[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 | |
---|
[58b6d1b] | 18 | #include "kernel.hfa" |
---|
| 19 | #include "thread.hfa" |
---|
[75f3522] | 20 | |
---|
[73abe95] | 21 | #include "alarm.hfa" |
---|
[8834751] | 22 | #include "stats.hfa" |
---|
[fa21ac9] | 23 | |
---|
[75f3522] | 24 | //----------------------------------------------------------------------------- |
---|
| 25 | // Scheduler |
---|
[1c273d0] | 26 | |
---|
[37ba662] | 27 | struct __attribute__((aligned(128))) __scheduler_lock_id_t; |
---|
[9b1dcc2] | 28 | |
---|
[1c273d0] | 29 | extern "C" { |
---|
[2026bb6] | 30 | void disable_interrupts() OPTIONAL_THREAD; |
---|
[969b3fe] | 31 | void enable_interrupts_noPoll(); |
---|
[36982fc] | 32 | void enable_interrupts( __cfaabi_dbg_ctx_param ); |
---|
[1c273d0] | 33 | } |
---|
| 34 | |
---|
[efc171d1] | 35 | void __schedule_thread( struct __processor_id_t *, $thread * ) |
---|
| 36 | #if defined(NDEBUG) || (!defined(__CFA_DEBUG__) && !defined(__CFA_VERIFY__)) |
---|
| 37 | __attribute__((nonnull (2))) |
---|
| 38 | #endif |
---|
| 39 | ; |
---|
[75f3522] | 40 | |
---|
[e60e0dc] | 41 | //Block current thread and release/wake-up the following resources |
---|
[b0c7419] | 42 | void __leave_thread() __attribute__((noreturn)); |
---|
[db6f06a] | 43 | |
---|
[75f3522] | 44 | //----------------------------------------------------------------------------- |
---|
| 45 | // Processor |
---|
| 46 | void main(processorCtx_t *); |
---|
[85b1deb] | 47 | |
---|
[8c50aed] | 48 | void * __create_pthread( pthread_t *, void * (*)(void *), void * ); |
---|
[1805b1b] | 49 | |
---|
[85b1deb] | 50 | |
---|
[75f3522] | 51 | |
---|
[6502a2b] | 52 | extern cluster * mainCluster; |
---|
| 53 | |
---|
[75f3522] | 54 | //----------------------------------------------------------------------------- |
---|
| 55 | // Threads |
---|
| 56 | extern "C" { |
---|
[c7a900a] | 57 | void __cfactx_invoke_thread(void (*main)(void *), void * this); |
---|
[75f3522] | 58 | } |
---|
| 59 | |
---|
[f7d6bb0] | 60 | __cfaabi_dbg_debug_do( |
---|
[ac2b598] | 61 | extern void __cfaabi_dbg_thread_register ( $thread * thrd ); |
---|
| 62 | extern void __cfaabi_dbg_thread_unregister( $thread * thrd ); |
---|
[f7d6bb0] | 63 | ) |
---|
| 64 | |
---|
[2d8f7b0] | 65 | // KERNEL ONLY unpark with out disabling interrupts |
---|
[e235429] | 66 | void __unpark( struct __processor_id_t *, $thread * thrd ); |
---|
[2d8f7b0] | 67 | |
---|
[f00b26d4] | 68 | static inline bool __post(single_sem & this, struct __processor_id_t * id) { |
---|
| 69 | for() { |
---|
| 70 | struct $thread * expected = this.ptr; |
---|
| 71 | if(expected == 1p) return false; |
---|
| 72 | if(expected == 0p) { |
---|
| 73 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 1p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { |
---|
| 74 | return false; |
---|
| 75 | } |
---|
| 76 | } |
---|
| 77 | else { |
---|
| 78 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { |
---|
[e235429] | 79 | __unpark( id, expected ); |
---|
[f00b26d4] | 80 | return true; |
---|
| 81 | } |
---|
| 82 | } |
---|
| 83 | } |
---|
| 84 | } |
---|
[92976d9] | 85 | |
---|
[969b3fe] | 86 | //----------------------------------------------------------------------------- |
---|
| 87 | // Utils |
---|
[ac2b598] | 88 | void doregister( struct cluster * cltr, struct $thread & thrd ); |
---|
| 89 | void unregister( struct cluster * cltr, struct $thread & thrd ); |
---|
[de94a60] | 90 | |
---|
[f00b26d4] | 91 | //----------------------------------------------------------------------------- |
