Linux Audio

Check our new training course

Loading...
v6.13.7
   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
  28struct padata_work {
  29	struct work_struct	pw_work;
  30	struct list_head	pw_list;  /* padata_free_works linkage */
  31	void			*pw_data;
  32};
  33
  34static DEFINE_SPINLOCK(padata_works_lock);
  35static struct padata_work *padata_works;
  36static LIST_HEAD(padata_free_works);
  37
  38struct 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
  47static void padata_free_pd(struct parallel_data *pd);
  48static void __init padata_mt_helper(struct work_struct *work);
  49
  50static inline void padata_get_pd(struct parallel_data *pd)
  51{
  52	refcount_inc(&pd->refcnt);
  53}
  54
  55static inline void padata_put_pd_cnt(struct parallel_data *pd, int cnt)
  56{
  57	if (refcount_sub_and_test(cnt, &pd->refcnt))
  58		padata_free_pd(pd);
  59}
  60
  61static inline void padata_put_pd(struct parallel_data *pd)
  62{
  63	padata_put_pd_cnt(pd, 1);
  64}
  65
  66static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
  67{
  68	int cpu, target_cpu;
  69
  70	target_cpu = cpumask_first(pd->cpumask.pcpu);
  71	for (cpu = 0; cpu < cpu_index; cpu++)
  72		target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
  73
  74	return target_cpu;
  75}
  76
  77static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr)
  78{
  79	/*
  80	 * Hash the sequence numbers to the cpus by taking
  81	 * seq_nr mod. number of cpus in use.
  82	 */
  83	int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
  84
  85	return padata_index_to_cpu(pd, cpu_index);
  86}
  87
  88static struct padata_work *padata_work_alloc(void)
  89{
  90	struct padata_work *pw;
  91
  92	lockdep_assert_held(&padata_works_lock);
  93
  94	if (list_empty(&padata_free_works))
  95		return NULL;	/* No more work items allowed to be queued. */
  96
  97	pw = list_first_entry(&padata_free_works, struct padata_work, pw_list);
  98	list_del(&pw->pw_list);
  99	return pw;
 100}
 101
 102/*
 103 * This function is marked __ref because this function may be optimized in such
 104 * a way that it directly refers to work_fn's address, which causes modpost to
 105 * complain when work_fn is marked __init. This scenario was observed with clang
 106 * LTO, where padata_work_init() was optimized to refer directly to
 107 * padata_mt_helper() because the calls to padata_work_init() with other work_fn
 108 * values were eliminated or inlined.
 109 */
 110static void __ref padata_work_init(struct padata_work *pw, work_func_t work_fn,
 111				   void *data, int flags)
 112{
 113	if (flags & PADATA_WORK_ONSTACK)
 114		INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
 115	else
 116		INIT_WORK(&pw->pw_work, work_fn);
 117	pw->pw_data = data;
 118}
 119
 120static int __init padata_work_alloc_mt(int nworks, void *data,
 121				       struct list_head *head)
 122{
 123	int i;
 124
 125	spin_lock_bh(&padata_works_lock);
 126	/* Start at 1 because the current task participates in the job. */
 127	for (i = 1; i < nworks; ++i) {
 128		struct padata_work *pw = padata_work_alloc();
 129
 130		if (!pw)
 131			break;
 132		padata_work_init(pw, padata_mt_helper, data, 0);
 133		list_add(&pw->pw_list, head);
 134	}
 135	spin_unlock_bh(&padata_works_lock);
 136
 137	return i;
 138}
 
 139
 140static void padata_work_free(struct padata_work *pw)
 141{
 142	lockdep_assert_held(&padata_works_lock);
 143	list_add(&pw->pw_list, &padata_free_works);
 144}
 145
 146static void __init padata_works_free(struct list_head *works)
 147{
 148	struct padata_work *cur, *next;
 149
 150	if (list_empty(works))
 151		return;
 152
 153	spin_lock_bh(&padata_works_lock);
 154	list_for_each_entry_safe(cur, next, works, pw_list) {
 155		list_del(&cur->pw_list);
 156		padata_work_free(cur);
 157	}
 158	spin_unlock_bh(&padata_works_lock);
 159}
 160
 161static void padata_parallel_worker(struct work_struct *parallel_work)
 162{
 163	struct padata_work *pw = container_of(parallel_work, struct padata_work,
 164					      pw_work);
 165	struct padata_priv *padata = pw->pw_data;
 166
 167	local_bh_disable();
 168	padata->parallel(padata);
 169	spin_lock(&padata_works_lock);
 170	padata_work_free(pw);
 171	spin_unlock(&padata_works_lock);
 172	local_bh_enable();
 173}
 174
 175/**
 176 * padata_do_parallel - padata parallelization function
 177 *
 178 * @ps: padatashell
 179 * @padata: object to be parallelized
 180 * @cb_cpu: pointer to the CPU that the serialization callback function should
 181 *          run on.  If it's not in the serial cpumask of @pinst
 182 *          (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
 183 *          none found, returns -EINVAL.
 184 *
 185 * The parallelization callback function will run with BHs off.
 186 * Note: Every object which is parallelized by padata_do_parallel
 187 * must be seen by padata_do_serial.
 188 *
 189 * Return: 0 on success or else negative error code.
 190 */
 191int padata_do_parallel(struct padata_shell *ps,
 192		       struct padata_priv *padata, int *cb_cpu)
 193{
 194	struct padata_instance *pinst = ps->pinst;
 195	int i, cpu, cpu_index, err;
 196	struct parallel_data *pd;
 197	struct padata_work *pw;
 198
 199	rcu_read_lock_bh();
 200
 201	pd = rcu_dereference_bh(ps->pd);
 202
 203	err = -EINVAL;
 204	if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
 205		goto out;
 206
 207	if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
 208		if (cpumask_empty(pd->cpumask.cbcpu))
 209			goto out;
 210
 211		/* Select an alternate fallback CPU and notify the caller. */
 212		cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
 213
 214		cpu = cpumask_first(pd->cpumask.cbcpu);
 215		for (i = 0; i < cpu_index; i++)
 216			cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
 217
 218		*cb_cpu = cpu;
 219	}
 220
 221	err = -EBUSY;
 222	if ((pinst->flags & PADATA_RESET))
 223		goto out;
 224
 225	padata_get_pd(pd);
 
 
 
 
 226	padata->pd = pd;
 227	padata->cb_cpu = *cb_cpu;
 228
 229	spin_lock(&padata_works_lock);
 230	padata->seq_nr = ++pd->seq_nr;
 231	pw = padata_work_alloc();
 232	spin_unlock(&padata_works_lock);
 
 
 
 
 233
 234	if (!pw) {
 235		/* Maximum works limit exceeded, run in the current task. */
 236		padata->parallel(padata);
 237	}
 238
 239	rcu_read_unlock_bh();
 240
 241	if (pw) {
 242		padata_work_init(pw, padata_parallel_worker, padata, 0);
 243		queue_work(pinst->parallel_wq, &pw->pw_work);
 244	}
 245
 246	return 0;
 247out:
 248	rcu_read_unlock_bh();
 249
 250	return err;
 251}
 252EXPORT_SYMBOL(padata_do_parallel);
 253
 254/*
 255 * padata_find_next - Find the next object that needs serialization.
 256 *
 257 * Return:
 258 * * A pointer to the control struct of the next object that needs
 259 *   serialization, if present in one of the percpu reorder queues.
 260 * * NULL, if the next object that needs serialization will
 261 *   be parallel processed by another cpu and is not yet present in
 262 *   the cpu's reorder queue.
 
 
 263 */
 264static struct padata_priv *padata_find_next(struct parallel_data *pd,
 265					    bool remove_object)
 266{
 
 267	struct padata_priv *padata;
 268	struct padata_list *reorder;
 269	int cpu = pd->cpu;
 270
 271	reorder = per_cpu_ptr(pd->reorder_list, cpu);
 
 272
 273	spin_lock(&reorder->lock);
 274	if (list_empty(&reorder->list)) {
 275		spin_unlock(&reorder->lock);
 276		return NULL;
 277	}
 278
 279	padata = list_entry(reorder->list.next, struct padata_priv, list);
 280
 281	/*
 282	 * Checks the rare case where two or more parallel jobs have hashed to
 283	 * the same CPU and one of the later ones finishes first.
 284	 */
 285	if (padata->seq_nr != pd->processed) {
 286		spin_unlock(&reorder->lock);
 287		return NULL;
 288	}
 289
 290	if (remove_object) {
 291		list_del_init(&padata->list);
 
 292		++pd->processed;
 293		pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
 294	}
 295
 296	spin_unlock(&reorder->lock);
 297	return padata;
 298}
 299
 300static void padata_reorder(struct parallel_data *pd)
 301{
 302	struct padata_instance *pinst = pd->ps->pinst;
 303	int cb_cpu;
 304	struct padata_priv *padata;
 305	struct padata_serial_queue *squeue;
 306	struct padata_list *reorder;
 
