1 /* SPDX-License-Identifier: (GPL-2.0-only or LGPL-2.1-only)
5 * Linux Trace Toolkit Next Generation Kernel State Dump
7 * Copyright 2005 Jean-Hugues Deschenes <jean-hugues.deschenes@polymtl.ca>
8 * Copyright 2006-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
11 * Eric Clement: Add listing of network IP interface
12 * 2006, 2007 Mathieu Desnoyers Fix kernel threads
16 #include <linux/init.h>
17 #include <linux/module.h>
18 #include <linux/netlink.h>
19 #include <linux/inet.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>
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>
37 #include <lttng-events.h>
38 #include <lttng-tracer.h>
39 #include <wrapper/namespace.h>
40 #include <wrapper/irq.h>
41 #include <wrapper/tracepoint.h>
42 #include <wrapper/genhd.h>
43 #include <wrapper/file.h>
45 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
46 #include <linux/irq.h>
49 /* Define the tracepoints, but do not build the probes */
50 #define CREATE_TRACE_POINTS
51 #define TRACE_INCLUDE_PATH instrumentation/events/lttng-module
52 #define TRACE_INCLUDE_FILE lttng-statedump
53 #define LTTNG_INSTRUMENTATION
54 #include <instrumentation/events/lttng-module/lttng-statedump.h>
56 DEFINE_TRACE(lttng_statedump_block_device
);
57 DEFINE_TRACE(lttng_statedump_end
);
58 DEFINE_TRACE(lttng_statedump_interrupt
);
59 DEFINE_TRACE(lttng_statedump_file_descriptor
);
60 DEFINE_TRACE(lttng_statedump_start
);
61 DEFINE_TRACE(lttng_statedump_process_state
);
62 DEFINE_TRACE(lttng_statedump_process_pid_ns
);
63 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
64 DEFINE_TRACE(lttng_statedump_process_cgroup_ns
);
66 DEFINE_TRACE(lttng_statedump_process_ipc_ns
);
67 #ifndef LTTNG_MNT_NS_MISSING_HEADER
68 DEFINE_TRACE(lttng_statedump_process_mnt_ns
);
70 DEFINE_TRACE(lttng_statedump_process_net_ns
);
71 DEFINE_TRACE(lttng_statedump_process_user_ns
);
72 DEFINE_TRACE(lttng_statedump_process_uts_ns
);
73 DEFINE_TRACE(lttng_statedump_network_interface
);
74 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
75 DEFINE_TRACE(lttng_statedump_cpu_topology
);
80 struct lttng_session
*session
;
81 struct files_struct
*files
;
85 * Protected by the trace lock.
87 static struct delayed_work cpu_work
[NR_CPUS
];
88 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq
);
89 static atomic_t kernel_threads_to_run
;
91 enum lttng_thread_type
{
92 LTTNG_USER_THREAD
= 0,
93 LTTNG_KERNEL_THREAD
= 1,
96 enum lttng_execution_mode
{
102 LTTNG_MODE_UNKNOWN
= 5,
105 enum lttng_execution_submode
{
110 enum lttng_process_status
{
122 int lttng_enumerate_block_devices(struct lttng_session
*session
)
124 struct class *ptr_block_class
;
125 struct device_type
*ptr_disk_type
;
126 struct class_dev_iter iter
;
129 ptr_block_class
= wrapper_get_block_class();
130 if (!ptr_block_class
)
132 ptr_disk_type
= wrapper_get_disk_type();
133 if (!ptr_disk_type
) {
136 class_dev_iter_init(&iter
, ptr_block_class
, NULL
, ptr_disk_type
);
137 while ((dev
= class_dev_iter_next(&iter
))) {
138 struct disk_part_iter piter
;
139 struct gendisk
*disk
= dev_to_disk(dev
);
140 struct hd_struct
*part
;
143 * Don't show empty devices or things that have been
146 if (get_capacity(disk
) == 0 ||
147 (disk
->flags
& GENHD_FL_SUPPRESS_PARTITION_INFO
))
150 disk_part_iter_init(&piter
