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/namespace.h>
  41 #include <wrapper/genhd.h>
  42
  43 /* Define the tracepoints, but do not build the probes */
  44 #define CREATE_TRACE_POINTS
  45 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
  46 #define TRACE_INCLUDE_FILE lttng-statedump
  47 #define LTTNG_INSTRUMENTATION
  48 #include <instrumentation/events/lttng-module/lttng-statedump.h>
  49
  50 DEFINE_TRACE(lttng_statedump_block_device);
  51 DEFINE_TRACE(lttng_statedump_end);
  52 DEFINE_TRACE(lttng_statedump_interrupt);
  53 DEFINE_TRACE(lttng_statedump_file_descriptor);
  54 DEFINE_TRACE(lttng_statedump_start);
  55 DEFINE_TRACE(lttng_statedump_process_state);
  56 DEFINE_TRACE(lttng_statedump_process_pid_ns);
  57 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
  58 DEFINE_TRACE(lttng_statedump_process_cgroup_ns);
  59 #endif
  60 DEFINE_TRACE(lttng_statedump_process_ipc_ns);
  61 #ifndef LTTNG_MNT_NS_MISSING_HEADER
  62 DEFINE_TRACE(lttng_statedump_process_mnt_ns);
  63 #endif
  64 DEFINE_TRACE(lttng_statedump_process_net_ns);
  65 DEFINE_TRACE(lttng_statedump_process_user_ns);
  66 DEFINE_TRACE(lttng_statedump_process_uts_ns);
  67 DEFINE_TRACE(lttng_statedump_network_interface);
  68 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
  69 DEFINE_TRACE(lttng_statedump_cpu_topology);
  70 #endif
  71
  72 struct lttng_fd_ctx {
  73         char *page;
  74         struct lttng_session *session;
  75         struct files_struct *files;
  76 };
  77
  78 /*
  79  * Protected by the trace lock.
  80  */
  81 static struct delayed_work cpu_work[NR_CPUS];
  82 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
  83 static atomic_t kernel_threads_to_run;
  84
  85 enum lttng_thread_type {
  86         LTTNG_USER_THREAD = 0,
  87         LTTNG_KERNEL_THREAD = 1,
  88 };
  89
  90 enum lttng_execution_mode {
  91         LTTNG_USER_MODE = 0,
  92         LTTNG_SYSCALL = 1,
  93         LTTNG_TRAP = 2,
  94         LTTNG_IRQ = 3,
  95         LTTNG_SOFTIRQ = 4,
  96         LTTNG_MODE_UNKNOWN = 5,
  97 };
  98
  99 enum lttng_execution_submode {
 100         LTTNG_NONE = 0,
 101         LTTNG_UNKNOWN = 1,
 102 };
 103
 104 enum lttng_process_status {
 105         LTTNG_UNNAMED = 0,
 106         LTTNG_WAIT_FORK = 1,
 107         LTTNG_WAIT_CPU = 2,
 108         LTTNG_EXIT = 3,
 109         LTTNG_ZOMBIE = 4,
 110         LTTNG_WAIT = 5,
 111         LTTNG_RUN = 6,
 112         LTTNG_DEAD = 7,
 113 };
 114
 115 static
 116 int lttng_enumerate_block_devices(struct lttng_session *session)
 117 {
 118         struct class *ptr_block_class;
 119         struct device_type *ptr_disk_type;
 120         struct class_dev_iter iter;
 121         struct device *dev;
 122
 123         ptr_block_class = wrapper_get_block_class();
 124         if (!ptr_block_class)
 125                 return -ENOSYS;
 126         ptr_disk_type = wrapper_get_disk_type();
 127         if (!ptr_disk_type) {
 128                 return -ENOSYS;
 129         }
 130         class_dev_iter_init(&iter, ptr_block_class, NULL, ptr_disk_type);
 131         while ((dev = class_dev_iter_next(&iter))) {
 132                 struct disk_part_iter piter;
 133                 struct gendisk *disk = dev_to_disk(dev);
 134                 struct hd_struct *part;
 135
 136                 /*
 137                  * Don't show empty devices or things that have been
 138                  * suppressed
 139                  */
 140                 if (get_capacity(disk) == 0 ||
 141                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 142                         continue;
 143
 144                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 145                 while ((part = disk_part_iter_next(&piter))) {
 146                         char name_buf[BDEVNAME_SIZE];
 147                         char *p;
 148
 149                         p = wrapper_disk_name(disk, part->partno, name_buf);
 150                         if (!p) {
 151                                 disk_part_iter_exit(&piter);
 152                                 class_dev_iter_exit(&iter);
 153                                 return -ENOSYS;
 154                         }
 155                         trace_lttng_statedump_block_device(session,
 156                                         part_devt(part), name_buf);
 157                 }
 158                 disk_part_iter_exit(&piter);
 159         }
