replay: Show callsite location with --srcline by junyeong0619 · Pull Request #2047 · namhyung/uftrace

junyeong0619 · 2026-05-18T07:04:31Z

closes #2027

I've added the callsite trigger action proposed in the issue.
Setting -T <func>@callsite records the return address of the matched
call, and uftrace replay --srcline shows it after the existing source
line, e.g. a() { /* file.c:11 from file.c:33 */.

Changes:

Added TRIGGER_FL_CALLSITE / EVENT_ID_CALLSITE and the callsite
trigger action parser.
libmcount emits the event (8-byte return address payload) right
before the matching ENTRY when the trigger fires.
Added task_find_exact_loc_addr() that uses libdw's
dwarf_addrdie() + dwarf_getsrc_die() for PC-precise line lookup
(the existing task_find_loc_addr() only knows the symbol's entry
line). load_debug_info() now keeps libdw open at replay time so
this lookup works there.
replay intercepts the CALLSITE event, stashes the address on the
task, and appends from file:line to the source comment on the
next ENTRY.
Added t305 covering both the live and --srcline paths.
Documented the new action in doc/uftrace-record.md with an example
and a short note about line precision under optimization.

Testing on Rocky Linux 9.7 (GCC 11.5, x86_64):

DURATION     TID      FUNCTION [SOURCE]
   0.510 us [3615820] | __monstartup();
   0.187 us [3615820] | __cxa_atexit();
            [3615820] | main() { /* s-abc.c:26 */
            [3615820] |   a() { /* s-abc.c:11 from s-abc.c:33 */
            [3615820] |     b() { /* s-abc.c:16 from s-abc.c:13 */
            [3615820] |       c() { /* s-abc.c:21 from s-abc.c:18 */
   0.250 us [3615820] |         getpid();
   1.449 us [3615820] |       } /* c */
   1.744 us [3615820] |     } /* b */
   2.317 us [3615820] |   } /* a */
   2.604 us [3615820] | } /* main */

One thing I'd like your feedback on:
save_callsite_event() writes the 8-byte payload with a direct
*(uint64_t *)event->data = ... rather than the existing
mcount_memcpy4(). Using the memcpy helper here recorded zero on my
setup. My guess is strict aliasing — mcount_memcpy4 reinterprets the
source as unsigned int *, and the 8-byte return address write through
that path gets optimized away. The direct write goes through the
uint8_t event->data[] member, which is well-defined. I'm happy to add
an mcount_memcpy8 helper (or update mcount_memcpy4 to use byte-wise
loads) if you prefer either of those.

Also a precision note (also in the doc): the resolved call site line
depends on the dwarf line program. -O0 / -Og builds report the
exact line; with -O1 or higher the call instruction can be folded
with surrounding code so the reported line may shift by 1–2 lines.
-finstrument-functions builds are stable across optimization levels.
t305 forces -O0 in cflags for this reason.

Please let me know if you'd like changes to the approach or naming.
Thanks!

namhyung · 2026-05-19T05:06:35Z

Thanks for working on this! A few comments before looking at the code.

I feel like requiring --srcline looks like unnecessary. It'd be nice if it can always show callsite info once recorded. Maybe we can add --no-callsite option to disable it at replay time. Then the output becomes compact and can be used with --srcline if needed.

For memcpy, there's mcount_memcpy1() for byte-wise copy. I'm not sure if it's a strict alias issue but unaligned access on ARM may cause problems.

Regarding the precision of the location, I think it's out of scope. We can just show whatever compiler generated for the code and don't care about the value itself (unless it's definitely wrong). Please don't change the test code or settings just to pass the test, it'd be better to leave it failing so that we can be aware of the issue - or we can check whether it generates any callsite info.

junyeong0619 · 2026-05-19T05:27:26Z

Thanks for the review!

Decouple callsite from --srcline so that running uftrace record -T 'a@callsite' followed by a plain uftrace replay shows the callsite info on its own, as an independent trigger output.
Switch the payload write to mcount_memcpy1().
Keep the cflags untouched and adjust the test instead.

Will push the updates along these lines.

junyeong0619 · 2026-05-19T09:23:02Z

Pushed the revisions. Now we can use for 4 versions.

