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
  38 #include <lttng-events.h>
  39 #include <lttng-tracer.h>
  40 #include <wrapper/genhd.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 *ptr_block_class;
 118         struct device_type *ptr_disk_type;
 119         struct class_dev_iter iter;
 120         struct device *dev;
 121
 122         ptr_block_class = wrapper_get_block_class();
 123         if (!ptr_block_class)
 124                 return -ENOSYS;
 125         ptr_disk_type = wrapper_get_disk_type();
 126         if (!ptr_disk_type) {
 127                 return -ENOSYS;
 128         }
 129         class_dev_iter_init(&iter, ptr_block_class, NULL, ptr_disk_type);
 130         while ((dev = class_dev_iter_next(&iter))) {
 131                 struct disk_part_iter piter;
 132                 struct gendisk *disk = dev_to_disk(dev);
 133                 struct hd_struct *part;
 134
 135                 /*
 136                  * Don't show empty devices or things that have been
 137                  * suppressed
 138                  */
 139                 if (get_capacity(disk) == 0 ||
 140                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 141                         continue;
 142
 143                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 144                 while ((part = disk_part_iter_next(&piter))) {
 145                         char name_buf[BDEVNAME_SIZE];
 146                         char *p;
 147
 148                         p = wrapper_disk_name(disk, part->partno, name_buf);
 149                         if (!p) {
 150                                 disk_part_iter_exit(&piter);
 151                                 class_dev_iter_exit(&iter);
 152                                 return -ENOSYS;
 153                         }
 154                         trace_lttng_statedump_block_device(session,
 155                                         part_devt(part), name_buf);
 156                 }
 157                 disk_part_iter_exit(&piter);
 158         }
 159         class_dev_iter_exit(&iter);
 160         return 0;
 161 }
 162
 163 #ifdef CONFIG_INET
 164
 165 static
 166 void lttng_enumerate_device(struct lttng_session *session,
 167                 struct net_device *dev)
 168 {
 169         struct in_device *in_dev;
 170         struct in_ifaddr *ifa;
 171
 172         if (dev->flags & IFF_UP) {
 173                 in_dev = in_dev_get(dev);
 174                 if (in_dev) {
 175                         for (ifa = in_dev->ifa_list; ifa != NULL;
 176                              ifa = ifa->ifa_next) {
 177                                 trace_lttng_statedump_network_interface(
 178                                         session, dev, ifa);
 179                         }
 180                         in_dev_put(in_dev);
 181                 }
 182         } else {
 183                 trace_lttng_statedump_network_interface(
 184                         session, dev, NULL);
 185         }
 186 }
 187
 188 static
 189 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 190 {
 191         struct net_device *dev;
 192
 193         read_lock(&dev_base_lock);
 194         for_each_netdev(&init_net, dev)
 195                 lttng_enumerate_device(session, dev);
 196         read_unlock(&dev_base_lock);
 197
 198         return 0;
 199 }
 200 #else /* CONFIG_INET */
 201 static inline
 202 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 203 {
 204         return 0;
 205 }
 206 #endif /* CONFIG_INET */
 207
 208 static
 209 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 210 {
 211         const struct lttng_fd_ctx *ctx = p;
 212         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 213         unsigned int flags = file->f_flags;
 214         struct fdtable *fdt;
 215
 216         /*
 217          * We don't expose kernel internal flags, only userspace-visible
 218          * flags.
 219          */
 220         flags &= ~FMODE_NONOTIFY;
 221         fdt = files_fdtable(ctx->files);
 222         /*
 223          * We need to check here again whether fd is within the fdt
 224          * max_fds range, because we might be seeing a different
 225          * files_fdtable() than iterate_fd(), assuming only RCU is
 226          * protecting the read. In reality, iterate_fd() holds
 227          * file_lock, which should ensure the fdt does not change while
 228          * the lock is taken, but we are not aware whether this is
 229          * guaranteed or not, so play safe.
