liblttng-ust/lttng-ust-comm.c

   1 /*
   2  * lttng-ust-comm.c
   3  *
   4  * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
   5  * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   6  *
   7  * This library is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU Lesser General Public
   9  * License as published by the Free Software Foundation; only
  10  * version 2.1 of the License.
  11  *
  12  * This library is distributed in the hope that it will be useful,
  13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15  * Lesser General Public License for more details.
  16  *
  17  * You should have received a copy of the GNU Lesser General Public
  18  * License along with this library; if not, write to the Free Software
  19  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  20  */
  21
  22 #define _LGPL_SOURCE
  23 #define _GNU_SOURCE
  24 #include <sys/types.h>
  25 #include <sys/socket.h>
  26 #include <sys/mman.h>
  27 #include <sys/stat.h>
  28 #include <sys/types.h>
  29 #include <sys/wait.h>
  30 #include <fcntl.h>
  31 #include <unistd.h>
  32 #include <errno.h>
  33 #include <pthread.h>
  34 #include <semaphore.h>
  35 #include <time.h>
  36 #include <assert.h>
  37 #include <signal.h>
  38 #include <urcu/uatomic.h>
  39 #include <urcu/futex.h>
  40 #include <urcu/compiler.h>
  41
  42 #include <lttng/ust-events.h>
  43 #include <lttng/ust-abi.h>
  44 #include <lttng/ust.h>
  45 #include <lttng/ust-error.h>
  46 #include <lttng/ust-ctl.h>
  47 #include <urcu/tls-compat.h>
  48 #include <ust-comm.h>
  49 #include <ust-fd.h>
  50 #include <usterr-signal-safe.h>
  51 #include <helper.h>
  52 #include "tracepoint-internal.h"
  53 #include "lttng-tracer-core.h"
  54 #include "compat.h"
  55 #include "../libringbuffer/tlsfixup.h"
  56 #include "lttng-ust-statedump.h"
  57 #include "clock.h"
  58 #include "../libringbuffer/getcpu.h"
  59 #include "getenv.h"
  60
  61 /*
  62  * Has lttng ust comm constructor been called ?
  63  */
  64 static int initialized;
  65
  66 /*
  67  * The ust_lock/ust_unlock lock is used as a communication thread mutex.
  68  * Held when handling a command, also held by fork() to deal with
  69  * removal of threads, and by exit path.
  70  *
  71  * The UST lock is the centralized mutex across UST tracing control and
  72  * probe registration.
  73  *
  74  * ust_exit_mutex must never nest in ust_mutex.
  75  *
  76  * ust_fork_mutex must never nest in ust_mutex.
  77  *
  78  * ust_mutex_nest is a per-thread nesting counter, allowing the perf
  79  * counter lazy initialization called by events within the statedump,
  80  * which traces while the ust_mutex is held.
  81  *
  82  * ust_lock nests within the dynamic loader lock (within glibc) because
  83  * it is taken within the library constructor.
  84  */
  85 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
  86
  87 /* Allow nesting the ust_mutex within the same thread. */
  88 static DEFINE_URCU_TLS(int, ust_mutex_nest);
  89
  90 /*
  91  * ust_exit_mutex protects thread_active variable wrt thread exit. It
  92  * cannot be done by ust_mutex because pthread_cancel(), which takes an
  93  * internal libc lock, cannot nest within ust_mutex.
  94  *
  95  * It never nests within a ust_mutex.
  96  */
  97 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
  98
  99 /*
 100  * ust_fork_mutex protects base address statedump tracing against forks. It
 101  * prevents the dynamic loader lock to be taken (by base address statedump
 102  * tracing) while a fork is happening, thus preventing deadlock issues with
 103  * the dynamic loader lock.
 104  */
 105 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
 106
 107 /* Should the ust comm thread quit ? */
 108 static int lttng_ust_comm_should_quit;
 109
 110 /*
 111  * This variable can be tested by applications to check whether
 112  * lttng-ust is loaded. They simply have to define their own
 113  * "lttng_ust_loaded" weak symbol, and test it. It is set to 1 by the
 114  * library constructor.
 115  */
 116 int lttng_ust_loaded __attribute__((weak));
 117
 118 /*
 119  * Return 0 on success, -1 if should quit.
 120  * The lock is taken in both cases.
 121  * Signal-safe.
 122  */
 123 int ust_lock(void)
 124 {
 125         sigset_t sig_all_blocked, orig_mask;
 126         int ret, oldstate;
 127
 128         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 129         if (ret) {
 130                 ERR("pthread_setcancelstate: %s", strerror(ret));
 131         }
 132         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 133                 ERR("pthread_setcancelstate: unexpected oldstate");
 134         }
 135         sigfillset(&sig_all_blocked);
 136         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 137         if (ret) {
 138                 ERR("pthread_sigmask: %s", strerror(ret));
 139         }
 140         if (!URCU_TLS(ust_mutex_nest)++)
 141                 pthread_mutex_lock(&ust_mutex);
 142         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 143         if (ret) {
 144                 ERR("pthread_sigmask: %s", strerror(ret));
 145         }
 146         if (lttng_ust_comm_should_quit) {
 147                 return -1;
 148         } else {
 149                 return 0;
 150         }
 151 }
 152
 153 /*
 154  * ust_lock_nocheck() can be used in constructors/destructors, because
 155  * they are already nested within the dynamic loader lock, and therefore
 156  * have exclusive access against execution of liblttng-ust destructor.
 157  * Signal-safe.
 158  */
 159 void ust_lock_nocheck(void)
 160 {
 161         sigset_t sig_all_blocked, orig_mask;
 162         int ret, oldstate;
 163
 164         ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
 165         if (ret) {
 166                 ERR("pthread_setcancelstate: %s", strerror(ret));
 167         }
 168         if (oldstate != PTHREAD_CANCEL_ENABLE) {
 169                 ERR("pthread_setcancelstate: unexpected oldstate");
 170         }
 171         sigfillset(&sig_all_blocked);
 172         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 173         if (ret) {
 174                 ERR("pthread_sigmask: %s", strerror(ret));
 175         }
 176         if (!URCU_TLS(ust_mutex_nest)++)
 177                 pthread_mutex_lock(&ust_mutex);
 178         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 179         if (ret) {
 180                 ERR("pthread_sigmask: %s", strerror(ret));
 181         }
 182 }
 183
 184 /*
 185  * Signal-safe.
 186  */
 187 void ust_unlock(void)
 188 {
 189         sigset_t sig_all_blocked, orig_mask;
 190         int ret, oldstate;
 191
 192         sigfillset(&sig_all_blocked);
 193         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
 194         if (ret) {
 195                 ERR("pthread_sigmask: %s", strerror(ret));
 196         }
 197         if (!--URCU_TLS(ust_mutex_nest))
 198                 pthread_mutex_unlock(&ust_mutex);
 199         ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
 200         if (ret) {
 201                 ERR("pthread_sigmask: %s", strerror(ret));
 202         }
 203         ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
 204         if (ret) {
 205                 ERR("pthread_setcancelstate: %s", strerror(ret));
 206         }
 207         if (oldstate != PTHREAD_CANCEL_DISABLE) {
 208                 ERR("pthread_setcancelstate: unexpected oldstate");
 209         }
 210 }
 211
 212 /*
 213  * Wait for either of these before continuing to the main
 214  * program:
 215  * - the register_done message from sessiond daemon
 216  *   (will let the sessiond daemon enable sessions before main
 217  *   starts.)
 218  * - sessiond daemon is not reachable.
 219  * - timeout (ensuring applications are resilient to session
 220  *   daemon problems).
 221  */
 222 static sem_t constructor_wait;
 223 /*
 224  * Doing this for both the global and local sessiond.
 225  */
 226 static int sem_count = { 2 };
 227
 228 /*
 229  * Counting nesting within lttng-ust. Used to ensure that calling fork()
 230  * from liblttng-ust does not execute the pre/post fork handlers.
 231  */
 232 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
 233
 234 /*
 235  * Info about socket and associated listener thread.