Usage:

$ uftrace record -T 'a@callsite' tests/t-abc 

$ uftrace replay
# DURATION     TID      FUNCTION
   0.714 us [ 189072] | __monstartup();
  12.921 us [ 189072] | __cxa_atexit();
            [ 189072] | main() {
            [ 189072] |   a() { /* from s-abc.c:33 */
            [ 189072] |     b() {
            [ 189072] |       c() {
   0.390 us [ 189072] |         getpid();
   0.989 us [ 189072] |       } /* c */
   1.357 us [ 189072] |     } /* b */
   1.905 us [ 189072] |   } /* a */
   2.186 us [ 189072] | } /* main */

$ uftrace replay --srcline
# DURATION     TID      FUNCTION [SOURCE]
   0.714 us [ 189072] | __monstartup();
  12.921 us [ 189072] | __cxa_atexit();
            [ 189072] | main() { /* s-abc.c:26 */
            [ 189072] |   a() { /* s-abc.c:11 from s-abc.c:33 */
            [ 189072] |     b() { /* s-abc.c:16 */
            [ 189072] |       c() { /* s-abc.c:21 */
   0.390 us [ 189072] |         getpid();
   0.989 us [ 189072] |       } /* c */
   1.357 us [ 189072] |     } /* b */
   1.905 us [ 189072] |   } /* a */
   2.186 us [ 189072] | } /* main */

$ uftrace replay --no-callsite
# DURATION     TID      FUNCTION
   0.714 us [ 189072] | __monstartup();
  12.921 us [ 189072] | __cxa_atexit();
            [ 189072] | main() {
            [ 189072] |   a() {
            [ 189072] |     b() {
            [ 189072] |       c() {
   0.390 us [ 189072] |         getpid();
   0.989 us [ 189072] |       } /* c */
   1.357 us [ 189072] |     } /* b */
   1.905 us [ 189072] |   } /* a */
   2.186 us [ 189072] | } /* main */

$ uftrace replay --srcline --no-callsite
# DURATION     TID      FUNCTION [SOURCE]
   0.714 us [ 189072] | __monstartup();
  12.921 us [ 189072] | __cxa_atexit();
            [ 189072] | main() { /* s-abc.c:26 */
            [ 189072] |   a() { /* s-abc.c:11 */
            [ 189072] |     b() { /* s-abc.c:16 */
            [ 189072] |       c() { /* s-abc.c:21 */
   0.390 us [ 189072] |         getpid();
   0.989 us [ 189072] |       } /* c */
   1.357 us [ 189072] |     } /* b */
   1.905 us [ 189072] |   } /* a */
   2.186 us [ 189072] | } /* main */

Also Treated the write with mcount_memcpy1().

junyeong0619 · 2026-05-19T09:33:24Z

Rewrite the commit message for the changes.

namhyung

I have some concerns if it would

work well with other events and arguments. (I'm a bit worried about the timestamp management now)
print "function+offset" when DWARF is not available.

Also please pick a single prefix in the commit subject line. More general and/or important one would be preferred. Maybe you want to split the commit otherwise.

junyeong0619 · 2026-05-20T13:04:25Z

Here is full dump and replay output.

root@server:~/uftrace# uftrace record -W var:mydata -T 'bar@callsite' -T 'foo@callsite' t-watch
root@server:~/uftrace# uftrace dump | head -50
uftrace file header: magic         = 4674726163652100
uftrace file header: version       = 4
uftrace file header: header size   = 40
uftrace file header: endian        = 1 (little)
uftrace file header: class         = 2 (64 bit)
uftrace file header: features      = 0x1763 (PLTHOOK | TASK_SESSION | SYM_REL_ADDR | MAX_STACK | PERF_EVENT | AUTO_ARGS | DEBUG_INFO | SYM_SIZE)
uftrace file header: info          = 0xfbff (EXE_NAME | EXE_BUILD_ID | EXIT_STATUS | CMDLINE | CPUINFO | MEMINFO | OSINFO | TASKINFO | USAGEINFO | LOADINFO | RECORD_DATE | PATTERN_TYPE | VERSION | UTC_OFFSET | SYSCTL)

