1 /* SPDX-License-Identifier: (GPL-2.0 or LGPL-2.1)
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>
36 #include <lttng-events.h>
37 #include <lttng-tracer.h>
38 #include <wrapper/irqdesc.h>
39 #include <wrapper/fdtable.h>
40 #include <wrapper/namespace.h>
41 #include <wrapper/irq.h>
42 #include <wrapper/tracepoint.h>
43 #include <wrapper/genhd.h>
44 #include <wrapper/file.h>
45 #include <wrapper/fdtable.h>
47 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
48 #include <linux/irq.h>
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>
58 DEFINE_TRACE(lttng_statedump_block_device
);
59 DEFINE_TRACE(lttng_statedump_end
);
60 DEFINE_TRACE(lttng_statedump_interrupt
);
61 DEFINE_TRACE(lttng_statedump_file_descriptor
);
62 DEFINE_TRACE(lttng_statedump_start
);
63 DEFINE_TRACE(lttng_statedump_process_state
);
64 DEFINE_TRACE(lttng_statedump_process_pid_ns
);
65 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
66 DEFINE_TRACE(lttng_statedump_process_cgroup_ns
);
68 DEFINE_TRACE(lttng_statedump_process_ipc_ns
);
69 #ifndef LTTNG_MNT_NS_MISSING_HEADER
70 DEFINE_TRACE(lttng_statedump_process_mnt_ns
);
72 DEFINE_TRACE(lttng_statedump_process_net_ns
);
73 DEFINE_TRACE(lttng_statedump_process_user_ns
);
74 DEFINE_TRACE(lttng_statedump_process_uts_ns
);
75 DEFINE_TRACE(lttng_statedump_network_interface
);
76 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
77 DEFINE_TRACE(lttng_statedump_cpu_topology
);
82 struct lttng_session
*session
;
83 struct files_struct
*files
;
87 * Protected by the trace lock.
89 static struct delayed_work cpu_work
[NR_CPUS
];
90 static DECLARE_WAIT_QUEUE_HEAD(statedump_wq
);
91 static atomic_t kernel_threads_to_run
;
93 enum lttng_thread_type
{
94 LTTNG_USER_THREAD
= 0,
95 LTTNG_KERNEL_THREAD
= 1,
98 enum lttng_execution_mode
{
104 LTTNG_MODE_UNKNOWN
= 5,
107 enum lttng_execution_submode
{
112 enum lttng_process_status
{
124 int lttng_enumerate_block_devices(struct lttng_session
*session
)
126 struct class *ptr_block_class
;
127 struct device_type
*ptr_disk_type
;
128 struct class_dev_iter iter
;
131 ptr_block_class
= wrapper_get_block_class();
132 if (!ptr_block_class
)
134 ptr_disk_type
= wrapper_get_disk_type();
135 if (!ptr_disk_type
) {
138 class_dev_iter_init(&iter
, ptr_block_class
, NULL
, ptr_disk_type
);
139 while ((dev
= class_dev_iter_next(&iter
))) {
140 struct disk_part_iter piter
;
141 struct gendisk
*disk
= dev_to_disk(dev
);
142 struct hd_struct
*part
;
145 * Don't show empty devices or things that have been
148 if (get_capacity(disk
) == 0 ||
149 (disk
->flags
& GENHD_FL_SUPPRESS_PARTITION_INFO
))
152 disk_part_iter_init(&piter
, disk
, DISK_PITER_INCL_PART0
);
153 while ((part
= disk_part_iter_next(&piter
))) {
154 char name_buf
[BDEVNAME_SIZE
];
157 p
= wrapper_disk_name(disk
, part
->partno
, name_buf
);
159 disk_part_iter_exit(&piter
);
160 class_dev_iter_exit(&iter
);
163 trace_lttng_statedump_block_device(session
,
164 part_devt(part
), name_buf
);
166 disk_part_iter_exit(&piter
);
168 class_dev_iter_exit(&iter
);
175 void lttng_enumerate_device(struct lttng_session
*session
,
176 struct net_device
*dev
)
178 struct in_device
*in_dev
;
179 struct in_ifaddr
*ifa
;
181 if (dev
->flags
& IFF_UP
) {
182 in_dev
= in_dev_get(dev
);
184 for (ifa
= in_dev
->ifa_list
; ifa
!= NULL
;
185 ifa
= ifa
->ifa_next
) {
186 trace_lttng_statedump_network_interface(
192 trace_lttng_statedump_network_interface(
198 int lttng_enumerate_network_ip_interface(struct lttng_session
*session
)
200 struct net_device
*dev
;
202 read_lock(&dev_base_lock
);
203 for_each_netdev(&init_net
, dev
)
204 lttng_enumerate_device(session
, dev
);
205 read_unlock(&dev_base_lock
);
