1/* SPDX-License-Identifier: GPL-2.0 */ 2#ifndef __LINUX_PREEMPT_H 3#define __LINUX_PREEMPT_H 4 5/* 6 * include/linux/preempt.h - macros for accessing and manipulating 7 * preempt_count (used for kernel preemption, interrupt count, etc.) 8 */ 9 10#include <linux/linkage.h> 11#include <linux/list.h> 12 13/* 14 * We put the hardirq and softirq counter into the preemption 15 * counter. The bitmask has the following meaning: 16 * 17 * - bits 0-7 are the preemption count (max preemption depth: 256) 18 * - bits 8-15 are the softirq count (max # of softirqs: 256) 19 * 20 * The hardirq count could in theory be the same as the number of 21 * interrupts in the system, but we run all interrupt handlers with 22 * interrupts disabled, so we cannot have nesting interrupts. Though 23 * there are a few palaeontologic drivers which reenable interrupts in 24 * the handler, so we need more than one bit here. 25 * 26 * PREEMPT_MASK: 0x000000ff 27 * SOFTIRQ_MASK: 0x0000ff00 28 * HARDIRQ_MASK: 0x000f0000 29 * NMI_MASK: 0x00f00000 30 * PREEMPT_NEED_RESCHED: 0x80000000 31 */ 32#define PREEMPT_BITS 8 33#define SOFTIRQ_BITS 8 34#define HARDIRQ_BITS 4 35#define NMI_BITS 4 36 37#define PREEMPT_SHIFT 0 38#define SOFTIRQ_SHIFT (PREEMPT_SHIFT + PREEMPT_BITS) 39#define HARDIRQ_SHIFT (SOFTIRQ_SHIFT + SOFTIRQ_BITS) 40#define NMI_SHIFT (HARDIRQ_SHIFT + HARDIRQ_BITS) 41 42#define __IRQ_MASK(x) ((1UL << (x))-1) 43 44#define PREEMPT_MASK (__IRQ_MASK(PREEMPT_BITS) << PREEMPT_SHIFT) 45#define SOFTIRQ_MASK (__IRQ_MASK(SOFTIRQ_BITS) << SOFTIRQ_SHIFT) 46#define HARDIRQ_MASK (__IRQ_MASK(HARDIRQ_BITS) << HARDIRQ_SHIFT) 47#define NMI_MASK (__IRQ_MASK(NMI_BITS) << NMI_SHIFT) 48 49#define PREEMPT_OFFSET (1UL << PREEMPT_SHIFT) 50#define SOFTIRQ_OFFSET (1UL << SOFTIRQ_SHIFT) 51#define HARDIRQ_OFFSET (1UL << HARDIRQ_SHIFT) 52#define NMI_OFFSET (1UL << NMI_SHIFT) 53 54#define SOFTIRQ_DISABLE_OFFSET (2 * SOFTIRQ_OFFSET) 55 56#define PREEMPT_DISABLED (PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED) 57 58/* 59 * Disable preemption until the scheduler is running -- use an unconditional 60 * value so that it also works on !PREEMPT_COUNT kernels. 61 * 62 * Reset by start_kernel()->sched_init()->init_idle()->init_idle_preempt_count(). 63 */ 64#define INIT_PREEMPT_COUNT PREEMPT_OFFSET 65 66/* 67 * Initial preempt_count value; reflects the preempt_count schedule invariant 68 * which states that during context switches: 69 * 70 * preempt_count() == 2*PREEMPT_DISABLE_OFFSET 71 * 72 * Note: PREEMPT_DISABLE_OFFSET is 0 for !PREEMPT_COUNT kernels. 73 * Note: See finish_task_switch(). 74 */ 75#define FORK_PREEMPT_COUNT (2*PREEMPT_DISABLE_OFFSET + PREEMPT_ENABLED) 76 77/* preempt_count() and related functions, depends on PREEMPT_NEED_RESCHED */ 78#include <asm/preempt.h> 79 80#define nmi_count() (preempt_count() & NMI_MASK) 81#define hardirq_count() (preempt_count() & HARDIRQ_MASK) 82#define softirq_count() (preempt_count() & SOFTIRQ_MASK) 83#define irq_count() (nmi_count() | hardirq_count() | softirq_count()) 84 85/* 86 * Macros to retrieve the current execution context: 87 * 88 * in_nmi() - We're in NMI context 89 * in_hardirq() - We're in hard IRQ context 90 * in_serving_softirq() - We're in softirq context 91 * in_task() - We're in task context 92 */ 93#define in_nmi() (nmi_count()) 94#define in_hardirq() (hardirq_count()) 95#define in_serving_softirq() (softirq_count() & SOFTIRQ_OFFSET) 96#define in_task() (!(in_nmi() | in_hardirq() | in_serving_softirq())) 97 98/* 99 * The following macros are deprecated and should not be used in new code: 100 * in_irq() - Obsolete version of in_hardirq() 101 * in_softirq() - We have BH disabled, or are processing softirqs 102 * in_interrupt() - We're in NMI,IRQ,SoftIRQ context or have BH disabled 103 */ 104#define in_irq() (hardirq_count()) 105#define in_softirq() (softirq_count()) 106#define in_interrupt() (irq_count()) 107 108/* 109 * The preempt_count offset after preempt_disable(); 110 */ 111#if defined(CONFIG_PREEMPT_COUNT) 112# define PREEMPT_DISABLE_OFFSET PREEMPT_OFFSET 113#else 114# define PREEMPT_DISABLE_OFFSET 0 115#endif 116 117/* 118 * The preempt_count offset after spin_lock() 119 */ 120#define PREEMPT_LOCK_OFFSET PREEMPT_DISABLE_OFFSET 121 122/* 123 * The preempt_count offset needed for things like: 124 * 125 * spin_lock_bh() 126 * 127 * Which need to disable both preemption (CONFIG_PREEMPT_COUNT) and 128 * softirqs, such that unlock sequences of: 129 * 130 * spin_unlock(); 131 * local_bh_enable(); 132 * 133 * Work as expected. 