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-events.h>
  40 #include <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 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
  57 DEFINE_TRACE(lttng_statedump_process_cgroup_ns);
  58 #endif
  59 DEFINE_TRACE(lttng_statedump_process_ipc_ns);
  60 #ifndef LTTNG_MNT_NS_MISSING_HEADER
  61 DEFINE_TRACE(lttng_statedump_process_mnt_ns);
  62 #endif
  63 DEFINE_TRACE(lttng_statedump_process_net_ns);
  64 DEFINE_TRACE(lttng_statedump_process_user_ns);
  65 DEFINE_TRACE(lttng_statedump_process_uts_ns);
  66 DEFINE_TRACE(lttng_statedump_network_interface);
  67 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
  68 DEFINE_TRACE(lttng_statedump_cpu_topology);
  69 #endif
  70
  71 struct lttng_fd_ctx {
  72         char *page;
  73         struct lttng_session *session;
  74         struct files_struct *files;
  75 };
  76
  77 /*
  78  * Protected by the trace lock.
  79  */
  80 static struct delayed_work cpu_work[NR_CPUS];
  81 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
  82 static atomic_t kernel_threads_to_run;
  83
  84 enum lttng_thread_type {
  85         LTTNG_USER_THREAD = 0,
  86         LTTNG_KERNEL_THREAD = 1,
  87 };
  88
  89 enum lttng_execution_mode {
  90         LTTNG_USER_MODE = 0,
  91         LTTNG_SYSCALL = 1,
  92         LTTNG_TRAP = 2,
  93         LTTNG_IRQ = 3,
  94         LTTNG_SOFTIRQ = 4,
  95         LTTNG_MODE_UNKNOWN = 5,
  96 };
  97
  98 enum lttng_execution_submode {
  99         LTTNG_NONE = 0,
 100         LTTNG_UNKNOWN = 1,
 101 };
 102
 103 enum lttng_process_status {
 104         LTTNG_UNNAMED = 0,
 105         LTTNG_WAIT_FORK = 1,
 106         LTTNG_WAIT_CPU = 2,
 107         LTTNG_EXIT = 3,
 108         LTTNG_ZOMBIE = 4,
 109         LTTNG_WAIT = 5,
 110         LTTNG_RUN = 6,
 111         LTTNG_DEAD = 7,
 112 };
 113
 114 static
 115 int lttng_enumerate_block_devices(struct lttng_session *session)
 116 {
 117         struct class_dev_iter iter;
 118         struct device *dev;
 119
 120         class_dev_iter_init(&iter, gendisk_block_class(), NULL,
 121                             gendisk_device_type());
 122         while ((dev = class_dev_iter_next(&iter))) {
 123                 struct disk_part_iter piter;
 124                 struct gendisk *disk = dev_to_disk(dev);
 125                 struct hd_struct *part;
 126
 127                 /*
 128                  * Don't show empty devices or things that have been
 129                  * suppressed
 130                  */
 131                 if (get_capacity(disk) == 0 ||
 132                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 133                         continue;
 134
 135                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 136                 while ((part = disk_part_iter_next(&piter))) {
 137                         char name_buf[BDEVNAME_SIZE];
 138                         char *p;
 139
 140                         p = gendisk_name(disk, part->partno, name_buf);
 141                         if (!p) {
 142                                 disk_part_iter_exit(&piter);
 143                                 class_dev_iter_exit(&iter);
 144                                 return -ENOSYS;
 145                         }
 146                         trace_lttng_statedump_block_device(session,
 147                                         part_devt(part), name_buf);
 148                 }
 149                 disk_part_iter_exit(&piter);
 150         }
 151         class_dev_iter_exit(&iter);
 152         return 0;
 153 }
 154
 155 #ifdef CONFIG_INET
 156
 157 static
 158 void lttng_enumerate_device(struct lttng_session *session,
 159                 struct net_device *dev)
 160 {
 161         struct in_device *in_dev;
 162         struct in_ifaddr *ifa;
 163
 164         if (dev->flags & IFF_UP) {
 165                 in_dev = in_dev_get(dev);
 166                 if (in_dev) {
 167                         for (ifa = in_dev->ifa_list; ifa != NULL;
 168                              ifa = ifa->ifa_next) {
 169                                 trace_lttng_statedump_network_interface(
 170                                         session, dev, ifa);
 171                         }
 172                         in_dev_put(in_dev);
 173                 }
 174         } else {
 175                 trace_lttng_statedump_network_interface(
 176                         session, dev, NULL);
 177         }
 178 }
 179
 180 static
 181 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 182 {
 183         struct net_device *dev;
 184
 185         read_lock(&dev_base_lock);
 186         for_each_netdev(&init_net, dev)
 187                 lttng_enumerate_device(session, dev);
 188         read_unlock(&dev_base_lock);
 189
 190         return 0;
 191 }
 192 #else /* CONFIG_INET */
 193 static inline
 194 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 195 {
 196         return 0;
 197 }
 198 #endif /* CONFIG_INET */
 199
 200 static
 201 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 202 {
 203         const struct lttng_fd_ctx *ctx = p;
 204         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 205         unsigned int flags = file->f_flags;
 206         struct fdtable *fdt;
 207
 208         /*
 209          * We don't expose kernel internal flags, only userspace-visible
 210          * flags.