---|
| 92 | // I/O |
---|
| 93 | void ^?{}(io_context & this, bool ); |
---|
| 94 | |
---|
[7768b8d] | 95 | //======================================================================= |
---|
| 96 | // Cluster lock API |
---|
| 97 | //======================================================================= |
---|
[b388ee81] | 98 | // Cells use by the reader writer lock |
---|
| 99 | // while not generic it only relies on a opaque pointer |
---|
[37ba662] | 100 | struct __attribute__((aligned(128))) __scheduler_lock_id_t { |
---|
[64a7146] | 101 | // Spin lock used as the underlying lock |
---|
[7768b8d] | 102 | volatile bool lock; |
---|
[64a7146] | 103 | |
---|
| 104 | // Handle pointing to the proc owning this cell |
---|
| 105 | // Used for allocating cells and debugging |
---|
| 106 | __processor_id_t * volatile handle; |
---|
| 107 | |
---|
| 108 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 109 | // Debug, check if this is owned for reading |
---|
| 110 | bool owned; |
---|
| 111 | #endif |
---|
[7768b8d] | 112 | }; |
---|
| 113 | |
---|
[64a7146] | 114 | static_assert( sizeof(struct __scheduler_lock_id_t) <= __alignof(struct __scheduler_lock_id_t)); |
---|
| 115 | |
---|
[7768b8d] | 116 | // Lock-Free registering/unregistering of threads |
---|
| 117 | // Register a processor to a given cluster and get its unique id in return |
---|
[9b1dcc2] | 118 | unsigned doregister( struct __processor_id_t * proc ); |
---|
[7768b8d] | 119 | |
---|
| 120 | // Unregister a processor from a given cluster using its id, getting back the original pointer |
---|
[9b1dcc2] | 121 | void unregister( struct __processor_id_t * proc ); |
---|
[7768b8d] | 122 | |
---|
[1eb239e4] | 123 | //----------------------------------------------------------------------- |
---|
| 124 | // Cluster idle lock/unlock |
---|
| 125 | static inline void lock(__cluster_idles & this) { |
---|
| 126 | for() { |
---|
| 127 | uint64_t l = this.lock; |
---|
| 128 | if( |
---|
| 129 | (0 == (l % 2)) |
---|
| 130 | && __atomic_compare_exchange_n(&this.lock, &l, l + 1, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) |
---|
| 131 | ) return; |
---|
| 132 | Pause(); |
---|
| 133 | } |
---|
| 134 | } |
---|
| 135 | |
---|
| 136 | static inline void unlock(__cluster_idles & this) { |
---|
| 137 | /* paranoid */ verify( 1 == (this.lock % 2) ); |
---|
| 138 | __atomic_fetch_add( &this.lock, 1, __ATOMIC_SEQ_CST ); |
---|
| 139 | } |
---|
| 140 | |
---|
[7768b8d] | 141 | //======================================================================= |
---|
| 142 | // Reader-writer lock implementation |
---|
| 143 | // Concurrent with doregister/unregister, |
---|
| 144 | // i.e., threads can be added at any point during or between the entry/exit |
---|
[dca5802] | 145 | |
---|
| 146 | //----------------------------------------------------------------------- |
---|
| 147 | // simple spinlock underlying the RWLock |
---|
| 148 | // Blocking acquire |
---|
[7768b8d] | 149 | static inline void __atomic_acquire(volatile bool * ll) { |
---|
| 150 | while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) { |
---|
| 151 | while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED)) |
---|
[fd9b524] | 152 | Pause(); |
---|
[7768b8d] | 153 | } |
---|
| 154 | /* paranoid */ verify(*ll); |
---|
| 155 | } |
---|
| 156 | |
---|
[dca5802] | 157 | // Non-Blocking acquire |
---|
[7768b8d] | 158 | static inline bool __atomic_try_acquire(volatile bool * ll) { |
---|