 307
 308	/*
 309	 * We need to ensure that only one cpu can work on dequeueing of
 310	 * the reorder queue the time. Calculating in which percpu reorder
 311	 * queue the next object will arrive takes some time. A spinlock
 312	 * would be highly contended. Also it is not clear in which order
 313	 * the objects arrive to the reorder queues. So a cpu could wait to
 314	 * get the lock just to notice that there is nothing to do at the
 315	 * moment. Therefore we use a trylock and let the holder of the lock
 316	 * care for all the objects enqueued during the holdtime of the lock.
 317	 */
 318	if (!spin_trylock_bh(&pd->lock))
 319		return;
 320
 321	while (1) {
 322		padata = padata_find_next(pd, true);
 323
 324		/*
 325		 * If the next object that needs serialization is parallel
 326		 * processed by another cpu and is still on it's way to the
 327		 * cpu's reorder queue, nothing to do for now.
 328		 */
 329		if (!padata)
 330			break;
 331
 332		cb_cpu = padata->cb_cpu;
 333		squeue = per_cpu_ptr(pd->squeue, cb_cpu);
 334
 335		spin_lock(&squeue->serial.lock);
 336		list_add_tail(&padata->list, &squeue->serial.list);
 337		spin_unlock(&squeue->serial.lock);
 338
 339		queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
 340	}
 341
 342	spin_unlock_bh(&pd->lock);
 343
 344	/*
 345	 * The next object that needs serialization might have arrived to
 346	 * the reorder queues in the meantime.
 347	 *
 348	 * Ensure reorder queue is read after pd->lock is dropped so we see
 349	 * new objects from another task in padata_do_serial.  Pairs with
 350	 * smp_mb in padata_do_serial.
 351	 */
 352	smp_mb();
 353
 354	reorder = per_cpu_ptr(pd->reorder_list, pd->cpu);
 355	if (!list_empty(&reorder->list) && padata_find_next(pd, false)) {
 356		/*
 357		 * Other context(eg. the padata_serial_worker) can finish the request.
 358		 * To avoid UAF issue, add pd ref here, and put pd ref after reorder_work finish.
 359		 */
 360		padata_get_pd(pd);
 361		queue_work(pinst->serial_wq, &pd->reorder_work);
 362	}
 363}
 364
 365static void invoke_padata_reorder(struct work_struct *work)
 366{
 367	struct parallel_data *pd;
 368
 369	local_bh_disable();
 370	pd = container_of(work, struct parallel_data, reorder_work);
 371	padata_reorder(pd);
 372	local_bh_enable();
 373	/* Pairs with putting the reorder_work in the serial_wq */
 374	padata_put_pd(pd);
 375}
 376
 377static void padata_serial_worker(struct work_struct *serial_work)
 378{
 379	struct padata_serial_queue *squeue;
 380	struct parallel_data *pd;
 381	LIST_HEAD(local_list);
 382	int cnt;
 383
 384	local_bh_disable();
 385	squeue = container_of(serial_work, struct padata_serial_queue, work);
 386	pd = squeue->pd;
 387
 388	spin_lock(&squeue->serial.lock);
 389	list_replace_init(&squeue->serial.list, &local_list);
 390	spin_unlock(&squeue->serial.lock);
 391
 392	cnt = 0;
 393
 394	while (!list_empty(&local_list)) {
 395		struct padata_priv *padata;
 396
 397		padata = list_entry(local_list.next,
 398				    struct padata_priv, list);
 399
 400		list_del_init(&padata->list);
 401
 402		padata->serial(padata);
 403		cnt++;
 404	}
 405	local_bh_enable();
 406
 407	padata_put_pd_cnt(pd, cnt);
 408}
 409
 410/**
 411 * padata_do_serial - padata serialization function
 412 *
 413 * @padata: object to be serialized.
 414 *
 415 * padata_do_serial must be called for every parallelized object.
 416 * The serialization callback function will run with BHs off.
 417 */
 418void padata_do_serial(struct padata_priv *padata)
 419{
 420	struct parallel_data *pd = padata->pd;
 421	int hashed_cpu = padata_cpu_hash(pd, padata->seq_nr);
 422	struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu);
 423	struct padata_priv *cur;
 424	struct list_head *pos;
 425
 426	spin_lock(&reorder->lock);
 427	/* Sort in ascending order of sequence number. */
 428	list_for_each_prev(pos, &reorder->list) {
 429		cur = list_entry(pos, struct padata_priv, list);
 430		/* Compare by difference to consider integer wrap around */
 431		if ((signed int)(cur->seq_nr - padata->seq_nr) < 0)
 432			break;
 433	}
 434	list_add(&padata->list, pos);
 435	spin_unlock(&reorder->lock);
 436
 437	/*
 438	 * Ensure the addition to the reorder list is ordered correctly
 439	 * with the trylock of pd->lock in padata_reorder.  Pairs with smp_mb
 440	 * in padata_reorder.
 441	 */
 442	smp_mb();
 443
 444	padata_reorder(pd);
 445}
 446EXPORT_SYMBOL(padata_do_serial);
 447
 448static int padata_setup_cpumasks(struct padata_instance *pinst)
 
 
 449{
 450	struct workqueue_attrs *attrs;
 451	int err;
 
 
 
 
 
 
 
 
 452
 453	attrs = alloc_workqueue_attrs();
 454	if (!attrs)
 455		return -ENOMEM;
 456
 457	/* Restrict parallel_wq workers to pd->cpumask.pcpu. */
 458	cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
 459	err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
 460	free_workqueue_attrs(attrs);
 
 
 461
 462	return err;
 463}
 464
 465static void __init padata_mt_helper(struct work_struct *w)
 466{
 467	struct padata_work *pw = container_of(w, struct padata_work, pw_work);
 468	struct padata_mt_job_state *ps = pw->pw_data;
 469	struct padata_mt_job *job = ps->job;
 470	bool done;
 471
 472	spin_lock(&ps->lock);
 473
 474	while (job->size > 0) {
 475		unsigned long start, size, end;
 476
 477		start = job->start;
 478		/* So end is chunk size aligned if enough work remains. */
 479		size = roundup(start + 1, ps->chunk_size) - start;
 480		size = min(size, job->size);
 481		end = start + size;
 482
 483		job->start = end;
 484		job->size -= size;
 485
 486		spin_unlock(&ps->lock);
 487		job->thread_fn(start, end, job->fn_arg);
 488		spin_lock(&ps->lock);
 489	}
 490
 491	++ps->nworks_fini;
 492	done = (ps->nworks_fini == ps->nworks);
 493	spin_unlock(&ps->lock);
 494
 495	if (done)
 496		complete(&ps->completion);
 497}
 498
 499/**
 500 * padata_do_multithreaded - run a multithreaded job
 501 * @job: Description of the job.
 502 *
 503 * See the definition of struct padata_mt_job for more details.
 504 */
 505void __init padata_do_multithreaded(struct padata_mt_job *job)
 506{
 507	/* In case threads finish at different times. */
 508	static const unsigned long load_balance_factor = 4;
 509	struct padata_work my_work, *pw;
 510	struct padata_mt_job_state ps;
 511	LIST_HEAD(works);
 512	int nworks, nid;
 513	static atomic_t last_used_nid __initdata;
 514
 515	if (job->size == 0)
 516		return;
 517
 518	/* Ensure at least one thread when size < min_chunk. */
 519	nworks = max(job->size / max(job->min_chunk, job->align), 1ul);
 520	nworks = min(nworks, job->max_threads);
 521
 522	if (nworks == 1) {
 523		/* Single thread, no coordination needed, cut to the chase. */
 524		job->thread_fn(job->start, job->start + job->size, job->fn_arg);
 525		return;
 526	}
 527
 528	spin_lock_init(&ps.lock);
 529	init_completion(&ps.completion);
 530	ps.job	       = job;
 531	ps.nworks      = padata_work_alloc_mt(nworks, &ps, &works);
 532	ps.nworks_fini = 0;
 533
 534	/*
 535	 * Chunk size is the amount of work a helper does per call to the
 536	 * thread function.  Load balance large jobs between threads by
 537	 * increasing the number of chunks, guarantee at least the minimum
 538	 * chunk size from the caller, and honor the caller's alignment.
 539	 * Ensure chunk_size is at least 1 to prevent divide-by-0
 540	 * panic in padata_mt_helper().
 541	 */
 542	ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
 543	ps.chunk_size = max(ps.chunk_size, job->min_chunk);
 544	ps.chunk_size = max(ps.chunk_size, 1ul);
 545	ps.chunk_size = roundup(ps.chunk_size, job->align);
 546
 547	list_for_each_entry(pw, &works, pw_list)
 548		if (job->numa_aware) {
 549			int old_node = atomic_read(&last_used_nid);
 550
 551			do {
 552				nid = next_node_in(old_node, node_states[N_CPU]);
 553			} while (!atomic_try_cmpxchg(&last_used_nid, &old_node, nid));
 554			queue_work_node(nid, system_unbound_wq, &pw->pw_work);
 555		} else {
 556			queue_work(system_unbound_wq, &pw->pw_work);
 557		}
 558
 559	/* Use the current thread, which saves starting a workqueue worker. */
 560	padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
 561	padata_mt_helper(&my_work.pw_work);
 562
 563	/* Wait for all the helpers to finish. */
 564	wait_for_completion(&ps.completion);
 565
 566	destroy_work_on_stack(&my_work.pw_work);
 567	padata_works_free(&works);
 