 236  */
 237 struct sock_info {
 238         const char *name;
 239         pthread_t ust_listener; /* listener thread */
 240         int root_handle;
 241         int constructor_sem_posted;
 242         int allowed;
 243         int global;
 244         int thread_active;
 245
 246         char sock_path[PATH_MAX];
 247         int socket;
 248         int notify_socket;
 249
 250         char wait_shm_path[PATH_MAX];
 251         char *wait_shm_mmap;
 252         /* Keep track of lazy state dump not performed yet. */
 253         int statedump_pending;
 254 };
 255
 256 /* Socket from app (connect) to session daemon (listen) for communication */
 257 struct sock_info global_apps = {
 258         .name = "global",
 259         .global = 1,
 260
 261         .root_handle = -1,
 262         .allowed = 0,
 263         .thread_active = 0,
 264
 265         .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
 266         .socket = -1,
 267         .notify_socket = -1,
 268
 269         .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
 270
 271         .statedump_pending = 0,
 272 };
 273
 274 /* TODO: allow global_apps_sock_path override */
 275
 276 struct sock_info local_apps = {
 277         .name = "local",
 278         .global = 0,
 279         .root_handle = -1,
 280         .allowed = 0,   /* Check setuid bit first */
 281         .thread_active = 0,
 282
 283         .socket = -1,
 284         .notify_socket = -1,
 285
 286         .statedump_pending = 0,
 287 };
 288
 289 static int wait_poll_fallback;
 290
 291 static const char *cmd_name_mapping[] = {
 292         [ LTTNG_UST_RELEASE ] = "Release",
 293         [ LTTNG_UST_SESSION ] = "Create Session",
 294         [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
 295
 296         [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
 297         [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
 298         [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
 299         [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
 300
 301         /* Session FD commands */
 302         [ LTTNG_UST_CHANNEL ] = "Create Channel",
 303         [ LTTNG_UST_SESSION_START ] = "Start Session",
 304         [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
 305
 306         /* Channel FD commands */
 307         [ LTTNG_UST_STREAM ] = "Create Stream",
 308         [ LTTNG_UST_EVENT ] = "Create Event",
 309
 310         /* Event and Channel FD commands */
 311         [ LTTNG_UST_CONTEXT ] = "Create Context",
 312         [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
 313
 314         /* Event, Channel and Session commands */
 315         [ LTTNG_UST_ENABLE ] = "Enable",
 316         [ LTTNG_UST_DISABLE ] = "Disable",
 317
 318         /* Tracepoint list commands */
 319         [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
 320         [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
 321
 322         /* Event FD commands */
 323         [ LTTNG_UST_FILTER ] = "Create Filter",
 324         [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
 325 };
 326
 327 static const char *str_timeout;
 328 static int got_timeout_env;
 329
 330 extern void lttng_ring_buffer_client_overwrite_init(void);
 331 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
 332 extern void lttng_ring_buffer_client_discard_init(void);
 333 extern void lttng_ring_buffer_client_discard_rt_init(void);
 334 extern void lttng_ring_buffer_metadata_client_init(void);
 335 extern void lttng_ring_buffer_client_overwrite_exit(void);
 336 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
 337 extern void lttng_ring_buffer_client_discard_exit(void);
 338 extern void lttng_ring_buffer_client_discard_rt_exit(void);
 339 extern void lttng_ring_buffer_metadata_client_exit(void);
 340
 341 static char *get_map_shm(struct sock_info *sock_info);
 342
 343 ssize_t lttng_ust_read(int fd, void *buf, size_t len)
 344 {
 345         ssize_t ret;
 346         size_t copied = 0, to_copy = len;
 347
 348         do {
 349                 ret = read(fd, buf + copied, to_copy);
 350                 if (ret > 0) {
 351                         copied += ret;
 352                         to_copy -= ret;
 353                 }
 354         } while ((ret > 0 && to_copy > 0)
 355                 || (ret < 0 && errno == EINTR));
 356         if (ret > 0) {
 357                 ret = copied;
 358         }
 359         return ret;
 360 }
 361 /*
 362  * Returns the HOME directory path. Caller MUST NOT free(3) the returned
 363  * pointer.
 364  */
 365 static
 366 const char *get_lttng_home_dir(void)
 367 {
 368        const char *val;
 369
 370        val = (const char *) lttng_getenv("LTTNG_HOME");
 371        if (val != NULL) {
 372                return val;
 373        }
 374        return (const char *) lttng_getenv("HOME");
 375 }
 376
 377 /*
 378  * Force a read (imply TLS fixup for dlopen) of TLS variables.
 379  */
 380 static
 381 void lttng_fixup_nest_count_tls(void)
 382 {
 383         asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
 384 }
 385
 386 static
 387 void lttng_fixup_ust_mutex_nest_tls(void)
 388 {
 389         asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
 390 }
 391
 392 /*
 393  * Fixup urcu bp TLS.
 394  */
 395 static
 396 void lttng_fixup_urcu_bp_tls(void)
 397 {
 398         rcu_read_lock();
 399         rcu_read_unlock();
 400 }
 401
 402 void lttng_ust_fixup_tls(void)
 403 {
 404         lttng_fixup_urcu_bp_tls();
 405         lttng_fixup_ringbuffer_tls();
 406         lttng_fixup_vtid_tls();
 407         lttng_fixup_nest_count_tls();
 408         lttng_fixup_procname_tls();
 409         lttng_fixup_ust_mutex_nest_tls();
 410         lttng_ust_fixup_fd_tracker_tls();
 411 }
 412
 413 int lttng_get_notify_socket(void *owner)
 414 {
 415         struct sock_info *info = owner;
 416
 417         return info->notify_socket;
 418 }
 419
 420 static
 421 void print_cmd(int cmd, int handle)
 422 {
 423         const char *cmd_name = "Unknown";
 424
 425         if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
 426                         && cmd_name_mapping[cmd]) {
 427                 cmd_name = cmd_name_mapping[cmd];
 428         }
 429         DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
 430                 cmd_name, cmd,
 431                 lttng_ust_obj_get_name(handle), handle);
 432 }
 433
 434 static
 435 int setup_global_apps(void)
 436 {
 437         int ret = 0;
 438         assert(!global_apps.wait_shm_mmap);
 439
 440         global_apps.wait_shm_mmap = get_map_shm(&global_apps);
 441         if (!global_apps.wait_shm_mmap) {
 442                 WARN("Unable to get map shm for global apps. Disabling LTTng-UST global tracing.");
 443                 global_apps.allowed = 0;
 444                 ret = -EIO;
 445                 goto error;
 446         }
 447
 448         global_apps.allowed = 1;
 449 error:
 450         return ret;
 451 }
 452 static
 453 int setup_local_apps(void)
 454 {
 455         int ret = 0;
 456         const char *home_dir;
 457         uid_t uid;
 458
 459         assert(!local_apps.wait_shm_mmap);
 460
 461         uid = getuid();
 462         /*
 463          * Disallow per-user tracing for setuid binaries.
 464          */
 465         if (uid != geteuid()) {
 466                 assert(local_apps.allowed == 0);
 467                 ret = 0;
 468                 goto end;
 469         }
 470         home_dir = get_lttng_home_dir();
 471         if (!home_dir) {
 472                 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
 473                 assert(local_apps.allowed == 0);
 474                 ret = -ENOENT;
 475                 goto end;
 476         }
 477         local_apps.allowed = 1;
 478         snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
 479                 home_dir,
 480                 LTTNG_DEFAULT_HOME_RUNDIR,
 481                 LTTNG_UST_SOCK_FILENAME);
 482         snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
 483                 LTTNG_UST_WAIT_FILENAME,
 484                 uid);
 485
 486         local_apps.wait_shm_mmap = get_map_shm(&local_apps);
 487         if (!local_apps.wait_shm_mmap) {
 488                 WARN("Unable to get map shm for local apps. Disabling LTTng-UST per-user tracing.");
 489                 local_apps.allowed = 0;
 490                 ret = -EIO;
 491                 goto end;
 492         }
 493 end:
 494         return ret;
 495 }
 496
 497 /*
 498  * Get socket timeout, in ms.
 499  * -1: wait forever. 0: don't wait. >0: timeout, in ms.
