1.. SPDX-License-Identifier: GPL-2.0 2 3============================== 4Using RCU's CPU Stall Detector 5============================== 6 7This document first discusses what sorts of issues RCU's CPU stall 8detector can locate, and then discusses kernel parameters and Kconfig 9options that can be used to fine-tune the detector's operation. Finally, 10this document explains the stall detector's "splat" format. 11 12 13What Causes RCU CPU Stall Warnings? 14=================================== 15 16So your kernel printed an RCU CPU stall warning. The next question is 17"What caused it?" The following problems can result in RCU CPU stall 18warnings: 19 20- A CPU looping in an RCU read-side critical section. 21 22- A CPU looping with interrupts disabled. 23 24- A CPU looping with preemption disabled. 25 26- A CPU looping with bottom halves disabled. 27 28- For !CONFIG_PREEMPTION kernels, a CPU looping anywhere in the 29 kernel without potentially invoking schedule(). If the looping 30 in the kernel is really expected and desirable behavior, you 31 might need to add some calls to cond_resched(). 32 33- Booting Linux using a console connection that is too slow to 34 keep up with the boot-time console-message rate. For example, 35 a 115Kbaud serial console can be *way* too slow to keep up 36 with boot-time message rates, and will frequently result in 37 RCU CPU stall warning messages. Especially if you have added 38 debug printk()s. 39 40- Anything that prevents RCU's grace-period kthreads from running. 41 This can result in the "All QSes seen" console-log message. 42 This message will include information on when the kthread last 43 ran and how often it should be expected to run. It can also 44 result in the ``rcu_.*kthread starved for`` console-log message, 45 which will include additional debugging information. 46 47- A CPU-bound real-time task in a CONFIG_PREEMPTION kernel, which might 48 happen to preempt a low-priority task in the middle of an RCU 49 read-side critical section. This is especially damaging if 50 that low-priority task is not permitted to run on any other CPU, 51 in which case the next RCU grace period can never complete, which 52 will eventually cause the system to run out of memory and hang. 53 While the system is in the process of running itself out of 54 memory, you might see stall-warning messages. 55 56- A CPU-bound real-time task in a CONFIG_PREEMPT_RT kernel that 57 is running at a higher priority than the RCU softirq threads. 58 This will prevent RCU callbacks from ever being invoked, 59 and in a CONFIG_PREEMPT_RCU kernel will further prevent 60 RCU grace periods from ever completing. Either way, the 61 system will eventually run out of memory and hang. In the 62 CONFIG_PREEMPT_RCU case, you might see stall-warning 63 messages. 64 65 You can use the rcutree.kthread_prio kernel boot parameter to 66 increase the scheduling priority of RCU's kthreads, which can 67 help avoid this problem. However, please note that doing this 68 can increase your system's context-switch rate and thus degrade 69 performance. 70 71- A periodic interrupt whose handler takes longer than the time 72 interval between successive pairs of interrupts. This can 73 prevent RCU's kthreads and softirq handlers from running. 74 Note that certain high-overhead debugging options, for example 75 the function_graph tracer, can result in interrupt handler taking 76 considerably longer than normal, which can in turn result in 77 RCU CPU stall warnings. 78 79- Testing a workload on a fast system, tuning the stall-warning 80 timeout down to just barely avoid RCU CPU stall warnings, and then 81 running the same workload with the same stall-warning timeout on a 82 slow system. Note that thermal throttling and on-demand governors 83 can cause a single system to be sometimes fast and sometimes slow! 84 85- A hardware or software issue shuts off the scheduler-clock 86 interrupt on a CPU that is not in dyntick-idle mode. This 87 problem really has happened, and seems to be most likely to 88 result in RCU CPU stall warnings for CONFIG_NO_HZ_COMMON=n kernels. 89 90- A hardware or software issue that prevents time-based wakeups 91 from occurring. These issues can range from misconfigured or 92 buggy timer hardware through bugs in the interrupt or exception 93 path (whether hardware, firmware, or software) through bugs 94 in Linux's timer subsystem through bugs in the scheduler, and, 95 yes, even including bugs in RCU itself. It can also result in 96 the ``rcu_.*timer wakeup didn't happen for`` console-log message, 97 which will include additional debugging information. 