perf_reader: Change perf_reader_event_read to only read a maximum of

one ring buffer worth of data.

Signed-off-by: James Bartlett <[email protected]>

Author: JamesMBartlett

src/cc/perf_reader.c

@@ -172,8 +172,8 @@ void perf_reader_event_read(struct perf_reader *reader) {
   // Consume all the events on this ring, calling the cb function for each one.
   // The message may fall on the ring boundary, in which case copy the message
   // into a malloced buffer.
-  for (data_head = read_data_head(perf_header); perf_header->data_tail != data_head;
-      data_head = read_data_head(perf_header)) {
+  data_head = read_data_head(perf_header);
+  while (perf_header->data_tail != data_head) {
     uint64_t data_tail = perf_header->data_tail;
     uint8_t *ptr;
 

tests/cc/CMakeLists.txt

@@ -36,6 +36,7 @@ set(TEST_LIBBCC_SOURCES
 	test_sock_table.cc
 	test_usdt_args.cc
 	test_usdt_probes.cc
+	test_perf_buffer_poll_full_ring_buf.cc
 	utils.cc
 	test_parse_tracepoint.cc)
 
@@ -46,7 +47,7 @@ if(NOT CMAKE_USE_LIBBPF_PACKAGE)
   add_executable(test_libbcc ${TEST_LIBBCC_SOURCES})
   add_dependencies(test_libbcc bcc-shared)
 
-  target_link_libraries(test_libbcc ${PROJECT_BINARY_DIR}/src/cc/libbcc.so dl usdt_test_lib)
+  target_link_libraries(test_libbcc ${PROJECT_BINARY_DIR}/src/cc/libbcc.so dl usdt_test_lib pthread)
   set_target_properties(test_libbcc PROPERTIES INSTALL_RPATH ${PROJECT_BINARY_DIR}/src/cc)
   target_compile_definitions(test_libbcc PRIVATE -DLIBBCC_NAME=\"libbcc.so\")
 
@@ -57,7 +58,7 @@ if(LIBBPF_FOUND)
   add_executable(test_libbcc_no_libbpf ${TEST_LIBBCC_SOURCES})
   add_dependencies(test_libbcc_no_libbpf bcc-shared)
 
-  target_link_libraries(test_libbcc_no_libbpf ${PROJECT_BINARY_DIR}/src/cc/libbcc.so dl usdt_test_lib ${LIBBPF_LIBRARIES})
+  target_link_libraries(test_libbcc_no_libbpf ${PROJECT_BINARY_DIR}/src/cc/libbcc.so dl usdt_test_lib ${LIBBPF_LIBRARIES} pthread)
   set_target_properties(test_libbcc_no_libbpf PROPERTIES INSTALL_RPATH ${PROJECT_BINARY_DIR}/src/cc)
   target_compile_definitions(test_libbcc_no_libbpf PRIVATE -DLIBBCC_NAME=\"libbcc.so\")
 

