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
  37 #include <lttng-events.h>
  38 #include <lttng-tracer.h>
  39 #include <wrapper/namespace.h>
  40 #include <wrapper/irq.h>
  41 #include <wrapper/genhd.h>
  42 #include <wrapper/file.h>
  43
  44 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
  45 #include <linux/irq.h>
  46 #endif
  47
  48 /* Define the tracepoints, but do not build the probes */
  49 #define CREATE_TRACE_POINTS
  50 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
  51 #define TRACE_INCLUDE_FILE lttng-statedump
  52 #define LTTNG_INSTRUMENTATION
  53 #include <instrumentation/events/lttng-module/lttng-statedump.h>
  54
  55 DEFINE_TRACE(lttng_statedump_block_device);
  56 DEFINE_TRACE(lttng_statedump_end);
  57 DEFINE_TRACE(lttng_statedump_interrupt);
  58 DEFINE_TRACE(lttng_statedump_file_descriptor);
  59 DEFINE_TRACE(lttng_statedump_start);
  60 DEFINE_TRACE(lttng_statedump_process_state);
  61 DEFINE_TRACE(lttng_statedump_process_pid_ns);
  62 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
  63 DEFINE_TRACE(lttng_statedump_process_cgroup_ns);
  64 #endif
  65 DEFINE_TRACE(lttng_statedump_process_ipc_ns);
  66 #ifndef LTTNG_MNT_NS_MISSING_HEADER
  67 DEFINE_TRACE(lttng_statedump_process_mnt_ns);
  68 #endif
  69 DEFINE_TRACE(lttng_statedump_process_net_ns);
  70 DEFINE_TRACE(lttng_statedump_process_user_ns);
  71 DEFINE_TRACE(lttng_statedump_process_uts_ns);
  72 DEFINE_TRACE(lttng_statedump_network_interface);
  73 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
  74 DEFINE_TRACE(lttng_statedump_cpu_topology);
  75 #endif
  76
  77 struct lttng_fd_ctx {
  78         char *page;
  79         struct lttng_session *session;
  80         struct files_struct *files;
  81 };
  82
  83 /*
  84  * Protected by the trace lock.
  85  */
  86 static struct delayed_work cpu_work[NR_CPUS];
  87 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
  88 static atomic_t kernel_threads_to_run;
  89
  90 enum lttng_thread_type {
  91         LTTNG_USER_THREAD = 0,
  92         LTTNG_KERNEL_THREAD = 1,
  93 };
  94
  95 enum lttng_execution_mode {
  96         LTTNG_USER_MODE = 0,
  97         LTTNG_SYSCALL = 1,
  98         LTTNG_TRAP = 2,
  99         LTTNG_IRQ = 3,
 100         LTTNG_SOFTIRQ = 4,
 101         LTTNG_MODE_UNKNOWN = 5,
 102 };
 103
 104 enum lttng_execution_submode {
 105         LTTNG_NONE = 0,
 106         LTTNG_UNKNOWN = 1,
 107 };
 108
 109 enum lttng_process_status {
 110         LTTNG_UNNAMED = 0,
 111         LTTNG_WAIT_FORK = 1,
 112         LTTNG_WAIT_CPU = 2,
 113         LTTNG_EXIT = 3,
 114         LTTNG_ZOMBIE = 4,
 115         LTTNG_WAIT = 5,
 116         LTTNG_RUN = 6,
 117         LTTNG_DEAD = 7,
 118 };
 119
 120 static
 121 int lttng_enumerate_block_devices(struct lttng_session *session)
 122 {
 123         struct class *ptr_block_class;
 124         struct device_type *ptr_disk_type;
 125         struct class_dev_iter iter;
 126         struct device *dev;
 127
 128         ptr_block_class = wrapper_get_block_class();
 129         if (!ptr_block_class)
 130                 return -ENOSYS;
 131         ptr_disk_type = wrapper_get_disk_type();
 132         if (!ptr_disk_type) {
 133                 return -ENOSYS;
 134         }
 135         class_dev_iter_init(&iter, ptr_block_class, NULL, ptr_disk_type);
 136         while ((dev = class_dev_iter_next(&iter))) {
 137                 struct disk_part_iter piter;
 138                 struct gendisk *disk = dev_to_disk(dev);
 139                 struct hd_struct *part;
 140
 141                 /*
