| 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 | #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 | // Processor id, required for scheduling threads
 | 
|---|
| 51 | struct __processor_id_t {
 | 
|---|
| 52 |         unsigned id:24;
 | 
|---|
| 53 | 
 | 
|---|
| 54 |         #if !defined(__CFA_NO_STATISTICS__)
 | 
|---|
| 55 |                 struct __stats_t * stats;
 | 
|---|
| 56 |         #endif
 | 
|---|
| 57 | };
 | 
|---|
| 58 | 
 | 
|---|
| 59 | coroutine processorCtx_t {
 | 
|---|
| 60 |         struct processor * proc;
 | 
|---|
| 61 | };
 | 
|---|
| 62 | 
 | 
|---|
| 63 | // Wrapper around kernel threads
 | 
|---|
| 64 | struct __attribute__((aligned(128))) processor {
 | 
|---|
| 65 |         // Main state
 | 
|---|
| 66 |         inline __processor_id_t;
 | 
|---|
| 67 | 
 | 
|---|
| 68 |         // Cluster from which to get threads
 | 
|---|
| 69 |         struct cluster * cltr;
 | 
|---|
| 70 | 
 | 
|---|
| 71 |         // Id within the cluster
 | 
|---|
| 72 |         unsigned cltr_id;
 | 
|---|
| 73 | 
 | 
|---|
| 74 |         // Set to true to notify the processor should terminate
 | 
|---|
| 75 |         volatile bool do_terminate;
 | 
|---|
| 76 | 
 | 
|---|
| 77 |         // Coroutine ctx who does keeps the state of the processor
 | 
|---|
| 78 |         struct processorCtx_t runner;
 | 
|---|
| 79 | 
 | 
|---|
| 80 |         // Name of the processor
 | 
|---|
| 81 |         const char * name;
 | 
|---|
| 82 | 
 | 
|---|
| 83 |         // Handle to pthreads
 | 
|---|
| 84 |         pthread_t kernel_thread;
 | 
|---|
| 85 | 
 | 
|---|
| 86 |         struct {
 | 
|---|
| 87 |                 $io_context * ctx;
 | 
|---|
| 88 |                 bool pending;
 | 
|---|
| 89 |                 bool dirty;
 | 
|---|
| 90 |         } io;
 | 
|---|
| 91 | 
 | 
|---|
| 92 |         // Preemption data
 | 
|---|
| 93 |         // Node which is added in the discrete event simulaiton
 | 
|---|
| 94 |         struct alarm_node_t * preemption_alarm;
 | 
|---|
| 95 | 
 | 
|---|
| 96 |         // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
 | 
|---|
| 97 |         bool pending_preemption;
 | 
|---|
| 98 | 
 | 
|---|
| 99 |         // Idle lock (kernel semaphore)
 | 
|---|
| 100 |         int idle;
 | 
|---|
| 101 | 
 | 
|---|
| 102 |         // Termination synchronisation (user semaphore)
 | 
|---|
| 103 |         oneshot terminated;
 | 
|---|
| 104 | 
 | 
|---|
| 105 |         // pthread Stack
 | 
|---|
| 106 |         void * stack;
 | 
|---|
| 107 | 
 | 
|---|
| 108 |         // Link lists fields
 | 
|---|
| 109 |         DLISTED_MGD_IMPL_IN(processor)
 | 
|---|
| 110 | 
 | 
|---|
| 111 |         // special init fields
 | 
|---|
| 112 |         // This is needed for memcached integration
 | 
|---|
| 113 |         // once memcached experiments are done this should probably be removed
 | 
|---|
| 114 |         // it is not a particularly safe scheme as it can make processors less homogeneous
 | 
|---|
| 115 |         struct {
 | 
|---|
| 116 |                 $thread * thrd;
 | 
|---|
| 117 |         } init;
 | 
|---|
| 118 | 
 | 
|---|
| 119 |         #if !defined(__CFA_NO_STATISTICS__)
 | 
|---|
| 120 |                 int print_stats;
 | 
|---|
| 121 |                 bool print_halts;
 | 
|---|
| 122 |         #endif
 | 
|---|
| 123 | 
 | 
|---|
| 124 | #ifdef __CFA_DEBUG__
 | 
|---|
| 125 |         // Last function to enable preemption on this processor
 | 
|---|
| 126 |         const char * last_enable;
 | 
|---|
| 127 | #endif
 | 
|---|
| 128 | };
 | 
|---|
| 129 | 
 | 
|---|
| 130 | void  ?{}(processor & this, const char name[], struct cluster & cltr);
 | 
|---|
| 131 | void ^?{}(processor & this);
 | 
|---|
| 132 | 
 | 
|---|
| 133 | static inline void  ?{}(processor & this)                        { this{ "Anonymous Processor", *mainCluster}; }
 | 
|---|
