diff options
Diffstat (limited to 'litmus/sched_pfp.c')
-rw-r--r-- | litmus/sched_pfp.c | 1685 |
1 files changed, 1685 insertions, 0 deletions
diff --git a/litmus/sched_pfp.c b/litmus/sched_pfp.c new file mode 100644 index 000000000000..62be699629b1 --- /dev/null +++ b/litmus/sched_pfp.c | |||
@@ -0,0 +1,1685 @@ | |||
1 | /* | ||
2 | * litmus/sched_pfp.c | ||
3 | * | ||
4 | * Implementation of partitioned fixed-priority scheduling. | ||
5 | * Based on PSN-EDF. | ||
6 | */ | ||
7 | |||
8 | #include <linux/percpu.h> | ||
9 | #include <linux/sched.h> | ||
10 | #include <linux/list.h> | ||
11 | #include <linux/spinlock.h> | ||
12 | #include <linux/module.h> | ||
13 | |||
14 | #include <litmus/litmus.h> | ||
15 | #include <litmus/wait.h> | ||
16 | #include <litmus/jobs.h> | ||
17 | #include <litmus/preempt.h> | ||
18 | #include <litmus/fp_common.h> | ||
19 | #include <litmus/sched_plugin.h> | ||
20 | #include <litmus/sched_trace.h> | ||
21 | #include <litmus/trace.h> | ||
22 | #include <litmus/budget.h> | ||
23 | |||
24 | #include <linux/uaccess.h> | ||
25 | |||
26 | |||
27 | typedef struct { | ||
28 | rt_domain_t domain; | ||
29 | struct fp_prio_queue ready_queue; | ||
30 | int cpu; | ||
31 | struct task_struct* scheduled; /* only RT tasks */ | ||
32 | /* | ||
33 | * scheduling lock slock | ||
34 | * protects the domain and serializes scheduling decisions | ||
35 | */ | ||
36 | #define slock domain.ready_lock | ||
37 | |||
38 | } pfp_domain_t; | ||
39 | |||
40 | DEFINE_PER_CPU(pfp_domain_t, pfp_domains); | ||
41 | |||
42 | pfp_domain_t* pfp_doms[NR_CPUS]; | ||
43 | |||
44 | #define local_pfp (&__get_cpu_var(pfp_domains)) | ||
45 | #define remote_dom(cpu) (&per_cpu(pfp_domains, cpu).domain) | ||
46 | #define remote_pfp(cpu) (&per_cpu(pfp_domains, cpu)) | ||
47 | #define task_dom(task) remote_dom(get_partition(task)) | ||
48 | #define task_pfp(task) remote_pfp(get_partition(task)) | ||
49 | |||
50 | /* we assume the lock is being held */ | ||
51 | static void preempt(pfp_domain_t *pfp) | ||
52 | { | ||
53 | preempt_if_preemptable(pfp->scheduled, pfp->cpu); | ||
54 | } | ||
55 | |||
56 | static unsigned int priority_index(struct task_struct* t) | ||
57 | { | ||
58 | #ifdef CONFIG_LOCKING | ||
59 | if (unlikely(t->rt_param.inh_task)) | ||
60 | /* use effective priority */ | ||
61 | t = t->rt_param.inh_task; | ||
62 | |||
63 | if (is_priority_boosted(t)) { | ||
64 | /* zero is reserved for priority-boosted tasks */ | ||
65 | return 0; | ||
66 | } else | ||
67 | #endif | ||
68 | return get_priority(t); | ||
69 | } | ||
70 | |||
71 | |||
72 | static void pfp_release_jobs(rt_domain_t* rt, struct bheap* tasks) | ||
73 | { | ||
74 | pfp_domain_t *pfp = container_of(rt, pfp_domain_t, domain); | ||
75 | unsigned long flags; | ||
76 | struct task_struct* t; | ||
77 | struct bheap_node* hn; | ||
78 | |||
79 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
80 | |||
81 | while (!bheap_empty(tasks)) { | ||
82 | hn = bheap_take(fp_ready_order, tasks); | ||
83 | t = bheap2task(hn); | ||
84 | TRACE_TASK(t, "released (part:%d prio:%d)\n", | ||
85 | get_partition(t), get_priority(t)); | ||
86 | fp_prio_add(&pfp->ready_queue, t, priority_index(t)); | ||
87 | } | ||
88 | |||
89 | /* do we need to preempt? */ | ||
90 | if (fp_higher_prio(fp_prio_peek(&pfp->ready_queue), pfp->scheduled)) { | ||
91 | TRACE_CUR("preempted by new release\n"); | ||
92 | preempt(pfp); | ||
93 | } | ||
94 | |||
95 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
96 | } | ||
97 | |||
98 | static void pfp_domain_init(pfp_domain_t* pfp, | ||
99 | int cpu) | ||
100 | { | ||
101 | fp_domain_init(&pfp->domain, NULL, pfp_release_jobs); | ||
102 | pfp->cpu = cpu; | ||
103 | pfp->scheduled = NULL; | ||
104 | fp_prio_queue_init(&pfp->ready_queue); | ||
105 | } | ||
106 | |||
107 | static void requeue(struct task_struct* t, pfp_domain_t *pfp) | ||
108 | { | ||
109 | if (t->state != TASK_RUNNING) | ||
110 | TRACE_TASK(t, "requeue: !TASK_RUNNING\n"); | ||
111 | |||
112 | set_rt_flags(t, RT_F_RUNNING); | ||
113 | if (is_released(t, litmus_clock())) | ||
114 | fp_prio_add(&pfp->ready_queue, t, priority_index(t)); | ||
115 | else | ||
116 | add_release(&pfp->domain, t); /* it has got to wait */ | ||
117 | } | ||
118 | |||
119 | static void job_completion(struct task_struct* t, int forced) | ||
120 | { | ||
121 | sched_trace_task_completion(t,forced); | ||
122 | TRACE_TASK(t, "job_completion().\n"); | ||
123 | |||
124 | set_rt_flags(t, RT_F_SLEEP); | ||
125 | prepare_for_next_period(t); | ||
126 | } | ||
127 | |||
128 | static void pfp_tick(struct task_struct *t) | ||
129 | { | ||
130 | pfp_domain_t *pfp = local_pfp; | ||
131 | |||
132 | /* Check for inconsistency. We don't need the lock for this since | ||
133 | * ->scheduled is only changed in schedule, which obviously is not | ||
134 | * executing in parallel on this CPU | ||
135 | */ | ||
136 | BUG_ON(is_realtime(t) && t != pfp->scheduled); | ||
137 | |||
138 | if (is_realtime(t) && budget_enforced(t) && budget_exhausted(t)) { | ||
139 | if (!is_np(t)) { | ||
140 | litmus_reschedule_local(); | ||
141 | TRACE("pfp_scheduler_tick: " | ||
142 | "%d is preemptable " | ||
143 | " => FORCE_RESCHED\n", t->pid); | ||
144 | } else if (is_user_np(t)) { | ||
145 | TRACE("pfp_scheduler_tick: " | ||
146 | "%d is non-preemptable, " | ||
147 | "preemption delayed.\n", t->pid); | ||
148 | request_exit_np(t); | ||
149 | } | ||
150 | } | ||
151 | } | ||
152 | |||
153 | static struct task_struct* pfp_schedule(struct task_struct * prev) | ||
154 | { | ||
155 | pfp_domain_t* pfp = local_pfp; | ||
156 | struct task_struct* next; | ||
157 | |||
158 | int out_of_time, sleep, preempt, np, exists, blocks, resched, migrate; | ||
159 | |||
160 | raw_spin_lock(&pfp->slock); | ||
161 | |||
162 | /* sanity checking | ||
163 | * differently from gedf, when a task exits (dead) | ||
164 | * pfp->schedule may be null and prev _is_ realtime | ||
165 | */ | ||
166 | BUG_ON(pfp->scheduled && pfp->scheduled != prev); | ||
167 | BUG_ON(pfp->scheduled && !is_realtime(prev)); | ||
168 | |||
169 | /* (0) Determine state */ | ||
170 | exists = pfp->scheduled != NULL; | ||
171 | blocks = exists && !is_running(pfp->scheduled); | ||
172 | out_of_time = exists && | ||
173 | budget_enforced(pfp->scheduled) && | ||
174 | budget_exhausted(pfp->scheduled); | ||
175 | np = exists && is_np(pfp->scheduled); | ||
176 | sleep = exists && get_rt_flags(pfp->scheduled) == RT_F_SLEEP; | ||
177 | migrate = exists && get_partition(pfp->scheduled) != pfp->cpu; | ||
178 | preempt = migrate || fp_preemption_needed(&pfp->ready_queue, prev); | ||
179 | |||
180 | /* If we need to preempt do so. | ||
181 | * The following checks set resched to 1 in case of special | ||
182 | * circumstances. | ||
183 | */ | ||
184 | resched = preempt; | ||
185 | |||
186 | /* If a task blocks we have no choice but to reschedule. | ||
187 | */ | ||
188 | if (blocks) | ||
189 | resched = 1; | ||
190 | |||
191 | /* Request a sys_exit_np() call if we would like to preempt but cannot. | ||
