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/mm.h>
  30 #include <linux/swap.h>
  31 #include <linux/wait.h>
  32 #include <linux/mutex.h>
  33 #include <linux/device.h>
  34
  35 #include <lttng-events.h>
  36 #include <lttng-tracer.h>
  37 #include <wrapper/irqdesc.h>
  38 #include <wrapper/spinlock.h>
  39 #include <wrapper/fdtable.h>
  40 #include <wrapper/irq.h>
  41 #include <wrapper/tracepoint.h>
  42 #include <wrapper/genhd.h>
  43 #include <wrapper/file.h>
  44 #include <wrapper/fdtable.h>
  45 #include <wrapper/sched.h>
  46
  47 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
  48 #include <linux/irq.h>
  49 #endif
  50
  51 /* Define the tracepoints, but do not build the probes */
  52 #define CREATE_TRACE_POINTS
  53 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
  54 #define TRACE_INCLUDE_FILE lttng-statedump
  55 #define LTTNG_INSTRUMENTATION
  56 #include <instrumentation/events/lttng-module/lttng-statedump.h>
  57
  58 LTTNG_DEFINE_TRACE(lttng_statedump_block_device,
  59         TP_PROTO(struct lttng_session *session,
  60                 dev_t dev, const char *diskname),
  61         TP_ARGS(session, dev, diskname));
  62
  63 LTTNG_DEFINE_TRACE(lttng_statedump_end,
  64         TP_PROTO(struct lttng_session *session),
  65         TP_ARGS(session));
  66
  67 LTTNG_DEFINE_TRACE(lttng_statedump_interrupt,
  68         TP_PROTO(struct lttng_session *session,
  69                 unsigned int irq, const char *chip_name,
  70                 struct irqaction *action),
  71         TP_ARGS(session, irq, chip_name, action));
  72
  73 LTTNG_DEFINE_TRACE(lttng_statedump_file_descriptor,
  74         TP_PROTO(struct lttng_session *session,
  75                 struct task_struct *p, int fd, const char *filename,
  76                 unsigned int flags, fmode_t fmode),
  77         TP_ARGS(session, p, fd, filename, flags, fmode));
  78
  79 LTTNG_DEFINE_TRACE(lttng_statedump_start,
  80         TP_PROTO(struct lttng_session *session),
  81         TP_ARGS(session));
  82
  83 LTTNG_DEFINE_TRACE(lttng_statedump_process_state,
  84         TP_PROTO(struct lttng_session *session,
  85                 struct task_struct *p,
  86                 int type, int mode, int submode, int status,
  87                 struct pid_namespace *pid_ns),
  88         TP_ARGS(session, p, type, mode, submode, status, pid_ns));
  89
  90 LTTNG_DEFINE_TRACE(lttng_statedump_network_interface,
  91         TP_PROTO(struct lttng_session *session,
  92                 struct net_device *dev, struct in_ifaddr *ifa),
  93         TP_ARGS(session, dev, ifa));
  94
  95 struct lttng_fd_ctx {
  96         char *page;
  97         struct lttng_session *session;
  98         struct task_struct *p;
  99         struct files_struct *files;
 100 };
 101
 102 /*
 103  * Protected by the trace lock.
