[75f3522] | 1 | // |
---|
| 2 | // Cforall Version 1.0.0 Copyright (C) 2016 University of Waterloo |
---|
| 3 | // |
---|
| 4 | // The contents of this file are covered under the licence agreement in the |
---|
| 5 | // file "LICENCE" distributed with Cforall. |
---|
| 6 | // |
---|
[73abe95] | 7 | // kernel_private.hfa -- |
---|
[75f3522] | 8 | // |
---|
| 9 | // Author : Thierry Delisle |
---|
| 10 | // Created On : Mon Feb 13 12:27:26 2017 |
---|
[6b0b624] | 11 | // Last Modified By : Peter A. Buhr |
---|
[fd9b524] | 12 | // Last Modified On : Wed Aug 12 08:21:33 2020 |
---|
| 13 | // Update Count : 9 |
---|
[75f3522] | 14 | // |
---|
| 15 | |
---|
[6b0b624] | 16 | #pragma once |
---|
[75f3522] | 17 | |
---|
[3489ea6] | 18 | #if !defined(__cforall_thread__) |
---|
| 19 | #error kernel_private.hfa should only be included in libcfathread source |
---|
| 20 | #endif |
---|
| 21 | |
---|
[58b6d1b] | 22 | #include "kernel.hfa" |
---|
| 23 | #include "thread.hfa" |
---|
[75f3522] | 24 | |
---|
[73abe95] | 25 | #include "alarm.hfa" |
---|
[8834751] | 26 | #include "stats.hfa" |
---|
[fa21ac9] | 27 | |
---|
[3489ea6] | 28 | extern "C" { |
---|
| 29 | #if defined(CFA_HAVE_LINUX_LIBRSEQ) |
---|
| 30 | #include <rseq/rseq.h> |
---|
| 31 | #elif defined(CFA_HAVE_LINUX_RSEQ_H) |
---|
[f558b5f] | 32 | #include <linux/rseq.h> |
---|
[3489ea6] | 33 | #else |
---|
| 34 | #ifndef _GNU_SOURCE |
---|
| 35 | #error kernel_private requires gnu_source |
---|
| 36 | #endif |
---|
| 37 | #include <sched.h> |
---|
| 38 | #endif |
---|
| 39 | } |
---|
| 40 | |
---|
[d3605f8] | 41 | // Defines whether or not we *want* to use io_uring_enter as the idle_sleep blocking call |
---|
| 42 | #define CFA_WANT_IO_URING_IDLE |
---|
| 43 | |
---|
| 44 | // Defines whether or not we *can* use io_uring_enter as the idle_sleep blocking call |
---|
| 45 | #if defined(CFA_WANT_IO_URING_IDLE) && defined(CFA_HAVE_LINUX_IO_URING_H) |
---|
| 46 | #if defined(CFA_HAVE_IORING_OP_READ) || (defined(CFA_HAVE_READV) && defined(CFA_HAVE_IORING_OP_READV)) |
---|
| 47 | #define CFA_WITH_IO_URING_IDLE |
---|
| 48 | #endif |
---|
| 49 | #endif |
---|
[059ad16] | 50 | |
---|
[75f3522] | 51 | //----------------------------------------------------------------------------- |
---|
| 52 | // Scheduler |
---|
[1c273d0] | 53 | extern "C" { |
---|
[2026bb6] | 54 | void disable_interrupts() OPTIONAL_THREAD; |
---|
[a3821fa] | 55 | void enable_interrupts( bool poll = true ); |
---|
[1c273d0] | 56 | } |
---|
| 57 | |
---|
[24e321c] | 58 | void schedule_thread$( thread$ *, unpark_hint hint ) __attribute__((nonnull (1))); |
---|
[75f3522] | 59 | |
---|
[8fc652e0] | 60 | extern bool __preemption_enabled(); |
---|
| 61 | |
---|
[7d0ebd0] | 62 | static inline void __disable_interrupts_checked() { |
---|
| 63 | /* paranoid */ verify( __preemption_enabled() ); |
---|
| 64 | disable_interrupts(); |
---|
| 65 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 66 | } |
---|
| 67 | |
---|
| 68 | static inline void __enable_interrupts_checked( bool poll = true ) { |
---|
| 69 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 70 | enable_interrupts( poll ); |
---|
| 71 | /* paranoid */ verify( __preemption_enabled() ); |
---|
| 72 | } |
---|
| 73 | |
---|
[5afb49a] | 74 | //release/wake-up the following resources |
---|
[e84ab3d] | 75 | void __thread_finish( thread$ * thrd ); |
---|
[db6f06a] | 76 | |
---|
[3489ea6] | 77 | //----------------------------------------------------------------------------- |
---|
| 78 | // Hardware |
---|
| 79 | |
---|
| 80 | #if defined(CFA_HAVE_LINUX_LIBRSEQ) |
---|
| 81 | // No data needed |
---|
| 82 | #elif defined(CFA_HAVE_LINUX_RSEQ_H) |
---|
| 83 | extern "Cforall" { |
---|
