lttng-statedump-impl.c

   1 /* SPDX-License-Identifier: (GPL-2.0 or LGPL-2.1)
   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/fdtable.h>
  32 #include <linux/swap.h>
  33 #include <linux/wait.h>
  34 #include <linux/mutex.h>
  35 #include <linux/device.h>
  36
  37 #include <lttng-events.h>
  38 #include <lttng-tracer.h>
  39 #include <wrapper/irqdesc.h>
  40 #include <wrapper/fdtable.h>
  41 #include <wrapper/namespace.h>
  42 #include <wrapper/irq.h>
  43 #include <wrapper/tracepoint.h>
  44 #include <wrapper/genhd.h>
  45 #include <wrapper/file.h>
  46 #include <wrapper/time.h>
  47
  48 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
  49 #include <linux/irq.h>
  50 #endif
  51
  52 /* Define the tracepoints, but do not build the probes */
  53 #define CREATE_TRACE_POINTS
  54 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
  55 #define TRACE_INCLUDE_FILE lttng-statedump
  56 #define LTTNG_INSTRUMENTATION
  57 #include <instrumentation/events/lttng-module/lttng-statedump.h>
  58
  59 DEFINE_TRACE(lttng_statedump_block_device);
  60 DEFINE_TRACE(lttng_statedump_end);
  61 DEFINE_TRACE(lttng_statedump_interrupt);
  62 DEFINE_TRACE(lttng_statedump_file_descriptor);
  63 DEFINE_TRACE(lttng_statedump_start);
  64 DEFINE_TRACE(lttng_statedump_process_state);
  65 DEFINE_TRACE(lttng_statedump_process_pid_ns);
  66 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
  67 DEFINE_TRACE(lttng_statedump_process_cgroup_ns);
  68 #endif
  69 DEFINE_TRACE(lttng_statedump_process_ipc_ns);
  70 #ifndef LTTNG_MNT_NS_MISSING_HEADER
  71 DEFINE_TRACE(lttng_statedump_process_mnt_ns);
  72 #endif
  73 DEFINE_TRACE(lttng_statedump_process_net_ns);
  74 DEFINE_TRACE(lttng_statedump_process_user_ns);
  75 DEFINE_TRACE(lttng_statedump_process_uts_ns);
  76 DEFINE_TRACE(lttng_statedump_network_interface);
  77 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
  78 DEFINE_TRACE(lttng_statedump_cpu_topology);
  79 #endif
  80
  81 struct lttng_fd_ctx {
  82         char *page;
  83         struct lttng_session *session;
  84         struct task_struct *p;
  85         struct files_struct *files;
  86 };
  87
  88 /*
  89  * Protected by the trace lock.
  90  */
  91 static struct delayed_work cpu_work[NR_CPUS];
  92 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
  93 static atomic_t kernel_threads_to_run;
  94
  95 enum lttng_thread_type {
  96         LTTNG_USER_THREAD = 0,
  97         LTTNG_KERNEL_THREAD = 1,
  98 };
  99
 100 enum lttng_execution_mode {
 101         LTTNG_USER_MODE = 0,
 102         LTTNG_SYSCALL = 1,
 103         LTTNG_TRAP = 2,
 104         LTTNG_IRQ = 3,
 105         LTTNG_SOFTIRQ = 4,
 106         LTTNG_MODE_UNKNOWN = 5,
 107 };
 108
 109 enum lttng_execution_submode {
 110         LTTNG_NONE = 0,
 111         LTTNG_UNKNOWN = 1,
 112 };
 113
 114 enum lttng_process_status {
 115         LTTNG_UNNAMED = 0,
 116         LTTNG_WAIT_FORK = 1,
 117         LTTNG_WAIT_CPU = 2,
 118         LTTNG_EXIT = 3,
 119         LTTNG_ZOMBIE = 4,
 120         LTTNG_WAIT = 5,
 121         LTTNG_RUN = 6,
 122         LTTNG_DEAD = 7,
 123 };
 124
 125 static
 126 int lttng_enumerate_block_devices(struct lttng_session *session)
 127 {
 128         struct class *ptr_block_class;
 129         struct device_type *ptr_disk_type;
 130         struct class_dev_iter iter;
 131         struct device *dev;
 132
 133         ptr_block_class = wrapper_get_block_class();
 134         if (!ptr_block_class)
 135                 return -ENOSYS;
 136         ptr_disk_type = wrapper_get_disk_type();
 137         if (!ptr_disk_type) {
 138                 return -ENOSYS;
 139         }
 140         class_dev_iter_init(&iter, ptr_block_class, NULL, ptr_disk_type);
 141         while ((dev = class_dev_iter_next(&iter))) {