 160         class_dev_iter_exit(&iter);
 161         return 0;
 162 }
 163
 164 #ifdef CONFIG_INET
 165
 166 static
 167 void lttng_enumerate_device(struct lttng_session *session,
 168                 struct net_device *dev)
 169 {
 170         struct in_device *in_dev;
 171         struct in_ifaddr *ifa;
 172
 173         if (dev->flags & IFF_UP) {
 174                 in_dev = in_dev_get(dev);
 175                 if (in_dev) {
 176                         for (ifa = in_dev->ifa_list; ifa != NULL;
 177                              ifa = ifa->ifa_next) {
 178                                 trace_lttng_statedump_network_interface(
 179                                         session, dev, ifa);
 180                         }
 181                         in_dev_put(in_dev);
 182                 }
 183         } else {
 184                 trace_lttng_statedump_network_interface(
 185                         session, dev, NULL);
 186         }
 187 }
 188
 189 static
 190 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 191 {
 192         struct net_device *dev;
 193
 194         read_lock(&dev_base_lock);
 195         for_each_netdev(&init_net, dev)
 196                 lttng_enumerate_device(session, dev);
 197         read_unlock(&dev_base_lock);
 198
 199         return 0;
 200 }
 201 #else /* CONFIG_INET */
 202 static inline
 203 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 204 {
 205         return 0;
 206 }
 207 #endif /* CONFIG_INET */
 208
 209 static
 210 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 211 {
 212         const struct lttng_fd_ctx *ctx = p;
 213         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 214         unsigned int flags = file->f_flags;
 215         struct fdtable *fdt;
 216
 217         /*
 218          * We don't expose kernel internal flags, only userspace-visible
 219          * flags.
 220          */
 221         flags &= ~FMODE_NONOTIFY;
 222         fdt = files_fdtable(ctx->files);
 223         /*
 224          * We need to check here again whether fd is within the fdt
 225          * max_fds range, because we might be seeing a different
 226          * files_fdtable() than iterate_fd(), assuming only RCU is
 227          * protecting the read. In reality, iterate_fd() holds
 228          * file_lock, which should ensure the fdt does not change while
 229          * the lock is taken, but we are not aware whether this is
 230          * guaranteed or not, so play safe.
 231          */
 232         if (fd < fdt->max_fds && close_on_exec(fd, fdt))
 233                 flags |= O_CLOEXEC;
 234         if (IS_ERR(s)) {
 235                 struct dentry *dentry = file->f_path.dentry;
 236
 237                 /* Make sure we give at least some info */
 238                 spin_lock(&dentry->d_lock);
 239                 trace_lttng_statedump_file_descriptor(ctx->session,
 240                         ctx->files, fd, dentry->d_name.name, flags,
 241                         file->f_mode);
 242                 spin_unlock(&dentry->d_lock);
 243                 goto end;
 244         }
 245         trace_lttng_statedump_file_descriptor(ctx->session,
 246                 ctx->files, fd, s, flags, file->f_mode);
 247 end:
 248         return 0;
 249 }
 250
 251 /* Called with task lock held. */
 252 static
 253 void lttng_enumerate_files(struct lttng_session *session,
 254                 struct files_struct *files,
 255                 char *tmp)
 256 {
 257         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .files = files, };
 258
 259         iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 260 }
 261
 262 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 263 static
 264 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 265 {
 266         int cpu;
 267         const cpumask_t *cpumask = cpu_possible_mask;
 268
 269         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 270                         cpu = cpumask_next(cpu, cpumask)) {
 271                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 272         }
 273
 274         return 0;
 275 }
 276 #else
 277 static
 278 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 279 {
 280         return 0;
 281 }
 282 #endif
 283
 284 #if 0
 285 /*
 286  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 287  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 288  * iteration, but it is not exported to modules.
