1 CPU hotplug Support in Linux(tm) Kernel 2 3 Maintainers: 4 CPU Hotplug Core: 5 Rusty Russell <rusty@rustcorp.com.au> 6 Srivatsa Vaddagiri <vatsa@in.ibm.com> 7 i386: 8 Zwane Mwaikambo <zwanem@gmail.com> 9 ppc64: 10 Nathan Lynch <nathanl@austin.ibm.com> 11 Joel Schopp <jschopp@austin.ibm.com> 12 ia64/x86_64: 13 Ashok Raj <ashok.raj@intel.com> 14 s390: 15 Heiko Carstens <heiko.carstens@de.ibm.com> 16 17Authors: Ashok Raj <ashok.raj@intel.com> 18Lots of feedback: Nathan Lynch <nathanl@austin.ibm.com>, 19 Joel Schopp <jschopp@austin.ibm.com> 20 21Introduction 22 23Modern advances in system architectures have introduced advanced error 24reporting and correction capabilities in processors. CPU architectures permit 25partitioning support, where compute resources of a single CPU could be made 26available to virtual machine environments. There are couple OEMS that 27support NUMA hardware which are hot pluggable as well, where physical 28node insertion and removal require support for CPU hotplug. 29 30Such advances require CPUs available to a kernel to be removed either for 31provisioning reasons, or for RAS purposes to keep an offending CPU off 32system execution path. Hence the need for CPU hotplug support in the 33Linux kernel. 34 35A more novel use of CPU-hotplug support is its use today in suspend 36resume support for SMP. Dual-core and HT support makes even 37a laptop run SMP kernels which didn't support these methods. SMP support 38for suspend/resume is a work in progress. 39 40General Stuff about CPU Hotplug 41-------------------------------- 42 43Command Line Switches 44--------------------- 45maxcpus=n Restrict boot time cpus to n. Say if you have 4 cpus, using 46 maxcpus=2 will only boot 2. You can choose to bring the 47 other cpus later online, read FAQ's for more info. 48 49additional_cpus=n (*) Use this to limit hotpluggable cpus. This option sets 50 cpu_possible_mask = cpu_present_mask + additional_cpus 51 52cede_offline={"off","on"} Use this option to disable/enable putting offlined 53 processors to an extended H_CEDE state on 54 supported pseries platforms. 55 If nothing is specified, 56 cede_offline is set to "on". 57 58(*) Option valid only for following architectures 59- ia64 60 61ia64 uses the number of disabled local apics in ACPI tables MADT to 62determine the number of potentially hot-pluggable cpus. The implementation 63should only rely on this to count the # of cpus, but *MUST* not rely 64on the apicid values in those tables for disabled apics. In the event 65BIOS doesn't mark such hot-pluggable cpus as disabled entries, one could 66use this parameter "additional_cpus=x" to represent those cpus in the 67cpu_possible_mask. 68 69possible_cpus=n [s390,x86_64] use this to set hotpluggable cpus. 70 This option sets possible_cpus bits in 71 cpu_possible_mask. Thus keeping the numbers of bits set 72 constant even if the machine gets rebooted. 73 74CPU maps and such 75----------------- 76[More on cpumaps and primitive to manipulate, please check 77include/linux/cpumask.h that has more descriptive text.] 78 79cpu_possible_mask: Bitmap of possible CPUs that can ever be available in the 80system. This is used to allocate some boot time memory for per_cpu variables 81that aren't designed to grow/shrink as CPUs are made available or removed. 82Once set during boot time discovery phase, the map is static, i.e no bits 83are added or removed anytime. Trimming it accurately for your system needs 84upfront can save some boot time memory. See below for how we use heuristics 85in x86_64 case to keep this under check. 86 87cpu_online_mask: Bitmap of all CPUs currently online. Its set in __cpu_up() 88after a cpu is available for kernel scheduling and ready to receive 89interrupts from devices. Its cleared when a cpu is brought down using 90__cpu_disable(), before which all OS services including interrupts are 91migrated to another target CPU. 92 93cpu_present_mask: Bitmap of CPUs currently present in the system. Not all 94of them may be online. When physical hotplug is processed by the relevant 95subsystem (e.g ACPI) can change and new bit either be added or removed 96from the map depending on the event is hot-add/hot-remove. There are currently 97no locking rules as of now. Typical usage is to init topology during boot, 98at which time hotplug is disabled. 99 100You really dont need to manipulate any of the system cpu maps. They should 101be read-only for most use. When setting up per-cpu resources almost always use 102cpu_possible_mask/for_each_possible_cpu() to iterate. 