 142                  * Don't show empty devices or things that have been
 143                  * suppressed
 144                  */
 145                 if (get_capacity(disk) == 0 ||
 146                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 147                         continue;
 148
 149                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 150                 while ((part = disk_part_iter_next(&piter))) {
 151                         char name_buf[BDEVNAME_SIZE];
 152                         char *p;
 153
 154                         p = wrapper_disk_name(disk, part->partno, name_buf);
 155                         if (!p) {
 156                                 disk_part_iter_exit(&piter);
 157                                 class_dev_iter_exit(&iter);
 158                                 return -ENOSYS;
 159                         }
 160                         trace_lttng_statedump_block_device(session,
 161                                         part_devt(part), name_buf);
 162                 }
 163                 disk_part_iter_exit(&piter);
 164         }
 165         class_dev_iter_exit(&iter);
 166         return 0;
 167 }
 168
 169 #ifdef CONFIG_INET
 170
 171 static
 172 void lttng_enumerate_device(struct lttng_session *session,
 173                 struct net_device *dev)
 174 {
 175         struct in_device *in_dev;
 176         struct in_ifaddr *ifa;
 177
 178         if (dev->flags & IFF_UP) {
 179                 in_dev = in_dev_get(dev);
 180                 if (in_dev) {
 181                         for (ifa = in_dev->ifa_list; ifa != NULL;
 182                              ifa = ifa->ifa_next) {
 183                                 trace_lttng_statedump_network_interface(
 184                                         session, dev, ifa);
 185                         }
 186                         in_dev_put(in_dev);
 187                 }
 188         } else {
 189                 trace_lttng_statedump_network_interface(
 190                         session, dev, NULL);
 191         }
 192 }
 193
 194 static
 195 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 196 {
 197         struct net_device *dev;
 198
 199         read_lock(&dev_base_lock);
 200         for_each_netdev(&init_net, dev)
 201                 lttng_enumerate_device(session, dev);
 202         read_unlock(&dev_base_lock);
 203
 204         return 0;
 205 }
 206 #else /* CONFIG_INET */
 207 static inline
 208 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 209 {
 210         return 0;
 211 }
 212 #endif /* CONFIG_INET */
 213
 214 static
 215 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 216 {
 217         const struct lttng_fd_ctx *ctx = p;
 218         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 219         unsigned int flags = file->f_flags;
 220         struct fdtable *fdt;
 221
 222         /*
 223          * We don't expose kernel internal flags, only userspace-visible
 224          * flags.
 225          */
 226         flags &= ~FMODE_NONOTIFY;
 227         fdt = files_fdtable(ctx->files);
 228         /*
 229          * We need to check here again whether fd is within the fdt
 230          * max_fds range, because we might be seeing a different
 231          * files_fdtable() than iterate_fd(), assuming only RCU is
 232          * protecting the read. In reality, iterate_fd() holds
 233          * file_lock, which should ensure the fdt does not change while
 234          * the lock is taken, but we are not aware whether this is
 235          * guaranteed or not, so play safe.
