lttng-statedump-impl.c

   1 /* SPDX-License-Identifier: (GPL-2.0-only or LGPL-2.1-only)
   2  *
   3  * lttng-statedump.c
   4  *
   5  * Linux Trace Toolkit Next Generation Kernel State Dump
   6  *
   7  * Copyright 2005 Jean-Hugues Deschenes <jean-hugues.deschenes@polymtl.ca>
   8  * Copyright 2006-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
   9  *
  10  * Changes:
  11  *      Eric Clement:                   Add listing of network IP interface
  12  *      2006, 2007 Mathieu Desnoyers    Fix kernel threads
  13  *                                      Various updates
  14  */
  15
  16 #include <linux/init.h>
  17 #include <linux/module.h>
  18 #include <linux/netlink.h>
  19 #include <linux/inet.h>
  20 #include <linux/ip.h>
  21 #include <linux/kthread.h>
  22 #include <linux/proc_fs.h>
  23 #include <linux/file.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/irqnr.h>
  26 #include <linux/cpu.h>
  27 #include <linux/netdevice.h>
  28 #include <linux/inetdevice.h>
  29 #include <linux/sched.h>
  30 #include <linux/mm.h>
  31 #include <linux/swap.h>
  32 #include <linux/wait.h>
  33 #include <linux/mutex.h>
  34 #include <linux/device.h>
  35 #include <linux/fdtable.h>
  36 #include <linux/irq.h>
  37 #include <linux/genhd.h>
  38
  39 #include <lttng/lttng-events.h>
  40 #include <lttng/lttng-tracer.h>
  41
  42 /* Define the tracepoints, but do not build the probes */
  43 #define CREATE_TRACE_POINTS
  44 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
  45 #define TRACE_INCLUDE_FILE lttng-statedump
  46 #define LTTNG_INSTRUMENTATION
  47 #include <instrumentation/events/lttng-module/lttng-statedump.h>
  48
  49 DEFINE_TRACE(lttng_statedump_block_device);
  50 DEFINE_TRACE(lttng_statedump_end);
  51 DEFINE_TRACE(lttng_statedump_interrupt);
  52 DEFINE_TRACE(lttng_statedump_file_descriptor);
  53 DEFINE_TRACE(lttng_statedump_start);
  54 DEFINE_TRACE(lttng_statedump_process_state);
  55 DEFINE_TRACE(lttng_statedump_process_pid_ns);
  56 DEFINE_TRACE(lttng_statedump_process_cgroup_ns);
  57 DEFINE_TRACE(lttng_statedump_process_ipc_ns);
  58 #ifndef LTTNG_MNT_NS_MISSING_HEADER
  59 DEFINE_TRACE(lttng_statedump_process_mnt_ns);
  60 #endif
  61 DEFINE_TRACE(lttng_statedump_process_net_ns);
  62 DEFINE_TRACE(lttng_statedump_process_user_ns);
  63 DEFINE_TRACE(lttng_statedump_process_uts_ns);
  64 DEFINE_TRACE(lttng_statedump_network_interface);
  65 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
  66 DEFINE_TRACE(lttng_statedump_cpu_topology);
  67 #endif
  68
  69 struct lttng_fd_ctx {
  70         char *page;
  71         struct lttng_session *session;
  72         struct files_struct *files;
  73 };
  74
  75 /*
  76  * Protected by the trace lock.