 104  */
 105 static struct delayed_work cpu_work[NR_CPUS];
 106 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq);
 107 static atomic_t kernel_threads_to_run;
 108
 109 enum lttng_thread_type {
 110         LTTNG_USER_THREAD = 0,
 111         LTTNG_KERNEL_THREAD = 1,
 112 };
 113
 114 enum lttng_execution_mode {
 115         LTTNG_USER_MODE = 0,
 116         LTTNG_SYSCALL = 1,
 117         LTTNG_TRAP = 2,
 118         LTTNG_IRQ = 3,
 119         LTTNG_SOFTIRQ = 4,
 120         LTTNG_MODE_UNKNOWN = 5,
 121 };
 122
 123 enum lttng_execution_submode {
 124         LTTNG_NONE = 0,
 125         LTTNG_UNKNOWN = 1,
 126 };
 127
 128 enum lttng_process_status {
 129         LTTNG_UNNAMED = 0,
 130         LTTNG_WAIT_FORK = 1,
 131         LTTNG_WAIT_CPU = 2,
 132         LTTNG_EXIT = 3,
 133         LTTNG_ZOMBIE = 4,
 134         LTTNG_WAIT = 5,
 135         LTTNG_RUN = 6,
 136         LTTNG_DEAD = 7,
 137 };
 138
 139
 140 #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(5,11,0))
 141
 142 #define LTTNG_PART_STRUCT_TYPE struct block_device
 143
 144 static
 145 int lttng_get_part_name(struct gendisk *disk, struct block_device *part, char *name_buf)
 146 {
 147         const char *p;
 148
 149         p = bdevname(part, name_buf);
 150         if (!p)
 151                 return -ENOSYS;
 152
 153         return 0;
 154 }
 155
 156 static
 157 dev_t lttng_get_part_devt(struct block_device *part)
 158 {
 159         return part->bd_dev;
 160 }
 161
 162 #else
 163
 164 #define LTTNG_PART_STRUCT_TYPE struct hd_struct
 165
 166 static
 167 int lttng_get_part_name(struct gendisk *disk, struct hd_struct *part, char *name_buf)
 168 {
 169         const char *p;
 170         struct block_device bdev;
 171
 172         /*
 173          * Create a partial 'struct blockdevice' to use
 174          * 'bdevname()' which is a simple wrapper over
 175          * 'disk_name()' but has the honor to be EXPORT_SYMBOL.
 176          */
 177         bdev.bd_disk = disk;
 178         bdev.bd_part = part;
 179
 180         p = bdevname(&bdev, name_buf);
 181         if (!p)
 182                 return -ENOSYS;
 183
 184         return 0;
 185 }
 186
 187 static
 188 dev_t lttng_get_part_devt(struct hd_struct *part)
 189 {
 190         return part_devt(part);
 191 }
 192 #endif
 193
 194 #if (LTTNG_LINUX_VERSION_CODE >= LTTNG_KERNEL_VERSION(5,12,0))
 195 static
 196 int lttng_statedump_each_block_device(struct lttng_session *session, struct gendisk *disk)
 197 {
 198         struct block_device *part;
 199         unsigned long idx;
 200         int ret = 0;
 201
 202         /* Include partition 0 */
 203         idx = 0;
 204
 205         rcu_read_lock();
 206         xa_for_each(&disk->part_tbl, idx, part) {
 207                 char name_buf[BDEVNAME_SIZE];
 208
 209                 /* Exclude non-partitions bdev and empty partitions. */
 210                 if (bdev_is_partition(part) && !bdev_nr_sectors(part))
 211                         continue;
 212
 213                 if (lttng_get_part_name(disk, part, name_buf) == -ENOSYS) {
 214                         ret = -ENOSYS;
 215                         goto end;
 216                 }
 217                 trace_lttng_statedump_block_device(session, lttng_get_part_devt(part),
 218                                 name_buf);
 219         }
 220 end:
 221         rcu_read_unlock();
 222         return ret;
 223 }
 224 #else
 225 static
 226 int lttng_statedump_each_block_device(struct lttng_session *session, struct gendisk *disk)
 227 {
 228         struct disk_part_iter piter;
 229         LTTNG_PART_STRUCT_TYPE *part;
 230
 231         disk_part_iter_init(&piter, disk, DISK_PITER_INCL_PART0);
 232
 233         while ((part = disk_part_iter_next(&piter))) {
 234                 char name_buf[BDEVNAME_SIZE];
 235
 236                 if (lttng_get_part_name(disk, part, name_buf) == -ENOSYS) {
 237                         disk_part_iter_exit(&piter);
 238                         return -ENOSYS;
 239                 }