98 99- A low-level kernel issue that either fails to invoke one of the 100 variants of rcu_eqs_enter(true), rcu_eqs_exit(true), ct_idle_enter(), 101 ct_idle_exit(), ct_irq_enter(), or ct_irq_exit() on the one 102 hand, or that invokes one of them too many times on the other. 103 Historically, the most frequent issue has been an omission 104 of either irq_enter() or irq_exit(), which in turn invoke 105 ct_irq_enter() or ct_irq_exit(), respectively. Building your 106 kernel with CONFIG_RCU_EQS_DEBUG=y can help track down these types 107 of issues, which sometimes arise in architecture-specific code. 108 109- A bug in the RCU implementation. 110 111- A hardware failure. This is quite unlikely, but is not at all 112 uncommon in large datacenter. In one memorable case some decades 113 back, a CPU failed in a running system, becoming unresponsive, 114 but not causing an immediate crash. This resulted in a series 115 of RCU CPU stall warnings, eventually leading the realization 116 that the CPU had failed. 117 118The RCU, RCU-sched, RCU-tasks, and RCU-tasks-trace implementations have 119CPU stall warning. Note that SRCU does *not* have CPU stall warnings. 120Please note that RCU only detects CPU stalls when there is a grace period 121in progress. No grace period, no CPU stall warnings. 122 123To diagnose the cause of the stall, inspect the stack traces. 124The offending function will usually be near the top of the stack. 125If you have a series of stall warnings from a single extended stall, 126comparing the stack traces can often help determine where the stall 127is occurring, which will usually be in the function nearest the top of 128that portion of the stack which remains the same from trace to trace. 129If you can reliably trigger the stall, ftrace can be quite helpful. 130 131RCU bugs can often be debugged with the help of CONFIG_RCU_TRACE 132and with RCU's event tracing. For information on RCU's event tracing, 133see include/trace/events/rcu.h. 134 135 136Fine-Tuning the RCU CPU Stall Detector 137====================================== 138 139The rcuupdate.rcu_cpu_stall_suppress module parameter disables RCU's 140CPU stall detector, which detects conditions that unduly delay RCU grace 141periods. This module parameter enables CPU stall detection by default, 142but may be overridden via boot-time parameter or at runtime via sysfs. 143The stall detector's idea of what constitutes "unduly delayed" is 144controlled by a set of kernel configuration variables and cpp macros: 145 146CONFIG_RCU_CPU_STALL_TIMEOUT 147---------------------------- 148 149 This kernel configuration parameter defines the period of time 150 that RCU will wait from the beginning of a grace period until it 151 issues an RCU CPU stall warning. This time period is normally 152 21 seconds. 153 154 This configuration parameter may be changed at runtime via the 155 /sys/module/rcupdate/parameters/rcu_cpu_stall_timeout, however 156 this parameter is checked only at the beginning of a cycle. 157 So if you are 10 seconds into a 40-second stall, setting this 158 sysfs parameter to (say) five will shorten the timeout for the 159 *next* stall, or the following warning for the current stall 160 (assuming the stall lasts long enough). It will not affect the 161 timing of the next warning for the current stall. 162 163 Stall-warning messages may be enabled and disabled completely via 164 /sys/module/rcupdate/parameters/rcu_cpu_stall_suppress. 165 166CONFIG_RCU_EXP_CPU_STALL_TIMEOUT 167-------------------------------- 168 169 Same as the CONFIG_RCU_CPU_STALL_TIMEOUT parameter but only for 170 the expedited grace period. This parameter defines the period 171 of time that RCU will wait from the beginning of an expedited 172 grace period until it issues an RCU CPU stall warning. This time 173 period is normally 20 milliseconds on Android devices. A zero 174 value causes the CONFIG_RCU_CPU_STALL_TIMEOUT value to be used, 175 after conversion to milliseconds. 176 177 This configuration parameter may be changed at runtime via the 178 /sys/module/rcupdate/parameters/rcu_exp_cpu_stall_timeout, however 179 this parameter is checked only at the beginning of a cycle. If you 180 are in a current stall cycle, setting it to a new value will change 181 the timeout for the -next- stall. 182 183 Stall-warning messages may be enabled and disabled completely via 184 /sys/module/rcupdate/parameters/rcu_cpu_stall_suppress. 185 186RCU_STALL_DELAY_DELTA 187--------------------- 188 189 Although the lockdep facility is extremely useful, it does add 190 some overhead. Therefore, under CONFIG_PROVE_RCU, the 191 RCU_STALL_DELAY_DELTA macro allows five extra seconds before 192 giving an RCU CPU stall warning message. (This is a cpp 193 macro, not a kernel configuration parameter.) 