  77  */
  78 static struct delayed_work cpu_work[NR_CPUS];
  79 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
  80 static atomic_t kernel_threads_to_run;
  81
  82 enum lttng_thread_type {
  83         LTTNG_USER_THREAD = 0,
  84         LTTNG_KERNEL_THREAD = 1,
  85 };
  86
  87 enum lttng_execution_mode {
  88         LTTNG_USER_MODE = 0,
  89         LTTNG_SYSCALL = 1,
  90         LTTNG_TRAP = 2,
  91         LTTNG_IRQ = 3,
  92         LTTNG_SOFTIRQ = 4,
  93         LTTNG_MODE_UNKNOWN = 5,
  94 };
  95
  96 enum lttng_execution_submode {
  97         LTTNG_NONE = 0,
  98         LTTNG_UNKNOWN = 1,
  99 };
 100
 101 enum lttng_process_status {
 102         LTTNG_UNNAMED = 0,
 103         LTTNG_WAIT_FORK = 1,
 104         LTTNG_WAIT_CPU = 2,
 105         LTTNG_EXIT = 3,
 106         LTTNG_ZOMBIE = 4,
 107         LTTNG_WAIT = 5,
 108         LTTNG_RUN = 6,
 109         LTTNG_DEAD = 7,
 110 };
 111
 112 static
 113 int lttng_enumerate_block_devices(struct lttng_session *session)
 114 {
 115         struct class_dev_iter iter;
 116         struct device *dev;
 117
 118         class_dev_iter_init(&iter, gendisk_block_class(), NULL,
 119                             gendisk_device_type());
 120         while ((dev = class_dev_iter_next(&iter))) {
 121                 struct disk_part_iter piter;
 122                 struct gendisk *disk = dev_to_disk(dev);
 123                 struct hd_struct *part;
 124
 125                 /*
 126                  * Don't show empty devices or things that have been
 127                  * suppressed
 128                  */
 129                 if (get_capacity(disk) == 0 ||
 130                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 131                         continue;
 132
 133                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 134                 while ((part = disk_part_iter_next(&piter))) {
 135                         char name_buf[BDEVNAME_SIZE];
 136                         char *p;
 137
 138                         p = gendisk_name(disk, part->partno, name_buf);
 139                         if (!p) {
 140                                 disk_part_iter_exit(&piter);
 141                                 class_dev_iter_exit(&iter);
 142                                 return -ENOSYS;
 143                         }
 144                         trace_lttng_statedump_block_device(session,
 145                                         part_devt(part), name_buf);
 146                 }
 147                 disk_part_iter_exit(&piter);
 148         }
 149         class_dev_iter_exit(&iter);
 150         return 0;
 151 }
 152
 153 #ifdef CONFIG_INET
 154
 155 static
 156 void lttng_enumerate_device(struct lttng_session *session,
 157                 struct net_device *dev)
 158 {
 159         struct in_device *in_dev;
 160         struct in_ifaddr *ifa;
 161
 162         if (dev->flags & IFF_UP) {
 163                 in_dev = in_dev_get(dev);
 164                 if (in_dev) {
 165                         for (ifa = in_dev->ifa_list; ifa != NULL;
 166                              ifa = ifa->ifa_next) {
 167                                 trace_lttng_statedump_network_interface(
 168                                         session, dev, ifa);
 169                         }
 170                         in_dev_put(in_dev);
 171                 }
 172         } else {
 173                 trace_lttng_statedump_network_interface(
 174                         session, dev, NULL);
 175         }
 176 }
 177
 178 static
 179 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 180 {
 181         struct net_device *dev;
 182
 183         read_lock(&dev_base_lock);
 184         for_each_netdev(&init_net, dev)
 185                 lttng_enumerate_device(session, dev);
 186         read_unlock(&dev_base_lock);
 187
 188         return 0;
 189 }
 190 #else /* CONFIG_INET */
 191 static inline
 192 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 193 {
 194         return 0;
 195 }
 196 #endif /* CONFIG_INET */
 197
 198 static
 199 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 200 {
 201         const struct lttng_fd_ctx *ctx = p;
 202         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 203         unsigned int flags = file->f_flags;
 204         struct fdtable *fdt;
 205
 206         /*
 207          * We don't expose kernel internal flags, only userspace-visible
 208          * flags.