 500  */
 501 static
 502 long get_timeout(void)
 503 {
 504         long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
 505
 506         if (!got_timeout_env) {
 507                 str_timeout = lttng_getenv("LTTNG_UST_REGISTER_TIMEOUT");
 508                 got_timeout_env = 1;
 509         }
 510         if (str_timeout)
 511                 constructor_delay_ms = strtol(str_timeout, NULL, 10);
 512         /* All negative values are considered as "-1". */
 513         if (constructor_delay_ms < -1)
 514                 constructor_delay_ms = -1;
 515         return constructor_delay_ms;
 516 }
 517
 518 /* Timeout for notify socket send and recv. */
 519 static
 520 long get_notify_sock_timeout(void)
 521 {
 522         return get_timeout();
 523 }
 524
 525 /* Timeout for connecting to cmd and notify sockets. */
 526 static
 527 long get_connect_sock_timeout(void)
 528 {
 529         return get_timeout();
 530 }
 531
 532 /*
 533  * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
 534  */
 535 static
 536 int get_constructor_timeout(struct timespec *constructor_timeout)
 537 {
 538         long constructor_delay_ms;
 539         int ret;
 540
 541         constructor_delay_ms = get_timeout();
 542
 543         switch (constructor_delay_ms) {
 544         case -1:/* fall-through */
 545         case 0:
 546                 return constructor_delay_ms;
 547         default:
 548                 break;
 549         }
 550
 551         /*
 552          * If we are unable to find the current time, don't wait.
 553          */
 554         ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
 555         if (ret) {
 556                 /* Don't wait. */
 557                 return 0;
 558         }
 559         constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
 560         constructor_timeout->tv_nsec +=
 561                 (constructor_delay_ms % 1000UL) * 1000000UL;
 562         if (constructor_timeout->tv_nsec >= 1000000000UL) {
 563                 constructor_timeout->tv_sec++;
 564                 constructor_timeout->tv_nsec -= 1000000000UL;
 565         }
 566         /* Timeout wait (constructor_delay_ms). */
 567         return 1;
 568 }
 569
 570 static
 571 int register_to_sessiond(int socket, enum ustctl_socket_type type)
 572 {
 573         return ustcomm_send_reg_msg(socket,
 574                 type,
 575                 CAA_BITS_PER_LONG,
 576                 lttng_alignof(uint8_t) * CHAR_BIT,
 577                 lttng_alignof(uint16_t) * CHAR_BIT,
 578                 lttng_alignof(uint32_t) * CHAR_BIT,
 579                 lttng_alignof(uint64_t) * CHAR_BIT,
 580                 lttng_alignof(unsigned long) * CHAR_BIT);
 581 }
 582
 583 static
 584 int send_reply(int sock, struct ustcomm_ust_reply *lur)
 585 {
 586         ssize_t len;
 587
 588         len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
 589         switch (len) {
 590         case sizeof(*lur):
 591                 DBG("message successfully sent");
 592                 return 0;
 593         default:
 594                 if (len == -ECONNRESET) {
 595                         DBG("remote end closed connection");
 596                         return 0;
 597                 }
 598                 if (len < 0)
 599                         return len;
 600                 DBG("incorrect message size: %zd", len);
 601                 return -EINVAL;
 602         }
 603 }
 604
 605 static
 606 int handle_register_done(struct sock_info *sock_info)
 607 {
 608         int ret;
 609
 610         if (sock_info->constructor_sem_posted)
 611                 return 0;
 612         sock_info->constructor_sem_posted = 1;
 613         if (uatomic_read(&sem_count) <= 0) {
 614                 return 0;
 615         }
 616         ret = uatomic_add_return(&sem_count, -1);
 617         if (ret == 0) {
 618                 ret = sem_post(&constructor_wait);
 619                 assert(!ret);
 620         }
 621         return 0;
 622 }
 623
 624 /*
 625  * Only execute pending statedump after the constructor semaphore has
 626  * been posted by each listener thread. This means statedump will only
 627  * be performed after the "registration done" command is received from
 628  * each session daemon the application is connected to.
 629  *
 630  * This ensures we don't run into deadlock issues with the dynamic
 631  * loader mutex, which is held while the constructor is called and
 632  * waiting on the constructor semaphore. All operations requiring this
 633  * dynamic loader lock need to be postponed using this mechanism.
 634  */
 635 static
 636 void handle_pending_statedump(struct sock_info *sock_info)
 637 {
 638         int ctor_passed = sock_info->constructor_sem_posted;
 639
 640         if (ctor_passed && sock_info->statedump_pending) {
 641                 sock_info->statedump_pending = 0;
 642                 pthread_mutex_lock(&ust_fork_mutex);
 643                 lttng_handle_pending_statedump(sock_info);
 644                 pthread_mutex_unlock(&ust_fork_mutex);
 645         }
 646 }
 647
 648 static
 649 int handle_message(struct sock_info *sock_info,
 650                 int sock, struct ustcomm_ust_msg *lum)
 651 {
 652         int ret = 0;
 653         const struct lttng_ust_objd_ops *ops;
 654         struct ustcomm_ust_reply lur;
 655         union ust_args args;
 656         char ctxstr[LTTNG_UST_SYM_NAME_LEN];    /* App context string. */
 657         ssize_t len;
 658
 659         memset(&lur, 0, sizeof(lur));
 660
 661         if (ust_lock()) {
 662                 ret = -LTTNG_UST_ERR_EXITING;
 663                 goto error;
 664         }
 665
 666         ops = objd_ops(lum->handle);
 667         if (!ops) {
 668                 ret = -ENOENT;
 669                 goto error;
 670         }
 671
 672         switch (lum->cmd) {
 673         case LTTNG_UST_REGISTER_DONE:
 674                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 675                         ret = handle_register_done(sock_info);
 676                 else
 677                         ret = -EINVAL;
 678                 break;
 679         case LTTNG_UST_RELEASE:
 680                 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
 681                         ret = -EPERM;
 682                 else
 683                         ret = lttng_ust_objd_unref(lum->handle, 1);
 684                 break;
 685         case LTTNG_UST_FILTER:
 686         {
 687                 /* Receive filter data */
 688                 struct lttng_ust_filter_bytecode_node *bytecode;
 689
 690                 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
 691                         ERR("Filter data size is too large: %u bytes",
 692                                 lum->u.filter.data_size);
 693                         ret = -EINVAL;
 694                         goto error;
 695                 }
 696
 697                 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
 698                         ERR("Filter reloc offset %u is not within data",
 699                                 lum->u.filter.reloc_offset);
 700                         ret = -EINVAL;
 701                         goto error;
 702                 }
 703
 704                 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
 705                 if (!bytecode) {
 706                         ret = -ENOMEM;
 707                         goto error;
 708                 }
 709                 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
 710                                 lum->u.filter.data_size);
 711                 switch (len) {
 712                 case 0: /* orderly shutdown */
 713                         ret = 0;
 714                         free(bytecode);
 715                         goto error;
 716                 default:
 717                         if (len == lum->u.filter.data_size) {
 718                                 DBG("filter data received");
 719                                 break;
 720                         } else if (len < 0) {
 721                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 722                                 if (len == -ECONNRESET) {
 723                                         ERR("%s remote end closed connection", sock_info->name);
 724                                         ret = len;
 725                                         free(bytecode);
 726                                         goto error;
 727                                 }
 728                                 ret = len;
 729                                 free(bytecode);
 730                                 goto error;
 731                         } else {
 732                                 DBG("incorrect filter data message size: %zd", len);
 733                                 ret = -EINVAL;
 734                                 free(bytecode);
 735                                 goto error;
 736                         }
 737                 }
 738                 bytecode->bc.len = lum->u.filter.data_size;
 739                 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
 740                 bytecode->bc.seqnum = lum->u.filter.seqnum;
 741                 if (ops->cmd) {
 742                         ret = ops->cmd(lum->handle, lum->cmd,
 743                                         (unsigned long) bytecode,
 744                                         &args, sock_info);
 745                         if (ret) {
 746                                 free(bytecode);
 747                         }
 748                         /* don't free bytecode if everything went fine. */
 749                 } else {
 750                         ret = -ENOSYS;
 751                         free(bytecode);
 752                 }
 753                 break;
 754         }
 755         case LTTNG_UST_EXCLUSION:
 756         {
 757                 /* Receive exclusion names */
 758                 struct lttng_ust_excluder_node *node;
 759                 unsigned int count;
 760
 761                 count = lum->u.exclusion.count;
 762                 if (count == 0) {
 763                         /* There are no names to read */
 764                         ret = 0;
 765                         goto error;
 766                 }
 767                 node = zmalloc(sizeof(*node) +
 768                                 count * LTTNG_UST_SYM_NAME_LEN);
 769                 if (!node) {
 770                         ret = -ENOMEM;
 771                         goto error;
 772                 }
 773                 node->excluder.count = count;
 774                 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
 775                                 count * LTTNG_UST_SYM_NAME_LEN);
 776                 switch (len) {
 777                 case 0: /* orderly shutdown */
 778                         ret = 0;