194 195RCU_STALL_RAT_DELAY 196------------------- 197 198 The CPU stall detector tries to make the offending CPU print its 199 own warnings, as this often gives better-quality stack traces. 200 However, if the offending CPU does not detect its own stall in 201 the number of jiffies specified by RCU_STALL_RAT_DELAY, then 202 some other CPU will complain. This delay is normally set to 203 two jiffies. (This is a cpp macro, not a kernel configuration 204 parameter.) 205 206rcupdate.rcu_task_stall_timeout 207------------------------------- 208 209 This boot/sysfs parameter controls the RCU-tasks and 210 RCU-tasks-trace stall warning intervals. A value of zero or less 211 suppresses RCU-tasks stall warnings. A positive value sets the 212 stall-warning interval in seconds. An RCU-tasks stall warning 213 starts with the line: 214 215 INFO: rcu_tasks detected stalls on tasks: 216 217 And continues with the output of sched_show_task() for each 218 task stalling the current RCU-tasks grace period. 219 220 An RCU-tasks-trace stall warning starts (and continues) similarly: 221 222 INFO: rcu_tasks_trace detected stalls on tasks 223 224 225Interpreting RCU's CPU Stall-Detector "Splats" 226============================================== 227 228For non-RCU-tasks flavors of RCU, when a CPU detects that some other 229CPU is stalling, it will print a message similar to the following:: 230 231 INFO: rcu_sched detected stalls on CPUs/tasks: 232 2-...: (3 GPs behind) idle=06c/0/0 softirq=1453/1455 fqs=0 233 16-...: (0 ticks this GP) idle=81c/0/0 softirq=764/764 fqs=0 234 (detected by 32, t=2603 jiffies, g=7075, q=625) 235 236This message indicates that CPU 32 detected that CPUs 2 and 16 were both 237causing stalls, and that the stall was affecting RCU-sched. This message 238will normally be followed by stack dumps for each CPU. Please note that 239PREEMPT_RCU builds can be stalled by tasks as well as by CPUs, and that 240the tasks will be indicated by PID, for example, "P3421". It is even 241possible for an rcu_state stall to be caused by both CPUs *and* tasks, 242in which case the offending CPUs and tasks will all be called out in the list. 243In some cases, CPUs will detect themselves stalling, which will result 244in a self-detected stall. 245 246CPU 2's "(3 GPs behind)" indicates that this CPU has not interacted with 247the RCU core for the past three grace periods. In contrast, CPU 16's "(0 248ticks this GP)" indicates that this CPU has not taken any scheduling-clock 249interrupts during the current stalled grace period. 250 251The "idle=" portion of the message prints the dyntick-idle state. 252The hex number before the first "/" is the low-order 12 bits of the 253dynticks counter, which will have an even-numbered value if the CPU 254is in dyntick-idle mode and an odd-numbered value otherwise. The hex 255number between the two "/"s is the value of the nesting, which will be 256a small non-negative number if in the idle loop (as shown above) and a 257very large positive number otherwise. The number following the final 258"/" is the NMI nesting, which will be a small non-negative number. 259 260The "softirq=" portion of the message tracks the number of RCU softirq 261handlers that the stalled CPU has executed. The number before the "/" 262is the number that had executed since boot at the time that this CPU 263last noted the beginning of a grace period, which might be the current 264(stalled) grace period, or it might be some earlier grace period (for 265example, if the CPU might have been in dyntick-idle mode for an extended 266time period). The number after the "/" is the number that have executed 267since boot until the current time. If this latter number stays constant 268across repeated stall-warning messages, it is possible that RCU's softirq 269handlers are no longer able to execute on this CPU. This can happen if 270the stalled CPU is spinning with interrupts are disabled, or, in -rt 271kernels, if a high-priority process is starving RCU's softirq handler. 272 273The "fqs=" shows the number of force-quiescent-state idle/offline 274detection passes that the grace-period kthread has made across this 275CPU since the last time that this CPU noted the beginning of a grace 276period. 277 278The "detected by" line indicates which CPU detected the stall (in this 279case, CPU 32), how many jiffies have elapsed since the start of the grace 280period (in this case 2603), the grace-period sequence number (7075), and 281an estimate of the total number of RCU callbacks queued across all CPUs 282(625 in this case). 