source: libcfa/src/concurrency/kernel/cluster.cfa @ 8b74fa7

ADTast-experimental
Last change on this file since 8b74fa7 was 31c967b, checked in by Thierry Delisle <tdelisle@…>, 2 years ago

Changed ready-queue so I can easily change the averaging algorithm.
Changed averaging to use logscale.

  • Property mode set to 100644
File size: 18.3 KB
Line 
1//
2// Cforall Version 1.0.0 Copyright (C) 2022 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// cluster.cfa -- file that includes helpers for subsystem that need cluster wide support
8//
9// Author           : Thierry Delisle
10// Created On       : Fri Mar 11 12:39:24 2022
11// Last Modified By :
12// Last Modified On :
13// Update Count     :
14//
15
16#define __cforall_thread__
17#define _GNU_SOURCE
18
19#include "bits/defs.hfa"
20#include "device/cpu.hfa"
21#include "kernel/cluster.hfa"
22#include "kernel/private.hfa"
23
24#include "stdlib.hfa"
25#include "limits.hfa"
26#include "math.hfa"
27
28#include "ready_subqueue.hfa"
29#include "io/types.hfa"
30
31#include <errno.h>
32#include <unistd.h>
33
34extern "C" {
35        #include <sys/syscall.h>  // __NR_xxx
36}
37
38// No overriden function, no environment variable, no define
39// fall back to a magic number
40#ifndef __CFA_MAX_PROCESSORS__
41        #define __CFA_MAX_PROCESSORS__ 1024
42#endif
43
44#if !defined(__CFA_NO_STATISTICS__)
45        #define __STATS(...) __VA_ARGS__
46#else
47        #define __STATS(...)
48#endif
49
50// returns the maximum number of processors the RWLock support
51__attribute__((weak)) unsigned __max_processors() libcfa_public {
52        const char * max_cores_s = getenv("CFA_MAX_PROCESSORS");
53        if(!max_cores_s) {
54                __cfadbg_print_nolock(ready_queue, "No CFA_MAX_PROCESSORS in ENV\n");
55                return __CFA_MAX_PROCESSORS__;
56        }
57
58        char * endptr = 0p;
59        long int max_cores_l = strtol(max_cores_s, &endptr, 10);
60        if(max_cores_l < 1 || max_cores_l > 65535) {
61                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS out of range : %ld\n", max_cores_l);
62                return __CFA_MAX_PROCESSORS__;
63        }
64        if('\0' != *endptr) {
65                __cfadbg_print_nolock(ready_queue, "CFA_MAX_PROCESSORS not a decimal number : %s\n", max_cores_s);
66                return __CFA_MAX_PROCESSORS__;
67        }
68
69        return max_cores_l;
70}
71
72#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
73        // No forward declaration needed
74        #define __kernel_rseq_register rseq_register_current_thread
75        #define __kernel_rseq_unregister rseq_unregister_current_thread
76#elif defined(CFA_HAVE_LINUX_RSEQ_H)
77        static void __kernel_raw_rseq_register  (void);
78        static void __kernel_raw_rseq_unregister(void);
79
80        #define __kernel_rseq_register __kernel_raw_rseq_register
81        #define __kernel_rseq_unregister __kernel_raw_rseq_unregister
82#else
83        // No forward declaration needed
84        // No initialization needed
85        static inline void noop(void) {}
86
87        #define __kernel_rseq_register noop
88        #define __kernel_rseq_unregister noop
89#endif
90
91//=======================================================================
92// Cluster wide reader-writer lock
93//=======================================================================
94void  ?{}(__scheduler_RWLock_t & this) {
95        this.lock.max   = __max_processors();
96        this.lock.alloc = 0;
97        this.lock.ready = 0;
98        this.lock.data  = alloc(this.lock.max);
99        this.lock.write_lock  = false;
100
101        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.lock.alloc), &this.lock.alloc));
102        /*paranoid*/ verify(__atomic_is_lock_free(sizeof(this.lock.ready), &this.lock.ready));