 211          */
 212         flags &= ~FMODE_NONOTIFY;
 213         fdt = files_fdtable(ctx->files);
 214         /*
 215          * We need to check here again whether fd is within the fdt
 216          * max_fds range, because we might be seeing a different
 217          * files_fdtable() than iterate_fd(), assuming only RCU is
 218          * protecting the read. In reality, iterate_fd() holds
 219          * file_lock, which should ensure the fdt does not change while
 220          * the lock is taken, but we are not aware whether this is
 221          * guaranteed or not, so play safe.
 222          */
 223         if (fd < fdt->max_fds && close_on_exec(fd, fdt))
 224                 flags |= O_CLOEXEC;
 225         if (IS_ERR(s)) {
 226                 struct dentry *dentry = file->f_path.dentry;
 227
 228                 /* Make sure we give at least some info */
 229                 spin_lock(&dentry->d_lock);
 230                 trace_lttng_statedump_file_descriptor(ctx->session,
 231                         ctx->files, fd, dentry->d_name.name, flags,
 232                         file->f_mode);
 233                 spin_unlock(&dentry->d_lock);
 234                 goto end;
 235         }
 236         trace_lttng_statedump_file_descriptor(ctx->session,
 237                 ctx->files, fd, s, flags, file->f_mode);
 238 end:
 239         return 0;
 240 }
 241
 242 /* Called with task lock held. */
 243 static
 244 void lttng_enumerate_files(struct lttng_session *session,
 245                 struct files_struct *files,
 246                 char *tmp)
 247 {
 248         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .files = files, };
 249
 250         iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 251 }
 252
 253 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 254 static
 255 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 256 {
 257         int cpu;
 258         const cpumask_t *cpumask = cpu_possible_mask;
 259
 260         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 261                         cpu = cpumask_next(cpu, cpumask)) {
 262                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 263         }
 264
 265         return 0;
 266 }
 267 #else
 268 static
 269 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 270 {
 271         return 0;
 272 }
 273 #endif
 274
 275 #if 0
 276 /*
 277  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 278  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 279  * iteration, but it is not exported to modules.
 280  */
 281 static
 282 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 283                 struct task_struct *p)
 284 {
 285         struct mm_struct *mm;
 286         struct vm_area_struct *map;
 287         unsigned long ino;
 288
 289         /* get_task_mm does a task_lock... */
 290         mm = get_task_mm(p);
 291         if (!mm)
 292                 return;
 293
 294         map = mm->mmap;
 295         if (map) {
 296                 down_read(&mm->mmap_sem);
 297                 while (map) {
 298                         if (map->vm_file)
 299                                 ino = map->vm_file->f_path.dentry->d_inode->i_ino;
 300                         else
 301                                 ino = 0;
 302                         trace_lttng_statedump_vm_map(session, p, map, ino);
 303                         map = map->vm_next;
 304                 }
 305                 up_read(&mm->mmap_sem);
 306         }
 307         mmput(mm);
 308 }
 309
 310 static
 311 int lttng_enumerate_vm_maps(struct lttng_session *session)
 312 {
 313         struct task_struct *p;
 314
 315         rcu_read_lock();
 316         for_each_process(p)
 317                 lttng_enumerate_task_vm_maps(session, p);
 318         rcu_read_unlock();
 319         return 0;
 320 }
 321 #endif
 322
 323 static
 324 int lttng_list_interrupts(struct lttng_session *session)
 325 {
 326         unsigned int irq;
 327         unsigned long flags = 0;
 328         struct irq_desc *desc;
 329
 330         /* needs irq_desc */
 331         for_each_irq_desc(irq, desc) {
 332                 struct irqaction *action;
 333                 const char *irq_chip_name =
 334                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 335
 336                 local_irq_save(flags);
 337                 raw_spin_lock(&desc->lock);
 338                 for (action = desc->action; action; action = action->next) {
 339                         trace_lttng_statedump_interrupt(session,
 340                                 irq, irq_chip_name, action);
 341                 }
 342                 raw_spin_unlock(&desc->lock);
 343                 local_irq_restore(flags);
 344         }
 345         return 0;
 346 }
 347
 348 /*
 349  * Statedump the task's namespaces using the proc filesystem inode number as
 350  * the unique identifier. The user and pid ns are nested and will be dumped
 351  * recursively.