103 104Never use anything other than cpumask_t to represent bitmap of CPUs. 105 106 #include <linux/cpumask.h> 107 108 for_each_possible_cpu - Iterate over cpu_possible_mask 109 for_each_online_cpu - Iterate over cpu_online_mask 110 for_each_present_cpu - Iterate over cpu_present_mask 111 for_each_cpu(x,mask) - Iterate over some random collection of cpu mask. 112 113 #include <linux/cpu.h> 114 get_online_cpus() and put_online_cpus(): 115 116The above calls are used to inhibit cpu hotplug operations. While the 117cpu_hotplug.refcount is non zero, the cpu_online_mask will not change. 118If you merely need to avoid cpus going away, you could also use 119preempt_disable() and preempt_enable() for those sections. 120Just remember the critical section cannot call any 121function that can sleep or schedule this process away. The preempt_disable() 122will work as long as stop_machine_run() is used to take a cpu down. 123 124CPU Hotplug - Frequently Asked Questions. 125 126Q: How to enable my kernel to support CPU hotplug? 127A: When doing make defconfig, Enable CPU hotplug support 128 129 "Processor type and Features" -> Support for Hotpluggable CPUs 130 131Make sure that you have CONFIG_SMP turned on as well. 132 133You would need to enable CONFIG_HOTPLUG_CPU for SMP suspend/resume support 134as well. 135 136Q: What architectures support CPU hotplug? 137A: As of 2.6.14, the following architectures support CPU hotplug. 138 139i386 (Intel), ppc, ppc64, parisc, s390, ia64 and x86_64 140 141Q: How to test if hotplug is supported on the newly built kernel? 142A: You should now notice an entry in sysfs. 143 144Check if sysfs is mounted, using the "mount" command. You should notice 145an entry as shown below in the output. 146 147 .... 148 none on /sys type sysfs (rw) 149 .... 150 151If this is not mounted, do the following. 152 153 #mkdir /sysfs 154 #mount -t sysfs sys /sys 155 156Now you should see entries for all present cpu, the following is an example 157in a 8-way system. 158 159 #pwd 160 #/sys/devices/system/cpu 161 #ls -l 162 total 0 163 drwxr-xr-x 10 root root 0 Sep 19 07:44 . 164 drwxr-xr-x 13 root root 0 Sep 19 07:45 .. 165 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu0 166 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu1 167 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu2 168 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu3 169 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu4 170 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu5 171 drwxr-xr-x 3 root root 0 Sep 19 07:44 cpu6 172 drwxr-xr-x 3 root root 0 Sep 19 07:48 cpu7 173 174Under each directory you would find an "online" file which is the control 175file to logically online/offline a processor. 176 177Q: Does hot-add/hot-remove refer to physical add/remove of cpus? 178A: The usage of hot-add/remove may not be very consistently used in the code. 179CONFIG_HOTPLUG_CPU enables logical online/offline capability in the kernel. 180To support physical addition/removal, one would need some BIOS hooks and 181the platform should have something like an attention button in PCI hotplug. 182CONFIG_ACPI_HOTPLUG_CPU enables ACPI support for physical add/remove of CPUs. 183 184Q: How do i logically offline a CPU? 185A: Do the following. 186 187 #echo 0 > /sys/devices/system/cpu/cpuX/online 188 189Once the logical offline is successful, check 190 191 #cat /proc/interrupts 192 193You should now not see the CPU that you removed. Also online file will report 194the state as 0 when a cpu if offline and 1 when its online. 195 196 #To display the current cpu state. 197 #cat /sys/devices/system/cpu/cpuX/online 198 199Q: Why can't i remove CPU0 on some systems? 200A: Some architectures may have some special dependency on a certain CPU. 201 202For e.g in IA64 platforms we have ability to sent platform interrupts to the 203OS. a.k.a Corrected Platform Error Interrupts (CPEI). In current ACPI 204specifications, we didn't have a way to change the target CPU. Hence if the 205current ACPI version doesn't support such re-direction, we disable that CPU 206by making it not-removable. 207 208In such cases you will also notice that the online file is missing under cpu0. 209 210Q: Is CPU0 removable on X86? 211A: Yes. If kernel is compiled with CONFIG_BOOTPARAM_HOTPLUG_CPU0=y, CPU0 is 212removable by default. Otherwise, CPU0 is also removable by kernel option 213cpu0_hotplug. 