| 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 -- Header containing the core of the kernel API
|
|---|
| 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 : Wed Aug 30 21:28:46 2023
|
|---|
| 13 | // Update Count : 23
|
|---|
| 14 | //
|
|---|
| 15 |
|
|---|
| 16 | #pragma once
|
|---|
| 17 |
|
|---|
| 18 | #include "invoke.h"
|
|---|
| 19 | #include "time_t.hfa"
|
|---|
| 20 | #include "coroutine.hfa"
|
|---|
| 21 |
|
|---|
| 22 | #include "collections/list.hfa"
|
|---|
| 23 |
|
|---|
| 24 | extern "C" {
|
|---|
| 25 | #include <bits/pthreadtypes.h>
|
|---|
| 26 | #include <pthread.h>
|
|---|
| 27 | #include <linux/types.h>
|
|---|
| 28 | }
|
|---|
| 29 |
|
|---|
| 30 | #ifdef __CFA_WITH_VERIFY__
|
|---|
| 31 | extern bool __cfaabi_dbg_in_kernel();
|
|---|
| 32 | #endif
|
|---|
| 33 |
|
|---|
| 34 | //-----------------------------------------------------------------------------
|
|---|
| 35 | // I/O
|
|---|
| 36 | struct cluster;
|
|---|
| 37 | struct io_context$;
|
|---|
| 38 | struct io_arbiter$;
|
|---|
| 39 |
|
|---|
| 40 | struct io_context_params {
|
|---|
| 41 | int num_entries;
|
|---|
| 42 | };
|
|---|
| 43 |
|
|---|
| 44 | void ?{}(io_context_params & this);
|
|---|
| 45 |
|
|---|
| 46 | //-----------------------------------------------------------------------------
|
|---|
| 47 | // Processor
|
|---|
| 48 | extern struct cluster * mainCluster;
|
|---|
| 49 |
|
|---|
| 50 | // Coroutine used py processors for the 2-step context switch
|
|---|
| 51 |
|
|---|
| 52 | struct processorCtx_t {
|
|---|
| 53 | struct coroutine$ self;
|
|---|
| 54 | struct processor * proc;
|
|---|
| 55 | };
|
|---|
| 56 |
|
|---|
| 57 | struct io_future_t;
|
|---|
| 58 |
|
|---|
| 59 | // Information needed for idle sleep
|
|---|
| 60 | struct __fd_waitctx {
|
|---|
| 61 | // semaphore/future like object
|
|---|
| 62 | // values can be 0, 1 or some file descriptor.
|
|---|
| 63 | // 0 - is the default state
|
|---|
| 64 | // 1 - means the proc should wake-up immediately
|
|---|
| 65 | // FD - means the proc is going asleep and should be woken by writing to the FD.
|
|---|
| 66 | // The FD value should always be the evfd field just below.
|
|---|
| 67 | volatile int sem;
|
|---|
| 68 |
|
|---|
| 69 | // The event FD that corresponds to this processor
|
|---|
| 70 | int evfd;
|
|---|
| 71 |
|
|---|
| 72 | // Used for debugging, should be removed eventually.
