fix iteration over CPUs

Since kernel version 4.9.0 BPF stopped working in a KVM guest.
The problem are calls to perf_event_open with CPU identifiers which do
not exist (ENODEV). The root cause for this is that the current code
assumes ascending numbered CPUs. However, this is not always the case
(e.g. CPU hotplugging).

This patch introduces the get_online_cpus() and get_possible_cpus()
helper functions and uses the appropriate function for iterations over
CPUs. The BPF_MAP_TYPE_PERF_EVENT_ARRAY map contains now an entry for
each possible CPU instead of for each online CPU.

Fixes: iovisor#893
Signed-off-by: Andreas Gerstmayr <[email protected]>

src/cc/BPF.cc

@@ -30,6 +30,7 @@
 #include "bpf_module.h"
 #include "libbpf.h"
 #include "perf_reader.h"
+#include "common.h"
 #include "usdt.h"
 
 #include "BPF.h"
@@ -297,11 +298,12 @@ StatusTuple BPF::attach_perf_event(uint32_t ev_type, uint32_t ev_config,
  TRY2(load_func(probe_func, BPF_PROG_TYPE_PERF_EVENT, probe_fd));
 
  auto fds = new std::map<int, int>();
- int cpu_st = 0;
- int cpu_en = sysconf(_SC_NPROCESSORS_ONLN) - 1;
+ std::vector<int> cpus;
  if (cpu >= 0)
- cpu_st = cpu_en = cpu;
- for (int i = cpu_st; i <= cpu_en; i++) {
+ cpus.push_back(cpu);
+ else
+ cpus = get_online_cpus();
+ for (int i: cpus) {
  int fd = bpf_attach_perf_event(probe_fd, ev_type, ev_config, sample_period,
  sample_freq, pid, i, group_fd);
  if (fd < 0) {

src/cc/BPFTable.cc

@@ -25,6 +25,7 @@
 #include "bcc_syms.h"
 #include "libbpf.h"
 #include "perf_reader.h"
+#include "common.h"
 
 namespace ebpf {
 
@@ -89,7 +90,7 @@ StatusTuple BPFPerfBuffer::open_all_cpu(perf_reader_raw_cb cb,
  if (cpu_readers_.size() != 0 || readers_.size() != 0)
  return StatusTuple(-1, "Previously opened perf buffer not cleaned");
 
- for (int i = 0; i < sysconf(_SC_NPROCESSORS_ONLN); i++) {
+ for (int i: get_online_cpus()) {
  auto res = open_on_cpu(cb, i, cb_cookie);
  if (res.code() != 0) {
  TRY2(close_all_cpu());
@@ -113,7 +114,7 @@ StatusTuple BPFPerfBuffer::close_on_cpu(int cpu) {
 StatusTuple BPFPerfBuffer::close_all_cpu() {
  std::string errors;
  bool has_error = false;
- for (int i = 0; i < sysconf(_SC_NPROCESSORS_ONLN); i++) {
+ for (int i: get_online_cpus()) {
  auto res = close_on_cpu(i);
  if (res.code() != 0) {
  errors += "Failed to close CPU" + std::to_string(i) + " perf buffer: ";

src/cc/CMakeLists.txt

@@ -35,12 +35,12 @@ if (CMAKE_COMPILER_IS_GNUCC AND LIBCLANG_ISSTATIC)
  endif()
 endif()
 
-add_library(bcc-shared SHARED bpf_common.cc bpf_module.cc libbpf.c perf_reader.c shared_table.cc exported_files.cc bcc_elf.c bcc_perf_map.c bcc_proc.c bcc_syms.cc usdt_args.cc usdt.cc BPF.cc BPFTable.cc)
+add_library(bcc-shared SHARED bpf_common.cc bpf_module.cc libbpf.c perf_reader.c shared_table.cc exported_files.cc bcc_elf.c bcc_perf_map.c bcc_proc.c bcc_syms.cc usdt_args.cc usdt.cc common.cc BPF.cc BPFTable.cc)
 set_target_properties(bcc-shared PROPERTIES VERSION ${REVISION_LAST} SOVERSION 0)
 set_target_properties(bcc-shared PROPERTIES OUTPUT_NAME bcc)
 
 add_library(bcc-loader-static libbpf.c perf_reader.c bcc_elf.c bcc_perf_map.c bcc_proc.c)
-add_library(bcc-static STATIC bpf_common.cc bpf_module.cc shared_table.cc exported_files.cc bcc_syms.cc usdt_args.cc usdt.cc BPF.cc BPFTable.cc)
+add_library(bcc-static STATIC bpf_common.cc bpf_module.cc shared_table.cc exported_files.cc bcc_syms.cc usdt_args.cc usdt.cc common.cc BPF.cc BPFTable.cc)
 set_target_properties(bcc-static PROPERTIES OUTPUT_NAME bcc)
 