103
104}
105void ^?{}(__scheduler_RWLock_t & this) {
106        free(this.lock.data);
107}
108
109
110//=======================================================================
111// Lock-Free registering/unregistering of threads
112unsigned register_proc_id( void ) with(__scheduler_lock.lock) {
113        __kernel_rseq_register();
114
115        bool * handle = (bool *)&kernelTLS().sched_lock;
116
117        // Step - 1 : check if there is already space in the data
118        uint_fast32_t s = ready;
119
120        // Check among all the ready
121        for(uint_fast32_t i = 0; i < s; i++) {
122                bool * volatile * cell = (bool * volatile *)&data[i]; // Cforall is bugged and the double volatiles causes problems
123                /* paranoid */ verify( handle != *cell );
124
125                bool * null = 0p; // Re-write every loop since compare thrashes it
126                if( __atomic_load_n(cell, (int)__ATOMIC_RELAXED) == null
127                        && __atomic_compare_exchange_n( cell, &null, handle, false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST)) {
128                        /* paranoid */ verify(i < ready);
129                        /* paranoid */ verify( (kernelTLS().sched_id = i, true) );
130                        return i;
131                }
132        }
133
134        if(max <= alloc) abort("Trying to create more than %ud processors", __scheduler_lock.lock.max);
135
136        // Step - 2 : F&A to get a new spot in the array.
137        uint_fast32_t n = __atomic_fetch_add(&alloc, 1, __ATOMIC_SEQ_CST);
138        if(max <= n) abort("Trying to create more than %ud processors", __scheduler_lock.lock.max);
139
140        // Step - 3 : Mark space as used and then publish it.
141        data[n] = handle;
142        while() {
143                unsigned copy = n;
144                if( __atomic_load_n(&ready, __ATOMIC_RELAXED) == n
145                        && __atomic_compare_exchange_n(&ready, &copy, n + 1, true, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST))
146                        break;
147                Pause();
148        }
149
150        // Return new spot.
151        /* paranoid */ verify(n < ready);
152        /* paranoid */ verify( (kernelTLS().sched_id = n, true) );
153        return n;
154}
155
156void unregister_proc_id( unsigned id ) with(__scheduler_lock.lock) {
157        /* paranoid */ verify(id < ready);
158        /* paranoid */ verify(id == kernelTLS().sched_id);
159        /* paranoid */ verify(data[id] == &kernelTLS().sched_lock);
160
161        bool * volatile * cell = (bool * volatile *)&data[id]; // Cforall is bugged and the double volatiles causes problems
162
163        __atomic_store_n(cell, 0p, __ATOMIC_RELEASE);
164
165        __kernel_rseq_unregister();
166}
167
168//-----------------------------------------------------------------------
169// Writer side : acquire when changing the ready queue, e.g. adding more
170//  queues or removing them.
171uint_fast32_t ready_mutate_lock( void ) with(__scheduler_lock.lock) {
172        /* paranoid */ verify( ! __preemption_enabled() );
173
174        // Step 1 : lock global lock
175        // It is needed to avoid processors that register mid Critical-Section
176        //   to simply lock their own lock and enter.
177        __atomic_acquire( &write_lock );
178
179        // Make sure we won't deadlock ourself
180        // Checking before acquiring the writer lock isn't safe
181        // because someone else could have locked us.
182        /* paranoid */ verify( ! kernelTLS().sched_lock );
183
184        // Step 2 : lock per-proc lock
185        // Processors that are currently being registered aren't counted
186        //   but can't be in read_lock or in the critical section.
