1# SPDX-License-Identifier: GPL-2.0-only 2# 3# RCU-related configuration options 4# 5 6menu "RCU Subsystem" 7 8config TREE_RCU 9 bool 10 default y if SMP 11 # Dynticks-idle tracking 12 select CONTEXT_TRACKING_IDLE 13 help 14 This option selects the RCU implementation that is 15 designed for very large SMP system with hundreds or 16 thousands of CPUs. It also scales down nicely to 17 smaller systems. 18 19config PREEMPT_RCU 20 bool 21 default y if (PREEMPT || PREEMPT_RT || PREEMPT_DYNAMIC) 22 select TREE_RCU 23 help 24 This option selects the RCU implementation that is 25 designed for very large SMP systems with hundreds or 26 thousands of CPUs, but for which real-time response 27 is also required. It also scales down nicely to 28 smaller systems. 29 30 Select this option if you are unsure. 31 32config TINY_RCU 33 bool 34 default y if !PREEMPT_RCU && !SMP 35 help 36 This option selects the RCU implementation that is 37 designed for UP systems from which real-time response 38 is not required. This option greatly reduces the 39 memory footprint of RCU. 40 41config RCU_EXPERT 42 bool "Make expert-level adjustments to RCU configuration" 43 default n 44 help 45 This option needs to be enabled if you wish to make 46 expert-level adjustments to RCU configuration. By default, 47 no such adjustments can be made, which has the often-beneficial 48 side-effect of preventing "make oldconfig" from asking you all 49 sorts of detailed questions about how you would like numerous 50 obscure RCU options to be set up. 51 52 Say Y if you need to make expert-level adjustments to RCU. 53 54 Say N if you are unsure. 55 56config TINY_SRCU 57 bool 58 default y if TINY_RCU 59 help 60 This option selects the single-CPU non-preemptible version of SRCU. 61 62config TREE_SRCU 63 bool 64 default y if !TINY_RCU 65 help 66 This option selects the full-fledged version of SRCU. 67 68config FORCE_NEED_SRCU_NMI_SAFE 69 bool "Force selection of NEED_SRCU_NMI_SAFE" 70 depends on !TINY_SRCU 71 depends on RCU_EXPERT 72 depends on ARCH_HAS_NMI_SAFE_THIS_CPU_OPS 73 select NEED_SRCU_NMI_SAFE 74 default n 75 help 76 This option forces selection of the NEED_SRCU_NMI_SAFE 77 Kconfig option, allowing testing of srcu_read_lock_nmisafe() 78 and srcu_read_unlock_nmisafe() on architectures (like x86) 79 that select the ARCH_HAS_NMI_SAFE_THIS_CPU_OPS Kconfig option. 80 81config NEED_SRCU_NMI_SAFE 82 def_bool HAVE_NMI && !ARCH_HAS_NMI_SAFE_THIS_CPU_OPS && !TINY_SRCU 83 84config TASKS_RCU_GENERIC 85 def_bool TASKS_RCU || TASKS_RUDE_RCU 86 help 87 This option enables generic infrastructure code supporting 88 task-based RCU implementations. Not for manual selection. 89 90config FORCE_TASKS_RCU 91 bool "Force selection of TASKS_RCU" 92 depends on RCU_EXPERT 93 select TASKS_RCU 94 default n 95 help 96 This option force-enables a task-based RCU implementation 97 that uses only voluntary context switch (not preemption!), 98 idle, and user-mode execution as quiescent states. Not for 99 manual selection in most cases. 100 101config NEED_TASKS_RCU 102 bool 103 default n 104 105config TASKS_RCU 106 bool 107 default NEED_TASKS_RCU && PREEMPTION 108 select IRQ_WORK 109 110config FORCE_TASKS_RUDE_RCU 111 bool "Force selection of Tasks Rude RCU" 112 depends on RCU_EXPERT 113 select TASKS_RUDE_RCU 114 default n 115 help 116 This option force-enables a task-based RCU implementation 117 that uses only context switch (including preemption) and 118 user-mode execution as quiescent states. It forces IPIs and 119 context switches on all online CPUs, including idle ones, 120 so use with caution. Not for manual selection in most cases. 