 236          */
 237         if (fd < fdt->max_fds && close_on_exec(fd, fdt))
 238                 flags |= O_CLOEXEC;
 239         if (IS_ERR(s)) {
 240                 struct dentry *dentry = file->f_path.dentry;
 241
 242                 /* Make sure we give at least some info */
 243                 spin_lock(&dentry->d_lock);
 244                 trace_lttng_statedump_file_descriptor(ctx->session,
 245                         ctx->files, fd, dentry->d_name.name, flags,
 246                         file->f_mode);
 247                 spin_unlock(&dentry->d_lock);
 248                 goto end;
 249         }
 250         trace_lttng_statedump_file_descriptor(ctx->session,
 251                 ctx->files, fd, s, flags, file->f_mode);
 252 end:
 253         return 0;
 254 }
 255
 256 /* Called with task lock held. */
 257 static
 258 void lttng_enumerate_files(struct lttng_session *session,
 259                 struct files_struct *files,
 260                 char *tmp)
 261 {
 262         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .files = files, };
 263
 264         iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 265 }
 266
 267 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 268 static
 269 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 270 {
 271         int cpu;
 272         const cpumask_t *cpumask = cpu_possible_mask;
 273
 274         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 275                         cpu = cpumask_next(cpu, cpumask)) {
 276                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 277         }
 278
 279         return 0;
 280 }
 281 #else
 282 static
 283 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 284 {
 285         return 0;
 286 }
 287 #endif
 288
 289 #if 0
 290 /*
 291  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 292  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 293  * iteration, but it is not exported to modules.
 294  */
 295 static
 296 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 297                 struct task_struct *p)
 298 {
 299         struct mm_struct *mm;
 300         struct vm_area_struct *map;
 301         unsigned long ino;
 302
 303         /* get_task_mm does a task_lock... */
 304         mm = get_task_mm(p);
 305         if (!mm)
 306                 return;
 307
 308         map = mm->mmap;
 309         if (map) {
 310                 down_read(&mm->mmap_sem);
 311                 while (map) {
 312                         if (map->vm_file)
 313                                 ino = map->vm_file->lttng_f_dentry->d_inode->i_ino;
 314                         else
 315                                 ino = 0;
 316                         trace_lttng_statedump_vm_map(session, p, map, ino);
 317                         map = map->vm_next;
 318                 }
 319                 up_read(&mm->mmap_sem);
 320         }
 321         mmput(mm);
 322 }
 323
 324 static
 325 int lttng_enumerate_vm_maps(struct lttng_session *session)
 326 {
 327         struct task_struct *p;
 328
 329         rcu_read_lock();
 330         for_each_process(p)
 331                 lttng_enumerate_task_vm_maps(session, p);
 332         rcu_read_unlock();
 333         return 0;
 334 }
 335 #endif
 336
 337 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
 338
 339 static
 340 int lttng_list_interrupts(struct lttng_session *session)
 341 {
 342         unsigned int irq;
 343         unsigned long flags = 0;
 344         struct irq_desc *desc;
 345
 346         /* needs irq_desc */
 347         for_each_irq_desc(irq, desc) {
 348                 struct irqaction *action;
 349                 const char *irq_chip_name =
 350                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 351
 352                 local_irq_save(flags);
 353                 raw_spin_lock(&desc->lock);
 354                 for (action = desc->action; action; action = action->next) {
 355                         trace_lttng_statedump_interrupt(session,
 356                                 irq, irq_chip_name, action);
 357                 }
 358                 raw_spin_unlock(&desc->lock);
 359                 local_irq_restore(flags);
 360         }
 361         return 0;
 362 }
 363 #else
 364 static inline
 365 int lttng_list_interrupts(struct lttng_session *session)
 366 {
 367         return 0;
 368 }
 369 #endif
 370
 371 /*
 372  * Statedump the task's namespaces using the proc filesystem inode number as
 373  * the unique identifier. The user and pid ns are nested and will be dumped
 374  * recursively.
 375  *
 376  * Called with task lock held.
 377  */
 378 static
 379 void lttng_statedump_process_ns(struct lttng_session *session,
 380                 struct task_struct *p,
 381                 enum lttng_thread_type type,
 382                 enum lttng_execution_mode mode,
 383                 enum lttng_execution_submode submode,
 384                 enum lttng_process_status status)
 385 {
 386         struct nsproxy *proxy;
 387         struct pid_namespace *pid_ns;
 388         struct user_namespace *user_ns;
 389
 390         /*
 391          * The pid and user namespaces are special, they are nested and
 392          * accessed with specific functions instead of the nsproxy struct
 393          * like the other namespaces.
