1Using RCU's CPU Stall Detector 2 3This document first discusses what sorts of issues RCU's CPU stall 4detector can locate, and then discusses kernel parameters and Kconfig 5options that can be used to fine-tune the detector's operation. Finally, 6this document explains the stall detector's "splat" format. 7 8 9What Causes RCU CPU Stall Warnings? 10 11So your kernel printed an RCU CPU stall warning. The next question is 12"What caused it?" The following problems can result in RCU CPU stall 13warnings: 14 15o A CPU looping in an RCU read-side critical section. 16 17o A CPU looping with interrupts disabled. 18 19o A CPU looping with preemption disabled. 20 21o A CPU looping with bottom halves disabled. 22 23o For !CONFIG_PREEMPT kernels, a CPU looping anywhere in the kernel 24 without invoking schedule(). If the looping in the kernel is 25 really expected and desirable behavior, you might need to add 26 some calls to cond_resched(). 27 28o Booting Linux using a console connection that is too slow to 29 keep up with the boot-time console-message rate. For example, 30 a 115Kbaud serial console can be -way- too slow to keep up 31 with boot-time message rates, and will frequently result in 32 RCU CPU stall warning messages. Especially if you have added 33 debug printk()s. 34 35o Anything that prevents RCU's grace-period kthreads from running. 36 This can result in the "All QSes seen" console-log message. 37 This message will include information on when the kthread last 38 ran and how often it should be expected to run. It can also 39 result in the "rcu_.*kthread starved for" console-log message, 40 which will include additional debugging information. 41 42o A CPU-bound real-time task in a CONFIG_PREEMPT kernel, which might 43 happen to preempt a low-priority task in the middle of an RCU 44 read-side critical section. This is especially damaging if 45 that low-priority task is not permitted to run on any other CPU, 46 in which case the next RCU grace period can never complete, which 47 will eventually cause the system to run out of memory and hang. 48 While the system is in the process of running itself out of 49 memory, you might see stall-warning messages. 50 51o A CPU-bound real-time task in a CONFIG_PREEMPT_RT kernel that 52 is running at a higher priority than the RCU softirq threads. 53 This will prevent RCU callbacks from ever being invoked, 54 and in a CONFIG_PREEMPT_RCU kernel will further prevent 55 RCU grace periods from ever completing. Either way, the 56 system will eventually run out of memory and hang. In the 57 CONFIG_PREEMPT_RCU case, you might see stall-warning 58 messages. 59 60 You can use the rcutree.kthread_prio kernel boot parameter to 61 increase the scheduling priority of RCU's kthreads, which can 62 help avoid this problem. However, please note that doing this 63 can increase your system's context-switch rate and thus degrade 64 performance. 65 66o A periodic interrupt whose handler takes longer than the time 67 interval between successive pairs of interrupts. This can 68 prevent RCU's kthreads and softirq handlers from running. 69 Note that certain high-overhead debugging options, for example 70 the function_graph tracer, can result in interrupt handler taking 71 considerably longer than normal, which can in turn result in 72 RCU CPU stall warnings. 73 74o Testing a workload on a fast system, tuning the stall-warning 75 timeout down to just barely avoid RCU CPU stall warnings, and then 76 running the same workload with the same stall-warning timeout on a 77 slow system. Note that thermal throttling and on-demand governors 78 can cause a single system to be sometimes fast and sometimes slow! 79 80o A hardware or software issue shuts off the scheduler-clock 81 interrupt on a CPU that is not in dyntick-idle mode. This 82 problem really has happened, and seems to be most likely to 83 result in RCU CPU stall warnings for CONFIG_NO_HZ_COMMON=n kernels. 84 85o A bug in the RCU implementation. 86 87o A hardware failure. This is quite unlikely, but has occurred 88 at least once in real life. A CPU failed in a running system, 89 becoming unresponsive, but not causing an immediate crash. 90 This resulted in a series of RCU CPU stall warnings, eventually 91 leading the realization that the CPU had failed. 92 93The RCU, RCU-sched, and RCU-tasks implementations have CPU stall warning. 94Note that SRCU does -not- have CPU stall warnings. Please note that 95RCU only detects CPU stalls when there is a grace period in progress. 96No grace period, no CPU stall warnings. 97 98To diagnose the cause of the stall, inspect the stack traces. 99The offending function will usually be near the top of the stack. 100If you have a series of stall warnings from a single extended stall, 101comparing the stack traces can often help determine where the stall 102is occurring, which will usually be in the function nearest the top of 103that portion of the stack which remains the same from trace to trace. 104If you can reliably trigger the stall, ftrace can be quite helpful. 105 106RCU bugs can often be debugged with the help of CONFIG_RCU_TRACE 107and with RCU's event tracing. For information on RCU's event tracing, 108see include/trace/events/rcu.h. 109 110 111Fine-Tuning the RCU CPU Stall Detector 112 113The rcuupdate.rcu_cpu_stall_suppress module parameter disables RCU's 114CPU stall detector, which detects conditions that unduly delay RCU grace 115periods. This module parameter enables CPU stall detection by default, 116but may be overridden via boot-time parameter or at runtime via sysfs. 