 142                 struct disk_part_iter piter;
 143                 struct gendisk *disk = dev_to_disk(dev);
 144                 struct hd_struct *part;
 145
 146                 /*
 147                  * Don't show empty devices or things that have been
 148                  * suppressed
 149                  */
 150                 if (get_capacity(disk) == 0 ||
 151                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 152                         continue;
 153
 154                 disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 155                 while ((part = disk_part_iter_next(&piter))) {
 156                         char name_buf[BDEVNAME_SIZE];
 157                         char *p;
 158
 159                         p = wrapper_disk_name(disk, part->partno, name_buf);
 160                         if (!p) {
 161                                 disk_part_iter_exit(&piter);
 162                                 class_dev_iter_exit(&iter);
 163                                 return -ENOSYS;
 164                         }
 165                         trace_lttng_statedump_block_device(session,
 166                                         part_devt(part), name_buf);
 167                 }
 168                 disk_part_iter_exit(&piter);
 169         }
 170         class_dev_iter_exit(&iter);
 171         return 0;
 172 }
 173
 174 #ifdef CONFIG_INET
 175
 176 static
 177 void lttng_enumerate_device(struct lttng_session *session,
 178                 struct net_device *dev)
 179 {
 180         struct in_device *in_dev;
 181         struct in_ifaddr *ifa;
 182
 183         if (dev->flags & IFF_UP) {
 184                 in_dev = in_dev_get(dev);
 185                 if (in_dev) {
 186                         for (ifa = in_dev->ifa_list; ifa != NULL;
 187                              ifa = ifa->ifa_next) {
 188                                 trace_lttng_statedump_network_interface(
 189                                         session, dev, ifa);
 190                         }
 191                         in_dev_put(in_dev);
 192                 }
 193         } else {
 194                 trace_lttng_statedump_network_interface(
 195                         session, dev, NULL);
 196         }
 197 }
 198
 199 static
 200 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 201 {
 202         struct net_device *dev;
 203
 204         read_lock(&dev_base_lock);
 205         for_each_netdev(&init_net, dev)
 206                 lttng_enumerate_device(session, dev);
 207         read_unlock(&dev_base_lock);
 208
 209         return 0;
 210 }
 211 #else /* CONFIG_INET */
 212 static inline
 213 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 214 {
 215         return 0;
 216 }
 217 #endif /* CONFIG_INET */
 218
 219 static
 220 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 221 {
 222         const struct lttng_fd_ctx *ctx = p;
 223         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 224         unsigned int flags = file->f_flags;
 225         struct fdtable *fdt;
 226
 227         /*
 228          * We don't expose kernel internal flags, only userspace-visible
 229          * flags.
 230          */
 231         flags &= ~FMODE_NONOTIFY;
 232         fdt = files_fdtable(ctx->files);
 233         /*
 234          * We need to check here again whether fd is within the fdt
 235          * max_fds range, because we might be seeing a different
 236          * files_fdtable() than iterate_fd(), assuming only RCU is
 237          * protecting the read. In reality, iterate_fd() holds
 238          * file_lock, which should ensure the fdt does not change while
 239          * the lock is taken, but we are not aware whether this is
 240          * guaranteed or not, so play safe.