121 122config TASKS_RUDE_RCU 123 bool 124 default n 125 select IRQ_WORK 126 127config FORCE_TASKS_TRACE_RCU 128 bool "Force selection of Tasks Trace RCU" 129 depends on RCU_EXPERT 130 select TASKS_TRACE_RCU 131 default n 132 help 133 This option enables a task-based RCU implementation that uses 134 explicit rcu_read_lock_trace() read-side markers, and allows 135 these readers to appear in the idle loop as well as on the 136 CPU hotplug code paths. It can force IPIs on online CPUs, 137 including idle ones, so use with caution. Not for manual 138 selection in most cases. 139 140config TASKS_TRACE_RCU 141 bool 142 default n 143 144config TASKS_TRACE_RCU_NO_MB 145 bool "Override RCU Tasks Trace inclusion of read-side memory barriers" 146 depends on RCU_EXPERT && TASKS_TRACE_RCU 147 default ARCH_WANTS_NO_INSTR 148 help 149 This option prevents the use of read-side memory barriers in 150 rcu_read_lock_tasks_trace() and rcu_read_unlock_tasks_trace() 151 even in kernels built with CONFIG_ARCH_WANTS_NO_INSTR=n, that is, 152 in kernels that do not have noinstr set up in entry/exit code. 153 By setting this option, you are promising to carefully review 154 use of ftrace, BPF, and friends to ensure that no tracing 155 operation is attached to a function that runs in that portion 156 of the entry/exit code that RCU does not watch, that is, 157 where rcu_is_watching() returns false. Alternatively, you 158 might choose to never remove traces except by rebooting. 159 160 Those wishing to disable read-side memory barriers for an entire 161 architecture can select this Kconfig option, hence the polarity. 162 163 Say Y here if you need speed and will review use of tracing. 164 Say N here for certain esoteric testing of RCU itself. 165 Take the default if you are unsure. 166 167config RCU_STALL_COMMON 168 def_bool TREE_RCU 169 help 170 This option enables RCU CPU stall code that is common between 171 the TINY and TREE variants of RCU. The purpose is to allow 172 the tiny variants to disable RCU CPU stall warnings, while 173 making these warnings mandatory for the tree variants. 174 175config RCU_NEED_SEGCBLIST 176 def_bool ( TREE_RCU || TREE_SRCU || TASKS_RCU_GENERIC ) 177 178# The deferral (and the IRQ_WORK it uses) is only needed where call_rcu() / 179# call_srcu() can be invoked while a callback-list operation is in flight. 180config RCU_DEFER 181 def_bool HAVE_NMI || KPROBES || FUNCTION_TRACER || TRACEPOINTS 182 select IRQ_WORK 183 184config RCU_FANOUT 185 int "Tree-based hierarchical RCU fanout value" 186 range 2 64 if 64BIT 187 range 2 32 if !64BIT 188 depends on TREE_RCU && RCU_EXPERT 189 default 64 if 64BIT 190 default 32 if !64BIT 191 help 192 This option controls the fanout of hierarchical implementations 193 of RCU, allowing RCU to work efficiently on machines with 194 large numbers of CPUs. This value must be at least the fourth 195 root of NR_CPUS, which allows NR_CPUS to be insanely large. 196 The default value of RCU_FANOUT should be used for production 197 systems, but if you are stress-testing the RCU implementation 198 itself, small RCU_FANOUT values allow you to test large-system 199 code paths on small(er) systems. 200 201 Select a specific number if testing RCU itself. 202 Take the default if unsure. 