| 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 : 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/list.hfa"
|
|---|
| 23 |
|
|---|
| 24 | extern "C" {
|
|---|
| 25 | #include <bits/pthreadtypes.h>
|
|---|
| 26 | #include <pthread.h>
|
|---|
| 27 | #include <linux/types.h>
|
|---|
| 28 | }
|
|---|
| 29 |
|
|---|
| 30 | //-----------------------------------------------------------------------------
|
|---|
| 31 | // Underlying Locks
|
|---|
| 32 | #ifdef __CFA_WITH_VERIFY__
|
|---|
| 33 | extern bool __cfaabi_dbg_in_kernel();
|
|---|
| 34 | #endif
|
|---|
| 35 |
|
|---|
| 36 | extern "C" {
|
|---|
| 37 | char * strerror(int);
|
|---|
| 38 | }
|
|---|
| 39 | #define CHECKED(x) { int err = x; if( err != 0 ) abort("KERNEL ERROR: Operation \"" #x "\" return error %d - %s\n", err, strerror(err)); }
|
|---|
| 40 |
|
|---|
| 41 | struct __bin_sem_t {
|
|---|
| 42 | pthread_mutex_t lock;
|
|---|
| 43 | pthread_cond_t cond;
|
|---|
| 44 | int val;
|
|---|
| 45 | };
|
|---|
| 46 |
|
|---|
| 47 | static inline void ?{}(__bin_sem_t & this) with( this ) {
|
|---|
| 48 | // Create the mutex with error checking
|
|---|
| 49 | pthread_mutexattr_t mattr;
|
|---|
| 50 | pthread_mutexattr_init( &mattr );
|
|---|
| 51 | pthread_mutexattr_settype( &mattr, PTHREAD_MUTEX_ERRORCHECK_NP);
|
|---|
| 52 | pthread_mutex_init(&lock, &mattr);
|
|---|
| 53 |
|
|---|
| 54 | pthread_cond_init (&cond, (const pthread_condattr_t *)0p); // workaround trac#208: cast should not be required
|
|---|
| 55 | val = 0;
|
|---|
| 56 | }
|
|---|
| 57 |
|
|---|
| 58 | static inline void ^?{}(__bin_sem_t & this) with( this ) {
|
|---|
| 59 | CHECKED( pthread_mutex_destroy(&lock) );
|
|---|
| 60 | CHECKED( pthread_cond_destroy (&cond) );
|
|---|
| 61 | }
|
|---|
| 62 |
|
|---|
| 63 | static inline void wait(__bin_sem_t & this) with( this ) {
|
|---|
| 64 | verify(__cfaabi_dbg_in_kernel());
|
|---|
| 65 | CHECKED( pthread_mutex_lock(&lock) );
|
|---|
| 66 | while(val < 1) {
|
|---|
| 67 | pthread_cond_wait(&cond, &lock);
|
|---|
| 68 | }
|
|---|
| 69 | val -= 1;
|
|---|
| 70 | CHECKED( pthread_mutex_unlock(&lock) );
|
|---|
| 71 | }
|
|---|
| 72 |
|
|---|
| 73 | static inline bool post(__bin_sem_t & this) with( this ) {
|
|---|
| 74 | bool needs_signal = false;
|
|---|
| 75 |
|
|---|
| 76 | CHECKED( pthread_mutex_lock(&lock) );
|
|---|
| 77 | if(val < 1) {
|
|---|
| 78 | val += 1;
|
|---|
| 79 | pthread_cond_signal(&cond);
|
|---|
| 80 | needs_signal = true;
|
|---|
| 81 | }
|
|---|
| 82 | CHECKED( pthread_mutex_unlock(&lock) );
|
|---|
| 83 |
|
|---|
| 84 | return needs_signal;
|
|---|
| 85 | }
|
|---|
| 86 |
|
|---|
| 87 | #undef CHECKED
|
|---|
| 88 |
|
|---|
| 89 |
|
|---|
| 90 | //-----------------------------------------------------------------------------
|
|---|
| 91 | // Processor
|
|---|
| 92 | extern struct cluster * mainCluster;
|
|---|
| 93 |
|
|---|
| 94 | // Processor id, required for scheduling threads
|
|---|
| 95 | struct __processor_id_t {
|
|---|
| 96 | unsigned id:24;
|
|---|
| 97 | bool full_proc:1;
|
|---|
| 98 |
|
|---|
| 99 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 100 | struct __stats_t * stats;
|
|---|
| 101 | #endif
|
|---|
| 102 | };
|
|---|
| 103 |
|
|---|
| 104 | coroutine processorCtx_t {
|
|---|
| 105 | struct processor * proc;
|
|---|
| 106 | };
|
|---|
| 107 |
|
|---|
| 108 | // Wrapper around kernel threads