 241          */
 242         if (fd < fdt->max_fds && lttng_close_on_exec(fd, fdt))
 243                 flags |= O_CLOEXEC;
 244         if (IS_ERR(s)) {
 245                 struct dentry *dentry = file->f_path.dentry;
 246
 247                 /* Make sure we give at least some info */
 248                 spin_lock(&dentry->d_lock);
 249                 trace_lttng_statedump_file_descriptor(ctx->session, ctx->p, fd,
 250                         dentry->d_name.name, flags, file->f_mode);
 251                 spin_unlock(&dentry->d_lock);
 252                 goto end;
 253         }
 254         trace_lttng_statedump_file_descriptor(ctx->session, ctx->p, fd, s,
 255                 flags, file->f_mode);
 256 end:
 257         return 0;
 258 }
 259
 260 static
 261 void lttng_enumerate_task_fd(struct lttng_session *session,
 262                 struct task_struct *p, char *tmp)
 263 {
 264         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .p = p };
 265         struct files_struct *files;
 266
 267         task_lock(p);
 268         files = p->files;
 269         if (!files)
 270                 goto end;
 271         ctx.files = files;
 272         lttng_iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 273 end:
 274         task_unlock(p);
 275 }
 276
 277 static
 278 int lttng_enumerate_file_descriptors(struct lttng_session *session)
 279 {
 280         struct task_struct *p;
 281         char *tmp;
 282
 283         tmp = (char *) __get_free_page(GFP_KERNEL);
 284         if (!tmp)
 285                 return -ENOMEM;
 286
 287         /* Enumerate active file descriptors */
 288         rcu_read_lock();
 289         for_each_process(p)
 290                 lttng_enumerate_task_fd(session, p, tmp);
 291         rcu_read_unlock();
 292         free_page((unsigned long) tmp);
 293         return 0;
 294 }
 295
 296 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
 297 static
 298 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 299 {
 300         int cpu;
 301         const cpumask_t *cpumask = cpu_possible_mask;
 302
 303         for (cpu = cpumask_first(cpumask); cpu < nr_cpu_ids;
 304                         cpu = cpumask_next(cpu, cpumask)) {
 305                 trace_lttng_statedump_cpu_topology(session, &cpu_data(cpu));
 306         }
 307
 308         return 0;
 309 }
 310 #else
 311 static
 312 int lttng_enumerate_cpu_topology(struct lttng_session *session)
 313 {
 314         return 0;
 315 }
 316 #endif
 317
 318 #if 0
 319 /*
 320  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 321  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 322  * iteration, but it is not exported to modules.
 323  */
 324 static
 325 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 326                 struct task_struct *p)
 327 {
 328         struct mm_struct *mm;
 329         struct vm_area_struct *map;
 330         unsigned long ino;
 331
 332         /* get_task_mm does a task_lock... */
 333         mm = get_task_mm(p);
 334         if (!mm)
 335                 return;
 336
 337         map = mm->mmap;
 338         if (map) {
 339                 down_read(&mm->mmap_sem);
 340                 while (map) {
 341                         if (map->vm_file)
 342                                 ino = map->vm_file->lttng_f_dentry->d_inode->i_ino;
 343                         else
 344                                 ino = 0;
 345                         trace_lttng_statedump_vm_map(session, p, map, ino);
 346                         map = map->vm_next;
 347                 }
 348                 up_read(&mm->mmap_sem);
 349         }
 350         mmput(mm);
 351 }
 352
 353 static
 354 int lttng_enumerate_vm_maps(struct lttng_session *session)
 355 {
 356         struct task_struct *p;
 357
 358         rcu_read_lock();
 359         for_each_process(p)
 360                 lttng_enumerate_task_vm_maps(session, p);
 361         rcu_read_unlock();
 362         return 0;
 363 }
 364 #endif
 365
 366 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
 367
 368 static
 369 int lttng_list_interrupts(struct lttng_session *session)
 370 {
 371         unsigned int irq;
 372         unsigned long flags = 0;
 373         struct irq_desc *desc;
 374
 375 #define irq_to_desc     wrapper_irq_to_desc
 376         /* needs irq_desc */
 377         for_each_irq_desc(irq, desc) {
 378                 struct irqaction *action;
 379                 const char *irq_chip_name =
 380                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 381
 382                 local_irq_save(flags);
 383                 raw_spin_lock(&desc->lock);
 384                 for (action = desc->action; action; action = action->next) {
 385                         trace_lttng_statedump_interrupt(session,
 386                                 irq, irq_chip_name, action);
 387                 }
 388                 raw_spin_unlock(&desc->lock);
 389                 local_irq_restore(flags);
 390         }
 391         return 0;
 392 #undef irq_to_desc
 393 }
 394 #else
 395 static inline
 396 int lttng_list_interrupts(struct lttng_session *session)
 397 {
 398         return 0;
 399 }
 400 #endif
 401
 402 /*
 403  * Statedump the task's namespaces using the proc filesystem inode number as
 404  * the unique identifier. The user and pid ns are nested and will be dumped
 405  * recursively.