 240                 trace_lttng_statedump_block_device(session, lttng_get_part_devt(part),
 241                                 name_buf);
 242         }
 243         disk_part_iter_exit(&piter);
 244
 245         return 0;
 246 }
 247 #endif
 248
 249 static
 250 int lttng_enumerate_block_devices(struct lttng_session *session)
 251 {
 252         struct class *ptr_block_class;
 253         struct device_type *ptr_disk_type;
 254         struct class_dev_iter iter;
 255         struct device *dev;
 256         int ret = 0;
 257
 258         ptr_block_class = wrapper_get_block_class();
 259         if (!ptr_block_class) {
 260                 ret = -ENOSYS;
 261                 goto end;
 262         }
 263         ptr_disk_type = wrapper_get_disk_type();
 264         if (!ptr_disk_type) {
 265                 ret = -ENOSYS;
 266                 goto end;
 267         }
 268         class_dev_iter_init(&iter, ptr_block_class, NULL, ptr_disk_type);
 269         while ((dev = class_dev_iter_next(&iter))) {
 270                 struct gendisk *disk = dev_to_disk(dev);
 271
 272                 /*
 273                  * Don't show empty devices or things that have been
 274                  * suppressed
 275                  */
 276                 if (get_capacity(disk) == 0 ||
 277                     (disk->flags & GENHD_FL_SUPPRESS_PARTITION_INFO))
 278                         continue;
 279
 280                 ret = lttng_statedump_each_block_device(session, disk);
 281         }
 282         class_dev_iter_exit(&iter);
 283 end:
 284         return ret;
 285 }
 286
 287 #ifdef CONFIG_INET
 288
 289 static
 290 void lttng_enumerate_device(struct lttng_session *session,
 291                 struct net_device *dev)
 292 {
 293         struct in_device *in_dev;
 294         struct in_ifaddr *ifa;
 295
 296         if (dev->flags & IFF_UP) {
 297                 in_dev = in_dev_get(dev);
 298                 if (in_dev) {
 299                         for (ifa = in_dev->ifa_list; ifa != NULL;
 300                              ifa = ifa->ifa_next) {
 301                                 trace_lttng_statedump_network_interface(
 302                                         session, dev, ifa);
 303                         }
 304                         in_dev_put(in_dev);
 305                 }
 306         } else {
 307                 trace_lttng_statedump_network_interface(
 308                         session, dev, NULL);
 309         }
 310 }
 311
 312 static
 313 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 314 {
 315         struct net_device *dev;
 316
 317         read_lock(&dev_base_lock);
 318         for_each_netdev(&init_net, dev)
 319                 lttng_enumerate_device(session, dev);
 320         read_unlock(&dev_base_lock);
 321
 322         return 0;
 323 }
 324 #else /* CONFIG_INET */
 325 static inline
 326 int lttng_enumerate_network_ip_interface(struct lttng_session *session)
 327 {
 328         return 0;
 329 }
 330 #endif /* CONFIG_INET */
 331
 332 static
 333 int lttng_dump_one_fd(const void *p, struct file *file, unsigned int fd)
 334 {
 335         const struct lttng_fd_ctx *ctx = p;
 336         const char *s = d_path(&file->f_path, ctx->page, PAGE_SIZE);
 337         unsigned int flags = file->f_flags;
 338         struct fdtable *fdt;
 339
 340         /*
 341          * We don't expose kernel internal flags, only userspace-visible
 342          * flags.
 343          */
 344         flags &= ~FMODE_NONOTIFY;
 345         fdt = files_fdtable(ctx->files);
 346         /*
 347          * We need to check here again whether fd is within the fdt
 348          * max_fds range, because we might be seeing a different
 349          * files_fdtable() than iterate_fd(), assuming only RCU is
 350          * protecting the read. In reality, iterate_fd() holds
 351          * file_lock, which should ensure the fdt does not change while
 352          * the lock is taken, but we are not aware whether this is
 353          * guaranteed or not, so play safe.