|
|---|
| 109 | struct __attribute__((aligned(128))) processor {
|
|---|
| 110 | // Main state
|
|---|
| 111 | inline __processor_id_t;
|
|---|
| 112 |
|
|---|
| 113 | // Cluster from which to get threads
|
|---|
| 114 | struct cluster * cltr;
|
|---|
| 115 |
|
|---|
| 116 | // Set to true to notify the processor should terminate
|
|---|
| 117 | volatile bool do_terminate;
|
|---|
| 118 |
|
|---|
| 119 | // Coroutine ctx who does keeps the state of the processor
|
|---|
| 120 | struct processorCtx_t runner;
|
|---|
| 121 |
|
|---|
| 122 | // Name of the processor
|
|---|
| 123 | const char * name;
|
|---|
| 124 |
|
|---|
| 125 | // Handle to pthreads
|
|---|
| 126 | pthread_t kernel_thread;
|
|---|
| 127 |
|
|---|
| 128 | // Preemption data
|
|---|
| 129 | // Node which is added in the discrete event simulaiton
|
|---|
| 130 | struct alarm_node_t * preemption_alarm;
|
|---|
| 131 |
|
|---|
| 132 | // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
|
|---|
| 133 | bool pending_preemption;
|
|---|
| 134 |
|
|---|
| 135 | // Idle lock (kernel semaphore)
|
|---|
| 136 | __bin_sem_t idle;
|
|---|
| 137 |
|
|---|
| 138 | // Termination synchronisation (user semaphore)
|
|---|
| 139 | oneshot terminated;
|
|---|
| 140 |
|
|---|
| 141 | // pthread Stack
|
|---|
| 142 | void * stack;
|
|---|
| 143 |
|
|---|
| 144 | // Link lists fields
|
|---|
| 145 | DLISTED_MGD_IMPL_IN(processor)
|
|---|
| 146 |
|
|---|
| 147 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 148 | int print_stats;
|
|---|
| 149 | bool print_halts;
|
|---|
| 150 | #endif
|
|---|
| 151 |
|
|---|
| 152 | #ifdef __CFA_DEBUG__
|
|---|
| 153 | // Last function to enable preemption on this processor
|
|---|
| 154 | const char * last_enable;
|
|---|
| 155 | #endif
|
|---|
| 156 | };
|
|---|
| 157 |
|
|---|
| 158 | void ?{}(processor & this, const char name[], struct cluster & cltr);
|
|---|
| 159 | void ^?{}(processor & this);
|
|---|
| 160 |
|
|---|
| 161 | static inline void ?{}(processor & this) { this{ "Anonymous Processor", *mainCluster}; }
|
|---|
| 162 | static inline void ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
|
|---|
| 163 | static inline void ?{}(processor & this, const char name[]) { this{name, *mainCluster }; }
|
|---|
| 164 |
|
|---|
| 165 | DLISTED_MGD_IMPL_OUT(processor)
|
|---|
| 166 |
|
|---|
| 167 | //-----------------------------------------------------------------------------
|
|---|
| 168 | // I/O
|
|---|
| 169 | struct __io_data;
|
|---|
| 170 |
|
|---|
| 171 | // IO poller user-thread
|
|---|
| 172 | // Not using the "thread" keyword because we want to control
|
|---|
| 173 | // more carefully when to start/stop it
|
|---|
| 174 | struct $io_ctx_thread {
|
|---|
| 175 | struct __io_data * ring;
|
|---|
| 176 | single_sem sem;
|
|---|
| 177 | volatile bool done;
|
|---|
| 178 | $thread self;
|
|---|
| 179 | };
|
|---|
| 180 |
|
|---|
| 181 |
|
|---|
| 182 | struct io_context {
|
|---|
| 183 | $io_ctx_thread thrd;
|
|---|
| 184 | };
|
|---|
| 185 |
|
|---|
| 186 | struct io_context_params {
|
|---|
| 187 | int num_entries;
|
|---|
| 188 | int num_ready;
|
|---|
| 189 | int submit_aff;
|
|---|
| 190 | bool eager_submits:1;
|
|---|
| 191 | bool poller_submits:1;
|
|---|
| 192 | bool poll_submit:1;
|
|---|
| 193 | bool poll_complete:1;
|
|---|
| 194 | };
|
|---|
| 195 |
|
|---|
