1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * padata.c - generic interface to process data streams in parallel |
4 | * |
5 | * See Documentation/core-api/padata.rst for more information. |
6 | * |
7 | * Copyright (C) 2008, 2009 secunet Security Networks AG |
8 | * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com> |
9 | * |
10 | * Copyright (c) 2020 Oracle and/or its affiliates. |
11 | * Author: Daniel Jordan <daniel.m.jordan@oracle.com> |
12 | */ |
13 | |
14 | #include <linux/completion.h> |
15 | #include <linux/export.h> |
16 | #include <linux/cpumask.h> |
17 | #include <linux/err.h> |
18 | #include <linux/cpu.h> |
19 | #include <linux/padata.h> |
20 | #include <linux/mutex.h> |
21 | #include <linux/sched.h> |
22 | #include <linux/slab.h> |
23 | #include <linux/sysfs.h> |
24 | #include <linux/rcupdate.h> |
25 | |
26 | #define PADATA_WORK_ONSTACK 1 /* Work's memory is on stack */ |
27 | |
28 | struct padata_work { |
29 | struct work_struct pw_work; |
30 | struct list_head pw_list; /* padata_free_works linkage */ |
31 | void *pw_data; |
32 | }; |
33 | |
34 | static DEFINE_SPINLOCK(padata_works_lock); |
35 | static struct padata_work *padata_works; |
36 | static LIST_HEAD(padata_free_works); |
37 | |
38 | struct padata_mt_job_state { |
39 | spinlock_t lock; |
40 | struct completion completion; |
41 | struct padata_mt_job *job; |
42 | int nworks; |
43 | int nworks_fini; |
44 | unsigned long chunk_size; |
45 | }; |
46 | |
47 | static void padata_free_pd(struct parallel_data *pd); |
48 | static void __init padata_mt_helper(struct work_struct *work); |
49 | |
50 | static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index) |
51 | { |
52 | int cpu, target_cpu; |
53 | |
54 | target_cpu = cpumask_first(srcp: pd->cpumask.pcpu); |
55 | for (cpu = 0; cpu < cpu_index; cpu++) |
56 | target_cpu = cpumask_next(n: target_cpu, srcp: pd->cpumask.pcpu); |
57 | |
58 | return target_cpu; |
59 | } |
60 | |
61 | static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr) |
62 | { |
63 | /* |
64 | * Hash the sequence numbers to the cpus by taking |
65 | * seq_nr mod. number of cpus in use. |
66 | */ |
67 | int cpu_index = seq_nr % cpumask_weight(srcp: pd->cpumask.pcpu); |
68 | |
69 | return padata_index_to_cpu(pd, cpu_index); |
70 | } |
71 | |
72 | static struct padata_work *padata_work_alloc(void) |
73 | { |
74 | struct padata_work *pw; |
75 | |
76 | lockdep_assert_held(&padata_works_lock); |
77 | |
78 | if (list_empty(head: &padata_free_works)) |
79 | return NULL; /* No more work items allowed to be queued. */ |
80 | |
81 | pw = list_first_entry(&padata_free_works, struct padata_work, pw_list); |
82 | list_del(entry: &pw->pw_list); |
83 | return pw; |
84 | } |
85 | |
86 | /* |
87 | * This function is marked __ref because this function may be optimized in such |
88 | * a way that it directly refers to work_fn's address, which causes modpost to |
89 | * complain when work_fn is marked __init. This scenario was observed with clang |
90 | * LTO, where padata_work_init() was optimized to refer directly to |
91 | * padata_mt_helper() because the calls to padata_work_init() with other work_fn |
92 | * values were eliminated or inlined. |
93 | */ |
94 | static void __ref padata_work_init(struct padata_work *pw, work_func_t work_fn, |
95 | void *data, int flags) |
96 | { |
97 | if (flags & PADATA_WORK_ONSTACK) |
98 | INIT_WORK_ONSTACK(&pw->pw_work, work_fn); |
99 | else |
100 | INIT_WORK(&pw->pw_work, work_fn); |
101 | pw->pw_data = data; |
102 | } |
103 | |
104 | static int __init padata_work_alloc_mt(int nworks, void *data, |
105 | struct list_head *head) |
106 | { |
107 | int i; |
108 | |
109 | spin_lock(lock: &padata_works_lock); |
110 | /* Start at 1 because the current task participates in the job. */ |
111 | for (i = 1; i < nworks; ++i) { |
112 | struct padata_work *pw = padata_work_alloc(); |
113 | |
114 | if (!pw) |
115 | break; |
116 | padata_work_init(pw, work_fn: padata_mt_helper, data, flags: 0); |
117 | list_add(new: &pw->pw_list, head); |
118 | } |
119 | spin_unlock(lock: &padata_works_lock); |
120 | |
121 | return i; |
122 | } |
123 | |
124 | static void padata_work_free(struct padata_work *pw) |
125 | { |
126 | lockdep_assert_held(&padata_works_lock); |
127 | list_add(new: &pw->pw_list, head: &padata_free_works); |
128 | } |
129 | |
130 | static void __init padata_works_free(struct list_head *works) |
131 | { |
132 | struct padata_work *cur, *next; |
133 | |
134 | if (list_empty(head: works)) |
135 | return; |
136 | |
137 | spin_lock(lock: &padata_works_lock); |
138 | list_for_each_entry_safe(cur, next, works, pw_list) { |
139 | list_del(entry: &cur->pw_list); |
140 | padata_work_free(pw: cur); |
