overflow crash the perf tool?

[ ... ]

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=8
#18 lore raw

Re: [PATCH 9/9] perf mem record: Use the IBS swfilt filter when available

In reply to Arnaldo Carvalho de Melo
Sashiko has reviewed this patch and found no issues. It looks great!
Quoted 18 lines

From: Arnaldo Carvalho de Melo <[email protected]>

IBS events with exclude_{user,kernel} bits, as used in per-thread mode when kernel samples are not allowed, are rejected by the kernel on hardware without the privilege filter, so a per-thread 'perf mem record' fails:

  $ perf mem record -o /dev/null -- true
  Error:
  Failure to open event 'ibs_op/ldlat=0/u' on PMU 'ibs_op' which will be removed.
  Invalid event (ibs_op/ldlat=0/u) in per-thread mode, enable system wide with '-a'.

Kernel v6.14 added swfilt, a software privilege filter that makes those events usable per-thread, exposed as the 'swfilt' format term (d29e744c71673a71 "perf/x86: Relax privilege filter restriction on AMD [ ... ]
ldlat=150.

Quoted 3 lines

Sashiko has reviewed this patch and found no issues. It looks great!

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=9