 779                         free(node);
 780                         goto error;
 781                 default:
 782                         if (len == count * LTTNG_UST_SYM_NAME_LEN) {
 783                                 DBG("Exclusion data received");
 784                                 break;
 785                         } else if (len < 0) {
 786                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 787                                 if (len == -ECONNRESET) {
 788                                         ERR("%s remote end closed connection", sock_info->name);
 789                                         ret = len;
 790                                         free(node);
 791                                         goto error;
 792                                 }
 793                                 ret = len;
 794                                 free(node);
 795                                 goto error;
 796                         } else {
 797                                 DBG("Incorrect exclusion data message size: %zd", len);
 798                                 ret = -EINVAL;
 799                                 free(node);
 800                                 goto error;
 801                         }
 802                 }
 803                 if (ops->cmd) {
 804                         ret = ops->cmd(lum->handle, lum->cmd,
 805                                         (unsigned long) node,
 806                                         &args, sock_info);
 807                         if (ret) {
 808                                 free(node);
 809                         }
 810                         /* Don't free exclusion data if everything went fine. */
 811                 } else {
 812                         ret = -ENOSYS;
 813                         free(node);
 814                 }
 815                 break;
 816         }
 817         case LTTNG_UST_CHANNEL:
 818         {
 819                 void *chan_data;
 820                 int wakeup_fd;
 821
 822                 len = ustcomm_recv_channel_from_sessiond(sock,
 823                                 &chan_data, lum->u.channel.len,
 824                                 &wakeup_fd);
 825                 switch (len) {
 826                 case 0: /* orderly shutdown */
 827                         ret = 0;
 828                         goto error;
 829                 default:
 830                         if (len == lum->u.channel.len) {
 831                                 DBG("channel data received");
 832                                 break;
 833                         } else if (len < 0) {
 834                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 835                                 if (len == -ECONNRESET) {
 836                                         ERR("%s remote end closed connection", sock_info->name);
 837                                         ret = len;
 838                                         goto error;
 839                                 }
 840                                 ret = len;
 841                                 goto error;
 842                         } else {
 843                                 DBG("incorrect channel data message size: %zd", len);
 844                                 ret = -EINVAL;
 845                                 goto error;
 846                         }
 847                 }
 848                 args.channel.chan_data = chan_data;
 849                 args.channel.wakeup_fd = wakeup_fd;
 850                 if (ops->cmd)
 851                         ret = ops->cmd(lum->handle, lum->cmd,
 852                                         (unsigned long) &lum->u,
 853                                         &args, sock_info);
 854                 else
 855                         ret = -ENOSYS;
 856                 break;
 857         }
 858         case LTTNG_UST_STREAM:
 859         {
 860                 uint64_t memory_map_size;
 861
 862                 /* Receive shm_fd, wakeup_fd */
 863                 ret = ustcomm_recv_stream_from_sessiond(sock,
 864                         &memory_map_size,
 865                         &args.stream.shm_fd,
 866                         &args.stream.wakeup_fd);
 867                 if (ret) {
 868                         goto error;
 869                 }
 870                 lum->u.stream.len = memory_map_size;
 871
 872                 if (ops->cmd)
 873                         ret = ops->cmd(lum->handle, lum->cmd,
 874                                         (unsigned long) &lum->u,
 875                                         &args, sock_info);
 876                 else
 877                         ret = -ENOSYS;
 878                 break;
 879         }
 880         case LTTNG_UST_CONTEXT:
 881                 switch (lum->u.context.ctx) {
 882                 case LTTNG_UST_CONTEXT_APP_CONTEXT:
 883                 {
 884                         char *p;
 885                         size_t ctxlen, recvlen;
 886
 887                         ctxlen = strlen("$app.") + lum->u.context.u.app_ctx.provider_name_len - 1
 888                                         + strlen(":") + lum->u.context.u.app_ctx.ctx_name_len;
 889                         if (ctxlen >= LTTNG_UST_SYM_NAME_LEN) {
 890                                 ERR("Application context string length size is too large: %zu bytes",
 891                                         ctxlen);
 892                                 ret = -EINVAL;
 893                                 goto error;
 894                         }
 895                         strcpy(ctxstr, "$app.");
 896                         p = &ctxstr[strlen("$app.")];
 897                         recvlen = ctxlen - strlen("$app.");
 898                         len = ustcomm_recv_unix_sock(sock, p, recvlen);
 899                         switch (len) {
 900                         case 0: /* orderly shutdown */
 901                                 ret = 0;
 902                                 goto error;
 903                         default:
 904                                 if (len == recvlen) {
 905                                         DBG("app context data received");
 906                                         break;
 907                                 } else if (len < 0) {
 908                                         DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
 909                                         if (len == -ECONNRESET) {
 910                                                 ERR("%s remote end closed connection", sock_info->name);
 911                                                 ret = len;
 912                                                 goto error;
 913                                         }
 914                                         ret = len;
 915                                         goto error;
 916                                 } else {
 917                                         DBG("incorrect app context data message size: %zd", len);
 918                                         ret = -EINVAL;
 919                                         goto error;
 920                                 }
 921                         }
 922                         /* Put : between provider and ctxname. */
 923                         p[lum->u.context.u.app_ctx.provider_name_len - 1] = ':';
 924                         args.app_context.ctxname = ctxstr;
 925                         break;
 926                 }
 927                 default:
 928                         break;
 929                 }
 930                 if (ops->cmd) {
 931                         ret = ops->cmd(lum->handle, lum->cmd,
 932                                         (unsigned long) &lum->u,
 933                                         &args, sock_info);
 934                 } else {
 935                         ret = -ENOSYS;
 936                 }
 937                 break;
 938         default:
 939                 if (ops->cmd)
 940                         ret = ops->cmd(lum->handle, lum->cmd,
 941                                         (unsigned long) &lum->u,
 942                                         &args, sock_info);
 943                 else
 944                         ret = -ENOSYS;
 945                 break;
 946         }
 947
 948         lur.handle = lum->handle;
 949         lur.cmd = lum->cmd;
 950         lur.ret_val = ret;
 951         if (ret >= 0) {
 952                 lur.ret_code = LTTNG_UST_OK;
 953         } else {
 954                 /*
 955                  * Use -LTTNG_UST_ERR as wildcard for UST internal
 956                  * error that are not caused by the transport, except if
 957                  * we already have a more precise error message to
 958                  * report.
 959                  */
 960                 if (ret > -LTTNG_UST_ERR) {
 961                         /* Translate code to UST error. */
 962                         switch (ret) {
 963                         case -EEXIST:
 964                                 lur.ret_code = -LTTNG_UST_ERR_EXIST;
 965                                 break;
 966                         case -EINVAL:
 967                                 lur.ret_code = -LTTNG_UST_ERR_INVAL;
 968                                 break;
 969                         case -ENOENT:
 970                                 lur.ret_code = -LTTNG_UST_ERR_NOENT;
 971                                 break;
 972                         case -EPERM:
 973                                 lur.ret_code = -LTTNG_UST_ERR_PERM;
 974                                 break;
 975                         case -ENOSYS:
 976                                 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
 977                                 break;
 978                         default:
 979                                 lur.ret_code = -LTTNG_UST_ERR;
 980                                 break;
 981                         }
 982                 } else {
 983                         lur.ret_code = ret;
 984                 }
 985         }
 986         if (ret >= 0) {
 987                 switch (lum->cmd) {
 988                 case LTTNG_UST_TRACER_VERSION:
 989                         lur.u.version = lum->u.version;
 990                         break;
 991                 case LTTNG_UST_TRACEPOINT_LIST_GET:
 992                         memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
 993                         break;
 994                 }
 995         }
 996         DBG("Return value: %d", lur.ret_val);
 997
 998         ust_unlock();
 999
1000         /*
1001          * Performed delayed statedump operations outside of the UST
1002          * lock. We need to take the dynamic loader lock before we take
1003          * the UST lock internally within handle_pending_statedump().
