source: libcfa/src/concurrency/kernel.hfa@ c655650

ADT ast-experimental enum pthread-emulation qualifiedEnum
Last change on this file since c655650 was 7cf3b1d, checked in by Thierry Delisle <tdelisle@…>, 4 years ago

Added level of indirection to idle sleeps which helps statistics.
  • Property mode set to 100644
File size: 9.3 KB
Line 
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
24extern "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
36struct cluster;
37struct $io_context;
38struct $io_arbiter;
39
40struct io_context_params {
41 int num_entries;
42};
43
44void ?{}(io_context_params & this);
45
46//-----------------------------------------------------------------------------
47// Processor
48extern struct cluster * mainCluster;
49
50// Processor id, required for scheduling threads
51
52
53coroutine processorCtx_t {
54 struct processor * proc;
55};
56
57
58struct __fd_waitctx {
59 volatile int fd;
60};
61
62// Wrapper around kernel threads
63struct __attribute__((aligned(128))) processor {
64 // Cluster from which to get threads
65 struct cluster * cltr;
66
67 // Ready Queue state per processor
68 struct {
69 unsigned short its;
70 unsigned short itr;
71 unsigned id;
72 unsigned target;
73 unsigned last;
74 signed cpu;
75 } rdq;
76
77 // Set to true to notify the processor should terminate
78 volatile bool do_terminate;
79
80 // Coroutine ctx who does keeps the state of the processor
81 struct processorCtx_t runner;
82
83 // Name of the processor
84 const char * name;
85
86 // Handle to pthreads
87 pthread_t kernel_thread;
88
89 // Unique id for the processor (not per cluster)
90 unsigned unique_id;
91
92 struct {
93 $io_context * ctx;
94 bool pending;
95 bool dirty;
96 } io;
97
98 // Preemption data
99 // Node which is added in the discrete event simulaiton
100 struct alarm_node_t * preemption_alarm;
101
102 // If true, a preemption was triggered in an unsafe region, the processor must preempt as soon as possible
103 bool pending_preemption;
104
105 // Idle lock (kernel semaphore)
106 int idle_fd;
107
108 // Idle waitctx
109 struct __fd_waitctx idle_wctx;
110
111 // Termination synchronisation (user semaphore)
112 oneshot terminated;
113
114 // pthread Stack
115 void * stack;
116
117 // Link lists fields
118 inline dlink(processor);
119
120 // special init fields
121 // This is needed for memcached integration
122 // once memcached experiments are done this should probably be removed
123 // it is not a particularly safe scheme as it can make processors less homogeneous
124 struct {
125 thread$ * thrd;
126 } init;
127
128 struct KernelThreadData * local_data;
129
130 #if !defined(__CFA_NO_STATISTICS__)
131 int print_stats;
132 bool print_halts;
133 #endif
134
135#ifdef __CFA_DEBUG__
136 // Last function to enable preemption on this processor
137 const char * last_enable;
138#endif
139};
140P9_EMBEDDED( processor, dlink(processor) )
141
142void ?{}(processor & this, const char name[], struct cluster & cltr);
143void ^?{}(processor & this);
144
145static inline void ?{}(processor & this) { this{ "Anonymous Processor", *mainCluster}; }
146static inline void ?{}(processor & this, struct cluster & cltr) { this{ "Anonymous Processor", cltr}; }
147static inline void ?{}(processor & this, const char name[]) { this{name, *mainCluster}; }
148
149//-----------------------------------------------------------------------------
150// Cluster Tools
151
152// Intrusives lanes which are used by the ready queue
153struct __attribute__((aligned(128))) __intrusive_lane_t;
154void ?{}(__intrusive_lane_t & this);
155void ^?{}(__intrusive_lane_t & this);
156
157// Aligned timestamps which are used by the relaxed ready queue
158struct __attribute__((aligned(128))) __timestamp_t {
159 volatile unsigned long long tv;
160 volatile unsigned long long ma;
161};
162
163struct __attribute__((aligned(16))) __cache_id_t {
164 volatile unsigned id;
165};
166
167// Aligned timestamps which are used by the relaxed ready queue
168struct __attribute__((aligned(128))) __help_cnts_t {
169 volatile unsigned long long src;
170 volatile unsigned long long dst;
171 volatile unsigned long long tri;
172};
173
174static inline void ?{}(__timestamp_t & this) { this.tv = 0; this.ma = 0; }
175static inline void ^?{}(__timestamp_t & this) {}
176
177struct __attribute__((aligned(128))) __ready_queue_caches_t;
178void ?{}(__ready_queue_caches_t & this);
179void ^?{}(__ready_queue_caches_t & this);
180
181//TODO adjust cache size to ARCHITECTURE
182// Structure holding the ready queue
183struct __ready_queue_t {
184 // Data tracking the actual lanes
185 // On a seperate cacheline from the used struct since
186 // used can change on each push/pop but this data