 289  */
 290 static
 291 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 292                 struct task_struct *p)
 293 {
 294         struct mm_struct *mm;
 295         struct vm_area_struct *map;
 296         unsigned long ino;
 297
 298         /* get_task_mm does a task_lock... */
 299         mm = get_task_mm(p);
 300         if (!mm)
 301                 return;
 302
 303         map = mm->mmap;
 304         if (map) {
 305                 down_read(&mm->mmap_sem);
 306                 while (map) {
 307                         if (map->vm_file)
 308                                 ino = map->vm_file->f_path.dentry->d_inode->i_ino;
 309                         else
 310                                 ino = 0;
 311                         trace_lttng_statedump_vm_map(session, p, map, ino);
 312                         map = map->vm_next;
 313                 }
 314                 up_read(&mm->mmap_sem);
 315         }
 316         mmput(mm);
 317 }
 318
 319 static
 320 int lttng_enumerate_vm_maps(struct lttng_session *session)
 321 {
 322         struct task_struct *p;
 323
 324         rcu_read_lock();
 325         for_each_process(p)
 326                 lttng_enumerate_task_vm_maps(session, p);
 327         rcu_read_unlock();
 328         return 0;
 329 }
 330 #endif
 331
 332 static
 333 int lttng_list_interrupts(struct lttng_session *session)
 334 {
 335         unsigned int irq;
 336         unsigned long flags = 0;
 337         struct irq_desc *desc;
 338
 339         /* needs irq_desc */
 340         for_each_irq_desc(irq, desc) {
 341                 struct irqaction *action;
 342                 const char *irq_chip_name =
 343                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 344
 345                 local_irq_save(flags);
 346                 raw_spin_lock(&desc->lock);
 347                 for (action = desc->action; action; action = action->next) {
 348                         trace_lttng_statedump_interrupt(session,
 349                                 irq, irq_chip_name, action);
 350                 }
 351                 raw_spin_unlock(&desc->lock);
 352                 local_irq_restore(flags);
 353         }
 354         return 0;
 355 }
 356
 357 /*
 358  * Statedump the task's namespaces using the proc filesystem inode number as
 359  * the unique identifier. The user and pid ns are nested and will be dumped
 360  * recursively.
 361  *
 362  * Called with task lock held.
 363  */
 364 static
 365 void lttng_statedump_process_ns(struct lttng_session *session,
 366                 struct task_struct *p,
 367                 enum lttng_thread_type type,
 368                 enum lttng_execution_mode mode,
 369                 enum lttng_execution_submode submode,
 370                 enum lttng_process_status status)
 371 {
 372         struct nsproxy *proxy;
 373         struct pid_namespace *pid_ns;
 374         struct user_namespace *user_ns;
 375
 376         /*
 377          * The pid and user namespaces are special, they are nested and
 378          * accessed with specific functions instead of the nsproxy struct
 379          * like the other namespaces.
 380          */
 381         pid_ns = task_active_pid_ns(p);
 382         do {
 383                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 384                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 385         } while (pid_ns);
 386
 387
 388         user_ns = task_cred_xxx(p, user_ns);
 389         do {
 390                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 391                 /*
 392                  * trace_lttng_statedump_process_user_ns() internally
 393                  * checks whether user_ns is NULL. While this does not
 394                  * appear to be a possible return value for
 395                  * task_cred_xxx(), err on the safe side and check
 396                  * for NULL here as well to be consistent with the
 397                  * paranoid behavior of
 398                  * trace_lttng_statedump_process_user_ns().
 399                  */
 400                 user_ns = user_ns ? user_ns->lttng_user_ns_parent : NULL;
 401         } while (user_ns);
 402
 403         /*
 404          * Back and forth on locking strategy within Linux upstream for nsproxy.
 405          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 406          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 407          * for details.
 408          */
 409 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 410                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 411                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 412                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 413         proxy = p->nsproxy;
 414 #else
 415         rcu_read_lock();
 416         proxy = task_nsproxy(p);
 417 #endif
 418         if (proxy) {
 419 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
 420                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 421 #endif
 422                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 423 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 424                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 425 #endif
 426                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 427                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 428         }
 429 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 430                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 431                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 432                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 433         /* (nothing) */
 434 #else
 435         rcu_read_unlock();
 436 #endif
 437 }
 438
 439 static
 440 int lttng_enumerate_process_states(struct lttng_session *session)
 441 {
 442         struct task_struct *g, *p;
 443         char *tmp;
 444
 445         tmp = (char *) __get_free_page(GFP_KERNEL);
 446         if (!tmp)
 447                 return -ENOMEM;
 448
 449         rcu_read_lock();
 450         for_each_process(g) {
 451                 struct files_struct *prev_files = NULL;
 452
 453                 p = g;
 454                 do {
 455                         enum lttng_execution_mode mode =
 456                                 LTTNG_MODE_UNKNOWN;
 457                         enum lttng_execution_submode submode =
 458                                 LTTNG_UNKNOWN;
 459                         enum lttng_process_status status;
 460                         enum lttng_thread_type type;
 461                         struct files_struct *files;
 462
 463                         task_lock(p);
 464                         if (p->exit_state == EXIT_ZOMBIE)
 465                                 status = LTTNG_ZOMBIE;
 466                         else if (p->exit_state == EXIT_DEAD)
 467                                 status = LTTNG_DEAD;
 468                         else if (p->state == TASK_RUNNING) {
 469                                 /* Is this a forked child that has not run yet? */
 470                                 if (list_empty(&p->rt.run_list))
 471                                         status = LTTNG_WAIT_FORK;
 472                                 else
 473                                         /*
 474                                          * All tasks are considered as wait_cpu;
 475                                          * the viewer will sort out if the task
 476                                          * was really running at this time.