 230          */
 231         if (fd < fdt->max_fds && close_on_exec(fd, fdt))
 232                 flags |= O_CLOEXEC;
 233         if (IS_ERR(s)) {
 234                 struct dentry *dentry = file->f_path.dentry;
 235
 236                 /* Make sure we give at least some info */
 237                 spin_lock(&dentry->d_lock);
 238                 trace_lttng_statedump_file_descriptor(ctx->session,
 239                         ctx->files, fd, dentry->d_name.name, flags,
 240                         file->f_mode);
 241                 spin_unlock(&dentry->d_lock);
 242                 goto end;
 243         }
 244         trace_lttng_statedump_file_descriptor(ctx->session,
 245                 ctx->files, fd, s, flags, file->f_mode);
 246 end:
 247         return 0;
 248 }
 249
 250 /* Called with task lock held. */
 251 static
 252 void lttng_enumerate_files(struct lttng_session *session,
 253                 struct files_struct *files,
 254                 char *tmp)
 255 {
 256         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .files = files, };
 257
 258         iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 259 }
 260
 261 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 262 static
 263 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 264 {
 265         int cpu;
 266         const cpumask_t *cpumask = cpu_possible_mask;
 267
 268         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 269                         cpu = cpumask_next(cpu, cpumask)) {
 270                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 271         }
 272
 273         return 0;
 274 }
 275 #else
 276 static
 277 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 278 {
 279         return 0;
 280 }
 281 #endif
 282
 283 #if 0
 284 /*
 285  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 286  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 287  * iteration, but it is not exported to modules.
 288  */
 289 static
 290 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 291                 struct task_struct *p)
 292 {
 293         struct mm_struct *mm;
 294         struct vm_area_struct *map;
 295         unsigned long ino;
 296
 297         /* get_task_mm does a task_lock... */
 298         mm = get_task_mm(p);
 299         if (!mm)
 300                 return;
 301
 302         map = mm->mmap;
 303         if (map) {
 304                 down_read(&mm->mmap_sem);
 305                 while (map) {
 306                         if (map->vm_file)
 307                                 ino = map->vm_file->f_path.dentry->d_inode->i_ino;
 308                         else
 309                                 ino = 0;
 310                         trace_lttng_statedump_vm_map(session, p, map, ino);
 311                         map = map->vm_next;
 312                 }
 313                 up_read(&mm->mmap_sem);
 314         }
 315         mmput(mm);
 316 }
 317
 318 static
 319 int lttng_enumerate_vm_maps(struct lttng_session *session)
 320 {
 321         struct task_struct *p;
 322
 323         rcu_read_lock();
 324         for_each_process(p)
 325                 lttng_enumerate_task_vm_maps(session, p);
 326         rcu_read_unlock();
 327         return 0;
 328 }
 329 #endif
 330
 331 static
 332 int lttng_list_interrupts(struct lttng_session *session)
 333 {
 334         unsigned int irq;
 335         unsigned long flags = 0;
 336         struct irq_desc *desc;
 337
 338         /* needs irq_desc */
 339         for_each_irq_desc(irq, desc) {
 340                 struct irqaction *action;
 341                 const char *irq_chip_name =
 342                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 343
 344                 local_irq_save(flags);
 345                 raw_spin_lock(&desc->lock);
 346                 for (action = desc->action; action; action = action->next) {
 347                         trace_lttng_statedump_interrupt(session,
 348                                 irq, irq_chip_name, action);
 349                 }
 350                 raw_spin_unlock(&desc->lock);
 351                 local_irq_restore(flags);
 352         }
 353         return 0;
 354 }
 355
 356 /*
 357  * Statedump the task's namespaces using the proc filesystem inode number as
 358  * the unique identifier. The user and pid ns are nested and will be dumped
 359  * recursively.
 360  *
 361  * Called with task lock held.