 394          */
 395         pid_ns = task_active_pid_ns(p);
 396         do {
 397                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 398                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 399         } while (pid_ns);
 400
 401
 402         user_ns = task_cred_xxx(p, user_ns);
 403         do {
 404                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 405                 /*
 406                  * trace_lttng_statedump_process_user_ns() internally
 407                  * checks whether user_ns is NULL. While this does not
 408                  * appear to be a possible return value for
 409                  * task_cred_xxx(), err on the safe side and check
 410                  * for NULL here as well to be consistent with the
 411                  * paranoid behavior of
 412                  * trace_lttng_statedump_process_user_ns().
 413                  */
 414                 user_ns = user_ns ? user_ns->lttng_user_ns_parent : NULL;
 415         } while (user_ns);
 416
 417         /*
 418          * Back and forth on locking strategy within Linux upstream for nsproxy.
 419          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 420          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 421          * for details.
 422          */
 423 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 424                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 425                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 426                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 427         proxy = p->nsproxy;
 428 #else
 429         rcu_read_lock();
 430         proxy = task_nsproxy(p);
 431 #endif
 432         if (proxy) {
 433 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
 434                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 435 #endif
 436                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 437 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 438                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 439 #endif
 440                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 441                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 442         }
 443 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 444                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 445                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 446                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 447         /* (nothing) */
 448 #else
 449         rcu_read_unlock();
 450 #endif
 451 }
 452
 453 static
 454 int lttng_enumerate_process_states(struct lttng_session *session)
 455 {
 456         struct task_struct *g, *p;
 457         char *tmp;
 458
 459         tmp = (char *) __get_free_page(GFP_KERNEL);
 460         if (!tmp)
 461                 return -ENOMEM;
 462
 463         rcu_read_lock();
 464         for_each_process(g) {
 465                 struct files_struct *prev_files = NULL;
 466
 467                 p = g;
 468                 do {
 469                         enum lttng_execution_mode mode =
 470                                 LTTNG_MODE_UNKNOWN;
 471                         enum lttng_execution_submode submode =
 472                                 LTTNG_UNKNOWN;
 473                         enum lttng_process_status status;
 474                         enum lttng_thread_type type;
 475                         struct files_struct *files;
 476
 477                         task_lock(p);
 478                         if (p->exit_state == EXIT_ZOMBIE)
 479                                 status = LTTNG_ZOMBIE;
 480                         else if (p->exit_state == EXIT_DEAD)
 481                                 status = LTTNG_DEAD;
 482                         else if (p->state == TASK_RUNNING) {
 483                                 /* Is this a forked child that has not run yet? */
 484                                 if (list_empty(&p->rt.run_list))
 485                                         status = LTTNG_WAIT_FORK;
 486                                 else
 487                                         /*
 488                                          * All tasks are considered as wait_cpu;
 489                                          * the viewer will sort out if the task
 490                                          * was really running at this time.
 491                                          */
 492                                         status = LTTNG_WAIT_CPU;
 493                         } else if (p->state &
 494                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 495                                 /* Task is waiting for something to complete */
 496                                 status = LTTNG_WAIT;
 497                         } else
 498                                 status = LTTNG_UNNAMED;
 499                         submode = LTTNG_NONE;
 500
 501                         /*
 502                          * Verification of t->mm is to filter out kernel
 503                          * threads; Viewer will further filter out if a
 504                          * user-space thread was in syscall mode or not.
