Skip to contentSkip to conversation

perf record: Use pinned BPF program for filter (v1)

RFC New Namhyung Kim 1 participant 7 messages May 22, 2024

Subsystems (2)
lore
Share & tools
lore

Namhyung Kim

author maintainer
#0 lore raw

cover perf record: Use pinned BPF program for filter (v1)

Hello,

This is to support the unprivileged BPF filter for profiling per-task events. Until now only root (or any user with CAP_BPF) can use the filter and we cannot add a new unprivileged BPF program types. After talking with the BPF folks at LSF/MM/BPF 2024, I was told that this is the way to go. Finally I managed to make it working with pinned BPF objects. :)

But it requires the one-time setup (by root) before using it like below.

  # perf record --setup-filter pin

This will load the BPF program and maps and pin them in the BPF-fs. Then normal users can use the filter.

  $ perf record -o- -e cycles:u --filter 'period < 10000' perf test -w noploop | perf script -i-
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.011 MB - ]
        perf  759982 448227.214189:          1 cycles:u:      7f153719f4d0 _start+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2)
        perf  759982 448227.214195:          1 cycles:u:      7f153719f4d0 _start+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2)
        perf  759982 448227.214196:          7 cycles:u:      7f153719f4d0 _start+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2)
        perf  759982 448227.214196:        223 cycles:u:      7f153719f4d0 _start+0x0 (/usr/lib/x86_64-linux-gnu/ld-linux-x86-64.so.2)
        perf  759982 448227.214198:       9475 cycles:u:  ffffffff8ee012a0 [unknown] ([unknown])
        perf  759982 448227.548608:          1 cycles:u:      559a9f03c81c noploop+0x5c (/home/namhyung/linux/tools/perf/perf)
        perf  759982 448227.548611:          1 cycles:u:      559a9f03c81c noploop+0x5c (/home/namhyung/linux/tools/perf/perf)
        perf  759982 448227.548612:         12 cycles:u:      559a9f03c81c noploop+0x5c (/home/namhyung/linux/tools/perf/perf)
        perf  759982 448227.548613:        466 cycles:u:      559a9f03c81c noploop+0x5c (/home/namhyung/linux/tools/perf/perf)

It's also possible to unload (and unpin, of course) using this command:

  # perf record --setup-filter unpin

The code is avaiable in 'perf/pinned-filter-v1' branch at

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

Thanks,
Namhyung

Namhyung Kim (6):
perf bpf-filter: Make filters map a single entry hashmap perf bpf-filter: Pass 'target' to perf_bpf_filter__prepare() perf bpf-filter: Split per-task filter use case perf bpf-filter: Support pin/unpin BPF object
perf record: Fix a potential error handling issue perf record: Add --setup-filter option

 tools/perf/Documentation/perf-record.txt     |   5 +
 tools/perf/builtin-record.c                  |  23 +-
 tools/perf/builtin-stat.c                    |   2 +-
 tools/perf/builtin-trace.c                   |   2 +-
 tools/perf/util/bpf-filter.c                 | 369 +++++++++++++++++--
 tools/perf/util/bpf-filter.h                 |  19 +-
 tools/perf/util/bpf_skel/sample-filter.h     |   4 +-
 tools/perf/util/bpf_skel/sample_filter.bpf.c |  58 ++-
 tools/perf/util/evlist.c                     |   5 +-
 tools/perf/util/evlist.h                     |   4 +-
 tools/perf/util/python.c                     |   3 +-
 11 files changed, 428 insertions(+), 66 deletions(-)

base-commit: ea558c86248b4955e5c5f3c0c921df450880605e

Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#1 lore raw

PATCH 1/6 perf bpf-filter: Make filters map a single entry hashmap

And the value is now an array. This is to support multiple filter entries in the map later.

No functional changes intended.

---
 tools/perf/util/bpf-filter.c                 | 81 ++++++++++++++------
 tools/perf/util/bpf_skel/sample-filter.h     |  3 +-
 tools/perf/util/bpf_skel/sample_filter.bpf.c | 37 +++++----
 3 files changed, 78 insertions(+), 43 deletions(-)
tools/perf/util/bpf-filter.c +56 -25
index b51544996046..12e9c7dbb4dd 100644
@@ -87,71 +87,102 @@ static int check_sample_flags(struct evsel *evsel, struct perf_bpf_filter_expr *
 
 int perf_bpf_filter__prepare(struct evsel *evsel)
 {
-	int i, x, y, fd;
+	int i, x, y, fd, ret;
 	struct sample_filter_bpf *skel;
 	struct bpf_program *prog;
 	struct bpf_link *link;
 	struct perf_bpf_filter_expr *expr;
+	struct perf_bpf_filter_entry *entry;
+
+	entry = calloc(MAX_FILTERS, sizeof(*entry));
+	if (entry == NULL)
+		return -1;
 
 	skel = sample_filter_bpf__open_and_load();
 	if (!skel) {
 		pr_err("Failed to load perf sample-filter BPF skeleton\n");
-		return -1;
+		ret = -EPERM;
+		goto err;
 	}
 
 	i = 0;
 	fd = bpf_map__fd(skel->maps.filters);
 	list_for_each_entry(expr, &evsel->bpf_filters, list) {
-		struct perf_bpf_filter_entry entry = {
-			.op = expr->op,
-			.part = expr->part,
-			.flags = expr->sample_flags,
-			.value = expr->val,
-		};
+		if (check_sample_flags(evsel, expr) < 0) {
+			ret = -EINVAL;
+			goto err;
+		}
 
-		if (check_sample_flags(evsel, expr) < 0)
-			return -1;
+		if (i == MAX_FILTERS) {
+			ret = -E2BIG;
+			goto err;
+		}
 