 362  */
 363 static
 364 void lttng_statedump_process_ns(struct lttng_session *session,
 365                 struct task_struct *p,
 366                 enum lttng_thread_type type,
 367                 enum lttng_execution_mode mode,
 368                 enum lttng_execution_submode submode,
 369                 enum lttng_process_status status)
 370 {
 371         struct nsproxy *proxy;
 372         struct pid_namespace *pid_ns;
 373         struct user_namespace *user_ns;
 374
 375         /*
 376          * The pid and user namespaces are special, they are nested and
 377          * accessed with specific functions instead of the nsproxy struct
 378          * like the other namespaces.
 379          */
 380         pid_ns = task_active_pid_ns(p);
 381         do {
 382                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 383                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 384         } while (pid_ns);
 385
 386
 387         user_ns = task_cred_xxx(p, user_ns);
 388         do {
 389                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 390                 /*
 391                  * trace_lttng_statedump_process_user_ns() internally
 392                  * checks whether user_ns is NULL. While this does not
 393                  * appear to be a possible return value for
 394                  * task_cred_xxx(), err on the safe side and check
 395                  * for NULL here as well to be consistent with the
 396                  * paranoid behavior of
 397                  * trace_lttng_statedump_process_user_ns().
 398                  */
 399                 user_ns = user_ns ? user_ns->parent : NULL;
 400         } while (user_ns);
 401
 402         /*
 403          * Back and forth on locking strategy within Linux upstream for nsproxy.
 404          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 405          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 406          * for details.
 407          */
 408 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 409                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 410                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 411                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 412         proxy = p->nsproxy;
 413 #else
 414         rcu_read_lock();
 415         proxy = task_nsproxy(p);
 416 #endif
 417         if (proxy) {
 418 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
 419                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 420 #endif
 421                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 422 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 423                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 424 #endif
 425                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 426                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 427         }
 428 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 429                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 430                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 431                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 432         /* (nothing) */
 433 #else
 434         rcu_read_unlock();
 435 #endif
 436 }
 437
 438 static
 439 int lttng_enumerate_process_states(struct lttng_session *session)
 440 {
 441         struct task_struct *g, *p;
 442         char *tmp;
 443
 444         tmp = (char *) __get_free_page(GFP_KERNEL);
 445         if (!tmp)
 446                 return -ENOMEM;
 447
 448         rcu_read_lock();
 449         for_each_process(g) {
 450                 struct files_struct *prev_files = NULL;
 451
 452                 p = g;
 453                 do {
 454                         enum lttng_execution_mode mode =
 455                                 LTTNG_MODE_UNKNOWN;
 456                         enum lttng_execution_submode submode =
 457                                 LTTNG_UNKNOWN;
 458                         enum lttng_process_status status;
 459                         enum lttng_thread_type type;
 460                         struct files_struct *files;
 461
 462                         task_lock(p);
 463                         if (p->exit_state == EXIT_ZOMBIE)
 464                                 status = LTTNG_ZOMBIE;
 465                         else if (p->exit_state == EXIT_DEAD)
 466                                 status = LTTNG_DEAD;
 467                         else if (p->state == TASK_RUNNING) {
 468                                 /* Is this a forked child that has not run yet? */
 469                                 if (list_empty(&p->rt.run_list))
 470                                         status = LTTNG_WAIT_FORK;
 471                                 else
 472                                         /*
 473                                          * All tasks are considered as wait_cpu;
 474                                          * the viewer will sort out if the task
 475                                          * was really running at this time.
 476                                          */
 477                                         status = LTTNG_WAIT_CPU;
 478                         } else if (p->state &
 479                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 480                                 /* Task is waiting for something to complete */
 481                                 status = LTTNG_WAIT;
 482                         } else
 483                                 status = LTTNG_UNNAMED;
 484                         submode = LTTNG_NONE;
 485
 486                         /*
 487                          * Verification of t->mm is to filter out kernel
 488                          * threads; Viewer will further filter out if a
 489                          * user-space thread was in syscall mode or not.