 505                          */
 506                         if (p->mm)
 507                                 type = LTTNG_USER_THREAD;
 508                         else
 509                                 type = LTTNG_KERNEL_THREAD;
 510                         files = p->files;
 511
 512                         trace_lttng_statedump_process_state(session,
 513                                 p, type, mode, submode, status, files);
 514                         lttng_statedump_process_ns(session,
 515                                 p, type, mode, submode, status);
 516                         /*
 517                          * As an optimisation for the common case, do not
 518                          * repeat information for the same files_struct in
 519                          * two consecutive threads. This is the common case
 520                          * for threads sharing the same fd table. RCU guarantees
 521                          * that the same files_struct pointer is not re-used
 522                          * throughout processes/threads iteration.
 523                          */
 524                         if (files && files != prev_files) {
 525                                 lttng_enumerate_files(session, files, tmp);
 526                                 prev_files = files;
 527                         }
 528                         task_unlock(p);
 529                 } while_each_thread(g, p);
 530         }
 531         rcu_read_unlock();
 532
 533         free_page((unsigned long) tmp);
 534
 535         return 0;
 536 }
 537
 538 static
 539 void lttng_statedump_work_func(struct work_struct *work)
 540 {
 541         if (atomic_dec_and_test(&kernel_threads_to_run))
 542                 /* If we are the last thread, wake up do_lttng_statedump */
 543                 wake_up(&statedump_wq);
 544 }
 545
 546 static
 547 int do_lttng_statedump(struct lttng_session *session)
 548 {
 549         int cpu, ret;
 550
 551         trace_lttng_statedump_start(session);
 552         ret = lttng_enumerate_process_states(session);
 553         if (ret)
 554                 return ret;
 555         /*
 556          * FIXME
 557          * ret = lttng_enumerate_vm_maps(session);
 558          * if (ret)
 559          *      return ret;
 560          */
 561         ret = lttng_list_interrupts(session);
 562         if (ret)
 563                 return ret;
 564         ret = lttng_enumerate_network_ip_interface(session);
 565         if (ret)
 566                 return ret;
 567         ret = lttng_enumerate_block_devices(session);
 568         switch (ret) {
 569         case 0:
 570                 break;
 571         case -ENOSYS:
 572                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 573                 break;
 574         default:
 575                 return ret;
 576         }
 577         ret = lttng_enumerate_cpu_topology(session);
 578         if (ret)
 579                 return ret;
 580
 581         /* TODO lttng_dump_idt_table(session); */
 582         /* TODO lttng_dump_softirq_vec(session); */
 583         /* TODO lttng_list_modules(session); */
 584         /* TODO lttng_dump_swap_files(session); */
 585
 586         /*
 587          * Fire off a work queue on each CPU. Their sole purpose in life
 588          * is to guarantee that each CPU has been in a state where is was in
 589          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 590          */
 591         get_online_cpus();
 592         atomic_set(&kernel_threads_to_run, num_online_cpus());
 593         for_each_online_cpu(cpu) {
 594                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 595                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 596         }
 597         /* Wait for all threads to run */
 598         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 599         put_online_cpus();
 600         /* Our work is done */
 601         trace_lttng_statedump_end(session);
 602         return 0;
 603 }
 604
 605 /*
 606  * Called with session mutex held.
 607  */
 608 int lttng_statedump_start(struct lttng_session *session)
 609 {
 610         return do_lttng_statedump(session);
 611 }
 612 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 613
 614 static
 615 int __init lttng_statedump_init(void)
 616 {
 617         return 0;
 618 }
 619
 620 module_init(lttng_statedump_init);
 621
 622 static
 623 void __exit lttng_statedump_exit(void)
 624 {
 625 }
 626
 627 module_exit(lttng_statedump_exit);
 628
 629 MODULE_LICENSE("GPL and additional rights");
 630 MODULE_AUTHOR("Jean-Hugues Deschenes");
 631 MODULE_DESCRIPTION("LTTng statedump provider");
 632 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 633         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 634         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 635         LTTNG_MODULES_EXTRAVERSION);