|
|---|
| 73 | volatile unsigned long long wake__time;
|
|---|
| 74 | volatile unsigned long long sleep_time;
|
|---|
| 75 | volatile unsigned long long drain_time;
|
|---|
| 76 | };
|
|---|
| 77 |
|
|---|
| 78 | // Wrapper around kernel threads
|
|---|
| 79 | struct __attribute__((aligned(64))) processor {
|
|---|
| 80 | // Cluster from which to get threads
|
|---|
| 81 | struct cluster * cltr;
|
|---|
| 82 |
|
|---|
| 83 | // Ready Queue state per processor
|
|---|
| 84 | struct {
|
|---|
| 85 | unsigned short its;
|
|---|
| 86 | unsigned short itr;
|
|---|
| 87 | unsigned id;
|
|---|
| 88 | unsigned target;
|
|---|
| 89 | unsigned last;
|
|---|
| 90 | signed cpu;
|
|---|
| 91 | } rdq;
|
|---|
| 92 |
|
|---|
| 93 | // Set to true to notify the processor should terminate
|
|---|
| 94 | volatile bool do_terminate;
|
|---|
| 95 |
|
|---|
| 96 | // Coroutine ctx who does keeps the state of the processor
|
|---|
| 97 | struct processorCtx_t runner;
|
|---|
| 98 |
|
|---|
| 99 | // Name of the processor
|
|---|
| 100 | const char * name;
|
|---|
| 101 |
|
|---|
| 102 | // Handle to pthreads
|
|---|
| 103 | pthread_t kernel_thread;
|
|---|
| 104 |
|
|---|
| 105 | // Unique id for the processor (not per cluster)
|
|---|
| 106 | unsigned unique_id;
|
|---|
| 107 |
|
|---|
| 108 | struct {
|
|---|
| 109 | io_context$ * ctx;
|
|---|
| 110 | unsigned target;
|
|---|
| 111 | volatile bool pending;
|
|---|
| 112 | volatile bool dirty;
|
|---|
| 113 | } io;
|
|---|
| 114 |
|
|---|
| 115 | // Preemption data
|
|---|
| 116 | // Node which is added in the discrete event simulaiton
|
|---|
| 117 | struct alarm_node_t * preemption_alarm;
|
|---|
| 118 |
|
|---|
| 119 | // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
|
|---|
| 120 | bool pending_preemption;
|
|---|
| 121 |
|
|---|
| 122 | // context for idle sleep
|
|---|
| 123 | struct __fd_waitctx idle_wctx;
|
|---|
| 124 |
|
|---|
| 125 | // Termination synchronisation (user semaphore)
|
|---|
| 126 | oneshot terminated;
|
|---|
| 127 |
|
|---|
| 128 | // pthread Stack
|
|---|
| 129 | void * stack;
|
|---|
| 130 |
|
|---|
| 131 | // Link lists fields
|
|---|
| 132 | dlink(processor) link;
|
|---|
| 133 |
|
|---|
| 134 | // special init fields
|
|---|
| 135 | // This is needed for memcached integration
|
|---|
| 136 | // once memcached experiments are done this should probably be removed
|
|---|
| 137 | // it is not a particularly safe scheme as it can make processors less homogeneous
|
|---|
| 138 | struct {
|
|---|
| 139 | thread$ * thrd;
|
|---|
| 140 | } init;
|
|---|
| 141 |
|
|---|
| 142 | struct KernelThreadData * local_data;
|
|---|
| 143 |
|
|---|
| 144 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 145 | int print_stats;
|
|---|
| 146 | bool print_halts;
|
|---|
| 147 | #endif
|
|---|
| 148 |
|
|---|
| 149 | #ifdef __CFA_DEBUG__
|
|---|
| 150 | // Last function to enable preemption on this processor
|
|---|
| 151 | const char * last_enable;
|
|---|
| 152 | #endif
|
|---|
| 153 | };
|
|---|
| 154 | // P9_EMBEDDED( processor, dlink(processor) )
|
|---|
| 155 | static inline tytagref( dlink(processor), dlink(processor) ) ?`inner( processor & this ) {
|
|---|
| 156 | dlink(processor) & b = this.link;
|
|---|
| 157 | tytagref( dlink(processor), dlink(processor) ) result = { b };
|
|---|
| 158 | return result;
|
|---|
| 159 | }
|
|---|
| 160 |
|
|---|
| 161 | void ?{}(processor & this, const char name[], struct cluster & cltr);
|
|---|
| 162 | void ^?{}(processor & this);
|
|---|
| 163 |
|
|---|
| 164 | static inline void ?{}(processor & this) { this{ "Anonymous Processor", *mainCluster}; }
|
|---|