187        // All other processors are counted
188        uint_fast32_t s = ready;
189        for(uint_fast32_t i = 0; i < s; i++) {
190                volatile bool * llock = data[i];
191                if(llock) __atomic_acquire( llock );
192        }
193
194        /* paranoid */ verify( ! __preemption_enabled() );
195        return s;
196}
197
198void ready_mutate_unlock( uint_fast32_t last_s ) with(__scheduler_lock.lock) {
199        /* paranoid */ verify( ! __preemption_enabled() );
200
201        // Step 1 : release local locks
202        // This must be done while the global lock is held to avoid
203        //   threads that where created mid critical section
204        //   to race to lock their local locks and have the writer
205        //   immidiately unlock them
206        // Alternative solution : return s in write_lock and pass it to write_unlock
207        for(uint_fast32_t i = 0; i < last_s; i++) {
208                volatile bool * llock = data[i];
209                if(llock) __atomic_store_n(llock, (bool)false, __ATOMIC_RELEASE);
210        }
211
212        // Step 2 : release global lock
213        /*paranoid*/ assert(true == write_lock);
214        __atomic_store_n(&write_lock, (bool)false, __ATOMIC_RELEASE);
215
216        /* paranoid */ verify( ! __preemption_enabled() );
217}
218
219//=======================================================================
220// Cluster growth
221static const unsigned __readyq_single_shard = 2;
222
223void  ?{}(__timestamp_t & this) { this.t.tv = 0; this.t.ma = 0; }
224void ^?{}(__timestamp_t &) {}
225
226//-----------------------------------------------------------------------
227// Check that all the intrusive queues in the data structure are still consistent
228static void check_readyQ( cluster * cltr ) with (cltr->sched) {
229        #if defined(__CFA_WITH_VERIFY__)
230                {
231                        const unsigned lanes_count = readyQ.count;
232                        for( idx ; lanes_count ) {
233                                __intrusive_lane_t & sl = readyQ.data[idx];
234                                assert(!readyQ.data[idx].l.lock);
235
236                                        if(is_empty(sl)) {
237                                                assert( sl.l.anchor.next == 0p );
238                                                assert( sl.l.anchor.ts   == MAX );
239                                                assert( mock_head(sl)  == sl.l.prev );
240                                        } else {
241                                                assert( sl.l.anchor.next != 0p );
242                                                assert( sl.l.anchor.ts   != MAX );
243                                                assert( mock_head(sl)  != sl.l.prev );
244                                        }
245                        }
246                }
247        #endif
248}
249
250// Call this function of the intrusive list was moved using memcpy
251// fixes the list so that the pointers back to anchors aren't left dangling
252static inline void fix(__intrusive_lane_t & ll) {
253        if(is_empty(ll)) {
254                verify(ll.l.anchor.next == 0p);
255                ll.l.prev = mock_head(ll);
256        }
257}
258
259static void assign_list(unsigned & valrq, unsigned & valio, dlist(struct processor) & list, unsigned count) {
260        struct processor * it = &list`first;
261        for(unsigned i = 0; i < count; i++) {
262                /* paranoid */ verifyf( it, "Unexpected null iterator, at index %u of %u\n", i, count);
263                it->rdq.id = valrq;
264                it->rdq.target = UINT_MAX;
265                valrq += __shard_factor.readyq;
266                #if defined(CFA_HAVE_LINUX_IO_URING_H)
267                        it->io.ctx->cq.id = valio;
268                        it->io.target = UINT_MAX;
269                        valio += __shard_factor.io;
270                #endif
271                it = &(*it)`next;
272        }
273}
274
275static void reassign_cltr_id(struct cluster * cltr) {
276        unsigned prefrq = 0;
277        unsigned prefio = 0;
278        assign_list(prefrq, prefio, cltr->procs.actives, cltr->procs.total - cltr->procs.idle);
279        assign_list(prefrq, prefio, cltr->procs.idles  , cltr->procs.idle );
280}
281