 
 
 
 568}
 569
 570static void __padata_list_init(struct padata_list *pd_list)
 571{
 572	INIT_LIST_HEAD(&pd_list->list);
 573	spin_lock_init(&pd_list->lock);
 574}
 575
 576/* Initialize all percpu queues used by serial workers */
 577static void padata_init_squeues(struct parallel_data *pd)
 578{
 579	int cpu;
 580	struct padata_serial_queue *squeue;
 581
 582	for_each_cpu(cpu, pd->cpumask.cbcpu) {
 583		squeue = per_cpu_ptr(pd->squeue, cpu);
 584		squeue->pd = pd;
 585		__padata_list_init(&squeue->serial);
 586		INIT_WORK(&squeue->work, padata_serial_worker);
 587	}
 588}
 589
 590/* Initialize per-CPU reorder lists */
 591static void padata_init_reorder_list(struct parallel_data *pd)
 592{
 593	int cpu;
 594	struct padata_list *list;
 595
 596	for_each_cpu(cpu, pd->cpumask.pcpu) {
 597		list = per_cpu_ptr(pd->reorder_list, cpu);
 598		__padata_list_init(list);
 
 
 
 
 599	}
 600}
 601
 602/* Allocate and initialize the internal cpumask dependend resources. */
 603static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
 
 
 604{
 605	struct padata_instance *pinst = ps->pinst;
 606	struct parallel_data *pd;
 607
 608	pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
 609	if (!pd)
 610		goto err;
 611
 612	pd->reorder_list = alloc_percpu(struct padata_list);
 613	if (!pd->reorder_list)
 614		goto err_free_pd;
 615
 616	pd->squeue = alloc_percpu(struct padata_serial_queue);
 617	if (!pd->squeue)
 618		goto err_free_reorder_list;
 619
 620	pd->ps = ps;
 621
 622	if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
 
 623		goto err_free_squeue;
 624	if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
 625		goto err_free_pcpu;
 626
 627	cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
 628	cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
 629
 630	padata_init_reorder_list(pd);
 631	padata_init_squeues(pd);
 632	pd->seq_nr = -1;
 633	refcount_set(&pd->refcnt, 1);
 
 634	spin_lock_init(&pd->lock);
 635	pd->cpu = cpumask_first(pd->cpumask.pcpu);
 636	INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
 637
 638	return pd;
 639
 640err_free_pcpu:
 641	free_cpumask_var(pd->cpumask.pcpu);
 642err_free_squeue:
 643	free_percpu(pd->squeue);
 644err_free_reorder_list:
 645	free_percpu(pd->reorder_list);
 646err_free_pd:
 647	kfree(pd);
 648err:
 649	return NULL;
 650}
 651
 652static void padata_free_pd(struct parallel_data *pd)
 653{
 654	free_cpumask_var(pd->cpumask.pcpu);
 655	free_cpumask_var(pd->cpumask.cbcpu);
 656	free_percpu(pd->reorder_list);
 657	free_percpu(pd->squeue);
 658	kfree(pd);
 659}
 660
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 661static void __padata_start(struct padata_instance *pinst)
 662{
 663	pinst->flags |= PADATA_INIT;
 664}
 665
 666static void __padata_stop(struct padata_instance *pinst)
 667{
 668	if (!(pinst->flags & PADATA_INIT))
 669		return;
 670
 671	pinst->flags &= ~PADATA_INIT;
 672
 673	synchronize_rcu();
 674}
 675
 676/* Replace the internal control structure with a new one. */
 677static int padata_replace_one(struct padata_shell *ps)
 678{
 679	struct parallel_data *pd_new;
 680
 681	pd_new = padata_alloc_pd(ps);
 682	if (!pd_new)
 683		return -ENOMEM;
 684
 685	ps->opd = rcu_dereference_protected(ps->pd, 1);
 686	rcu_assign_pointer(ps->pd, pd_new);
 687
 688	return 0;
 689}
 690
 691static int padata_replace(struct padata_instance *pinst)
 
 
 692{
 693	struct padata_shell *ps;
 694	int err = 0;
 695
 696	pinst->flags |= PADATA_RESET;
 697
 698	list_for_each_entry(ps, &pinst->pslist, list) {
 699		err = padata_replace_one(ps);
 700		if (err)
 701			break;
 702	}
 703
 704	synchronize_rcu();
 705
 706	list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
 707		padata_put_pd(ps->opd);
 
 
 
 
 
 
 
 
 
 
 708
 709	pinst->flags &= ~PADATA_RESET;
 
 710
 711	return err;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 712}
 
 
 713
 714/* If cpumask contains no active cpu, we mark the instance as invalid. */
 715static bool padata_validate_cpumask(struct padata_instance *pinst,
 716				    const struct cpumask *cpumask)
 717{
 718	if (!cpumask_intersects(cpumask, cpu_online_mask)) {
 719		pinst->flags |= PADATA_INVALID;
 720		return false;
 721	}
 722
 723	pinst->flags &= ~PADATA_INVALID;
 724	return true;
 725}
 726
 727static int __padata_set_cpumasks(struct padata_instance *pinst,
 728				 cpumask_var_t pcpumask,
 729				 cpumask_var_t cbcpumask)
 730{
 731	int valid;
 732	int err;
 733
 734	valid = padata_validate_cpumask(pinst, pcpumask);
 735	if (!valid) {
 736		__padata_stop(pinst);
 737		goto out_replace;
 738	}
 739
 740	valid = padata_validate_cpumask(pinst, cbcpumask);
 741	if (!valid)
 742		__padata_stop(pinst);
 743
 744out_replace:
 
 
 
 
 745	cpumask_copy(pinst->cpumask.pcpu, pcpumask);
 746	cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 747
 748	err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
 749
 750	if (valid)
 751		__padata_start(pinst);
 752
 753	return err;
 754}
 755
 756/**
 757 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
 758 *                      equivalent to @cpumask.
 
 759 * @pinst: padata instance
 760 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
 761 *                to parallel and serial cpumasks respectively.
 762 * @cpumask: the cpumask to use
 763 *
 764 * Return: 0 on success or negative error code
 765 */
 766int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
 767		       cpumask_var_t cpumask)
 768{
 769	struct cpumask *serial_mask, *parallel_mask;
 770	int err = -EINVAL;
 771
 772	cpus_read_lock();
 773	mutex_lock(&pinst->lock);
 
 774
 775	switch (cpumask_type) {
 776	case PADATA_CPU_PARALLEL:
 777		serial_mask = pinst->cpumask.cbcpu;
 778		parallel_mask = cpumask;
 779		break;
 780	case PADATA_CPU_SERIAL:
 781		parallel_mask = pinst->cpumask.pcpu;
 782		serial_mask = cpumask;
 783		break;
 784	default:
 785		 goto out;
 786	}
 787
 788	err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
 789
 790out:
 
 791	mutex_unlock(&pinst->lock);
 792	cpus_read_unlock();
 793
 794	return err;
 795}
 796EXPORT_SYMBOL(padata_set_cpumask);
 797
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 798#ifdef CONFIG_HOTPLUG_CPU
 799
 800static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 801{
 802	int err = 0;
 803
 804	if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 805		err = padata_replace(pinst);
 
 
 
 
 
 806
 807		if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
 808		    padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 809			__padata_start(pinst);
 810	}
 811
 812	return err;
 813}
 814
 815static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 816{
 817	int err = 0;
 
 
 818
 819	if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
 820		if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
 821		    !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 822			__padata_stop(pinst);
 823
 824		err = padata_replace(pinst);
 
 
 
 
 
 
 
 
 825	}
 826
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 827	return err;
 828}
 