| 165 | static inline void ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
|
|---|
| 166 | static inline void ?{}(processor & this, const char name[]) { this{name, *mainCluster}; }
|
|---|
| 167 |
|
|---|
| 168 | //-----------------------------------------------------------------------------
|
|---|
| 169 | // Cluster Tools
|
|---|
| 170 |
|
|---|
| 171 | // Intrusives lanes which are used by the ready queue
|
|---|
| 172 | union __attribute__((aligned(64))) __intrusive_lane_t;
|
|---|
| 173 | void ?{}(__intrusive_lane_t & this);
|
|---|
| 174 | void ^?{}(__intrusive_lane_t & this);
|
|---|
| 175 |
|
|---|
| 176 | // Aligned timestamps which are used by the ready queue and io subsystem
|
|---|
| 177 | union __attribute__((aligned(64))) __timestamp_t;
|
|---|
| 178 |
|
|---|
| 179 | void ?{}(__timestamp_t & this);
|
|---|
| 180 | void ^?{}(__timestamp_t &);
|
|---|
| 181 |
|
|---|
| 182 |
|
|---|
| 183 | struct __attribute__((aligned(16))) __cache_id_t {
|
|---|
| 184 | volatile unsigned id;
|
|---|
| 185 | };
|
|---|
| 186 |
|
|---|
| 187 | // Idle Sleep
|
|---|
| 188 | struct __cluster_proc_list {
|
|---|
| 189 | // Spin lock protecting the queue
|
|---|
| 190 | __spinlock_t lock;
|
|---|
| 191 |
|
|---|
| 192 | // FD to use to wake a processor
|
|---|
| 193 | struct __fd_waitctx * volatile fdw;
|
|---|
| 194 |
|
|---|
| 195 | // Total number of processors
|
|---|
| 196 | unsigned total;
|
|---|
| 197 |
|
|---|
| 198 | // Number of processors constructed
|
|---|
| 199 | // incremented at construction time, total is incremented once the processor asyncronously registers
|
|---|
| 200 | unsigned constructed;
|
|---|
| 201 |
|
|---|
| 202 | // Total number of idle processors
|
|---|
| 203 | unsigned idle;
|
|---|
| 204 |
|
|---|
| 205 | // List of idle processors
|
|---|
| 206 | dlist(processor) idles;
|
|---|
| 207 |
|
|---|
| 208 | // List of active processors
|
|---|
| 209 | dlist(processor) actives;
|
|---|
| 210 | };
|
|---|
| 211 |
|
|---|
| 212 | //-----------------------------------------------------------------------------
|
|---|
| 213 | // Cluster
|
|---|
| 214 | struct __attribute__((aligned(64))) cluster {
|
|---|
| 215 | struct {
|
|---|
| 216 | struct {
|
|---|
| 217 | // Arary of subqueues
|
|---|
| 218 | __intrusive_lane_t * data;
|
|---|
| 219 |
|
|---|
| 220 | // Time since subqueues were processed
|
|---|
| 221 | __timestamp_t * tscs;
|
|---|
| 222 |
|
|---|
| 223 | // Number of subqueue / timestamps
|
|---|
| 224 | size_t count;
|
|---|
| 225 | } readyQ;
|
|---|
| 226 |
|
|---|
| 227 | struct {
|
|---|
| 228 | // Array of $io_
|
|---|
| 229 | io_context$ ** data;
|
|---|
| 230 |
|
|---|
| 231 | // Time since subqueues were processed
|
|---|
| 232 | __timestamp_t * tscs;
|
|---|
| 233 |
|
|---|
| 234 | // Number of I/O subqueues
|
|---|
| 235 | size_t count;
|
|---|
| 236 | } io;
|
|---|
| 237 |
|
|---|
| 238 | // Cache each kernel thread belongs to
|
|---|
| 239 | __cache_id_t * caches;
|
|---|
| 240 | } sched;
|
|---|
| 241 |
|
|---|
| 242 | // // Ready queue for threads
|
|---|
| 243 | // __ready_queue_t ready_queue;
|
|---|
| 244 |
|
|---|
| 245 | // Name of the cluster
|
|---|
| 246 | const char * name;
|
|---|
| 247 |
|
|---|
| 248 | // Preemption rate on this cluster
|
|---|
| 249 | Duration preemption_rate;
|
|---|
| 250 |
|
|---|
| 251 | // List of idle processors
|
|---|
| 252 | __cluster_proc_list procs;
|
|---|
| 253 |
|
|---|
| 254 | // List of threads
|
|---|
| 255 | __spinlock_t thread_list_lock;
|
|---|
| 256 | dlist(struct thread$, struct __thread_user_link) threads;
|
|---|
| 257 | unsigned int nthreads;
|
|---|
| 258 |
|
|---|
| 259 | // Link lists fields