 490                          */
 491                         if (p->mm)
 492                                 type = LTTNG_USER_THREAD;
 493                         else
 494                                 type = LTTNG_KERNEL_THREAD;
 495                         files = p->files;
 496
 497                         trace_lttng_statedump_process_state(session,
 498                                 p, type, mode, submode, status, files);
 499                         lttng_statedump_process_ns(session,
 500                                 p, type, mode, submode, status);
 501                         /*
 502                          * As an optimisation for the common case, do not
 503                          * repeat information for the same files_struct in
 504                          * two consecutive threads. This is the common case
 505                          * for threads sharing the same fd table. RCU guarantees
 506                          * that the same files_struct pointer is not re-used
 507                          * throughout processes/threads iteration.
 508                          */
 509                         if (files && files != prev_files) {
 510                                 lttng_enumerate_files(session, files, tmp);
 511                                 prev_files = files;
 512                         }
 513                         task_unlock(p);
 514                 } while_each_thread(g, p);
 515         }
 516         rcu_read_unlock();
 517
 518         free_page((unsigned long) tmp);
 519
 520         return 0;
 521 }
 522
 523 static
 524 void lttng_statedump_work_func(struct work_struct *work)
 525 {
 526         if (atomic_dec_and_test(&kernel_threads_to_run))
 527                 /* If we are the last thread, wake up do_lttng_statedump */
 528                 wake_up(&statedump_wq);
 529 }
 530
 531 static
 532 int do_lttng_statedump(struct lttng_session *session)
 533 {
 534         int cpu, ret;
 535
 536         trace_lttng_statedump_start(session);
 537         ret = lttng_enumerate_process_states(session);
 538         if (ret)
 539                 return ret;
 540         /*
 541          * FIXME
 542          * ret = lttng_enumerate_vm_maps(session);
 543          * if (ret)
 544          *      return ret;
 545          */
 546         ret = lttng_list_interrupts(session);
 547         if (ret)
 548                 return ret;
 549         ret = lttng_enumerate_network_ip_interface(session);
 550         if (ret)
 551                 return ret;
 552         ret = lttng_enumerate_block_devices(session);
 553         switch (ret) {
 554         case 0:
 555                 break;
 556         case -ENOSYS:
 557                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 558                 break;
 559         default:
 560                 return ret;
 561         }
 562         ret = lttng_enumerate_cpu_topology(session);
 563         if (ret)
 564                 return ret;
 565
 566         /* TODO lttng_dump_idt_table(session); */
 567         /* TODO lttng_dump_softirq_vec(session); */
 568         /* TODO lttng_list_modules(session); */
 569         /* TODO lttng_dump_swap_files(session); */
 570
 571         /*
 572          * Fire off a work queue on each CPU. Their sole purpose in life
 573          * is to guarantee that each CPU has been in a state where is was in
 574          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 575          */
 576         get_online_cpus();
 577         atomic_set(&kernel_threads_to_run, num_online_cpus());
 578         for_each_online_cpu(cpu) {
 579                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 580                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 581         }
 582         /* Wait for all threads to run */
 583         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 584         put_online_cpus();
 585         /* Our work is done */
 586         trace_lttng_statedump_end(session);
 587         return 0;
 588 }
 589
 590 /*
 591  * Called with session mutex held.
 592  */
 593 int lttng_statedump_start(struct lttng_session *session)
 594 {
 595         return do_lttng_statedump(session);
 596 }
 597 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 598
 599 static
 600 int __init lttng_statedump_init(void)
 601 {
 602         return 0;
 603 }
 604
 605 module_init(lttng_statedump_init);
 606
 607 static
 608 void __exit lttng_statedump_exit(void)
 609 {
 610 }
 611
 612 module_exit(lttng_statedump_exit);
 613
 614 MODULE_LICENSE("GPL and additional rights");
 615 MODULE_AUTHOR("Jean-Hugues Deschenes");
 616 MODULE_DESCRIPTION("LTTng statedump provider");
 617 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 618         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 619         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 620         LTTNG_MODULES_EXTRAVERSION);