282#if defined(CFA_HAVE_LINUX_IO_URING_H)
283        static void assign_io(io_context$ ** data, size_t count, dlist(struct processor) & list) {
284                struct processor * it = &list`first;
285                while(it) {
286                        /* paranoid */ verifyf( it, "Unexpected null iterator\n");
287                        /* paranoid */ verifyf( it->io.ctx->cq.id < count, "Processor %p has id %u above count %zu\n", it, it->rdq.id, count);
288                        data[it->io.ctx->cq.id] = it->io.ctx;
289                        it = &(*it)`next;
290                }
291        }
292
293        static void reassign_cltr_io(struct cluster * cltr) {
294                assign_io(cltr->sched.io.data, cltr->sched.io.count, cltr->procs.actives);
295                assign_io(cltr->sched.io.data, cltr->sched.io.count, cltr->procs.idles  );
296        }
297#else
298        static void reassign_cltr_io(struct cluster *) {}
299#endif
300
301static void fix_times( __timestamp_t * volatile & tscs, unsigned count ) {
302        tscs = alloc(count, tscs`realloc);
303        for(i; count) {
304                tscs[i].t.tv = rdtscl();
305                tscs[i].t.ma = 0;
306        }
307}
308
309// Grow the ready queue
310void ready_queue_grow(struct cluster * cltr) {
311        int target = cltr->procs.total;
312
313        /* paranoid */ verify( ready_mutate_islocked() );
314        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue\n");
315
316        // Make sure that everything is consistent
317        /* paranoid */ check_readyQ( cltr );
318
319
320        // Find new count
321        // Make sure we always have atleast 1 list
322        size_t ocount = cltr->sched.readyQ.count;
323        size_t ncount = max(target * __shard_factor.readyq, __readyq_single_shard);
324
325        // Do we have to do anything?
326        if( ocount != ncount ) {
327
328                // grow the ready queue
329                with( cltr->sched ) {
330
331                        // Allocate new array (uses realloc and memcpies the data)
332                        readyQ.data = alloc( ncount, readyQ.data`realloc );
333
334                        // Fix the moved data
335                        for( idx; ocount ) {
336                                fix(readyQ.data[idx]);
337                        }
338
339                        // Construct new data
340                        for( idx; ocount ~ ncount) {
341                                (readyQ.data[idx]){};
342                        }
343
344                        // Update original count
345                        readyQ.count = ncount;
346                }
347
348
349                fix_times(cltr->sched.readyQ.tscs, cltr->sched.readyQ.count);
350        }
351
352        // Fix the io times
353        cltr->sched.io.count = target * __shard_factor.io;
354        fix_times(cltr->sched.io.tscs, cltr->sched.io.count);
355
356        // realloc the caches
357        cltr->sched.caches = alloc( target, cltr->sched.caches`realloc );
358
359        // reassign the clusters.
360        reassign_cltr_id(cltr);
361
362        cltr->sched.io.data = alloc( cltr->sched.io.count, cltr->sched.io.data`realloc );
363        reassign_cltr_io(cltr);
364
365        // Make sure that everything is consistent
366        /* paranoid */ check_readyQ( cltr );
367        /* paranoid */ verify( (target == 0) == (cltr->sched.caches == 0p) );
368
369        __cfadbg_print_safe(ready_queue, "Kernel : Growing ready queue done\n");
370
371        /* paranoid */ verify( ready_mutate_islocked() );
372}
373
374// Shrink the ready queue
375void ready_queue_shrink(struct cluster * cltr) {
376        /* paranoid */ verify( ready_mutate_islocked() );
377        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue\n");
378
379        // Make sure that everything is consistent
380        /* paranoid */ check_readyQ( cltr );
381
382        int target = cltr->procs.total;
383
384        with( cltr->sched ) {
385                // Remember old count
386                size_t ocount = readyQ.count;
387
388                // Find new count
389                // Make sure we always have atleast 1 list
390                size_t ncount = max(target * __shard_factor.readyq, __readyq_single_shard);