[f558b5f] | 84 | extern __attribute__((aligned(128))) thread_local volatile struct rseq __cfaabi_rseq; |
---|
[3489ea6] | 85 | } |
---|
| 86 | #else |
---|
| 87 | // No data needed |
---|
| 88 | #endif |
---|
| 89 | |
---|
| 90 | static inline int __kernel_getcpu() { |
---|
| 91 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 92 | #if defined(CFA_HAVE_LINUX_LIBRSEQ) |
---|
[f558b5f] | 93 | return rseq_current_cpu(); |
---|
[3489ea6] | 94 | #elif defined(CFA_HAVE_LINUX_RSEQ_H) |
---|
[f558b5f] | 95 | int r = __cfaabi_rseq.cpu_id; |
---|
| 96 | /* paranoid */ verify( r >= 0 ); |
---|
| 97 | return r; |
---|
[3489ea6] | 98 | #else |
---|
| 99 | return sched_getcpu(); |
---|
| 100 | #endif |
---|
| 101 | } |
---|
| 102 | |
---|
[75f3522] | 103 | //----------------------------------------------------------------------------- |
---|
| 104 | // Processor |
---|
| 105 | void main(processorCtx_t *); |
---|
[85b1deb] | 106 | |
---|
[8c50aed] | 107 | void * __create_pthread( pthread_t *, void * (*)(void *), void * ); |
---|
[bfcf6b9] | 108 | void __destroy_pthread( pthread_t pthread, void * stack, void ** retval ); |
---|
[1805b1b] | 109 | |
---|
[6502a2b] | 110 | extern cluster * mainCluster; |
---|
| 111 | |
---|
[75f3522] | 112 | //----------------------------------------------------------------------------- |
---|
| 113 | // Threads |
---|
| 114 | extern "C" { |
---|
[c7a900a] | 115 | void __cfactx_invoke_thread(void (*main)(void *), void * this); |
---|
[75f3522] | 116 | } |
---|
| 117 | |
---|
[f7d6bb0] | 118 | __cfaabi_dbg_debug_do( |
---|
[e84ab3d] | 119 | extern void __cfaabi_dbg_thread_register ( thread$ * thrd ); |
---|
| 120 | extern void __cfaabi_dbg_thread_unregister( thread$ * thrd ); |
---|
[f7d6bb0] | 121 | ) |
---|
| 122 | |
---|
[6a77224] | 123 | #define TICKET_BLOCKED (-1) // thread is blocked |
---|
| 124 | #define TICKET_RUNNING ( 0) // thread is running |
---|
| 125 | #define TICKET_UNBLOCK ( 1) // thread should ignore next block |
---|
| 126 | |
---|
[969b3fe] | 127 | //----------------------------------------------------------------------------- |
---|
| 128 | // Utils |
---|
[e84ab3d] | 129 | void doregister( struct cluster * cltr, struct thread$ & thrd ); |
---|
| 130 | void unregister( struct cluster * cltr, struct thread$ & thrd ); |
---|
[de94a60] | 131 | |
---|
[f00b26d4] | 132 | //----------------------------------------------------------------------------- |
---|
| 133 | // I/O |
---|
[78da4ab] | 134 | $io_arbiter * create(void); |
---|
| 135 | void destroy($io_arbiter *); |
---|
[f00b26d4] | 136 | |
---|
[7768b8d] | 137 | //======================================================================= |
---|
| 138 | // Cluster lock API |
---|
| 139 | //======================================================================= |
---|
| 140 | // Lock-Free registering/unregistering of threads |
---|
| 141 | // Register a processor to a given cluster and get its unique id in return |
---|
[c993b15] | 142 | unsigned register_proc_id( void ); |
---|
[7768b8d] | 143 | |
---|
| 144 | // Unregister a processor from a given cluster using its id, getting back the original pointer |
---|
[c993b15] | 145 | void unregister_proc_id( unsigned ); |
---|
[7768b8d] | 146 | |
---|
| 147 | //======================================================================= |
---|
| 148 | // Reader-writer lock implementation |
---|
| 149 | // Concurrent with doregister/unregister, |
---|
| 150 | // i.e., threads can be added at any point during or between the entry/exit |
---|
[dca5802] | 151 | |
---|
| 152 | //----------------------------------------------------------------------- |
---|
| 153 | // simple spinlock underlying the RWLock |
---|
| 154 | // Blocking acquire |
---|