 352  *
 353  * Called with task lock held.
 354  */
 355 static
 356 void lttng_statedump_process_ns(struct lttng_session *session,
 357                 struct task_struct *p,
 358                 enum lttng_thread_type type,
 359                 enum lttng_execution_mode mode,
 360                 enum lttng_execution_submode submode,
 361                 enum lttng_process_status status)
 362 {
 363         struct nsproxy *proxy;
 364         struct pid_namespace *pid_ns;
 365         struct user_namespace *user_ns;
 366
 367         /*
 368          * The pid and user namespaces are special, they are nested and
 369          * accessed with specific functions instead of the nsproxy struct
 370          * like the other namespaces.
 371          */
 372         pid_ns = task_active_pid_ns(p);
 373         do {
 374                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 375                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 376         } while (pid_ns);
 377
 378
 379         user_ns = task_cred_xxx(p, user_ns);
 380         do {
 381                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 382                 /*
 383                  * trace_lttng_statedump_process_user_ns() internally
 384                  * checks whether user_ns is NULL. While this does not
 385                  * appear to be a possible return value for
 386                  * task_cred_xxx(), err on the safe side and check
 387                  * for NULL here as well to be consistent with the
 388                  * paranoid behavior of
 389                  * trace_lttng_statedump_process_user_ns().
 390                  */
 391                 user_ns = user_ns ? user_ns->parent : NULL;
 392         } while (user_ns);
 393
 394         /*
 395          * Back and forth on locking strategy within Linux upstream for nsproxy.
 396          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 397          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 398          * for details.
 399          */
 400 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 401                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 402                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 403                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 404         proxy = p->nsproxy;
 405 #else
 406         rcu_read_lock();
 407         proxy = task_nsproxy(p);
 408 #endif
 409         if (proxy) {
 410 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
 411                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 412 #endif
 413                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 414 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 415                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 416 #endif
 417                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 418                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 419         }
 420 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 421                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 422                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 423                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 424         /* (nothing) */
 425 #else
 426         rcu_read_unlock();
 427 #endif
 428 }
 429
 430 static
 431 int lttng_enumerate_process_states(struct lttng_session *session)
 432 {
 433         struct task_struct *g, *p;
 434         char *tmp;
 435
 436         tmp = (char *) __get_free_page(GFP_KERNEL);
 437         if (!tmp)
 438                 return -ENOMEM;
 439
 440         rcu_read_lock();
 441         for_each_process(g) {
 442                 struct files_struct *prev_files = NULL;
 443
 444                 p = g;
 445                 do {
 446                         enum lttng_execution_mode mode =
 447                                 LTTNG_MODE_UNKNOWN;
 448                         enum lttng_execution_submode submode =
 449                                 LTTNG_UNKNOWN;
 450                         enum lttng_process_status status;
 451                         enum lttng_thread_type type;
 452                         struct files_struct *files;
 453
 454                         task_lock(p);
 455                         if (p->exit_state == EXIT_ZOMBIE)
 456                                 status = LTTNG_ZOMBIE;
 457                         else if (p->exit_state == EXIT_DEAD)
 458                                 status = LTTNG_DEAD;
 459                         else if (p->state == TASK_RUNNING) {
 460                                 /* Is this a forked child that has not run yet? */
 461                                 if (list_empty(&p->rt.run_list))
 462                                         status = LTTNG_WAIT_FORK;
 463                                 else
 464                                         /*
 465                                          * All tasks are considered as wait_cpu;
 466                                          * the viewer will sort out if the task
 467                                          * was really running at this time.