192 | * Multiple calls to request_exit_np() don't hurt. | ||
193 | */ | ||
194 | if (np && (out_of_time || preempt || sleep)) | ||
195 | request_exit_np(pfp->scheduled); | ||
196 | |||
197 | /* Any task that is preemptable and either exhausts its execution | ||
198 | * budget or wants to sleep completes. We may have to reschedule after | ||
199 | * this. | ||
200 | */ | ||
201 | if (!np && (out_of_time || sleep) && !blocks && !migrate) { | ||
202 | job_completion(pfp->scheduled, !sleep); | ||
203 | resched = 1; | ||
204 | } | ||
205 | |||
206 | /* The final scheduling decision. Do we need to switch for some reason? | ||
207 | * Switch if we are in RT mode and have no task or if we need to | ||
208 | * resched. | ||
209 | */ | ||
210 | next = NULL; | ||
211 | if ((!np || blocks) && (resched || !exists)) { | ||
212 | /* When preempting a task that does not block, then | ||
213 | * re-insert it into either the ready queue or the | ||
214 | * release queue (if it completed). requeue() picks | ||
215 | * the appropriate queue. | ||
216 | */ | ||
217 | if (pfp->scheduled && !blocks && !migrate) | ||
218 | requeue(pfp->scheduled, pfp); | ||
219 | next = fp_prio_take(&pfp->ready_queue); | ||
220 | } else | ||
221 | /* Only override Linux scheduler if we have a real-time task | ||
222 | * scheduled that needs to continue. | ||
223 | */ | ||
224 | if (exists) | ||
225 | next = prev; | ||
226 | |||
227 | if (next) { | ||
228 | TRACE_TASK(next, "scheduled at %llu\n", litmus_clock()); | ||
229 | set_rt_flags(next, RT_F_RUNNING); | ||
230 | } else { | ||
231 | TRACE("becoming idle at %llu\n", litmus_clock()); | ||
232 | } | ||
233 | |||
234 | pfp->scheduled = next; | ||
235 | sched_state_task_picked(); | ||
236 | raw_spin_unlock(&pfp->slock); | ||
237 | |||
238 | return next; | ||
239 | } | ||
240 | |||
241 | #ifdef CONFIG_LITMUS_LOCKING | ||
242 | |||
243 | /* prev is no longer scheduled --- see if it needs to migrate */ | ||
244 | static void pfp_finish_switch(struct task_struct *prev) | ||
245 | { | ||
246 | pfp_domain_t *to; | ||
247 | |||
248 | if (is_realtime(prev) && | ||
249 | is_running(prev) && | ||
250 | get_partition(prev) != smp_processor_id()) { | ||
251 | TRACE_TASK(prev, "needs to migrate from P%d to P%d\n", | ||
252 | smp_processor_id(), get_partition(prev)); | ||
253 | |||
254 | to = task_pfp(prev); | ||
255 | |||
256 | raw_spin_lock(&to->slock); | ||
257 | |||
258 | TRACE_TASK(prev, "adding to queue on P%d\n", to->cpu); | ||
259 | requeue(prev, to); | ||
260 | if (fp_preemption_needed(&to->ready_queue, to->scheduled)) | ||
261 | preempt(to); | ||
262 | |||
263 | raw_spin_unlock(&to->slock); | ||
264 | |||
265 | } | ||
266 | } | ||
267 | |||
268 | #endif | ||
269 | |||
270 | /* Prepare a task for running in RT mode | ||
271 | */ | ||
272 | static void pfp_task_new(struct task_struct * t, int on_rq, int running) | ||
273 | { | ||
274 | pfp_domain_t* pfp = task_pfp(t); | ||
275 | unsigned long flags; | ||
276 | |||
277 | TRACE_TASK(t, "P-FP: task new, cpu = %d\n", | ||
278 | t->rt_param.task_params.cpu); | ||
279 | |||
280 | /* setup job parameters */ | ||
281 | release_at(t, litmus_clock()); | ||
282 | |||
283 | /* The task should be running in the queue, otherwise signal | ||
284 | * code will try to wake it up with fatal consequences. | ||
285 | */ | ||
286 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
287 | if (running) { | ||
288 | /* there shouldn't be anything else running at the time */ | ||
289 | BUG_ON(pfp->scheduled); | ||
290 | pfp->scheduled = t; | ||
291 | } else { | ||
292 | requeue(t, pfp); | ||
293 | /* maybe we have to reschedule */ | ||
294 | preempt(pfp); | ||
295 | } | ||
296 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
297 | } | ||
298 | |||
299 | static void pfp_task_wake_up(struct task_struct *task) | ||
300 | { | ||
301 | unsigned long flags; | ||
302 | pfp_domain_t* pfp = task_pfp(task); | ||
303 | lt_t now; | ||
304 | |||
305 | TRACE_TASK(task, "wake_up at %llu\n", litmus_clock()); | ||
306 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
307 | |||
308 | #ifdef CONFIG_LITMUS_LOCKING | ||
309 | /* Should only be queued when processing a fake-wake up due to a | ||
310 | * migration-related state change. */ | ||
311 | if (unlikely(is_queued(task))) { | ||
312 | TRACE_TASK(task, "WARNING: waking task still queued. Is this right?\n"); | ||
313 | goto out_unlock; | ||
314 | } | ||
315 | #else | ||
316 | BUG_ON(is_queued(task)); | ||
317 | #endif | ||
318 | now = litmus_clock(); | ||
319 | if (is_tardy(task, now) | ||
320 | #ifdef CONFIG_LITMUS_LOCKING | ||
321 | /* We need to take suspensions because of semaphores into | ||
322 | * account! If a job resumes after being suspended due to acquiring | ||
323 | * a semaphore, it should never be treated as a new job release. | ||
324 | */ | ||
325 | && !is_priority_boosted(task) | ||
326 | #endif | ||
327 | ) { | ||
328 | /* new sporadic release */ | ||
329 | release_at(task, now); | ||
330 | sched_trace_task_release(task); | ||
331 | } | ||
332 | |||
333 | /* Only add to ready queue if it is not the currently-scheduled | ||
334 | * task. This could be the case if a task was woken up concurrently | ||
335 | * on a remote CPU before the executing CPU got around to actually | ||
336 | * de-scheduling the task, i.e., wake_up() raced with schedule() | ||
337 | * and won. Also, don't requeue if it is still queued, which can | ||
338 | * happen under the DPCP due wake-ups racing with migrations. | ||
339 | */ | ||
340 | if (pfp->scheduled != task) | ||
341 | requeue(task, pfp); | ||
342 | |||
343 | out_unlock: | ||
344 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
345 | TRACE_TASK(task, "wake up done\n"); | ||
346 | } | ||
347 | |||
348 | static void pfp_task_block(struct task_struct *t) | ||
349 | { | ||
350 | /* only running tasks can block, thus t is in no queue */ | ||
351 | TRACE_TASK(t, "block at %llu, state=%d\n", litmus_clock(), t->state); | ||
352 | |||
353 | BUG_ON(!is_realtime(t)); | ||
354 | |||
355 | /* If this task blocked normally, it shouldn't be queued. The exception is | ||
356 | * if this is a simulated block()/wakeup() pair from the pull-migration code path. | ||
357 | * This should only happen if the DPCP is being used. | ||
358 | */ | ||
359 | #ifdef CONFIG_LITMUS_LOCKING | ||
360 | if (unlikely(is_queued(t))) | ||
361 | TRACE_TASK(t, "WARNING: blocking task still queued. Is this right?\n"); | ||
362 | #else | ||
363 | BUG_ON(is_queued(t)); | ||
364 | #endif | ||
365 | } | ||
366 | |||
367 | static void pfp_task_exit(struct task_struct * t) | ||
368 | { | ||
369 | unsigned long flags; | ||
370 | pfp_domain_t* pfp = task_pfp(t); | ||
371 | rt_domain_t* dom; | ||
372 | |||
373 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
374 | if (is_queued(t)) { | ||
375 | BUG(); /* This currently doesn't work. */ | ||
376 | /* dequeue */ | ||
377 | dom = task_dom(t); | ||
378 | remove(dom, t); | ||
379 | } | ||