391                /* paranoid */ verifyf( ocount >= ncount, "Error in shrinking size calculation, %zu >= %zu", ocount, ncount );
392                /* paranoid */ verifyf( ncount == target * __shard_factor.readyq || ncount == __readyq_single_shard,
393                /* paranoid */          "Error in shrinking size calculation, expected %u or %u, got %zu", target * __shard_factor.readyq, __readyq_single_shard, ncount );
394
395                readyQ.count = ncount;
396
397                // for printing count the number of displaced threads
398                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
399                        __attribute__((unused)) size_t displaced = 0;
400                #endif
401
402                // redistribute old data
403                for( idx; ncount ~ ocount) {
404                        // Lock is not strictly needed but makes checking invariants much easier
405                        __attribute__((unused)) bool locked = __atomic_try_acquire(&readyQ.data[idx].l.lock);
406                        verify(locked);
407
408                        // As long as we can pop from this lane to push the threads somewhere else in the queue
409                        while(!is_empty(readyQ.data[idx])) {
410                                struct thread$ * thrd;
411                                unsigned long long _;
412                                [thrd, _] = pop(readyQ.data[idx]);
413
414                                push(cltr, thrd, true);
415
416                                // for printing count the number of displaced threads
417                                #if defined(__CFA_DEBUG_PRINT__) || defined(__CFA_DEBUG_PRINT_READY_QUEUE__)
418                                        displaced++;
419                                #endif
420                        }
421
422                        // Unlock the lane
423                        __atomic_unlock(&readyQ.data[idx].l.lock);
424
425                        // TODO print the queue statistics here
426
427                        ^(readyQ.data[idx]){};
428                }
429
430                __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue displaced %zu threads\n", displaced);
431
432                // Allocate new array (uses realloc and memcpies the data)
433                readyQ.data = alloc( ncount, readyQ.data`realloc );
434
435                // Fix the moved data
436                for( idx; ncount ) {
437                        fix(readyQ.data[idx]);
438                }
439
440                fix_times(readyQ.tscs, ncount);
441        }
442        cltr->sched.caches = alloc( target, cltr->sched.caches`realloc );
443
444        // Fix the io times
445        cltr->sched.io.count = target * __shard_factor.io;
446        fix_times(cltr->sched.io.tscs, cltr->sched.io.count);
447
448        reassign_cltr_id(cltr);
449
450        cltr->sched.io.data = alloc( cltr->sched.io.count, cltr->sched.io.data`realloc );
451        reassign_cltr_io(cltr);
452
453        // Make sure that everything is consistent
454        /* paranoid */ verify( (target == 0) == (cltr->sched.caches == 0p) );
455        /* paranoid */ check_readyQ( cltr );
456
457        __cfadbg_print_safe(ready_queue, "Kernel : Shrinking ready queue done\n");
458        /* paranoid */ verify( ready_mutate_islocked() );
459}
460
461void ready_queue_close(struct cluster * cltr) {
462        free( cltr->sched.readyQ.data );
463        free( cltr->sched.readyQ.tscs );
464        cltr->sched.readyQ.data = 0p;
465        cltr->sched.readyQ.tscs = 0p;
466        cltr->sched.readyQ.count = 0;
467
468        free( cltr->sched.io.tscs );
469        free( cltr->sched.caches );
470}
471
472#define nested_offsetof(type, field) ((off_t)(&(((type*)0)-> field)))
473
474// Ctor
475void ?{}( __intrusive_lane_t & this ) {
476        this.l.lock = false;
477        this.l.prev = mock_head(this);
478        this.l.anchor.next = 0p;
479        this.l.anchor.ts   = MAX;
480        #if !defined(__CFA_NO_STATISTICS__)
481                this.l.cnt  = 0;
482        #endif
483
484        // We add a boat-load of assertions here because the anchor code is very fragile
485        /* paranoid */ _Static_assert( offsetof( thread$, link ) == nested_offsetof(__intrusive_lane_t, l.anchor) );
486        /* paranoid */ verify( offsetof( thread$, link ) == nested_offsetof(__intrusive_lane_t, l.anchor) );