[b798713] | 159 | return !__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST); |
---|
[7768b8d] | 160 | } |
---|
| 161 | |
---|
[dca5802] | 162 | // Release |
---|
[7768b8d] | 163 | static inline void __atomic_unlock(volatile bool * ll) { |
---|
| 164 | /* paranoid */ verify(*ll); |
---|
| 165 | __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE); |
---|
| 166 | } |
---|
| 167 | |
---|
[b388ee81] | 168 | //----------------------------------------------------------------------- |
---|
| 169 | // Reader-Writer lock protecting the ready-queues |
---|
| 170 | // while this lock is mostly generic some aspects |
---|
| 171 | // have been hard-coded to for the ready-queue for |
---|
| 172 | // simplicity and performance |
---|
| 173 | struct __scheduler_RWLock_t { |
---|
| 174 | // total cachelines allocated |
---|
| 175 | unsigned int max; |
---|
| 176 | |
---|
| 177 | // cachelines currently in use |
---|
| 178 | volatile unsigned int alloc; |
---|
| 179 | |
---|
| 180 | // cachelines ready to itereate over |
---|
| 181 | // (!= to alloc when thread is in second half of doregister) |
---|
| 182 | volatile unsigned int ready; |
---|
| 183 | |
---|
| 184 | // writer lock |
---|
| 185 | volatile bool lock; |
---|
| 186 | |
---|
| 187 | // data pointer |
---|
[9b1dcc2] | 188 | __scheduler_lock_id_t * data; |
---|
[b388ee81] | 189 | }; |
---|
| 190 | |
---|
| 191 | void ?{}(__scheduler_RWLock_t & this); |
---|
| 192 | void ^?{}(__scheduler_RWLock_t & this); |
---|
| 193 | |
---|
| 194 | extern __scheduler_RWLock_t * __scheduler_lock; |
---|
| 195 | |
---|
[7768b8d] | 196 | //----------------------------------------------------------------------- |
---|
| 197 | // Reader side : acquire when using the ready queue to schedule but not |
---|
| 198 | // creating/destroying queues |
---|
[9b1dcc2] | 199 | static inline void ready_schedule_lock( struct __processor_id_t * proc) with(*__scheduler_lock) { |
---|
[7768b8d] | 200 | unsigned iproc = proc->id; |
---|
| 201 | /*paranoid*/ verify(data[iproc].handle == proc); |
---|
| 202 | /*paranoid*/ verify(iproc < ready); |
---|
| 203 | |
---|
| 204 | // Step 1 : make sure no writer are in the middle of the critical section |
---|
| 205 | while(__atomic_load_n(&lock, (int)__ATOMIC_RELAXED)) |
---|
[fd9b524] | 206 | Pause(); |
---|
[7768b8d] | 207 | |
---|
| 208 | // Fence needed because we don't want to start trying to acquire the lock |
---|
| 209 | // before we read a false. |
---|
| 210 | // Not needed on x86 |
---|
| 211 | // std::atomic_thread_fence(std::memory_order_seq_cst); |
---|
| 212 | |
---|
| 213 | // Step 2 : acquire our local lock |
---|
| 214 | __atomic_acquire( &data[iproc].lock ); |
---|
| 215 | /*paranoid*/ verify(data[iproc].lock); |
---|
[64a7146] | 216 | |
---|
| 217 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 218 | // Debug, check if this is owned for reading |
---|
| 219 | data[iproc].owned = true; |
---|
| 220 | #endif |
---|
[7768b8d] | 221 | } |
---|
| 222 | |
---|
[9b1dcc2] | 223 | static inline void ready_schedule_unlock( struct __processor_id_t * proc) with(*__scheduler_lock) { |
---|
[7768b8d] | 224 | unsigned iproc = proc->id; |
---|
| 225 | /*paranoid*/ verify(data[iproc].handle == proc); |
---|
| 226 | /*paranoid*/ verify(iproc < ready); |
---|
| 227 | /*paranoid*/ verify(data[iproc].lock); |
---|
[64a7146] | 228 | /*paranoid*/ verify(data[iproc].owned); |
---|
| 229 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 230 | // Debug, check if this is owned for reading |
---|
| 231 | data[iproc].owned = false; |