reading 1355229.dat
646618.687034679  1355229: [entry] __monstartup(62efea147090) depth: 0
646618.687036364  1355229: [exit ] __monstartup(62efea147090) depth: 0
646618.687037419  1355229: [entry] __cxa_atexit(62efea1470a0) depth: 0
646618.687037792  1355229: [exit ] __cxa_atexit(62efea1470a0) depth: 0
646618.687038474  1355229: [entry] main(62efea147269) depth: 0
646618.687038640  1355229: [event] trigger:callsite(186ad) depth: 0
646618.687038640  1355229: [data ] length = 8
  trigger:callsite: callsite=0x62efea14729d
646618.687038641  1355229: [entry] foo(62efea147233) depth: 1
646618.687038895  1355229: [event] trigger:callsite(186ad) depth: 0
646618.687038895  1355229: [data ] length = 8
  trigger:callsite: callsite=0x62efea14724a
646618.687038895  1355229: [event] watch:var(186ac) depth: 0
646618.687038895  1355229: [data ] length = 12
  watch:var: mydata=1
646618.687038896  1355229: [entry] bar(62efea14720b) depth: 2
646618.687040199  1355229: [exit ] bar(62efea14720b) depth: 2
646618.687041376  1355229: [event] watch:var(186ac) depth: 0
646618.687041376  1355229: [data ] length = 12
  watch:var: mydata=2
646618.687041377  1355229: [exit ] foo(62efea147233) depth: 1
646618.687041612  1355229: [exit ] main(62efea147269) depth: 0
reading perf-cpu4.dat
646618.675870741  1355229: [event] linux:task-name(200006)
646618.685026396  1355229: [event] linux:sched-out (pre-empted)(200007)
646618.685077278  1355229: [event] linux:sched-in(200001)
646618.687172426  1355229: [event] linux:task-exit(200005)

root@server:~/uftrace# uftrace replay
# DURATION     TID      FUNCTION
   1.685 us [1355229] | __monstartup();
   0.373 us [1355229] | __cxa_atexit();
            [1355229] | main() {
            [1355229] |   foo() { /* from tests/s-watch-global.c:26 */
            [1355229] |     /* watch:var (mydata=1) */
   1.303 us [1355229] |     bar(); /* from tests/s-watch-global.c:15 */
            [1355229] |     /* watch:var (mydata=2) */
   2.736 us [1355229] |   } /* foo */
   3.138 us [1355229] | } /* main */

junyeong0619 · 2026-05-20T13:06:12Z

Followed up with a deliberate-collision case using watch:var on a global variable changed inside the traced functions

(tests/s-watch-global.c):

  $ uftrace record -W var:mydata -T 'foo@callsite' -T 'bar@callsite' t-watch
  $ uftrace dump | head -40

  646618.687038895  [event] trigger:callsite        # entry-1 for bar
  646618.687038895  [event] watch:var (mydata=1)    # same nanosecond
  646618.687038896  [entry] bar

Two events do share a timestamp here. The ordering is still deterministic: mtdp->event[] is push-ordered, and
mcount_entry_filter_record() calls save_callsite_event() before save_watchpoint(), so the callsite is written to the trace first and replay processes them in the same order.

The resulting replay output keeps both visible without any mixing:

  $ uftrace replay
              [1355229] | main() {
              [1355229] |   foo() { /* from tests/s-watch-global.c:26 /
              [1355229] |     / watch:var (mydata=1) /
     1.303 us [1355229] |     bar(); / from tests/s-watch-global.c:15 /
              [1355229] |     / watch:var (mydata=2) /
     2.736 us [1355229] |   } / foo /
     3.138 us [1355229] | } / main */

junyeong0619 · 2026-05-20T13:08:49Z

Also treated the case that Dwarf is not available.

root@server:~/uftrace# uftrace record -T 'a@callsite' tests/t-abc-nodbg (t-abc with no -g)
root@server:~/uftrace# uftrace replay
# DURATION     TID      FUNCTION
   0.632 us [1360511] | __monstartup();
   0.282 us [1360511] | __cxa_atexit();
            [1360511] | main() {
            [1360511] |   a() { /* from main+0x41 */
            [1360511] |     b() {
            [1360511] |       c() {
   0.265 us [1360511] |         getpid();
   0.801 us [1360511] |       } /* c */
   1.066 us [1360511] |     } /* b */
   1.477 us [1360511] |   } /* a */
   1.747 us [1360511] | } /* main */

namhyung · 2026-05-22T07:20:43Z

Thanks, I've check it with other options and it has some side effects.

