| 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 | //
|
|---|
| 7 | // kernel --
|
|---|
| 8 | //
|
|---|
| 9 | // Author : Thierry Delisle
|
|---|
| 10 | // Created On : Tue Jan 17 12:27:26 2017
|
|---|
| 11 | // Last Modified By : Peter A. Buhr
|
|---|
| 12 | // Last Modified On : Tue Feb 4 12:29:26 2020
|
|---|
| 13 | // Update Count : 22
|
|---|
| 14 | //
|
|---|
| 15 |
|
|---|
| 16 | #pragma once
|
|---|
| 17 |
|
|---|
| 18 | #include "invoke.h"
|
|---|
| 19 | #include "time_t.hfa"
|
|---|
| 20 | #include "coroutine.hfa"
|
|---|
| 21 |
|
|---|
| 22 | #include "containers/stackLockFree.hfa"
|
|---|
| 23 |
|
|---|
| 24 | extern "C" {
|
|---|
| 25 | #include <bits/pthreadtypes.h>
|
|---|
| 26 | }
|
|---|
| 27 |
|
|---|
| 28 | //-----------------------------------------------------------------------------
|
|---|
| 29 | // Locks
|
|---|
| 30 | struct semaphore {
|
|---|
| 31 | __spinlock_t lock;
|
|---|
| 32 | int count;
|
|---|
| 33 | __queue_t($thread) waiting;
|
|---|
| 34 | };
|
|---|
| 35 |
|
|---|
| 36 | void ?{}(semaphore & this, int count = 1);
|
|---|
| 37 | void ^?{}(semaphore & this);
|
|---|
| 38 | bool P (semaphore & this);
|
|---|
| 39 | bool V (semaphore & this);
|
|---|
| 40 | bool V (semaphore & this, unsigned count);
|
|---|
| 41 |
|
|---|
| 42 |
|
|---|
| 43 | //-----------------------------------------------------------------------------
|
|---|
| 44 | // Processor
|
|---|
| 45 | extern struct cluster * mainCluster;
|
|---|
| 46 |
|
|---|
| 47 | // Processor id, required for scheduling threads
|
|---|
| 48 | struct __processor_id_t {
|
|---|
| 49 | unsigned id;
|
|---|
| 50 |
|
|---|
| 51 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 52 | struct __stats_t * stats;
|
|---|
| 53 | #endif
|
|---|
| 54 | };
|
|---|
| 55 |
|
|---|
| 56 | coroutine processorCtx_t {
|
|---|
| 57 | struct processor * proc;
|
|---|
| 58 | };
|
|---|
| 59 |
|
|---|
| 60 | // Wrapper around kernel threads
|
|---|
| 61 | struct __attribute__((aligned(128))) processor {
|
|---|
| 62 | // Main state
|
|---|
| 63 | inline __processor_id_t;
|
|---|
| 64 |
|
|---|
| 65 | // Cluster from which to get threads
|
|---|
| 66 | struct cluster * cltr;
|
|---|
| 67 |
|
|---|
| 68 | // Set to true to notify the processor should terminate
|
|---|
| 69 | volatile bool do_terminate;
|
|---|
| 70 |
|
|---|
| 71 | // Coroutine ctx who does keeps the state of the processor
|
|---|
| 72 | struct processorCtx_t runner;
|
|---|
| 73 |
|
|---|
| 74 | // Name of the processor
|
|---|
| 75 | const char * name;
|
|---|
| 76 |
|
|---|
| 77 | // Handle to pthreads
|
|---|
| 78 | pthread_t kernel_thread;
|
|---|
| 79 |
|
|---|
| 80 | // RunThread data
|
|---|
| 81 | // Action to do after a thread is ran
|
|---|
| 82 | $thread * destroyer;
|
|---|
| 83 |
|
|---|
| 84 | // Preemption data
|
|---|
| 85 | // Node which is added in the discrete event simulaiton
|
|---|
| 86 | struct alarm_node_t * preemption_alarm;
|
|---|
| 87 |
|
|---|
| 88 | // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
|
|---|
| 89 | bool pending_preemption;
|
|---|
| 90 |
|
|---|
| 91 | // Idle lock (kernel semaphore)
|
|---|
| 92 | __bin_sem_t idle;
|
|---|
| 93 |
|
|---|
| 94 | // Termination synchronisation (user semaphore)
|
|---|
| 95 | semaphore terminated;
|
|---|
| 96 |
|
|---|
| 97 | // pthread Stack
|
|---|
| 98 | void * stack;
|
|---|
| 99 |
|
|---|
| 100 | // Link lists fields
|
|---|
| 101 | Link(processor) link;
|
|---|
| 102 |
|
|---|
| 103 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 104 | int print_stats;
|
|---|
| 105 | bool print_halts;
|
|---|
| 106 | #endif
|
|---|
| 107 |
|
|---|
| 108 | #ifdef __CFA_DEBUG__
|
|---|
| 109 | // Last function to enable preemption on this processor
|
|---|
| 110 | const char * last_enable;
|
|---|
| 111 | #endif
|
|---|
| 112 | };
|
|---|
| 113 |
|
|---|
| 114 | void ?{}(processor & this, const char name[], struct cluster & cltr);
|
|---|
| 115 | void ^?{}(processor & this);
|
|---|
| 116 |