487        /* paranoid */ verify( ((uintptr_t)( mock_head(this) ) + offsetof( thread$, link )) == (uintptr_t)(&this.l.anchor) );
488        /* paranoid */ verify( &mock_head(this)->link.next == &this.l.anchor.next );
489        /* paranoid */ verify( &mock_head(this)->link.ts   == &this.l.anchor.ts   );
490        /* paranoid */ verify( mock_head(this)->link.next == 0p );
491        /* paranoid */ verify( mock_head(this)->link.ts   == MAX );
492        /* paranoid */ verify( mock_head(this) == this.l.prev );
493        /* paranoid */ verify( __alignof__(__intrusive_lane_t) == 64 );
494        /* paranoid */ verify( __alignof__(this) == 64 );
495        /* paranoid */ verifyf( ((intptr_t)(&this) % 64) == 0, "Expected address to be aligned %p %% 64 == %zd", &this, ((intptr_t)(&this) % 64) );
496}
497
498#undef nested_offsetof
499
500// Dtor is trivial
501void ^?{}( __intrusive_lane_t & this ) {
502        // Make sure the list is empty
503        /* paranoid */ verify( this.l.anchor.next == 0p );
504        /* paranoid */ verify( this.l.anchor.ts   == MAX );
505        /* paranoid */ verify( mock_head(this)    == this.l.prev );
506}
507
508#if   defined(CFA_HAVE_LINUX_LIBRSEQ)
509        // No definition needed
510#elif defined(CFA_HAVE_LINUX_RSEQ_H)
511
512        #if defined( __x86_64 ) || defined( __i386 )
513                #define RSEQ_SIG        0x53053053
514        #elif defined( __ARM_ARCH )
515                #ifdef __ARMEB__
516                #define RSEQ_SIG    0xf3def5e7      /* udf    #24035    ; 0x5de3 (ARMv6+) */
517                #else
518                #define RSEQ_SIG    0xe7f5def3      /* udf    #24035    ; 0x5de3 */
519                #endif
520        #endif
521
522        extern void __disable_interrupts_hard();
523        extern void __enable_interrupts_hard();
524
525        static void __kernel_raw_rseq_register  (void) {
526                /* paranoid */ verify( __cfaabi_rseq.cpu_id == RSEQ_CPU_ID_UNINITIALIZED );
527
528                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, (sigset_t *)0p, _NSIG / 8);
529                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), 0, RSEQ_SIG);
530                if(ret != 0) {
531                        int e = errno;
532                        switch(e) {
533                        case EINVAL: abort("KERNEL ERROR: rseq register invalid argument");
534                        case ENOSYS: abort("KERNEL ERROR: rseq register no supported");
535                        case EFAULT: abort("KERNEL ERROR: rseq register with invalid argument");
536                        case EBUSY : abort("KERNEL ERROR: rseq register already registered");
537                        case EPERM : abort("KERNEL ERROR: rseq register sig  argument  on unregistration does not match the signature received on registration");
538                        default: abort("KERNEL ERROR: rseq register unexpected return %d", e);
539                        }
540                }
541        }
542
543        static void __kernel_raw_rseq_unregister(void) {
544                /* paranoid */ verify( __cfaabi_rseq.cpu_id >= 0 );
545
546                // int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, (sigset_t *)0p, _NSIG / 8);
547                int ret = syscall(__NR_rseq, &__cfaabi_rseq, sizeof(struct rseq), RSEQ_FLAG_UNREGISTER, RSEQ_SIG);
548                if(ret != 0) {
549                        int e = errno;
550                        switch(e) {
551                        case EINVAL: abort("KERNEL ERROR: rseq unregister invalid argument");
552                        case ENOSYS: abort("KERNEL ERROR: rseq unregister no supported");
553                        case EFAULT: abort("KERNEL ERROR: rseq unregister with invalid argument");
554                        case EBUSY : abort("KERNEL ERROR: rseq unregister already registered");
555                        case EPERM : abort("KERNEL ERROR: rseq unregister sig  argument  on unregistration does not match the signature received on registration");
556                        default: abort("KERNEL ERROR: rseq unregisteunexpected return %d", e);
557                        }
558                }
559        }
560#else
561        // No definition needed
562#endif
Note: See TracBrowser for help on using the repository browser.