 829
 830static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
 831{
 832	return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
 833		cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
 834}
 835
 836static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
 837{
 838	struct padata_instance *pinst;
 839	int ret;
 840
 841	pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
 842	if (!pinst_has_cpu(pinst, cpu))
 843		return 0;
 844
 845	mutex_lock(&pinst->lock);
 846	ret = __padata_add_cpu(pinst, cpu);
 847	mutex_unlock(&pinst->lock);
 848	return ret;
 849}
 850
 851static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
 852{
 853	struct padata_instance *pinst;
 854	int ret;
 855
 856	pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
 857	if (!pinst_has_cpu(pinst, cpu))
 858		return 0;
 859
 860	mutex_lock(&pinst->lock);
 861	ret = __padata_remove_cpu(pinst, cpu);
 862	mutex_unlock(&pinst->lock);
 863	return ret;
 864}
 865
 866static enum cpuhp_state hp_online;
 867#endif
 868
 869static void __padata_free(struct padata_instance *pinst)
 870{
 871#ifdef CONFIG_HOTPLUG_CPU
 872	cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
 873					    &pinst->cpu_dead_node);
 874	cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
 875#endif
 876
 877	WARN_ON(!list_empty(&pinst->pslist));
 878
 879	free_cpumask_var(pinst->cpumask.pcpu);
 880	free_cpumask_var(pinst->cpumask.cbcpu);
 881	destroy_workqueue(pinst->serial_wq);
 882	destroy_workqueue(pinst->parallel_wq);
 883	kfree(pinst);
 884}
 885
 886#define kobj2pinst(_kobj)					\
 887	container_of(_kobj, struct padata_instance, kobj)
 888#define attr2pentry(_attr)					\
 889	container_of(_attr, struct padata_sysfs_entry, attr)
 890
 891static void padata_sysfs_release(struct kobject *kobj)
 892{
 893	struct padata_instance *pinst = kobj2pinst(kobj);
 894	__padata_free(pinst);
 895}
 896
 897struct padata_sysfs_entry {
 898	struct attribute attr;
 899	ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
 900	ssize_t (*store)(struct padata_instance *, struct attribute *,
 901			 const char *, size_t);
 902};
 903
 904static ssize_t show_cpumask(struct padata_instance *pinst,
 905			    struct attribute *attr,  char *buf)
 906{
 907	struct cpumask *cpumask;
 908	ssize_t len;
 909
 910	mutex_lock(&pinst->lock);
 911	if (!strcmp(attr->name, "serial_cpumask"))
 912		cpumask = pinst->cpumask.cbcpu;
 913	else
 914		cpumask = pinst->cpumask.pcpu;
 915
 916	len = snprintf(buf, PAGE_SIZE, "%*pb\n",
 917		       nr_cpu_ids, cpumask_bits(cpumask));
 918	mutex_unlock(&pinst->lock);
 919	return len < PAGE_SIZE ? len : -EINVAL;
 920}
 921
 922static ssize_t store_cpumask(struct padata_instance *pinst,
 923			     struct attribute *attr,
 924			     const char *buf, size_t count)
 925{
 926	cpumask_var_t new_cpumask;
 927	ssize_t ret;
 928	int mask_type;
 929
 930	if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
 931		return -ENOMEM;
 932
 933	ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
 934			   nr_cpumask_bits);
 935	if (ret < 0)
 936		goto out;
 937
 938	mask_type = !strcmp(attr->name, "serial_cpumask") ?
 939		PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
 940	ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
 941	if (!ret)
 942		ret = count;
 943
 944out:
 945	free_cpumask_var(new_cpumask);
 946	return ret;
 947}
 948
 949#define PADATA_ATTR_RW(_name, _show_name, _store_name)		\
 950	static struct padata_sysfs_entry _name##_attr =		\
 951		__ATTR(_name, 0644, _show_name, _store_name)
 952#define PADATA_ATTR_RO(_name, _show_name)		\
 953	static struct padata_sysfs_entry _name##_attr = \
 954		__ATTR(_name, 0400, _show_name, NULL)
 955
 956PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
 957PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
 958
 959/*
 960 * Padata sysfs provides the following objects:
 961 * serial_cpumask   [RW] - cpumask for serial workers
 962 * parallel_cpumask [RW] - cpumask for parallel workers
 963 */
 964static struct attribute *padata_default_attrs[] = {
 965	&serial_cpumask_attr.attr,
 966	&parallel_cpumask_attr.attr,
 967	NULL,
 968};
 969ATTRIBUTE_GROUPS(padata_default);
 970
 971static ssize_t padata_sysfs_show(struct kobject *kobj,
 972				 struct attribute *attr, char *buf)
 973{
 974	struct padata_instance *pinst;
 975	struct padata_sysfs_entry *pentry;
 976	ssize_t ret = -EIO;
 977
 978	pinst = kobj2pinst(kobj);
 979	pentry = attr2pentry(attr);
 980	if (pentry->show)
 981		ret = pentry->show(pinst, attr, buf);
 982
 983	return ret;
 984}
 985
 986static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
 987				  const char *buf, size_t count)
 988{
 989	struct padata_instance *pinst;
 990	struct padata_sysfs_entry *pentry;
 991	ssize_t ret = -EIO;
 992
 993	pinst = kobj2pinst(kobj);
 994	pentry = attr2pentry(attr);
 995	if (pentry->store)
 996		ret = pentry->store(pinst, attr, buf, count);
 997
 998	return ret;
 999}
1000
1001static const struct sysfs_ops padata_sysfs_ops = {
1002	.show = padata_sysfs_show,
1003	.store = padata_sysfs_store,
1004};
1005
1006static const struct kobj_type padata_attr_type = {
1007	.sysfs_ops = &padata_sysfs_ops,
1008	.default_groups = padata_default_groups,
1009	.release = padata_sysfs_release,
1010};
1011
1012/**
1013 * padata_alloc - allocate and initialize a padata instance
1014 * @name: used to identify the instance
1015 *
1016 * Return: new instance on success, NULL on error
 
 
1017 */
1018struct padata_instance *padata_alloc(const char *name)
 
 
1019{
1020	struct padata_instance *pinst;
 
1021
1022	pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1023	if (!pinst)
1024		goto err;
1025
1026	pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
1027					     name);
1028	if (!pinst->parallel_wq)
1029		goto err_free_inst;
1030
1031	cpus_read_lock();
1032
1033	pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
1034					   WQ_CPU_INTENSIVE, 1, name);
1035	if (!pinst->serial_wq)
1036		goto err_put_cpus;
1037
1038	if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1039		goto err_free_serial_wq;
1040	if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1041		free_cpumask_var(pinst->cpumask.pcpu);
1042		goto err_free_serial_wq;
1043	}
 
 
 
1044
1045	INIT_LIST_HEAD(&pinst->pslist);
 
 
1046
1047	cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1048	cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1049
1050	if (padata_setup_cpumasks(pinst))
1051		goto err_free_masks;
1052
1053	__padata_start(pinst);
1054
 
1055	kobject_init(&pinst->kobj, &padata_attr_type);
1056	mutex_init(&pinst->lock);
1057
1058#ifdef CONFIG_HOTPLUG_CPU
1059	cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1060						    &pinst->cpu_online_node);
1061	cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1062						    &pinst->cpu_dead_node);
1063#endif
1064
1065	cpus_read_unlock();
1066
1067	return pinst;
1068
1069err_free_masks:
1070	free_cpumask_var(pinst->cpumask.pcpu);
1071	free_cpumask_var(pinst->cpumask.cbcpu);
1072err_free_serial_wq:
1073	destroy_workqueue(pinst->serial_wq);
1074err_put_cpus:
1075	cpus_read_unlock();
1076	destroy_workqueue(pinst->parallel_wq);
1077err_free_inst:
1078	kfree(pinst);
1079err:
1080	return NULL;
1081}
1082EXPORT_SYMBOL(padata_alloc);
1083
1084/**
1085 * padata_free - free a padata instance
 
 
1086 *
1087 * @pinst: padata instance to free
1088 */
1089void padata_free(struct padata_instance *pinst)
1090{
1091	kobject_put(&pinst->kobj);
1092}
1093EXPORT_SYMBOL(padata_free);
1094
1095/**
1096 * padata_alloc_shell - Allocate and initialize padata shell.
1097 *
1098 * @pinst: Parent padata_instance object.
1099 *
1100 * Return: new shell on success, NULL on error
1101 */
1102struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1103{
1104	struct parallel_data *pd;
1105	struct padata_shell *ps;
1106
1107	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1108	if (!ps)
1109		goto out;
1110
1111	ps->pinst = pinst;
1112
1113	cpus_read_lock();
1114	pd = padata_alloc_pd(ps);
1115	cpus_read_unlock();
1116
1117	if (!pd)
1118		goto out_free_ps;
1119
1120	mutex_lock(&pinst->lock);
1121	RCU_INIT_POINTER(ps->pd, pd);
1122	list_add(&ps->list, &pinst->pslist);
1123	mutex_unlock(&pinst->lock);
1124
1125	return ps;
1126
1127out_free_ps:
1128	kfree(ps);
1129out:
1130	return NULL;
1131}
1132EXPORT_SYMBOL(padata_alloc_shell);
1133
1134/**
1135 * padata_free_shell - free a padata shell
1136 *
1137 * @ps: padata shell to free
1138 */
1139void padata_free_shell(struct padata_shell *ps)
1140{
1141	struct parallel_data *pd;
1142
1143	if (!ps)
1144		return;
1145
1146	/*
1147	 * Wait for all _do_serial calls to finish to avoid touching
1148	 * freed pd's and ps's.
1149	 */
1150	synchronize_rcu();
1151
1152	mutex_lock(&ps->pinst->lock);
1153	list_del(&ps->list);
1154	pd = rcu_dereference_protected(ps->pd, 1);
1155	padata_put_pd(pd);
1156	mutex_unlock(&ps->pinst->lock);
1157
1158	kfree(ps);
1159}
1160EXPORT_SYMBOL(padata_free_shell);
1161
1162void __init padata_init(void)
1163{
1164	unsigned int i, possible_cpus;
1165#ifdef CONFIG_HOTPLUG_CPU
1166	int ret;
1167
1168	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1169				      padata_cpu_online, NULL);
 