203 204config RCU_FANOUT_LEAF 205 int "Tree-based hierarchical RCU leaf-level fanout value" 206 range 2 64 if 64BIT && !RCU_STRICT_GRACE_PERIOD 207 range 2 32 if !64BIT && !RCU_STRICT_GRACE_PERIOD 208 range 2 3 if RCU_STRICT_GRACE_PERIOD 209 depends on TREE_RCU && RCU_EXPERT 210 default 16 if !RCU_STRICT_GRACE_PERIOD 211 default 2 if RCU_STRICT_GRACE_PERIOD 212 help 213 This option controls the leaf-level fanout of hierarchical 214 implementations of RCU, and allows trading off cache misses 215 against lock contention. Systems that synchronize their 216 scheduling-clock interrupts for energy-efficiency reasons will 217 want the default because the smaller leaf-level fanout keeps 218 lock contention levels acceptably low. Very large systems 219 (hundreds or thousands of CPUs) will instead want to set this 220 value to the maximum value possible in order to reduce the 221 number of cache misses incurred during RCU's grace-period 222 initialization. These systems tend to run CPU-bound, and thus 223 are not helped by synchronized interrupts, and thus tend to 224 skew them, which reduces lock contention enough that large 225 leaf-level fanouts work well. That said, setting leaf-level 226 fanout to a large number will likely cause problematic 227 lock contention on the leaf-level rcu_node structures unless 228 you boot with the skew_tick kernel parameter. 229 230 Select a specific number if testing RCU itself. 231 232 Select the maximum permissible value for large systems, but 233 please understand that you may also need to set the skew_tick 234 kernel boot parameter to avoid contention on the rcu_node 235 structure's locks. 236 237 Take the default if unsure. 238 239config RCU_BOOST 240 bool "Enable RCU priority boosting" 241 depends on (RT_MUTEXES && PREEMPT_RCU && RCU_EXPERT) || PREEMPT_RT 242 default y if PREEMPT_RT 243 help 244 This option boosts the priority of preempted RCU readers that 245 block the current preemptible RCU grace period for too long. 246 This option also prevents heavy loads from blocking RCU 247 callback invocation. 248 249 Say Y here if you are working with real-time apps or heavy loads 250 Say N here if you are unsure. 251 252config RCU_BOOST_DELAY 253 int "Milliseconds to delay boosting after RCU grace-period start" 254 range 0 3000 255 depends on RCU_BOOST 256 default 500 257 help 258 This option specifies the time to wait after the beginning of 259 a given grace period before priority-boosting preempted RCU 260 readers blocking that grace period. Note that any RCU reader 261 blocking an expedited RCU grace period is boosted immediately. 262 263 Accept the default if unsure. 264 265config RCU_EXP_KTHREAD 266 bool "Perform RCU expedited work in a real-time kthread" 267 depends on RCU_BOOST && RCU_EXPERT 268 default !PREEMPT_RT && NR_CPUS <= 32 269 help 270 Use this option to further reduce the latencies of expedited 271 grace periods at the expense of being more disruptive. 272 273 This option is disabled by default on PREEMPT_RT=y kernels which 274 disable expedited grace periods after boot by unconditionally 275 setting rcupdate.rcu_normal_after_boot=1. 276 277 Accept the default if unsure. 278 279config RCU_NOCB_CPU 280 bool "Offload RCU callback processing from boot-selected CPUs" 281 depends on TREE_RCU 282 depends on RCU_EXPERT || NO_HZ_FULL 283 default n 284 help 285 Use this option to reduce OS jitter for aggressive HPC or 286 real-time workloads. It can also be used to offload RCU 287 callback invocation to energy-efficient CPUs in battery-powered 288 asymmetric multiprocessors. The price of this reduced jitter 289 is that the overhead of call_rcu() increases and that some 290 workloads will incur significant increases in context-switch 291 rates. 292 293 This option offloads callback invocation from the set of 294 CPUs specified at boot time by the rcu_nocbs parameter. 295 For each such CPU, a kthread ("rcuox/N") will be created to 296 invoke callbacks, where the "N" is the CPU being offloaded, 297 and where the "x" is "p" for RCU-preempt (PREEMPTION kernels) 298 and "s" for RCU-sched (!PREEMPTION kernels). This option 299 also creates another kthread for each sqrt(nr_cpu_ids) CPUs 300 ("rcuog/N", where N is the first CPU in that group to come 301 online), which handles grace periods for its group. Nothing 302 prevents these kthreads from running on the specified CPUs, 303 but (1) the kthreads may be preempted between each callback, 304 and (2) affinity or cgroups can be used to force the kthreads 305 to run on whatever set of CPUs is desired. 