, disk
, DISK_PITER_INCL_PART0
);
151 while ((part
= disk_part_iter_next(&piter
))) {
152 char name_buf
[BDEVNAME_SIZE
];
155 p
= wrapper_disk_name(disk
, part
->partno
, name_buf
);
157 disk_part_iter_exit(&piter
);
158 class_dev_iter_exit(&iter
);
161 trace_lttng_statedump_block_device(session
,
162 part_devt(part
), name_buf
);
164 disk_part_iter_exit(&piter
);
166 class_dev_iter_exit(&iter
);
173 void lttng_enumerate_device(struct lttng_session
*session
,
174 struct net_device
*dev
)
176 struct in_device
*in_dev
;
177 struct in_ifaddr
*ifa
;
179 if (dev
->flags
& IFF_UP
) {
180 in_dev
= in_dev_get(dev
);
182 for (ifa
= in_dev
->ifa_list
; ifa
!= NULL
;
183 ifa
= ifa
->ifa_next
) {
184 trace_lttng_statedump_network_interface(
190 trace_lttng_statedump_network_interface(
196 int lttng_enumerate_network_ip_interface(struct lttng_session
*session
)
198 struct net_device
*dev
;
200 read_lock(&dev_base_lock
);
201 for_each_netdev(&init_net
, dev
)
202 lttng_enumerate_device(session
, dev
);
203 read_unlock(&dev_base_lock
);
207 #else /* CONFIG_INET */
209 int lttng_enumerate_network_ip_interface(struct lttng_session
*session
)
213 #endif /* CONFIG_INET */
216 int lttng_dump_one_fd(const void *p
, struct file
*file
, unsigned int fd
)
218 const struct lttng_fd_ctx
*ctx
= p
;
219 const char *s
= d_path(&file
->f_path
, ctx
->page
, PAGE_SIZE
);
220 unsigned int flags
= file
->f_flags
;
224 * We don't expose kernel internal flags, only userspace-visible
227 flags
&= ~FMODE_NONOTIFY
;
228 fdt
= files_fdtable(ctx
->files
);
230 * We need to check here again whether fd is within the fdt
231 * max_fds range, because we might be seeing a different
232 * files_fdtable() than iterate_fd(), assuming only RCU is
233 * protecting the read. In reality, iterate_fd() holds
234 * file_lock, which should ensure the fdt does not change while
235 * the lock is taken, but we are not aware whether this is
236 * guaranteed or not, so play safe.
238 if (fd
< fdt
->max_fds
&& close_on_exec(fd
, fdt
))
241 struct dentry
*dentry
= file
->f_path
.dentry
;
243 /* Make sure we give at least some info */
244 spin_lock(&dentry
->d_lock
);
245 trace_lttng_statedump_file_descriptor(ctx
->session
,
246 ctx
->files
, fd
, dentry
->d_name
.name
, flags
,
248 spin_unlock(&dentry
->d_lock
);
251 trace_lttng_statedump_file_descriptor(ctx
->session
,
252 ctx
->files
, fd
, s
, flags
, file
->f_mode
);
257 /* Called with task lock held. */
259 void lttng_enumerate_files(struct lttng_session
*session
,
260 struct files_struct
*files
,
263 struct lttng_fd_ctx ctx
= { .page
= tmp
, .session
= session
, .files
= files
, };
265 iterate_fd(files
, 0, lttng_dump_one_fd
, &ctx
);
268 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
270 int lttng_enumerate_cpu_topology(struct lttng_session
*session
)
273 const cpumask_t
*cpumask
= cpu_possible_mask
;
275 for (cpu
= cpumask_first(cpumask
); cpu
< nr_cpu_ids
;
276 cpu
= cpumask_next(cpu
, cpumask
)) {
277 trace_lttng_statedump_cpu_topology(session
, &cpu_data(cpu
));
284 int lttng_enumerate_cpu_topology(struct lttng_session
*session
)
292 * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
293 * (scheduling in atomic). Normally, the tasklist lock protects this kind of
294 * iteration, but it is not exported to modules.