-		bpf_map_update_elem(fd, &i, &entry, BPF_ANY);
+		entry[i].op = expr->op;
+		entry[i].part = expr->part;
+		entry[i].flags = expr->sample_flags;
+		entry[i].value = expr->val;
 		i++;
 
 		if (expr->op == PBF_OP_GROUP_BEGIN) {
 			struct perf_bpf_filter_expr *group;
 
 			list_for_each_entry(group, &expr->groups, list) {
-				struct perf_bpf_filter_entry group_entry = {
-					.op = group->op,
-					.part = group->part,
-					.flags = group->sample_flags,
-					.value = group->val,
-				};
-				bpf_map_update_elem(fd, &i, &group_entry, BPF_ANY);
+				if (i == MAX_FILTERS) {
+					ret = -E2BIG;
+					goto err;
+				}
+
+				entry[i].op = group->op;
+				entry[i].part = group->part;
+				entry[i].flags = group->sample_flags;
+				entry[i].value = group->val;
 				i++;
 			}
 
-			memset(&entry, 0, sizeof(entry));
-			entry.op = PBF_OP_GROUP_END;
-			bpf_map_update_elem(fd, &i, &entry, BPF_ANY);
+			if (i == MAX_FILTERS) {
+				ret = -E2BIG;
+				goto err;
+			}
+
+			entry[i].op = PBF_OP_GROUP_END;
 			i++;
 		}
 	}
 
-	if (i > MAX_FILTERS) {
-		pr_err("Too many filters: %d (max = %d)\n", i, MAX_FILTERS);
-		return -1;
+	if (i < MAX_FILTERS) {
+		/* to terminate the loop early */
+		entry[i].op = PBF_OP_DONE;
+		i++;
+	}
+
+	/* The filters map has only one entry for now */
+	i = 0;
+	if (bpf_map_update_elem(fd, &i, entry, BPF_ANY) < 0) {
+		ret = -errno;
+		pr_err("Failed to update the filter map\n");
+		goto err;
 	}
+
 	prog = skel->progs.perf_sample_filter;
 	for (x = 0; x < xyarray__max_x(evsel->core.fd); x++) {
 		for (y = 0; y < xyarray__max_y(evsel->core.fd); y++) {
 			link = bpf_program__attach_perf_event(prog, FD(evsel, x, y));
 			if (IS_ERR(link)) {
 				pr_err("Failed to attach perf sample-filter program\n");
-				return PTR_ERR(link);
+				ret = PTR_ERR(link);
+				goto err;
 			}
 		}
 	}
+	free(entry);
 	evsel->bpf_skel = skel;
 	return 0;
+
+err:
+	free(entry);
+	sample_filter_bpf__destroy(skel);
+	return ret;
 }
 
 int perf_bpf_filter__destroy(struct evsel *evsel)
tools/perf/util/bpf_skel/sample-filter.h +2 -1
index 2e96e1ab084a..cf18f570eef4 100644
@@ -14,6 +14,7 @@ enum perf_bpf_filter_op {
 	PBF_OP_AND,
 	PBF_OP_GROUP_BEGIN,
 	PBF_OP_GROUP_END,
+	PBF_OP_DONE,
 };
 
 /* BPF map entry for filtering */
@@ -24,4 +25,4 @@ struct perf_bpf_filter_entry {
 	__u64 value;
 };
 
-#endif /* PERF_UTIL_BPF_SKEL_SAMPLE_FILTER_H */
\ No newline at end of file
+#endif /* PERF_UTIL_BPF_SKEL_SAMPLE_FILTER_H */
tools/perf/util/bpf_skel/sample_filter.bpf.c +20 -17
index fb94f5280626..5f17cd6458b7 100644
@@ -9,10 +9,10 @@
 
 /* BPF map that will be filled by user space */
 struct filters {
-	__uint(type, BPF_MAP_TYPE_ARRAY);
+	__uint(type, BPF_MAP_TYPE_HASH);
 	__type(key, int);
-	__type(value, struct perf_bpf_filter_entry);
-	__uint(max_entries, MAX_FILTERS);
+	__type(value, struct perf_bpf_filter_entry[MAX_FILTERS]);
+	__uint(max_entries, 1);
 } filters SEC(".maps");
 
 int dropped;
@@ -144,35 +144,35 @@ int perf_sample_filter(void *ctx)
 
 	kctx = bpf_cast_to_kern_ctx(ctx);
 
-	for (i = 0; i < MAX_FILTERS; i++) {
-		int key = i; /* needed for verifier :( */
+	i = 0;
+	entry = bpf_map_lookup_elem(&filters, &i);
+	if (entry == NULL)
+		goto drop;
 
-		entry = bpf_map_lookup_elem(&filters, &key);
-		if (entry == NULL)
-			break;
-		sample_data = perf_get_sample(kctx, entry);
+	for (i = 0; i < MAX_FILTERS; i++) {
+		sample_data = perf_get_sample(kctx, &entry[i]);
 
-		switch (entry->op) {
+		switch (entry[i].op) {
 		case PBF_OP_EQ:
-			CHECK_RESULT(sample_data, ==, entry->value)
+			CHECK_RESULT(sample_data, ==, entry[i].value)
 			break;
 		case PBF_OP_NEQ:
-			CHECK_RESULT(sample_data, !=, entry->value)
+			CHECK_RESULT(sample_data, !=, entry[i].value)
 			break;
 		case PBF_OP_GT:
-			CHECK_RESULT(sample_data, >, entry->value)
+			CHECK_RESULT(sample_data, >, entry[i].value)
 			break;
 		case PBF_OP_GE:
-			CHECK_RESULT(sample_data, >=, entry->value)
+			CHECK_RESULT(sample_data, >=, entry[i].value)
 			break;
 		case PBF_OP_LT:
-			CHECK_RESULT(sample_data, <, entry->value)
+			CHECK_RESULT(sample_data, <, entry[i].value)
 			break;
 		case PBF_OP_LE:
-			CHECK_RESULT(sample_data, <=, entry->value)
+			CHECK_RESULT(sample_data, <=, entry[i].value)
 			break;
 		case PBF_OP_AND:
-			CHECK_RESULT(sample_data, &, entry->value)
+			CHECK_RESULT(sample_data, &, entry[i].value)
 			break;
 		case PBF_OP_GROUP_BEGIN:
 			in_group = 1;
@@ -183,6 +183,9 @@ int perf_sample_filter(void *ctx)
 				goto drop;
 			in_group = 0;
 			break;
+		case PBF_OP_DONE:
+			/* no failures so far, accept it */
+			return 1;
 		}
 	}
 	/* generate sample data */
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#2 lore raw