[7768b8d] | 155 | static inline void __atomic_acquire(volatile bool * ll) { |
---|
| 156 | while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) { |
---|
| 157 | while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED)) |
---|
[fd9b524] | 158 | Pause(); |
---|
[7768b8d] | 159 | } |
---|
| 160 | /* paranoid */ verify(*ll); |
---|
| 161 | } |
---|
| 162 | |
---|
[dca5802] | 163 | // Non-Blocking acquire |
---|
[7768b8d] | 164 | static inline bool __atomic_try_acquire(volatile bool * ll) { |
---|
[b798713] | 165 | return !__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST); |
---|
[7768b8d] | 166 | } |
---|
| 167 | |
---|
[dca5802] | 168 | // Release |
---|
[7768b8d] | 169 | static inline void __atomic_unlock(volatile bool * ll) { |
---|
| 170 | /* paranoid */ verify(*ll); |
---|
| 171 | __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE); |
---|
| 172 | } |
---|
| 173 | |
---|
[b388ee81] | 174 | //----------------------------------------------------------------------- |
---|
| 175 | // Reader-Writer lock protecting the ready-queues |
---|
| 176 | // while this lock is mostly generic some aspects |
---|
| 177 | // have been hard-coded to for the ready-queue for |
---|
| 178 | // simplicity and performance |
---|
| 179 | struct __scheduler_RWLock_t { |
---|
| 180 | // total cachelines allocated |
---|
| 181 | unsigned int max; |
---|
| 182 | |
---|
| 183 | // cachelines currently in use |
---|
| 184 | volatile unsigned int alloc; |
---|
| 185 | |
---|
| 186 | // cachelines ready to itereate over |
---|
| 187 | // (!= to alloc when thread is in second half of doregister) |
---|
| 188 | volatile unsigned int ready; |
---|
| 189 | |
---|
| 190 | // writer lock |
---|
[c993b15] | 191 | volatile bool write_lock; |
---|
[b388ee81] | 192 | |
---|
| 193 | // data pointer |
---|
[c993b15] | 194 | volatile bool * volatile * data; |
---|
[b388ee81] | 195 | }; |
---|
| 196 | |
---|
| 197 | void ?{}(__scheduler_RWLock_t & this); |
---|
| 198 | void ^?{}(__scheduler_RWLock_t & this); |
---|
| 199 | |
---|
| 200 | extern __scheduler_RWLock_t * __scheduler_lock; |
---|
| 201 | |
---|
[7768b8d] | 202 | //----------------------------------------------------------------------- |
---|
| 203 | // Reader side : acquire when using the ready queue to schedule but not |
---|
| 204 | // creating/destroying queues |
---|
[e873838] | 205 | static inline void ready_schedule_lock(void) with(*__scheduler_lock) { |
---|
[8fc652e0] | 206 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[c993b15] | 207 | /* paranoid */ verify( ! kernelTLS().in_sched_lock ); |
---|
| 208 | /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock ); |
---|
| 209 | /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id ); |
---|
[7768b8d] | 210 | |
---|
| 211 | // Step 1 : make sure no writer are in the middle of the critical section |
---|
[c993b15] | 212 | while(__atomic_load_n(&write_lock, (int)__ATOMIC_RELAXED)) |
---|
[fd9b524] | 213 | Pause(); |
---|
[7768b8d] | 214 | |
---|
| 215 | // Fence needed because we don't want to start trying to acquire the lock |
---|
| 216 | // before we read a false. |
---|
| 217 | // Not needed on x86 |
---|
| 218 | // std::atomic_thread_fence(std::memory_order_seq_cst); |
---|
| 219 | |
---|
| 220 | // Step 2 : acquire our local lock |
---|
[c993b15] | 221 | __atomic_acquire( &kernelTLS().sched_lock ); |
---|
| 222 | /*paranoid*/ verify(kernelTLS().sched_lock); |
---|
[64a7146] | 223 | |
---|
| 224 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 225 | // Debug, check if this is owned for reading |
---|
[c993b15] | 226 | kernelTLS().in_sched_lock = true; |
---|
[64a7146] | 227 | #endif |
---|
[7768b8d] | 228 | } |
---|
| 229 | |
---|