380 | if (pfp->scheduled == t) { | ||
381 | pfp->scheduled = NULL; | ||
382 | preempt(pfp); | ||
383 | } | ||
384 | TRACE_TASK(t, "RIP, now reschedule\n"); | ||
385 | |||
386 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
387 | } | ||
388 | |||
389 | #ifdef CONFIG_LITMUS_LOCKING | ||
390 | |||
391 | #include <litmus/fdso.h> | ||
392 | #include <litmus/srp.h> | ||
393 | |||
394 | static void fp_dequeue(pfp_domain_t* pfp, struct task_struct* t) | ||
395 | { | ||
396 | BUG_ON(pfp->scheduled == t && is_queued(t)); | ||
397 | if (is_queued(t)) | ||
398 | fp_prio_remove(&pfp->ready_queue, t, priority_index(t)); | ||
399 | } | ||
400 | |||
401 | static void fp_set_prio_inh(pfp_domain_t* pfp, struct task_struct* t, | ||
402 | struct task_struct* prio_inh) | ||
403 | { | ||
404 | int requeue; | ||
405 | |||
406 | if (!t || t->rt_param.inh_task == prio_inh) { | ||
407 | /* no update required */ | ||
408 | if (t) | ||
409 | TRACE_TASK(t, "no prio-inh update required\n"); | ||
410 | return; | ||
411 | } | ||
412 | |||
413 | requeue = is_queued(t); | ||
414 | TRACE_TASK(t, "prio-inh: is_queued:%d\n", requeue); | ||
415 | |||
416 | if (requeue) | ||
417 | /* first remove */ | ||
418 | fp_dequeue(pfp, t); | ||
419 | |||
420 | t->rt_param.inh_task = prio_inh; | ||
421 | |||
422 | if (requeue) | ||
423 | /* add again to the right queue */ | ||
424 | fp_prio_add(&pfp->ready_queue, t, priority_index(t)); | ||
425 | } | ||
426 | |||
427 | static int effective_agent_priority(int prio) | ||
428 | { | ||
429 | /* make sure agents have higher priority */ | ||
430 | return prio - LITMUS_MAX_PRIORITY; | ||
431 | } | ||
432 | |||
433 | static lt_t prio_point(int eprio) | ||
434 | { | ||
435 | /* make sure we have non-negative prio points */ | ||
436 | return eprio + LITMUS_MAX_PRIORITY; | ||
437 | } | ||
438 | |||
439 | static int prio_from_point(lt_t prio_point) | ||
440 | { | ||
441 | return ((int) prio_point) - LITMUS_MAX_PRIORITY; | ||
442 | } | ||
443 | |||
444 | static void boost_priority(struct task_struct* t, lt_t priority_point) | ||
445 | { | ||
446 | unsigned long flags; | ||
447 | pfp_domain_t* pfp = task_pfp(t); | ||
448 | |||
449 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
450 | |||
451 | |||
452 | TRACE_TASK(t, "priority boosted at %llu\n", litmus_clock()); | ||
453 | |||
454 | tsk_rt(t)->priority_boosted = 1; | ||
455 | /* tie-break by protocol-specific priority point */ | ||
456 | tsk_rt(t)->boost_start_time = priority_point; | ||
457 | |||
458 | if (pfp->scheduled != t) { | ||
459 | /* holder may be queued: first stop queue changes */ | ||
460 | raw_spin_lock(&pfp->domain.release_lock); | ||
461 | if (is_queued(t) && | ||
462 | /* If it is queued, then we need to re-order. */ | ||
463 | bheap_decrease(fp_ready_order, tsk_rt(t)->heap_node) && | ||
464 | /* If we bubbled to the top, then we need to check for preemptions. */ | ||
465 | fp_preemption_needed(&pfp->ready_queue, pfp->scheduled)) | ||
466 | preempt(pfp); | ||
467 | raw_spin_unlock(&pfp->domain.release_lock); | ||
468 | } /* else: nothing to do since the job is not queued while scheduled */ | ||
469 | |||
470 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
471 | } | ||
472 | |||
473 | static void unboost_priority(struct task_struct* t) | ||
474 | { | ||
475 | unsigned long flags; | ||
476 | pfp_domain_t* pfp = task_pfp(t); | ||
477 | lt_t now; | ||
478 | |||
479 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
480 | now = litmus_clock(); | ||
481 | |||
482 | /* assumption: this only happens when the job is scheduled */ | ||
483 | BUG_ON(pfp->scheduled != t); | ||
484 | |||
485 | TRACE_TASK(t, "priority restored at %llu\n", now); | ||
486 | |||
487 | /* priority boosted jobs must be scheduled */ | ||
488 | BUG_ON(pfp->scheduled != t); | ||
489 | |||
490 | tsk_rt(t)->priority_boosted = 0; | ||
491 | tsk_rt(t)->boost_start_time = 0; | ||
492 | |||
493 | /* check if this changes anything */ | ||
494 | if (fp_preemption_needed(&pfp->ready_queue, pfp->scheduled)) | ||
495 | preempt(pfp); | ||
496 | |||
497 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
498 | } | ||
499 | |||
500 | /* ******************** SRP support ************************ */ | ||
501 | |||
502 | static unsigned int pfp_get_srp_prio(struct task_struct* t) | ||
503 | { | ||
504 | return get_priority(t); | ||
505 | } | ||
506 | |||
507 | /* ******************** FMLP support ********************** */ | ||
508 | |||
509 | struct fmlp_semaphore { | ||
510 | struct litmus_lock litmus_lock; | ||
511 | |||
512 | /* current resource holder */ | ||
513 | struct task_struct *owner; | ||
514 | |||
515 | /* FIFO queue of waiting tasks */ | ||
516 | wait_queue_head_t wait; | ||
517 | }; | ||
518 | |||
519 | static inline struct fmlp_semaphore* fmlp_from_lock(struct litmus_lock* lock) | ||
520 | { | ||
521 | return container_of(lock, struct fmlp_semaphore, litmus_lock); | ||
522 | } | ||
523 | int pfp_fmlp_lock(struct litmus_lock* l) | ||
524 | { | ||
525 | struct task_struct* t = current; | ||
526 | struct fmlp_semaphore *sem = fmlp_from_lock(l); | ||
527 | wait_queue_t wait; | ||
528 | unsigned long flags; | ||
529 | lt_t time_of_request; | ||
530 | |||
531 | if (!is_realtime(t)) | ||
532 | return -EPERM; | ||
533 | |||
534 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
535 | |||
536 | /* tie-break by this point in time */ | ||
537 | time_of_request = litmus_clock(); | ||
538 | |||
539 | /* Priority-boost ourself *before* we suspend so that | ||
540 | * our priority is boosted when we resume. */ | ||
541 | boost_priority(t, time_of_request); | ||
542 | |||
543 | if (sem->owner) { | ||
544 | /* resource is not free => must suspend and wait */ | ||
545 | |||
546 | init_waitqueue_entry(&wait, t); | ||
547 | |||
548 | /* FIXME: interruptible would be nice some day */ | ||
549 | set_task_state(t, TASK_UNINTERRUPTIBLE); | ||
550 | |||
551 | __add_wait_queue_tail_exclusive(&sem->wait, &wait); | ||
552 | |||
553 | TS_LOCK_SUSPEND; | ||
554 | |||
555 | /* release lock before sleeping */ | ||
556 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
557 | |||
558 | /* We depend on the FIFO order. Thus, we don't need to recheck | ||
559 | * when we wake up; we are guaranteed to have the lock since | ||
560 | * there is only one wake up per release. | ||
561 | */ | ||
562 | |||
563 | schedule(); | ||
564 | |||
565 | TS_LOCK_RESUME; | ||
566 | |||
567 | /* Since we hold the lock, no other task will change | ||
568 | * ->owner. We can thus check it without acquiring the spin | ||
569 | * lock. */ | ||
570 | BUG_ON(sem->owner != t); | ||
571 | } else { | ||
572 | /* it's ours now */ | ||
573 | sem->owner = t; | ||
574 | |||
575 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
576 | } | ||
577 | |||
578 | return 0; | ||
579 | } | ||
580 | |||
581 | int pfp_fmlp_unlock(struct litmus_lock* l) | ||
582 | { | ||
583 | struct task_struct *t = current, *next; | ||
584 | struct fmlp_semaphore *sem = fmlp_from_lock(l); | ||