PATCH 2/6 perf bpf-filter: Pass 'target' to perf_bpf_filter__prepare()

In reply to Namhyung Kim

This is needed to prepare target-specific actions in the later patch. We want to reuse the pinned BPF program and map for regular users to profile their own processes.

---
 tools/perf/builtin-record.c  | 2 +-
 tools/perf/builtin-stat.c    | 2 +-
 tools/perf/builtin-trace.c   | 2 +-
 tools/perf/util/bpf-filter.c | 2 +-
 tools/perf/util/bpf-filter.h | 6 ++++--
 tools/perf/util/evlist.c     | 5 +++--
 tools/perf/util/evlist.h     | 4 +++-
 tools/perf/util/python.c     | 3 ++-
 8 files changed, 16 insertions(+), 10 deletions(-)
tools/perf/builtin-record.c +1 -1
index 66a3de8ac661..8ec0b1607603 100644
@@ -1389,7 +1389,7 @@ static int record__open(struct record *rec)
 "even with a suitable vmlinux or kallsyms file.\n\n");
 	}
 
-	if (evlist__apply_filters(evlist, &pos)) {
+	if (evlist__apply_filters(evlist, &pos, &opts->target)) {
 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 			pos->filter ?: "BPF", evsel__name(pos), errno,
 			str_error_r(errno, msg, sizeof(msg)));
tools/perf/builtin-stat.c +1 -1
index 35f79b48e8dc..23ba1e54291b 100644
@@ -813,7 +813,7 @@ static int __run_perf_stat(int argc, const char **argv, int run_idx)
 			return -1;
 	}
 
-	if (evlist__apply_filters(evsel_list, &counter)) {
+	if (evlist__apply_filters(evsel_list, &counter, &target)) {
 		pr_err("failed to set filter \"%s\" on event %s with %d (%s)\n",
 			counter->filter, evsel__name(counter), errno,
 			str_error_r(errno, msg, sizeof(msg)));
tools/perf/builtin-trace.c +1 -1
index 51eca671c797..fd005ebdec04 100644
@@ -3974,7 +3974,7 @@ static int trace__run(struct trace *trace, int argc, const char **argv)
 	err = trace__expand_filters(trace, &evsel);
 	if (err)
 		goto out_delete_evlist;
-	err = evlist__apply_filters(evlist, &evsel);
+	err = evlist__apply_filters(evlist, &evsel, &trace->opts.target);
 	if (err < 0)
 		goto out_error_apply_filters;
 
tools/perf/util/bpf-filter.c +1 -1
index 12e9c7dbb4dd..f43b9e61bf42 100644
@@ -85,7 +85,7 @@ static int check_sample_flags(struct evsel *evsel, struct perf_bpf_filter_expr *
 	return -1;
 }
 
-int perf_bpf_filter__prepare(struct evsel *evsel)
+int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target __maybe_unused)
 {
 	int i, x, y, fd, ret;
 	struct sample_filter_bpf *skel;
tools/perf/util/bpf-filter.h +4 -2
index 7afd159411b8..955ef4e3a931 100644
@@ -16,13 +16,14 @@ struct perf_bpf_filter_expr {
 };
 
 struct evsel;
+struct target;
 
 #ifdef HAVE_BPF_SKEL
 struct perf_bpf_filter_expr *perf_bpf_filter_expr__new(unsigned long sample_flags, int part,
 						       enum perf_bpf_filter_op op,
 						       unsigned long val);
 int perf_bpf_filter__parse(struct list_head *expr_head, const char *str);
-int perf_bpf_filter__prepare(struct evsel *evsel);
+int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target);
 int perf_bpf_filter__destroy(struct evsel *evsel);
 u64 perf_bpf_filter__lost_count(struct evsel *evsel);
 
@@ -33,7 +34,8 @@ static inline int perf_bpf_filter__parse(struct list_head *expr_head __maybe_unu
 {
 	return -EOPNOTSUPP;
 }
-static inline int perf_bpf_filter__prepare(struct evsel *evsel __maybe_unused)
+static inline int perf_bpf_filter__prepare(struct evsel *evsel __maybe_unused,
+					   struct target *target __maybe_unused)
 {
 	return -EOPNOTSUPP;
 }
tools/perf/util/evlist.c +3 -2
index 3a719edafc7a..1417f9a23083 100644
@@ -1086,7 +1086,8 @@ int evlist__create_maps(struct evlist *evlist, struct target *target)
 	return -1;
 }
 
-int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
+int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel,
+			  struct target *target)
 {
 	struct evsel *evsel;
 	int err = 0;
@@ -1108,7 +1109,7 @@ int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel)
 		 * non-tracepoint events can have BPF filters.
 		 */
 		if (!list_empty(&evsel->bpf_filters)) {
-			err = perf_bpf_filter__prepare(evsel);
+			err = perf_bpf_filter__prepare(evsel, target);
 			if (err) {
 				*err_evsel = evsel;
 				break;
tools/perf/util/evlist.h +3 -1
index cb91dc9117a2..cccc34da5a02 100644
@@ -20,6 +20,7 @@ struct pollfd;
 struct thread_map;
 struct perf_cpu_map;
 struct record_opts;
+struct target;
 