$ uftrace -a -F main ~/tmp/progs/abc
# DURATION     TID      FUNCTION
            [  10623] | main(1, 0x7ffd623464f8) {
            [  10623] |   a() {
            [  10623] |     b() {
            [  10623] |       c() {
   0.427 us [  10623] |         getpid() = 10623;
   2.495 us [  10623] |       } = 10623; /* c */
   2.913 us [  10623] |     } = 10624; /* b */
   3.169 us [  10623] |   } = 10623; /* a */
   3.950 us [  10623] | } = 0; /* main */

$ uftrace -T a@read=pmu-cycle -F main ~/tmp/progs/abc
# DURATION     TID      FUNCTION
            [  10657] | main() {
            [  10657] |   a() {
            [  10657] |     /* read:pmu-cycle (cycles=604 instructions=48) */
            [  10657] |     b() {
            [  10657] |       c() {
   0.582 us [  10657] |         getpid();
   1.334 us [  10657] |       } /* c */
   1.579 us [  10657] |     } /* b */
            [  10657] |     /* diff:pmu-cycle (cycles=+5857 instructions=+4554 IPC=0.78) */
  24.601 us [  10657] |   } /* a */
  30.448 us [  10657] | } /* main */

And if they are used together, it doesn't work well.

$ uftrace -a -T a@read=pmu-cycle,callsite -F main ~/tmp/progs/abc
# DURATION     TID      FUNCTION
            [  10690] | main(1, 0x7ffd169dab78) {
            [  10690] |   a() { /* from tests/s-abc.c:33 */
            [  10690] |     /* read:pmu-cycle (cycles=591 instructions=48) */
            [  10690] |     b() {
            [  10690] |       c() {
   0.446 us [  10690] |         getpid() = 10690;
   2.304 us [  10690] |       } = 10690; /* c */
   2.678 us [  10690] |     } = 10691; /* b */
  23.815 us [  10690] |   } = 10690; /* a */
  24.946 us [  10690] | } = 0; /* main */

So it lost the diff:pmu-cycle event. Removing the argument (-a) gives me the correct result.

$ uftrace -T a@read=pmu-cycle,callsite -F main ~/tmp/progs/abc
# DURATION     TID      FUNCTION
            [  10710] | main() {
            [  10710] |   a() { /* from tests/s-abc.c:33 */
            [  10710] |     /* read:pmu-cycle (cycles=582 instructions=48) */
            [  10710] |     b() {
            [  10710] |       c() {
   0.468 us [  10710] |         getpid();
   1.228 us [  10710] |       } /* c */
   1.496 us [  10710] |     } /* b */
            [  10710] |     /* diff:pmu-cycle (cycles=+6834 instructions=+4784 IPC=0.70) */
  25.378 us [  10710] |   } /* a */
  30.669 us [  10710] | } /* main */

junyeong0619 · 2026-05-22T08:32:21Z

Thanks for review and testing! The lost diff event turned out to be a pre-existing bug from combining -a/-A/-R with @READ, not from the callsite trigger. I sent a separate PR with the fix: #2048.

namhyung

Thanks for the update!

namhyung · 2026-05-23T06:17:26Z

 			xasprintf(&evt_name, "watch:var");
 			break;
+		case EVENT_ID_CALLSITE:
+			xasprintf(&evt_name, "trigger:callsite");


Let's show it as just "callsite".

Changed it to callsite.

namhyung · 2026-05-23T06:42:46Z

+		 * @callsite trigger), unless the user opted out
+		 */
+		needs_dbg = opts->srcline || (opts->loc_filter != NULL);
+		if (!opts->no_callsite && (handle->hdr.feat_mask & DEBUG_INFO))


The location filter check is new. Also I think checking DEBUG_INFO is not enough and you need to add a new CALLSITE feature bit. Loading dwarf file should depend on the CALLSITE flag.

Also opts->no_callsite doesn't exist at this point which means it'd break the build and bisect.