585 | unsigned long flags; | ||
586 | int err = 0; | ||
587 | |||
588 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
589 | |||
590 | if (sem->owner != t) { | ||
591 | err = -EINVAL; | ||
592 | goto out; | ||
593 | } | ||
594 | |||
595 | /* we lose the benefit of priority boosting */ | ||
596 | |||
597 | unboost_priority(t); | ||
598 | |||
599 | /* check if there are jobs waiting for this resource */ | ||
600 | next = __waitqueue_remove_first(&sem->wait); | ||
601 | if (next) { | ||
602 | /* next becomes the resouce holder */ | ||
603 | sem->owner = next; | ||
604 | |||
605 | /* Wake up next. The waiting job is already priority-boosted. */ | ||
606 | wake_up_process(next); | ||
607 | } else | ||
608 | /* resource becomes available */ | ||
609 | sem->owner = NULL; | ||
610 | |||
611 | out: | ||
612 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
613 | return err; | ||
614 | } | ||
615 | |||
616 | int pfp_fmlp_close(struct litmus_lock* l) | ||
617 | { | ||
618 | struct task_struct *t = current; | ||
619 | struct fmlp_semaphore *sem = fmlp_from_lock(l); | ||
620 | unsigned long flags; | ||
621 | |||
622 | int owner; | ||
623 | |||
624 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
625 | |||
626 | owner = sem->owner == t; | ||
627 | |||
628 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
629 | |||
630 | if (owner) | ||
631 | pfp_fmlp_unlock(l); | ||
632 | |||
633 | return 0; | ||
634 | } | ||
635 | |||
636 | void pfp_fmlp_free(struct litmus_lock* lock) | ||
637 | { | ||
638 | kfree(fmlp_from_lock(lock)); | ||
639 | } | ||
640 | |||
641 | static struct litmus_lock_ops pfp_fmlp_lock_ops = { | ||
642 | .close = pfp_fmlp_close, | ||
643 | .lock = pfp_fmlp_lock, | ||
644 | .unlock = pfp_fmlp_unlock, | ||
645 | .deallocate = pfp_fmlp_free, | ||
646 | }; | ||
647 | |||
648 | static struct litmus_lock* pfp_new_fmlp(void) | ||
649 | { | ||
650 | struct fmlp_semaphore* sem; | ||
651 | |||
652 | sem = kmalloc(sizeof(*sem), GFP_KERNEL); | ||
653 | if (!sem) | ||
654 | return NULL; | ||
655 | |||
656 | sem->owner = NULL; | ||
657 | init_waitqueue_head(&sem->wait); | ||
658 | sem->litmus_lock.ops = &pfp_fmlp_lock_ops; | ||
659 | |||
660 | return &sem->litmus_lock; | ||
661 | } | ||
662 | |||
663 | /* ******************** MPCP support ********************** */ | ||
664 | |||
665 | struct mpcp_semaphore { | ||
666 | struct litmus_lock litmus_lock; | ||
667 | |||
668 | /* current resource holder */ | ||
669 | struct task_struct *owner; | ||
670 | |||
671 | /* priority queue of waiting tasks */ | ||
672 | wait_queue_head_t wait; | ||
673 | |||
674 | /* priority ceiling per cpu */ | ||
675 | unsigned int prio_ceiling[NR_CPUS]; | ||
676 | |||
677 | /* should jobs spin "virtually" for this resource? */ | ||
678 | int vspin; | ||
679 | }; | ||
680 | |||
681 | #define OMEGA_CEILING UINT_MAX | ||
682 | |||
683 | /* Since jobs spin "virtually" while waiting to acquire a lock, | ||
684 | * they first must aquire a local per-cpu resource. | ||
685 | */ | ||
686 | static DEFINE_PER_CPU(wait_queue_head_t, mpcpvs_vspin_wait); | ||
687 | static DEFINE_PER_CPU(struct task_struct*, mpcpvs_vspin); | ||
688 | |||
689 | /* called with preemptions off <=> no local modifications */ | ||
690 | static void mpcp_vspin_enter(void) | ||
691 | { | ||
692 | struct task_struct* t = current; | ||
693 | |||
694 | while (1) { | ||
695 | if (__get_cpu_var(mpcpvs_vspin) == NULL) { | ||
696 | /* good, we get to issue our request */ | ||
697 | __get_cpu_var(mpcpvs_vspin) = t; | ||
698 | break; | ||
699 | } else { | ||
700 | /* some job is spinning => enqueue in request queue */ | ||
701 | prio_wait_queue_t wait; | ||
702 | wait_queue_head_t* vspin = &__get_cpu_var(mpcpvs_vspin_wait); | ||
703 | unsigned long flags; | ||
704 | |||
705 | /* ordered by regular priority */ | ||
706 | init_prio_waitqueue_entry(&wait, t, prio_point(get_priority(t))); | ||
707 | |||
708 | spin_lock_irqsave(&vspin->lock, flags); | ||
709 | |||
710 | set_task_state(t, TASK_UNINTERRUPTIBLE); | ||
711 | |||
712 | __add_wait_queue_prio_exclusive(vspin, &wait); | ||
713 | |||
714 | spin_unlock_irqrestore(&vspin->lock, flags); | ||
715 | |||
716 | TS_LOCK_SUSPEND; | ||
717 | |||
718 | preempt_enable_no_resched(); | ||
719 | |||
720 | schedule(); | ||
721 | |||
722 | preempt_disable(); | ||
723 | |||
724 | TS_LOCK_RESUME; | ||
725 | /* Recheck if we got it --- some higher-priority process might | ||
726 | * have swooped in. */ | ||
727 | } | ||
728 | } | ||
729 | /* ok, now it is ours */ | ||
730 | } | ||
731 | |||
732 | /* called with preemptions off */ | ||
733 | static void mpcp_vspin_exit(void) | ||
734 | { | ||
735 | struct task_struct* t = current, *next; | ||
736 | unsigned long flags; | ||
737 | wait_queue_head_t* vspin = &__get_cpu_var(mpcpvs_vspin_wait); | ||
738 | |||
739 | BUG_ON(__get_cpu_var(mpcpvs_vspin) != t); | ||
740 | |||
741 | /* no spinning job */ | ||
742 | __get_cpu_var(mpcpvs_vspin) = NULL; | ||
743 | |||
744 | /* see if anyone is waiting for us to stop "spinning" */ | ||
745 | spin_lock_irqsave(&vspin->lock, flags); | ||
746 | next = __waitqueue_remove_first(vspin); | ||
747 | |||
748 | if (next) | ||
749 | wake_up_process(next); | ||
750 | |||
751 | spin_unlock_irqrestore(&vspin->lock, flags); | ||
752 | } | ||
753 | |||
754 | static inline struct mpcp_semaphore* mpcp_from_lock(struct litmus_lock* lock) | ||
755 | { | ||
756 | return container_of(lock, struct mpcp_semaphore, litmus_lock); | ||
757 | } | ||
758 | |||
759 | int pfp_mpcp_lock(struct litmus_lock* l) | ||
760 | { | ||
761 | struct task_struct* t = current; | ||
762 | struct mpcp_semaphore *sem = mpcp_from_lock(l); | ||
763 | prio_wait_queue_t wait; | ||
764 | unsigned long flags; | ||
765 | |||
766 | if (!is_realtime(t)) | ||
767 | return -EPERM; | ||
768 | |||
769 | preempt_disable(); | ||
770 | |||
771 | if (sem->vspin) | ||
772 | mpcp_vspin_enter(); | ||
773 | |||
774 | /* Priority-boost ourself *before* we suspend so that | ||
775 | * our priority is boosted when we resume. Use the priority | ||
776 | * ceiling for the local partition. */ | ||
777 | boost_priority(t, sem->prio_ceiling[get_partition(t)]); | ||
778 | |||
779 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
780 | |||
781 | preempt_enable_no_resched(); | ||
782 | |||
783 | if (sem->owner) { | ||
784 | /* resource is not free => must suspend and wait */ | ||
785 | |||
786 | /* ordered by regular priority */ | ||
787 | init_prio_waitqueue_entry(&wait, t, prio_point(get_priority(t))); | ||
788 | |||
789 | /* FIXME: interruptible would be nice some day */ | ||
790 | set_task_state(t, TASK_UNINTERRUPTIBLE); | ||
791 | |||
792 | __add_wait_queue_prio_exclusive(&sem->wait, &wait); | ||
793 | |||
794 | TS_LOCK_SUSPEND; | ||
795 | |||
796 | /* release lock before sleeping */ | ||
797 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
798 | |||
799 | /* We depend on the FIFO order. Thus, we don't need to recheck | ||
800 | * when we wake up; we are guaranteed to have the lock since | ||