141 | } |
142 | spin_unlock(lock: &padata_works_lock); |
143 | } |
144 | |
145 | static void padata_parallel_worker(struct work_struct *parallel_work) |
146 | { |
147 | struct padata_work *pw = container_of(parallel_work, struct padata_work, |
148 | pw_work); |
149 | struct padata_priv *padata = pw->pw_data; |
150 | |
151 | local_bh_disable(); |
152 | padata->parallel(padata); |
153 | spin_lock(lock: &padata_works_lock); |
154 | padata_work_free(pw); |
155 | spin_unlock(lock: &padata_works_lock); |
156 | local_bh_enable(); |
157 | } |
158 | |
159 | /** |
160 | * padata_do_parallel - padata parallelization function |
161 | * |
162 | * @ps: padatashell |
163 | * @padata: object to be parallelized |
164 | * @cb_cpu: pointer to the CPU that the serialization callback function should |
165 | * run on. If it's not in the serial cpumask of @pinst |
166 | * (i.e. cpumask.cbcpu), this function selects a fallback CPU and if |
167 | * none found, returns -EINVAL. |
168 | * |
169 | * The parallelization callback function will run with BHs off. |
170 | * Note: Every object which is parallelized by padata_do_parallel |
171 | * must be seen by padata_do_serial. |
172 | * |
173 | * Return: 0 on success or else negative error code. |
174 | */ |
175 | int padata_do_parallel(struct padata_shell *ps, |
176 | struct padata_priv *padata, int *cb_cpu) |
177 | { |
178 | struct padata_instance *pinst = ps->pinst; |
179 | int i, cpu, cpu_index, err; |
180 | struct parallel_data *pd; |
181 | struct padata_work *pw; |
182 | |
183 | rcu_read_lock_bh(); |
184 | |
185 | pd = rcu_dereference_bh(ps->pd); |
186 | |
187 | err = -EINVAL; |
188 | if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID) |
189 | goto out; |
190 | |
191 | if (!cpumask_test_cpu(cpu: *cb_cpu, cpumask: pd->cpumask.cbcpu)) { |
192 | if (cpumask_empty(srcp: pd->cpumask.cbcpu)) |
193 | goto out; |
194 | |
195 | /* Select an alternate fallback CPU and notify the caller. */ |
196 | cpu_index = *cb_cpu % cpumask_weight(srcp: pd->cpumask.cbcpu); |
197 | |
198 | cpu = cpumask_first(srcp: pd->cpumask.cbcpu); |
199 | for (i = 0; i < cpu_index; i++) |
200 | cpu = cpumask_next(n: cpu, srcp: pd->cpumask.cbcpu); |
201 | |
202 | *cb_cpu = cpu; |
203 | } |
204 | |
205 | err = -EBUSY; |
206 | if ((pinst->flags & PADATA_RESET)) |
207 | goto out; |
208 | |
209 | refcount_inc(r: &pd->refcnt); |
210 | padata->pd = pd; |
211 | padata->cb_cpu = *cb_cpu; |
212 | |
213 | spin_lock(lock: &padata_works_lock); |
214 | padata->seq_nr = ++pd->seq_nr; |
215 | pw = padata_work_alloc(); |
216 | spin_unlock(lock: &padata_works_lock); |
217 | |
218 | if (!pw) { |
219 | /* Maximum works limit exceeded, run in the current task. */ |
220 | padata->parallel(padata); |
221 | } |
222 | |
223 | rcu_read_unlock_bh(); |
224 | |
225 | if (pw) { |
226 | padata_work_init(pw, work_fn: padata_parallel_worker, data: padata, flags: 0); |
227 | queue_work(wq: pinst->parallel_wq, work: &pw->pw_work); |
228 | } |
229 | |
230 | return 0; |
231 | out: |
232 | rcu_read_unlock_bh(); |
233 | |
234 | return err; |
235 | } |
236 | EXPORT_SYMBOL(padata_do_parallel); |
237 | |
238 | /* |
239 | * padata_find_next - Find the next object that needs serialization. |
240 | * |
241 | * Return: |
242 | * * A pointer to the control struct of the next object that needs |
243 | * serialization, if present in one of the percpu reorder queues. |
244 | * * NULL, if the next object that needs serialization will |
245 | * be parallel processed by another cpu and is not yet present in |
246 | * the cpu's reorder queue. |
247 | */ |
248 | static struct padata_priv *padata_find_next(struct parallel_data *pd, |
249 | bool remove_object) |
250 | { |
251 | struct padata_priv *padata; |
252 | struct padata_list *reorder; |
253 | int cpu = pd->cpu; |
254 | |
255 | reorder = per_cpu_ptr(pd->reorder_list, cpu); |
256 | |
257 | spin_lock(lock: &reorder->lock); |
258 | if (list_empty(head: &reorder->list)) { |
259 | spin_unlock(lock: &reorder->lock); |
260 | return NULL; |
261 | } |
262 | |
263 | padata = list_entry(reorder->list.next, struct padata_priv, list); |
264 | |
265 | /* |
266 | * Checks the rare case where two or more parallel jobs have hashed to |
267 | * the same CPU and one of the later ones finishes first. |
268 | */ |
269 | if (padata->seq_nr != pd->processed) { |
270 | spin_unlock(lock: &reorder->lock); |
271 | return NULL; |
272 | } |
273 | |
274 | if (remove_object) { |
275 | list_del_init(entry: &padata->list); |
276 | ++pd->processed; |
277 | pd->cpu = cpumask_next_wrap(n: cpu, mask: pd->cpumask.pcpu, start: -1, wrap: false); |
278 | } |
279 | |
280 | spin_unlock(lock: &reorder->lock); |
281 | return padata; |
282 | } |
283 | |