 /*
  * State machine of bkw_mmap_state:
@@ -212,7 +213,8 @@ void evlist__enable_non_dummy(struct evlist *evlist);
 void evlist__set_selected(struct evlist *evlist, struct evsel *evsel);
 
 int evlist__create_maps(struct evlist *evlist, struct target *target);
-int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel);
+int evlist__apply_filters(struct evlist *evlist, struct evsel **err_evsel,
+			  struct target *target);
 
 u64 __evlist__combined_sample_type(struct evlist *evlist);
 u64 evlist__combined_sample_type(struct evlist *evlist);
tools/perf/util/python.c +2 -1
index 0aeb97c11c03..5e015d0c0df5 100644
@@ -175,7 +175,8 @@ int bpf_counter__disable(struct evsel *evsel __maybe_unused)
 
 // not to drag util/bpf-filter.c
 #ifdef HAVE_BPF_SKEL
-int perf_bpf_filter__prepare(struct evsel *evsel __maybe_unused)
+int perf_bpf_filter__prepare(struct evsel *evsel __maybe_unused,
+			     struct target *target __maybe_unused)
 {
 	return 0;
 }
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#3 lore raw

PATCH 3/6 perf bpf-filter: Split per-task filter use case

In reply to Namhyung Kim

If the target is a list of tasks, it can use a shared hash map for filter expressions. The key of the filter map is an integer index like in an array. A separate pid_hash map is added to get the index for the filter map using the tgid.

For system-wide mode including per-cpu or per-user targets are handled by the single entry map like before.

---
 tools/perf/util/bpf-filter.c                 | 186 +++++++++++++++----
 tools/perf/util/bpf_skel/sample-filter.h     |   1 +
 tools/perf/util/bpf_skel/sample_filter.bpf.c |  23 ++-
 3 files changed, 168 insertions(+), 42 deletions(-)
tools/perf/util/bpf-filter.c +146 -40
index f43b9e61bf42..2187975189c9 100644
@@ -3,10 +3,13 @@
 
 #include <bpf/bpf.h>
 #include <linux/err.h>
+#include <api/fs/fs.h>
 #include <internal/xyarray.h>
+#include <perf/threadmap.h>
 
 #include "util/debug.h"
 #include "util/evsel.h"
+#include "util/target.h"
 
 #include "util/bpf-filter.h"
 #include <util/bpf-filter-flex.h>
@@ -85,38 +88,17 @@ static int check_sample_flags(struct evsel *evsel, struct perf_bpf_filter_expr *
 	return -1;
 }
 
-int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target __maybe_unused)
+static int get_filter_entries(struct evsel *evsel, struct perf_bpf_filter_entry *entry)
 {
-	int i, x, y, fd, ret;
-	struct sample_filter_bpf *skel;
-	struct bpf_program *prog;
-	struct bpf_link *link;
+	int i = 0;
 	struct perf_bpf_filter_expr *expr;
-	struct perf_bpf_filter_entry *entry;
-
-	entry = calloc(MAX_FILTERS, sizeof(*entry));
-	if (entry == NULL)
-		return -1;
-
-	skel = sample_filter_bpf__open_and_load();
-	if (!skel) {
-		pr_err("Failed to load perf sample-filter BPF skeleton\n");
-		ret = -EPERM;
-		goto err;
-	}
 
-	i = 0;
-	fd = bpf_map__fd(skel->maps.filters);
 	list_for_each_entry(expr, &evsel->bpf_filters, list) {
-		if (check_sample_flags(evsel, expr) < 0) {
-			ret = -EINVAL;
-			goto err;
-		}
+		if (check_sample_flags(evsel, expr) < 0)
+			return -EINVAL;
 
-		if (i == MAX_FILTERS) {
-			ret = -E2BIG;
-			goto err;
-		}
+		if (i == MAX_FILTERS)
+			return -E2BIG;
 
 		entry[i].op = expr->op;
 		entry[i].part = expr->part;
@@ -128,10 +110,8 @@ int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target __maybe_
 			struct perf_bpf_filter_expr *group;
 
 			list_for_each_entry(group, &expr->groups, list) {
-				if (i == MAX_FILTERS) {
-					ret = -E2BIG;
-					goto err;
-				}
+				if (i == MAX_FILTERS)
+					return -E2BIG;
 
 				entry[i].op = group->op;
 				entry[i].part = group->part;
@@ -140,10 +120,8 @@ int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target __maybe_
 				i++;
 			}
 
-			if (i == MAX_FILTERS) {
-				ret = -E2BIG;
-				goto err;
-			}
+			if (i == MAX_FILTERS)
+				return -E2BIG;
 
 			entry[i].op = PBF_OP_GROUP_END;
 			i++;
@@ -155,15 +133,143 @@ int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target __maybe_
 		entry[i].op = PBF_OP_DONE;
 		i++;
 	}
+	return 0;
+}
+
+static int convert_to_tgid(int tid)
+{
+	char path[128];
+	char *buf, *p, *q;
+	int tgid;
+	size_t len;
+
+	scnprintf(path, sizeof(path), "%d/status", tid);
+	if (procfs__read_str(path, &buf, &len) < 0)
+		return -1;
 
