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_private.hfa -- |
---|
8 | // |
---|
9 | // Author : Thierry Delisle |
---|
10 | // Created On : Mon Feb 13 12:27:26 2017 |
---|
11 | // Last Modified By : Peter A. Buhr |
---|
12 | // Last Modified On : Sat Nov 30 19:25:02 2019 |
---|
13 | // Update Count : 8 |
---|
14 | // |
---|
15 | |
---|
16 | #pragma once |
---|
17 | |
---|
18 | #include "kernel.hfa" |
---|
19 | #include "thread.hfa" |
---|
20 | |
---|
21 | #include "alarm.hfa" |
---|
22 | #include "stats.hfa" |
---|
23 | |
---|
24 | //----------------------------------------------------------------------------- |
---|
25 | // Scheduler |
---|
26 | |
---|
27 | struct __attribute__((aligned(128))) __scheduler_lock_id_t; |
---|
28 | |
---|
29 | extern "C" { |
---|
30 | void disable_interrupts() OPTIONAL_THREAD; |
---|
31 | void enable_interrupts_noPoll(); |
---|
32 | void enable_interrupts( __cfaabi_dbg_ctx_param ); |
---|
33 | } |
---|
34 | |
---|
35 | void __schedule_thread( struct __processor_id_t *, $thread * ) |
---|
36 | #if defined(NDEBUG) || (!defined(__CFA_DEBUG__) && !defined(__CFA_VERIFY__)) |
---|
37 | __attribute__((nonnull (2))) |
---|
38 | #endif |
---|
39 | ; |
---|
40 | |
---|
41 | //Block current thread and release/wake-up the following resources |
---|
42 | void __leave_thread() __attribute__((noreturn)); |
---|
43 | |
---|
44 | //----------------------------------------------------------------------------- |
---|
45 | // Processor |
---|
46 | void main(processorCtx_t *); |
---|
47 | |
---|
48 | void * __create_pthread( pthread_t *, void * (*)(void *), void * ); |
---|
49 | |
---|
50 | |
---|
51 | |
---|
52 | extern cluster * mainCluster; |
---|
53 | |
---|
54 | //----------------------------------------------------------------------------- |
---|
55 | // Threads |
---|
56 | extern "C" { |
---|
57 | void __cfactx_invoke_thread(void (*main)(void *), void * this); |
---|
58 | } |
---|
59 | |
---|
60 | __cfaabi_dbg_debug_do( |
---|
61 | extern void __cfaabi_dbg_thread_register ( $thread * thrd ); |
---|
62 | extern void __cfaabi_dbg_thread_unregister( $thread * thrd ); |
---|
63 | ) |
---|
64 | |
---|
65 | // KERNEL ONLY unpark with out disabling interrupts |
---|
66 | void __unpark( struct __processor_id_t *, $thread * thrd __cfaabi_dbg_ctx_param2 ); |
---|
67 | |
---|
68 | static inline bool __post(single_sem & this, struct __processor_id_t * id) { |
---|
69 | for() { |
---|
70 | struct $thread * expected = this.ptr; |
---|
71 | if(expected == 1p) return false; |
---|
72 | if(expected == 0p) { |
---|
73 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 1p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { |
---|
74 | return false; |
---|
75 | } |
---|
76 | } |
---|
77 | else { |
---|
78 | if(__atomic_compare_exchange_n(&this.ptr, &expected, 0p, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) { |
---|
79 | __unpark( id, expected __cfaabi_dbg_ctx2 ); |
---|
80 | return true; |
---|
81 | } |
---|
82 | } |
---|
83 | } |
---|
84 | } |
---|
85 | |
---|
86 | //----------------------------------------------------------------------------- |
---|
87 | // Utils |
---|
88 | void doregister( struct cluster * cltr, struct $thread & thrd ); |
---|
89 | void unregister( struct cluster * cltr, struct $thread & thrd ); |
---|
90 | |
---|
91 | //----------------------------------------------------------------------------- |
---|
92 | // I/O |
---|
93 | void ^?{}(io_context & this, bool ); |
---|
94 | |
---|
95 | //======================================================================= |
---|
96 | // Cluster lock API |
---|
97 | //======================================================================= |
---|
98 | // Cells use by the reader writer lock |
---|
99 | // while not generic it only relies on a opaque pointer |
---|
100 | struct __attribute__((aligned(128))) __scheduler_lock_id_t { |
---|
101 | // Spin lock used as the underlying lock |
---|
102 | volatile bool lock; |
---|
103 | |
---|
104 | // Handle pointing to the proc owning this cell |
---|
105 | // Used for allocating cells and debugging |
---|
106 | __processor_id_t * volatile handle; |
---|
107 | |
---|
108 | #ifdef __CFA_WITH_VERIFY__ |
---|
109 | // Debug, check if this is owned for reading |
---|
110 | bool owned; |
---|
111 | #endif |