801 | * there is only one wake up per release. | ||
802 | */ | ||
803 | |||
804 | schedule(); | ||
805 | |||
806 | TS_LOCK_RESUME; | ||
807 | |||
808 | /* Since we hold the lock, no other task will change | ||
809 | * ->owner. We can thus check it without acquiring the spin | ||
810 | * lock. */ | ||
811 | BUG_ON(sem->owner != t); | ||
812 | } else { | ||
813 | /* it's ours now */ | ||
814 | sem->owner = t; | ||
815 | |||
816 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
817 | } | ||
818 | |||
819 | return 0; | ||
820 | } | ||
821 | |||
822 | int pfp_mpcp_unlock(struct litmus_lock* l) | ||
823 | { | ||
824 | struct task_struct *t = current, *next; | ||
825 | struct mpcp_semaphore *sem = mpcp_from_lock(l); | ||
826 | unsigned long flags; | ||
827 | int err = 0; | ||
828 | |||
829 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
830 | |||
831 | if (sem->owner != t) { | ||
832 | err = -EINVAL; | ||
833 | goto out; | ||
834 | } | ||
835 | |||
836 | /* we lose the benefit of priority boosting */ | ||
837 | |||
838 | unboost_priority(t); | ||
839 | |||
840 | /* check if there are jobs waiting for this resource */ | ||
841 | next = __waitqueue_remove_first(&sem->wait); | ||
842 | if (next) { | ||
843 | /* next becomes the resouce holder */ | ||
844 | sem->owner = next; | ||
845 | |||
846 | /* Wake up next. The waiting job is already priority-boosted. */ | ||
847 | wake_up_process(next); | ||
848 | } else | ||
849 | /* resource becomes available */ | ||
850 | sem->owner = NULL; | ||
851 | |||
852 | out: | ||
853 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
854 | |||
855 | if (sem->vspin && err == 0) { | ||
856 | preempt_disable(); | ||
857 | mpcp_vspin_exit(); | ||
858 | preempt_enable(); | ||
859 | } | ||
860 | |||
861 | return err; | ||
862 | } | ||
863 | |||
864 | int pfp_mpcp_open(struct litmus_lock* l, void* config) | ||
865 | { | ||
866 | struct task_struct *t = current; | ||
867 | struct mpcp_semaphore *sem = mpcp_from_lock(l); | ||
868 | int cpu, local_cpu; | ||
869 | unsigned long flags; | ||
870 | |||
871 | if (!is_realtime(t)) | ||
872 | /* we need to know the real-time priority */ | ||
873 | return -EPERM; | ||
874 | |||
875 | local_cpu = get_partition(t); | ||
876 | |||
877 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
878 | |||
879 | for (cpu = 0; cpu < NR_CPUS; cpu++) | ||
880 | if (cpu != local_cpu) | ||
881 | { | ||
882 | sem->prio_ceiling[cpu] = min(sem->prio_ceiling[cpu], | ||
883 | get_priority(t)); | ||
884 | TRACE_CUR("priority ceiling for sem %p is now %d on cpu %d\n", | ||
885 | sem, sem->prio_ceiling[cpu], cpu); | ||
886 | } | ||
887 | |||
888 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
889 | |||
890 | return 0; | ||
891 | } | ||
892 | |||
893 | int pfp_mpcp_close(struct litmus_lock* l) | ||
894 | { | ||
895 | struct task_struct *t = current; | ||
896 | struct mpcp_semaphore *sem = mpcp_from_lock(l); | ||
897 | unsigned long flags; | ||
898 | |||
899 | int owner; | ||
900 | |||
901 | spin_lock_irqsave(&sem->wait.lock, flags); | ||
902 | |||
903 | owner = sem->owner == t; | ||
904 | |||
905 | spin_unlock_irqrestore(&sem->wait.lock, flags); | ||
906 | |||
907 | if (owner) | ||
908 | pfp_mpcp_unlock(l); | ||
909 | |||
910 | return 0; | ||
911 | } | ||
912 | |||
913 | void pfp_mpcp_free(struct litmus_lock* lock) | ||
914 | { | ||
915 | kfree(mpcp_from_lock(lock)); | ||
916 | } | ||
917 | |||
918 | static struct litmus_lock_ops pfp_mpcp_lock_ops = { | ||
919 | .close = pfp_mpcp_close, | ||
920 | .lock = pfp_mpcp_lock, | ||
921 | .open = pfp_mpcp_open, | ||
922 | .unlock = pfp_mpcp_unlock, | ||
923 | .deallocate = pfp_mpcp_free, | ||
924 | }; | ||
925 | |||
926 | static struct litmus_lock* pfp_new_mpcp(int vspin) | ||
927 | { | ||
928 | struct mpcp_semaphore* sem; | ||
929 | int cpu; | ||
930 | |||
931 | sem = kmalloc(sizeof(*sem), GFP_KERNEL); | ||
932 | if (!sem) | ||
933 | return NULL; | ||
934 | |||
935 | sem->owner = NULL; | ||
936 | init_waitqueue_head(&sem->wait); | ||
937 | sem->litmus_lock.ops = &pfp_mpcp_lock_ops; | ||
938 | |||
939 | for (cpu = 0; cpu < NR_CPUS; cpu++) | ||
940 | sem->prio_ceiling[cpu] = OMEGA_CEILING; | ||
941 | |||
942 | /* mark as virtual spinning */ | ||
943 | sem->vspin = vspin; | ||
944 | |||
945 | return &sem->litmus_lock; | ||
946 | } | ||
947 | |||
948 | |||
949 | /* ******************** PCP support ********************** */ | ||
950 | |||
951 | |||
952 | struct pcp_semaphore { | ||
953 | struct litmus_lock litmus_lock; | ||
954 | |||
955 | struct list_head ceiling; | ||
956 | |||
957 | /* current resource holder */ | ||
958 | struct task_struct *owner; | ||
959 | |||
960 | /* priority ceiling --- can be negative due to DPCP support */ | ||
961 | int prio_ceiling; | ||
962 | |||
963 | /* on which processor is this PCP semaphore allocated? */ | ||
964 | int on_cpu; | ||
965 | }; | ||
966 | |||
967 | static inline struct pcp_semaphore* pcp_from_lock(struct litmus_lock* lock) | ||
968 | { | ||
969 | return container_of(lock, struct pcp_semaphore, litmus_lock); | ||
970 | } | ||
971 | |||
972 | |||
973 | struct pcp_state { | ||
974 | struct list_head system_ceiling; | ||
975 | |||
976 | /* highest-priority waiting task */ | ||
977 | struct task_struct* hp_waiter; | ||
978 | |||
979 | /* list of jobs waiting to get past the system ceiling */ | ||
980 | wait_queue_head_t ceiling_blocked; | ||
981 | }; | ||
982 | |||
983 | static void pcp_init_state(struct pcp_state* s) | ||
984 | { | ||
985 | INIT_LIST_HEAD(&s->system_ceiling); | ||
986 | s->hp_waiter = NULL; | ||
987 | init_waitqueue_head(&s->ceiling_blocked); | ||
988 | } | ||
989 | |||
990 | static DEFINE_PER_CPU(struct pcp_state, pcp_state); | ||
991 | |||
992 | /* assumes preemptions are off */ | ||
993 | static struct pcp_semaphore* pcp_get_ceiling(void) | ||
994 | { | ||
995 | struct list_head* top = __get_cpu_var(pcp_state).system_ceiling.next; | ||
996 | |||
997 | if (top) | ||
998 | return list_entry(top, struct pcp_semaphore, ceiling); | ||
999 | else | ||
1000 | return NULL; | ||
1001 | } | ||
1002 | |||
1003 | /* assumes preempt off */ | ||
1004 | static void pcp_add_ceiling(struct pcp_semaphore* sem) | ||
1005 | { | ||
1006 | struct list_head *pos; | ||
1007 | struct list_head *in_use = &__get_cpu_var(pcp_state).system_ceiling; | ||
1008 | struct pcp_semaphore* held; | ||
1009 | |||
1010 | BUG_ON(sem->on_cpu != smp_processor_id()); | ||
1011 | BUG_ON(in_list(&sem->ceiling)); | ||
1012 | |||
1013 | list_for_each(pos, in_use) { | ||
1014 | held = list_entry(pos, struct pcp_semaphore, ceiling); | ||
1015 | if (held->prio_ceiling >= sem->prio_ceiling) { | ||
1016 | __list_add(&sem->ceiling, pos->prev, pos); | ||
1017 | return; | ||
1018 | } | ||
1019 | } | ||
1020 | |||
1021 | /* we hit the end of the list */ | ||
1022 | |||
1023 | list_add_tail(&sem->ceiling, in_use); | ||
1024 | } | ||
1025 | |||
1026 | /* assumes preempt off */ | ||
1027 | static int pcp_exceeds_ceiling(struct pcp_semaphore* ceiling, | ||