1004           */
1005         handle_pending_statedump(sock_info);
1006
1007         if (ust_lock()) {
1008                 ret = -LTTNG_UST_ERR_EXITING;
1009                 goto error;
1010         }
1011
1012         ret = send_reply(sock, &lur);
1013         if (ret < 0) {
1014                 DBG("error sending reply");
1015                 goto error;
1016         }
1017
1018         /*
1019          * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
1020          * after the reply.
1021          */
1022         if (lur.ret_code == LTTNG_UST_OK) {
1023                 switch (lum->cmd) {
1024                 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
1025                         len = ustcomm_send_unix_sock(sock,
1026                                 &args.field_list.entry,
1027                                 sizeof(args.field_list.entry));
1028                         if (len < 0) {
1029                                 ret = len;
1030                                 goto error;
1031                         }
1032                         if (len != sizeof(args.field_list.entry)) {
1033                                 ret = -EINVAL;
1034                                 goto error;
1035                         }
1036                 }
1037         }
1038
1039 error:
1040         ust_unlock();
1041
1042         return ret;
1043 }
1044
1045 static
1046 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
1047 {
1048         int ret;
1049
1050         if (sock_info->root_handle != -1) {
1051                 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
1052                 if (ret) {
1053                         ERR("Error unref root handle");
1054                 }
1055                 sock_info->root_handle = -1;
1056         }
1057         sock_info->constructor_sem_posted = 0;
1058
1059         /*
1060          * wait_shm_mmap, socket and notify socket are used by listener
1061          * threads outside of the ust lock, so we cannot tear them down
1062          * ourselves, because we cannot join on these threads. Leave
1063          * responsibility of cleaning up these resources to the OS
1064          * process exit.
1065          */
1066         if (exiting)
1067                 return;
1068
1069         if (sock_info->socket != -1) {
1070                 ret = ustcomm_close_unix_sock(sock_info->socket);
1071                 if (ret) {
1072                         ERR("Error closing ust cmd socket");
1073                 }
1074                 sock_info->socket = -1;
1075         }
1076         if (sock_info->notify_socket != -1) {
1077                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1078                 if (ret) {
1079                         ERR("Error closing ust notify socket");
1080                 }
1081                 sock_info->notify_socket = -1;
1082         }
1083         if (sock_info->wait_shm_mmap) {
1084                 long page_size;
1085
1086                 page_size = sysconf(_SC_PAGE_SIZE);
1087                 if (page_size <= 0) {
1088                         if (!page_size) {
1089                                 errno = EINVAL;
1090                         }
1091                         PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1092                 } else {
1093                         ret = munmap(sock_info->wait_shm_mmap, page_size);
1094                         if (ret) {
1095                                 ERR("Error unmapping wait shm");
1096                         }
1097                 }
1098                 sock_info->wait_shm_mmap = NULL;
1099         }
1100 }
1101
1102 /*
1103  * Using fork to set umask in the child process (not multi-thread safe).
1104  * We deal with the shm_open vs ftruncate race (happening when the
1105  * sessiond owns the shm and does not let everybody modify it, to ensure
1106  * safety against shm_unlink) by simply letting the mmap fail and
1107  * retrying after a few seconds.
1108  * For global shm, everybody has rw access to it until the sessiond
1109  * starts.
1110  */
1111 static
1112 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
1113 {
1114         int wait_shm_fd, ret;
1115         pid_t pid;
1116
1117         /*
1118          * Try to open read-only.
1119          */
1120         wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1121         if (wait_shm_fd >= 0) {
1122                 int32_t tmp_read;
1123                 ssize_t len;
1124                 size_t bytes_read = 0;
1125
1126                 /*
1127                  * Try to read the fd. If unable to do so, try opening
1128                  * it in write mode.
1129                  */
1130                 do {
1131                         len = read(wait_shm_fd,
1132                                 &((char *) &tmp_read)[bytes_read],
1133                                 sizeof(tmp_read) - bytes_read);
1134                         if (len > 0) {
1135                                 bytes_read += len;
1136                         }
1137                 } while ((len < 0 && errno == EINTR)
1138                         || (len > 0 && bytes_read < sizeof(tmp_read)));
1139                 if (bytes_read != sizeof(tmp_read)) {
1140                         ret = close(wait_shm_fd);
1141                         if (ret) {
1142                                 ERR("close wait_shm_fd");
1143                         }
1144                         goto open_write;
1145                 }
1146                 goto end;
1147         } else if (wait_shm_fd < 0 && errno != ENOENT) {
1148                 /*
1149                  * Real-only open did not work, and it's not because the
1150                  * entry was not present. It's a failure that prohibits
1151                  * using shm.
1152                  */
1153                 ERR("Error opening shm %s", sock_info->wait_shm_path);
1154                 goto end;
1155         }
1156
1157 open_write:
1158         /*
1159          * If the open failed because the file did not exist, or because
1160          * the file was not truncated yet, try creating it ourself.
1161          */
1162         URCU_TLS(lttng_ust_nest_count)++;
1163         pid = fork();
1164         URCU_TLS(lttng_ust_nest_count)--;
1165         if (pid > 0) {
1166                 int status;
1167
1168                 /*
1169                  * Parent: wait for child to return, in which case the
1170                  * shared memory map will have been created.
1171                  */
1172                 pid = wait(&status);
1173                 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1174                         wait_shm_fd = -1;
1175                         goto end;
1176                 }
1177                 /*
1178                  * Try to open read-only again after creation.
1179                  */
1180                 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1181                 if (wait_shm_fd < 0) {
1182                         /*
1183                          * Real-only open did not work. It's a failure
1184                          * that prohibits using shm.
1185                          */
1186                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1187                         goto end;
1188                 }
1189                 goto end;
1190         } else if (pid == 0) {
1191                 int create_mode;
1192
1193                 /* Child */
1194                 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1195                 if (sock_info->global)
1196                         create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1197                 /*
1198                  * We're alone in a child process, so we can modify the
1199                  * process-wide umask.
1200                  */
1201                 umask(~create_mode);
1202                 /*
1203                  * Try creating shm (or get rw access).
1204                  * We don't do an exclusive open, because we allow other
1205                  * processes to create+ftruncate it concurrently.
1206                  */
1207                 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1208                                 O_RDWR | O_CREAT, create_mode);
1209                 if (wait_shm_fd >= 0) {
1210                         ret = ftruncate(wait_shm_fd, mmap_size);
1211                         if (ret) {
1212                                 PERROR("ftruncate");
1213                                 _exit(EXIT_FAILURE);
1214                         }
1215                         _exit(EXIT_SUCCESS);
1216                 }
1217                 /*
1218                  * For local shm, we need to have rw access to accept
1219                  * opening it: this means the local sessiond will be
1220                  * able to wake us up. For global shm, we open it even
1221                  * if rw access is not granted, because the root.root
1222                  * sessiond will be able to override all rights and wake
1223                  * us up.
1224                  */
1225                 if (!sock_info->global && errno != EACCES) {
1226                         ERR("Error opening shm %s", sock_info->wait_shm_path);
1227                         _exit(EXIT_FAILURE);
1228                 }
1229                 /*
1230                  * The shm exists, but we cannot open it RW. Report
1231                  * success.
1232                  */
1233                 _exit(EXIT_SUCCESS);
1234         } else {
1235                 return -1;
1236         }
1237 end:
1238         if (wait_shm_fd >= 0 && !sock_info->global) {
1239                 struct stat statbuf;
1240
1241                 /*
1242                  * Ensure that our user is the owner of the shm file for
1243                  * local shm. If we do not own the file, it means our
1244                  * sessiond will not have access to wake us up (there is
1245                  * probably a rogue process trying to fake our
1246                  * sessiond). Fallback to polling method in this case.
