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