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 <common/pipe.hpp>
  11 #include <common/utils.hpp>
  12
  13 #include "manage-kernel.hpp"
  14 #include "testpoint.hpp"
  15 #include "health-sessiond.hpp"
  16 #include "utils.hpp"
  17 #include "thread.hpp"
  18 #include "kernel.hpp"
  19 #include "kernel-consumer.hpp"
  20
  21 namespace {
  22 struct thread_notifiers {
  23         struct lttng_pipe *quit_pipe;
  24         int kernel_poll_pipe_read_fd;
  25 };
  26 } /* namespace */
  27
  28 /*
  29  * Update the kernel poll set of all channel fd available over all tracing
  30  * session. Add the wakeup pipe at the end of the set.
  31  */
  32 static int update_kernel_poll(struct lttng_poll_event *events)
  33 {
  34         int ret;
  35         struct ltt_kernel_channel *channel;
  36         struct ltt_session *session;
  37         const struct ltt_session_list *session_list = session_get_list();
  38
  39         DBG("Updating kernel poll set");
  40
  41         session_lock_list();
  42         cds_list_for_each_entry(session, &session_list->head, list) {
  43                 if (!session_get(session)) {
  44                         continue;
  45                 }
  46                 session_lock(session);
  47                 if (session->kernel_session == NULL) {
  48                         session_unlock(session);
  49                         session_put(session);
  50                         continue;
  51                 }
  52
  53                 cds_list_for_each_entry(channel,
  54                                 &session->kernel_session->channel_list.head, list) {
  55                         /* Add channel fd to the kernel poll set */
  56                         ret = lttng_poll_add(events, channel->fd, LPOLLIN | LPOLLRDNORM);
  57                         if (ret < 0) {
  58                                 session_unlock(session);
  59                                 session_put(session);
  60                                 goto error;
  61                         }
  62                         DBG("Channel fd %d added to kernel set", channel->fd);
  63                 }
  64                 session_unlock(session);
  65                 session_put(session);
  66         }
  67         session_unlock_list();
  68
  69         return 0;
  70
  71 error:
  72         session_unlock_list();
  73         return -1;
  74 }
  75
  76 /*
  77  * Find the channel fd from 'fd' over all tracing session. When found, check
  78  * for new channel stream and send those stream fds to the kernel consumer.
  79  *
  80  * Useful for CPU hotplug feature.
  81  */
  82 static int update_kernel_stream(int fd)
  83 {
  84         int ret = 0;
  85         struct ltt_session *session;
  86         struct ltt_kernel_session *ksess;
  87         struct ltt_kernel_channel *channel;
  88         const struct ltt_session_list *session_list = session_get_list();
  89
  90         DBG("Updating kernel streams for channel fd %d", fd);
  91
  92         session_lock_list();
  93         cds_list_for_each_entry(session, &session_list->head, list) {
  94                 if (!session_get(session)) {
  95                         continue;
  96                 }
  97                 session_lock(session);
  98                 if (session->kernel_session == NULL) {
  99                         session_unlock(session);
 100                         session_put(session);
 101                         continue;
 102                 }
 103                 ksess = session->kernel_session;
 104
 105                 cds_list_for_each_entry(channel,
 106                                 &ksess->channel_list.head, list) {
 107                         struct lttng_ht_iter iter;
 108                         struct consumer_socket *socket;
 109
 110                         if (channel->fd != fd) {
 111                                 continue;
 112                         }
 113                         DBG("Channel found, updating kernel streams");
 114                         ret = kernel_open_channel_stream(channel);
 115                         if (ret < 0) {
 116                                 goto error;
 117                         }
 118                         /* Update the stream global counter */
 119                         ksess->stream_count_global += ret;
 120
 121                         /*
 122                          * Have we already sent fds to the consumer? If yes, it
 123                          * means that tracing is started so it is safe to send
 124                          * our updated stream fds.