284 | static void padata_reorder(struct parallel_data *pd) |
285 | { |
286 | struct padata_instance *pinst = pd->ps->pinst; |
287 | int cb_cpu; |
288 | struct padata_priv *padata; |
289 | struct padata_serial_queue *squeue; |
290 | struct padata_list *reorder; |
291 | |
292 | /* |
293 | * We need to ensure that only one cpu can work on dequeueing of |
294 | * the reorder queue the time. Calculating in which percpu reorder |
295 | * queue the next object will arrive takes some time. A spinlock |
296 | * would be highly contended. Also it is not clear in which order |
297 | * the objects arrive to the reorder queues. So a cpu could wait to |
298 | * get the lock just to notice that there is nothing to do at the |
299 | * moment. Therefore we use a trylock and let the holder of the lock |
300 | * care for all the objects enqueued during the holdtime of the lock. |
301 | */ |
302 | if (!spin_trylock_bh(lock: &pd->lock)) |
303 | return; |
304 | |
305 | while (1) { |
306 | padata = padata_find_next(pd, remove_object: true); |
307 | |
308 | /* |
309 | * If the next object that needs serialization is parallel |
310 | * processed by another cpu and is still on it's way to the |
311 | * cpu's reorder queue, nothing to do for now. |
312 | */ |
313 | if (!padata) |
314 | break; |
315 | |
316 | cb_cpu = padata->cb_cpu; |
317 | squeue = per_cpu_ptr(pd->squeue, cb_cpu); |
318 | |
319 | spin_lock(lock: &squeue->serial.lock); |
320 | list_add_tail(new: &padata->list, head: &squeue->serial.list); |
321 | spin_unlock(lock: &squeue->serial.lock); |
322 | |
323 | queue_work_on(cpu: cb_cpu, wq: pinst->serial_wq, work: &squeue->work); |
324 | } |
325 | |
326 | spin_unlock_bh(lock: &pd->lock); |
327 | |
328 | /* |
329 | * The next object that needs serialization might have arrived to |
330 | * the reorder queues in the meantime. |
331 | * |
332 | * Ensure reorder queue is read after pd->lock is dropped so we see |
333 | * new objects from another task in padata_do_serial. Pairs with |
334 | * smp_mb in padata_do_serial. |
335 | */ |
336 | smp_mb(); |
337 | |
338 | reorder = per_cpu_ptr(pd->reorder_list, pd->cpu); |
339 | if (!list_empty(head: &reorder->list) && padata_find_next(pd, remove_object: false)) |
340 | queue_work(wq: pinst->serial_wq, work: &pd->reorder_work); |
341 | } |
342 | |
343 | static void invoke_padata_reorder(struct work_struct *work) |
344 | { |
345 | struct parallel_data *pd; |
346 | |
347 | local_bh_disable(); |
348 | pd = container_of(work, struct parallel_data, reorder_work); |
349 | padata_reorder(pd); |
350 | local_bh_enable(); |
351 | } |
352 | |
353 | static void padata_serial_worker(struct work_struct *serial_work) |
354 | { |
355 | struct padata_serial_queue *squeue; |
356 | struct parallel_data *pd; |
357 | LIST_HEAD(local_list); |
358 | int cnt; |
359 | |
360 | local_bh_disable(); |
361 | squeue = container_of(serial_work, struct padata_serial_queue, work); |
362 | pd = squeue->pd; |
363 | |
364 | spin_lock(lock: &squeue->serial.lock); |
365 | list_replace_init(old: &squeue->serial.list, new: &local_list); |
366 | spin_unlock(lock: &squeue->serial.lock); |
367 | |
368 | cnt = 0; |
369 | |
370 | while (!list_empty(head: &local_list)) { |
371 | struct padata_priv *padata; |
372 | |
373 | padata = list_entry(local_list.next, |
374 | struct padata_priv, list); |
375 | |
376 | list_del_init(entry: &padata->list); |
377 | |
378 | padata->serial(padata); |
379 | cnt++; |
380 | } |
381 | local_bh_enable(); |
382 | |
383 | if (refcount_sub_and_test(i: cnt, r: &pd->refcnt)) |
384 | padata_free_pd(pd); |
385 | } |
386 | |
387 | /** |
388 | * padata_do_serial - padata serialization function |
389 | * |
390 | * @padata: object to be serialized. |
391 | * |
392 | * padata_do_serial must be called for every parallelized object. |
393 | * The serialization callback function will run with BHs off. |
394 | */ |
395 | void padata_do_serial(struct padata_priv *padata) |
396 | { |
397 | struct parallel_data *pd = padata->pd; |
398 | int hashed_cpu = padata_cpu_hash(pd, seq_nr: padata->seq_nr); |
399 | struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu); |
400 | struct padata_priv *cur; |
401 | struct list_head *pos; |
402 | |
403 | spin_lock(lock: &reorder->lock); |
404 | /* Sort in ascending order of sequence number. */ |
405 | list_for_each_prev(pos, &reorder->list) { |
406 | cur = list_entry(pos, struct padata_priv, list); |
407 | if (cur->seq_nr < padata->seq_nr) |
408 | break; |
409 | } |
410 | list_add(new: &padata->list, head: pos); |
411 | spin_unlock(lock: &reorder->lock); |
412 | |
413 | /* |
414 | * Ensure the addition to the reorder list is ordered correctly |
415 | * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb |
416 | * in padata_reorder. |
417 | */ |
418 | smp_mb(); |
419 | |
420 | padata_reorder(pd); |
421 | } |
422 | EXPORT_SYMBOL(padata_do_serial); |
423 | |
424 | static int padata_setup_cpumasks(struct padata_instance *pinst) |
425 | { |
426 | struct workqueue_attrs *attrs; |
427 | int err; |
428 | |
429 | attrs = alloc_workqueue_attrs(); |
430 | if (!attrs) |
431 | return -ENOMEM; |
432 | |
433 | /* Restrict parallel_wq workers to pd->cpumask.pcpu. */ |
434 | cpumask_copy(dstp: attrs->cpumask, srcp: pinst->cpumask.pcpu); |
435 | err = apply_workqueue_attrs(wq: pinst->parallel_wq, attrs); |
436 | free_workqueue_attrs(attrs); |
437 | |
438 | return err; |
439 | } |
440 | |
441 | static void __init padata_mt_helper(struct work_struct *w) |
442 | { |
443 | struct padata_work *pw = container_of(w, struct padata_work, pw_work); |
444 | struct padata_mt_job_state *ps = pw->pw_data; |
445 | struct padata_mt_job *job = ps->job; |
446 | bool done; |
447 | |
448 | spin_lock(lock: &ps->lock); |
449 | |
450 | while (job->size > 0) { |
451 | unsigned long start, size, end; |
452 | |
453 | start = job->start; |
454 | /* So end is chunk size aligned if enough work remains. */ |
455 | size = roundup(start + 1, ps->chunk_size) - start; |
456 | size = min(size, job->size); |
457 | end = start + size; |
458 | |
459 | job->start = end; |
460 | job->size -= size; |
461 | |
462 | spin_unlock(lock: &ps->lock); |
463 | job->thread_fn(start, end, job->fn_arg); |
464 | spin_lock(lock: &ps->lock); |
465 | } |
466 | |
467 | ++ps->nworks_fini; |
468 | done = (ps->nworks_fini == ps->nworks); |
469 | spin_unlock(lock: &ps->lock); |
470 | |
471 | if (done) |
472 | complete(&ps->completion); |
473 | } |
474 | |
475 | /** |
476 | * padata_do_multithreaded - run a multithreaded job |
477 | * @job: Description of the job. |
478 | * |
479 | * See the definition of struct padata_mt_job for more details. |
480 | */ |
481 | void __init padata_do_multithreaded(struct padata_mt_job *job) |
482 | { |
483 | /* In case threads finish at different times. */ |
484 | static const unsigned long load_balance_factor = 4; |
485 | struct padata_work my_work, *pw; |
486 | struct padata_mt_job_state ps; |
487 | LIST_HEAD(works); |
488 | int nworks; |
489 | |
490 | if (job->size == 0) |
491 | return; |
492 | |
493 | /* Ensure at least one thread when size < min_chunk. */ |
494 | nworks = max(job->size / max(job->min_chunk, job->align), 1ul); |
495 | nworks = min(nworks, job->max_threads); |
496 | |
497 | if (nworks == 1) { |
498 | /* Single thread, no coordination needed, cut to the chase. */ |
499 | job->thread_fn(job->start, job->start + job->size, job->fn_arg); |
500 | return; |
501 | } |
502 | |
503 | spin_lock_init(&ps.lock); |
504 | init_completion(x: &ps.completion); |
505 | ps.job = job; |
506 | ps.nworks = padata_work_alloc_mt(nworks, data: &ps, head: &works); |
507 | ps.nworks_fini = 0; |
508 | |
509 | /* |
510 | * Chunk size is the amount of work a helper does per call to the |
511 | * thread function. Load balance large jobs between threads by |
512 | * increasing the number of chunks, guarantee at least the minimum |
513 | * chunk size from the caller, and honor the caller's alignment. |
514 | */ |
515 | ps.chunk_size = job->size / (ps.nworks * load_balance_factor); |
516 | ps.chunk_size = max(ps.chunk_size, job->min_chunk); |
517 | ps.chunk_size = roundup(ps.chunk_size, job->align); |
518 | |
519 | list_for_each_entry(pw, &works, pw_list) |
520 | queue_work(wq: system_unbound_wq, work: &pw->pw_work); |
521 | |
522 | /* Use the current thread, which saves starting a workqueue worker. */ |
523 | padata_work_init(pw: &my_work, work_fn: padata_mt_helper, data: &ps, PADATA_WORK_ONSTACK); |
524 | padata_mt_helper(w: &my_work.pw_work); |
525 | |
526 | /* Wait for all the helpers to finish. */ |
527 | wait_for_completion(&ps.completion); |
528 | |
529 | destroy_work_on_stack(work: &my_work.pw_work); |
530 | padata_works_free(works: &works); |
531 | } |
532 | |
533 | static void __padata_list_init(struct padata_list *pd_list) |
534 | { |
535 | INIT_LIST_HEAD(list: &pd_list->list); |
536 | spin_lock_init(&pd_list->lock); |
537 | } |
538 | |
539 | /* Initialize all percpu queues used by serial workers */ |
540 | static void padata_init_squeues(struct parallel_data *pd) |
541 | { |
542 | int cpu; |
543 | struct padata_serial_queue *squeue; |
544 | |
545 | for_each_cpu(cpu, pd->cpumask.cbcpu) { |
546 | squeue = per_cpu_ptr(pd->squeue, cpu); |
547 | squeue->pd = pd; |
548 | __padata_list_init(pd_list: &squeue->serial); |
549 | INIT_WORK(&squeue->work, padata_serial_worker); |
550 | } |
551 | } |
552 | |
553 | /* Initialize per-CPU reorder lists */ |