297 void lttng_enumerate_task_vm_maps(struct lttng_session
*session
,
298 struct task_struct
*p
)
300 struct mm_struct
*mm
;
301 struct vm_area_struct
*map
;
304 /* get_task_mm does a task_lock... */
311 down_read(&mm
->mmap_sem
);
314 ino
= map
->vm_file
->lttng_f_dentry
->d_inode
->i_ino
;
317 trace_lttng_statedump_vm_map(session
, p
, map
, ino
);
320 up_read(&mm
->mmap_sem
);
326 int lttng_enumerate_vm_maps(struct lttng_session
*session
)
328 struct task_struct
*p
;
332 lttng_enumerate_task_vm_maps(session
, p
);
338 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
341 int lttng_list_interrupts(struct lttng_session
*session
)
344 unsigned long flags
= 0;
345 struct irq_desc
*desc
;
348 for_each_irq_desc(irq
, desc
) {
349 struct irqaction
*action
;
350 const char *irq_chip_name
=
351 irq_desc_get_chip(desc
)->name
? : "unnamed_irq_chip";
353 local_irq_save(flags
);
354 raw_spin_lock(&desc
->lock
);
355 for (action
= desc
->action
; action
; action
= action
->next
) {
356 trace_lttng_statedump_interrupt(session
,
357 irq
, irq_chip_name
, action
);
359 raw_spin_unlock(&desc
->lock
);
360 local_irq_restore(flags
);
366 int lttng_list_interrupts(struct lttng_session
*session
)
373 * Statedump the task's namespaces using the proc filesystem inode number as
374 * the unique identifier. The user and pid ns are nested and will be dumped
377 * Called with task lock held.
380 void lttng_statedump_process_ns(struct lttng_session
*session
,
381 struct task_struct
*p
,
382 enum lttng_thread_type type
,
383 enum lttng_execution_mode mode
,
384 enum lttng_execution_submode submode
,
385 enum lttng_process_status status
)
387 struct nsproxy
*proxy
;
388 struct pid_namespace
*pid_ns
;
389 struct user_namespace
*user_ns
;
392 * The pid and user namespaces are special, they are nested and
393 * accessed with specific functions instead of the nsproxy struct
394 * like the other namespaces.
396 pid_ns
= task_active_pid_ns(p
);
398 trace_lttng_statedump_process_pid_ns(session
, p
, pid_ns
);
399 pid_ns
= pid_ns
? pid_ns
->parent
: NULL
;
403 user_ns
= task_cred_xxx(p
, user_ns
);
405 trace_lttng_statedump_process_user_ns(session
, p
, user_ns
);
407 * trace_lttng_statedump_process_user_ns() internally
408 * checks whether user_ns is NULL. While this does not
409 * appear to be a possible return value for
410 * task_cred_xxx(), err on the safe side and check
411 * for NULL here as well to be consistent with the
412 * paranoid behavior of
413 * trace_lttng_statedump_process_user_ns().
415 user_ns
= user_ns
? user_ns
->lttng_user_ns_parent
: NULL
;
419 * Back and forth on locking strategy within Linux upstream for nsproxy.
420 * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
421 * "namespaces: Use task_lock and not rcu to protect nsproxy"
424 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
425 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
426 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
427 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
431 proxy
= task_nsproxy(p
);
434 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
435 trace_lttng_statedump_process_cgroup_ns(session
, p
, proxy
->cgroup_ns
);
437 trace_lttng_statedump_process_ipc_ns(session
, p
, proxy
->ipc_ns
);
438 #ifndef LTTNG_MNT_NS_MISSING_HEADER
439 trace_lttng_statedump_process_mnt_ns(session
, p
, proxy
->mnt_ns
);
441 trace_lttng_statedump_process_net_ns(session
, p
, proxy
->net_ns
);
442 trace_lttng_statedump_process_uts_ns(session
, p
, proxy
->uts_ns
);
444 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
445 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
446 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
447 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
455 int lttng_enumerate_process_states(struct lttng_session
*session
)
457 struct task_struct
*g
, *p
;
460 tmp
= (char *) __get_free_page(GFP_KERNEL
);
465 for_each_process(g
) {
466 struct files_struct
*prev_files
= NULL
;
470 enum lttng_execution_mode mode
=
472 enum lttng_execution_submode submode
=
474 enum lttng_process_status status
;
475 enum lttng_thread_type type
;
476 struct files_struct
*files
;
479 if (p
->exit_state
== EXIT_ZOMBIE
)
480 status
= LTTNG_ZOMBIE
;
481 else if (p
->exit_state
== EXIT_DEAD
)
483 else if (p
->state
== TASK_RUNNING
) {
484 /* Is this a forked child that has not run yet? */
485 if (list_empty(&p
->rt
.run_list
))
486 status
= LTTNG_WAIT_FORK
;
489 * All tasks are considered as wait_cpu;
490 * the viewer will sort out if the task
491 * was really running at this time.