 477                                          */
 478                                         status = LTTNG_WAIT_CPU;
 479                         } else if (p->state &
 480                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 481                                 /* Task is waiting for something to complete */
 482                                 status = LTTNG_WAIT;
 483                         } else
 484                                 status = LTTNG_UNNAMED;
 485                         submode = LTTNG_NONE;
 486
 487                         /*
 488                          * Verification of t->mm is to filter out kernel
 489                          * threads; Viewer will further filter out if a
 490                          * user-space thread was in syscall mode or not.
 491                          */
 492                         if (p->mm)
 493                                 type = LTTNG_USER_THREAD;
 494                         else
 495                                 type = LTTNG_KERNEL_THREAD;
 496                         files = p->files;
 497
 498                         trace_lttng_statedump_process_state(session,
 499                                 p, type, mode, submode, status, files);
 500                         lttng_statedump_process_ns(session,
 501                                 p, type, mode, submode, status);
 502                         /*
 503                          * As an optimisation for the common case, do not
 504                          * repeat information for the same files_struct in
 505                          * two consecutive threads. This is the common case
 506                          * for threads sharing the same fd table. RCU guarantees
 507                          * that the same files_struct pointer is not re-used
 508                          * throughout processes/threads iteration.
 509                          */
 510                         if (files && files != prev_files) {
 511                                 lttng_enumerate_files(session, files, tmp);
 512                                 prev_files = files;
 513                         }
 514                         task_unlock(p);
 515                 } while_each_thread(g, p);
 516         }
 517         rcu_read_unlock();
 518
 519         free_page((unsigned long) tmp);
 520
 521         return 0;
 522 }
 523
 524 static
 525 void lttng_statedump_work_func(struct work_struct *work)
 526 {
 527         if (atomic_dec_and_test(&kernel_threads_to_run))
 528                 /* If we are the last thread, wake up do_lttng_statedump */
 529                 wake_up(&statedump_wq);
 530 }
 531
 532 static
 533 int do_lttng_statedump(struct lttng_session *session)
 534 {
 535         int cpu, ret;
 536
 537         trace_lttng_statedump_start(session);
 538         ret = lttng_enumerate_process_states(session);
 539         if (ret)
 540                 return ret;
 541         /*
 542          * FIXME
 543          * ret = lttng_enumerate_vm_maps(session);
 544          * if (ret)
 545          *      return ret;
 546          */
 547         ret = lttng_list_interrupts(session);
 548         if (ret)
 549                 return ret;
 550         ret = lttng_enumerate_network_ip_interface(session);
 551         if (ret)
 552                 return ret;
 553         ret = lttng_enumerate_block_devices(session);
 554         switch (ret) {
 555         case 0:
 556                 break;
 557         case -ENOSYS:
 558                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 559                 break;
 560         default:
 561                 return ret;
 562         }
 563         ret = lttng_enumerate_cpu_topology(session);
 564         if (ret)
 565                 return ret;
 566
 567         /* TODO lttng_dump_idt_table(session); */
 568         /* TODO lttng_dump_softirq_vec(session); */
 569         /* TODO lttng_list_modules(session); */
 570         /* TODO lttng_dump_swap_files(session); */
 571
 572         /*
 573          * Fire off a work queue on each CPU. Their sole purpose in life
 574          * is to guarantee that each CPU has been in a state where is was in
 575          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 576          */
 577         get_online_cpus();
 578         atomic_set(&kernel_threads_to_run, num_online_cpus());
 579         for_each_online_cpu(cpu) {
 580                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 581                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 582         }
 583         /* Wait for all threads to run */
 584         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 585         put_online_cpus();
 586         /* Our work is done */
 587         trace_lttng_statedump_end(session);
 588         return 0;
 589 }
 590
 591 /*
 592  * Called with session mutex held.
 593  */
 594 int lttng_statedump_start(struct lttng_session *session)
 595 {
 596         return do_lttng_statedump(session);
 597 }
 598 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 599
 600 static
 601 int __init lttng_statedump_init(void)
 602 {
 603         return 0;
 604 }
 605
 606 module_init(lttng_statedump_init);
 607
 608 static
 609 void __exit lttng_statedump_exit(void)
 610 {
 611 }
 612
 613 module_exit(lttng_statedump_exit);
 614
 615 MODULE_LICENSE("GPL and additional rights");
 616 MODULE_AUTHOR("Jean-Hugues Deschenes");
 617 MODULE_DESCRIPTION("LTTng statedump provider");
 618 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 619         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 620         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 621         LTTNG_MODULES_EXTRAVERSION);