1170	if (ret < 0)
1171		goto err;
1172	hp_online = ret;
 
 
 
1173
1174	ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1175				      NULL, padata_cpu_dead);
1176	if (ret < 0)
1177		goto remove_online_state;
1178#endif
1179
1180	possible_cpus = num_possible_cpus();
1181	padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1182				     GFP_KERNEL);
1183	if (!padata_works)
1184		goto remove_dead_state;
1185
1186	for (i = 0; i < possible_cpus; ++i)
1187		list_add(&padata_works[i].pw_list, &padata_free_works);
1188
1189	return;
1190
1191remove_dead_state:
1192#ifdef CONFIG_HOTPLUG_CPU
1193	cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1194remove_online_state:
1195	cpuhp_remove_multi_state(hp_online);
1196err:
1197#endif
1198	pr_warn("padata: initialization failed\n");
1199}
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * padata.c - generic interface to process data streams in parallel
   4 *
   5 * See Documentation/padata.txt for an api documentation.
   6 *
   7 * Copyright (C) 2008, 2009 secunet Security Networks AG
   8 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
   9 *
  10 * This program is free software; you can redistribute it and/or modify it
  11 * under the terms and conditions of the GNU General Public License,
  12 * version 2, as published by the Free Software Foundation.
  13 *
  14 * This program is distributed in the hope it will be useful, but WITHOUT
  15 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  16 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  17 * more details.
  18 *
  19 * You should have received a copy of the GNU General Public License along with
  20 * this program; if not, write to the Free Software Foundation, Inc.,
  21 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  22 */
  23
 
  24#include <linux/export.h>
  25#include <linux/cpumask.h>
  26#include <linux/err.h>
  27#include <linux/cpu.h>
  28#include <linux/padata.h>
  29#include <linux/mutex.h>
  30#include <linux/sched.h>
  31#include <linux/slab.h>
  32#include <linux/sysfs.h>
  33#include <linux/rcupdate.h>
  34#include <linux/module.h>
  35
  36#define MAX_OBJ_NUM 1000
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  37
  38static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
  39{
  40	int cpu, target_cpu;
  41
  42	target_cpu = cpumask_first(pd->cpumask.pcpu);
  43	for (cpu = 0; cpu < cpu_index; cpu++)
  44		target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
  45
  46	return target_cpu;
  47}
  48
  49static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr)
  50{
  51	/*
  52	 * Hash the sequence numbers to the cpus by taking
  53	 * seq_nr mod. number of cpus in use.
  54	 */
  55	int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
  56
  57	return padata_index_to_cpu(pd, cpu_index);
  58}
  59
  60static void padata_parallel_worker(struct work_struct *parallel_work)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  61{
  62	struct padata_parallel_queue *pqueue;
  63	LIST_HEAD(local_list);
 
 
 
 
  64
  65	local_bh_disable();
  66	pqueue = container_of(parallel_work,
  67			      struct padata_parallel_queue, work);
 
 
 
  68
  69	spin_lock(&pqueue->parallel.lock);
  70	list_replace_init(&pqueue->parallel.list, &local_list);
  71	spin_unlock(&pqueue->parallel.lock);
  72
  73	while (!list_empty(&local_list)) {
  74		struct padata_priv *padata;
 
 
 
  75
  76		padata = list_entry(local_list.next,
  77				    struct padata_priv, list);
 
  78
  79		list_del_init(&padata->list);
 
  80
  81		padata->parallel(padata);
 
 
 
  82	}
 
 
 
 
 
 
 
 
  83
 
 
 
 
 
  84	local_bh_enable();
  85}
  86
  87/**
  88 * padata_do_parallel - padata parallelization function
  89 *
  90 * @pinst: padata instance
  91 * @padata: object to be parallelized
  92 * @cb_cpu: pointer to the CPU that the serialization callback function should
  93 *          run on.  If it's not in the serial cpumask of @pinst
  94 *          (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
  95 *          none found, returns -EINVAL.
  96 *
  97 * The parallelization callback function will run with BHs off.
  98 * Note: Every object which is parallelized by padata_do_parallel
  99 * must be seen by padata_do_serial.
 
 
 100 */
 101int padata_do_parallel(struct padata_instance *pinst,
 102		       struct padata_priv *padata, int *cb_cpu)
 103{
 104	int i, cpu, cpu_index, target_cpu, err;
 105	struct padata_parallel_queue *queue;
 106	struct parallel_data *pd;
 
 107
 108	rcu_read_lock_bh();
 109
 110	pd = rcu_dereference_bh(pinst->pd);
 111
 112	err = -EINVAL;
 113	if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
 114		goto out;
 115
 116	if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
 117		if (!cpumask_weight(pd->cpumask.cbcpu))
 118			goto out;
 119
 120		/* Select an alternate fallback CPU and notify the caller. */
 121		cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
 122
 123		cpu = cpumask_first(pd->cpumask.cbcpu);
 124		for (i = 0; i < cpu_index; i++)
 125			cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
 126
 127		*cb_cpu = cpu;
 128	}
 129
 130	err =  -EBUSY;
 131	if ((pinst->flags & PADATA_RESET))
 132		goto out;
 133
 134	if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
 135		goto out;
 136
 137	err = 0;
 138	atomic_inc(&pd->refcnt);
 139	padata->pd = pd;
 140	padata->cb_cpu = *cb_cpu;
 141
 142	padata->seq_nr = atomic_inc_return(&pd->seq_nr);
 143	target_cpu = padata_cpu_hash(pd, padata->seq_nr);
 144	padata->cpu = target_cpu;
 145	queue = per_cpu_ptr(pd->pqueue, target_cpu);
 146
 147	spin_lock(&queue->parallel.lock);
 148	list_add_tail(&padata->list, &queue->parallel.list);
 149	spin_unlock(&queue->parallel.lock);
 150
 151	queue_work(pinst->parallel_wq, &queue->work);
 
 
 
 
 
 152
 
 
 
 
 
 
 153out:
 154	rcu_read_unlock_bh();
 155
 156	return err;
 157}
 158EXPORT_SYMBOL(padata_do_parallel);
 159
 160/*
 161 * padata_find_next - Find the next object that needs serialization.
 162 *
 163 * Return values are:
 164 *
 165 * A pointer to the control struct of the next object that needs
 166 * serialization, if present in one of the percpu reorder queues.
 167 *
 168 * NULL, if the next object that needs serialization will
 169 *  be parallel processed by another cpu and is not yet present in
 170 *  the cpu's reorder queue.
 171 */
 172static struct padata_priv *padata_find_next(struct parallel_data *pd,
 173					    bool remove_object)
 174{
 175	struct padata_parallel_queue *next_queue;
 176	struct padata_priv *padata;
 177	struct padata_list *reorder;
 178	int cpu = pd->cpu;
 179
 180	next_queue = per_cpu_ptr(pd->pqueue, cpu);
 181	reorder = &next_queue->reorder;
 182
 183	spin_lock(&reorder->lock);
 184	if (list_empty(&reorder->list)) {
 185		spin_unlock(&reorder->lock);
 186		return NULL;
 187	}
 188
 189	padata = list_entry(reorder->list.next, struct padata_priv, list);
 190
 191	/*
 192	 * Checks the rare case where two or more parallel jobs have hashed to
 193	 * the same CPU and one of the later ones finishes first.
 194	 */
 195	if (padata->seq_nr != pd->processed) {
 196		spin_unlock(&reorder->lock);
 197		return NULL;
 198	}
 199
 200	if (remove_object) {
 201		list_del_init(&padata->list);
 202		atomic_dec(&pd->reorder_objects);
 203		++pd->processed;
 204		pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
 205	}
 206
 207	spin_unlock(&reorder->lock);
 208	return padata;
 209}
 210
 211static void padata_reorder(struct parallel_data *pd)
 212{
 