 set(llvm_raw_libs bitwriter bpfcodegen irreader linker

src/cc/common.cc

@@ -0,0 +1,51 @@
+/*
+ * Copyright (c) 2016 Catalysts GmbH
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http:https://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <fstream>
+#include <sstream>
+#include "common.h"
+
+namespace ebpf {
+
+std::vector<int> read_cpu_range(std::string path) {
+ std::ifstream cpus_range_stream { path };
+ std::vector<int> cpus;
+ std::string cpu_range;
+
+ while (std::getline(cpus_range_stream, cpu_range, ',')) {
+ std::size_t rangeop = cpu_range.find('-');
+ if (rangeop == std::string::npos) {
+ cpus.push_back(std::stoi(cpu_range));
+ }
+ else {
+ int start = std::stoi(cpu_range.substr(0, rangeop));
+ int end = std::stoi(cpu_range.substr(rangeop + 1));
+ for (int i = start; i <= end; i++)
+ cpus.push_back(i);
+ }
+ }
+ return cpus;
+}
+
+std::vector<int> get_online_cpus() {
+ return read_cpu_range("/sys/devices/system/cpu/online");
+}
+
+std::vector<int> get_possible_cpus() {
+ return read_cpu_range("/sys/devices/system/cpu/possible");
+}
+
+
+} // namespace ebpf

src/cc/common.h

@@ -19,6 +19,7 @@
 #include <memory>
 #include <string>
 #include <tuple>
+#include <vector>
 
 namespace ebpf {
 
@@ -28,4 +29,8 @@ make_unique(Args &&... args) {
  return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
 }
 
+std::vector<int> get_online_cpus();
+
+std::vector<int> get_possible_cpus();
+
 } // namespace ebpf

src/cc/frontends/clang/CMakeLists.txt

@@ -2,4 +2,4 @@
 # Licensed under the Apache License, Version 2.0 (the "License")
 
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DKERNEL_MODULES_DIR='\"${BCC_KERNEL_MODULES_DIR}\"'")
-add_library(clang_frontend loader.cc b_frontend_action.cc tp_frontend_action.cc kbuild_helper.cc)
+add_library(clang_frontend loader.cc b_frontend_action.cc tp_frontend_action.cc kbuild_helper.cc ../../common.cc)

src/cc/frontends/clang/b_frontend_action.cc

@@ -27,6 +27,7 @@
 
 #include "b_frontend_action.h"
 #include "shared_table.h"
+#include "common.h"
 
 #include "libbpf.h"
 
@@ -656,7 +657,7 @@ bool BTypeVisitor::VisitVarDecl(VarDecl *Decl) {
  map_type = BPF_MAP_TYPE_PROG_ARRAY;
  } else if (A->getName() == "maps/perf_output") {
  map_type = BPF_MAP_TYPE_PERF_EVENT_ARRAY;
- int numcpu = sysconf(_SC_NPROCESSORS_ONLN);
+ int numcpu = get_possible_cpus().size();
  if (numcpu <= 0)
  numcpu = 1;
  table.max_entries = numcpu;

src/python/bcc/__init__.py

@@ -17,7 +17,6 @@
 import ctypes as ct
 import fcntl
 import json
-import multiprocessing
 import os
 import re
 import struct
@@ -29,6 +28,7 @@
 from .table import Table
 from .perf import Perf
 from .usyms import ProcessSymbols
+from .utils import get_online_cpus
 
 _kprobe_limit = 1000
 _num_open_probes = 0
@@ -660,7 +660,7 @@ def attach_perf_event(self, ev_type=-1, ev_config=-1, fn_name="",
  res[cpu] = self._attach_perf_event(fn.fd, ev_type, ev_config,
  sample_period, sample_freq, pid, cpu, group_fd)
  else:
- for i in range(0, multiprocessing.cpu_count()):
+ for i in get_online_cpus():
  res[i] = self._attach_perf_event(fn.fd, ev_type, ev_config,
  sample_period, sample_freq, pid, i, group_fd)
  self.open_perf_events[(ev_type, ev_config)] = res

src/python/bcc/perf.py

@@ -13,8 +13,8 @@
 # limitations under the License.
 
 import ctypes as ct
-import multiprocessing
 import os
+from .utils import get_online_cpus
 
 class Perf(object):
  class perf_event_attr(ct.Structure):
@@ -105,5 +105,5 @@ def perf_event_open(tpoint_id, pid=-1, ptype=PERF_TYPE_TRACEPOINT,
  attr.sample_period = 1
  attr.wakeup_events = 9999999 # don't wake up
 
- for cpu in range(0, multiprocessing.cpu_count()):
+ for cpu in get_online_cpus():
  Perf._open_for_cpu(cpu, attr)

src/python/bcc/table.py

@@ -19,6 +19,7 @@
 
 from .libbcc import lib, _RAW_CB_TYPE
 from .perf import Perf
+from .utils import get_online_cpus
 from subprocess import check_output
 
 BPF_MAP_TYPE_HASH = 1
@@ -509,7 +510,7 @@ def open_perf_buffer(self, callback):
  event submitted from the kernel, up to millions per second.
  """
 