209 #else /* CONFIG_INET */
211 int lttng_enumerate_network_ip_interface(struct lttng_session
*session
)
215 #endif /* CONFIG_INET */
218 int lttng_dump_one_fd(const void *p
, struct file
*file
, unsigned int fd
)
220 const struct lttng_fd_ctx
*ctx
= p
;
221 const char *s
= d_path(&file
->f_path
, ctx
->page
, PAGE_SIZE
);
222 unsigned int flags
= file
->f_flags
;
226 * We don't expose kernel internal flags, only userspace-visible
229 flags
&= ~FMODE_NONOTIFY
;
230 fdt
= files_fdtable(ctx
->files
);
232 * We need to check here again whether fd is within the fdt
233 * max_fds range, because we might be seeing a different
234 * files_fdtable() than iterate_fd(), assuming only RCU is
235 * protecting the read. In reality, iterate_fd() holds
236 * file_lock, which should ensure the fdt does not change while
237 * the lock is taken, but we are not aware whether this is
238 * guaranteed or not, so play safe.
240 if (fd
< fdt
->max_fds
&& lttng_close_on_exec(fd
, fdt
))
243 struct dentry
*dentry
= file
->f_path
.dentry
;
245 /* Make sure we give at least some info */
246 spin_lock(&dentry
->d_lock
);
247 trace_lttng_statedump_file_descriptor(ctx
->session
,
248 ctx
->files
, fd
, dentry
->d_name
.name
, flags
,
250 spin_unlock(&dentry
->d_lock
);
253 trace_lttng_statedump_file_descriptor(ctx
->session
,
254 ctx
->files
, fd
, s
, flags
, file
->f_mode
);
259 /* Called with task lock held. */
261 void lttng_enumerate_files(struct lttng_session
*session
,
262 struct files_struct
*files
,
265 struct lttng_fd_ctx ctx
= { .page
= tmp
, .session
= session
, .files
= files
, };
267 lttng_iterate_fd(files
, 0, lttng_dump_one_fd
, &ctx
);
270 #ifdef LTTNG_HAVE_STATEDUMP_CPU_TOPOLOGY
272 int lttng_enumerate_cpu_topology(struct lttng_session
*session
)
275 const cpumask_t
*cpumask
= cpu_possible_mask
;
277 for (cpu
= cpumask_first(cpumask
); cpu
< nr_cpu_ids
;
278 cpu
= cpumask_next(cpu
, cpumask
)) {
279 trace_lttng_statedump_cpu_topology(session
, &cpu_data(cpu
));
286 int lttng_enumerate_cpu_topology(struct lttng_session
*session
)
294 * FIXME: we cannot take a mmap_sem while in a RCU read-side critical section
295 * (scheduling in atomic). Normally, the tasklist lock protects this kind of
296 * iteration, but it is not exported to modules.
299 void lttng_enumerate_task_vm_maps(struct lttng_session
*session
,
300 struct task_struct
*p
)
302 struct mm_struct
*mm
;
303 struct vm_area_struct
*map
;
306 /* get_task_mm does a task_lock... */
313 down_read(&mm
->mmap_sem
);
316 ino
= map
->vm_file
->lttng_f_dentry
->d_inode
->i_ino
;
319 trace_lttng_statedump_vm_map(session
, p
, map
, ino
);
322 up_read(&mm
->mmap_sem
);
328 int lttng_enumerate_vm_maps(struct lttng_session
*session
)
330 struct task_struct
*p
;
334 lttng_enumerate_task_vm_maps(session
, p
);
340 #ifdef CONFIG_LTTNG_HAS_LIST_IRQ
343 int lttng_list_interrupts(struct lttng_session
*session
)
346 unsigned long flags
= 0;
347 struct irq_desc
*desc
;
349 #define irq_to_desc wrapper_irq_to_desc
351 for_each_irq_desc(irq
, desc
) {
352 struct irqaction
*action
;
353 const char *irq_chip_name
=
354 irq_desc_get_chip(desc
)->name
? : "unnamed_irq_chip";
356 local_irq_save(flags
);
357 raw_spin_lock(&desc
->lock
);
358 for (action
= desc
->action
; action
; action
= action
->next
) {
359 trace_lttng_statedump_interrupt(session
,
360 irq
, irq_chip_name
, action
);
362 raw_spin_unlock(&desc
->lock
);
363 local_irq_restore(flags
);
370 int lttng_list_interrupts(struct lttng_session
*session
)
377 * Statedump the task's namespaces using the proc filesystem inode number as
378 * the unique identifier. The user and pid ns are nested and will be dumped
381 * Called with task lock held.