 209          */
 210         flags &= ~FMODE_NONOTIFY;
 211         fdt = files_fdtable(ctx->files);
 212         /*
 213          * We need to check here again whether fd is within the fdt
 214          * max_fds range, because we might be seeing a different
 215          * files_fdtable() than iterate_fd(), assuming only RCU is
 216          * protecting the read. In reality, iterate_fd() holds
 217          * file_lock, which should ensure the fdt does not change while
 218          * the lock is taken, but we are not aware whether this is
 219          * guaranteed or not, so play safe.
 220          */
 221         if (fd < fdt->max_fds && close_on_exec(fd, fdt))
 222                 flags |= O_CLOEXEC;
 223         if (IS_ERR(s)) {
 224                 struct dentry *dentry = file->f_path.dentry;
 225
 226                 /* Make sure we give at least some info */
 227                 spin_lock(&dentry->d_lock);
 228                 trace_lttng_statedump_file_descriptor(ctx->session,
 229                         ctx->files, fd, dentry->d_name.name, flags,
 230                         file->f_mode);
 231                 spin_unlock(&dentry->d_lock);
 232                 goto end;
 233         }
 234         trace_lttng_statedump_file_descriptor(ctx->session,
 235                 ctx->files, fd, s, flags, file->f_mode);
 236 end:
 237         return 0;
 238 }
 239
 240 /* Called with task lock held. */
 241 static
 242 void lttng_enumerate_files(struct lttng_session *session,
 243                 struct files_struct *files,
 244                 char *tmp)
 245 {
 246         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .files = files, };
 247
 248         iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 249 }
 250
 251 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 252 static
 253 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 254 {
 255         int cpu;
 256         const cpumask_t *cpumask = cpu_possible_mask;
 257
 258         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 259                         cpu = cpumask_next(cpu, cpumask)) {
 260                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 261         }
 262
 263         return 0;
 264 }
 265 #else
 266 static
 267 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 268 {
 269         return 0;
 270 }
 271 #endif
 272
 273 #if 0
 274 /*
 275  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 276  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 277  * iteration, but it is not exported to modules.
 278  */
 279 static
 280 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 281                 struct task_struct *p)
 282 {
 283         struct mm_struct *mm;
 284         struct vm_area_struct *map;
 285         unsigned long ino;
 286
 287         /* get_task_mm does a task_lock... */
 288         mm = get_task_mm(p);
 289         if (!mm)
 290                 return;
 291
 292         map = mm->mmap;
 293         if (map) {
 294                 down_read(&mm->mmap_sem);
 295                 while (map) {
 296                         if (map->vm_file)
 297                                 ino = map->vm_file->f_path.dentry->d_inode->i_ino;
 298                         else
 299                                 ino = 0;
 300                         trace_lttng_statedump_vm_map(session, p, map, ino);
 301                         map = map->vm_next;
 302                 }
 303                 up_read(&mm->mmap_sem);
 304         }
 305         mmput(mm);
 306 }
 307
 308 static
 309 int lttng_enumerate_vm_maps(struct lttng_session *session)
 310 {
 311         struct task_struct *p;
 312
 313         rcu_read_lock();
 314         for_each_process(p)
 315                 lttng_enumerate_task_vm_maps(session, p);
 316         rcu_read_unlock();
 317         return 0;
 318 }
 319 #endif
 320
 321 static
 322 int lttng_list_interrupts(struct lttng_session *session)
 323 {
 324         unsigned int irq;
 325         unsigned long flags = 0;
 326         struct irq_desc *desc;
 327
 328         /* needs irq_desc */
 329         for_each_irq_desc(irq, desc) {
 330                 struct irqaction *action;
 331                 const char *irq_chip_name =
 332                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 333
 334                 local_irq_save(flags);
 335                 raw_spin_lock(&desc->lock);
 336                 for (action = desc->action; action; action = action->next) {
 337                         trace_lttng_statedump_interrupt(session,
 338                                 irq, irq_chip_name, action);
 339                 }
 340                 raw_spin_unlock(&desc->lock);
 341                 local_irq_restore(flags);
 342         }
 343         return 0;
 344 }
 345
 346 /*
 347  * Statedump the task's namespaces using the proc filesystem inode number as
 348  * the unique identifier. The user and pid ns are nested and will be dumped
 349  * recursively.