tests/cc/test_perf_buffer_poll_full_ring_buf.cc

@@ -0,0 +1,161 @@
+#include <linux/perf_event.h>
+#include <linux/version.h>
+#include <unistd.h>
+
+#include <atomic>
+#include <chrono>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <thread>
+
+#include "BPF.h"
+#include "catch.hpp"
+
+struct cb_data {
+  uint64_t data_received;
+  std::atomic<size_t> sleep_ms;
+};
+
+void handle_data_fn(void* cb_cookie, void*, int data_size) {
+  struct cb_data* cookie_data = static_cast<struct cb_data*>(cb_cookie);
+  cookie_data->data_received += data_size;
+  // Force the handler to take a little while so that the ring buffer consumer
+  // is slower than the producer.
+  std::this_thread::sleep_for(
+      std::chrono::milliseconds(cookie_data->sleep_ms.load()));
+}
+void handle_data_loss_fn(void*, unsigned long) {}
+
+class RAIIThreadCloser {
+ public:
+  RAIIThreadCloser(std::atomic<bool>* done, std::thread* thread)
+      : done_(done), thread_(thread) {}
+  ~RAIIThreadCloser() {
+    done_->store(true);
+    thread_->join();
+  }
+
+ private:
+  std::atomic<bool>* done_;
+  std::thread* thread_;
+};
+
+/**
+ * This test demonstrates a bug in perf_reader where perf_reader_event_read can
+ * loop over the ring buffer more than once in a single call, if the consumer of
+ * the event data (i.e. raw_cb) is slower than the producer (the kernel pushing
+ * events from ebpf). To demonstrate this bug we have a thread that countinually
+ * writes to /dev/null, then we deploy a BPF program that looks for writes from
+ * this PID and for each write it will submit 30kB to the perf buffer. Then we
+ * artificially slow down the perf buffer data callback so that its slower than
+ * the kernel producing data. If we removed the timeout in the test the
+ * perf_reader->poll() call could potentially run indefinitely (depending on the
+ * num_pages and sleep_ms it might not run forever but if sleep_ms is large
+ * enough it will). Instead we set a timeout and check that the amount of data
+ * read from a single `poll()` call is less than the size of the kernel ring
+ * buffer.
+ */
+TEST_CASE("test perf buffer poll full ring buf", "[bpf_perf_event]") {
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(4, 13, 0)
+  std::atomic<bool> write_done(false);
+  // This thread writes to /dev/null continuouosly so it should trigger many
+  // write syscalls which will fill up the perf ring buf.
+  std::thread write_thread([&]() {
+    std::ofstream out("/dev/null");
+    while (!write_done.load()) {
+      out << "test";
+      out.flush();
+    }
+  });
+  RAIIThreadCloser closer(&write_done, &write_thread);
+
+#define MSG_SIZE (32 * 1024)
+  const std::string BPF_PROGRAM_FORMAT_STR = R"(
+    #include <uapi/linux/ptrace.h>
+    #define MSG_SIZE %d
+    struct event_t {
+      char msg[MSG_SIZE];
+    };
+    BPF_PERF_OUTPUT(events);
+    BPF_PERCPU_ARRAY(events_heap, struct event_t, 1);
+
+
+    // Probe that submits a 32kB event every time write is called.
+    int syscall__probe_entry_write(struct pt_regs* ctx, int fd, char* buf, size_t count) {
+      uint32_t kZero = 0;
+      struct event_t* event = events_heap.lookup(&kZero);
+      if (event == NULL) {
+        return 0;
+      }
+      uint32_t tgid = bpf_get_current_pid_tgid() >> 32;
+      if (tgid != %lu) {
+        return 0;
+      }
+      events.perf_submit(ctx, event, sizeof(struct event_t));
+      return 0;
+    }
+  )";
+  unsigned long pid = getpid();
+
+  // Use the BPF program as a format string to insert the current test
+  // processes' PID so that we only submit events generated by the current
+  // process.
+  auto extra_length = snprintf(nullptr, 0, "%d %lu", MSG_SIZE, pid);
+  char bpf_program[BPF_PROGRAM_FORMAT_STR.size() + extra_length];
+  auto n = snprintf(bpf_program, sizeof(bpf_program),
+                    BPF_PROGRAM_FORMAT_STR.c_str(), MSG_SIZE, pid);
+  const std::string BPF_PROGRAM(bpf_program, n);
+
+  auto num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
+  auto page_size = sysconf(_SC_PAGE_SIZE);
+
+  ebpf::BPF bpf;
+  ebpf::StatusTuple res(0);
+  res = bpf.init(BPF_PROGRAM, {"-DNUM_CPUS=" + std::to_string(num_cpus)}, {});
+  REQUIRE(res.code() == 0);
+
+  std::string write_fnname = bpf.get_syscall_fnname("write");
+  res = bpf.attach_kprobe(write_fnname, "syscall__probe_entry_write");
+  REQUIRE(res.code() == 0);
+
+  struct cb_data cb_cookie = {};
+  cb_cookie.sleep_ms = 200;
+  cb_cookie.data_received = 0;
+
+  int num_pages = 64;
+  std::string perf_buffer_name("events");
+  res = bpf.open_perf_buffer(perf_buffer_name, handle_data_fn,
+                             handle_data_loss_fn, &cb_cookie, num_pages);
+  REQUIRE(res.code() == 0);
+
+  std::atomic<bool> poll_done(false);
+  int cnt = 0;
+  std::thread poll_thread([&]() {
+    auto perf_buffer = bpf.get_perf_buffer(perf_buffer_name);
+    REQUIRE(perf_buffer != nullptr);
+    cnt = perf_buffer->poll(0);
+    poll_done = true;
+  });
+
+  auto start = std::chrono::steady_clock::now();
+  std::chrono::seconds timeout(20);
+  while (!poll_done.load() &&
+         (std::chrono::steady_clock::now() - start) < timeout) {
+    std::this_thread::sleep_for(std::chrono::milliseconds{10});
+  }
+  // After the timeout we set the sleep time to 0, so the reader should catch up
+  // and terminate.
+  cb_cookie.sleep_ms = 0;
+  poll_thread.join();
+
+  res = bpf.close_perf_buffer(perf_buffer_name);
+  REQUIRE(res.code() == 0);
+  res = bpf.detach_kprobe(write_fnname);
+  REQUIRE(res.code() == 0);
+
+  // cnt is the number of perf_readers the poll() call read from. So we should
+  // not have received more data then 1 full ring buffer per perf_reader.
+  REQUIRE(cb_cookie.data_received <= (cnt * num_pages * page_size));
+#endif
+}