[e873838] | 230 | static inline void ready_schedule_unlock(void) with(*__scheduler_lock) { |
---|
[8fc652e0] | 231 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[c993b15] | 232 | /* paranoid */ verify( data[kernelTLS().sched_id] == &kernelTLS().sched_lock ); |
---|
| 233 | /* paranoid */ verify( !kernelTLS().this_processor || kernelTLS().this_processor->unique_id == kernelTLS().sched_id ); |
---|
| 234 | /* paranoid */ verify( kernelTLS().sched_lock ); |
---|
| 235 | /* paranoid */ verify( kernelTLS().in_sched_lock ); |
---|
[64a7146] | 236 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 237 | // Debug, check if this is owned for reading |
---|
[c993b15] | 238 | kernelTLS().in_sched_lock = false; |
---|
[64a7146] | 239 | #endif |
---|
[c993b15] | 240 | __atomic_unlock(&kernelTLS().sched_lock); |
---|
[7768b8d] | 241 | } |
---|
| 242 | |
---|
[64a7146] | 243 | #ifdef __CFA_WITH_VERIFY__ |
---|
[e873838] | 244 | static inline bool ready_schedule_islocked(void) { |
---|
[8fc652e0] | 245 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
[c993b15] | 246 | /* paranoid */ verify( (!kernelTLS().in_sched_lock) || kernelTLS().sched_lock ); |
---|
| 247 | return kernelTLS().sched_lock; |
---|
[64a7146] | 248 | } |
---|
| 249 | |
---|
| 250 | static inline bool ready_mutate_islocked() { |
---|
[c993b15] | 251 | return __scheduler_lock->write_lock; |
---|
[64a7146] | 252 | } |
---|
| 253 | #endif |
---|
| 254 | |
---|
[7768b8d] | 255 | //----------------------------------------------------------------------- |
---|
| 256 | // Writer side : acquire when changing the ready queue, e.g. adding more |
---|
| 257 | // queues or removing them. |
---|
[b388ee81] | 258 | uint_fast32_t ready_mutate_lock( void ); |
---|
[7768b8d] | 259 | |
---|
[b388ee81] | 260 | void ready_mutate_unlock( uint_fast32_t /* value returned by lock */ ); |
---|
[7768b8d] | 261 | |
---|
[a33c113] | 262 | //----------------------------------------------------------------------- |
---|
| 263 | // Lock-Free registering/unregistering of threads |
---|
| 264 | // Register a processor to a given cluster and get its unique id in return |
---|
| 265 | // For convenience, also acquires the lock |
---|
[c993b15] | 266 | static inline [unsigned, uint_fast32_t] ready_mutate_register() { |
---|
| 267 | unsigned id = register_proc_id(); |
---|
| 268 | uint_fast32_t last = ready_mutate_lock(); |
---|
| 269 | return [id, last]; |
---|
[a33c113] | 270 | } |
---|
| 271 | |
---|
| 272 | // Unregister a processor from a given cluster using its id, getting back the original pointer |
---|
| 273 | // assumes the lock is acquired |
---|
[c993b15] | 274 | static inline void ready_mutate_unregister( unsigned id, uint_fast32_t last_s ) { |
---|
[a33c113] | 275 | ready_mutate_unlock( last_s ); |
---|
[c993b15] | 276 | unregister_proc_id( id ); |
---|
[a33c113] | 277 | } |
---|
| 278 | |
---|
[a7504db] | 279 | //----------------------------------------------------------------------- |
---|
| 280 | // Cluster idle lock/unlock |
---|
[6a9b12b] | 281 | static inline void lock(__cluster_proc_list & this) { |
---|
[a7504db] | 282 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 283 | |
---|
| 284 | // Start by locking the global RWlock so that we know no-one is |
---|
| 285 | // adding/removing processors while we mess with the idle lock |
---|
| 286 | ready_schedule_lock(); |
---|
| 287 | |
---|
[a633f6f] | 288 | lock( this.lock __cfaabi_dbg_ctx2 ); |
---|
[a7504db] | 289 | |
---|
| 290 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 291 | } |
---|
| 292 | |
---|
[5f5a729] | 293 | static inline bool try_lock(__cluster_proc_list & this) { |
---|
| 294 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 295 | |
---|