 350  *
 351  * Called with task lock held.
 352  */
 353 static
 354 void lttng_statedump_process_ns(struct lttng_session *session,
 355                 struct task_struct *p,
 356                 enum lttng_thread_type type,
 357                 enum lttng_execution_mode mode,
 358                 enum lttng_execution_submode submode,
 359                 enum lttng_process_status status)
 360 {
 361         struct nsproxy *proxy;
 362         struct pid_namespace *pid_ns;
 363         struct user_namespace *user_ns;
 364
 365         /*
 366          * The pid and user namespaces are special, they are nested and
 367          * accessed with specific functions instead of the nsproxy struct
 368          * like the other namespaces.
 369          */
 370         pid_ns = task_active_pid_ns(p);
 371         do {
 372                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 373                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 374         } while (pid_ns);
 375
 376
 377         user_ns = task_cred_xxx(p, user_ns);
 378         do {
 379                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 380                 /*
 381                  * trace_lttng_statedump_process_user_ns() internally
 382                  * checks whether user_ns is NULL. While this does not
 383                  * appear to be a possible return value for
 384                  * task_cred_xxx(), err on the safe side and check
 385                  * for NULL here as well to be consistent with the
 386                  * paranoid behavior of
 387                  * trace_lttng_statedump_process_user_ns().
 388                  */
 389                 user_ns = user_ns ? user_ns->parent : NULL;
 390         } while (user_ns);
 391
 392         /*
 393          * Back and forth on locking strategy within Linux upstream for nsproxy.
 394          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 395          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 396          * for details.
 397          */
 398         proxy = p->nsproxy;
 399         if (proxy) {
 400                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 401                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 402 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 403                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 404 #endif
 405                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 406                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 407         }
 408 }
 409
 410 static
 411 int lttng_enumerate_process_states(struct lttng_session *session)
 412 {
 413         struct task_struct *g, *p;
 414         char *tmp;
 415
 416         tmp = (char *) __get_free_page(GFP_KERNEL);
 417         if (!tmp)
 418                 return -ENOMEM;
 419
 420         rcu_read_lock();
 421         for_each_process(g) {
 422                 struct files_struct *prev_files = NULL;
 423
 424                 p = g;
 425                 do {
 426                         enum lttng_execution_mode mode =
 427                                 LTTNG_MODE_UNKNOWN;
 428                         enum lttng_execution_submode submode =
 429                                 LTTNG_UNKNOWN;
 430                         enum lttng_process_status status;
 431                         enum lttng_thread_type type;
 432                         struct files_struct *files;
 433
 434                         task_lock(p);
 435                         if (p->exit_state == EXIT_ZOMBIE)
 436                                 status = LTTNG_ZOMBIE;
 437                         else if (p->exit_state == EXIT_DEAD)
 438                                 status = LTTNG_DEAD;
 439                         else if (p->state == TASK_RUNNING) {
 440                                 /* Is this a forked child that has not run yet? */
 441                                 if (list_empty(&p->rt.run_list))
 442                                         status = LTTNG_WAIT_FORK;
 443                                 else
 444                                         /*
 445                                          * All tasks are considered as wait_cpu;
 446                                          * the viewer will sort out if the task
 447                                          * was really running at this time.
 448                                          */
 449                                         status = LTTNG_WAIT_CPU;
 450                         } else if (p->state &
 451                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 452                                 /* Task is waiting for something to complete */
 453                                 status = LTTNG_WAIT;
 454                         } else
 455                                 status = LTTNG_UNNAMED;
 456                         submode = LTTNG_NONE;
 457
 458                         /*
 459                          * Verification of t->mm is to filter out kernel
 460                          * threads; Viewer will further filter out if a
 461                          * user-space thread was in syscall mode or not.