Added a CALLSITE feature bit (dwarf auto-load now depends on it), reverted the accidental loc_filter check in read_task_file(), and verified the entire series bisects cleanly.

namhyung · 2026-05-23T06:56:07Z

+ * match the trailing path against existing entries so callsite filenames
+ * stay consistent with the --srcline output
+ */
+static struct uftrace_dbg_file *find_or_get_debug_file(struct uftrace_dbg_info *dinfo,


Maybe rename to get_debug_file_suffix()?

Renamed it to get_debug_file_suffix.

namhyung · 2026-05-23T06:57:12Z


+/*
+ * PC-precise variant of task_find_loc_addr() — returns the source line
+ * at the instruction, not the enclosing function's entry


It returns a boolean. Also you need to mention that it uses DWARF directly.

Updated the comment to mention DWARF usage and the bool return.

namhyung · 2026-05-23T07:00:39Z

+			if (task_find_exact_loc_addr(sessions, task, rstack->time,
+						     task->callsite_ip, &csloc)) {
+				snprintf(cs_buf, sizeof(cs_buf), "%s:%d", csloc.file->name,
+					 csloc.line);


You need to make sure it has NUL character at the end.

Added explicit cs_buf[sizeof(cs_buf) - 1] = '\0' after the snprintf (and same in the function+offset helper).

namhyung · 2026-05-23T07:04:01Z

+							snprintf(cs_buf, sizeof(cs_buf), "%s+%#lx",
+								 cs_sym->name,
+								 (unsigned long)(mod_addr -
+										 cs_sym->addr));


Doesn't look pretty. Can we reduce the indention level or factor it out as a function?

Extracted into a format_callsite_symbol() helper.

namhyung · 2026-05-23T07:08:38Z

 	bool list_event;
 	bool event_skip_out;
 	bool no_event;
+	bool no_callsite;


Personally, I'd like to keep the options as positive even if the option name itself is negative. For example, I think it's natural to check if (opts->callsite) rather than if (!opts->no_callsite).

I don't plan to change the existing ones, but I want new ones in positive form.

Treated it as positive.

namhyung · 2026-05-23T07:12:22Z

+        return TestBase.build(self, name, cflags, ldflags)
+
+    def setup(self):
+        self.option = '-T a@callsite -T b@callsite -T c@callsite'


Couldn't it be -T '^.$@callsite'?

Done. Tests still pass.

namhyung · 2026-05-23T07:14:35Z

+            # strip directory prefix from filenames in comments
+            rest = re.sub(r'/[\w./-]+/(\S+\.c)', r'\1', rest)
+            # normalize source line numbers
+            rest = re.sub(r'(\.c):\d+', r'\1:N', rest)


Maybe you can just combine two lines as re.sub(r'/[\w./-]+/(\S+\.c):\d+', r'\1:N', rest).

namhyung

Thanks for the update, I think it's almost there. Please make sure to rebase your changes on to the latest commit whenever you push them.

namhyung · 2026-05-24T17:45:42Z

+The call site display can be suppressed with `--no-callsite` at replay
+time.  The reported line itself comes from the compiler's dwarf line
+program and is shown as-is.
+


Please add this to uftrace live man page too.

namhyung · 2026-05-24T17:55:07Z

 		    read_task_txt_file(sessions, opts->dirname, opts->with_syms ?: opts->dirname,
 				       true, sym_rel,
-				       opts->srcline | (opts->loc_filter != NULL)) < 0) {
+				       opts->srcline | (opts->loc_filter != NULL) | callsite_dbg) <


I think it's better to split callsite and srcline/loc-filter as the latter only requires the dbg file. You may introduce a local variable for srcline/loc-filter like needs_srcline. That means you need to pass it to load_debug_info() through create_session().

For this to reach load_debug_info(), create_session() would need to take both flags:

create_session(sess, &smsg, dirname, symdir, exename, sym_rel_addr, needs_symtab, needs_srcline, needs_callsite_dbg);

Does that look right?

namhyung · 2026-05-24T17:56:24Z

+		 * program at replay time.  finish_debug_info() releases it.
+		 */
+		if (needs_srcline && dinfo->dw == NULL)
+			setup_dwarf_info(map->libname, dinfo, map->start, true);


I think this part should belong to the next commit. Also I don't think it's using callsite info precisely.

To make this precise, would you prefer introducing a separate flag like needs_callsite_dbg?