-	/* The filters map has only one entry for now */
-	i = 0;
-	if (bpf_map_update_elem(fd, &i, entry, BPF_ANY) < 0) {
-		ret = -errno;
-		pr_err("Failed to update the filter map\n");
+	p = strstr(buf, "Tgid:");
+	if (p == NULL) {
+		free(buf);
+		return -1;
+	}
+
+	tgid = strtol(p + 6, &q, 0);
+	free(buf);
+	if (*q != '\n')
+		return -1;
+
+	return tgid;
+}
+
+static int update_pid_hash(struct sample_filter_bpf *skel, struct evsel *evsel,
+			   struct perf_bpf_filter_entry *entry)
+{
+	int filter_idx;
+	int nr, last;
+	int fd = bpf_map__fd(skel->maps.filters);
+	struct perf_thread_map *threads;
+
+	/* Find the first available entry in the filters map */
+	for (filter_idx = 0; filter_idx < MAX_FILTERS; filter_idx++) {
+		if (bpf_map_update_elem(fd, &filter_idx, entry, BPF_NOEXIST) == 0)
+			break;
+	}
+
+	if (filter_idx == MAX_FILTERS) {
+		pr_err("Too many users for the filter map\n");
+		return -EBUSY;
+	}
+
+	threads = perf_evsel__threads(&evsel->core);
+	if (threads == NULL) {
+		pr_err("Cannot get the thread list of the event\n");
+		return -EINVAL;
+	}
+
+	/* save the index to a hash map */
+	fd = bpf_map__fd(skel->maps.pid_hash);
+
+	last = -1;
+	nr = perf_thread_map__nr(threads);
+	for (int i = 0; i < nr; i++) {
+		int pid = perf_thread_map__pid(threads, i);
+		int tgid;
+
+		/* it actually needs tgid, let's get tgid from /proc. */
+		tgid = convert_to_tgid(pid);
+		if (tgid < 0) {
+			/* the thread may be dead, ignore. */
+			continue;
+		}
+
+		if (tgid == last)
+			continue;
+		last = tgid;
+
+		if (bpf_map_update_elem(fd, &tgid, &filter_idx, BPF_ANY) < 0) {
+			pr_err("Failed to update the pid hash\n");
+			return -errno;
+		}
+		pr_debug("pid hash: %d -> %d\n", tgid, filter_idx);
+	}
+	return 0;
+}
+
+int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target)
+{
+	int i, x, y, fd, ret;
+	struct sample_filter_bpf *skel = NULL;
+	struct bpf_program *prog;
+	struct bpf_link *link;
+	struct perf_bpf_filter_entry *entry;
+	bool needs_pid_hash = !target__has_cpu(target) && !target->uid_str;
+
+	entry = calloc(MAX_FILTERS, sizeof(*entry));
+	if (entry == NULL)
+		return -1;
+
+	ret = get_filter_entries(evsel, entry);
+	if (ret < 0) {
+		pr_err("Failed to process filter entries\n");
+		goto err;
+	}
+
+	skel = sample_filter_bpf__open();
+	if (!skel) {
+		pr_err("Failed to open perf sample-filter BPF skeleton\n");
+		ret = -EPERM;
 		goto err;
 	}
 
+	if (needs_pid_hash) {
+		bpf_map__set_max_entries(skel->maps.filters, MAX_FILTERS);
+		bpf_map__set_max_entries(skel->maps.pid_hash, MAX_PIDS);
+		skel->rodata->use_pid_hash = 1;
+	}
+
+	if (sample_filter_bpf__load(skel) < 0) {
+		pr_err("Failed to load perf sample-filter BPF skeleton\n");
+		ret = -EPERM;
+		goto err;
+	}
+
+	if (needs_pid_hash) {
+		/* The filters map is shared among other processes  */
+		ret = update_pid_hash(skel, evsel, entry);
+		if (ret < 0)
+			goto err;
+	} else {
+		i = 0;
+		fd = bpf_map__fd(skel->maps.filters);
+
+		/* The filters map has only one entry in this case */
+		if (bpf_map_update_elem(fd, &i, entry, BPF_ANY) < 0) {
+			ret = -errno;
+			pr_err("Failed to update the filter map\n");
+			goto err;
+		}
+	}
+
 	prog = skel->progs.perf_sample_filter;
 	for (x = 0; x < xyarray__max_x(evsel->core.fd); x++) {
 		for (y = 0; y < xyarray__max_y(evsel->core.fd); y++) {
tools/perf/util/bpf_skel/sample-filter.h +1 -0
index cf18f570eef4..375e27206640 100644
@@ -2,6 +2,7 @@
 #define PERF_UTIL_BPF_SKEL_SAMPLE_FILTER_H
 
 #define MAX_FILTERS  64
+#define MAX_PIDS     (16 * 1024)
 
 /* supported filter operations */
 enum perf_bpf_filter_op {
tools/perf/util/bpf_skel/sample_filter.bpf.c +21 -2
index 5f17cd6458b7..1ccb0e8be73b 100644
@@ -15,7 +15,16 @@ struct filters {
 	__uint(max_entries, 1);
 } filters SEC(".maps");
 
+/* tgid to filter index */
+struct pid_hash {
+	__uint(type, BPF_MAP_TYPE_HASH);
+	__type(key, int);
+	__type(value, int);
+	__uint(max_entries, 1);
+} pid_hash SEC(".maps");
+
 int dropped;
+volatile const int use_pid_hash;
 
 void *bpf_cast_to_kern_ctx(void *) __ksym;
 
@@ -140,11 +149,21 @@ int perf_sample_filter(void *ctx)
 	__u64 sample_data;
 	int in_group = 0;
 	int group_result = 0;
-	int i;
+	int i = 0;
 
 	kctx = bpf_cast_to_kern_ctx(ctx);
 
-	i = 0;
+	if (use_pid_hash) {
+		int tgid = bpf_get_current_pid_tgid() >> 32;
+		int *idx;
+
+		idx = bpf_map_lookup_elem(&pid_hash, &tgid);
+		if (idx)
+			i = *idx;
+		else
+			goto drop;
+	}
+
 	entry = bpf_map_lookup_elem(&filters, &i);
 	if (entry == NULL)
 		goto drop;
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#4 lore raw

PATCH 4/6 perf bpf-filter: Support pin/unpin BPF object

In reply to Namhyung Kim

And use the pinned objects for unprivileged users to profile their own tasks. The BPF objects need to be pinned in the BPF-fs by root first and it'll be handled in the later patch.