---|
112 | }; |
---|
113 | |
---|
114 | static_assert( sizeof(struct __scheduler_lock_id_t) <= __alignof(struct __scheduler_lock_id_t)); |
---|
115 | |
---|
116 | // Lock-Free registering/unregistering of threads |
---|
117 | // Register a processor to a given cluster and get its unique id in return |
---|
118 | unsigned doregister( struct __processor_id_t * proc ); |
---|
119 | |
---|
120 | // Unregister a processor from a given cluster using its id, getting back the original pointer |
---|
121 | void unregister( struct __processor_id_t * proc ); |
---|
122 | |
---|
123 | //----------------------------------------------------------------------- |
---|
124 | // Cluster idle lock/unlock |
---|
125 | static inline void lock(__cluster_idles & this) { |
---|
126 | for() { |
---|
127 | uint64_t l = this.lock; |
---|
128 | if( |
---|
129 | (0 == (l % 2)) |
---|
130 | && __atomic_compare_exchange_n(&this.lock, &l, l + 1, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST) |
---|
131 | ) return; |
---|
132 | Pause(); |
---|
133 | } |
---|
134 | } |
---|
135 | |
---|
136 | static inline void unlock(__cluster_idles & this) { |
---|
137 | /* paranoid */ verify( 1 == (this.lock % 2) ); |
---|
138 | __atomic_fetch_add( &this.lock, 1, __ATOMIC_SEQ_CST ); |
---|
139 | } |
---|
140 | |
---|
141 | //======================================================================= |
---|
142 | // Reader-writer lock implementation |
---|
143 | // Concurrent with doregister/unregister, |
---|
144 | // i.e., threads can be added at any point during or between the entry/exit |
---|
145 | |
---|
146 | //----------------------------------------------------------------------- |
---|
147 | // simple spinlock underlying the RWLock |
---|
148 | // Blocking acquire |
---|
149 | static inline void __atomic_acquire(volatile bool * ll) { |
---|
150 | while( __builtin_expect(__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST), false) ) { |
---|
151 | while(__atomic_load_n(ll, (int)__ATOMIC_RELAXED)) |
---|
152 | asm volatile("pause"); |
---|
153 | } |
---|
154 | /* paranoid */ verify(*ll); |
---|
155 | } |
---|
156 | |
---|
157 | // Non-Blocking acquire |
---|
158 | static inline bool __atomic_try_acquire(volatile bool * ll) { |
---|
159 | return !__atomic_exchange_n(ll, (bool)true, __ATOMIC_SEQ_CST); |
---|
160 | } |
---|
161 | |
---|
162 | // Release |
---|
163 | static inline void __atomic_unlock(volatile bool * ll) { |
---|
164 | /* paranoid */ verify(*ll); |
---|
165 | __atomic_store_n(ll, (bool)false, __ATOMIC_RELEASE); |
---|
166 | } |
---|
167 | |
---|
168 | //----------------------------------------------------------------------- |
---|
169 | // Reader-Writer lock protecting the ready-queues |
---|
170 | // while this lock is mostly generic some aspects |
---|
171 | // have been hard-coded to for the ready-queue for |
---|
172 | // simplicity and performance |
---|
173 | struct __scheduler_RWLock_t { |
---|
174 | // total cachelines allocated |
---|
175 | unsigned int max; |
---|
176 | |
---|
177 | // cachelines currently in use |
---|
178 | volatile unsigned int alloc; |
---|
179 | |
---|
180 | // cachelines ready to itereate over |
---|
181 | // (!= to alloc when thread is in second half of doregister) |
---|
182 | volatile unsigned int ready; |
---|
183 | |
---|
184 | // writer lock |
---|
185 | volatile bool lock; |
---|
186 | |
---|
187 | // data pointer |
---|
188 | __scheduler_lock_id_t * data; |
---|
189 | }; |
---|
190 | |
---|
191 | void ?{}(__scheduler_RWLock_t & this); |
---|
192 | void ^?{}(__scheduler_RWLock_t & this); |
---|
193 | |
---|
194 | extern __scheduler_RWLock_t * __scheduler_lock; |
---|
195 | |
---|
196 | //----------------------------------------------------------------------- |
---|
197 | // Reader side : acquire when using the ready queue to schedule but not |
---|
198 | // creating/destroying queues |
---|
199 | static inline void ready_schedule_lock( struct __processor_id_t * proc) with(*__scheduler_lock) { |
---|
200 | unsigned iproc = proc->id; |
---|
201 | /*paranoid*/ verify(data[iproc].handle == proc); |
---|
202 | /*paranoid*/ verify(iproc < ready); |
---|
203 | |
---|
204 | // Step 1 : make sure no writer are in the middle of the critical section |
---|
205 | while(__atomic_load_n(&lock, (int)__ATOMIC_RELAXED)) |