1247                  */
1248                 ret = fstat(wait_shm_fd, &statbuf);
1249                 if (ret) {
1250                         PERROR("fstat");
1251                         goto error_close;
1252                 }
1253                 if (statbuf.st_uid != getuid())
1254                         goto error_close;
1255         }
1256         return wait_shm_fd;
1257
1258 error_close:
1259         ret = close(wait_shm_fd);
1260         if (ret) {
1261                 PERROR("Error closing fd");
1262         }
1263         return -1;
1264 }
1265
1266 static
1267 char *get_map_shm(struct sock_info *sock_info)
1268 {
1269         long page_size;
1270         int wait_shm_fd, ret;
1271         char *wait_shm_mmap;
1272
1273         page_size = sysconf(_SC_PAGE_SIZE);
1274         if (page_size <= 0) {
1275                 if (!page_size) {
1276                         errno = EINVAL;
1277                 }
1278                 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1279                 goto error;
1280         }
1281
1282         lttng_ust_lock_fd_tracker();
1283         wait_shm_fd = get_wait_shm(sock_info, page_size);
1284         if (wait_shm_fd < 0) {
1285                 lttng_ust_unlock_fd_tracker();
1286                 goto error;
1287         }
1288
1289         ret = lttng_ust_add_fd_to_tracker(wait_shm_fd);
1290         if (ret < 0) {
1291                 ret = close(wait_shm_fd);
1292                 if (!ret) {
1293                         PERROR("Error closing fd");
1294                 }
1295                 lttng_ust_unlock_fd_tracker();
1296                 goto error;
1297         }
1298
1299         wait_shm_fd = ret;
1300         lttng_ust_unlock_fd_tracker();
1301
1302         wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1303                   MAP_SHARED, wait_shm_fd, 0);
1304
1305         /* close shm fd immediately after taking the mmap reference */
1306         lttng_ust_lock_fd_tracker();
1307         ret = close(wait_shm_fd);
1308         if (!ret) {
1309                 lttng_ust_delete_fd_from_tracker(wait_shm_fd);
1310         } else {
1311                 PERROR("Error closing fd");
1312         }
1313         lttng_ust_unlock_fd_tracker();
1314
1315         if (wait_shm_mmap == MAP_FAILED) {
1316                 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1317                 goto error;
1318         }
1319         return wait_shm_mmap;
1320
1321 error:
1322         return NULL;
1323 }
1324
1325 static
1326 void wait_for_sessiond(struct sock_info *sock_info)
1327 {
1328         /* Use ust_lock to check if we should quit. */
1329         if (ust_lock()) {
1330                 goto quit;
1331         }
1332         if (wait_poll_fallback) {
1333                 goto error;
1334         }
1335         ust_unlock();
1336
1337         assert(sock_info->wait_shm_mmap);
1338
1339         DBG("Waiting for %s apps sessiond", sock_info->name);
1340         /* Wait for futex wakeup */
1341         if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1342                 goto end_wait;
1343
1344         while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1345                         FUTEX_WAIT, 0, NULL, NULL, 0)) {
1346                 switch (errno) {
1347                 case EWOULDBLOCK:
1348                         /* Value already changed. */
1349                         goto end_wait;
1350                 case EINTR:
1351                         /* Retry if interrupted by signal. */
1352                         break;  /* Get out of switch. */
1353                 case EFAULT:
1354                         wait_poll_fallback = 1;
1355                         DBG(
1356 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1357 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1358 "Please upgrade your kernel "
1359 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1360 "mainline). LTTng-UST will use polling mode fallback.");
1361                         if (ust_debug())
1362                                 PERROR("futex");
1363                         goto end_wait;
1364                 }
1365         }
1366 end_wait:
1367         return;
1368
1369 quit:
1370         ust_unlock();
1371         return;
1372
1373 error:
1374         ust_unlock();
1375         return;
1376 }
1377
1378 /*
1379  * This thread does not allocate any resource, except within
1380  * handle_message, within mutex protection. This mutex protects against
1381  * fork and exit.
1382  * The other moment it allocates resources is at socket connection, which
1383  * is also protected by the mutex.
1384  */
1385 static
1386 void *ust_listener_thread(void *arg)
1387 {
1388         struct sock_info *sock_info = arg;
1389         int sock, ret, prev_connect_failed = 0, has_waited = 0, fd;
1390         long timeout;
1391
1392         lttng_ust_fixup_tls();
1393         /*
1394          * If available, add '-ust' to the end of this thread's
1395          * process name
1396          */
1397         ret = lttng_ust_setustprocname();
1398         if (ret) {
1399                 ERR("Unable to set UST process name");
1400         }
1401
1402         /* Restart trying to connect to the session daemon */
1403 restart:
1404         if (prev_connect_failed) {
1405                 /* Wait for sessiond availability with pipe */
1406                 wait_for_sessiond(sock_info);
1407                 if (has_waited) {
1408                         has_waited = 0;
1409                         /*
1410                          * Sleep for 5 seconds before retrying after a
1411                          * sequence of failure / wait / failure. This
1412                          * deals with a killed or broken session daemon.
1413                          */
1414                         sleep(5);
1415                 } else {
1416                         has_waited = 1;
1417                 }
1418                 prev_connect_failed = 0;
1419         }
1420
1421         if (ust_lock()) {
1422                 goto quit;
1423         }
1424
1425         if (sock_info->socket != -1) {
1426                 /* FD tracker is updated by ustcomm_close_unix_sock() */
1427                 ret = ustcomm_close_unix_sock(sock_info->socket);
1428                 if (ret) {
1429                         ERR("Error closing %s ust cmd socket",
1430                                 sock_info->name);
1431                 }
1432                 sock_info->socket = -1;
1433         }
1434         if (sock_info->notify_socket != -1) {
1435                 /* FD tracker is updated by ustcomm_close_unix_sock() */
1436                 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1437                 if (ret) {
1438                         ERR("Error closing %s ust notify socket",
1439                                 sock_info->name);
1440                 }
1441                 sock_info->notify_socket = -1;
1442         }
1443
1444
1445         /*
1446          * Register. We need to perform both connect and sending
1447          * registration message before doing the next connect otherwise
1448          * we may reach unix socket connect queue max limits and block
1449          * on the 2nd connect while the session daemon is awaiting the
1450          * first connect registration message.
1451          */
1452         /* Connect cmd socket */
1453         lttng_ust_lock_fd_tracker();
1454         ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1455                 get_connect_sock_timeout());
1456         if (ret < 0) {
1457                 lttng_ust_unlock_fd_tracker();
1458                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1459                 prev_connect_failed = 1;
1460
1461                 /*
1462                  * If we cannot find the sessiond daemon, don't delay
1463                  * constructor execution.
1464                  */
1465                 ret = handle_register_done(sock_info);
1466                 assert(!ret);
1467                 ust_unlock();
1468                 goto restart;
1469         }
1470         fd = ret;
1471         ret = lttng_ust_add_fd_to_tracker(fd);
1472         if (ret < 0) {
1473                 ret = close(fd);
1474                 if (ret) {
1475                         PERROR("close on sock_info->socket");
1476                 }
1477                 ret = -1;
1478                 lttng_ust_unlock_fd_tracker();
1479                 ust_unlock();
1480                 goto quit;
1481         }
1482
1483         sock_info->socket = ret;
1484         lttng_ust_unlock_fd_tracker();
1485
1486         ust_unlock();
1487         /*
1488          * Unlock/relock ust lock because connect is blocking (with
1489          * timeout). Don't delay constructors on the ust lock for too
1490          * long.
1491          */
1492         if (ust_lock()) {
1493                 goto quit;
1494         }
1495
1496         /*
1497          * Create only one root handle per listener thread for the whole
1498          * process lifetime, so we ensure we get ID which is statically
1499          * assigned to the root handle.
1500          */
1501         if (sock_info->root_handle == -1) {
1502                 ret = lttng_abi_create_root_handle();
1503                 if (ret < 0) {
1504                         ERR("Error creating root handle");
1505                         goto quit;
1506                 }
1507                 sock_info->root_handle = ret;
1508         }
1509
1510         ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1511         if (ret < 0) {
1512                 ERR("Error registering to %s ust cmd socket",
1513                         sock_info->name);
1514                 prev_connect_failed = 1;
1515                 /*
1516                  * If we cannot register to the sessiond daemon, don't
1517                  * delay constructor execution.
1518                  */
1519                 ret = handle_register_done(sock_info);
1520                 assert(!ret);
1521                 ust_unlock();
1522                 goto restart;
1523         }
1524
1525         ust_unlock();
1526         /*
1527          * Unlock/relock ust lock because connect is blocking (with
1528          * timeout). Don't delay constructors on the ust lock for too
1529          * long.
1530          */
1531         if (ust_lock()) {
1532                 goto quit;
1533         }
1534
1535         /* Connect notify socket */
1536         lttng_ust_lock_fd_tracker();
1537         ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1538                 get_connect_sock_timeout());
1539         if (ret < 0) {
1540                 lttng_ust_unlock_fd_tracker();
1541                 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1542                 prev_connect_failed = 1;
1543
1544                 /*
1545                  * If we cannot find the sessiond daemon, don't delay
1546                  * constructor execution.