| 296 | // Start by locking the global RWlock so that we know no-one is |
---|
| 297 | // adding/removing processors while we mess with the idle lock |
---|
| 298 | ready_schedule_lock(); |
---|
| 299 | |
---|
[a633f6f] | 300 | if(try_lock( this.lock __cfaabi_dbg_ctx2 )) { |
---|
[5f5a729] | 301 | // success |
---|
| 302 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 303 | return true; |
---|
| 304 | } |
---|
| 305 | |
---|
| 306 | // failed to lock |
---|
| 307 | ready_schedule_unlock(); |
---|
| 308 | |
---|
| 309 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 310 | return false; |
---|
| 311 | } |
---|
| 312 | |
---|
[6a9b12b] | 313 | static inline void unlock(__cluster_proc_list & this) { |
---|
[a7504db] | 314 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 315 | |
---|
[a633f6f] | 316 | unlock(this.lock); |
---|
[a7504db] | 317 | |
---|
| 318 | // Release the global lock, which we acquired when locking |
---|
| 319 | ready_schedule_unlock(); |
---|
| 320 | |
---|
| 321 | /* paranoid */ verify( ! __preemption_enabled() ); |
---|
| 322 | } |
---|
| 323 | |
---|
[b798713] | 324 | //======================================================================= |
---|
| 325 | // Ready-Queue API |
---|
[dca5802] | 326 | //----------------------------------------------------------------------- |
---|
| 327 | // push thread onto a ready queue for a cluster |
---|
| 328 | // returns true if the list was previously empty, false otherwise |
---|
[24e321c] | 329 | __attribute__((hot)) void push(struct cluster * cltr, struct thread$ * thrd, unpark_hint hint); |
---|
[dca5802] | 330 | |
---|
| 331 | //----------------------------------------------------------------------- |
---|
[fc59df78] | 332 | // pop thread from the local queues of a cluster |
---|
[dca5802] | 333 | // returns 0p if empty |
---|
[1eb239e4] | 334 | // May return 0p spuriously |
---|
[e84ab3d] | 335 | __attribute__((hot)) struct thread$ * pop_fast(struct cluster * cltr); |
---|
[dca5802] | 336 | |
---|
[1eb239e4] | 337 | //----------------------------------------------------------------------- |
---|
[fc59df78] | 338 | // pop thread from any ready queue of a cluster |
---|
[1eb239e4] | 339 | // returns 0p if empty |
---|
[fc59df78] | 340 | // May return 0p spuriously |
---|
[e84ab3d] | 341 | __attribute__((hot)) struct thread$ * pop_slow(struct cluster * cltr); |
---|
[1eb239e4] | 342 | |
---|
[fc59df78] | 343 | //----------------------------------------------------------------------- |
---|
| 344 | // search all ready queues of a cluster for any thread |
---|
| 345 | // returns 0p if empty |
---|
| 346 | // guaranteed to find any threads added before this call |
---|
[e84ab3d] | 347 | __attribute__((hot)) struct thread$ * pop_search(struct cluster * cltr); |
---|
[fc59df78] | 348 | |
---|
[24e321c] | 349 | //----------------------------------------------------------------------- |
---|
| 350 | // get preferred ready for new thread |
---|
| 351 | unsigned ready_queue_new_preferred(); |
---|
| 352 | |
---|
[dca5802] | 353 | //----------------------------------------------------------------------- |
---|
| 354 | // Increase the width of the ready queue (number of lanes) by 4 |
---|
[a017ee7] | 355 | void ready_queue_grow (struct cluster * cltr); |
---|
[dca5802] | 356 | |
---|
| 357 | //----------------------------------------------------------------------- |
---|
| 358 | // Decrease the width of the ready queue (number of lanes) by 4 |
---|
[a017ee7] | 359 | void ready_queue_shrink(struct cluster * cltr); |
---|
[b798713] | 360 | |
---|
[de94a60] | 361 | |
---|
[75f3522] | 362 | // Local Variables: // |
---|
| 363 | // mode: c // |
---|
| 364 | // tab-width: 4 // |
---|
[4aa2fb2] | 365 | // End: // |
---|