 213	int cb_cpu;
 214	struct padata_priv *padata;
 215	struct padata_serial_queue *squeue;
 216	struct padata_instance *pinst = pd->pinst;
 217	struct padata_parallel_queue *next_queue;
 218
 219	/*
 220	 * We need to ensure that only one cpu can work on dequeueing of
 221	 * the reorder queue the time. Calculating in which percpu reorder
 222	 * queue the next object will arrive takes some time. A spinlock
 223	 * would be highly contended. Also it is not clear in which order
 224	 * the objects arrive to the reorder queues. So a cpu could wait to
 225	 * get the lock just to notice that there is nothing to do at the
 226	 * moment. Therefore we use a trylock and let the holder of the lock
 227	 * care for all the objects enqueued during the holdtime of the lock.
 228	 */
 229	if (!spin_trylock_bh(&pd->lock))
 230		return;
 231
 232	while (1) {
 233		padata = padata_find_next(pd, true);
 234
 235		/*
 236		 * If the next object that needs serialization is parallel
 237		 * processed by another cpu and is still on it's way to the
 238		 * cpu's reorder queue, nothing to do for now.
 239		 */
 240		if (!padata)
 241			break;
 242
 243		cb_cpu = padata->cb_cpu;
 244		squeue = per_cpu_ptr(pd->squeue, cb_cpu);
 245
 246		spin_lock(&squeue->serial.lock);
 247		list_add_tail(&padata->list, &squeue->serial.list);
 248		spin_unlock(&squeue->serial.lock);
 249
 250		queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
 251	}
 252
 253	spin_unlock_bh(&pd->lock);
 254
 255	/*
 256	 * The next object that needs serialization might have arrived to
 257	 * the reorder queues in the meantime.
 258	 *
 259	 * Ensure reorder queue is read after pd->lock is dropped so we see
 260	 * new objects from another task in padata_do_serial.  Pairs with
 261	 * smp_mb__after_atomic in padata_do_serial.
 262	 */
 263	smp_mb();
 264
 265	next_queue = per_cpu_ptr(pd->pqueue, pd->cpu);
 266	if (!list_empty(&next_queue->reorder.list) &&
 267	    padata_find_next(pd, false))
 
 
 
 
 268		queue_work(pinst->serial_wq, &pd->reorder_work);
 
 269}
 270
 271static void invoke_padata_reorder(struct work_struct *work)
 272{
 273	struct parallel_data *pd;
 274
 275	local_bh_disable();
 276	pd = container_of(work, struct parallel_data, reorder_work);
 277	padata_reorder(pd);
 278	local_bh_enable();
 
 
 279}
 280
 281static void padata_serial_worker(struct work_struct *serial_work)
 282{
 283	struct padata_serial_queue *squeue;
 284	struct parallel_data *pd;
 285	LIST_HEAD(local_list);
 
 286
 287	local_bh_disable();
 288	squeue = container_of(serial_work, struct padata_serial_queue, work);
 289	pd = squeue->pd;
 290
 291	spin_lock(&squeue->serial.lock);
 292	list_replace_init(&squeue->serial.list, &local_list);
 293	spin_unlock(&squeue->serial.lock);
 294
 
 
 295	while (!list_empty(&local_list)) {
 296		struct padata_priv *padata;
 297
 298		padata = list_entry(local_list.next,
 299				    struct padata_priv, list);
 300
 301		list_del_init(&padata->list);
 302
 303		padata->serial(padata);
 304		atomic_dec(&pd->refcnt);
 305	}
 306	local_bh_enable();
 
 
 307}
 308
 309/**
 310 * padata_do_serial - padata serialization function
 311 *
 312 * @padata: object to be serialized.
 313 *
 314 * padata_do_serial must be called for every parallelized object.
 315 * The serialization callback function will run with BHs off.
 316 */
 317void padata_do_serial(struct padata_priv *padata)
 318{
 319	struct parallel_data *pd = padata->pd;
 320	struct padata_parallel_queue *pqueue = per_cpu_ptr(pd->pqueue,
 321							   padata->cpu);
 322	struct padata_priv *cur;
 
 323
 324	spin_lock(&pqueue->reorder.lock);
 325	/* Sort in ascending order of sequence number. */
 326	list_for_each_entry_reverse(cur, &pqueue->reorder.list, list)
 327		if (cur->seq_nr < padata->seq_nr)
 
 
 328			break;
 329	list_add(&padata->list, &cur->list);
 330	atomic_inc(&pd->reorder_objects);
 331	spin_unlock(&pqueue->reorder.lock);
 332
 333	/*
 334	 * Ensure the addition to the reorder list is ordered correctly
 335	 * with the trylock of pd->lock in padata_reorder.  Pairs with smp_mb
 336	 * in padata_reorder.
 337	 */
 338	smp_mb__after_atomic();
 339
 340	padata_reorder(pd);
 341}
 342EXPORT_SYMBOL(padata_do_serial);
 343
 344static int padata_setup_cpumasks(struct parallel_data *pd,
 345				 const struct cpumask *pcpumask,
 346				 const struct cpumask *cbcpumask)
 347{
 348	struct workqueue_attrs *attrs;
 349	int err = -ENOMEM;
 350
 351	if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
 352		goto out;
 353	cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
 354
 355	if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
 356		goto free_pcpu_mask;
 357	cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
 358
 359	attrs = alloc_workqueue_attrs();
 360	if (!attrs)
 361		goto free_cbcpu_mask;
 362
 363	/* Restrict parallel_wq workers to pd->cpumask.pcpu. */
 364	cpumask_copy(attrs->cpumask, pd->cpumask.pcpu);
 365	err = apply_workqueue_attrs(pd->pinst->parallel_wq, attrs);
 366	free_workqueue_attrs(attrs);
 367	if (err < 0)
 368		goto free_cbcpu_mask;
 369
 370	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 371
 372free_cbcpu_mask:
 373	free_cpumask_var(pd->cpumask.cbcpu);
 374free_pcpu_mask:
 375	free_cpumask_var(pd->cpumask.pcpu);
 376out:
 377	return err;
 378}
 379
 380static void __padata_list_init(struct padata_list *pd_list)
 381{
 382	INIT_LIST_HEAD(&pd_list->list);
 383	spin_lock_init(&pd_list->lock);
 384}
 385
 386/* Initialize all percpu queues used by serial workers */
 387static void padata_init_squeues(struct parallel_data *pd)
 388{
 389	int cpu;
 390	struct padata_serial_queue *squeue;
 391
 392	for_each_cpu(cpu, pd->cpumask.cbcpu) {
 393		squeue = per_cpu_ptr(pd->squeue, cpu);
 394		squeue->pd = pd;
 395		__padata_list_init(&squeue->serial);
 396		INIT_WORK(&squeue->work, padata_serial_worker);
 397	}
 398}
 399
 400/* Initialize all percpu queues used by parallel workers */
 401static void padata_init_pqueues(struct parallel_data *pd)
 402{
 403	int cpu;
 404	struct padata_parallel_queue *pqueue;
 405
 406	for_each_cpu(cpu, pd->cpumask.pcpu) {
 407		pqueue = per_cpu_ptr(pd->pqueue, cpu);
 408
 409		__padata_list_init(&pqueue->reorder);
 410		__padata_list_init(&pqueue->parallel);
 411		INIT_WORK(&pqueue->work, padata_parallel_worker);
 412		atomic_set(&pqueue->num_obj, 0);
 413	}
 414}
 415
 416/* Allocate and initialize the internal cpumask dependend resources. */
 417static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
 418					     const struct cpumask *pcpumask,
 419					     const struct cpumask *cbcpumask)
 420{
 
 421	struct parallel_data *pd;
 422
 423	pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
 424	if (!pd)
 425		goto err;
 426
 427	pd->pqueue = alloc_percpu(struct padata_parallel_queue);
 428	if (!pd->pqueue)
 429		goto err_free_pd;
 430
 431	pd->squeue = alloc_percpu(struct padata_serial_queue);
 432	if (!pd->squeue)
 433		goto err_free_pqueue;
 
 
 434
 435	pd->pinst = pinst;
 436	if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
 437		goto err_free_squeue;
 
 
 438
 439	padata_init_pqueues(pd);
 
 
 
 440	padata_init_squeues(pd);
 441	atomic_set(&pd->seq_nr, -1);
 442	atomic_set(&pd->reorder_objects, 0);
 443	atomic_set(&pd->refcnt, 0);
 444	spin_lock_init(&pd->lock);
 445	pd->cpu = cpumask_first(pd->cpumask.pcpu);
 446	INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
 447
 448	return pd;
 449
 
 
 450err_free_squeue:
 451	free_percpu(pd->squeue);
 452err_free_pqueue:
 453	free_percpu(pd->pqueue);
 454err_free_pd:
 455	kfree(pd);
 456err:
 457	return NULL;
 458}
 459
 460static void padata_free_pd(struct parallel_data *pd)
 461{
 462	free_cpumask_var(pd->cpumask.pcpu);
 463	free_cpumask_var(pd->cpumask.cbcpu);
 464	free_percpu(pd->pqueue);
 465	free_percpu(pd->squeue);
 466	kfree(pd);
 467}
 468
 469/* Flush all objects out of the padata queues. */
 470static void padata_flush_queues(struct parallel_data *pd)
 471{
 472	int cpu;
 473	struct padata_parallel_queue *pqueue;
 474	struct padata_serial_queue *squeue;
 475
 476	for_each_cpu(cpu, pd->cpumask.pcpu) {
 477		pqueue = per_cpu_ptr(pd->pqueue, cpu);
 478		flush_work(&pqueue->work);
 479	}
 480
 481	if (atomic_read(&pd->reorder_objects))
 482		padata_reorder(pd);
 483
 484	for_each_cpu(cpu, pd->cpumask.cbcpu) {
 485		squeue = per_cpu_ptr(pd->squeue, cpu);
 486		flush_work(&squeue->work);
 487	}
 488
 489	BUG_ON(atomic_read(&pd->refcnt) != 0);
 490}
 491
 492static void __padata_start(struct padata_instance *pinst)
 493{
 494	pinst->flags |= PADATA_INIT;
 495}
 496
 497static void __padata_stop(struct padata_instance *pinst)
 498{
 499	if (!(pinst->flags & PADATA_INIT))
 500		return;
 501
 502	pinst->flags &= ~PADATA_INIT;
 503
 504	synchronize_rcu();
 