 468                                          */
 469                                         status = LTTNG_WAIT_CPU;
 470                         } else if (p->state &
 471                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 472                                 /* Task is waiting for something to complete */
 473                                 status = LTTNG_WAIT;
 474                         } else
 475                                 status = LTTNG_UNNAMED;
 476                         submode = LTTNG_NONE;
 477
 478                         /*
 479                          * Verification of t->mm is to filter out kernel
 480                          * threads; Viewer will further filter out if a
 481                          * user-space thread was in syscall mode or not.
 482                          */
 483                         if (p->mm)
 484                                 type = LTTNG_USER_THREAD;
 485                         else
 486                                 type = LTTNG_KERNEL_THREAD;
 487                         files = p->files;
 488
 489                         trace_lttng_statedump_process_state(session,
 490                                 p, type, mode, submode, status, files);
 491                         lttng_statedump_process_ns(session,
 492                                 p, type, mode, submode, status);
 493                         /*
 494                          * As an optimisation for the common case, do not
 495                          * repeat information for the same files_struct in
 496                          * two consecutive threads. This is the common case
 497                          * for threads sharing the same fd table. RCU guarantees
 498                          * that the same files_struct pointer is not re-used
 499                          * throughout processes/threads iteration.
 500                          */
 501                         if (files && files != prev_files) {
 502                                 lttng_enumerate_files(session, files, tmp);
 503                                 prev_files = files;
 504                         }
 505                         task_unlock(p);
 506                 } while_each_thread(g, p);
 507         }
 508         rcu_read_unlock();
 509
 510         free_page((unsigned long) tmp);
 511
 512         return 0;
 513 }
 514
 515 static
 516 void lttng_statedump_work_func(struct work_struct *work)
 517 {
 518         if (atomic_dec_and_test(&kernel_threads_to_run))
 519                 /* If we are the last thread, wake up do_lttng_statedump */
 520                 wake_up(&statedump_wq);
 521 }
 522
 523 static
 524 int do_lttng_statedump(struct lttng_session *session)
 525 {
 526         int cpu, ret;
 527
 528         trace_lttng_statedump_start(session);
 529         ret = lttng_enumerate_process_states(session);
 530         if (ret)
 531                 return ret;
 532         /*
 533          * FIXME
 534          * ret = lttng_enumerate_vm_maps(session);
 535          * if (ret)
 536          *      return ret;
 537          */
 538         ret = lttng_list_interrupts(session);
 539         if (ret)
 540                 return ret;
 541         ret = lttng_enumerate_network_ip_interface(session);
 542         if (ret)
 543                 return ret;
 544         ret = lttng_enumerate_block_devices(session);
 545         switch (ret) {
 546         case 0:
 547                 break;
 548         case -ENOSYS:
 549                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 550                 break;
 551         default:
 552                 return ret;
 553         }
 554         ret = lttng_enumerate_cpu_topology(session);
 555         if (ret)
 556                 return ret;
 557
 558         /* TODO lttng_dump_idt_table(session); */
 559         /* TODO lttng_dump_softirq_vec(session); */
 560         /* TODO lttng_list_modules(session); */
 561         /* TODO lttng_dump_swap_files(session); */
 562
 563         /*
 564          * Fire off a work queue on each CPU. Their sole purpose in life
 565          * is to guarantee that each CPU has been in a state where is was in
 566          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 567          */
 568         get_online_cpus();
 569         atomic_set(&kernel_threads_to_run, num_online_cpus());
 570         for_each_online_cpu(cpu) {
 571                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 572                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 573         }
 574         /* Wait for all threads to run */
 575         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 576         put_online_cpus();
 577         /* Our work is done */
 578         trace_lttng_statedump_end(session);
 579         return 0;
 580 }
 581
 582 /*
 583  * Called with session mutex held.
 584  */
 585 int lttng_statedump_start(struct lttng_session *session)
 586 {
 587         return do_lttng_statedump(session);
 588 }
 589 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 590
 591 static
 592 int __init lttng_statedump_init(void)
 593 {
 594         return 0;
 595 }
 596
 597 module_init(lttng_statedump_init);
 598
 599 static
 600 void __exit lttng_statedump_exit(void)
 601 {
 602 }
 603
 604 module_exit(lttng_statedump_exit);
 605
 606 MODULE_LICENSE("GPL and additional rights");
 607 MODULE_AUTHOR("Jean-Hugues Deschenes");
 608 MODULE_DESCRIPTION("LTTng statedump provider");
 609 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 610         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 611         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 612         LTTNG_MODULES_EXTRAVERSION);