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