| 134 | static inline void  ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
 | 
|---|
| 135 | static inline void  ?{}(processor & this, const char name[])     { this{name, *mainCluster}; }
 | 
|---|
| 136 | 
 | 
|---|
| 137 | DLISTED_MGD_IMPL_OUT(processor)
 | 
|---|
| 138 | 
 | 
|---|
| 139 | //-----------------------------------------------------------------------------
 | 
|---|
| 140 | // Cluster Tools
 | 
|---|
| 141 | 
 | 
|---|
| 142 | // Intrusives lanes which are used by the relaxed ready queue
 | 
|---|
| 143 | struct __attribute__((aligned(128))) __intrusive_lane_t;
 | 
|---|
| 144 | void  ?{}(__intrusive_lane_t & this);
 | 
|---|
| 145 | void ^?{}(__intrusive_lane_t & this);
 | 
|---|
| 146 | 
 | 
|---|
| 147 | // Counter used for wether or not the lanes are all empty
 | 
|---|
| 148 | struct __attribute__((aligned(128))) __snzi_node_t;
 | 
|---|
| 149 | struct __snzi_t {
 | 
|---|
| 150 |         unsigned mask;
 | 
|---|
| 151 |         int root;
 | 
|---|
| 152 |         __snzi_node_t * nodes;
 | 
|---|
| 153 | };
 | 
|---|
| 154 | 
 | 
|---|
| 155 | void  ?{}( __snzi_t & this, unsigned depth );
 | 
|---|
| 156 | void ^?{}( __snzi_t & this );
 | 
|---|
| 157 | 
 | 
|---|
| 158 | //TODO adjust cache size to ARCHITECTURE
 | 
|---|
| 159 | // Structure holding the relaxed ready queue
 | 
|---|
| 160 | struct __ready_queue_t {
 | 
|---|
| 161 |         // Data tracking how many/which lanes are used
 | 
|---|
| 162 |         // Aligned to 128 for cache locality
 | 
|---|
| 163 |         __snzi_t snzi;
 | 
|---|
| 164 | 
 | 
|---|
| 165 |         // Data tracking the actual lanes
 | 
|---|
| 166 |         // On a seperate cacheline from the used struct since
 | 
|---|
| 167 |         // used can change on each push/pop but this data
 | 
|---|
| 168 |         // only changes on shrink/grow
 | 
|---|
| 169 |         struct {
 | 
|---|
| 170 |                 // Arary of lanes
 | 
|---|
| 171 |                 __intrusive_lane_t * volatile data;
 | 
|---|
| 172 | 
 | 
|---|
| 173 |                 // Number of lanes (empty or not)
 | 
|---|
| 174 |                 volatile size_t count;
 | 
|---|
| 175 |         } lanes;
 | 
|---|
| 176 | };
 | 
|---|
| 177 | 
 | 
|---|
| 178 | void  ?{}(__ready_queue_t & this);
 | 
|---|
| 179 | void ^?{}(__ready_queue_t & this);
 | 
|---|
| 180 | 
 | 
|---|
| 181 | // Idle Sleep
 | 
|---|
| 182 | struct __cluster_idles {
 | 
|---|
| 183 |         // Spin lock protecting the queue
 | 
|---|
| 184 |         volatile uint64_t lock;
 | 
|---|
| 185 | 
 | 
|---|
| 186 |         // Total number of processors
 | 
|---|
| 187 |         unsigned total;
 | 
|---|
| 188 | 
 | 
|---|
| 189 |         // Total number of idle processors
 | 
|---|
| 190 |         unsigned idle;
 | 
|---|
| 191 | 
 | 
|---|
| 192 |         // List of idle processors
 | 
|---|
| 193 |         dlist(processor, processor) list;
 | 
|---|
| 194 | };
 | 
|---|
| 195 | 
 | 
|---|
| 196 | //-----------------------------------------------------------------------------
 | 
|---|
| 197 | // Cluster
 | 
|---|
| 198 | struct __attribute__((aligned(128))) cluster {
 | 
|---|
| 199 |         // Ready queue for threads
 | 
|---|
| 200 |         __ready_queue_t ready_queue;
 | 
|---|
| 201 | 
 | 
|---|
| 202 |         // Name of the cluster
 | 
|---|
| 203 |         const char * name;
 | 
|---|
| 204 | 
 | 
|---|
| 205 |         // Preemption rate on this cluster
 | 
|---|
| 206 |         Duration preemption_rate;
 | 
|---|
| 207 | 
 | 
|---|
| 208 |         // List of idle processors
 | 
|---|
| 209 |         __cluster_idles idles;
 | 
|---|
| 210 | 
 | 
|---|
| 211 |         // List of threads
 | 
|---|
| 212 |         __spinlock_t thread_list_lock;
 | 
|---|
| 213 |         __dllist_t(struct $thread) threads;
 | 
|---|
| 214 |         unsigned int nthreads;
 | 
|---|
| 215 | 
 | 
|---|
| 216 |         // Link lists fields
 | 
|---|
| 217 |         struct __dbg_node_cltr {