For example:

if (needs_callsite_dbg && dinfo->dw == NULL) setup_dwarf_info(map->libname, dinfo, map->start, true);

Yes, something like needs_callsite.

Thanks for the quick review. I will change the check line in that way.

junyeong0619 · 2026-05-24T19:28:09Z

Rebased onto master. If there is something wrong let me know. Thanks.

namhyung · 2026-05-24T23:27:00Z

-		    0 &&
-	    read_task_file(&link, fha->opts->dirname, false, false, false) < 0)
+	if (read_task_txt_file(&link, fha->opts->dirname, fha->opts->dirname, false, false, false,
+			       false) < 0 &&


Looks like a whitespace error.

Whitespace is up to clang-format, so I shortened the line by introducing
char *dir = fha->opts->dirname;. Under 100 columns, clang-format leaves it alone.

namhyung · 2026-05-24T23:31:54Z

 		    read_task_txt_file(sessions, opts->dirname, opts->with_syms ?: opts->dirname,
-				       true, sym_rel,
-				       opts->srcline | (opts->loc_filter != NULL)) < 0) {
+				       true, sym_rel, opts->srcline | (opts->loc_filter != NULL),


Similarly, let's introduce needs_srcline variable.

Done — introduced needs_srcline = opts->srcline | (opts->loc_filter != NULL).

namhyung · 2026-05-24T23:34:19Z

 		if (!debug_info_has_location(dinfo) && !debug_info_has_argspec(dinfo)) {
 			load_debug_file(dinfo, stab, sinfo->symdir, map->libname, map->build_id,
-					needs_srcline);
+					needs_srcline || needs_callsite);


namhyung · 2026-05-24T23:34:56Z

 	if (!debug_info_has_location(dinfo) && !debug_info_has_argspec(dinfo)) {
 		load_debug_file(dinfo, &mod->symtab, dirname, mod->name, mod->build_id,
-				needs_srcline);
+				needs_srcline || needs_callsite);


I don't think this is necessary as callsite doesn't use the debug (.dbg) file.

Right, callsite doesn't read the .dbg file. Dropped needs_callsite from load_debug_file() in both load_debug_info() and load_module_debug_info().

Add the `callsite` trigger action which makes uftrace remember the return address of a matched function call. At record time it just sets a flag (TRIGGER_FL_CALLSITE); the actual address is captured later by libmcount. A new event id (EVENT_ID_CALLSITE) is reserved to carry the return address inside the trace data. This commit only wires up the option plumbing; the recording side and the replay-side display are added in subsequent commits. Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

@callsite

When the @callsite trigger fires at function entry, libmcount emits an EVENT_ID_CALLSITE record whose payload is the caller's return address (rstack->parent_ip). The event is timestamped one nanosecond before the entry record so it always precedes the matching ENTRY in the trace. The @callsite trigger also implicitly turns on debug info saving at record time and a new CALLSITE feature bit is set on the file header so the replay side can later resolve the recorded address to a source line without requiring --srcline. The replay-side reader decodes the 8-byte payload into task->args so existing event consumers (dump, replay, ...) can find it. event.c gains "callsite" naming and "callsite=0xADDR" formatting for the raw dump output. Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

task_find_loc_addr() returns the symbol's entry line stored in dinfo->locs[], which is fine for a function definition but loses the precision needed for a call site. Add task_find_exact_loc_addr() which goes through libdw directly: dwarf_addrdie() + dwarf_getsrc_die() against the function's module-relative address. This drills into the dwarf line program so a PC anywhere inside a function resolves to the matching source line. To make this work at replay time, load_debug_info() now also opens the libdw handle per module (kept until finish_debug_info()). The shared session/map lookup is factored into task_find_addr_module() used by both helpers. Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