1028 | struct task_struct* task, | ||
1029 | int effective_prio) | ||
1030 | { | ||
1031 | return ceiling == NULL || | ||
1032 | ceiling->prio_ceiling > effective_prio || | ||
1033 | ceiling->owner == task; | ||
1034 | } | ||
1035 | |||
1036 | /* assumes preempt off */ | ||
1037 | static void pcp_priority_inheritance(void) | ||
1038 | { | ||
1039 | unsigned long flags; | ||
1040 | pfp_domain_t* pfp = local_pfp; | ||
1041 | |||
1042 | struct pcp_semaphore* ceiling = pcp_get_ceiling(); | ||
1043 | struct task_struct *blocker, *blocked; | ||
1044 | |||
1045 | blocker = ceiling ? ceiling->owner : NULL; | ||
1046 | blocked = __get_cpu_var(pcp_state).hp_waiter; | ||
1047 | |||
1048 | raw_spin_lock_irqsave(&pfp->slock, flags); | ||
1049 | |||
1050 | /* Current is no longer inheriting anything by default. This should be | ||
1051 | * the currently scheduled job, and hence not currently queued. */ | ||
1052 | BUG_ON(current != pfp->scheduled); | ||
1053 | |||
1054 | fp_set_prio_inh(pfp, current, NULL); | ||
1055 | fp_set_prio_inh(pfp, blocked, NULL); | ||
1056 | fp_set_prio_inh(pfp, blocker, NULL); | ||
1057 | |||
1058 | |||
1059 | /* Let blocking job inherit priority of blocked job, if required. */ | ||
1060 | if (blocker && blocked && | ||
1061 | fp_higher_prio(blocked, blocker)) { | ||
1062 | TRACE_TASK(blocker, "PCP inherits from %s/%d (prio %u -> %u) \n", | ||
1063 | blocked->comm, blocked->pid, | ||
1064 | get_priority(blocker), get_priority(blocked)); | ||
1065 | fp_set_prio_inh(pfp, blocker, blocked); | ||
1066 | } | ||
1067 | |||
1068 | /* check if anything changed */ | ||
1069 | if (fp_higher_prio(fp_prio_peek(&pfp->ready_queue), pfp->scheduled)) | ||
1070 | preempt(pfp); | ||
1071 | |||
1072 | raw_spin_unlock_irqrestore(&pfp->slock, flags); | ||
1073 | } | ||
1074 | |||
1075 | /* called with preemptions off */ | ||
1076 | static void pcp_raise_ceiling(struct pcp_semaphore* sem, | ||
1077 | int effective_prio) | ||
1078 | { | ||
1079 | struct task_struct* t = current; | ||
1080 | struct pcp_semaphore* ceiling; | ||
1081 | prio_wait_queue_t wait; | ||
1082 | unsigned int waiting_higher_prio; | ||
1083 | |||
1084 | do { | ||
1085 | ceiling = pcp_get_ceiling(); | ||
1086 | if (pcp_exceeds_ceiling(ceiling, t, effective_prio)) | ||
1087 | break; | ||
1088 | |||
1089 | TRACE_CUR("PCP ceiling-blocked, wanted sem %p, but %s/%d has the ceiling \n", | ||
1090 | sem, ceiling->owner->comm, ceiling->owner->pid); | ||
1091 | |||
1092 | /* we need to wait until the ceiling is lowered */ | ||
1093 | |||
1094 | /* enqueue in priority order */ | ||
1095 | init_prio_waitqueue_entry(&wait, t, prio_point(effective_prio)); | ||
1096 | set_task_state(t, TASK_UNINTERRUPTIBLE); | ||
1097 | waiting_higher_prio = add_wait_queue_prio_exclusive( | ||
1098 | &__get_cpu_var(pcp_state).ceiling_blocked, &wait); | ||
1099 | |||
1100 | if (waiting_higher_prio == 0) { | ||
1101 | TRACE_CUR("PCP new highest-prio waiter => prio inheritance\n"); | ||
1102 | |||
1103 | /* we are the new highest-priority waiting job | ||
1104 | * => update inheritance */ | ||
1105 | __get_cpu_var(pcp_state).hp_waiter = t; | ||
1106 | pcp_priority_inheritance(); | ||
1107 | } | ||
1108 | |||
1109 | TS_LOCK_SUSPEND; | ||
1110 | |||
1111 | preempt_enable_no_resched(); | ||
1112 | schedule(); | ||
1113 | preempt_disable(); | ||
1114 | |||
1115 | /* pcp_resume_unblocked() removed us from wait queue */ | ||
1116 | |||
1117 | TS_LOCK_RESUME; | ||
1118 | } while(1); | ||
1119 | |||
1120 | TRACE_CUR("PCP got the ceiling and sem %p\n", sem); | ||
1121 | |||
1122 | /* We are good to go. The semaphore should be available. */ | ||
1123 | BUG_ON(sem->owner != NULL); | ||
1124 | |||
1125 | sem->owner = t; | ||
1126 | |||
1127 | pcp_add_ceiling(sem); | ||
1128 | } | ||
1129 | |||
1130 | static void pcp_resume_unblocked(void) | ||
1131 | { | ||
1132 | wait_queue_head_t *blocked = &__get_cpu_var(pcp_state).ceiling_blocked; | ||
1133 | unsigned long flags; | ||
1134 | prio_wait_queue_t* q; | ||
1135 | struct task_struct* t = NULL; | ||
1136 | |||
1137 | struct pcp_semaphore* ceiling = pcp_get_ceiling(); | ||
1138 | |||
1139 | spin_lock_irqsave(&blocked->lock, flags); | ||
1140 | |||
1141 | while (waitqueue_active(blocked)) { | ||
1142 | /* check first == highest-priority waiting job */ | ||
1143 | q = list_entry(blocked->task_list.next, | ||
1144 | prio_wait_queue_t, wq.task_list); | ||
1145 | t = (struct task_struct*) q->wq.private; | ||
1146 | |||
1147 | /* can it proceed now? => let it go */ | ||
1148 | if (pcp_exceeds_ceiling(ceiling, t, | ||
1149 | prio_from_point(q->priority))) { | ||
1150 | __remove_wait_queue(blocked, &q->wq); | ||
1151 | wake_up_process(t); | ||
1152 | } else { | ||
1153 | /* We are done. Update highest-priority waiter. */ | ||
1154 | __get_cpu_var(pcp_state).hp_waiter = t; | ||
1155 | goto out; | ||
1156 | } | ||
1157 | } | ||
1158 | /* If we get here, then there are no more waiting | ||
1159 | * jobs. */ | ||
1160 | __get_cpu_var(pcp_state).hp_waiter = NULL; | ||
1161 | out: | ||
1162 | spin_unlock_irqrestore(&blocked->lock, flags); | ||
1163 | } | ||
1164 | |||
1165 | /* assumes preempt off */ | ||
1166 | static void pcp_lower_ceiling(struct pcp_semaphore* sem) | ||
1167 | { | ||
1168 | BUG_ON(!in_list(&sem->ceiling)); | ||
1169 | BUG_ON(sem->owner != current); | ||
1170 | BUG_ON(sem->on_cpu != smp_processor_id()); | ||
1171 | |||
1172 | /* remove from ceiling list */ | ||
1173 | list_del(&sem->ceiling); | ||
1174 | |||
1175 | /* release */ | ||
1176 | sem->owner = NULL; | ||
1177 | |||
1178 | TRACE_CUR("PCP released sem %p\n", sem); | ||
1179 | |||
1180 | /* Wake up all ceiling-blocked jobs that now pass the ceiling. */ | ||
1181 | pcp_resume_unblocked(); | ||
1182 | |||
1183 | pcp_priority_inheritance(); | ||
1184 | } | ||
1185 | |||
1186 | static void pcp_update_prio_ceiling(struct pcp_semaphore* sem, | ||
1187 | int effective_prio) | ||
1188 | { | ||
1189 | /* This needs to be synchronized on something. | ||
1190 | * Might as well use waitqueue lock for the processor. | ||
1191 | * We assume this happens only before the task set starts execution, | ||
1192 | * (i.e., during initialization), but it may happen on multiple processors | ||
1193 | * at the same time. | ||
1194 | */ | ||
1195 | unsigned long flags; | ||
1196 | |||
1197 | struct pcp_state* s = &per_cpu(pcp_state, sem->on_cpu); | ||
1198 | |||
1199 | spin_lock_irqsave(&s->ceiling_blocked.lock, flags); | ||
1200 | |||
1201 | sem->prio_ceiling = min(sem->prio_ceiling, effective_prio); | ||
1202 | |||
1203 | spin_unlock_irqrestore(&s->ceiling_blocked.lock, flags); | ||
1204 | } | ||
1205 | |||
1206 | static void pcp_init_semaphore(struct pcp_semaphore* sem, int cpu) | ||
1207 | { | ||
1208 | sem->owner = NULL; | ||
1209 | INIT_LIST_HEAD(&sem->ceiling); | ||
1210 | sem->prio_ceiling = INT_MAX; | ||
1211 | sem->on_cpu = cpu; | ||
1212 | } | ||
1213 | |||
1214 | int pfp_pcp_lock(struct litmus_lock* l) | ||
1215 | { | ||