 406  *
 407  * Called with task lock held.
 408  */
 409 static
 410 void lttng_statedump_process_ns(struct lttng_session *session,
 411                 struct task_struct *p,
 412                 enum lttng_thread_type type,
 413                 enum lttng_execution_mode mode,
 414                 enum lttng_execution_submode submode,
 415                 enum lttng_process_status status)
 416 {
 417         struct nsproxy *proxy;
 418         struct pid_namespace *pid_ns;
 419         struct user_namespace *user_ns;
 420
 421         /*
 422          * The pid and user namespaces are special, they are nested and
 423          * accessed with specific functions instead of the nsproxy struct
 424          * like the other namespaces.
 425          */
 426         pid_ns = task_active_pid_ns(p);
 427         do {
 428                 trace_lttng_statedump_process_pid_ns(session, p, pid_ns);
 429                 pid_ns = pid_ns->parent;
 430         } while (pid_ns);
 431
 432
 433         user_ns = task_cred_xxx(p, user_ns);
 434         do {
 435                 trace_lttng_statedump_process_user_ns(session, p, user_ns);
 436                 user_ns = user_ns->lttng_user_ns_parent;
 437         } while (user_ns);
 438
 439         /*
 440          * Back and forth on locking strategy within Linux upstream for nsproxy.
 441          * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
 442          * "namespaces: Use task_lock and not rcu to protect nsproxy"
 443          * for details.
 444          */
 445 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 446                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 447                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 448                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 449         proxy = p->nsproxy;
 450 #else
 451         rcu_read_lock();
 452         proxy = task_nsproxy(p);
 453 #endif
 454         if (proxy) {
 455 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
 456                 trace_lttng_statedump_process_cgroup_ns(session, p, proxy->cgroup_ns);
 457 #endif
 458                 trace_lttng_statedump_process_ipc_ns(session, p, proxy->ipc_ns);
 459 #ifndef LTTNG_MNT_NS_MISSING_HEADER
 460                 trace_lttng_statedump_process_mnt_ns(session, p, proxy->mnt_ns);
 461 #endif
 462                 trace_lttng_statedump_process_net_ns(session, p, proxy->net_ns);
 463                 trace_lttng_statedump_process_uts_ns(session, p, proxy->uts_ns);
 464         }
 465 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
 466                 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
 467                 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
 468                 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
 469         /* (nothing) */
 470 #else
 471         rcu_read_unlock();
 472 #endif
 473 }
 474
 475 static
 476 int lttng_enumerate_process_states(struct lttng_session *session)
 477 {
 478         struct task_struct *g, *p;
 479
 480         rcu_read_lock();
 481         for_each_process(g) {
 482                 p = g;
 483                 do {
 484                         enum lttng_execution_mode mode =
 485                                 LTTNG_MODE_UNKNOWN;
 486                         enum lttng_execution_submode submode =
 487                                 LTTNG_UNKNOWN;
 488                         enum lttng_process_status status;
 489                         enum lttng_thread_type type;
 490
 491                         task_lock(p);
 492                         if (p->exit_state == EXIT_ZOMBIE)
 493                                 status = LTTNG_ZOMBIE;
 494                         else if (p->exit_state == EXIT_DEAD)
 495                                 status = LTTNG_DEAD;
 496                         else if (p->state == TASK_RUNNING) {
 497                                 /* Is this a forked child that has not run yet? */
 498                                 if (list_empty(&p->rt.run_list))
 499                                         status = LTTNG_WAIT_FORK;
 500                                 else
 501                                         /*
 502                                          * All tasks are considered as wait_cpu;
 503                                          * the viewer will sort out if the task
 504                                          * was really running at this time.