- for i in range(0, multiprocessing.cpu_count()):
+ for i in get_online_cpus():
  self._open_perf_buffer(i, callback)
 
  def _open_perf_buffer(self, cpu, callback):
@@ -550,7 +551,7 @@ def open_perf_event(self, ev):
  if not isinstance(ev, self.Event):
  raise Exception("argument must be an Event, got %s", type(ev))
 
- for i in range(0, multiprocessing.cpu_count()):
+ for i in get_online_cpus():
  self._open_perf_event(i, ev.typ, ev.config)
 
 

src/python/bcc/utils.py

@@ -0,0 +1,33 @@
+# Copyright 2016 Catalysts GmbH
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http:https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+def _read_cpu_range(path):
+ cpus = []
+ with open(path, 'r') as f:
+ cpus_range_str = f.read()
+ for cpu_range in cpus_range_str.split(','):
+ rangeop = cpu_range.find('-')
+ if rangeop == -1:
+ cpus.append(int(cpu_range))
+ else:
+ start = int(cpu_range[:rangeop])
+ end = int(cpu_range[rangeop+1:])
+ cpus.extend(range(start, end+1))
+ return cpus
+
+def get_online_cpus():
+ return _read_cpu_range('/sys/devices/system/cpu/online')
+
+def get_possible_cpus():
+ return _read_cpu_range('/sys/devices/system/cpu/possible')

tests/cc/test_c_api.cc

@@ -23,6 +23,7 @@
 #include "bcc_perf_map.h"
 #include "bcc_proc.h"
 #include "bcc_syms.h"
+#include "common.h"
 #include "vendor/tinyformat.hpp"
 
 #include "catch.hpp"
@@ -196,3 +197,10 @@ TEST_CASE("resolve symbols using /tmp/perf-pid.map", "[c_api]") {
 
  munmap(map_addr, map_sz);
 }
+
+
+TEST_CASE("get online CPUs", "[c_api]") {
+ std::vector<int> cpus = ebpf::get_online_cpus();
+ int num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+ REQUIRE(cpus.size() == num_cpus);
+}

tests/python/CMakeLists.txt

@@ -56,6 +56,8 @@ add_test(NAME py_test_tracepoint WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
  COMMAND ${TEST_WRAPPER} py_test_tracepoint sudo ${CMAKE_CURRENT_SOURCE_DIR}/test_tracepoint.py)
 add_test(NAME py_test_perf_event WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
  COMMAND ${TEST_WRAPPER} py_test_perf_event sudo ${CMAKE_CURRENT_SOURCE_DIR}/test_perf_event.py)
+add_test(NAME py_test_utils WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+ COMMAND ${TEST_WRAPPER} py_test_utils sudo ${CMAKE_CURRENT_SOURCE_DIR}/test_utils.py)
 
 add_test(NAME py_test_dump_func WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
  COMMAND ${TEST_WRAPPER} py_dump_func simple ${CMAKE_CURRENT_SOURCE_DIR}/test_dump_func.py)

tests/python/test_array.py

@@ -6,6 +6,8 @@
 import ctypes as ct
 import random
 import time
+import subprocess
+from bcc.utils import get_online_cpus
 from unittest import main, TestCase
 
 class TestArray(TestCase):
@@ -62,6 +64,37 @@ def cb(cpu, data, size):
  time.sleep(0.1)
  b.kprobe_poll()
  self.assertGreater(self.counter, 0)
+ b.cleanup()
+
+ def test_perf_buffer_for_each_cpu(self):
+ self.events = []
+
+ class Data(ct.Structure):
+ _fields_ = [("cpu", ct.c_ulonglong)]
+
+ def cb(cpu, data, size):
+ self.assertGreater(size, ct.sizeof(Data))
+ event = ct.cast(data, ct.POINTER(Data)).contents
+ self.events.append(event)
+
+ text = """
+BPF_PERF_OUTPUT(events);
+int kprobe__sys_nanosleep(void *ctx) {
+ struct {
+ u64 cpu;
+ } data = {bpf_get_smp_processor_id()};
+ events.perf_submit(ctx, &data, sizeof(data));
+ return 0;
+}
+"""
+ b = BPF(text=text)
+ b["events"].open_perf_buffer(cb)
+ online_cpus = get_online_cpus()
+ for cpu in online_cpus:
+ subprocess.call(['taskset', '-c', str(cpu), 'sleep', '0.1'])
+ b.kprobe_poll()
+ b.cleanup()
+ self.assertGreaterEqual(len(self.events), len(online_cpus), 'Received only {}/{} events'.format(len(self.events), len(online_cpus)))
 
 if __name__ == "__main__":
  main()

tests/python/test_utils.py

@@ -0,0 +1,18 @@
+#!/usr/bin/python
+# Copyright (c) Catalysts GmbH
+# Licensed under the Apache License, Version 2.0 (the "License")
+
+from bcc.utils import get_online_cpus
+import multiprocessing
+import unittest
+
+class TestUtils(unittest.TestCase):
+ def test_get_online_cpus(self):
+ online_cpus = get_online_cpus()
+ num_cores = multiprocessing.cpu_count()
+
+ self.assertEqual(len(online_cpus), num_cores)
+
+
+if __name__ == "__main__":
+ unittest.main()