src/bin/lttng-sessiond/manage-kernel.cpp

   1 /*
   2  * Copyright (C) 2011 EfficiOS Inc.
   3  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   4  * Copyright (C) 2013 Jérémie Galarneau <jeremie.galarneau@efficios.com>
   5  *
   6  * SPDX-License-Identifier: GPL-2.0-only
   7  *
   8  */
   9
  10 #include "health-sessiond.hpp"
  11 #include "kernel-consumer.hpp"
  12 #include "kernel.hpp"
  13 #include "manage-kernel.hpp"
  14 #include "testpoint.hpp"
  15 #include "thread.hpp"
  16 #include "utils.hpp"
  17
  18 #include <common/pipe.hpp>
  19 #include <common/urcu.hpp>
  20 #include <common/utils.hpp>
  21
  22 #include <fcntl.h>
  23
  24 namespace {
  25 struct thread_notifiers {
  26         struct lttng_pipe *quit_pipe;
  27         int kernel_poll_pipe_read_fd;
  28 };
  29 } /* namespace */
  30
  31 /*
  32  * Update the kernel poll set of all channel fd available over all tracing
  33  * session. Add the wakeup pipe at the end of the set.
  34  */
  35 static int update_kernel_poll(struct lttng_poll_event *events)
  36 {
  37         int ret;
  38         struct ltt_kernel_channel *channel;
  39         struct ltt_session *session;
  40
  41         DBG("Updating kernel poll set");
  42
  43         const auto list_lock = lttng::sessiond::lock_session_list();
  44         const struct ltt_session_list *session_list = session_get_list();
  45
  46         cds_list_for_each_entry (session, &session_list->head, list) {
  47                 if (!session_get(session)) {
  48                         continue;
  49                 }
  50                 session_lock(session);
  51                 if (session->kernel_session == nullptr) {
  52                         session_unlock(session);
  53                         session_put(session);
  54                         continue;
  55                 }
  56
  57                 cds_list_for_each_entry (
  58                         channel, &session->kernel_session->channel_list.head, list) {
  59                         /* Add channel fd to the kernel poll set */
  60                         ret = lttng_poll_add(events, channel->fd, LPOLLIN | LPOLLRDNORM);
  61                         if (ret < 0) {
  62                                 session_unlock(session);
  63                                 session_put(session);
  64                                 return -1;
  65                         }
  66                         DBG("Channel fd %d added to kernel set", channel->fd);
  67                 }
  68                 session_unlock(session);
  69                 session_put(session);
  70         }
  71
  72         return 0;
  73 }
  74
  75 /*
  76  * Find the channel fd from 'fd' over all tracing session. When found, check
  77  * for new channel stream and send those stream fds to the kernel consumer.
  78  *
  79  * Useful for CPU hotplug feature.
  80  */
  81 static int update_kernel_stream(int fd)
  82 {
  83         int ret = 0;
  84         struct ltt_session *session;
  85         struct ltt_kernel_session *ksess;
  86         struct ltt_kernel_channel *channel;
  87
  88         DBG("Updating kernel streams for channel fd %d", fd);
  89
  90         const auto list_lock = lttng::sessiond::lock_session_list();
  91         const struct ltt_session_list *session_list = session_get_list();
  92
  93         cds_list_for_each_entry (session, &session_list->head, list) {
  94                 if (!session_get(session)) {
  95                         continue;
  96                 }
  97
  98                 session_lock(session);
  99                 if (session->kernel_session == nullptr) {
 100                         session_unlock(session);
 101                         session_put(session);
 102                         continue;
 103                 }
 104
 105                 ksess = session->kernel_session;
 106
 107                 cds_list_for_each_entry (channel, &ksess->channel_list.head, list) {
 108                         struct lttng_ht_iter iter;
 109                         struct consumer_socket *socket;
 110
 111                         if (channel->fd != fd) {
 112                                 continue;
 113                         }
 114                         DBG("Channel found, updating kernel streams");
 115                         ret = kernel_open_channel_stream(channel);
 116                         if (ret < 0) {
 117                                 goto error;
 118                         }
 119                         /* Update the stream global counter */
 120                         ksess->stream_count_global += ret;
 121
 122                         /*
 123                          * Have we already sent fds to the consumer? If yes, it
 124                          * means that tracing is started so it is safe to send
 125                          * our updated stream fds.