187 // only changes on shrink/grow
188 struct {
189 // Arary of lanes
190 __intrusive_lane_t * volatile data;
191
192 // Array of times
193 __timestamp_t * volatile tscs;
194
195 __cache_id_t * volatile caches;
196
197 // Array of stats
198 __help_cnts_t * volatile help;
199
200 // Number of lanes (empty or not)
201 volatile size_t count;
202 } lanes;
203};
204
205void ?{}(__ready_queue_t & this);
206void ^?{}(__ready_queue_t & this);
207#if !defined(__CFA_NO_STATISTICS__)
208 unsigned cnt(const __ready_queue_t & this, unsigned idx);
209#endif
210
211// Idle Sleep
212struct __cluster_proc_list {
213 // Spin lock protecting the queue
214 __spinlock_t lock;
215
216 // FD to use to wake a processor
217 struct __fd_waitctx * volatile fdw;
218
219 // Total number of processors
220 unsigned total;
221
222 // Total number of idle processors
223 unsigned idle;
224
225 // List of idle processors
226 dlist(processor) idles;
227
228 // List of active processors
229 dlist(processor) actives;
230};
231
232//-----------------------------------------------------------------------------
233// Cluster
234struct __attribute__((aligned(128))) cluster {
235 // Ready queue for threads
236 __ready_queue_t ready_queue;
237
238 // Name of the cluster
239 const char * name;
240
241 // Preemption rate on this cluster
242 Duration preemption_rate;
243
244 // List of idle processors
245 __cluster_proc_list procs;
246
247 // List of threads
248 __spinlock_t thread_list_lock;
249 __dllist_t(struct thread$) threads;
250 unsigned int nthreads;
251
252 // Link lists fields
253 struct __dbg_node_cltr {
254 cluster * next;
255 cluster * prev;
256 } node;
257
258 struct {
259 $io_arbiter * arbiter;
260 io_context_params params;
261 } io;
262
263 #if !defined(__CFA_NO_STATISTICS__)
264 struct __stats_t * stats;
265 int print_stats;
266 #endif
267};
268extern Duration default_preemption();
269
270void ?{} (cluster & this, const char name[], Duration preemption_rate, unsigned num_io, const io_context_params & io_params);
271void ^?{}(cluster & this);
272
273static inline void ?{} (cluster & this) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), 1, default_params}; }
274static inline void ?{} (cluster & this, Duration preemption_rate) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, 1, default_params}; }
275static inline void ?{} (cluster & this, const char name[]) { io_context_params default_params; this{name, default_preemption(), 1, default_params}; }
276static inline void ?{} (cluster & this, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", default_preemption(), num_io, default_params}; }
277static inline void ?{} (cluster & this, Duration preemption_rate, unsigned num_io) { io_context_params default_params; this{"Anonymous Cluster", preemption_rate, num_io, default_params}; }
278static inline void ?{} (cluster & this, const char name[], unsigned num_io) { io_context_params default_params; this{name, default_preemption(), num_io, default_params}; }
279static inline void ?{} (cluster & this, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), 1, io_params}; }
280static inline void ?{} (cluster & this, Duration preemption_rate, const io_context_params & io_params) { this{"Anonymous Cluster", preemption_rate, 1, io_params}; }
281static inline void ?{} (cluster & this, const char name[], const io_context_params & io_params) { this{name, default_preemption(), 1, io_params}; }
282static inline void ?{} (cluster & this, unsigned num_io, const io_context_params & io_params) { this{"Anonymous Cluster", default_preemption(), num_io, io_params}; }
283static 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}; }
284static inline void ?{} (cluster & this, const char name[], unsigned num_io, const io_context_params & io_params) { this{name, default_preemption(), num_io, io_params}; }
285
286static inline [cluster *&, cluster *& ] __get( cluster & this ) __attribute__((const)) { return this.node.[next, prev]; }
287
288static inline struct processor * active_processor() { return publicTLS_get( this_processor ); } // UNSAFE
289static inline struct cluster * active_cluster () { return publicTLS_get( this_processor )->cltr; }
290
291#if !defined(__CFA_NO_STATISTICS__)
292 void print_stats_now( cluster & this, int flags );
293
294 static inline void print_stats_at_exit( cluster & this, int flags ) {
295 this.print_stats |= flags;
296 }
297
298 static inline void print_stats_at_exit( processor & this, int flags ) {
299 this.print_stats |= flags;
300 }
301
302 void print_halts( processor & this );
303#endif
304
305// Local Variables: //
306// mode: c //
307// tab-width: 4 //
308// End: //
Note: See TracBrowser for help on using the repository browser.