384 void lttng_statedump_process_ns(struct lttng_session
*session
,
385 struct task_struct
*p
,
386 enum lttng_thread_type type
,
387 enum lttng_execution_mode mode
,
388 enum lttng_execution_submode submode
,
389 enum lttng_process_status status
)
391 struct nsproxy
*proxy
;
392 struct pid_namespace
*pid_ns
;
393 struct user_namespace
*user_ns
;
396 * The pid and user namespaces are special, they are nested and
397 * accessed with specific functions instead of the nsproxy struct
398 * like the other namespaces.
400 pid_ns
= task_active_pid_ns(p
);
402 trace_lttng_statedump_process_pid_ns(session
, p
, pid_ns
);
403 pid_ns
= pid_ns
? pid_ns
->parent
: NULL
;
407 user_ns
= task_cred_xxx(p
, user_ns
);
409 trace_lttng_statedump_process_user_ns(session
, p
, user_ns
);
411 * trace_lttng_statedump_process_user_ns() internally
412 * checks whether user_ns is NULL. While this does not
413 * appear to be a possible return value for
414 * task_cred_xxx(), err on the safe side and check
415 * for NULL here as well to be consistent with the
416 * paranoid behavior of
417 * trace_lttng_statedump_process_user_ns().
419 user_ns
= user_ns
? user_ns
->lttng_user_ns_parent
: NULL
;
423 * Back and forth on locking strategy within Linux upstream for nsproxy.
424 * See Linux upstream commit 728dba3a39c66b3d8ac889ddbe38b5b1c264aec3
425 * "namespaces: Use task_lock and not rcu to protect nsproxy"
428 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
429 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
430 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
431 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
435 proxy
= task_nsproxy(p
);
438 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,6,0))
439 trace_lttng_statedump_process_cgroup_ns(session
, p
, proxy
->cgroup_ns
);
441 trace_lttng_statedump_process_ipc_ns(session
, p
, proxy
->ipc_ns
);
442 #ifndef LTTNG_MNT_NS_MISSING_HEADER
443 trace_lttng_statedump_process_mnt_ns(session
, p
, proxy
->mnt_ns
);
445 trace_lttng_statedump_process_net_ns(session
, p
, proxy
->net_ns
);
446 trace_lttng_statedump_process_uts_ns(session
, p
, proxy
->uts_ns
);
448 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) || \
449 LTTNG_UBUNTU_KERNEL_RANGE(3,13,11,36, 3,14,0,0) || \
450 LTTNG_UBUNTU_KERNEL_RANGE(3,16,1,11, 3,17,0,0) || \
451 LTTNG_RHEL_KERNEL_RANGE(3,10,0,229,13,0, 3,11,0,0,0,0))
459 int lttng_enumerate_process_states(struct lttng_session
*session
)
461 struct task_struct
*g
, *p
;
464 tmp
= (char *) __get_free_page(GFP_KERNEL
);
469 for_each_process(g
) {
470 struct files_struct
*prev_files
= NULL
;
474 enum lttng_execution_mode mode
=
476 enum lttng_execution_submode submode
=
478 enum lttng_process_status status
;
479 enum lttng_thread_type type
;
480 struct files_struct
*files
;
483 if (p
->exit_state
== EXIT_ZOMBIE
)
484 status
= LTTNG_ZOMBIE
;
485 else if (p
->exit_state
== EXIT_DEAD
)
487 else if (p
->state
== TASK_RUNNING
) {
488 /* Is this a forked child that has not run yet? */
489 if (list_empty(&p
->rt
.run_list
))
490 status
= LTTNG_WAIT_FORK
;
493 * All tasks are considered as wait_cpu;
494 * the viewer will sort out if the task
495 * was really running at this time.