134 */ 135#define SOFTIRQ_LOCK_OFFSET (SOFTIRQ_DISABLE_OFFSET + PREEMPT_LOCK_OFFSET) 136 137/* 138 * Are we running in atomic context? WARNING: this macro cannot 139 * always detect atomic context; in particular, it cannot know about 140 * held spinlocks in non-preemptible kernels. Thus it should not be 141 * used in the general case to determine whether sleeping is possible. 142 * Do not use in_atomic() in driver code. 143 */ 144#define in_atomic() (preempt_count() != 0) 145 146/* 147 * Check whether we were atomic before we did preempt_disable(): 148 * (used by the scheduler) 149 */ 150#define in_atomic_preempt_off() (preempt_count() != PREEMPT_DISABLE_OFFSET) 151 152#if defined(CONFIG_DEBUG_PREEMPT) || defined(CONFIG_TRACE_PREEMPT_TOGGLE) 153extern void preempt_count_add(int val); 154extern void preempt_count_sub(int val); 155#define preempt_count_dec_and_test() \ 156 ({ preempt_count_sub(1); should_resched(0); }) 157#else 158#define preempt_count_add(val) __preempt_count_add(val) 159#define preempt_count_sub(val) __preempt_count_sub(val) 160#define preempt_count_dec_and_test() __preempt_count_dec_and_test() 161#endif 162 163#define __preempt_count_inc() __preempt_count_add(1) 164#define __preempt_count_dec() __preempt_count_sub(1) 165 166#define preempt_count_inc() preempt_count_add(1) 167#define preempt_count_dec() preempt_count_sub(1) 168 169#ifdef CONFIG_PREEMPT_COUNT 170 171#define preempt_disable() \ 172do { \ 173 preempt_count_inc(); \ 174 barrier(); \ 175} while (0) 176 177#define sched_preempt_enable_no_resched() \ 178do { \ 179 barrier(); \ 180 preempt_count_dec(); \ 181} while (0) 182 183#define preempt_enable_no_resched() sched_preempt_enable_no_resched() 184 185#define preemptible() (preempt_count() == 0 && !irqs_disabled()) 186 187#ifdef CONFIG_PREEMPTION 188#define preempt_enable() \ 189do { \ 190 barrier(); \ 191 if (unlikely(preempt_count_dec_and_test())) \ 192 __preempt_schedule(); \ 193} while (0) 194 195#define preempt_enable_notrace() \ 196do { \ 197 barrier(); \ 198 if (unlikely(__preempt_count_dec_and_test())) \ 199 __preempt_schedule_notrace(); \ 200} while (0) 201 202#define preempt_check_resched() \ 203do { \ 204 if (should_resched(0)) \ 205 __preempt_schedule(); \ 206} while (0) 207 208#else /* !CONFIG_PREEMPTION */ 209#define preempt_enable() \ 210do { \ 211 barrier(); \ 212 preempt_count_dec(); \ 213} while (0) 214 215#define preempt_enable_notrace() \ 216do { \ 217 barrier(); \ 218 __preempt_count_dec(); \ 219} while (0) 220 221#define preempt_check_resched() do { } while (0) 222#endif /* CONFIG_PREEMPTION */ 223 224#define preempt_disable_notrace() \ 225do { \ 226 __preempt_count_inc(); \ 227 barrier(); \ 228} while (0) 229 230#define preempt_enable_no_resched_notrace() \ 231do { \ 232 barrier(); \ 233 __preempt_count_dec(); \ 234} while (0) 235 236#else /* !CONFIG_PREEMPT_COUNT */ 237 238/* 239 * Even if we don't have any preemption, we need preempt disable/enable 240 * to be barriers, so that we don't have things like get_user/put_user 241 * that can cause faults and scheduling migrate into our preempt-protected 242 * region. 243 */ 244#define preempt_disable() barrier() 245#define sched_preempt_enable_no_resched() barrier() 246#define preempt_enable_no_resched() barrier() 247#define preempt_enable() barrier() 248#define preempt_check_resched() do { } while (0) 249 250#define preempt_disable_notrace() barrier() 251#define preempt_enable_no_resched_notrace() barrier() 252#define preempt_enable_notrace() barrier() 253#define preemptible() 0 254 255#endif /* CONFIG_PREEMPT_COUNT */ 256 257#ifdef MODULE 258/* 259 * Modules have no business playing preemption tricks. 