 126                          */
 127                         if (ksess->consumer_fds_sent != 1 || ksess->consumer == nullptr) {
 128                                 ret = -1;
 129                                 goto error;
 130                         }
 131
 132                         {
 133                                 const lttng::urcu::read_lock_guard read_lock;
 134
 135                                 cds_lfht_for_each_entry (
 136                                         ksess->consumer->socks->ht, &iter.iter, socket, node.node) {
 137                                         pthread_mutex_lock(socket->lock);
 138                                         ret = kernel_consumer_send_channel_streams(
 139                                                 socket,
 140                                                 channel,
 141                                                 ksess,
 142                                                 session->output_traces ? 1 : 0);
 143                                         pthread_mutex_unlock(socket->lock);
 144                                         if (ret < 0) {
 145                                                 goto error;
 146                                         }
 147                                 }
 148                         }
 149                 }
 150
 151                 session_unlock(session);
 152                 session_put(session);
 153         }
 154
 155         return ret;
 156
 157 error:
 158         session_unlock(session);
 159         session_put(session);
 160         return ret;
 161 }
 162
 163 /*
 164  * This thread manage event coming from the kernel.
 165  *
 166  * Features supported in this thread:
 167  *    -) CPU Hotplug
 168  */
 169 static void *thread_kernel_management(void *data)
 170 {
 171         int ret, i, update_poll_flag = 1, err = -1;
 172         uint32_t nb_fd;
 173         char tmp;
 174         struct lttng_poll_event events;
 175         struct thread_notifiers *notifiers = (thread_notifiers *) data;
 176         const auto thread_quit_pipe_fd = lttng_pipe_get_readfd(notifiers->quit_pipe);
 177
 178         DBG("[thread] Thread manage kernel started");
 179
 180         health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_KERNEL);
 181
 182         /*
 183          * This first step of the while is to clean this structure which could free
 184          * non NULL pointers so initialize it before the loop.
 185          */
 186         lttng_poll_init(&events);
 187
 188         if (testpoint(sessiond_thread_manage_kernel)) {
 189                 goto error_testpoint;
 190         }
 191
 192         health_code_update();
 193
 194         if (testpoint(sessiond_thread_manage_kernel_before_loop)) {
 195                 goto error_testpoint;
 196         }
 197
 198         while (true) {
 199                 health_code_update();
 200
 201                 if (update_poll_flag == 1) {
 202                         /* Clean events object. We are about to populate it again. */
 203                         lttng_poll_clean(&events);
 204
 205                         ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
 206                         if (ret < 0) {
 207                                 goto error_poll_create;
 208                         }
 209
 210                         ret = lttng_poll_add(&events, notifiers->kernel_poll_pipe_read_fd, LPOLLIN);
 211                         if (ret < 0) {
 212                                 goto error;
 213                         }
 214
 215                         ret = lttng_poll_add(&events, thread_quit_pipe_fd, LPOLLIN);
 216                         if (ret < 0) {
 217                                 goto error;
 218                         }
 219
 220                         /* This will add the available kernel channel if any. */
 221                         ret = update_kernel_poll(&events);
 222                         if (ret < 0) {
 223                                 goto error;
 224                         }
 225                         update_poll_flag = 0;
 226                 }
 227
 228                 DBG("Thread kernel polling");
 229
 230                 /* Poll infinite value of time */
 231         restart:
 232                 health_poll_entry();
 233                 ret = lttng_poll_wait(&events, -1);
 234                 DBG("Thread kernel return from poll on %d fds", LTTNG_POLL_GETNB(&events));
 235                 health_poll_exit();
 236                 if (ret < 0) {
 237                         /*
 238                          * Restart interrupted system call.