554 | static void padata_init_reorder_list(struct parallel_data *pd) |
555 | { |
556 | int cpu; |
557 | struct padata_list *list; |
558 | |
559 | for_each_cpu(cpu, pd->cpumask.pcpu) { |
560 | list = per_cpu_ptr(pd->reorder_list, cpu); |
561 | __padata_list_init(pd_list: list); |
562 | } |
563 | } |
564 | |
565 | /* Allocate and initialize the internal cpumask dependend resources. */ |
566 | static struct parallel_data *padata_alloc_pd(struct padata_shell *ps) |
567 | { |
568 | struct padata_instance *pinst = ps->pinst; |
569 | struct parallel_data *pd; |
570 | |
571 | pd = kzalloc(size: sizeof(struct parallel_data), GFP_KERNEL); |
572 | if (!pd) |
573 | goto err; |
574 | |
575 | pd->reorder_list = alloc_percpu(struct padata_list); |
576 | if (!pd->reorder_list) |
577 | goto err_free_pd; |
578 | |
579 | pd->squeue = alloc_percpu(struct padata_serial_queue); |
580 | if (!pd->squeue) |
581 | goto err_free_reorder_list; |
582 | |
583 | pd->ps = ps; |
584 | |
585 | if (!alloc_cpumask_var(mask: &pd->cpumask.pcpu, GFP_KERNEL)) |
586 | goto err_free_squeue; |
587 | if (!alloc_cpumask_var(mask: &pd->cpumask.cbcpu, GFP_KERNEL)) |
588 | goto err_free_pcpu; |
589 | |
590 | cpumask_and(dstp: pd->cpumask.pcpu, src1p: pinst->cpumask.pcpu, cpu_online_mask); |
591 | cpumask_and(dstp: pd->cpumask.cbcpu, src1p: pinst->cpumask.cbcpu, cpu_online_mask); |
592 | |
593 | padata_init_reorder_list(pd); |
594 | padata_init_squeues(pd); |
595 | pd->seq_nr = -1; |
596 | refcount_set(r: &pd->refcnt, n: 1); |
597 | spin_lock_init(&pd->lock); |
598 | pd->cpu = cpumask_first(srcp: pd->cpumask.pcpu); |
599 | INIT_WORK(&pd->reorder_work, invoke_padata_reorder); |
600 | |
601 | return pd; |
602 | |
603 | err_free_pcpu: |
604 | free_cpumask_var(mask: pd->cpumask.pcpu); |
605 | err_free_squeue: |
606 | free_percpu(pdata: pd->squeue); |
607 | err_free_reorder_list: |
608 | free_percpu(pdata: pd->reorder_list); |
609 | err_free_pd: |
610 | kfree(objp: pd); |
611 | err: |
612 | return NULL; |
613 | } |
614 | |
615 | static void padata_free_pd(struct parallel_data *pd) |
616 | { |
617 | free_cpumask_var(mask: pd->cpumask.pcpu); |
618 | free_cpumask_var(mask: pd->cpumask.cbcpu); |
619 | free_percpu(pdata: pd->reorder_list); |
620 | free_percpu(pdata: pd->squeue); |
621 | kfree(objp: pd); |
622 | } |
623 | |
624 | static void __padata_start(struct padata_instance *pinst) |
625 | { |
626 | pinst->flags |= PADATA_INIT; |
627 | } |
628 | |
629 | static void __padata_stop(struct padata_instance *pinst) |
630 | { |
631 | if (!(pinst->flags & PADATA_INIT)) |
632 | return; |
633 | |
634 | pinst->flags &= ~PADATA_INIT; |
635 | |
636 | synchronize_rcu(); |
637 | } |
638 | |
639 | /* Replace the internal control structure with a new one. */ |
640 | static int padata_replace_one(struct padata_shell *ps) |
641 | { |
642 | struct parallel_data *pd_new; |
643 | |
644 | pd_new = padata_alloc_pd(ps); |
645 | if (!pd_new) |
646 | return -ENOMEM; |
647 | |
648 | ps->opd = rcu_dereference_protected(ps->pd, 1); |
649 | rcu_assign_pointer(ps->pd, pd_new); |
650 | |
651 | return 0; |
652 | } |
653 | |
654 | static int padata_replace(struct padata_instance *pinst) |
655 | { |
656 | struct padata_shell *ps; |
657 | int err = 0; |
658 | |
659 | pinst->flags |= PADATA_RESET; |
660 | |
661 | list_for_each_entry(ps, &pinst->pslist, list) { |
662 | err = padata_replace_one(ps); |
663 | if (err) |
664 | break; |
665 | } |
666 | |
667 | synchronize_rcu(); |
668 | |
669 | list_for_each_entry_continue_reverse(ps, &pinst->pslist, list) |
670 | if (refcount_dec_and_test(r: &ps->opd->refcnt)) |
671 | padata_free_pd(pd: ps->opd); |
672 | |
673 | pinst->flags &= ~PADATA_RESET; |
674 | |
675 | return err; |
676 | } |
677 | |
678 | /* If cpumask contains no active cpu, we mark the instance as invalid. */ |
679 | static bool padata_validate_cpumask(struct padata_instance *pinst, |
680 | const struct cpumask *cpumask) |
681 | { |
682 | if (!cpumask_intersects(src1p: cpumask, cpu_online_mask)) { |
683 | pinst->flags |= PADATA_INVALID; |
684 | return false; |
685 | } |
686 | |
687 | pinst->flags &= ~PADATA_INVALID; |
688 | return true; |
689 | } |
690 | |
691 | static int __padata_set_cpumasks(struct padata_instance *pinst, |
692 | cpumask_var_t pcpumask, |
693 | cpumask_var_t cbcpumask) |
694 | { |
695 | int valid; |
696 | int err; |
697 | |
698 | valid = padata_validate_cpumask(pinst, cpumask: pcpumask); |
699 | if (!valid) { |
700 | __padata_stop(pinst); |
701 | goto out_replace; |
702 | } |
703 | |
704 | valid = padata_validate_cpumask(pinst, cpumask: cbcpumask); |
705 | if (!valid) |
706 | __padata_stop(pinst); |
707 | |
708 | out_replace: |
709 | cpumask_copy(dstp: pinst->cpumask.pcpu, srcp: pcpumask); |