 505
 506	get_online_cpus();
 507	padata_flush_queues(pinst->pd);
 508	put_online_cpus();
 
 
 
 
 
 
 
 
 
 
 509}
 510
 511/* Replace the internal control structure with a new one. */
 512static void padata_replace(struct padata_instance *pinst,
 513			   struct parallel_data *pd_new)
 514{
 515	struct parallel_data *pd_old = pinst->pd;
 516	int notification_mask = 0;
 517
 518	pinst->flags |= PADATA_RESET;
 519
 520	rcu_assign_pointer(pinst->pd, pd_new);
 
 
 
 
 521
 522	synchronize_rcu();
 523
 524	if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
 525		notification_mask |= PADATA_CPU_PARALLEL;
 526	if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
 527		notification_mask |= PADATA_CPU_SERIAL;
 528
 529	padata_flush_queues(pd_old);
 530	padata_free_pd(pd_old);
 531
 532	if (notification_mask)
 533		blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
 534					     notification_mask,
 535					     &pd_new->cpumask);
 536
 537	pinst->flags &= ~PADATA_RESET;
 538}
 539
 540/**
 541 * padata_register_cpumask_notifier - Registers a notifier that will be called
 542 *                             if either pcpu or cbcpu or both cpumasks change.
 543 *
 544 * @pinst: A poineter to padata instance
 545 * @nblock: A pointer to notifier block.
 546 */
 547int padata_register_cpumask_notifier(struct padata_instance *pinst,
 548				     struct notifier_block *nblock)
 549{
 550	return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
 551						nblock);
 552}
 553EXPORT_SYMBOL(padata_register_cpumask_notifier);
 554
 555/**
 556 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
 557 *        registered earlier  using padata_register_cpumask_notifier
 558 *
 559 * @pinst: A pointer to data instance.
 560 * @nlock: A pointer to notifier block.
 561 */
 562int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
 563				       struct notifier_block *nblock)
 564{
 565	return blocking_notifier_chain_unregister(
 566		&pinst->cpumask_change_notifier,
 567		nblock);
 568}
 569EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
 570
 571
 572/* If cpumask contains no active cpu, we mark the instance as invalid. */
 573static bool padata_validate_cpumask(struct padata_instance *pinst,
 574				    const struct cpumask *cpumask)
 575{
 576	if (!cpumask_intersects(cpumask, cpu_online_mask)) {
 577		pinst->flags |= PADATA_INVALID;
 578		return false;
 579	}
 580
 581	pinst->flags &= ~PADATA_INVALID;
 582	return true;
 583}
 584
 585static int __padata_set_cpumasks(struct padata_instance *pinst,
 586				 cpumask_var_t pcpumask,
 587				 cpumask_var_t cbcpumask)
 588{
 589	int valid;
 590	struct parallel_data *pd;
 591
 592	valid = padata_validate_cpumask(pinst, pcpumask);
 593	if (!valid) {
 594		__padata_stop(pinst);
 595		goto out_replace;
 596	}
 597
 598	valid = padata_validate_cpumask(pinst, cbcpumask);
 599	if (!valid)
 600		__padata_stop(pinst);
 601
 602out_replace:
 603	pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 604	if (!pd)
 605		return -ENOMEM;
 606
 607	cpumask_copy(pinst->cpumask.pcpu, pcpumask);
 608	cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
 609
 610	padata_replace(pinst, pd);
 611
 612	if (valid)
 613		__padata_start(pinst);
 614
 615	return 0;
 616}
 617
 618/**
 619 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
 620 *                     equivalent to @cpumask.
 621 *
 622 * @pinst: padata instance
 623 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
 624 *                to parallel and serial cpumasks respectively.
 625 * @cpumask: the cpumask to use
 
 
 626 */
 627int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
 628		       cpumask_var_t cpumask)
 629{
 630	struct cpumask *serial_mask, *parallel_mask;
 631	int err = -EINVAL;
 632
 
 633	mutex_lock(&pinst->lock);
 634	get_online_cpus();
 635
 636	switch (cpumask_type) {
 637	case PADATA_CPU_PARALLEL:
 638		serial_mask = pinst->cpumask.cbcpu;
 639		parallel_mask = cpumask;
 640		break;
 641	case PADATA_CPU_SERIAL:
 642		parallel_mask = pinst->cpumask.pcpu;
 643		serial_mask = cpumask;
 644		break;
 645	default:
 646		 goto out;
 647	}
 648
 649	err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
 650
 651out:
 652	put_online_cpus();
 653	mutex_unlock(&pinst->lock);
 
 654
 655	return err;
 656}
 657EXPORT_SYMBOL(padata_set_cpumask);
 658
 659/**
 660 * padata_start - start the parallel processing
 661 *
 662 * @pinst: padata instance to start
 663 */
 664int padata_start(struct padata_instance *pinst)
 665{
 666	int err = 0;
 667
 668	mutex_lock(&pinst->lock);
 669
 670	if (pinst->flags & PADATA_INVALID)
 671		err = -EINVAL;
 672
 673	__padata_start(pinst);
 674
 675	mutex_unlock(&pinst->lock);
 676
 677	return err;
 678}
 679EXPORT_SYMBOL(padata_start);
 680
 681/**
 682 * padata_stop - stop the parallel processing
 683 *
 684 * @pinst: padata instance to stop
 685 */
 686void padata_stop(struct padata_instance *pinst)
 687{
 688	mutex_lock(&pinst->lock);
 689	__padata_stop(pinst);
 690	mutex_unlock(&pinst->lock);
 691}
 692EXPORT_SYMBOL(padata_stop);
 693
 694#ifdef CONFIG_HOTPLUG_CPU
 695
 696static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
 697{
 698	struct parallel_data *pd;
 699
 700	if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 701		pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 702				     pinst->cpumask.cbcpu);
 703		if (!pd)
 704			return -ENOMEM;
 705
 706		padata_replace(pinst, pd);
 707
 708		if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
 709		    padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 710			__padata_start(pinst);
 711	}
 712
 713	return 0;
 714}
 715
 716static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
 717{
 718	struct parallel_data *pd = NULL;
 719
 720	if (cpumask_test_cpu(cpu, cpu_online_mask)) {
 721
 
 722		if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
 723		    !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
 724			__padata_stop(pinst);
 725
 726		pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
 727				     pinst->cpumask.cbcpu);
 728		if (!pd)
 729			return -ENOMEM;
 730
 731		padata_replace(pinst, pd);
 732
 733		cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
 734		cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
 735	}
 736
 737	return 0;
 738}
 739
 740 /**
 741 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
 742 *                     padata cpumasks.
 743 *
 744 * @pinst: padata instance
 745 * @cpu: cpu to remove
 746 * @mask: bitmask specifying from which cpumask @cpu should be removed
 747 *        The @mask may be any combination of the following flags:
 748 *          PADATA_CPU_SERIAL   - serial cpumask
 749 *          PADATA_CPU_PARALLEL - parallel cpumask
 750 */
 751int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
 752{
 753	int err;
 754
 755	if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
 756		return -EINVAL;
 757
 758	mutex_lock(&pinst->lock);
 759
 760	get_online_cpus();
 761	if (mask & PADATA_CPU_SERIAL)
 762		cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
 763	if (mask & PADATA_CPU_PARALLEL)
 764		cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
 765
 766	err = __padata_remove_cpu(pinst, cpu);
 767	put_online_cpus();
 768
 769	mutex_unlock(&pinst->lock);
 770
 771	return err;
 772}
 773EXPORT_SYMBOL(padata_remove_cpu);
 774
 775static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
 776{
 777	return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
 778		cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
 779}
 780
 781static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
 782{
 783	struct padata_instance *pinst;
 784	int ret;
 785
 786	pinst = hlist_entry_safe(node, struct padata_instance, node);
 787	if (!pinst_has_cpu(pinst, cpu))
 788		return 0;
 789
 790	mutex_lock(&pinst->lock);
 791	ret = __padata_add_cpu(pinst, cpu);
 792	mutex_unlock(&pinst->lock);
 793	return ret;
 794}
 795
 796static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
 797{
 798	struct padata_instance *pinst;
 799	int ret;
 800
 801	pinst = hlist_entry_safe(node, struct padata_instance, node);
 802	if (!pinst_has_cpu(pinst, cpu))
 803		return 0;
 804
 805	mutex_lock(&pinst->lock);
 806	ret = __padata_remove_cpu(pinst, cpu);
 807	mutex_unlock(&pinst->lock);
 808	return ret;
 809}
 810
 811static enum cpuhp_state hp_online;
 812#endif
 813
 814static void __padata_free(struct padata_instance *pinst)
 815{
 816#ifdef CONFIG_HOTPLUG_CPU
 817	cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
 