1547                  */
1548                 ret = handle_register_done(sock_info);
1549                 assert(!ret);
1550                 ust_unlock();
1551                 goto restart;
1552         }
1553
1554         fd = ret;
1555         ret = lttng_ust_add_fd_to_tracker(fd);
1556         if (ret < 0) {
1557                 ret = close(fd);
1558                 if (ret) {
1559                         PERROR("close on sock_info->notify_socket");
1560                 }
1561                 ret = -1;
1562                 lttng_ust_unlock_fd_tracker();
1563                 ust_unlock();
1564                 goto quit;
1565         }
1566
1567         sock_info->notify_socket = ret;
1568         lttng_ust_unlock_fd_tracker();
1569
1570         ust_unlock();
1571         /*
1572          * Unlock/relock ust lock because connect is blocking (with
1573          * timeout). Don't delay constructors on the ust lock for too
1574          * long.
1575          */
1576         if (ust_lock()) {
1577                 goto quit;
1578         }
1579
1580         timeout = get_notify_sock_timeout();
1581         if (timeout >= 0) {
1582                 /*
1583                  * Give at least 10ms to sessiond to reply to
1584                  * notifications.
1585                  */
1586                 if (timeout < 10)
1587                         timeout = 10;
1588                 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1589                                 timeout);
1590                 if (ret < 0) {
1591                         WARN("Error setting socket receive timeout");
1592                 }
1593                 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1594                                 timeout);
1595                 if (ret < 0) {
1596                         WARN("Error setting socket send timeout");
1597                 }
1598         } else if (timeout < -1) {
1599                 WARN("Unsupported timeout value %ld", timeout);
1600         }
1601
1602         ret = register_to_sessiond(sock_info->notify_socket,
1603                         USTCTL_SOCKET_NOTIFY);
1604         if (ret < 0) {
1605                 ERR("Error registering to %s ust notify socket",
1606                         sock_info->name);
1607                 prev_connect_failed = 1;
1608                 /*
1609                  * If we cannot register to the sessiond daemon, don't
1610                  * delay constructor execution.
1611                  */
1612                 ret = handle_register_done(sock_info);
1613                 assert(!ret);
1614                 ust_unlock();
1615                 goto restart;
1616         }
1617         sock = sock_info->socket;
1618
1619         ust_unlock();
1620
1621         for (;;) {
1622                 ssize_t len;
1623                 struct ustcomm_ust_msg lum;
1624
1625                 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1626                 switch (len) {
1627                 case 0: /* orderly shutdown */
1628                         DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1629                         if (ust_lock()) {
1630                                 goto quit;
1631                         }
1632                         /*
1633                          * Either sessiond has shutdown or refused us by closing the socket.
1634                          * In either case, we don't want to delay construction execution,
1635                          * and we need to wait before retry.
1636                          */
1637                         prev_connect_failed = 1;
1638                         /*
1639                          * If we cannot register to the sessiond daemon, don't
1640                          * delay constructor execution.
1641                          */
1642                         ret = handle_register_done(sock_info);
1643                         assert(!ret);
1644                         ust_unlock();
1645                         goto end;
1646                 case sizeof(lum):
1647                         print_cmd(lum.cmd, lum.handle);
1648                         ret = handle_message(sock_info, sock, &lum);
1649                         if (ret) {
1650                                 ERR("Error handling message for %s socket",
1651                                         sock_info->name);
1652                                 /*
1653                                  * Close socket if protocol error is
1654                                  * detected.
1655                                  */
1656                                 goto end;
1657                         }
1658                         continue;
1659                 default:
1660                         if (len < 0) {
1661                                 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1662                         } else {
1663                                 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1664                         }
1665                         if (len == -ECONNRESET) {
1666                                 DBG("%s remote end closed connection", sock_info->name);
1667                                 goto end;
1668                         }
1669                         goto end;
1670                 }
1671
1672         }
1673 end:
1674         if (ust_lock()) {
1675                 goto quit;
1676         }
1677         /* Cleanup socket handles before trying to reconnect */
1678         lttng_ust_objd_table_owner_cleanup(sock_info);
1679         ust_unlock();
1680         goto restart;   /* try to reconnect */
1681
1682 quit:
1683         ust_unlock();
1684
1685         pthread_mutex_lock(&ust_exit_mutex);
1686         sock_info->thread_active = 0;
1687         pthread_mutex_unlock(&ust_exit_mutex);
1688         return NULL;
1689 }
1690
1691 /*
1692  * Weak symbol to call when the ust malloc wrapper is not loaded.
1693  */
1694 __attribute__((weak))
1695 void lttng_ust_malloc_wrapper_init(void)
1696 {
1697 }
1698
1699 /*
1700  * sessiond monitoring thread: monitor presence of global and per-user
1701  * sessiond by polling the application common named pipe.
1702  */
1703 void __attribute__((constructor)) lttng_ust_init(void)
1704 {
1705         struct timespec constructor_timeout;
1706         sigset_t sig_all_blocked, orig_parent_mask;
1707         pthread_attr_t thread_attr;
1708         int timeout_mode;
1709         int ret;
1710
1711         if (uatomic_xchg(&initialized, 1) == 1)
1712                 return;
1713
1714         /*
1715          * Fixup interdependency between TLS fixup mutex (which happens
1716          * to be the dynamic linker mutex) and ust_lock, taken within
1717          * the ust lock.
1718          */
1719         lttng_ust_fixup_tls();
1720
1721         lttng_ust_loaded = 1;
1722
1723         /*
1724          * We want precise control over the order in which we construct
1725          * our sub-libraries vs starting to receive commands from
1726          * sessiond (otherwise leading to errors when trying to create
1727          * sessiond before the init functions are completed).
1728          */
1729         init_usterr();
1730         lttng_ust_getenv_init();        /* Needs init_usterr() to be completed. */
1731         init_tracepoint();
1732         lttng_ust_init_fd_tracker();
1733         lttng_ust_clock_init();
1734         lttng_ust_getcpu_init();
1735         lttng_ust_statedump_init();
1736         lttng_ring_buffer_metadata_client_init();
1737         lttng_ring_buffer_client_overwrite_init();
1738         lttng_ring_buffer_client_overwrite_rt_init();
1739         lttng_ring_buffer_client_discard_init();
1740         lttng_ring_buffer_client_discard_rt_init();
1741         lttng_perf_counter_init();
1742         /*
1743          * Invoke ust malloc wrapper init before starting other threads.
1744          */
1745         lttng_ust_malloc_wrapper_init();
1746
1747         timeout_mode = get_constructor_timeout(&constructor_timeout);
1748
1749         ret = sem_init(&constructor_wait, 0, 0);
1750         if (ret) {
1751                 PERROR("sem_init");
1752         }
1753
1754         ret = setup_global_apps();
1755         if (ret) {
1756                 assert(global_apps.allowed == 0);
1757                 DBG("global apps setup returned %d", ret);
1758         }
1759
1760         ret = setup_local_apps();
1761         if (ret) {
1762                 assert(local_apps.allowed == 0);
1763                 DBG("local apps setup returned %d", ret);
1764         }
1765
1766         /* A new thread created by pthread_create inherits the signal mask
1767          * from the parent. To avoid any signal being received by the
1768          * listener thread, we block all signals temporarily in the parent,
1769          * while we create the listener thread.