 | 
|---|
| 218 |                 cluster * next;
 | 
|---|
| 219 |                 cluster * prev;
 | 
|---|
| 220 |         } node;
 | 
|---|
| 221 | 
 | 
|---|
| 222 |         struct {
 | 
|---|
| 223 |                 $io_arbiter * arbiter;
 | 
|---|
| 224 |                 io_context_params params;
 | 
|---|
| 225 |         } io;
 | 
|---|
| 226 | 
 | 
|---|
| 227 |         #if !defined(__CFA_NO_STATISTICS__)
 | 
|---|
| 228 |                 struct __stats_t * stats;
 | 
|---|
| 229 |                 int print_stats;
 | 
|---|
| 230 |         #endif
 | 
|---|
| 231 | };
 | 
|---|
| 232 | extern Duration default_preemption();
 | 
|---|
| 233 | 
 | 
|---|
| 234 | void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
 | 
|---|
| 235 | void ^?{}(cluster & this);
 | 
|---|
| 236 | 
 | 
|---|
| 237 | static inline void ?{} (cluster & this)                                            { io_context_params default_params;    this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
 | 
|---|
| 238 | static inline void ?{} (cluster & this, Duration preemption_rate)                  { io_context_params default_params;    this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
 | 
|---|
| 239 | static inline void ?{} (cluster & this, const char name[])                         { io_context_params default_params;    this{name, default_preemption(), 1, default_params}; }
 | 
|---|
| 240 | static inline void ?{} (cluster & this, unsigned num_io)                           { io_context_params default_params;    this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
 | 
|---|
| 241 | 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}; }
 | 
|---|
| 242 | static inline void ?{} (cluster & this, const char name[], unsigned num_io)        { io_context_params default_params;    this{name, default_preemption(), num_io, default_params}; }
 | 
|---|
| 243 | static inline void ?{} (cluster & this, const io_context_params & io_params)                                            { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
 | 
|---|
| 244 | static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params)                  { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
 | 
|---|
| 245 | static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params)                         { this{name, default_preemption(), 1, io_params}; }
 | 
|---|
| 246 | static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params)                           { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
 | 
|---|
| 247 | 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}; }
 | 
|---|
| 248 | 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}; }
 | 
|---|
| 249 | 
 | 
|---|
| 250 | static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }
 | 
|---|
| 251 | 
 | 
|---|
| 252 | static inline struct processor * active_processor() { return publicTLS_get( this_processor ); } // UNSAFE
 | 
|---|
| 253 | static inline struct cluster   * active_cluster  () { return publicTLS_get( this_processor )->cltr; }
 | 
|---|
| 254 | 
 | 
|---|
| 255 | #if !defined(__CFA_NO_STATISTICS__)
 | 
|---|
| 256 |         void print_stats_now( cluster & this, int flags );
 | 
|---|
| 257 | 
 | 
|---|
| 258 |         static inline void print_stats_at_exit( cluster & this, int flags ) {
 | 
|---|
| 259 |                 this.print_stats |= flags;
 | 
|---|
| 260 |         }
 | 
|---|
| 261 | 
 | 
|---|
| 262 |         static inline void print_stats_at_exit( processor & this, int flags ) {
 | 
|---|
| 263 |                 this.print_stats |= flags;
 | 
|---|
| 264 |         }
 | 
|---|
| 265 | 
 | 
|---|
| 266 |         void print_halts( processor & this );
 | 
|---|
| 267 | #endif
 | 
|---|
| 268 | 
 | 
|---|
| 269 | // Local Variables: //
 | 
|---|
| 270 | // mode: c //
 | 
|---|
| 271 | // tab-width: 4 //
 | 
|---|
| 272 | // End: //
 | 
|---|