260 */ 261#undef sched_preempt_enable_no_resched 262#undef preempt_enable_no_resched 263#undef preempt_enable_no_resched_notrace 264#undef preempt_check_resched 265#endif 266 267#define preempt_set_need_resched() \ 268do { \ 269 set_preempt_need_resched(); \ 270} while (0) 271#define preempt_fold_need_resched() \ 272do { \ 273 if (tif_need_resched()) \ 274 set_preempt_need_resched(); \ 275} while (0) 276 277#ifdef CONFIG_PREEMPT_NOTIFIERS 278 279struct preempt_notifier; 280 281/** 282 * preempt_ops - notifiers called when a task is preempted and rescheduled 283 * @sched_in: we're about to be rescheduled: 284 * notifier: struct preempt_notifier for the task being scheduled 285 * cpu: cpu we're scheduled on 286 * @sched_out: we've just been preempted 287 * notifier: struct preempt_notifier for the task being preempted 288 * next: the task that's kicking us out 289 * 290 * Please note that sched_in and out are called under different 291 * contexts. sched_out is called with rq lock held and irq disabled 292 * while sched_in is called without rq lock and irq enabled. This 293 * difference is intentional and depended upon by its users. 294 */ 295struct preempt_ops { 296 void (*sched_in)(struct preempt_notifier *notifier, int cpu); 297 void (*sched_out)(struct preempt_notifier *notifier, 298 struct task_struct *next); 299}; 300 301/** 302 * preempt_notifier - key for installing preemption notifiers 303 * @link: internal use 304 * @ops: defines the notifier functions to be called 305 * 306 * Usually used in conjunction with container_of(). 307 */ 308struct preempt_notifier { 309 struct hlist_node link; 310 struct preempt_ops *ops; 311}; 312 313void preempt_notifier_inc(void); 314void preempt_notifier_dec(void); 315void preempt_notifier_register(struct preempt_notifier *notifier); 316void preempt_notifier_unregister(struct preempt_notifier *notifier); 317 318static inline void preempt_notifier_init(struct preempt_notifier *notifier, 319 struct preempt_ops *ops) 320{ 321 INIT_HLIST_NODE(¬ifier->link); 322 notifier->ops = ops; 323} 324 325#endif 326 327#ifdef CONFIG_SMP 328 329/* 330 * Migrate-Disable and why it is undesired. 331 * 332 * When a preempted task becomes elegible to run under the ideal model (IOW it 333 * becomes one of the M highest priority tasks), it might still have to wait 334 * for the preemptee's migrate_disable() section to complete. Thereby suffering 335 * a reduction in bandwidth in the exact duration of the migrate_disable() 336 * section. 337 * 338 * Per this argument, the change from preempt_disable() to migrate_disable() 339 * gets us: 340 * 341 * - a higher priority tasks gains reduced wake-up latency; with preempt_disable() 342 * it would have had to wait for the lower priority task. 343 * 344 * - a lower priority tasks; which under preempt_disable() could've instantly 345 * migrated away when another CPU becomes available, is now constrained 346 * by the ability to push the higher priority task away, which might itself be 347 * in a migrate_disable() section, reducing it's available bandwidth. 348 * 349 * IOW it trades latency / moves the interference term, but it stays in the 350 * system, and as long as it remains unbounded, the system is not fully 351 * deterministic. 352 * 353 * 354 * The reason we have it anyway. 355 * 356 * PREEMPT_RT breaks a number of assumptions traditionally held. By forcing a 357 * number of primitives into becoming preemptible, they would also allow 358 * migration. This turns out to break a bunch of per-cpu usage. To this end, 359 * all these primitives employ migirate_disable() to restore this implicit 360 * assumption. 361 * 362 * This is a 'temporary' work-around at best. The correct solution is getting 363 * rid of the above assumptions and reworking the code to employ explicit 364 * per-cpu locking or short preempt-disable regions. 365 * 366 * The end goal must be to get rid of migrate_disable(), alternatively we need 367 * a schedulability theory that does not depend on abritrary migration. 368 * 369 * 370 * Notes on the implementation. 371 * 372 * The implementation is particularly tricky since existing code patterns 373 * dictate neither migrate_disable() nor migrate_enable() is allowed to block. 374 * This means that it cannot use cpus_read_lock() to serialize against hotplug, 375 * nor can it easily migrate itself into a pending affinity mask change on 376 * migrate_enable(). 377 * 378 * 379 * Note: even non-work-conserving schedulers like semi-partitioned depends on 380 * migration, so migrate_disable() is not only a problem for 381 * work-conserving schedulers. 382 * 383 */ 384extern void migrate_disable(void); 385extern void migrate_enable(void); 386 387#else 388 389static inline void migrate_disable(void) { } 390static inline void migrate_enable(void) { } 391 392#endif /* CONFIG_SMP */ 393 394#endif /* __LINUX_PREEMPT_H */ 395