306 307 The sqrt(nr_cpu_ids) grouping may be overridden using the 308 rcutree.rcu_nocb_gp_stride kernel boot parameter. This can 309 be especially helpful for smaller numbers of CPUs, where 310 sqrt(nr_cpu_ids) can be a bit of a blunt instrument. 311 312 Say Y here if you need reduced OS jitter, despite added overhead. 313 Say N here if you are unsure. 314 315config RCU_NOCB_CPU_DEFAULT_ALL 316 bool "Offload RCU callback processing from all CPUs by default" 317 depends on RCU_NOCB_CPU 318 default n 319 help 320 Use this option to offload callback processing from all CPUs 321 by default, in the absence of the rcu_nocbs or nohz_full boot 322 parameter. This also avoids the need to use any boot parameters 323 to achieve the effect of offloading all CPUs on boot. 324 325 Say Y here if you want offload all CPUs by default on boot. 326 Say N here if you are unsure. 327 328config RCU_NOCB_CPU_CB_BOOST 329 bool "Offload RCU callback from real-time kthread" 330 depends on RCU_NOCB_CPU && RCU_BOOST 331 default y if PREEMPT_RT 332 help 333 Use this option to invoke offloaded callbacks as SCHED_FIFO 334 to avoid starvation by heavy SCHED_OTHER background load. 335 Of course, running as SCHED_FIFO during callback floods will 336 cause the rcuo[ps] kthreads to monopolize the CPU for hundreds 337 of milliseconds or more. Therefore, when enabling this option, 338 it is your responsibility to ensure that latency-sensitive 339 tasks either run with higher priority or run on some other CPU. 340 341 Say Y here if you want to set RT priority for offloading kthreads. 342 Say N here if you are building a !PREEMPT_RT kernel and are unsure. 343 344config RCU_LAZY 345 bool "RCU callback lazy invocation functionality" 346 depends on RCU_NOCB_CPU 347 default n 348 help 349 To save power, batch RCU callbacks and delay starting the 350 corresponding grace period for multiple seconds. The grace 351 period will be started after this delay, in case of memory 352 pressure, or if the corresponding CPU's callback list grows 353 too large. 354 355 These delays happen only on rcu_nocbs CPUs, that is, CPUs 356 whose callbacks have been offloaded. 357 358 Use the rcutree.enable_rcu_lazy=0 kernel-boot parameter to 359 globally disable these delays. 360 361config RCU_LAZY_DEFAULT_OFF 362 bool "Turn RCU lazy invocation off by default" 363 depends on RCU_LAZY 364 default n 365 help 366 Build the kernel with CONFIG_RCU_LAZY=y, but cause the kernel 367 to boot with these energy-efficiency delays disabled. Use the 368 rcutree.enable_rcu_lazy=0 kernel-boot parameter to override 369 the this option at boot time, thus re-enabling these delays. 370 371config RCU_DOUBLE_CHECK_CB_TIME 372 bool "RCU callback-batch backup time check" 373 depends on RCU_EXPERT 374 default n 375 help 376 Use this option to provide more precise enforcement of the 377 rcutree.rcu_resched_ns module parameter in situations where 378 a single RCU callback might run for hundreds of microseconds, 379 thus defeating the 32-callback batching used to amortize the 380 cost of the fine-grained but expensive local_clock() function. 381 382 This option rounds rcutree.rcu_resched_ns up to the next 383 jiffy, and overrides the 32-callback batching if this limit 384 is exceeded. 385 386 Say Y here if you need tighter callback-limit enforcement. 387 Say N here if you are unsure. 388 389endmenu # "RCU Subsystem" 390