710 | cpumask_copy(dstp: pinst->cpumask.cbcpu, srcp: cbcpumask); |
711 | |
712 | err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst); |
713 | |
714 | if (valid) |
715 | __padata_start(pinst); |
716 | |
717 | return err; |
718 | } |
719 | |
720 | /** |
721 | * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value |
722 | * equivalent to @cpumask. |
723 | * @pinst: padata instance |
724 | * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding |
725 | * to parallel and serial cpumasks respectively. |
726 | * @cpumask: the cpumask to use |
727 | * |
728 | * Return: 0 on success or negative error code |
729 | */ |
730 | int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type, |
731 | cpumask_var_t cpumask) |
732 | { |
733 | struct cpumask *serial_mask, *parallel_mask; |
734 | int err = -EINVAL; |
735 | |
736 | cpus_read_lock(); |
737 | mutex_lock(&pinst->lock); |
738 | |
739 | switch (cpumask_type) { |
740 | case PADATA_CPU_PARALLEL: |
741 | serial_mask = pinst->cpumask.cbcpu; |
742 | parallel_mask = cpumask; |
743 | break; |
744 | case PADATA_CPU_SERIAL: |
745 | parallel_mask = pinst->cpumask.pcpu; |
746 | serial_mask = cpumask; |
747 | break; |
748 | default: |
749 | goto out; |
750 | } |
751 | |
752 | err = __padata_set_cpumasks(pinst, pcpumask: parallel_mask, cbcpumask: serial_mask); |
753 | |
754 | out: |
755 | mutex_unlock(lock: &pinst->lock); |
756 | cpus_read_unlock(); |
757 | |
758 | return err; |
759 | } |
760 | EXPORT_SYMBOL(padata_set_cpumask); |
761 | |
762 | #ifdef CONFIG_HOTPLUG_CPU |
763 | |
764 | static int __padata_add_cpu(struct padata_instance *pinst, int cpu) |
765 | { |
766 | int err = 0; |
767 | |
768 | if (cpumask_test_cpu(cpu, cpu_online_mask)) { |
769 | err = padata_replace(pinst); |
770 | |
771 | if (padata_validate_cpumask(pinst, cpumask: pinst->cpumask.pcpu) && |
772 | padata_validate_cpumask(pinst, cpumask: pinst->cpumask.cbcpu)) |
773 | __padata_start(pinst); |
774 | } |
775 | |
776 | return err; |
777 | } |
778 | |
779 | static int __padata_remove_cpu(struct padata_instance *pinst, int cpu) |
780 | { |
781 | int err = 0; |
782 | |
783 | if (!cpumask_test_cpu(cpu, cpu_online_mask)) { |
784 | if (!padata_validate_cpumask(pinst, cpumask: pinst->cpumask.pcpu) || |
785 | !padata_validate_cpumask(pinst, cpumask: pinst->cpumask.cbcpu)) |
786 | __padata_stop(pinst); |
787 | |
788 | err = padata_replace(pinst); |
789 | } |
790 | |
791 | return err; |
792 | } |
793 | |
794 | static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu) |
795 | { |
796 | return cpumask_test_cpu(cpu, cpumask: pinst->cpumask.pcpu) || |
797 | cpumask_test_cpu(cpu, cpumask: pinst->cpumask.cbcpu); |
798 | } |
799 | |
800 | static int padata_cpu_online(unsigned int cpu, struct hlist_node *node) |
801 | { |
802 | struct padata_instance *pinst; |
803 | int ret; |
804 | |
805 | pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node); |
806 | if (!pinst_has_cpu(pinst, cpu)) |
807 | return 0; |
808 | |
809 | mutex_lock(&pinst->lock); |
810 | ret = __padata_add_cpu(pinst, cpu); |
811 | mutex_unlock(lock: &pinst->lock); |
812 | return ret; |
813 | } |
814 | |
815 | static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node) |
816 | { |
817 | struct padata_instance *pinst; |
818 | int ret; |
819 | |
820 | pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node); |
821 | if (!pinst_has_cpu(pinst, cpu)) |
822 | return 0; |
823 | |
824 | mutex_lock(&pinst->lock); |
825 | ret = __padata_remove_cpu(pinst, cpu); |
826 | mutex_unlock(lock: &pinst->lock); |
827 | return ret; |
828 | } |
829 | |
830 | static enum cpuhp_state hp_online; |
831 | #endif |
832 | |
833 | static void __padata_free(struct padata_instance *pinst) |
834 | { |
835 | #ifdef CONFIG_HOTPLUG_CPU |
836 | cpuhp_state_remove_instance_nocalls(state: CPUHP_PADATA_DEAD, |
837 | node: &pinst->cpu_dead_node); |
838 | cpuhp_state_remove_instance_nocalls(state: hp_online, node: &pinst->cpu_online_node); |
839 | #endif |
840 | |
841 | WARN_ON(!list_empty(&pinst->pslist)); |
842 | |
843 | free_cpumask_var(mask: pinst->cpumask.pcpu); |
844 | free_cpumask_var(mask: pinst->cpumask.cbcpu); |
845 | destroy_workqueue(wq: pinst->serial_wq); |
846 | destroy_workqueue(wq: pinst->parallel_wq); |
847 | kfree(objp: pinst); |
848 | } |
849 | |
850 | #define kobj2pinst(_kobj) \ |
851 | container_of(_kobj, struct padata_instance, kobj) |
852 | #define attr2pentry(_attr) \ |
853 | container_of(_attr, struct padata_sysfs_entry, attr) |
854 | |
855 | static void padata_sysfs_release(struct kobject *kobj) |
856 | { |
857 | struct padata_instance *pinst = kobj2pinst(kobj); |
858 | __padata_free(pinst); |
859 | } |
860 | |
861 | struct padata_sysfs_entry { |