 354          */
 355         if (fd < fdt->max_fds && lttng_close_on_exec(fd, fdt))
 356                 flags |= O_CLOEXEC;
 357         if (IS_ERR(s)) {
 358                 struct dentry *dentry = file->f_path.dentry;
 359
 360                 /* Make sure we give at least some info */
 361                 spin_lock(&dentry->d_lock);
 362                 trace_lttng_statedump_file_descriptor(ctx->session, ctx->p, fd,
 363                         dentry->d_name.name, flags, file->f_mode);
 364                 spin_unlock(&dentry->d_lock);
 365                 goto end;
 366         }
 367         trace_lttng_statedump_file_descriptor(ctx->session, ctx->p, fd, s,
 368                 flags, file->f_mode);
 369 end:
 370         return 0;
 371 }
 372
 373 static
 374 void lttng_enumerate_task_fd(struct lttng_session *session,
 375                 struct task_struct *p, char *tmp)
 376 {
 377         struct lttng_fd_ctx ctx = { .page = tmp, .session = session, .p = p };
 378         struct files_struct *files;
 379
 380         task_lock(p);
 381         files = p->files;
 382         if (!files)
 383                 goto end;
 384         ctx.files = files;
 385         lttng_iterate_fd(files, 0, lttng_dump_one_fd, &ctx);
 386 end:
 387         task_unlock(p);
 388 }
 389
 390 static
 391 int lttng_enumerate_file_descriptors(struct lttng_session *session)
 392 {
 393         struct task_struct *p;
 394         char *tmp;
 395
 396         tmp = (char *) __get_free_page(GFP_KERNEL);
 397         if (!tmp)
 398                 return -ENOMEM;
 399
 400         /* Enumerate active file descriptors */
 401         rcu_read_lock();
 402         for_each_process(p)
 403                 lttng_enumerate_task_fd(session, p, tmp);
 404         rcu_read_unlock();
 405         free_page((unsigned long) tmp);
 406         return 0;
 407 }
 408
 409 #if 0
 410 /*
 411  * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
 412  * (scheduling in atomic). Normally, the tasklist lock protects this kind of
 413  * iteration, but it is not exported to modules.
 414  */
 415 static
 416 void lttng_enumerate_task_vm_maps(struct lttng_session *session,
 417                 struct task_struct *p)
 418 {
 419         struct mm_struct *mm;
 420         struct vm_area_struct *map;
 421         unsigned long ino;
 422
 423         /* get_task_mm does a task_lock... */
 424         mm = get_task_mm(p);
 425         if (!mm)
 426                 return;
 427
 428         map = mm->mmap;
 429         if (map) {
 430                 down_read(&mm->mmap_sem);
 431                 while (map) {
 432                         if (map->vm_file)
 433                                 ino = map->vm_file->lttng_f_dentry->d_inode->i_ino;
 434                         else
 435                                 ino = 0;
 436                         trace_lttng_statedump_vm_map(session, p, map, ino);
 437                         map = map->vm_next;
 438                 }
 439                 up_read(&mm->mmap_sem);
 440         }
 441         mmput(mm);
 442 }
 443
 444 static
 445 int lttng_enumerate_vm_maps(struct lttng_session *session)
 446 {
 447         struct task_struct *p;
 448
 449         rcu_read_lock();
 450         for_each_process(p)
 451                 lttng_enumerate_task_vm_maps(session, p);
 452         rcu_read_unlock();
 453         return 0;
 454 }
 455 #endif
 456
 457 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
 458
 459 #if (LTTNG_LINUX_VERSION_CODE < LTTNG_KERNEL_VERSION(2,6,39))
 460 #define irq_desc_get_chip(desc) get_irq_desc_chip(desc)
 461 #endif
 462
 463 static
 464 int lttng_list_interrupts(struct lttng_session *session)
 465 {
 466         unsigned int irq;
 467         unsigned long flags = 0;
 468         struct irq_desc *desc;
 469
 470 #define irq_to_desc     wrapper_irq_to_desc
 471         /* needs irq_desc */
 472         for_each_irq_desc(irq, desc) {
 473                 struct irqaction *action;
 474                 const char *irq_chip_name =
 475                         irq_desc_get_chip(desc)->name ? : "unnamed_irq_chip";
 476
 477                 local_irq_save(flags);
 478                 wrapper_desc_spin_lock(&desc->lock);
 479                 for (action = desc->action; action; action = action->next) {
 480                         trace_lttng_statedump_interrupt(session,
 481                                 irq, irq_chip_name, action);
 482                 }
 483                 wrapper_desc_spin_unlock(&desc->lock);
 484                 local_irq_restore(flags);
 485         }
 486         return 0;
 487 #undef irq_to_desc
 488 }
 489 #else
 490 static inline
 491 int lttng_list_interrupts(struct lttng_session *session)
 492 {
 493         return 0;
 494 }
 495 #endif
 496
 497 /*
 498  * Called with task lock held.