1216 | struct task_struct* t = current; | ||
1217 | struct pcp_semaphore *sem = pcp_from_lock(l); | ||
1218 | |||
1219 | int eprio = effective_agent_priority(get_priority(t)); | ||
1220 | int from = get_partition(t); | ||
1221 | int to = sem->on_cpu; | ||
1222 | |||
1223 | if (!is_realtime(t) || from != to) | ||
1224 | return -EPERM; | ||
1225 | |||
1226 | preempt_disable(); | ||
1227 | |||
1228 | pcp_raise_ceiling(sem, eprio); | ||
1229 | |||
1230 | preempt_enable(); | ||
1231 | |||
1232 | return 0; | ||
1233 | } | ||
1234 | |||
1235 | int pfp_pcp_unlock(struct litmus_lock* l) | ||
1236 | { | ||
1237 | struct task_struct *t = current; | ||
1238 | struct pcp_semaphore *sem = pcp_from_lock(l); | ||
1239 | |||
1240 | int err = 0; | ||
1241 | |||
1242 | preempt_disable(); | ||
1243 | |||
1244 | if (sem->on_cpu != smp_processor_id() || sem->owner != t) { | ||
1245 | err = -EINVAL; | ||
1246 | goto out; | ||
1247 | } | ||
1248 | |||
1249 | /* give it back */ | ||
1250 | pcp_lower_ceiling(sem); | ||
1251 | |||
1252 | out: | ||
1253 | preempt_enable(); | ||
1254 | |||
1255 | return err; | ||
1256 | } | ||
1257 | |||
1258 | int pfp_pcp_open(struct litmus_lock* l, void* __user config) | ||
1259 | { | ||
1260 | struct task_struct *t = current; | ||
1261 | struct pcp_semaphore *sem = pcp_from_lock(l); | ||
1262 | |||
1263 | int cpu, eprio; | ||
1264 | |||
1265 | if (!is_realtime(t)) | ||
1266 | /* we need to know the real-time priority */ | ||
1267 | return -EPERM; | ||
1268 | |||
1269 | if (get_user(cpu, (int*) config)) | ||
1270 | return -EFAULT; | ||
1271 | |||
1272 | /* make sure the resource location matches */ | ||
1273 | if (cpu != sem->on_cpu) | ||
1274 | return -EINVAL; | ||
1275 | |||
1276 | eprio = effective_agent_priority(get_priority(t)); | ||
1277 | |||
1278 | pcp_update_prio_ceiling(sem, eprio); | ||
1279 | |||
1280 | return 0; | ||
1281 | } | ||
1282 | |||
1283 | int pfp_pcp_close(struct litmus_lock* l) | ||
1284 | { | ||
1285 | struct task_struct *t = current; | ||
1286 | struct pcp_semaphore *sem = pcp_from_lock(l); | ||
1287 | |||
1288 | int owner = 0; | ||
1289 | |||
1290 | preempt_disable(); | ||
1291 | |||
1292 | if (sem->on_cpu == smp_processor_id()) | ||
1293 | owner = sem->owner == t; | ||
1294 | |||
1295 | preempt_enable(); | ||
1296 | |||
1297 | if (owner) | ||
1298 | pfp_pcp_unlock(l); | ||
1299 | |||
1300 | return 0; | ||
1301 | } | ||
1302 | |||
1303 | void pfp_pcp_free(struct litmus_lock* lock) | ||
1304 | { | ||
1305 | kfree(pcp_from_lock(lock)); | ||
1306 | } | ||
1307 | |||
1308 | |||
1309 | static struct litmus_lock_ops pfp_pcp_lock_ops = { | ||
1310 | .close = pfp_pcp_close, | ||
1311 | .lock = pfp_pcp_lock, | ||
1312 | .open = pfp_pcp_open, | ||
1313 | .unlock = pfp_pcp_unlock, | ||
1314 | .deallocate = pfp_pcp_free, | ||
1315 | }; | ||
1316 | |||
1317 | |||
1318 | static struct litmus_lock* pfp_new_pcp(int on_cpu) | ||
1319 | { | ||
1320 | struct pcp_semaphore* sem; | ||
1321 | |||
1322 | sem = kmalloc(sizeof(*sem), GFP_KERNEL); | ||
1323 | if (!sem) | ||
1324 | return NULL; | ||
1325 | |||
1326 | sem->litmus_lock.ops = &pfp_pcp_lock_ops; | ||
1327 | pcp_init_semaphore(sem, on_cpu); | ||
1328 | |||
1329 | return &sem->litmus_lock; | ||
1330 | } | ||
1331 | |||
1332 | /* ******************** DPCP support ********************** */ | ||
1333 | |||
1334 | struct dpcp_semaphore { | ||
1335 | struct litmus_lock litmus_lock; | ||
1336 | struct pcp_semaphore pcp; | ||
1337 | int owner_cpu; | ||
1338 | }; | ||
1339 | |||
1340 | static inline struct dpcp_semaphore* dpcp_from_lock(struct litmus_lock* lock) | ||
1341 | { | ||
1342 | return container_of(lock, struct dpcp_semaphore, litmus_lock); | ||
1343 | } | ||
1344 | |||
1345 | /* called with preemptions disabled */ | ||
1346 | static void pfp_migrate_to(int target_cpu) | ||
1347 | { | ||
1348 | struct task_struct* t = current; | ||
1349 | pfp_domain_t *from; | ||
1350 | |||
1351 | if (get_partition(t) == target_cpu) | ||
1352 | return; | ||
1353 | |||
1354 | /* make sure target_cpu makes sense */ | ||
1355 | BUG_ON(!cpu_online(target_cpu)); | ||
1356 | |||
1357 | local_irq_disable(); | ||
1358 | |||
1359 | /* scheduled task should not be in any ready or release queue */ | ||
1360 | BUG_ON(is_queued(t)); | ||
1361 | |||
1362 | /* lock both pfp domains in order of address */ | ||
1363 | from = task_pfp(t); | ||
1364 | |||
1365 | raw_spin_lock(&from->slock); | ||
1366 | |||
1367 | /* switch partitions */ | ||
1368 | tsk_rt(t)->task_params.cpu = target_cpu; | ||
1369 | |||
1370 | raw_spin_unlock(&from->slock); | ||
1371 | |||
1372 | /* Don't trace scheduler costs as part of | ||
1373 | * locking overhead. Scheduling costs are accounted for | ||
1374 | * explicitly. */ | ||
1375 | TS_LOCK_SUSPEND; | ||
1376 | |||
1377 | local_irq_enable(); | ||
1378 | preempt_enable_no_resched(); | ||
1379 | |||
1380 | /* deschedule to be migrated */ | ||
1381 | schedule(); | ||
1382 | |||
1383 | /* we are now on the target processor */ | ||
1384 | preempt_disable(); | ||
1385 | |||
1386 | /* start recording costs again */ | ||
1387 | TS_LOCK_RESUME; | ||
1388 | |||
1389 | BUG_ON(smp_processor_id() != target_cpu); | ||
1390 | } | ||
1391 | |||
1392 | int pfp_dpcp_lock(struct litmus_lock* l) | ||
1393 | { | ||
1394 | struct task_struct* t = current; | ||
1395 | struct dpcp_semaphore *sem = dpcp_from_lock(l); | ||
1396 | int eprio = effective_agent_priority(get_priority(t)); | ||
1397 | int from = get_partition(t); | ||
1398 | int to = sem->pcp.on_cpu; | ||
1399 | |||
1400 | if (!is_realtime(t)) | ||
1401 | return -EPERM; | ||
1402 | |||
1403 | preempt_disable(); | ||
1404 | |||
1405 | /* Priority-boost ourself *before* we suspend so that | ||
1406 | * our priority is boosted when we resume. */ | ||
1407 | |||
1408 | boost_priority(t, get_priority(t)); | ||
1409 | |||
1410 | pfp_migrate_to(to); | ||
1411 | |||
1412 | pcp_raise_ceiling(&sem->pcp, eprio); | ||
1413 | |||
1414 | /* yep, we got it => execute request */ | ||
1415 | sem->owner_cpu = from; | ||
1416 | |||
1417 | preempt_enable(); | ||
1418 | |||
1419 | return 0; | ||
1420 | } | ||
1421 | |||
1422 | int pfp_dpcp_unlock(struct litmus_lock* l) | ||
1423 | { | ||
1424 | struct task_struct *t = current; | ||
1425 | struct dpcp_semaphore *sem = dpcp_from_lock(l); | ||
1426 | int err = 0; | ||
1427 | int home; | ||
1428 | |||
1429 | preempt_disable(); | ||
1430 | |||
1431 | if (sem->pcp.on_cpu != smp_processor_id() || sem->pcp.owner != t) { | ||
1432 | err = -EINVAL; | ||
1433 | goto out; | ||
1434 | } | ||
1435 | |||
1436 | home = sem->owner_cpu; | ||
1437 | |||
1438 | /* give it back */ | ||
1439 | pcp_lower_ceiling(&sem->pcp); | ||
1440 | |||
1441 | /* we lose the benefit of priority boosting */ | ||
1442 | unboost_priority(t); | ||
1443 | |||
1444 | pfp_migrate_to(home); | ||
1445 | |||
1446 | out: | ||
1447 | preempt_enable(); | ||
1448 | |||
1449 | return err; | ||
1450 | } | ||
1451 | |||
1452 | int pfp_dpcp_open(struct litmus_lock* l, void* __user config) | ||
