[PATCH v3 07/10] perf trace: Switch user option to use BPF filter

[Ian Rogers <irogers@google.com>]

The existing --uid option finds user processes by scanning
/proc. Scanning /proc is inherently racy due to processes being in
/proc but then exiting before perf does the perf_event_open. When the
perf_event_open fails perf will terminate reporting errors which is a
disappointing user experience. Scanning /proc when perf starts also
cannot inform perf of new user processes starting.

The ability to filter perf events with BPF isn't new, and has been in
the perf tool for 10 years:
https://lore.kernel.org/all/1444826502-49291-8-git-send-email-wangnan0@huawei.com/
An ability to do filtering on the command line with a BPF program
that's part of perf was added 2 years ago:
https://lore.kernel.org/all/20230314234237.3008956-1-namhyung@kernel.org/
This was then extended to support uids as a way of filtering:
https://lore.kernel.org/all/20240524205227.244375-1-irogers@google.com/

This change switches the perf trace --uid option to use the BPF filter
code to avoid the inherent race and existing failures. To ensure all
processes are considered by the filter, the change forces system-wide
mode.

Using BPF has permission issues in loading the BPF program not present
in scanning /proc. As the scanning approach would miss new programs
and fail due to the race, this is considered preferable. The change
also avoids opening a perf event per PID, which is less overhead in
the kernel.

Signed-off-by: Ian Rogers <irogers@google.com>
---
 tools/perf/builtin-trace.c | 26 +++++++++++++++++---------
 1 file changed, 17 insertions(+), 9 deletions(-)

diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
index 6ac51925ea42..1f7d2b3d8b3d 100644
--- a/tools/perf/builtin-trace.c
+++ b/tools/perf/builtin-trace.c
@@ -240,6 +240,7 @@ struct trace {
 		struct ordered_events	data;
 		u64			last;
 	} oe;
+	const char		*uid_str;
 };
 
 static void trace__load_vmlinux_btf(struct trace *trace __maybe_unused)
@@ -4401,8 +4402,8 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
 		evlist__add(evlist, pgfault_min);
 	}
 
-	/* Enable ignoring missing threads when -u/-p option is defined. */
-	trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
+	/* Enable ignoring missing threads when -p option is defined. */
+	trace->opts.ignore_missing_thread = trace->opts.target.pid;
 
 	if (trace->sched &&
 	    evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
@@ -5420,8 +5421,7 @@ int cmd_trace(int argc, const char **argv)
 		    "child tasks do not inherit counters"),
 	OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
 		     "number of mmap data pages", evlist__parse_mmap_pages),
-	OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
-		   "user to profile"),
+	OPT_STRING('u', "uid", &trace.uid_str, "user", "user to profile"),
 	OPT_CALLBACK(0, "duration", &trace, "float",
 		     "show only events with duration > N.M ms",
 		     trace__set_duration),
@@ -5762,11 +5762,19 @@ int cmd_trace(int argc, const char **argv)
 		goto out_close;
 	}
 
-	err = target__parse_uid(&trace.opts.target);
-	if (err) {
-		target__strerror(&trace.opts.target, err, bf, sizeof(bf));
-		fprintf(trace.output, "%s", bf);
-		goto out_close;
+	if (trace.uid_str) {
+		uid_t uid = parse_uid(trace.uid_str);
+
+		if (uid == UINT_MAX) {
+			ui__error("Invalid User: %s", trace.uid_str);
+			err = -EINVAL;
+			goto out_close;
+		}
+		err = parse_uid_filter(trace.evlist, uid);
+		if (err)
+			goto out_close;
+
+		trace.opts.target.system_wide = true;
 	}
 
 	if (!argc && target__none(&trace.opts.target))
-- 
2.49.0.850.g28803427d3-goog