117The stall detector's idea of what constitutes "unduly delayed" is 118controlled by a set of kernel configuration variables and cpp macros: 119 120CONFIG_RCU_CPU_STALL_TIMEOUT 121 122 This kernel configuration parameter defines the period of time 123 that RCU will wait from the beginning of a grace period until it 124 issues an RCU CPU stall warning. This time period is normally 125 21 seconds. 126 127 This configuration parameter may be changed at runtime via the 128 /sys/module/rcupdate/parameters/rcu_cpu_stall_timeout, however 129 this parameter is checked only at the beginning of a cycle. 130 So if you are 10 seconds into a 40-second stall, setting this 131 sysfs parameter to (say) five will shorten the timeout for the 132 -next- stall, or the following warning for the current stall 133 (assuming the stall lasts long enough). It will not affect the 134 timing of the next warning for the current stall. 135 136 Stall-warning messages may be enabled and disabled completely via 137 /sys/module/rcupdate/parameters/rcu_cpu_stall_suppress. 138 139RCU_STALL_DELAY_DELTA 140 141 Although the lockdep facility is extremely useful, it does add 142 some overhead. Therefore, under CONFIG_PROVE_RCU, the 143 RCU_STALL_DELAY_DELTA macro allows five extra seconds before 144 giving an RCU CPU stall warning message. (This is a cpp 145 macro, not a kernel configuration parameter.) 146 147RCU_STALL_RAT_DELAY 148 149 The CPU stall detector tries to make the offending CPU print its 150 own warnings, as this often gives better-quality stack traces. 151 However, if the offending CPU does not detect its own stall in 152 the number of jiffies specified by RCU_STALL_RAT_DELAY, then 153 some other CPU will complain. This delay is normally set to 154 two jiffies. (This is a cpp macro, not a kernel configuration 155 parameter.) 156 157rcupdate.rcu_task_stall_timeout 158 159 This boot/sysfs parameter controls the RCU-tasks stall warning 160 interval. A value of zero or less suppresses RCU-tasks stall 161 warnings. A positive value sets the stall-warning interval 162 in seconds. An RCU-tasks stall warning starts with the line: 163 164 INFO: rcu_tasks detected stalls on tasks: 165 166 And continues with the output of sched_show_task() for each 167 task stalling the current RCU-tasks grace period. 168 169 170Interpreting RCU's CPU Stall-Detector "Splats" 171 172For non-RCU-tasks flavors of RCU, when a CPU detects that it is stalling, 173it will print a message similar to the following: 174 175 INFO: rcu_sched detected stalls on CPUs/tasks: 176 2-...: (3 GPs behind) idle=06c/0/0 softirq=1453/1455 fqs=0 177 16-...: (0 ticks this GP) idle=81c/0/0 softirq=764/764 fqs=0 178 (detected by 32, t=2603 jiffies, g=7075, q=625) 179 180This message indicates that CPU 32 detected that CPUs 2 and 16 were both 181causing stalls, and that the stall was affecting RCU-sched. This message 182will normally be followed by stack dumps for each CPU. Please note that 183PREEMPT_RCU builds can be stalled by tasks as well as by CPUs, and that 184the tasks will be indicated by PID, for example, "P3421". It is even 185possible for an rcu_state stall to be caused by both CPUs -and- tasks, 186in which case the offending CPUs and tasks will all be called out in the list. 187 188CPU 2's "(3 GPs behind)" indicates that this CPU has not interacted with 189the RCU core for the past three grace periods. In contrast, CPU 16's "(0 190ticks this GP)" indicates that this CPU has not taken any scheduling-clock 191interrupts during the current stalled grace period. 192 193The "idle=" portion of the message prints the dyntick-idle state. 194The hex number before the first "/" is the low-order 12 bits of the 195dynticks counter, which will have an even-numbered value if the CPU 196is in dyntick-idle mode and an odd-numbered value otherwise. The hex 197number between the two "/"s is the value of the nesting, which will be 198a small non-negative number if in the idle loop (as shown above) and a 199very large positive number otherwise. 200 201The "softirq=" portion of the message tracks the number of RCU softirq 202handlers that the stalled CPU has executed. The number before the "/" 203is the number that had executed since boot at the time that this CPU 204last noted the beginning of a grace period, which might be the current 205(stalled) grace period, or it might be some earlier grace period (for 206example, if the CPU might have been in dyntick-idle mode for an extended 207time period. The number after the "/" is the number that have executed 208since boot until the current time. If this latter number stays constant 209across repeated stall-warning messages, it is possible that RCU's softirq 210handlers are no longer able to execute on this CPU. This can happen if 211the stalled CPU is spinning with interrupts are disabled, or, in -rt 212kernels, if a high-priority process is starving RCU's softirq handler. 213 214The "fqs=" shows the number of force-quiescent-state idle/offline 215detection passes that the grace-period kthread has made across this 216CPU since the last time that this CPU noted the beginning of a grace 217period. 