---|
| 232 | #endif |
---|
[dca5802] | 233 | __atomic_unlock(&data[iproc].lock); |
---|
[7768b8d] | 234 | } |
---|
| 235 | |
---|
[64a7146] | 236 | #ifdef __CFA_WITH_VERIFY__ |
---|
| 237 | static inline bool ready_schedule_islocked( struct __processor_id_t * proc) { |
---|
| 238 | return __scheduler_lock->data[proc->id].owned; |
---|
| 239 | } |
---|
| 240 | |
---|
| 241 | static inline bool ready_mutate_islocked() { |
---|
| 242 | return __scheduler_lock->lock; |
---|
| 243 | } |
---|
| 244 | #endif |
---|
| 245 | |
---|
[7768b8d] | 246 | //----------------------------------------------------------------------- |
---|
| 247 | // Writer side : acquire when changing the ready queue, e.g. adding more |
---|
| 248 | // queues or removing them. |
---|
[b388ee81] | 249 | uint_fast32_t ready_mutate_lock( void ); |
---|
[7768b8d] | 250 | |
---|
[b388ee81] | 251 | void ready_mutate_unlock( uint_fast32_t /* value returned by lock */ ); |
---|
[7768b8d] | 252 | |
---|
[b798713] | 253 | //======================================================================= |
---|
| 254 | // Ready-Queue API |
---|
[64a7146] | 255 | //----------------------------------------------------------------------- |
---|
| 256 | // pop thread from the ready queue of a cluster |
---|
| 257 | // returns 0p if empty |
---|
| 258 | __attribute__((hot)) bool query(struct cluster * cltr); |
---|
| 259 | |
---|
[dca5802] | 260 | //----------------------------------------------------------------------- |
---|
| 261 | // push thread onto a ready queue for a cluster |
---|
| 262 | // returns true if the list was previously empty, false otherwise |
---|
[504a7dc] | 263 | __attribute__((hot)) bool push(struct cluster * cltr, struct $thread * thrd); |
---|
[dca5802] | 264 | |
---|
| 265 | //----------------------------------------------------------------------- |
---|
| 266 | // pop thread from the ready queue of a cluster |
---|
| 267 | // returns 0p if empty |
---|
[1eb239e4] | 268 | // May return 0p spuriously |
---|
[504a7dc] | 269 | __attribute__((hot)) struct $thread * pop(struct cluster * cltr); |
---|
[dca5802] | 270 | |
---|
[1eb239e4] | 271 | //----------------------------------------------------------------------- |
---|
| 272 | // pop thread from the ready queue of a cluster |
---|
| 273 | // returns 0p if empty |
---|
| 274 | // guaranteed to find any threads added before this call |
---|
| 275 | __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr); |
---|
| 276 | |
---|
[13c5e19] | 277 | //----------------------------------------------------------------------- |
---|
| 278 | // remove thread from the ready queue of a cluster |
---|
| 279 | // returns bool if it wasn't found |
---|
| 280 | bool remove_head(struct cluster * cltr, struct $thread * thrd); |
---|
| 281 | |
---|
[dca5802] | 282 | //----------------------------------------------------------------------- |
---|
| 283 | // Increase the width of the ready queue (number of lanes) by 4 |
---|
[320ec6fc] | 284 | void ready_queue_grow (struct cluster * cltr, int target); |
---|
[dca5802] | 285 | |
---|
| 286 | //----------------------------------------------------------------------- |
---|
| 287 | // Decrease the width of the ready queue (number of lanes) by 4 |
---|
[320ec6fc] | 288 | void ready_queue_shrink(struct cluster * cltr, int target); |
---|
[b798713] | 289 | |
---|
[de94a60] | 290 | |
---|
[75f3522] | 291 | // Local Variables: // |
---|
| 292 | // mode: c // |
---|
| 293 | // tab-width: 4 // |
---|
[4aa2fb2] | 294 | // End: // |
---|