214 215But some features depend on CPU0. Two known dependencies are: 216 2171. Resume from hibernate/suspend depends on CPU0. Hibernate/suspend will fail if 218CPU0 is offline and you need to online CPU0 before hibernate/suspend can 219continue. 2202. PIC interrupts also depend on CPU0. CPU0 can't be removed if a PIC interrupt 221is detected. 222 223It's said poweroff/reboot may depend on CPU0 on some machines although I haven't 224seen any poweroff/reboot failure so far after CPU0 is offline on a few tested 225machines. 226 227Please let me know if you know or see any other dependencies of CPU0. 228 229If the dependencies are under your control, you can turn on CPU0 hotplug feature 230either by CONFIG_BOOTPARAM_HOTPLUG_CPU0 or by kernel parameter cpu0_hotplug. 231 232--Fenghua Yu <fenghua.yu@intel.com> 233 234Q: How do i find out if a particular CPU is not removable? 235A: Depending on the implementation, some architectures may show this by the 236absence of the "online" file. This is done if it can be determined ahead of 237time that this CPU cannot be removed. 238 239In some situations, this can be a run time check, i.e if you try to remove the 240last CPU, this will not be permitted. You can find such failures by 241investigating the return value of the "echo" command. 242 243Q: What happens when a CPU is being logically offlined? 244A: The following happen, listed in no particular order :-) 245 246- A notification is sent to in-kernel registered modules by sending an event 247 CPU_DOWN_PREPARE or CPU_DOWN_PREPARE_FROZEN, depending on whether or not the 248 CPU is being offlined while tasks are frozen due to a suspend operation in 249 progress 250- All processes are migrated away from this outgoing CPU to new CPUs. 251 The new CPU is chosen from each process' current cpuset, which may be 252 a subset of all online CPUs. 253- All interrupts targeted to this CPU is migrated to a new CPU 254- timers/bottom half/task lets are also migrated to a new CPU 255- Once all services are migrated, kernel calls an arch specific routine 256 __cpu_disable() to perform arch specific cleanup. 257- Once this is successful, an event for successful cleanup is sent by an event 258 CPU_DEAD (or CPU_DEAD_FROZEN if tasks are frozen due to a suspend while the 259 CPU is being offlined). 260 261 "It is expected that each service cleans up when the CPU_DOWN_PREPARE 262 notifier is called, when CPU_DEAD is called its expected there is nothing 263 running on behalf of this CPU that was offlined" 264 265Q: If i have some kernel code that needs to be aware of CPU arrival and 266 departure, how to i arrange for proper notification? 267A: This is what you would need in your kernel code to receive notifications. 268 269 #include <linux/cpu.h> 270 static int foobar_cpu_callback(struct notifier_block *nfb, 271 unsigned long action, void *hcpu) 272 { 273 unsigned int cpu = (unsigned long)hcpu; 274 275 switch (action) { 276 case CPU_ONLINE: 277 case CPU_ONLINE_FROZEN: 278 foobar_online_action(cpu); 279 break; 280 case CPU_DEAD: 281 case CPU_DEAD_FROZEN: 282 foobar_dead_action(cpu); 283 break; 284 } 285 return NOTIFY_OK; 286 } 287 288 static struct notifier_block foobar_cpu_notifier = 289 { 290 .notifier_call = foobar_cpu_callback, 291 }; 292 293You need to call register_cpu_notifier() from your init function. 294Init functions could be of two types: 2951. early init (init function called when only the boot processor is online). 2962. late init (init function called _after_ all the CPUs are online). 297 298For the first case, you should add the following to your init function 299 300 register_cpu_notifier(&foobar_cpu_notifier); 301 302For the second case, you should add the following to your init function 303 304 register_hotcpu_notifier(&foobar_cpu_notifier); 305 306You can fail PREPARE notifiers if something doesn't work to prepare resources. 307This will stop the activity and send a following CANCELED event back. 308 309CPU_DEAD should not be failed, its just a goodness indication, but bad 310things will happen if a notifier in path sent a BAD notify code. 311 312Q: I don't see my action being called for all CPUs already up and running? 313A: Yes, CPU notifiers are called only when new CPUs are on-lined or offlined. 