When the trace contains an EVENT_ID_CALLSITE record, intercept it in the UFTRACE_EVENT branch, stash the return address on the task, and suppress the default event-as-comment output. The next ENTRY consumes that address: replay resolves it to file:line via task_find_exact_loc_addr() and appends it to the source comment. The call site is shown by default and can be turned off with the new --no-callsite option. When --srcline is also given, both the callee definition line and the call site are shown. When dwarf info is unavailable (e.g. when the binary is built without -g), replay falls back to a "function+offset" form so the call site is still informative. The fallback path is factored into a small helper format_callsite_symbol() to keep the caller flat. open_data_file() also loads debug info automatically when the trace carries the CALLSITE feature bit (unless --no-callsite is given), so the call site is resolved without having to pass --srcline. Simple usage: $ uftrace record -T 'a@callsite' t-abc $ uftrace replay # DURATION TID FUNCTION [3615820] | main() { [3615820] | a() { /* from s-abc.c:33 */ [3615820] | b() { [3615820] | c() { ... $ uftrace replay --srcline # DURATION TID FUNCTION [SOURCE] [3615820] | main() { /* s-abc.c:26 */ [3615820] | a() { /* s-abc.c:11 from s-abc.c:33 */ ... $ uftrace replay --no-callsite # DURATION TID FUNCTION [3615820] | main() { [3615820] | a() { ... $ uftrace replay # binary built without -g [3615820] | a() { /* from main+0x41 */ ... When the call site cannot be resolved at all (e.g. coming from libc) the suffix is silently dropped. closes namhyung#2027 Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

@callsite

t305 records s-abc with -T '^.$@callsite' and verifies that the replay output contains a "from <file>:<line>" suffix for each function. The expected output normalizes line numbers and strips directory prefixes (in sort()) so the test stays stable across optimization levels and build environments — the exact line reported by the compiler's dwarf line program is not asserted, only the presence of the callsite info itself. The default --no-event option is dropped via a runcmd() override so the EVENT_ID_CALLSITE records reach replay (same pattern as the existing read-trigger tests). The test is skipped when dwarf info is unavailable. Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

Add `callsite` to the trigger action BNF in uftrace-record.md and uftrace-live.md and describe the new behaviour with examples showing the "from file:N" suffix in the default replay output and the combined "file:N from file:N" output under --srcline. uftrace-replay.md gains a new entry for --no-callsite that explains how to suppress the call site display, and clarifies that --srcline only controls the callee definition line. Signed-off-by: Jun Yeong Kim <junyeonggim5@gmail.com>

namhyung

LGTM! I'm curious when you sleep.. take care. :)

junyeong0619 · 2026-05-25T02:47:45Z

Haha, just happened to wake up and couldn't help picking it up :) Thanks for the kind words — and for the thorough reviews!

junyeong0619 force-pushed the callsite-trigger branch from 06c462b to 8fceea8 Compare May 19, 2026 09:18

junyeong0619 force-pushed the callsite-trigger branch from 8fceea8 to 535551f Compare May 19, 2026 09:31

namhyung reviewed May 20, 2026

View reviewed changes

junyeong0619 force-pushed the callsite-trigger branch from 535551f to 9afd1c5 Compare May 20, 2026 13:07

namhyung reviewed May 23, 2026

View reviewed changes

junyeong0619 force-pushed the callsite-trigger branch from 9afd1c5 to 1b715ec Compare May 24, 2026 02:01

namhyung reviewed May 24, 2026

View reviewed changes

junyeong0619 force-pushed the callsite-trigger branch from 1b715ec to c4e3e83 Compare May 24, 2026 19:25

namhyung reviewed May 24, 2026

View reviewed changes

junyeong0619 added 6 commits May 25, 2026 10:28

junyeong0619 force-pushed the callsite-trigger branch from c4e3e83 to d28e9f5 Compare May 25, 2026 01:36

namhyung approved these changes May 25, 2026

View reviewed changes

namhyung merged commit 0a4f285 into namhyung:master May 25, 2026
3 checks passed

Conversation

junyeong0619 commented May 18, 2026

Uh oh!

namhyung commented May 19, 2026

Uh oh!

junyeong0619 commented May 19, 2026

Uh oh!

junyeong0619 commented May 19, 2026

Uh oh!

junyeong0619 commented May 19, 2026

Uh oh!

namhyung left a comment

Choose a reason for hiding this comment

Uh oh!

junyeong0619 commented May 20, 2026

Uh oh!

junyeong0619 commented May 20, 2026

Uh oh!

junyeong0619 commented May 20, 2026

Uh oh!

namhyung commented May 22, 2026

Uh oh!

junyeong0619 commented May 22, 2026

Uh oh!

namhyung left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

namhyung left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!