493 status
= LTTNG_WAIT_CPU
;
494 } else if (p
->state
&
495 (TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
)) {
496 /* Task is waiting for something to complete */
499 status
= LTTNG_UNNAMED
;
500 submode
= LTTNG_NONE
;
503 * Verification of t->mm is to filter out kernel
504 * threads; Viewer will further filter out if a
505 * user-space thread was in syscall mode or not.
508 type
= LTTNG_USER_THREAD
;
510 type
= LTTNG_KERNEL_THREAD
;
513 trace_lttng_statedump_process_state(session
,
514 p
, type
, mode
, submode
, status
, files
);
515 lttng_statedump_process_ns(session
,
516 p
, type
, mode
, submode
, status
);
518 * As an optimisation for the common case, do not
519 * repeat information for the same files_struct in
520 * two consecutive threads. This is the common case
521 * for threads sharing the same fd table. RCU guarantees
522 * that the same files_struct pointer is not re-used
523 * throughout processes/threads iteration.
525 if (files
&& files
!= prev_files
) {
526 lttng_enumerate_files(session
, files
, tmp
);
530 } while_each_thread(g
, p
);
534 free_page((unsigned long) tmp
);
540 void lttng_statedump_work_func(struct work_struct
*work
)
542 if (atomic_dec_and_test(&kernel_threads_to_run
))
543 /* If we are the last thread, wake up do_lttng_statedump */
544 wake_up(&statedump_wq
);
548 int do_lttng_statedump(struct lttng_session
*session
)
552 trace_lttng_statedump_start(session
);
553 ret
= lttng_enumerate_process_states(session
);
558 * ret = lttng_enumerate_vm_maps(session);
562 ret
= lttng_list_interrupts(session
);
565 ret
= lttng_enumerate_network_ip_interface(session
);
568 ret
= lttng_enumerate_block_devices(session
);
573 printk(KERN_WARNING
"LTTng: block device enumeration is not supported by kernel\n");
578 ret
= lttng_enumerate_cpu_topology(session
);
582 /* TODO lttng_dump_idt_table(session); */
583 /* TODO lttng_dump_softirq_vec(session); */
584 /* TODO lttng_list_modules(session); */
585 /* TODO lttng_dump_swap_files(session); */
588 * Fire off a work queue on each CPU. Their sole purpose in life
589 * is to guarantee that each CPU has been in a state where is was in
590 * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
593 atomic_set(&kernel_threads_to_run
, num_online_cpus());
594 for_each_online_cpu(cpu
) {
595 INIT_DELAYED_WORK(&cpu_work
[cpu
], lttng_statedump_work_func
);
596 schedule_delayed_work_on(cpu
, &cpu_work
[cpu
], 0);
598 /* Wait for all threads to run */
599 __wait_event(statedump_wq
, (atomic_read(&kernel_threads_to_run
) == 0));
601 /* Our work is done */
602 trace_lttng_statedump_end(session
);
607 * Called with session mutex held.
609 int lttng_statedump_start(struct lttng_session
*session
)
611 return do_lttng_statedump(session
);
613 EXPORT_SYMBOL_GPL(lttng_statedump_start
);
616 int __init
lttng_statedump_init(void)
619 * Allow module to load even if the fixup cannot be done. This
620 * will allow seemless transition when the underlying issue fix
621 * is merged into the Linux kernel, and when tracepoint.c
622 * "tracepoint_module_notify" is turned into a static function.
624 (void) wrapper_lttng_fixup_sig(THIS_MODULE
);
628 module_init(lttng_statedump_init
);
631 void __exit
lttng_statedump_exit(void)
635 module_exit(lttng_statedump_exit
);
637 MODULE_LICENSE("GPL and additional rights");
638 MODULE_AUTHOR("Jean-Hugues Deschenes");
639 MODULE_DESCRIPTION("LTTng statedump provider");
640 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION
) "."
641 __stringify(LTTNG_MODULES_MINOR_VERSION
) "."
642 __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION
)
643 LTTNG_MODULES_EXTRAVERSION
);