---
 tools/perf/util/bpf-filter.c | 230 +++++++++++++++++++++++++++++------
 tools/perf/util/bpf-filter.h |  13 ++
 2 files changed, 209 insertions(+), 34 deletions(-)
tools/perf/util/bpf-filter.c +196 -34
index 2187975189c9..ac84260082a7 100644
@@ -1,5 +1,8 @@
 /* SPDX-License-Identifier: GPL-2.0 */
 #include <stdlib.h>
+#include <fcntl.h>
+#include <sys/ioctl.h>
+#include <sys/stat.h>
 
 #include <bpf/bpf.h>
 #include <linux/err.h>
@@ -23,6 +26,9 @@
 #define __PERF_SAMPLE_TYPE(st, opt)	{ st, #st, opt }
 #define PERF_SAMPLE_TYPE(_st, opt)	__PERF_SAMPLE_TYPE(PERF_SAMPLE_##_st, opt)
 
+/* Index in the pinned 'filters' map.  Should be released after use. */
+static int pinned_filter_idx = -1;
+
 static const struct perf_sample_info {
 	u64 type;
 	const char *name;
@@ -47,6 +53,8 @@ static const struct perf_sample_info {
 	PERF_SAMPLE_TYPE(DATA_PAGE_SIZE, "--data-page-size"),
 };
 
+static int get_pinned_fd(const char *name);
+
 static const struct perf_sample_info *get_sample_info(u64 flags)
 {
 	size_t i;
@@ -161,19 +169,26 @@ static int convert_to_tgid(int tid)
 	return tgid;
 }
 
-static int update_pid_hash(struct sample_filter_bpf *skel, struct evsel *evsel,
-			   struct perf_bpf_filter_entry *entry)
+static int update_pid_hash(struct evsel *evsel, struct perf_bpf_filter_entry *entry)
 {
 	int filter_idx;
-	int nr, last;
-	int fd = bpf_map__fd(skel->maps.filters);
+	int fd, nr, last;
 	struct perf_thread_map *threads;
 
+	fd = get_pinned_fd("filters");
+	if (fd < 0) {
+		pr_debug("cannot get fd for 'filters' map\n");
+		return fd;
+	}
+
 	/* Find the first available entry in the filters map */
 	for (filter_idx = 0; filter_idx < MAX_FILTERS; filter_idx++) {
-		if (bpf_map_update_elem(fd, &filter_idx, entry, BPF_NOEXIST) == 0)
+		if (bpf_map_update_elem(fd, &filter_idx, entry, BPF_NOEXIST) == 0) {
+			pinned_filter_idx = filter_idx;
 			break;
+		}
 	}
+	close(fd);
 
 	if (filter_idx == MAX_FILTERS) {
 		pr_err("Too many users for the filter map\n");
@@ -187,7 +202,9 @@ static int update_pid_hash(struct sample_filter_bpf *skel, struct evsel *evsel,
 	}
 
 	/* save the index to a hash map */
-	fd = bpf_map__fd(skel->maps.pid_hash);
+	fd = get_pinned_fd("pid_hash");
+	if (fd < 0)
+		return fd;
 
 	last = -1;
 	nr = perf_thread_map__nr(threads);
@@ -208,10 +225,12 @@ static int update_pid_hash(struct sample_filter_bpf *skel, struct evsel *evsel,
 
 		if (bpf_map_update_elem(fd, &tgid, &filter_idx, BPF_ANY) < 0) {
 			pr_err("Failed to update the pid hash\n");
-			return -errno;
+			close(fd);
+			return -1;
 		}
 		pr_debug("pid hash: %d -> %d\n", tgid, filter_idx);
 	}
+	close(fd);
 	return 0;
 }
 
@@ -234,40 +253,47 @@ int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target)
 		goto err;
 	}
 
-	skel = sample_filter_bpf__open();
-	if (!skel) {
-		pr_err("Failed to open perf sample-filter BPF skeleton\n");
-		ret = -EPERM;
-		goto err;
-	}
+	if (needs_pid_hash && geteuid() != 0) {
+		/* The filters map is shared among other processes */
+		ret = update_pid_hash(evsel, entry);
+		if (ret < 0)
+			goto err;
 
-	if (needs_pid_hash) {
-		bpf_map__set_max_entries(skel->maps.filters, MAX_FILTERS);
-		bpf_map__set_max_entries(skel->maps.pid_hash, MAX_PIDS);
-		skel->rodata->use_pid_hash = 1;
+		fd = get_pinned_fd("perf_sample_filter");
+		if (fd < 0) {
+			ret = fd;
+			goto err;
+		}
+
+		for (x = 0; x < xyarray__max_x(evsel->core.fd); x++) {
+			for (y = 0; y < xyarray__max_y(evsel->core.fd); y++) {
+				ret = ioctl(FD(evsel, x, y), PERF_EVENT_IOC_SET_BPF, fd);
+				if (ret < 0) {
+					pr_err("Failed to attach perf sample-filter\n");
+					goto err;
+				}
+			}
+		}
+
+		free(entry);
+		return 0;
 	}
 
-	if (sample_filter_bpf__load(skel) < 0) {
+	skel = sample_filter_bpf__open_and_load();
+	if (!skel) {
+		ret = -errno;
 		pr_err("Failed to load perf sample-filter BPF skeleton\n");
-		ret = -EPERM;
 		goto err;
 	}
 