1770          */
1771         sigfillset(&sig_all_blocked);
1772         ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1773         if (ret) {
1774                 ERR("pthread_sigmask: %s", strerror(ret));
1775         }
1776
1777         ret = pthread_attr_init(&thread_attr);
1778         if (ret) {
1779                 ERR("pthread_attr_init: %s", strerror(ret));
1780         }
1781         ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1782         if (ret) {
1783                 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1784         }
1785
1786         if (global_apps.allowed) {
1787                 pthread_mutex_lock(&ust_exit_mutex);
1788                 ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1789                                 ust_listener_thread, &global_apps);
1790                 if (ret) {
1791                         ERR("pthread_create global: %s", strerror(ret));
1792                 }
1793                 global_apps.thread_active = 1;
1794                 pthread_mutex_unlock(&ust_exit_mutex);
1795         } else {
1796                 handle_register_done(&global_apps);
1797         }
1798
1799         if (local_apps.allowed) {
1800                 pthread_mutex_lock(&ust_exit_mutex);
1801                 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1802                                 ust_listener_thread, &local_apps);
1803                 if (ret) {
1804                         ERR("pthread_create local: %s", strerror(ret));
1805                 }
1806                 local_apps.thread_active = 1;
1807                 pthread_mutex_unlock(&ust_exit_mutex);
1808         } else {
1809                 handle_register_done(&local_apps);
1810         }
1811         ret = pthread_attr_destroy(&thread_attr);
1812         if (ret) {
1813                 ERR("pthread_attr_destroy: %s", strerror(ret));
1814         }
1815
1816         /* Restore original signal mask in parent */
1817         ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1818         if (ret) {
1819                 ERR("pthread_sigmask: %s", strerror(ret));
1820         }
1821
1822         switch (timeout_mode) {
1823         case 1: /* timeout wait */
1824                 do {
1825                         ret = sem_timedwait(&constructor_wait,
1826                                         &constructor_timeout);
1827                 } while (ret < 0 && errno == EINTR);
1828                 if (ret < 0) {
1829                         switch (errno) {
1830                         case ETIMEDOUT:
1831                                 ERR("Timed out waiting for lttng-sessiond");
1832                                 break;
1833                         case EINVAL:
1834                                 PERROR("sem_timedwait");
1835                                 break;
1836                         default:
1837                                 ERR("Unexpected error \"%s\" returned by sem_timedwait",
1838                                         strerror(errno));
1839                         }
1840                 }
1841                 break;
1842         case -1:/* wait forever */
1843                 do {
1844                         ret = sem_wait(&constructor_wait);
1845                 } while (ret < 0 && errno == EINTR);
1846                 if (ret < 0) {
1847                         switch (errno) {
1848                         case EINVAL:
1849                                 PERROR("sem_wait");
1850                                 break;
1851                         default:
1852                                 ERR("Unexpected error \"%s\" returned by sem_wait",
1853                                         strerror(errno));
1854                         }
1855                 }
1856                 break;
1857         case 0: /* no timeout */
1858                 break;
1859         }
1860 }
1861
1862 static
1863 void lttng_ust_cleanup(int exiting)
1864 {
1865         cleanup_sock_info(&global_apps, exiting);
1866         cleanup_sock_info(&local_apps, exiting);
1867         local_apps.allowed = 0;
1868         global_apps.allowed = 0;
1869         /*
1870          * The teardown in this function all affect data structures
1871          * accessed under the UST lock by the listener thread. This
1872          * lock, along with the lttng_ust_comm_should_quit flag, ensure
1873          * that none of these threads are accessing this data at this
1874          * point.
1875          */
1876         lttng_ust_abi_exit();
1877         lttng_ust_events_exit();
1878         lttng_perf_counter_exit();
1879         lttng_ring_buffer_client_discard_rt_exit();
1880         lttng_ring_buffer_client_discard_exit();
1881         lttng_ring_buffer_client_overwrite_rt_exit();
1882         lttng_ring_buffer_client_overwrite_exit();
1883         lttng_ring_buffer_metadata_client_exit();
1884         lttng_ust_statedump_destroy();
1885         exit_tracepoint();
1886         if (!exiting) {
1887                 /* Reinitialize values for fork */
1888                 sem_count = 2;
1889                 lttng_ust_comm_should_quit = 0;
1890                 initialized = 0;
1891         }
1892 }
1893
1894 void __attribute__((destructor)) lttng_ust_exit(void)
1895 {
1896         int ret;
1897
1898         /*
1899          * Using pthread_cancel here because:
1900          * A) we don't want to hang application teardown.
1901          * B) the thread is not allocating any resource.
1902          */
1903
1904         /*
1905          * Require the communication thread to quit. Synchronize with
1906          * mutexes to ensure it is not in a mutex critical section when
1907          * pthread_cancel is later called.
1908          */
1909         ust_lock_nocheck();
1910         lttng_ust_comm_should_quit = 1;
1911         ust_unlock();
1912
1913         pthread_mutex_lock(&ust_exit_mutex);
1914         /* cancel threads */
1915         if (global_apps.thread_active) {
1916                 ret = pthread_cancel(global_apps.ust_listener);
1917                 if (ret) {
1918                         ERR("Error cancelling global ust listener thread: %s",
1919                                 strerror(ret));
1920                 } else {
1921                         global_apps.thread_active = 0;
1922                 }
1923         }
1924         if (local_apps.thread_active) {
1925                 ret = pthread_cancel(local_apps.ust_listener);
1926                 if (ret) {
1927                         ERR("Error cancelling local ust listener thread: %s",
1928                                 strerror(ret));
1929                 } else {
1930                         local_apps.thread_active = 0;
1931                 }
1932         }
1933         pthread_mutex_unlock(&ust_exit_mutex);
1934
1935         /*
1936          * Do NOT join threads: use of sys_futex makes it impossible to
1937          * join the threads without using async-cancel, but async-cancel
1938          * is delivered by a signal, which could hit the target thread
1939          * anywhere in its code path, including while the ust_lock() is
1940          * held, causing a deadlock for the other thread. Let the OS
1941          * cleanup the threads if there are stalled in a syscall.
1942          */
1943         lttng_ust_cleanup(1);
1944 }
1945
1946 /*
1947  * We exclude the worker threads across fork and clone (except
1948  * CLONE_VM), because these system calls only keep the forking thread
1949  * running in the child.  Therefore, we don't want to call fork or clone
1950  * in the middle of an tracepoint or ust tracing state modification.
1951  * Holding this mutex protects these structures across fork and clone.
1952  */
1953 void ust_before_fork(sigset_t *save_sigset)
1954 {
1955         /*
1956          * Disable signals. This is to avoid that the child intervenes
1957          * before it is properly setup for tracing. It is safer to
1958          * disable all signals, because then we know we are not breaking
1959          * anything by restoring the original mask.
1960          */
1961         sigset_t all_sigs;
1962         int ret;
1963
1964         /* Fixup lttng-ust TLS. */
1965         lttng_ust_fixup_tls();
1966
1967         if (URCU_TLS(lttng_ust_nest_count))
1968                 return;
1969         /* Disable signals */
1970         sigfillset(&all_sigs);
1971         ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1972         if (ret == -1) {
1973                 PERROR("sigprocmask");
1974         }
1975
1976         pthread_mutex_lock(&ust_fork_mutex);
1977
1978         ust_lock_nocheck();
1979         rcu_bp_before_fork();
1980 }
1981
1982 static void ust_after_fork_common(sigset_t *restore_sigset)
1983 {
1984         int ret;
1985
1986         DBG("process %d", getpid());
1987         ust_unlock();
1988
1989         pthread_mutex_unlock(&ust_fork_mutex);
1990
1991         /* Restore signals */
1992         ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1993         if (ret == -1) {
1994                 PERROR("sigprocmask");
1995         }
1996 }
1997
1998 void ust_after_fork_parent(sigset_t *restore_sigset)
1999 {
2000         if (URCU_TLS(lttng_ust_nest_count))
2001                 return;
2002         DBG("process %d", getpid());
2003         rcu_bp_after_fork_parent();
2004         /* Release mutexes and reenable signals */
2005         ust_after_fork_common(restore_sigset);
2006 }
2007
2008 /*
2009  * After fork, in the child, we need to cleanup all the leftover state,
2010  * except the worker thread which already magically disappeared thanks
2011  * to the weird Linux fork semantics. After tyding up, we call
2012  * lttng_ust_init() again to start over as a new PID.
2013  *
2014  * This is meant for forks() that have tracing in the child between the
2015  * fork and following exec call (if there is any).
2016  */
2017 void ust_after_fork_child(sigset_t *restore_sigset)
2018 {
2019         if (URCU_TLS(lttng_ust_nest_count))
2020                 return;
2021         lttng_context_vpid_reset();
2022         lttng_context_vtid_reset();
2023         lttng_context_procname_reset();
2024         DBG("process %d", getpid());
2025         /* Release urcu mutexes */
2026         rcu_bp_after_fork_child();
2027         lttng_ust_cleanup(0);
2028         /* Release mutexes and reenable signals */
2029         ust_after_fork_common(restore_sigset);
2030         lttng_ust_init();
2031 }
2032
2033 void lttng_ust_sockinfo_session_enabled(void *owner)
2034 {
2035         struct sock_info *sock_info = owner;
2036         sock_info->statedump_pending = 1;
2037 }