 239                          */
 240                         if (errno == EINTR) {
 241                                 goto restart;
 242                         }
 243                         goto error;
 244                 } else if (ret == 0) {
 245                         /* Should not happen since timeout is infinite */
 246                         ERR("Return value of poll is 0 with an infinite timeout.\n"
 247                             "This should not have happened! Continuing...");
 248                         continue;
 249                 }
 250
 251                 nb_fd = ret;
 252
 253                 for (i = 0; i < nb_fd; i++) {
 254                         /* Fetch once the poll data */
 255                         const auto revents = LTTNG_POLL_GETEV(&events, i);
 256                         const auto pollfd = LTTNG_POLL_GETFD(&events, i);
 257
 258                         health_code_update();
 259
 260                         /* Activity on thread quit pipe, exiting. */
 261                         if (pollfd == thread_quit_pipe_fd) {
 262                                 DBG("Activity on thread quit pipe");
 263                                 err = 0;
 264                                 goto exit;
 265                         }
 266
 267                         /* Check for data on kernel pipe */
 268                         if (revents & LPOLLIN) {
 269                                 if (pollfd == notifiers->kernel_poll_pipe_read_fd) {
 270                                         (void) lttng_read(
 271                                                 notifiers->kernel_poll_pipe_read_fd, &tmp, 1);
 272                                         /*
 273                                          * Ret value is useless here, if this pipe gets any actions
 274                                          * an update is required anyway.
 275                                          */
 276                                         update_poll_flag = 1;
 277                                         continue;
 278                                 } else {
 279                                         /*
 280                                          * New CPU detected by the kernel. Adding kernel stream to
 281                                          * kernel session and updating the kernel consumer
 282                                          */
 283                                         ret = update_kernel_stream(pollfd);
 284                                         if (ret < 0) {
 285                                                 continue;
 286                                         }
 287                                         break;
 288                                 }
 289                         } else if (revents & (LPOLLERR | LPOLLHUP | LPOLLRDHUP)) {
 290                                 update_poll_flag = 1;
 291                                 continue;
 292                         } else {
 293                                 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
 294                                 goto error;
 295                         }
 296                 }
 297         }
 298
 299 exit:
 300 error:
 301         lttng_poll_clean(&events);
 302 error_poll_create:
 303 error_testpoint:
 304         if (err) {
 305                 health_error();
 306                 ERR("Health error occurred in %s", __func__);
 307                 WARN("Kernel thread died unexpectedly. "
 308                      "Kernel tracing can continue but CPU hotplug is disabled.");
 309         }
 310         health_unregister(the_health_sessiond);
 311         DBG("Kernel thread dying");
 312         return nullptr;
 313 }
 314
 315 static bool shutdown_kernel_management_thread(void *data)
 316 {
 317         struct thread_notifiers *notifiers = (thread_notifiers *) data;
 318         const int write_fd = lttng_pipe_get_writefd(notifiers->quit_pipe);
 319
 320         return notify_thread_pipe(write_fd) == 1;
 321 }
 322
 323 static void cleanup_kernel_management_thread(void *data)
 324 {
 325         struct thread_notifiers *notifiers = (thread_notifiers *) data;
 326
 327         lttng_pipe_destroy(notifiers->quit_pipe);
 328         free(notifiers);
 329 }
 330
 331 bool launch_kernel_management_thread(int kernel_poll_pipe_read_fd)
 332 {
 333         struct lttng_pipe *quit_pipe;
 334         struct thread_notifiers *notifiers = nullptr;
 335         struct lttng_thread *thread;
 336
 337         notifiers = zmalloc<thread_notifiers>();
 338         if (!notifiers) {
 339                 goto error_alloc;
 340         }
 341         quit_pipe = lttng_pipe_open(FD_CLOEXEC);
 342         if (!quit_pipe) {
 343                 goto error;
 344         }
 345         notifiers->quit_pipe = quit_pipe;
 346         notifiers->kernel_poll_pipe_read_fd = kernel_poll_pipe_read_fd;
 347
 348         thread = lttng_thread_create("Kernel management",
 349                                      thread_kernel_management,
 350                                      shutdown_kernel_management_thread,
 351                                      cleanup_kernel_management_thread,
 352                                      notifiers);
 353         if (!thread) {
 354                 goto error;
 355         }
 356         lttng_thread_put(thread);
 357         return true;
 358 error:
 359         cleanup_kernel_management_thread(notifiers);
 360 error_alloc:
 361         return false;
 362 }