 499  */
 500 static
 501 void lttng_statedump_process_ns(struct lttng_session *session,
 502                 struct task_struct *p,
 503                 enum lttng_thread_type type,
 504                 enum lttng_execution_mode mode,
 505                 enum lttng_execution_submode submode,
 506                 enum lttng_process_status status)
 507 {
 508         struct pid_namespace *pid_ns;
 509
 510         pid_ns = task_active_pid_ns(p);
 511         do {
 512                 trace_lttng_statedump_process_state(session,
 513                         p, type, mode, submode, status, pid_ns);
 514                 pid_ns = pid_ns ? pid_ns->parent : NULL;
 515         } while (pid_ns);
 516 }
 517
 518 static
 519 int lttng_enumerate_process_states(struct lttng_session *session)
 520 {
 521         struct task_struct *g, *p;
 522
 523         rcu_read_lock();
 524         for_each_process(g) {
 525                 p = g;
 526                 do {
 527                         enum lttng_execution_mode mode =
 528                                 LTTNG_MODE_UNKNOWN;
 529                         enum lttng_execution_submode submode =
 530                                 LTTNG_UNKNOWN;
 531                         enum lttng_process_status status;
 532                         enum lttng_thread_type type;
 533
 534                         task_lock(p);
 535                         if (p->exit_state == EXIT_ZOMBIE)
 536                                 status = LTTNG_ZOMBIE;
 537                         else if (p->exit_state == EXIT_DEAD)
 538                                 status = LTTNG_DEAD;
 539                         else if (lttng_task_is_running(p)) {
 540                                 /* Is this a forked child that has not run yet? */
 541                                 if (list_empty(&p->rt.run_list))
 542                                         status = LTTNG_WAIT_FORK;
 543                                 else
 544                                         /*
 545                                          * All tasks are considered as wait_cpu;
 546                                          * the viewer will sort out if the task
 547                                          * was really running at this time.
 548                                          */
 549                                         status = LTTNG_WAIT_CPU;
 550                         } else if (lttng_get_task_state(p) &
 551                                 (TASK_INTERRUPTIBLE | TASK_UNINTERRUPTIBLE)) {
 552                                 /* Task is waiting for something to complete */
 553                                 status = LTTNG_WAIT;
 554                         } else
 555                                 status = LTTNG_UNNAMED;
 556                         submode = LTTNG_NONE;
 557
 558                         /*
 559                          * Verification of t->mm is to filter out kernel
 560                          * threads; Viewer will further filter out if a
 561                          * user-space thread was in syscall mode or not.