|
|---|
| 260 | struct __dbg_node_cltr {
|
|---|
| 261 | cluster * next;
|
|---|
| 262 | cluster * prev;
|
|---|
| 263 | } node;
|
|---|
| 264 |
|
|---|
| 265 | struct {
|
|---|
| 266 | io_arbiter$ * arbiter;
|
|---|
| 267 | io_context_params params;
|
|---|
| 268 | } io;
|
|---|
| 269 |
|
|---|
| 270 | struct {
|
|---|
| 271 | struct processor ** procs;
|
|---|
| 272 | unsigned cnt;
|
|---|
| 273 | } managed;
|
|---|
| 274 |
|
|---|
| 275 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 276 | struct __stats_t * stats;
|
|---|
| 277 | int print_stats;
|
|---|
| 278 | #endif
|
|---|
| 279 | };
|
|---|
| 280 | extern Duration default_preemption();
|
|---|
| 281 |
|
|---|
| 282 | void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
|
|---|
| 283 | void ^?{}(cluster & this);
|
|---|
| 284 |
|
|---|
| 285 | static inline void ?{} (cluster & this) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
|
|---|
| 286 | static inline void ?{} (cluster & this, Duration preemption_rate) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
|
|---|
| 287 | static inline void ?{} (cluster & this, const char name[]) { io_context_params default_params; this{name, default_preemption(), 1, default_params}; }
|
|---|
| 288 | static inline void ?{} (cluster & this, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
|
|---|
| 289 | 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}; }
|
|---|
| 290 | static inline void ?{} (cluster & this, const char name[], unsigned num_io) { io_context_params default_params; this{name, default_preemption(), num_io, default_params}; }
|
|---|
| 291 | static inline void ?{} (cluster & this, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
|
|---|
| 292 | static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
|
|---|
| 293 | static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params) { this{name, default_preemption(), 1, io_params}; }
|
|---|
| 294 | static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
|
|---|
| 295 | 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}; }
|
|---|
| 296 | 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}; }
|
|---|
| 297 |
|
|---|
| 298 | static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }
|
|---|
| 299 |
|
|---|
| 300 | static inline struct processor * active_processor() { return publicTLS_get( this_processor ); } // UNSAFE
|
|---|
| 301 | static inline struct cluster * active_cluster () { return publicTLS_get( this_processor )->cltr; }
|
|---|
| 302 |
|
|---|
| 303 | // gets the number of constructed processors on the cluster
|
|---|
| 304 | static inline unsigned get_proc_count( cluster & this ) { return this.procs.constructed; }
|
|---|
| 305 |
|
|---|
| 306 | // set the number of internal processors
|
|---|
| 307 | // these processors are in addition to any explicitly declared processors
|
|---|
| 308 | unsigned set_concurrency( cluster & this, unsigned new_count );
|
|---|
| 309 |
|
|---|
| 310 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 311 | void print_stats_now( cluster & this, int flags );
|
|---|
| 312 |
|
|---|
| 313 | static inline void print_stats_at_exit( cluster & this, int flags ) {
|
|---|
| 314 | this.print_stats |= flags;
|
|---|
| 315 | }
|
|---|
| 316 |
|
|---|
| 317 | static inline void print_stats_at_exit( processor & this, int flags ) {
|
|---|
| 318 | this.print_stats |= flags;
|
|---|
| 319 | }
|
|---|
| 320 |
|
|---|
| 321 | void print_halts( processor & this );
|
|---|
| 322 | #endif
|
|---|
| 323 |
|
|---|
| 324 | // Local Variables: //
|
|---|
| 325 | // mode: c //
|
|---|
| 326 | // tab-width: 4 //
|
|---|
| 327 | // End: //
|
|---|