-	if (needs_pid_hash) {
-		/* The filters map is shared among other processes  */
-		ret = update_pid_hash(skel, evsel, entry);
-		if (ret < 0)
-			goto err;
-	} else {
-		i = 0;
-		fd = bpf_map__fd(skel->maps.filters);
-
-		/* The filters map has only one entry in this case */
-		if (bpf_map_update_elem(fd, &i, entry, BPF_ANY) < 0) {
-			ret = -errno;
-			pr_err("Failed to update the filter map\n");
-			goto err;
-		}
+	i = 0;
+	fd = bpf_map__fd(skel->maps.filters);
+
+	/* The filters map has only one entry in this case */
+	if (bpf_map_update_elem(fd, &i, entry, BPF_ANY) < 0) {
+		ret = -errno;
+		pr_err("Failed to update the filter map\n");
+		goto err;
 	}
 
 	prog = skel->progs.perf_sample_filter;
@@ -300,6 +326,15 @@ int perf_bpf_filter__destroy(struct evsel *evsel)
 		free(expr);
 	}
 	sample_filter_bpf__destroy(evsel->bpf_skel);
+
+	if (pinned_filter_idx >= 0) {
+		int fd = get_pinned_fd("filters");
+
+		bpf_map_delete_elem(fd, &pinned_filter_idx);
+		pinned_filter_idx = -1;
+		close(fd);
+	}
+
 	return 0;
 }
 
@@ -342,3 +377,130 @@ int perf_bpf_filter__parse(struct list_head *expr_head, const char *str)
 
 	return ret;
 }
+
+int perf_bpf_filter__pin(void)
+{
+	struct sample_filter_bpf *skel;
+	char *path = NULL;
+	int dir_fd, ret = -1;
+
+	skel = sample_filter_bpf__open();
+	if (!skel) {
+		ret = -errno;
+		pr_err("Failed to open perf sample-filter BPF skeleton\n");
+		goto err;
+	}
+
+	/* pinned program will use pid-hash */
+	bpf_map__set_max_entries(skel->maps.filters, MAX_FILTERS);
+	bpf_map__set_max_entries(skel->maps.pid_hash, MAX_PIDS);
+	skel->rodata->use_pid_hash = 1;
+
+	if (sample_filter_bpf__load(skel) < 0) {
+		ret = -errno;
+		pr_err("Failed to load perf sample-filter BPF skeleton\n");
+		goto err;
+	}
+
+	if (asprintf(&path, "%s/fs/bpf/%s", sysfs__mountpoint(),
+		     PERF_BPF_FILTER_PIN_PATH) < 0) {
+		ret = -errno;
+		pr_err("Failed to allocate pathname in the BPF-fs\n");
+		goto err;
+	}
+
+	ret = bpf_object__pin(skel->obj, path);
+	if (ret < 0) {
+		pr_err("Failed to pin BPF filter objects\n");
+		goto err;
+	}
+
+	/* setup access permissions for the pinned objects */
+	dir_fd = open(path, O_PATH);
+	if (dir_fd < 0) {
+		bpf_object__unpin(skel->obj, path);
+		ret = dir_fd;
+		goto err;
+	}
+
+	/* BPF-fs root has the sticky bit */
+	if (fchmodat(dir_fd, "..", 01755, 0) < 0) {
+		pr_debug("chmod for BPF-fs failed\n");
+		ret = -errno;
+	}
+
+	/* perf_filter directory */
+	if (fchmod(dir_fd, 0755) < 0) {
+		pr_debug("chmod for perf_filter directory failed?\n");
+		/*
+		 * On my machine, the operation succeeds but returns failure.
+		 * Don't know why, let's ignore the failure for now.  If it's
+		 * a real error, the following fchmodat() should fail too.
+		 */
+	}
+
+	/* programs need write permission for some reason */
+	if (fchmodat(dir_fd, "perf_sample_filter", 0777, 0) < 0) {
+		pr_debug("chmod for perf_sample_filter failed\n");
+		ret = -errno;
+	}
+	/* maps */
+	if (fchmodat(dir_fd, "filters", 0666, 0) < 0) {
+		pr_debug("chmod for filters failed\n");
+		ret = -errno;
+	}
+	if (fchmodat(dir_fd, "pid_hash", 0666, 0) < 0) {
+		pr_debug("chmod for pid_hash failed\n");
+		ret = -errno;
+	}
+
+	close(dir_fd);
+
+err:
+	free(path);
+	sample_filter_bpf__destroy(skel);
+	return ret;
+}
+
+int perf_bpf_filter__unpin(void)
+{
+	struct sample_filter_bpf *skel;
+	char *path = NULL;
+	int ret = -1;
+
+	skel = sample_filter_bpf__open_and_load();
+	if (!skel) {
+		ret = -errno;
+		pr_err("Failed to open perf sample-filter BPF skeleton\n");
+		goto err;
+	}
+
+	if (asprintf(&path, "%s/fs/bpf/%s", sysfs__mountpoint(),
+		     PERF_BPF_FILTER_PIN_PATH) < 0) {
+		ret = -errno;
+		pr_err("Failed to allocate pathname in the BPF-fs\n");
+		goto err;
+	}
+
+	ret = bpf_object__unpin(skel->obj, path);
+
+err:
+	free(path);
+	sample_filter_bpf__destroy(skel);
+	return ret;
+}
+
+static int get_pinned_fd(const char *name)
+{
+	char *path = NULL;
+	int fd;
+
+	if (asprintf(&path, "%s/fs/bpf/%s/%s", sysfs__mountpoint(),
+		     PERF_BPF_FILTER_PIN_PATH, name) < 0)
+		return -1;
+
+	fd = bpf_obj_get(path);
+
+	free(path);
+	return fd;
+}
tools/perf/util/bpf-filter.h +13 -0
index 955ef4e3a931..32ff8299a5f3 100644
@@ -18,6 +18,9 @@ struct perf_bpf_filter_expr {
 struct evsel;
 struct target;
 