314 If you need to perform some action for each cpu already in the system, then 315 do this: 316 317 for_each_online_cpu(i) { 318 foobar_cpu_callback(&foobar_cpu_notifier, CPU_UP_PREPARE, i); 319 foobar_cpu_callback(&foobar_cpu_notifier, CPU_ONLINE, i); 320 } 321 322 However, if you want to register a hotplug callback, as well as perform 323 some initialization for CPUs that are already online, then do this: 324 325 Version 1: (Correct) 326 --------- 327 328 cpu_notifier_register_begin(); 329 330 for_each_online_cpu(i) { 331 foobar_cpu_callback(&foobar_cpu_notifier, 332 CPU_UP_PREPARE, i); 333 foobar_cpu_callback(&foobar_cpu_notifier, 334 CPU_ONLINE, i); 335 } 336 337 /* Note the use of the double underscored version of the API */ 338 __register_cpu_notifier(&foobar_cpu_notifier); 339 340 cpu_notifier_register_done(); 341 342 Note that the following code is *NOT* the right way to achieve this, 343 because it is prone to an ABBA deadlock between the cpu_add_remove_lock 344 and the cpu_hotplug.lock. 345 346 Version 2: (Wrong!) 347 --------- 348 349 get_online_cpus(); 350 351 for_each_online_cpu(i) { 352 foobar_cpu_callback(&foobar_cpu_notifier, 353 CPU_UP_PREPARE, i); 354 foobar_cpu_callback(&foobar_cpu_notifier, 355 CPU_ONLINE, i); 356 } 357 358 register_cpu_notifier(&foobar_cpu_notifier); 359 360 put_online_cpus(); 361 362 So always use the first version shown above when you want to register 363 callbacks as well as initialize the already online CPUs. 364 365 366Q: If i would like to develop cpu hotplug support for a new architecture, 367 what do i need at a minimum? 368A: The following are what is required for CPU hotplug infrastructure to work 369 correctly. 370 371 - Make sure you have an entry in Kconfig to enable CONFIG_HOTPLUG_CPU 372 - __cpu_up() - Arch interface to bring up a CPU 373 - __cpu_disable() - Arch interface to shutdown a CPU, no more interrupts 374 can be handled by the kernel after the routine 375 returns. Including local APIC timers etc are 376 shutdown. 377 - __cpu_die() - This actually supposed to ensure death of the CPU. 378 Actually look at some example code in other arch 379 that implement CPU hotplug. The processor is taken 380 down from the idle() loop for that specific 381 architecture. __cpu_die() typically waits for some 382 per_cpu state to be set, to ensure the processor 383 dead routine is called to be sure positively. 384 385Q: I need to ensure that a particular cpu is not removed when there is some 386 work specific to this cpu is in progress. 387A: There are two ways. If your code can be run in interrupt context, use 388 smp_call_function_single(), otherwise use work_on_cpu(). Note that 389 work_on_cpu() is slow, and can fail due to out of memory: 390 391 int my_func_on_cpu(int cpu) 392 { 393 int err; 394 get_online_cpus(); 395 if (!cpu_online(cpu)) 396 err = -EINVAL; 397 else 398#if NEEDS_BLOCKING 399 err = work_on_cpu(cpu, __my_func_on_cpu, NULL); 400#else 401 smp_call_function_single(cpu, __my_func_on_cpu, &err, 402 true); 403#endif 404 put_online_cpus(); 405 return err; 406 } 407 408Q: How do we determine how many CPUs are available for hotplug. 409A: There is no clear spec defined way from ACPI that can give us that 410 information today. Based on some input from Natalie of Unisys, 411 that the ACPI MADT (Multiple APIC Description Tables) marks those possible 412 CPUs in a system with disabled status. 413 414 Andi implemented some simple heuristics that count the number of disabled 415 CPUs in MADT as hotpluggable CPUS. In the case there are no disabled CPUS 416 we assume 1/2 the number of CPUs currently present can be hotplugged. 417 418 Caveat: ACPI MADT can only provide 256 entries in systems with only ACPI 2.0c 419 or earlier ACPI version supported, because the apicid field in MADT is only 420 8 bits. From ACPI 3.0, this limitation was removed since the apicid field 421 was extended to 32 bits with x2APIC introduced. 422 423User Space Notification 424 425Hotplug support for devices is common in Linux today. Its being used today to 426support automatic configuration of network, usb and pci devices. A hotplug 427event can be used to invoke an agent script to perform the configuration task. 428 429You can add /etc/hotplug/cpu.agent to handle hotplug notification user space 430scripts. 431 432 #!/bin/bash 433 # $Id: cpu.agent 434 # Kernel hotplug params include: 435 #ACTION=%s [online or offline] 436 #DEVPATH=%s 437 # 438 cd /etc/hotplug 439 . ./hotplug.functions 440 441 case $ACTION in 442 online) 443 echo `date` ":cpu.agent" add cpu >> /tmp/hotplug.txt 444 ;; 445 offline) 446 echo `date` ":cpu.agent" remove cpu >>/tmp/hotplug.txt 447 ;; 448 *) 449 debug_mesg CPU $ACTION event not supported 450 exit 1 451 ;; 452 esac 453