862 | struct attribute attr; |
863 | ssize_t (*show)(struct padata_instance *, struct attribute *, char *); |
864 | ssize_t (*store)(struct padata_instance *, struct attribute *, |
865 | const char *, size_t); |
866 | }; |
867 | |
868 | static ssize_t show_cpumask(struct padata_instance *pinst, |
869 | struct attribute *attr, char *buf) |
870 | { |
871 | struct cpumask *cpumask; |
872 | ssize_t len; |
873 | |
874 | mutex_lock(&pinst->lock); |
875 | if (!strcmp(attr->name, "serial_cpumask" )) |
876 | cpumask = pinst->cpumask.cbcpu; |
877 | else |
878 | cpumask = pinst->cpumask.pcpu; |
879 | |
880 | len = snprintf(buf, PAGE_SIZE, fmt: "%*pb\n" , |
881 | nr_cpu_ids, cpumask_bits(cpumask)); |
882 | mutex_unlock(lock: &pinst->lock); |
883 | return len < PAGE_SIZE ? len : -EINVAL; |
884 | } |
885 | |
886 | static ssize_t store_cpumask(struct padata_instance *pinst, |
887 | struct attribute *attr, |
888 | const char *buf, size_t count) |
889 | { |
890 | cpumask_var_t new_cpumask; |
891 | ssize_t ret; |
892 | int mask_type; |
893 | |
894 | if (!alloc_cpumask_var(mask: &new_cpumask, GFP_KERNEL)) |
895 | return -ENOMEM; |
896 | |
897 | ret = bitmap_parse(buf, buflen: count, cpumask_bits(new_cpumask), |
898 | nr_cpumask_bits); |
899 | if (ret < 0) |
900 | goto out; |
901 | |
902 | mask_type = !strcmp(attr->name, "serial_cpumask" ) ? |
903 | PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL; |
904 | ret = padata_set_cpumask(pinst, mask_type, new_cpumask); |
905 | if (!ret) |
906 | ret = count; |
907 | |
908 | out: |
909 | free_cpumask_var(mask: new_cpumask); |
910 | return ret; |
911 | } |
912 | |
913 | #define PADATA_ATTR_RW(_name, _show_name, _store_name) \ |
914 | static struct padata_sysfs_entry _name##_attr = \ |
915 | __ATTR(_name, 0644, _show_name, _store_name) |
916 | #define PADATA_ATTR_RO(_name, _show_name) \ |
917 | static struct padata_sysfs_entry _name##_attr = \ |
918 | __ATTR(_name, 0400, _show_name, NULL) |
919 | |
920 | PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask); |
921 | PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask); |
922 | |
923 | /* |
924 | * Padata sysfs provides the following objects: |
925 | * serial_cpumask [RW] - cpumask for serial workers |
926 | * parallel_cpumask [RW] - cpumask for parallel workers |
927 | */ |
928 | static struct attribute *padata_default_attrs[] = { |
929 | &serial_cpumask_attr.attr, |
930 | ¶llel_cpumask_attr.attr, |
931 | NULL, |
932 | }; |
933 | ATTRIBUTE_GROUPS(padata_default); |
934 | |
935 | static ssize_t padata_sysfs_show(struct kobject *kobj, |
936 | struct attribute *attr, char *buf) |
937 | { |
938 | struct padata_instance *pinst; |
939 | struct padata_sysfs_entry *pentry; |
940 | ssize_t ret = -EIO; |
941 | |
942 | pinst = kobj2pinst(kobj); |
943 | pentry = attr2pentry(attr); |
944 | if (pentry->show) |
945 | ret = pentry->show(pinst, attr, buf); |
946 | |
947 | return ret; |
948 | } |
949 | |
950 | static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr, |
951 | const char *buf, size_t count) |
952 | { |
953 | struct padata_instance *pinst; |
954 | struct padata_sysfs_entry *pentry; |
955 | ssize_t ret = -EIO; |
956 | |
957 | pinst = kobj2pinst(kobj); |
958 | pentry = attr2pentry(attr); |
959 | if (pentry->show) |
960 | ret = pentry->store(pinst, attr, buf, count); |
961 | |
962 | return ret; |
963 | } |
964 | |
965 | static const struct sysfs_ops padata_sysfs_ops = { |
966 | .show = padata_sysfs_show, |
967 | .store = padata_sysfs_store, |
968 | }; |
969 | |
970 | static const struct kobj_type padata_attr_type = { |
971 | .sysfs_ops = &padata_sysfs_ops, |
972 | .default_groups = padata_default_groups, |
973 | .release = padata_sysfs_release, |
974 | }; |
975 | |
976 | /** |
977 | * padata_alloc - allocate and initialize a padata instance |
978 | * @name: used to identify the instance |
979 | * |
980 | * Return: new instance on success, NULL on error |
981 | */ |
982 | struct padata_instance *padata_alloc(const char *name) |
983 | { |
984 | struct padata_instance *pinst; |
985 | |
986 | pinst = kzalloc(size: sizeof(struct padata_instance), GFP_KERNEL); |
987 | if (!pinst) |
988 | goto err; |
989 | |
990 | pinst->parallel_wq = alloc_workqueue(fmt: "%s_parallel" , flags: WQ_UNBOUND, max_active: 0, |
991 | name); |
992 | if (!pinst->parallel_wq) |
993 | goto err_free_inst; |
994 | |
995 | cpus_read_lock(); |
996 | |
997 | pinst->serial_wq = alloc_workqueue(fmt: "%s_serial" , flags: WQ_MEM_RECLAIM | |
998 | WQ_CPU_INTENSIVE, max_active: 1, name); |
999 | if (!pinst->serial_wq) |
1000 | goto err_put_cpus; |
1001 | |
1002 | if (!alloc_cpumask_var(mask: &pinst->cpumask.pcpu, GFP_KERNEL)) |
1003 | goto err_free_serial_wq; |
