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