283 284If the grace period ends just as the stall warning starts printing, 285there will be a spurious stall-warning message, which will include 286the following:: 287 288 INFO: Stall ended before state dump start 289 290This is rare, but does happen from time to time in real life. It is also 291possible for a zero-jiffy stall to be flagged in this case, depending 292on how the stall warning and the grace-period initialization happen to 293interact. Please note that it is not possible to entirely eliminate this 294sort of false positive without resorting to things like stop_machine(), 295which is overkill for this sort of problem. 296 297If all CPUs and tasks have passed through quiescent states, but the 298grace period has nevertheless failed to end, the stall-warning splat 299will include something like the following:: 300 301 All QSes seen, last rcu_preempt kthread activity 23807 (4297905177-4297881370), jiffies_till_next_fqs=3, root ->qsmask 0x0 302 303The "23807" indicates that it has been more than 23 thousand jiffies 304since the grace-period kthread ran. The "jiffies_till_next_fqs" 305indicates how frequently that kthread should run, giving the number 306of jiffies between force-quiescent-state scans, in this case three, 307which is way less than 23807. Finally, the root rcu_node structure's 308->qsmask field is printed, which will normally be zero. 309 310If the relevant grace-period kthread has been unable to run prior to 311the stall warning, as was the case in the "All QSes seen" line above, 312the following additional line is printed:: 313 314 rcu_sched kthread starved for 23807 jiffies! g7075 f0x0 RCU_GP_WAIT_FQS(3) ->state=0x1 ->cpu=5 315 Unless rcu_sched kthread gets sufficient CPU time, OOM is now expected behavior. 316 317Starving the grace-period kthreads of CPU time can of course result 318in RCU CPU stall warnings even when all CPUs and tasks have passed 319through the required quiescent states. The "g" number shows the current 320grace-period sequence number, the "f" precedes the ->gp_flags command 321to the grace-period kthread, the "RCU_GP_WAIT_FQS" indicates that the 322kthread is waiting for a short timeout, the "state" precedes value of the 323task_struct ->state field, and the "cpu" indicates that the grace-period 324kthread last ran on CPU 5. 325 326If the relevant grace-period kthread does not wake from FQS wait in a 327reasonable time, then the following additional line is printed:: 328 329 kthread timer wakeup didn't happen for 23804 jiffies! g7076 f0x0 RCU_GP_WAIT_FQS(5) ->state=0x402 330 331The "23804" indicates that kthread's timer expired more than 23 thousand 332jiffies ago. The rest of the line has meaning similar to the kthread 333starvation case. 334 335Additionally, the following line is printed:: 336 337 Possible timer handling issue on cpu=4 timer-softirq=11142 338 339Here "cpu" indicates that the grace-period kthread last ran on CPU 4, 340where it queued the fqs timer. The number following the "timer-softirq" 341is the current ``TIMER_SOFTIRQ`` count on cpu 4. If this value does not 342change on successive RCU CPU stall warnings, there is further reason to 343suspect a timer problem. 344 345These messages are usually followed by stack dumps of the CPUs and tasks 346involved in the stall. These stack traces can help you locate the cause 347of the stall, keeping in mind that the CPU detecting the stall will have 348an interrupt frame that is mainly devoted to detecting the stall. 349 350 351Multiple Warnings From One Stall 352================================ 353 354If a stall lasts long enough, multiple stall-warning messages will 355be printed for it. The second and subsequent messages are printed at 356longer intervals, so that the time between (say) the first and second 357message will be about three times the interval between the beginning 358of the stall and the first message. It can be helpful to compare the 359stack dumps for the different messages for the same stalled grace period. 360 361 362Stall Warnings for Expedited Grace Periods 363========================================== 364 365If an expedited grace period detects a stall, it will place a message 366like the following in dmesg:: 367 368 INFO: rcu_sched detected expedited stalls on CPUs/tasks: { 7-... } 21119 jiffies s: 73 root: 0x2/. 