1004 | if (!alloc_cpumask_var(mask: &pinst->cpumask.cbcpu, GFP_KERNEL)) { |
1005 | free_cpumask_var(mask: pinst->cpumask.pcpu); |
1006 | goto err_free_serial_wq; |
1007 | } |
1008 | |
1009 | INIT_LIST_HEAD(list: &pinst->pslist); |
1010 | |
1011 | cpumask_copy(dstp: pinst->cpumask.pcpu, cpu_possible_mask); |
1012 | cpumask_copy(dstp: pinst->cpumask.cbcpu, cpu_possible_mask); |
1013 | |
1014 | if (padata_setup_cpumasks(pinst)) |
1015 | goto err_free_masks; |
1016 | |
1017 | __padata_start(pinst); |
1018 | |
1019 | kobject_init(kobj: &pinst->kobj, ktype: &padata_attr_type); |
1020 | mutex_init(&pinst->lock); |
1021 | |
1022 | #ifdef CONFIG_HOTPLUG_CPU |
1023 | cpuhp_state_add_instance_nocalls_cpuslocked(state: hp_online, |
1024 | node: &pinst->cpu_online_node); |
1025 | cpuhp_state_add_instance_nocalls_cpuslocked(state: CPUHP_PADATA_DEAD, |
1026 | node: &pinst->cpu_dead_node); |
1027 | #endif |
1028 | |
1029 | cpus_read_unlock(); |
1030 | |
1031 | return pinst; |
1032 | |
1033 | err_free_masks: |
1034 | free_cpumask_var(mask: pinst->cpumask.pcpu); |
1035 | free_cpumask_var(mask: pinst->cpumask.cbcpu); |
1036 | err_free_serial_wq: |
1037 | destroy_workqueue(wq: pinst->serial_wq); |
1038 | err_put_cpus: |
1039 | cpus_read_unlock(); |
1040 | destroy_workqueue(wq: pinst->parallel_wq); |
1041 | err_free_inst: |
1042 | kfree(objp: pinst); |
1043 | err: |
1044 | return NULL; |
1045 | } |
1046 | EXPORT_SYMBOL(padata_alloc); |
1047 | |
1048 | /** |
1049 | * padata_free - free a padata instance |
1050 | * |
1051 | * @pinst: padata instance to free |
1052 | */ |
1053 | void padata_free(struct padata_instance *pinst) |
1054 | { |
1055 | kobject_put(kobj: &pinst->kobj); |
1056 | } |
1057 | EXPORT_SYMBOL(padata_free); |
1058 | |
1059 | /** |
1060 | * padata_alloc_shell - Allocate and initialize padata shell. |
1061 | * |
1062 | * @pinst: Parent padata_instance object. |
1063 | * |
1064 | * Return: new shell on success, NULL on error |
1065 | */ |
1066 | struct padata_shell *padata_alloc_shell(struct padata_instance *pinst) |
1067 | { |
1068 | struct parallel_data *pd; |
1069 | struct padata_shell *ps; |
1070 | |
1071 | ps = kzalloc(size: sizeof(*ps), GFP_KERNEL); |
1072 | if (!ps) |
1073 | goto out; |
1074 | |
1075 | ps->pinst = pinst; |
1076 | |
1077 | cpus_read_lock(); |
1078 | pd = padata_alloc_pd(ps); |
1079 | cpus_read_unlock(); |
1080 | |
1081 | if (!pd) |
1082 | goto out_free_ps; |
1083 | |
1084 | mutex_lock(&pinst->lock); |
1085 | RCU_INIT_POINTER(ps->pd, pd); |
1086 | list_add(new: &ps->list, head: &pinst->pslist); |
1087 | mutex_unlock(lock: &pinst->lock); |
1088 | |
1089 | return ps; |
1090 | |
1091 | out_free_ps: |
1092 | kfree(objp: ps); |
1093 | out: |
1094 | return NULL; |
1095 | } |
1096 | EXPORT_SYMBOL(padata_alloc_shell); |
1097 | |
1098 | /** |
1099 | * padata_free_shell - free a padata shell |
1100 | * |
1101 | * @ps: padata shell to free |
1102 | */ |
1103 | void padata_free_shell(struct padata_shell *ps) |
1104 | { |
1105 | struct parallel_data *pd; |
1106 | |
1107 | if (!ps) |
1108 | return; |
1109 | |
1110 | mutex_lock(&ps->pinst->lock); |
1111 | list_del(entry: &ps->list); |
1112 | pd = rcu_dereference_protected(ps->pd, 1); |
1113 | if (refcount_dec_and_test(r: &pd->refcnt)) |
1114 | padata_free_pd(pd); |
1115 | mutex_unlock(lock: &ps->pinst->lock); |
1116 | |
1117 | kfree(objp: ps); |
1118 | } |
1119 | EXPORT_SYMBOL(padata_free_shell); |
1120 | |
1121 | void __init padata_init(void) |
1122 | { |
1123 | unsigned int i, possible_cpus; |
1124 | #ifdef CONFIG_HOTPLUG_CPU |
1125 | int ret; |
1126 | |
1127 | ret = cpuhp_setup_state_multi(state: CPUHP_AP_ONLINE_DYN, name: "padata:online" , |
1128 | startup: padata_cpu_online, NULL); |
1129 | if (ret < 0) |
1130 | goto err; |
1131 | hp_online = ret; |
1132 | |
1133 | ret = cpuhp_setup_state_multi(state: CPUHP_PADATA_DEAD, name: "padata:dead" , |
1134 | NULL, teardown: padata_cpu_dead); |
1135 | if (ret < 0) |
1136 | goto remove_online_state; |
1137 | #endif |
1138 | |
1139 | possible_cpus = num_possible_cpus(); |
1140 | padata_works = kmalloc_array(n: possible_cpus, size: sizeof(struct padata_work), |
1141 | GFP_KERNEL); |
1142 | if (!padata_works) |
1143 | goto remove_dead_state; |
1144 | |
1145 | for (i = 0; i < possible_cpus; ++i) |
1146 | list_add(new: &padata_works[i].pw_list, head: &padata_free_works); |
1147 | |
1148 | return; |
1149 | |
1150 | remove_dead_state: |
1151 | #ifdef CONFIG_HOTPLUG_CPU |
1152 | cpuhp_remove_multi_state(state: CPUHP_PADATA_DEAD); |
1153 | remove_online_state: |
1154 | cpuhp_remove_multi_state(state: hp_online); |
1155 | err: |
1156 | #endif |
1157 | pr_warn("padata: initialization failed\n" ); |
1158 | } |
1159 | |