+/* path in BPF-fs for the pinned program and maps */
+#define PERF_BPF_FILTER_PIN_PATH  "perf_filter"
+
 #ifdef HAVE_BPF_SKEL
 struct perf_bpf_filter_expr *perf_bpf_filter_expr__new(unsigned long sample_flags, int part,
 						       enum perf_bpf_filter_op op,
@@ -26,6 +29,8 @@ int perf_bpf_filter__parse(struct list_head *expr_head, const char *str);
 int perf_bpf_filter__prepare(struct evsel *evsel, struct target *target);
 int perf_bpf_filter__destroy(struct evsel *evsel);
 u64 perf_bpf_filter__lost_count(struct evsel *evsel);
+int perf_bpf_filter__pin(void);
+int perf_bpf_filter__unpin(void);
 
 #else /* !HAVE_BPF_SKEL */
 
@@ -47,5 +52,13 @@ static inline u64 perf_bpf_filter__lost_count(struct evsel *evsel __maybe_unused
 {
 	return 0;
 }
+static inline int perf_bpf_filter__pin(void)
+{
+	return -EOPNOTSUPP;
+}
+static inline int perf_bpf_filter__unpin(void)
+{
+	return -EOPNOTSUPP;
+}
 #endif /* HAVE_BPF_SKEL*/
 #endif /* PERF_UTIL_BPF_FILTER_H */
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#5 lore raw

PATCH 5/6 perf record: Fix a potential error handling issue

In reply to Namhyung Kim

The evlist is allocated at the beginning of cmd_record(). Also free-ing thread masks should be paired with record__init_thread_masks() which is called right before __cmd_record().

Let's change the order of these functions to release the resources correctly in case of errors. This is maybe fine as the process exits, but it might be a problem if it manages some system-wide resources that live longer than the process.

---
 tools/perf/builtin-record.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
tools/perf/builtin-record.c +3 -3
index 8ec0b1607603..3a5a24dec356 100644
@@ -4258,13 +4258,13 @@ int cmd_record(int argc, const char **argv)
 
 	err = __cmd_record(&record, argc, argv);
 out:
-	evlist__delete(rec->evlist);
+	record__free_thread_masks(rec, rec->nr_threads);
+	rec->nr_threads = 0;
 	symbol__exit();
 	auxtrace_record__free(rec->itr);
 out_opts:
-	record__free_thread_masks(rec, rec->nr_threads);
-	rec->nr_threads = 0;
 	evlist__close_control(rec->opts.ctl_fd, rec->opts.ctl_fd_ack, &rec->opts.ctl_fd_close);
+	evlist__delete(rec->evlist);
 	return err;
 }
 
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog
author maintainer
#6 lore raw

PATCH 6/6 perf record: Add --setup-filter option

In reply to Namhyung Kim

To allow BPF filters for unprivileged users it needs to pin the BPF objects to BPF-fs first. Let's add a new option to pin and unpin the objects easily.

  $ sudo perf record --setup-filter pin

The above command would pin BPF program and maps for the filter when the system has BPF-fs (usually at /sys/fs/bpf/). To unpin the objects, users can run the following command (as root).

  $ sudo perf record --setup-filter unpin
---
 tools/perf/Documentation/perf-record.txt |  5 +++++
 tools/perf/builtin-record.c              | 15 +++++++++++++++
 2 files changed, 20 insertions(+)
tools/perf/Documentation/perf-record.txt +5 -0
index 6015fdd08fb6..e51a492dc8e0 100644
@@ -828,6 +828,11 @@ filtered through the mask provided by -C option.
 	only, as of now.  So the applications built without the frame
 	pointer might see bogus addresses.
 
+--setup-filter=<action>::
+	Prepare BPF filter to be used by regular users.  The action should be
+	either "pin" or "unpin".  The filter can be used after it's pinned.
+
+
 include::intel-hybrid.txt[]
 
 SEE ALSO
tools/perf/builtin-record.c +15 -0
index 3a5a24dec356..4dababd0d338 100644
@@ -171,6 +171,7 @@ struct record {
 	bool			timestamp_filename;
 	bool			timestamp_boundary;
 	bool			off_cpu;
+	const char		*filter_action;
 	struct switch_output	switch_output;
 	unsigned long long	samples;
 	unsigned long		output_max_size;	/* = 0: unlimited */
@@ -3573,6 +3574,8 @@ static struct option __record_options[] = {
 			    "write collected trace data into several data files using parallel threads",
 			    record__parse_threads),
 	OPT_BOOLEAN(0, "off-cpu", &record.off_cpu, "Enable off-cpu analysis"),
+	OPT_STRING(0, "setup-filter", &record.filter_action, "pin|unpin",
+		   "BPF filter action"),
 	OPT_END()
 };
 
@@ -4102,6 +4105,18 @@ int cmd_record(int argc, const char **argv)
 		pr_warning("WARNING: --timestamp-filename option is not available in parallel streaming mode.\n");
 	}
 
+	if (rec->filter_action) {
+		if (!strcmp(rec->filter_action, "pin"))
+			err = perf_bpf_filter__pin();
+		else if (!strcmp(rec->filter_action, "unpin"))
+			err = perf_bpf_filter__unpin();
+		else {
+			pr_warning("Unknown BPF filter action: %s\n", rec->filter_action);
+			err = -EINVAL;
+		}
+		goto out_opts;
+	}
+
 	/*
 	 * Allow aliases to facilitate the lookup of symbols for address
 	 * filters. Refer to auxtrace_parse_filters().
Signature · 1 line
2.45.1.288.g0e0cd299f1-goog