369 370This indicates that CPU 7 has failed to respond to a reschedule IPI. 371The three periods (".") following the CPU number indicate that the CPU 372is online (otherwise the first period would instead have been "O"), 373that the CPU was online at the beginning of the expedited grace period 374(otherwise the second period would have instead been "o"), and that 375the CPU has been online at least once since boot (otherwise, the third 376period would instead have been "N"). The number before the "jiffies" 377indicates that the expedited grace period has been going on for 21,119 378jiffies. The number following the "s:" indicates that the expedited 379grace-period sequence counter is 73. The fact that this last value is 380odd indicates that an expedited grace period is in flight. The number 381following "root:" is a bitmask that indicates which children of the root 382rcu_node structure correspond to CPUs and/or tasks that are blocking the 383current expedited grace period. If the tree had more than one level, 384additional hex numbers would be printed for the states of the other 385rcu_node structures in the tree. 386 387As with normal grace periods, PREEMPT_RCU builds can be stalled by 388tasks as well as by CPUs, and that the tasks will be indicated by PID, 389for example, "P3421". 390 391It is entirely possible to see stall warnings from normal and from 392expedited grace periods at about the same time during the same run. 393 394RCU_CPU_STALL_CPUTIME 395===================== 396 397In kernels built with CONFIG_RCU_CPU_STALL_CPUTIME=y or booted with 398rcupdate.rcu_cpu_stall_cputime=1, the following additional information 399is supplied with each RCU CPU stall warning:: 400 401 rcu: hardirqs softirqs csw/system 402 rcu: number: 624 45 0 403 rcu: cputime: 69 1 2425 ==> 2500(ms) 404 405These statistics are collected during the sampling period. The values 406in row "number:" are the number of hard interrupts, number of soft 407interrupts, and number of context switches on the stalled CPU. The 408first three values in row "cputime:" indicate the CPU time in 409milliseconds consumed by hard interrupts, soft interrupts, and tasks 410on the stalled CPU. The last number is the measurement interval, again 411in milliseconds. Because user-mode tasks normally do not cause RCU CPU 412stalls, these tasks are typically kernel tasks, which is why only the 413system CPU time are considered. 414 415The sampling period is shown as follows:: 416 417 |<------------first timeout---------->|<-----second timeout----->| 418 |<--half timeout-->|<--half timeout-->| | 419 | |<--first period-->| | 420 | |<-----------second sampling period---------->| 421 | | | | 422 snapshot time point 1st-stall 2nd-stall 423 424The following describes four typical scenarios: 425 4261. A CPU looping with interrupts disabled. 427 428 :: 429 430 rcu: hardirqs softirqs csw/system 431 rcu: number: 0 0 0 432 rcu: cputime: 0 0 0 ==> 2500(ms) 433 434 Because interrupts have been disabled throughout the measurement 435 interval, there are no interrupts and no context switches. 436 Furthermore, because CPU time consumption was measured using interrupt 437 handlers, the system CPU consumption is misleadingly measured as zero. 438 This scenario will normally also have "(0 ticks this GP)" printed on 439 this CPU's summary line. 440 4412. A CPU looping with bottom halves disabled. 442 443 This is similar to the previous example, but with non-zero number of 444 and CPU time consumed by hard interrupts, along with non-zero CPU 445 time consumed by in-kernel execution:: 446 447 rcu: hardirqs softirqs csw/system 448 rcu: number: 624 0 0 449 rcu: cputime: 49 0 2446 ==> 2500(ms) 450 451 The fact that there are zero softirqs gives a hint that these were 452 disabled, perhaps via local_bh_disable(). It is of course possible 453 that there were no softirqs, perhaps because all events that would 454 result in softirq execution are confined to other CPUs. In this case, 455 the diagnosis should continue as shown in the next example. 456 4573. A CPU looping with preemption disabled. 458 459 Here, only the number of context switches is zero:: 460 461 rcu: hardirqs softirqs csw/system 462 rcu: number: 624 45 0 463 rcu: cputime: 69 1 2425 ==> 2500(ms) 464 465 This situation hints that the stalled CPU was looping with preemption 466 disabled. 467 4684. No looping, but massive hard and soft interrupts. 469 470 :: 471 472 rcu: hardirqs softirqs csw/system 473 rcu: number: xx xx 0 474 rcu: cputime: xx xx 0 ==> 2500(ms) 475 476 Here, the number and CPU time of hard interrupts are all non-zero, 477 but the number of context switches and the in-kernel CPU time consumed 478 are zero. The number and cputime of soft interrupts will usually be 479 non-zero, but could be zero, for example, if the CPU was spinning 480 within a single hard interrupt handler. 481 482 If this type of RCU CPU stall warning can be reproduced, you can 483 narrow it down by looking at /proc/interrupts or by writing code to 484 trace each interrupt, for example, by referring to show_interrupts(). 485