| 196 | void ?{}(io_context_params & this);
|
|---|
| 197 |
|
|---|
| 198 | void ?{}(io_context & this, struct cluster & cl);
|
|---|
| 199 | void ?{}(io_context & this, struct cluster & cl, const io_context_params & params);
|
|---|
| 200 | void ^?{}(io_context & this);
|
|---|
| 201 |
|
|---|
| 202 | struct io_cancellation {
|
|---|
| 203 | __u64 target;
|
|---|
| 204 | };
|
|---|
| 205 |
|
|---|
| 206 | static inline void ?{}(io_cancellation & this) { this.target = -1u; }
|
|---|
| 207 | static inline void ^?{}(io_cancellation &) {}
|
|---|
| 208 | bool cancel(io_cancellation & this);
|
|---|
| 209 |
|
|---|
| 210 | //-----------------------------------------------------------------------------
|
|---|
| 211 | // Cluster Tools
|
|---|
| 212 |
|
|---|
| 213 | // Intrusives lanes which are used by the relaxed ready queue
|
|---|
| 214 | struct __attribute__((aligned(128))) __intrusive_lane_t;
|
|---|
| 215 | void ?{}(__intrusive_lane_t & this);
|
|---|
| 216 | void ^?{}(__intrusive_lane_t & this);
|
|---|
| 217 |
|
|---|
| 218 | // Counter used for wether or not the lanes are all empty
|
|---|
| 219 | struct __attribute__((aligned(128))) __snzi_node_t;
|
|---|
| 220 | struct __snzi_t {
|
|---|
| 221 | unsigned mask;
|
|---|
| 222 | int root;
|
|---|
| 223 | __snzi_node_t * nodes;
|
|---|
| 224 | };
|
|---|
| 225 |
|
|---|
| 226 | void ?{}( __snzi_t & this, unsigned depth );
|
|---|
| 227 | void ^?{}( __snzi_t & this );
|
|---|
| 228 |
|
|---|
| 229 | //TODO adjust cache size to ARCHITECTURE
|
|---|
| 230 | // Structure holding the relaxed ready queue
|
|---|
| 231 | struct __ready_queue_t {
|
|---|
| 232 | // Data tracking how many/which lanes are used
|
|---|
| 233 | // Aligned to 128 for cache locality
|
|---|
| 234 | __snzi_t snzi;
|
|---|
| 235 |
|
|---|
| 236 | // Data tracking the actual lanes
|
|---|
| 237 | // On a seperate cacheline from the used struct since
|
|---|
| 238 | // used can change on each push/pop but this data
|
|---|
| 239 | // only changes on shrink/grow
|
|---|
| 240 | struct {
|
|---|
| 241 | // Arary of lanes
|
|---|
| 242 | __intrusive_lane_t * volatile data;
|
|---|
| 243 |
|
|---|
| 244 | // Number of lanes (empty or not)
|
|---|
| 245 | volatile size_t count;
|
|---|
| 246 | } lanes;
|
|---|
| 247 | };
|
|---|
| 248 |
|
|---|
| 249 | void ?{}(__ready_queue_t & this);
|
|---|
| 250 | void ^?{}(__ready_queue_t & this);
|
|---|
| 251 |
|
|---|
| 252 | // Idle Sleep
|
|---|
| 253 | struct __cluster_idles {
|
|---|
| 254 | // Spin lock protecting the queue
|
|---|
| 255 | volatile uint64_t lock;
|
|---|
| 256 |
|
|---|
| 257 | // Total number of processors
|
|---|
| 258 | unsigned total;
|
|---|
| 259 |
|
|---|
| 260 | // Total number of idle processors
|
|---|
| 261 | unsigned idle;
|
|---|
| 262 |
|
|---|
| 263 | // List of idle processors
|
|---|
| 264 | dlist(processor, processor) list;
|
|---|
| 265 | };
|
|---|
| 266 |
|
|---|
| 267 | //-----------------------------------------------------------------------------
|
|---|
| 268 | // Cluster
|
|---|
| 269 | struct __attribute__((aligned(128))) cluster {
|
|---|
| 270 | // Ready queue for threads
|
|---|
| 271 | __ready_queue_t ready_queue;
|
|---|
| 272 |
|
|---|
| 273 | // Name of the cluster
|
|---|
| 274 | const char * name;
|
|---|
| 275 |
|
|---|
| 276 | // Preemption rate on this cluster
|
|---|
| 277 | Duration preemption_rate;
|
|---|