497 status
= LTTNG_WAIT_CPU
;
498 } else if (p
->state
&
499 (TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
)) {
500 /* Task is waiting for something to complete */
503 status
= LTTNG_UNNAMED
;
504 submode
= LTTNG_NONE
;
507 * Verification of t->mm is to filter out kernel
508 * threads; Viewer will further filter out if a
509 * user-space thread was in syscall mode or not.
512 type
= LTTNG_USER_THREAD
;
514 type
= LTTNG_KERNEL_THREAD
;
517 trace_lttng_statedump_process_state(session
,
518 p
, type
, mode
, submode
, status
, files
);
519 lttng_statedump_process_ns(session
,
520 p
, type
, mode
, submode
, status
);
522 * As an optimisation for the common case, do not
523 * repeat information for the same files_struct in
524 * two consecutive threads. This is the common case
525 * for threads sharing the same fd table. RCU guarantees
526 * that the same files_struct pointer is not re-used
527 * throughout processes/threads iteration.
529 if (files
&& files
!= prev_files
) {
530 lttng_enumerate_files(session
, files
, tmp
);
534 } while_each_thread(g
, p
);
538 free_page((unsigned long) tmp
);
544 void lttng_statedump_work_func(struct work_struct
*work
)
546 if (atomic_dec_and_test(&kernel_threads_to_run
))
547 /* If we are the last thread, wake up do_lttng_statedump */
548 wake_up(&statedump_wq
);
552 int do_lttng_statedump(struct lttng_session
*session
)
556 trace_lttng_statedump_start(session
);
557 ret
= lttng_enumerate_process_states(session
);
562 * ret = lttng_enumerate_vm_maps(session);
566 ret
= lttng_list_interrupts(session
);
569 ret
= lttng_enumerate_network_ip_interface(session
);
572 ret
= lttng_enumerate_block_devices(session
);
577 printk(KERN_WARNING
"LTTng: block device enumeration is not supported by kernel\n");
582 ret
= lttng_enumerate_cpu_topology(session
);
586 /* TODO lttng_dump_idt_table(session); */
587 /* TODO lttng_dump_softirq_vec(session); */
588 /* TODO lttng_list_modules(session); */
589 /* TODO lttng_dump_swap_files(session); */
592 * Fire off a work queue on each CPU. Their sole purpose in life
593 * is to guarantee that each CPU has been in a state where is was in
594 * syscall mode (i.e. not in a trap, an IRQ or a soft IRQ).
597 atomic_set(&kernel_threads_to_run
, num_online_cpus());
598 for_each_online_cpu(cpu
) {
599 INIT_DELAYED_WORK(&cpu_work
[cpu
], lttng_statedump_work_func
);
600 schedule_delayed_work_on(cpu
, &cpu_work
[cpu
], 0);
602 /* Wait for all threads to run */
603 __wait_event(statedump_wq
, (atomic_read(&kernel_threads_to_run
) == 0));
605 /* Our work is done */
606 trace_lttng_statedump_end(session
);
611 * Called with session mutex held.
613 int lttng_statedump_start(struct lttng_session
*session
)
615 return do_lttng_statedump(session
);
617 EXPORT_SYMBOL_GPL(lttng_statedump_start
);
620 int __init
lttng_statedump_init(void)
623 * Allow module to load even if the fixup cannot be done. This
624 * will allow seemless transition when the underlying issue fix
625 * is merged into the Linux kernel, and when tracepoint.c
626 * "tracepoint_module_notify" is turned into a static function.
628 (void) wrapper_lttng_fixup_sig(THIS_MODULE
);
632 module_init(lttng_statedump_init
);
635 void __exit
lttng_statedump_exit(void)
639 module_exit(lttng_statedump_exit
);
641 MODULE_LICENSE("GPL and additional rights");
642 MODULE_AUTHOR("Jean-Hugues Deschenes");
643 MODULE_DESCRIPTION("LTTng statedump provider");
644 MODULE_VERSION(__stringify(LTTNG_MODULES_MAJOR_VERSION
) "."
645 __stringify(LTTNG_MODULES_MINOR_VERSION
) "."
646 __stringify(LTTNG_MODULES_PATCHLEVEL_VERSION
)
647 LTTNG_MODULES_EXTRAVERSION
);