---|
206 | asm volatile("pause"); |
---|
207 | |
---|
208 | // Fence needed because we don't want to start trying to acquire the lock |
---|
209 | // before we read a false. |
---|
210 | // Not needed on x86 |
---|
211 | // std::atomic_thread_fence(std::memory_order_seq_cst); |
---|
212 | |
---|
213 | // Step 2 : acquire our local lock |
---|
214 | __atomic_acquire( &data[iproc].lock ); |
---|
215 | /*paranoid*/ verify(data[iproc].lock); |
---|
216 | |
---|
217 | #ifdef __CFA_WITH_VERIFY__ |
---|
218 | // Debug, check if this is owned for reading |
---|
219 | data[iproc].owned = true; |
---|
220 | #endif |
---|
221 | } |
---|
222 | |
---|
223 | static inline void ready_schedule_unlock( struct __processor_id_t * proc) with(*__scheduler_lock) { |
---|
224 | unsigned iproc = proc->id; |
---|
225 | /*paranoid*/ verify(data[iproc].handle == proc); |
---|
226 | /*paranoid*/ verify(iproc < ready); |
---|
227 | /*paranoid*/ verify(data[iproc].lock); |
---|
228 | /*paranoid*/ verify(data[iproc].owned); |
---|
229 | #ifdef __CFA_WITH_VERIFY__ |
---|
230 | // Debug, check if this is owned for reading |
---|
231 | data[iproc].owned = false; |
---|
232 | #endif |
---|
233 | __atomic_unlock(&data[iproc].lock); |
---|
234 | } |
---|
235 | |
---|
236 | #ifdef __CFA_WITH_VERIFY__ |
---|
237 | static inline bool ready_schedule_islocked( struct __processor_id_t * proc) { |
---|
238 | return __scheduler_lock->data[proc->id].owned; |
---|
239 | } |
---|
240 | |
---|
241 | static inline bool ready_mutate_islocked() { |
---|
242 | return __scheduler_lock->lock; |
---|
243 | } |
---|
244 | #endif |
---|
245 | |
---|
246 | //----------------------------------------------------------------------- |
---|
247 | // Writer side : acquire when changing the ready queue, e.g. adding more |
---|
248 | // queues or removing them. |
---|
249 | uint_fast32_t ready_mutate_lock( void ); |
---|
250 | |
---|
251 | void ready_mutate_unlock( uint_fast32_t /* value returned by lock */ ); |
---|
252 | |
---|
253 | //======================================================================= |
---|
254 | // Ready-Queue API |
---|
255 | //----------------------------------------------------------------------- |
---|
256 | // pop thread from the ready queue of a cluster |
---|
257 | // returns 0p if empty |
---|
258 | __attribute__((hot)) bool query(struct cluster * cltr); |
---|
259 | |
---|
260 | //----------------------------------------------------------------------- |
---|
261 | // push thread onto a ready queue for a cluster |
---|
262 | // returns true if the list was previously empty, false otherwise |
---|
263 | __attribute__((hot)) bool push(struct cluster * cltr, struct $thread * thrd); |
---|
264 | |
---|
265 | //----------------------------------------------------------------------- |
---|
266 | // pop thread from the ready queue of a cluster |
---|
267 | // returns 0p if empty |
---|
268 | // May return 0p spuriously |
---|
269 | __attribute__((hot)) struct $thread * pop(struct cluster * cltr); |
---|
270 | |
---|
271 | //----------------------------------------------------------------------- |
---|
272 | // pop thread from the ready queue of a cluster |
---|
273 | // returns 0p if empty |
---|
274 | // guaranteed to find any threads added before this call |
---|
275 | __attribute__((hot)) struct $thread * pop_slow(struct cluster * cltr); |
---|
276 | |
---|
277 | //----------------------------------------------------------------------- |
---|
278 | // remove thread from the ready queue of a cluster |
---|
279 | // returns bool if it wasn't found |
---|
280 | bool remove_head(struct cluster * cltr, struct $thread * thrd); |
---|
281 | |
---|
282 | //----------------------------------------------------------------------- |
---|
283 | // Increase the width of the ready queue (number of lanes) by 4 |
---|
284 | void ready_queue_grow (struct cluster * cltr, int target); |
---|
285 | |
---|
286 | //----------------------------------------------------------------------- |
---|
287 | // Decrease the width of the ready queue (number of lanes) by 4 |
---|
288 | void ready_queue_shrink(struct cluster * cltr, int target); |
---|
289 | |
---|
290 | |
---|
291 | // Local Variables: // |
---|
292 | // mode: c // |
---|
293 | // tab-width: 4 // |
---|
294 | // End: // |
---|