218 219The "detected by" line indicates which CPU detected the stall (in this 220case, CPU 32), how many jiffies have elapsed since the start of the grace 221period (in this case 2603), the grace-period sequence number (7075), and 222an estimate of the total number of RCU callbacks queued across all CPUs 223(625 in this case). 224 225In kernels with CONFIG_RCU_FAST_NO_HZ, more information is printed 226for each CPU: 227 228 0: (64628 ticks this GP) idle=dd5/3fffffffffffffff/0 softirq=82/543 last_accelerate: a345/d342 Nonlazy posted: ..D 229 230The "last_accelerate:" prints the low-order 16 bits (in hex) of the 231jiffies counter when this CPU last invoked rcu_try_advance_all_cbs() 232from rcu_needs_cpu() or last invoked rcu_accelerate_cbs() from 233rcu_prepare_for_idle(). The "Nonlazy posted:" indicates lazy-callback 234status, so that an "l" indicates that all callbacks were lazy at the start 235of the last idle period and an "L" indicates that there are currently 236no non-lazy callbacks (in both cases, "." is printed otherwise, as 237shown above) and "D" indicates that dyntick-idle processing is enabled 238("." is printed otherwise, for example, if disabled via the "nohz=" 239kernel boot parameter). 240 241If the grace period ends just as the stall warning starts printing, 242there will be a spurious stall-warning message, which will include 243the following: 244 245 INFO: Stall ended before state dump start 246 247This is rare, but does happen from time to time in real life. It is also 248possible for a zero-jiffy stall to be flagged in this case, depending 249on how the stall warning and the grace-period initialization happen to 250interact. Please note that it is not possible to entirely eliminate this 251sort of false positive without resorting to things like stop_machine(), 252which is overkill for this sort of problem. 253 254If all CPUs and tasks have passed through quiescent states, but the 255grace period has nevertheless failed to end, the stall-warning splat 256will include something like the following: 257 258 All QSes seen, last rcu_preempt kthread activity 23807 (4297905177-4297881370), jiffies_till_next_fqs=3, root ->qsmask 0x0 259 260The "23807" indicates that it has been more than 23 thousand jiffies 261since the grace-period kthread ran. The "jiffies_till_next_fqs" 262indicates how frequently that kthread should run, giving the number 263of jiffies between force-quiescent-state scans, in this case three, 264which is way less than 23807. Finally, the root rcu_node structure's 265->qsmask field is printed, which will normally be zero. 266 267If the relevant grace-period kthread has been unable to run prior to 268the stall warning, as was the case in the "All QSes seen" line above, 269the following additional line is printed: 270 271 kthread starved for 23807 jiffies! g7075 f0x0 RCU_GP_WAIT_FQS(3) ->state=0x1 ->cpu=5 272 273Starving the grace-period kthreads of CPU time can of course result 274in RCU CPU stall warnings even when all CPUs and tasks have passed 275through the required quiescent states. The "g" number shows the current 276grace-period sequence number, the "f" precedes the ->gp_flags command 277to the grace-period kthread, the "RCU_GP_WAIT_FQS" indicates that the 278kthread is waiting for a short timeout, the "state" precedes value of the 279task_struct ->state field, and the "cpu" indicates that the grace-period 280kthread last ran on CPU 5. 281 282 283Multiple Warnings From One Stall 284 285If a stall lasts long enough, multiple stall-warning messages will be 286printed for it. The second and subsequent messages are printed at 287longer intervals, so that the time between (say) the first and second 288message will be about three times the interval between the beginning 289of the stall and the first message. 290 291 292Stall Warnings for Expedited Grace Periods 293 294If an expedited grace period detects a stall, it will place a message 295like the following in dmesg: 296 297 INFO: rcu_sched detected expedited stalls on CPUs/tasks: { 7-... } 21119 jiffies s: 73 root: 0x2/. 298 299This indicates that CPU 7 has failed to respond to a reschedule IPI. 300The three periods (".") following the CPU number indicate that the CPU 301is online (otherwise the first period would instead have been "O"), 302that the CPU was online at the beginning of the expedited grace period 303(otherwise the second period would have instead been "o"), and that 304the CPU has been online at least once since boot (otherwise, the third 305period would instead have been "N"). The number before the "jiffies" 306indicates that the expedited grace period has been going on for 21,119 307jiffies. The number following the "s:" indicates that the expedited 308grace-period sequence counter is 73. The fact that this last value is 309odd indicates that an expedited grace period is in flight. The number 310following "root:" is a bitmask that indicates which children of the root 311rcu_node structure correspond to CPUs and/or tasks that are blocking the 312current expedited grace period. If the tree had more than one level, 313additional hex numbers would be printed for the states of the other 314rcu_node structures in the tree. 315 316As with normal grace periods, PREEMPT_RCU builds can be stalled by 317tasks as well as by CPUs, and that the tasks will be indicated by PID, 318for example, "P3421". 319 320It is entirely possible to see stall warnings from normal and from 321expedited grace periods at about the same time during the same run. 322