1453 | { | ||
1454 | struct task_struct *t = current; | ||
1455 | struct dpcp_semaphore *sem = dpcp_from_lock(l); | ||
1456 | int cpu, eprio; | ||
1457 | |||
1458 | if (!is_realtime(t)) | ||
1459 | /* we need to know the real-time priority */ | ||
1460 | return -EPERM; | ||
1461 | |||
1462 | if (get_user(cpu, (int*) config)) | ||
1463 | return -EFAULT; | ||
1464 | |||
1465 | /* make sure the resource location matches */ | ||
1466 | if (cpu != sem->pcp.on_cpu) | ||
1467 | return -EINVAL; | ||
1468 | |||
1469 | eprio = effective_agent_priority(get_priority(t)); | ||
1470 | |||
1471 | pcp_update_prio_ceiling(&sem->pcp, eprio); | ||
1472 | |||
1473 | return 0; | ||
1474 | } | ||
1475 | |||
1476 | int pfp_dpcp_close(struct litmus_lock* l) | ||
1477 | { | ||
1478 | struct task_struct *t = current; | ||
1479 | struct dpcp_semaphore *sem = dpcp_from_lock(l); | ||
1480 | int owner = 0; | ||
1481 | |||
1482 | preempt_disable(); | ||
1483 | |||
1484 | if (sem->pcp.on_cpu == smp_processor_id()) | ||
1485 | owner = sem->pcp.owner == t; | ||
1486 | |||
1487 | preempt_enable(); | ||
1488 | |||
1489 | if (owner) | ||
1490 | pfp_dpcp_unlock(l); | ||
1491 | |||
1492 | return 0; | ||
1493 | } | ||
1494 | |||
1495 | void pfp_dpcp_free(struct litmus_lock* lock) | ||
1496 | { | ||
1497 | kfree(dpcp_from_lock(lock)); | ||
1498 | } | ||
1499 | |||
1500 | static struct litmus_lock_ops pfp_dpcp_lock_ops = { | ||
1501 | .close = pfp_dpcp_close, | ||
1502 | .lock = pfp_dpcp_lock, | ||
1503 | .open = pfp_dpcp_open, | ||
1504 | .unlock = pfp_dpcp_unlock, | ||
1505 | .deallocate = pfp_dpcp_free, | ||
1506 | }; | ||
1507 | |||
1508 | static struct litmus_lock* pfp_new_dpcp(int on_cpu) | ||
1509 | { | ||
1510 | struct dpcp_semaphore* sem; | ||
1511 | |||
1512 | sem = kmalloc(sizeof(*sem), GFP_KERNEL); | ||
1513 | if (!sem) | ||
1514 | return NULL; | ||
1515 | |||
1516 | sem->litmus_lock.ops = &pfp_dpcp_lock_ops; | ||
1517 | sem->owner_cpu = NO_CPU; | ||
1518 | pcp_init_semaphore(&sem->pcp, on_cpu); | ||
1519 | |||
1520 | return &sem->litmus_lock; | ||
1521 | } | ||
1522 | |||
1523 | |||
1524 | /* **** lock constructor **** */ | ||
1525 | |||
1526 | |||
1527 | static long pfp_allocate_lock(struct litmus_lock **lock, int type, | ||
1528 | void* __user config) | ||
1529 | { | ||
1530 | int err = -ENXIO, cpu; | ||
1531 | struct srp_semaphore* srp; | ||
1532 | |||
1533 | /* P-FP currently supports the SRP for local resources and the FMLP | ||
1534 | * for global resources. */ | ||
1535 | switch (type) { | ||
1536 | case FMLP_SEM: | ||
1537 | /* FIFO Mutex Locking Protocol */ | ||
1538 | *lock = pfp_new_fmlp(); | ||
1539 | if (*lock) | ||
1540 | err = 0; | ||
1541 | else | ||
1542 | err = -ENOMEM; | ||
1543 | break; | ||
1544 | |||
1545 | case MPCP_SEM: | ||
1546 | /* Multiprocesor Priority Ceiling Protocol */ | ||
1547 | *lock = pfp_new_mpcp(0); | ||
1548 | if (*lock) | ||
1549 | err = 0; | ||
1550 | else | ||
1551 | err = -ENOMEM; | ||
1552 | break; | ||
1553 | |||
1554 | case MPCP_VS_SEM: | ||
1555 | /* Multiprocesor Priority Ceiling Protocol with virtual spinning */ | ||
1556 | *lock = pfp_new_mpcp(1); | ||
1557 | if (*lock) | ||
1558 | err = 0; | ||
1559 | else | ||
1560 | err = -ENOMEM; | ||
1561 | break; | ||
1562 | |||
1563 | case DPCP_SEM: | ||
1564 | /* Distributed Priority Ceiling Protocol */ | ||
1565 | if (get_user(cpu, (int*) config)) | ||
1566 | return -EFAULT; | ||
1567 | |||
1568 | if (!cpu_online(cpu)) | ||
1569 | return -EINVAL; | ||
1570 | |||
1571 | *lock = pfp_new_dpcp(cpu); | ||
1572 | if (*lock) | ||
1573 | err = 0; | ||
1574 | else | ||
1575 | err = -ENOMEM; | ||
1576 | break; | ||
1577 | |||
1578 | case SRP_SEM: | ||
1579 | /* Baker's Stack Resource Policy */ | ||
1580 | srp = allocate_srp_semaphore(); | ||
1581 | if (srp) { | ||
1582 | *lock = &srp->litmus_lock; | ||
1583 | err = 0; | ||
1584 | } else | ||
1585 | err = -ENOMEM; | ||
1586 | break; | ||
1587 | |||
1588 | case PCP_SEM: | ||
1589 | /* Priority Ceiling Protocol */ | ||
1590 | if (get_user(cpu, (int*) config)) | ||
1591 | return -EFAULT; | ||
1592 | |||
1593 | if (!cpu_online(cpu)) | ||
1594 | return -EINVAL; | ||
1595 | |||
1596 | *lock = pfp_new_pcp(cpu); | ||
1597 | if (*lock) | ||
1598 | err = 0; | ||
1599 | else | ||
1600 | err = -ENOMEM; | ||
1601 | break; | ||
1602 | }; | ||
1603 | |||
1604 | return err; | ||
1605 | } | ||
1606 | |||
1607 | #endif | ||
1608 | |||
1609 | static long pfp_admit_task(struct task_struct* tsk) | ||
1610 | { | ||
1611 | if (task_cpu(tsk) == tsk->rt_param.task_params.cpu && | ||
1612 | #ifdef CONFIG_RELEASE_MASTER | ||
1613 | /* don't allow tasks on release master CPU */ | ||
1614 | task_cpu(tsk) != remote_dom(task_cpu(tsk))->release_master && | ||
1615 | #endif | ||
1616 | litmus_is_valid_fixed_prio(get_priority(tsk))) | ||
1617 | return 0; | ||
1618 | else | ||
1619 | return -EINVAL; | ||
1620 | } | ||
1621 | |||
1622 | static long pfp_activate_plugin(void) | ||
1623 | { | ||
1624 | #if defined(CONFIG_RELEASE_MASTER) || defined(CONFIG_LITMUS_LOCKING) | ||
1625 | int cpu; | ||
1626 | #endif | ||
1627 | |||
1628 | #ifdef CONFIG_RELEASE_MASTER | ||
1629 | for_each_online_cpu(cpu) { | ||
1630 | remote_dom(cpu)->release_master = atomic_read(&release_master_cpu); | ||
1631 | } | ||
1632 | #endif | ||
1633 | |||
1634 | #ifdef CONFIG_LITMUS_LOCKING | ||
1635 | get_srp_prio = pfp_get_srp_prio; | ||
1636 | |||
1637 | for_each_online_cpu(cpu) { | ||
1638 | init_waitqueue_head(&per_cpu(mpcpvs_vspin_wait, cpu)); | ||
1639 | per_cpu(mpcpvs_vspin, cpu) = NULL; | ||
1640 | |||
1641 | pcp_init_state(&per_cpu(pcp_state, cpu)); | ||
1642 | pfp_doms[cpu] = remote_pfp(cpu); | ||
1643 | } | ||
1644 | |||
1645 | #endif | ||
1646 | |||
1647 | return 0; | ||
1648 | } | ||
1649 | |||
1650 | |||
1651 | /* Plugin object */ | ||
1652 | static struct sched_plugin pfp_plugin __cacheline_aligned_in_smp = { | ||
1653 | .plugin_name = "P-FP", | ||
1654 | .tick = pfp_tick, | ||
1655 | .task_new = pfp_task_new, | ||
1656 | .complete_job = complete_job, | ||
1657 | .task_exit = pfp_task_exit, | ||
1658 | .schedule = pfp_schedule, | ||
1659 | .task_wake_up = pfp_task_wake_up, | ||
1660 | .task_block = pfp_task_block, | ||
1661 | .admit_task = pfp_admit_task, | ||
1662 | .activate_plugin = pfp_activate_plugin, | ||
1663 | #ifdef CONFIG_LITMUS_LOCKING | ||
1664 | .allocate_lock = pfp_allocate_lock, | ||
1665 | .finish_switch = pfp_finish_switch, | ||
1666 | #endif | ||
1667 | }; | ||
1668 | |||
1669 | |||
1670 | static int __init init_pfp(void) | ||
1671 | { | ||
1672 | int i; | ||
1673 | |||
1674 | /* We do not really want to support cpu hotplug, do we? ;) | ||
1675 | * However, if we are so crazy to do so, | ||
1676 | * we cannot use num_online_cpu() | ||
1677 | */ | ||
1678 | for (i = 0; i < num_online_cpus(); i++) { | ||
1679 | pfp_domain_init(remote_pfp(i), i); | ||
1680 | } | ||
1681 | return register_sched_plugin(&pfp_plugin); | ||
1682 | } | ||
1683 | |||
1684 | module_init(init_pfp); | ||
1685 | |||