 562                          */
 563                         if (p->mm)
 564                                 type = LTTNG_USER_THREAD;
 565                         else
 566                                 type = LTTNG_KERNEL_THREAD;
 567                         lttng_statedump_process_ns(session,
 568                                 p, type, mode, submode, status);
 569                         task_unlock(p);
 570                 } while_each_thread(g, p);
 571         }
 572         rcu_read_unlock();
 573
 574         return 0;
 575 }
 576
 577 static
 578 void lttng_statedump_work_func(struct work_struct *work)
 579 {
 580         if (atomic_dec_and_test(&kernel_threads_to_run))
 581                 /* If we are the last thread, wake up do_lttng_statedump */
 582                 wake_up(&statedump_wq);
 583 }
 584
 585 static
 586 int do_lttng_statedump(struct lttng_session *session)
 587 {
 588         int cpu, ret;
 589
 590         trace_lttng_statedump_start(session);
 591         ret = lttng_enumerate_process_states(session);
 592         if (ret)
 593                 return ret;
 594         ret = lttng_enumerate_file_descriptors(session);
 595         if (ret)
 596                 return ret;
 597         /*
 598          * FIXME
 599          * ret = lttng_enumerate_vm_maps(session);
 600          * if (ret)
 601          *      return ret;
 602          */
 603         ret = lttng_list_interrupts(session);
 604         if (ret)
 605                 return ret;
 606         ret = lttng_enumerate_network_ip_interface(session);
 607         if (ret)
 608                 return ret;
 609         ret = lttng_enumerate_block_devices(session);
 610         switch (ret) {
 611         case 0:
 612                 break;
 613         case -ENOSYS:
 614                 printk(KERN_WARNING "LTTng: block device enumeration is not supported by kernel\n");
 615                 break;
 616         default:
 617                 return ret;
 618         }
 619
 620         /* TODO lttng_dump_idt_table(session); */
 621         /* TODO lttng_dump_softirq_vec(session); */
 622         /* TODO lttng_list_modules(session); */
 623         /* TODO lttng_dump_swap_files(session); */
 624
 625         /*
 626          * Fire off a work queue on each CPU. Their sole purpose in life
 627          * is to guarantee that each CPU has been in a state where is was in
 628          * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
 629          */
 630         get_online_cpus();
 631         atomic_set(&kernel_threads_to_run, num_online_cpus());
 632         for_each_online_cpu(cpu) {
 633                 INIT_DELAYED_WORK(&cpu_work[cpu], lttng_statedump_work_func);
 634                 schedule_delayed_work_on(cpu, &cpu_work[cpu], 0);
 635         }
 636         /* Wait for all threads to run */
 637         __wait_event(statedump_wq, (atomic_read(&kernel_threads_to_run) == 0));
 638         put_online_cpus();
 639         /* Our work is done */
 640         trace_lttng_statedump_end(session);
 641         return 0;
 642 }
 643
 644 /*
 645  * Called with session mutex held.
 646  */
 647 int lttng_statedump_start(struct lttng_session *session)
 648 {
 649         return do_lttng_statedump(session);
 650 }
 651 EXPORT_SYMBOL_GPL(lttng_statedump_start);
 652
 653 static
 654 int __init lttng_statedump_init(void)
 655 {
 656         /*
 657          * Allow module to load even if the fixup cannot be done. This
 658          * will allow seemless transition when the underlying issue fix
 659          * is merged into the Linux kernel, and when tracepoint.c
 660          * "tracepoint_module_notify" is turned into a static function.
 661          */
 662         (void) wrapper_lttng_fixup_sig(THIS_MODULE);
 663         return 0;
 664 }
 665
 666 module_init(lttng_statedump_init);
 667
 668 static
 669 void __exit lttng_statedump_exit(void)
 670 {
 671 }
 672
 673 module_exit(lttng_statedump_exit);
 674
 675 MODULE_LICENSE("GPL and additional rights");
 676 MODULE_AUTHOR("Jean-Hugues Deschenes");
 677 MODULE_DESCRIPTION("LTTng statedump provider");
 678 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION) "."
 679         __stringify(LTTNG_MODULES_MINOR_VERSION) "."
 680         __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION)
 681         LTTNG_MODULES_EXTRAVERSION);