| 278 |
|
|---|
| 279 | // List of idle processors
|
|---|
| 280 | __cluster_idles idles;
|
|---|
| 281 |
|
|---|
| 282 | // List of threads
|
|---|
| 283 | __spinlock_t thread_list_lock;
|
|---|
| 284 | __dllist_t(struct $thread) threads;
|
|---|
| 285 | unsigned int nthreads;
|
|---|
| 286 |
|
|---|
| 287 | // Link lists fields
|
|---|
| 288 | struct __dbg_node_cltr {
|
|---|
| 289 | cluster * next;
|
|---|
| 290 | cluster * prev;
|
|---|
| 291 | } node;
|
|---|
| 292 |
|
|---|
| 293 | struct {
|
|---|
| 294 | io_context * ctxs;
|
|---|
| 295 | unsigned cnt;
|
|---|
| 296 | } io;
|
|---|
| 297 |
|
|---|
| 298 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 299 | struct __stats_t * stats;
|
|---|
| 300 | int print_stats;
|
|---|
| 301 | #endif
|
|---|
| 302 | };
|
|---|
| 303 | extern Duration default_preemption();
|
|---|
| 304 |
|
|---|
| 305 | void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
|
|---|
| 306 | void ^?{}(cluster & this);
|
|---|
| 307 |
|
|---|
| 308 | static inline void ?{} (cluster & this) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
|
|---|
| 309 | static inline void ?{} (cluster & this, Duration preemption_rate) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
|
|---|
| 310 | static inline void ?{} (cluster & this, const char name[]) { io_context_params default_params; this{name, default_preemption(), 1, default_params}; }
|
|---|
| 311 | static inline void ?{} (cluster & this, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
|
|---|
| 312 | 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}; }
|
|---|
| 313 | static inline void ?{} (cluster & this, const char name[], unsigned num_io) { io_context_params default_params; this{name, default_preemption(), num_io, default_params}; }
|
|---|
| 314 | static inline void ?{} (cluster & this, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
|
|---|
| 315 | static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
|
|---|
| 316 | static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params) { this{name, default_preemption(), 1, io_params}; }
|
|---|
| 317 | static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
|
|---|
| 318 | 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}; }
|
|---|
| 319 | 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}; }
|
|---|
| 320 |
|
|---|
| 321 | static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }
|
|---|
| 322 |
|
|---|
| 323 | static inline struct processor * active_processor() { return publicTLS_get( this_processor ); } // UNSAFE
|
|---|
| 324 | static inline struct cluster * active_cluster () { return publicTLS_get( this_processor )->cltr; }
|
|---|
| 325 |
|
|---|
| 326 | #if !defined(__CFA_NO_STATISTICS__)
|
|---|
| 327 | void print_stats_now( cluster & this, int flags );
|
|---|
| 328 |
|
|---|
| 329 | static inline void print_stats_at_exit( cluster & this, int flags ) {
|
|---|
| 330 | this.print_stats |= flags;
|
|---|
| 331 | }
|
|---|
| 332 |
|
|---|
| 333 | static inline void print_stats_at_exit( processor & this, int flags ) {
|
|---|
| 334 | this.print_stats |= flags;
|
|---|
| 335 | }
|
|---|
| 336 |
|
|---|
| 337 | void print_halts( processor & this );
|
|---|
| 338 | #endif
|
|---|
| 339 |
|
|---|
| 340 | // Local Variables: //
|
|---|
| 341 | // mode: c //
|
|---|
| 342 | // tab-width: 4 //
|
|---|
| 343 | // End: //
|
|---|