|
|---|
| 117 | static inline void ?{}(processor & this) { this{ "Anonymous Processor", *mainCluster}; }
|
|---|
| 118 | static inline void ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
|
|---|
| 119 | static inline void ?{}(processor & this, const char name[]) { this{name, *mainCluster }; }
|
|---|
| 120 |
|
|---|
| 121 | static inline Link(processor) * ?`next( processor * this ) { return &this->link; }
|
|---|
| 122 |
|
|---|
| 123 | //-----------------------------------------------------------------------------
|
|---|
| 124 | // I/O
|
|---|
| 125 | struct __io_data;
|
|---|
| 126 |
|
|---|
| 127 | // IO poller user-thread
|
|---|
| 128 | // Not using the "thread" keyword because we want to control
|
|---|
| 129 | // more carefully when to start/stop it
|
|---|
| 130 | struct $io_ctx_thread {
|
|---|
| 131 | struct __io_data * ring;
|
|---|
| 132 | single_sem sem;
|
|---|
| 133 | volatile bool done;
|
|---|
| 134 | $thread self;
|
|---|
| 135 | };
|
|---|
| 136 |
|
|---|
| 137 |
|
|---|
| 138 | struct io_context {
|
|---|
| 139 | $io_ctx_thread thrd;
|
|---|
| 140 | };
|
|---|
| 141 |
|
|---|
| 142 | struct io_context_params {
|
|---|
| 143 | int num_entries;
|
|---|
| 144 | int num_ready;
|
|---|
| 145 | int submit_aff;
|
|---|
| 146 | bool eager_submits:1;
|
|---|
| 147 | bool poller_submits:1;
|
|---|
| 148 | bool poll_submit:1;
|
|---|
| 149 | bool poll_complete:1;
|
|---|
| 150 | };
|
|---|
| 151 |
|
|---|
| 152 | void ?{}(io_context_params & this);
|
|---|
| 153 |
|
|---|
| 154 | void ?{}(io_context & this, struct cluster & cl);
|
|---|
| 155 | void ?{}(io_context & this, struct cluster & cl, const io_context_params & params);
|
|---|
| 156 | void ^?{}(io_context & this);
|
|---|
| 157 |
|
|---|
| 158 | struct io_cancellation {
|
|---|
| 159 | uint32_t target;
|
|---|
| 160 | };
|
|---|
| 161 |
|
|---|
| 162 | static inline void ?{}(io_cancellation & this) { this.target = -1u; }
|
|---|
| 163 | static inline void ^?{}(io_cancellation & this) {}
|
|---|
| 164 | bool cancel(io_cancellation & this);
|
|---|
| 165 |
|
|---|
| 166 | //-----------------------------------------------------------------------------
|
|---|
| 167 | // Cluster Tools
|
|---|
| 168 |
|
|---|
| 169 | // Intrusives lanes which are used by the relaxed ready queue
|
|---|
| 170 | struct __attribute__((aligned(128))) __intrusive_lane_t;
|
|---|
| 171 | void ?{}(__intrusive_lane_t & this);
|
|---|
| 172 | void ^?{}(__intrusive_lane_t & this);
|
|---|
| 173 |
|
|---|
| 174 | // Counter used for wether or not the lanes are all empty
|
|---|
| 175 | struct __attribute__((aligned(128))) __snzi_node_t;
|
|---|
| 176 | struct __snzi_t {
|
|---|
| 177 | unsigned mask;
|
|---|
| 178 | int root;
|
|---|
| 179 | __snzi_node_t * nodes;
|
|---|
| 180 | };
|
|---|
| 181 |
|
|---|
| 182 | void ?{}( __snzi_t & this, unsigned depth );
|
|---|
| 183 | void ^?{}( __snzi_t & this );
|
|---|
| 184 |
|
|---|
| 185 | //TODO adjust cache size to ARCHITECTURE
|
|---|
| 186 | // Structure holding the relaxed ready queue
|
|---|
| 187 | struct __ready_queue_t {
|
|---|
| 188 | // Data tracking how many/which lanes are used
|
|---|
| 189 | // Aligned to 128 for cache locality
|
|---|
| 190 | __snzi_t snzi;
|
|---|
| 191 |
|
|---|
| 192 | // Data tracking the actual lanes
|
|---|
| 193 | // On a seperate cacheline from the used struct since
|
|---|
| 194 | // used can change on each push/pop but this data
|
|---|
| 195 | // only changes on shrink/grow
|
|---|
| 196 | struct {
|
|---|
| 197 | // Arary of lanes
|
|---|
| 198 | __intrusive_lane_t * volatile data;
|
|---|
| 199 |
|
|---|
| 200 | // Number of lanes (empty or not)
|
|---|
| 201 | volatile size_t count;
|
|---|
| 202 | } lanes;
|
|---|
| 203 | };
|
|---|
| 204 |
|
|---|
| 205 | void ?{}(__ready_queue_t & this);
|
|---|
| 206 | void ^?{}(__ready_queue_t & this);
|
|---|
| 207 |
|
|---|
| 208 | //-----------------------------------------------------------------------------
|
|---|
| 209 | // Cluster