 505                                          */
 506                                         status = LTTNG_WAIT_CPU;
 507                         } else if (p->state &
 508                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 509                                 /* Task is waiting for something to complete */
 510                                 status = LTTNG_WAIT;
 511                         } else
 512                                 status = LTTNG_UNNAMED;
 513                         submode = LTTNG_NONE;
 514
 515                         /*
 516                          * Verification of t->mm is to filter out kernel
 517                          * threads; Viewer will further filter out if a
 518                          * user-space thread was in syscall mode or not.
 519                          */
 520                         if (p->mm)
 521                                 type = LTTNG_USER_THREAD;
 522                         else
 523                                 type = LTTNG_KERNEL_THREAD;
 524
 525                         trace_lttng_statedump_process_state(session,
 526                                 p, type, mode, submode, status);
 527                         lttng_statedump_process_ns(session,
 528                                 p, type, mode, submode, status);
 529                         task_unlock(p);
 530                 } while_each_thread(g, p);
 531         }
 532         rcu_read_unlock();
 533
 534         return 0;
 535 }
 536
 537 static
 538 void lttng_statedump_work_func(struct work_struct *work)
 539 {
 540         if (atomic_dec_and_test(&kernel_threads_to_run))
 541                 /* If we are the last thread, wake up do_lttng_statedump */
 542                 wake_up(&statedump_wq);
 543 }
 544
 545 static
 546 int do_lttng_statedump(struct lttng_session *session)
 547 {
 548         int cpu, ret;
 549
 550         trace_lttng_statedump_start(session);
 551         ret = lttng_enumerate_process_states(session);
 552         if (ret)
 553                 return ret;
 554         ret = lttng_enumerate_file_descriptors(session);
 555         if (ret)
 556                 return ret;
 557         /*
 558          * FIXME
 559          * ret = lttng_enumerate_vm_maps(session);
 560          * if (ret)
 561          *      return ret;
 562          */
 563         ret = lttng_list_interrupts(session);
 564         if (ret)
 565                 return ret;
 566         ret = lttng_enumerate_network_ip_interface(session);
 567         if (ret)
 568                 return ret;
 569         ret = lttng_enumerate_block_devices(session);
 570         switch (ret) {
 571         case 0:
 572                 break;
 573         case -ENOSYS:
 574                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 575                 break;
 576         default:
 577                 return ret;
 578         }
 579         ret = lttng_enumerate_cpu_topology(session);
 580         if (ret)
 581                 return ret;
 582
 583         /* TODO lttng_dump_idt_table(session); */
 584         /* TODO lttng_dump_softirq_vec(session); */
 585         /* TODO lttng_list_modules(session); */
 586         /* TODO lttng_dump_swap_files(session); */
 587
 588         /*
 589          * Fire off a work queue on each CPU. Their sole purpose in life
 590          * is to guarantee that each CPU has been in a state where is was in
 591          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 592          */
 593         get_online_cpus();
 594         atomic_set(&kernel_threads_to_run, num_online_cpus());
 595         for_each_online_cpu(cpu) {
 596                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 597                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 598         }
 599         /* Wait for all threads to run */
 600         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 601         put_online_cpus();
 602         /* Our work is done */
 603         trace_lttng_statedump_end(session);
 604         return 0;
 605 }
 606
 607 /*
 608  * Called with session mutex held.
 609  */
 610 int lttng_statedump_start(struct lttng_session *session)
 611 {
 612         return do_lttng_statedump(session);
 613 }
 614 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 615
 616 static
 617 int __init lttng_statedump_init(void)
 618 {
 619         /*
 620          * Allow module to load even if the fixup cannot be done. This
 621          * will allow seemless transition when the underlying issue fix
 622          * is merged into the Linux kernel, and when tracepoint.c
 623          * "tracepoint_module_notify" is turned into a static function.
 624          */
 625         (void) wrapper_lttng_fixup_sig(THIS_MODULE);
 626         return 0;
 627 }
 628
 629 module_init(lttng_statedump_init);
 630
 631 static
 632 void __exit lttng_statedump_exit(void)
 633 {
 634 }
 635
 636 module_exit(lttng_statedump_exit);
 637
 638 MODULE_LICENSE("GPL and additional rights");
 639 MODULE_AUTHOR("Jean-Hugues Deschenes");
 640 MODULE_DESCRIPTION("LTTng statedump provider");
 641 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 642         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 643         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 644         LTTNG_MODULES_EXTRAVERSION);