 462                          */
 463                         if (p->mm)
 464                                 type = LTTNG_USER_THREAD;
 465                         else
 466                                 type = LTTNG_KERNEL_THREAD;
 467                         files = p->files;
 468
 469                         trace_lttng_statedump_process_state(session,
 470                                 p, type, mode, submode, status, files);
 471                         lttng_statedump_process_ns(session,
 472                                 p, type, mode, submode, status);
 473                         /*
 474                          * As an optimisation for the common case, do not
 475                          * repeat information for the same files_struct in
 476                          * two consecutive threads. This is the common case
 477                          * for threads sharing the same fd table. RCU guarantees
 478                          * that the same files_struct pointer is not re-used
 479                          * throughout processes/threads iteration.
 480                          */
 481                         if (files && files != prev_files) {
 482                                 lttng_enumerate_files(session, files, tmp);
 483                                 prev_files = files;
 484                         }
 485                         task_unlock(p);
 486                 } while_each_thread(g, p);
 487         }
 488         rcu_read_unlock();
 489
 490         free_page((unsigned long) tmp);
 491
 492         return 0;
 493 }
 494
 495 static
 496 void lttng_statedump_work_func(struct work_struct *work)
 497 {
 498         if (atomic_dec_and_test(&kernel_threads_to_run))
 499                 /* If we are the last thread, wake up do_lttng_statedump */
 500                 wake_up(&statedump_wq);
 501 }
 502
 503 static
 504 int do_lttng_statedump(struct lttng_session *session)
 505 {
 506         int cpu, ret;
 507
 508         trace_lttng_statedump_start(session);
 509         ret = lttng_enumerate_process_states(session);
 510         if (ret)
 511                 return ret;
 512         /*
 513          * FIXME
 514          * ret = lttng_enumerate_vm_maps(session);
 515          * if (ret)
 516          *      return ret;
 517          */
 518         ret = lttng_list_interrupts(session);
 519         if (ret)
 520                 return ret;
 521         ret = lttng_enumerate_network_ip_interface(session);
 522         if (ret)
 523                 return ret;
 524         ret = lttng_enumerate_block_devices(session);
 525         switch (ret) {
 526         case 0:
 527                 break;
 528         case -ENOSYS:
 529                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 530                 break;
 531         default:
 532                 return ret;
 533         }
 534         ret = lttng_enumerate_cpu_topology(session);
 535         if (ret)
 536                 return ret;
 537
 538         /* TODO lttng_dump_idt_table(session); */
 539         /* TODO lttng_dump_softirq_vec(session); */
 540         /* TODO lttng_list_modules(session); */
 541         /* TODO lttng_dump_swap_files(session); */
 542
 543         /*
 544          * Fire off a work queue on each CPU. Their sole purpose in life
 545          * is to guarantee that each CPU has been in a state where is was in
 546          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 547          */
 548         get_online_cpus();
 549         atomic_set(&kernel_threads_to_run, num_online_cpus());
 550         for_each_online_cpu(cpu) {
 551                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 552                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 553         }
 554         /* Wait for all threads to run */
 555         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 556         put_online_cpus();
 557         /* Our work is done */
 558         trace_lttng_statedump_end(session);
 559         return 0;
 560 }
 561
 562 /*
 563  * Called with session mutex held.
 564  */
 565 int lttng_statedump_start(struct lttng_session *session)
 566 {
 567         return do_lttng_statedump(session);
 568 }
 569 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 570
 571 static
 572 int __init lttng_statedump_init(void)
 573 {
 574         return 0;
 575 }
 576
 577 module_init(lttng_statedump_init);
 578
 579 static
 580 void __exit lttng_statedump_exit(void)
 581 {
 582 }
 583
 584 module_exit(lttng_statedump_exit);
 585
 586 MODULE_LICENSE("GPL and additional rights");
 587 MODULE_AUTHOR("Jean-Hugues Deschenes");
 588 MODULE_DESCRIPTION("LTTng statedump provider");
 589 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 590         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 591         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 592         LTTNG_MODULES_EXTRAVERSION);