 125                          */
 126                         if (ksess->consumer_fds_sent != 1
 127                                         || ksess->consumer == NULL) {
 128                                 ret = -1;
 129                                 goto error;
 130                         }
 131
 132                         rcu_read_lock();
 133                         cds_lfht_for_each_entry(ksess->consumer->socks->ht,
 134                                         &iter.iter, socket, node.node) {
 135                                 pthread_mutex_lock(socket->lock);
 136                                 ret = kernel_consumer_send_channel_streams(socket,
 137                                                 channel, ksess,
 138                                                 session->output_traces ? 1 : 0);
 139                                 pthread_mutex_unlock(socket->lock);
 140                                 if (ret < 0) {
 141                                         rcu_read_unlock();
 142                                         goto error;
 143                                 }
 144                         }
 145                         rcu_read_unlock();
 146                 }
 147                 session_unlock(session);
 148                 session_put(session);
 149         }
 150         session_unlock_list();
 151         return ret;
 152
 153 error:
 154         session_unlock(session);
 155         session_put(session);
 156         session_unlock_list();
 157         return ret;
 158 }
 159
 160 /*
 161  * This thread manage event coming from the kernel.
 162  *
 163  * Features supported in this thread:
 164  *    -) CPU Hotplug
 165  */
 166 static void *thread_kernel_management(void *data)
 167 {
 168         int ret, i, pollfd, update_poll_flag = 1, err = -1;
 169         uint32_t revents, nb_fd;
 170         char tmp;
 171         struct lttng_poll_event events;
 172         struct thread_notifiers *notifiers = (thread_notifiers *) data;
 173         const int quit_pipe_read_fd = lttng_pipe_get_readfd(notifiers->quit_pipe);
 174
 175         DBG("[thread] Thread manage kernel started");
 176
 177         health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_KERNEL);
 178
 179         /*
 180          * This first step of the while is to clean this structure which could free
 181          * non NULL pointers so initialize it before the loop.
 182          */
 183         lttng_poll_init(&events);
 184
 185         if (testpoint(sessiond_thread_manage_kernel)) {
 186                 goto error_testpoint;
 187         }
 188
 189         health_code_update();
 190
 191         if (testpoint(sessiond_thread_manage_kernel_before_loop)) {
 192                 goto error_testpoint;
 193         }
 194
 195         while (1) {
 196                 health_code_update();
 197
 198                 if (update_poll_flag == 1) {
 199                         /* Clean events object. We are about to populate it again. */
 200                         lttng_poll_clean(&events);
 201
 202                         ret = lttng_poll_create(&events, 2, LTTNG_CLOEXEC);
 203                         if (ret < 0) {
 204                                 goto error_poll_create;
 205                         }
 206
 207                         ret = lttng_poll_add(&events,
 208                                         notifiers->kernel_poll_pipe_read_fd,
 209                                         LPOLLIN);
 210                         if (ret < 0) {
 211                                 goto error;
 212                         }
 213
 214                         ret = lttng_poll_add(&events,
 215                                         quit_pipe_read_fd,
 216                                         LPOLLIN);
 217                         if (ret < 0) {
 218                                 goto error;
 219                         }
 220
 221                         /* This will add the available kernel channel if any. */
 222                         ret = update_kernel_poll(&events);
 223                         if (ret < 0) {
 224                                 goto error;
 225                         }
 226                         update_poll_flag = 0;
 227                 }
 228
 229                 DBG("Thread kernel polling");
 230
 231                 /* Poll infinite value of time */
 232         restart:
 233                 health_poll_entry();
 234                 ret = lttng_poll_wait(&events, -1);
 235                 DBG("Thread kernel return from poll on %d fds",
 236                                 LTTNG_POLL_GETNB(&events));
 237                 health_poll_exit();
 238                 if (ret < 0) {
 239                         /*
 240                          * Restart interrupted system call.
 241                          */
 242                         if (errno == EINTR) {
 243                                 goto restart;
 244                         }
 245                         goto error;
 246                 } else if (ret == 0) {
 247                         /* Should not happen since timeout is infinite */
 248                         ERR("Return value of poll is 0 with an infinite timeout.\n"
 249                                 "This should not have happened! Continuing...");
 250                         continue;
 251                 }
 252
 253                 nb_fd = ret;
 254
 255                 for (i = 0; i < nb_fd; i++) {
 256                         /* Fetch once the poll data */
 257                         revents = LTTNG_POLL_GETEV(&events, i);
 258                         pollfd = LTTNG_POLL_GETFD(&events, i);
 259
 260                         health_code_update();
 261
 262                         if (pollfd == quit_pipe_read_fd) {
 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(notifiers->kernel_poll_pipe_read_fd,
 271                                                 &tmp, 1);
 272                                         /*
 273                                          * Ret value is useless here, if this pipe gets any actions an
 274                                          * 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 NULL;
 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 = NULL;
 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 }