 
 818#endif
 819
 820	padata_stop(pinst);
 821	padata_free_pd(pinst->pd);
 822	free_cpumask_var(pinst->cpumask.pcpu);
 823	free_cpumask_var(pinst->cpumask.cbcpu);
 824	destroy_workqueue(pinst->serial_wq);
 825	destroy_workqueue(pinst->parallel_wq);
 826	kfree(pinst);
 827}
 828
 829#define kobj2pinst(_kobj)					\
 830	container_of(_kobj, struct padata_instance, kobj)
 831#define attr2pentry(_attr)					\
 832	container_of(_attr, struct padata_sysfs_entry, attr)
 833
 834static void padata_sysfs_release(struct kobject *kobj)
 835{
 836	struct padata_instance *pinst = kobj2pinst(kobj);
 837	__padata_free(pinst);
 838}
 839
 840struct padata_sysfs_entry {
 841	struct attribute attr;
 842	ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
 843	ssize_t (*store)(struct padata_instance *, struct attribute *,
 844			 const char *, size_t);
 845};
 846
 847static ssize_t show_cpumask(struct padata_instance *pinst,
 848			    struct attribute *attr,  char *buf)
 849{
 850	struct cpumask *cpumask;
 851	ssize_t len;
 852
 853	mutex_lock(&pinst->lock);
 854	if (!strcmp(attr->name, "serial_cpumask"))
 855		cpumask = pinst->cpumask.cbcpu;
 856	else
 857		cpumask = pinst->cpumask.pcpu;
 858
 859	len = snprintf(buf, PAGE_SIZE, "%*pb\n",
 860		       nr_cpu_ids, cpumask_bits(cpumask));
 861	mutex_unlock(&pinst->lock);
 862	return len < PAGE_SIZE ? len : -EINVAL;
 863}
 864
 865static ssize_t store_cpumask(struct padata_instance *pinst,
 866			     struct attribute *attr,
 867			     const char *buf, size_t count)
 868{
 869	cpumask_var_t new_cpumask;
 870	ssize_t ret;
 871	int mask_type;
 872
 873	if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
 874		return -ENOMEM;
 875
 876	ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
 877			   nr_cpumask_bits);
 878	if (ret < 0)
 879		goto out;
 880
 881	mask_type = !strcmp(attr->name, "serial_cpumask") ?
 882		PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
 883	ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
 884	if (!ret)
 885		ret = count;
 886
 887out:
 888	free_cpumask_var(new_cpumask);
 889	return ret;
 890}
 891
 892#define PADATA_ATTR_RW(_name, _show_name, _store_name)		\
 893	static struct padata_sysfs_entry _name##_attr =		\
 894		__ATTR(_name, 0644, _show_name, _store_name)
 895#define PADATA_ATTR_RO(_name, _show_name)		\
 896	static struct padata_sysfs_entry _name##_attr = \
 897		__ATTR(_name, 0400, _show_name, NULL)
 898
 899PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
 900PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
 901
 902/*
 903 * Padata sysfs provides the following objects:
 904 * serial_cpumask   [RW] - cpumask for serial workers
 905 * parallel_cpumask [RW] - cpumask for parallel workers
 906 */
 907static struct attribute *padata_default_attrs[] = {
 908	&serial_cpumask_attr.attr,
 909	&parallel_cpumask_attr.attr,
 910	NULL,
 911};
 912ATTRIBUTE_GROUPS(padata_default);
 913
 914static ssize_t padata_sysfs_show(struct kobject *kobj,
 915				 struct attribute *attr, char *buf)
 916{
 917	struct padata_instance *pinst;
 918	struct padata_sysfs_entry *pentry;
 919	ssize_t ret = -EIO;
 920
 921	pinst = kobj2pinst(kobj);
 922	pentry = attr2pentry(attr);
 923	if (pentry->show)
 924		ret = pentry->show(pinst, attr, buf);
 925
 926	return ret;
 927}
 928
 929static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
 930				  const char *buf, size_t count)
 931{
 932	struct padata_instance *pinst;
 933	struct padata_sysfs_entry *pentry;
 934	ssize_t ret = -EIO;
 935
 936	pinst = kobj2pinst(kobj);
 937	pentry = attr2pentry(attr);
 938	if (pentry->show)
 939		ret = pentry->store(pinst, attr, buf, count);
 940
 941	return ret;
 942}
 943
 944static const struct sysfs_ops padata_sysfs_ops = {
 945	.show = padata_sysfs_show,
 946	.store = padata_sysfs_store,
 947};
 948
 949static struct kobj_type padata_attr_type = {
 950	.sysfs_ops = &padata_sysfs_ops,
 951	.default_groups = padata_default_groups,
 952	.release = padata_sysfs_release,
 953};
 954
 955/**
 956 * padata_alloc - allocate and initialize a padata instance and specify
 957 *                cpumasks for serial and parallel workers.
 958 *
 959 * @name: used to identify the instance
 960 * @pcpumask: cpumask that will be used for padata parallelization
 961 * @cbcpumask: cpumask that will be used for padata serialization
 962 */
 963static struct padata_instance *padata_alloc(const char *name,
 964					    const struct cpumask *pcpumask,
 965					    const struct cpumask *cbcpumask)
 966{
 967	struct padata_instance *pinst;
 968	struct parallel_data *pd = NULL;
 969
 970	pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
 971	if (!pinst)
 972		goto err;
 973
 974	pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
 975					     name);
 976	if (!pinst->parallel_wq)
 977		goto err_free_inst;
 978
 979	get_online_cpus();
 980
 981	pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
 982					   WQ_CPU_INTENSIVE, 1, name);
 983	if (!pinst->serial_wq)
 984		goto err_put_cpus;
 985
 986	if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
 987		goto err_free_serial_wq;
 988	if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
 989		free_cpumask_var(pinst->cpumask.pcpu);
 990		goto err_free_serial_wq;
 991	}
 992	if (!padata_validate_cpumask(pinst, pcpumask) ||
 993	    !padata_validate_cpumask(pinst, cbcpumask))
 994		goto err_free_masks;
 995
 996	pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
 997	if (!pd)
 998		goto err_free_masks;
 999
1000	rcu_assign_pointer(pinst->pd, pd);
 
1001
1002	cpumask_copy(pinst->cpumask.pcpu, pcpumask);
1003	cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
1004
1005	pinst->flags = 0;
1006
1007	BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1008	kobject_init(&pinst->kobj, &padata_attr_type);
1009	mutex_init(&pinst->lock);
1010
1011#ifdef CONFIG_HOTPLUG_CPU
1012	cpuhp_state_add_instance_nocalls_cpuslocked(hp_online, &pinst->node);
 
 
 
1013#endif
1014
1015	put_online_cpus();
1016
1017	return pinst;
1018
1019err_free_masks:
1020	free_cpumask_var(pinst->cpumask.pcpu);
1021	free_cpumask_var(pinst->cpumask.cbcpu);
1022err_free_serial_wq:
1023	destroy_workqueue(pinst->serial_wq);
1024err_put_cpus:
1025	put_online_cpus();
1026	destroy_workqueue(pinst->parallel_wq);
1027err_free_inst:
1028	kfree(pinst);
1029err:
1030	return NULL;
1031}
 
1032
1033/**
1034 * padata_alloc_possible - Allocate and initialize padata instance.
1035 *                         Use the cpu_possible_mask for serial and
1036 *                         parallel workers.
1037 *
1038 * @name: used to identify the instance
1039 */
1040struct padata_instance *padata_alloc_possible(const char *name)
1041{
1042	return padata_alloc(name, cpu_possible_mask, cpu_possible_mask);
1043}
1044EXPORT_SYMBOL(padata_alloc_possible);
1045
1046/**
1047 * padata_free - free a padata instance
 
 
1048 *
1049 * @padata_inst: padata instance to free
1050 */
1051void padata_free(struct padata_instance *pinst)
1052{
1053	kobject_put(&pinst->kobj);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1054}
1055EXPORT_SYMBOL(padata_free);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1056
1057#ifdef CONFIG_HOTPLUG_CPU
 
 
1058
1059static __init int padata_driver_init(void)
1060{
 
 
1061	int ret;
1062
1063	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1064				      padata_cpu_online,
1065				      padata_cpu_prep_down);
1066	if (ret < 0)
1067		return ret;
1068	hp_online = ret;
1069	return 0;
1070}
1071module_init(padata_driver_init);
1072
1073static __exit void padata_driver_exit(void)
1074{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1075	cpuhp_remove_multi_state(hp_online);
 
 
 
1076}
1077module_exit(padata_driver_exit);
1078#endif