|
|---|
| 210 | struct __attribute__((aligned(128))) cluster {
|
|---|
| 211 | // Ready queue for threads
|
|---|
| 212 | __ready_queue_t ready_queue;
|
|---|
| 213 |
|
|---|
| 214 | // Name of the cluster
|
|---|
| 215 | const char * name;
|
|---|
| 216 |
|
|---|
| 217 | // Preemption rate on this cluster
|
|---|
| 218 | Duration preemption_rate;
|
|---|
| 219 |
|
|---|
| 220 | // List of idle processors
|
|---|
| 221 | StackLF(processor) idles;
|
|---|
| 222 | volatile unsigned int nprocessors;
|
|---|
| 223 |
|
|---|
| 224 | // List of threads
|
|---|
| 225 | __spinlock_t thread_list_lock;
|
|---|
| 226 | __dllist_t(struct $thread) threads;
|
|---|
| 227 | unsigned int nthreads;
|
|---|
| 228 |
|
|---|
| 229 | // Link lists fields
|
|---|
| 230 | struct __dbg_node_cltr {
|
|---|
| 231 | cluster * next;
|
|---|
| 232 | cluster * prev;
|
|---|
| 233 | } node;
|
|---|
| 234 |
|
|---|
| 235 | struct {
|
|---|
| 236 | io_context * ctxs;
|
|---|
| 237 | unsigned cnt;
|
|---|
| 238 | } io;
|
|---|
| 239 |
|
|---|
| 240 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 241 | struct __stats_t * stats;
|
|---|
| 242 | int print_stats;
|
|---|
| 243 | #endif
|
|---|
| 244 | };
|
|---|
| 245 | extern Duration default_preemption();
|
|---|
| 246 |
|
|---|
| 247 | void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
|
|---|
| 248 | void ^?{}(cluster & this);
|
|---|
| 249 |
|
|---|
| 250 | static inline void ?{} (cluster & this) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
|
|---|
| 251 | static inline void ?{} (cluster & this, Duration preemption_rate) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
|
|---|
| 252 | static inline void ?{} (cluster & this, const char name[]) { io_context_params default_params; this{name, default_preemption(), 1, default_params}; }
|
|---|
| 253 | static inline void ?{} (cluster & this, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
|
|---|
| 254 | static inline void ?{} (cluster & this, Duration preemption_rate, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, num_io, default_params}; }
|
|---|
| 255 | static inline void ?{} (cluster & this, const char name[], unsigned num_io) { io_context_params default_params; this{name, default_preemption(), num_io, default_params}; }
|
|---|
| 256 | static inline void ?{} (cluster & this, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
|
|---|
| 257 | static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
|
|---|
| 258 | static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params) { this{name, default_preemption(), 1, io_params}; }
|
|---|
| 259 | static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
|
|---|
| 260 | static inline void ?{} (cluster & this, Duration preemption_rate, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, num_io, io_params}; }
|
|---|
| 261 | static inline void ?{} (cluster & this, const char name[], unsigned num_io, const io_context_params & io_params) { this{name, default_preemption(), num_io, io_params}; }
|
|---|
| 262 |
|
|---|
| 263 | static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }
|
|---|
| 264 |
|
|---|
| 265 | static inline struct processor * active_processor() { return TL_GET( this_processor ); } // UNSAFE
|
|---|
| 266 | static inline struct cluster * active_cluster () { return TL_GET( this_processor )->cltr; }
|
|---|
| 267 |
|
|---|
| 268 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 269 | static inline void print_stats_at_exit( cluster & this, int flags ) {
|
|---|
| 270 | this.print_stats |= flags;
|
|---|
| 271 | }
|
|---|
| 272 |
|
|---|
| 273 | static inline void print_stats_at_exit( processor & this, int flags ) {
|
|---|
| 274 | this.print_stats |= flags;
|
|---|
| 275 | }
|
|---|
| 276 |
|
|---|
| 277 | void print_halts( processor & this );
|
|---|
| 278 | #endif
|
|---|
| 279 |
|
|---|
| 280 | // Local Variables: //
|
|---|
| 281 | // mode: c //
|
|---|
| 282 | // tab-width: 4 //
|
|---|
| 283 | // End: //
|
|---|
