blob: 4012ff541b7b3248bdb3e7140fc929095d332aab [file] [log] [blame]
Jens Axboe771b53d02019-10-22 10:25:58 -06001// SPDX-License-Identifier: GPL-2.0
2/*
3 * Basic worker thread pool for io_uring
4 *
5 * Copyright (C) 2019 Jens Axboe
6 *
7 */
8#include <linux/kernel.h>
9#include <linux/init.h>
10#include <linux/errno.h>
11#include <linux/sched/signal.h>
12#include <linux/mm.h>
Jens Axboe771b53d02019-10-22 10:25:58 -060013#include <linux/sched/mm.h>
14#include <linux/percpu.h>
15#include <linux/slab.h>
16#include <linux/kthread.h>
17#include <linux/rculist_nulls.h>
Jens Axboe9392a272020-02-06 21:42:51 -070018#include <linux/fs_struct.h>
Jens Axboeaa96bf82020-04-03 11:26:26 -060019#include <linux/task_work.h>
Dennis Zhou91d8f512020-09-16 13:41:05 -070020#include <linux/blk-cgroup.h>
Jens Axboe4ea33a92020-10-15 13:46:44 -060021#include <linux/audit.h>
Jens Axboe771b53d02019-10-22 10:25:58 -060022
23#include "io-wq.h"
24
25#define WORKER_IDLE_TIMEOUT (5 * HZ)
26
27enum {
28 IO_WORKER_F_UP = 1, /* up and active */
29 IO_WORKER_F_RUNNING = 2, /* account as running */
30 IO_WORKER_F_FREE = 4, /* worker on free list */
Jens Axboe145cc8c2020-09-26 12:37:46 -060031 IO_WORKER_F_FIXED = 8, /* static idle worker */
32 IO_WORKER_F_BOUND = 16, /* is doing bounded work */
Jens Axboe771b53d02019-10-22 10:25:58 -060033};
34
35enum {
36 IO_WQ_BIT_EXIT = 0, /* wq exiting */
37 IO_WQ_BIT_CANCEL = 1, /* cancel work on list */
Jens Axboeb60fda62019-11-19 08:37:07 -070038 IO_WQ_BIT_ERROR = 2, /* error on setup */
Jens Axboe771b53d02019-10-22 10:25:58 -060039};
40
41enum {
42 IO_WQE_FLAG_STALLED = 1, /* stalled on hash */
43};
44
45/*
46 * One for each thread in a wqe pool
47 */
48struct io_worker {
49 refcount_t ref;
50 unsigned flags;
51 struct hlist_nulls_node nulls_node;
Jens Axboee61df662019-11-13 13:54:49 -070052 struct list_head all_list;
Jens Axboe771b53d02019-10-22 10:25:58 -060053 struct task_struct *task;
Jens Axboe771b53d02019-10-22 10:25:58 -060054 struct io_wqe *wqe;
Jens Axboe36c2f922019-11-13 09:43:34 -070055
Jens Axboe771b53d02019-10-22 10:25:58 -060056 struct io_wq_work *cur_work;
Jens Axboe36c2f922019-11-13 09:43:34 -070057 spinlock_t lock;
Jens Axboe771b53d02019-10-22 10:25:58 -060058
59 struct rcu_head rcu;
60 struct mm_struct *mm;
Dennis Zhou91d8f512020-09-16 13:41:05 -070061#ifdef CONFIG_BLK_CGROUP
62 struct cgroup_subsys_state *blkcg_css;
63#endif
Jens Axboecccf0ee2020-01-27 16:34:48 -070064 const struct cred *cur_creds;
65 const struct cred *saved_creds;
Jens Axboefcb323c2019-10-24 12:39:47 -060066 struct files_struct *restore_files;
Jens Axboe9b828492020-09-18 20:13:06 -060067 struct nsproxy *restore_nsproxy;
Jens Axboe9392a272020-02-06 21:42:51 -070068 struct fs_struct *restore_fs;
Jens Axboe771b53d02019-10-22 10:25:58 -060069};
70
Jens Axboe771b53d02019-10-22 10:25:58 -060071#if BITS_PER_LONG == 64
72#define IO_WQ_HASH_ORDER 6
73#else
74#define IO_WQ_HASH_ORDER 5
75#endif
76
Pavel Begunkov86f3cd12020-03-23 22:57:22 +030077#define IO_WQ_NR_HASH_BUCKETS (1u << IO_WQ_HASH_ORDER)
78
Jens Axboec5def4a2019-11-07 11:41:16 -070079struct io_wqe_acct {
80 unsigned nr_workers;
81 unsigned max_workers;
82 atomic_t nr_running;
83};
84
85enum {
86 IO_WQ_ACCT_BOUND,
87 IO_WQ_ACCT_UNBOUND,
88};
89
Jens Axboe771b53d02019-10-22 10:25:58 -060090/*
91 * Per-node worker thread pool
92 */
93struct io_wqe {
94 struct {
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +020095 raw_spinlock_t lock;
Jens Axboe6206f0e2019-11-26 11:59:32 -070096 struct io_wq_work_list work_list;
Jens Axboe771b53d02019-10-22 10:25:58 -060097 unsigned long hash_map;
98 unsigned flags;
99 } ____cacheline_aligned_in_smp;
100
101 int node;
Jens Axboec5def4a2019-11-07 11:41:16 -0700102 struct io_wqe_acct acct[2];
Jens Axboe771b53d02019-10-22 10:25:58 -0600103
Jens Axboe021d1cd2019-11-14 08:00:41 -0700104 struct hlist_nulls_head free_list;
Jens Axboee61df662019-11-13 13:54:49 -0700105 struct list_head all_list;
Jens Axboe771b53d02019-10-22 10:25:58 -0600106
107 struct io_wq *wq;
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300108 struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
Jens Axboe771b53d02019-10-22 10:25:58 -0600109};
110
111/*
112 * Per io_wq state
113 */
114struct io_wq {
115 struct io_wqe **wqes;
116 unsigned long state;
Jens Axboe771b53d02019-10-22 10:25:58 -0600117
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300118 free_work_fn *free_work;
Pavel Begunkovf5fa38c2020-06-08 21:08:20 +0300119 io_wq_work_fn *do_work;
Jens Axboe7d723062019-11-12 22:31:31 -0700120
Jens Axboe771b53d02019-10-22 10:25:58 -0600121 struct task_struct *manager;
Jens Axboec5def4a2019-11-07 11:41:16 -0700122 struct user_struct *user;
Jens Axboe771b53d02019-10-22 10:25:58 -0600123 refcount_t refs;
124 struct completion done;
Jens Axboe848f7e12020-01-23 15:33:32 -0700125
126 refcount_t use_refs;
Jens Axboe771b53d02019-10-22 10:25:58 -0600127};
128
Jens Axboe771b53d02019-10-22 10:25:58 -0600129static bool io_worker_get(struct io_worker *worker)
130{
131 return refcount_inc_not_zero(&worker->ref);
132}
133
134static void io_worker_release(struct io_worker *worker)
135{
136 if (refcount_dec_and_test(&worker->ref))
137 wake_up_process(worker->task);
138}
139
140/*
141 * Note: drops the wqe->lock if returning true! The caller must re-acquire
142 * the lock in that case. Some callers need to restart handling if this
143 * happens, so we can't just re-acquire the lock on behalf of the caller.
144 */
145static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
146{
Jens Axboefcb323c2019-10-24 12:39:47 -0600147 bool dropped_lock = false;
148
Jens Axboecccf0ee2020-01-27 16:34:48 -0700149 if (worker->saved_creds) {
150 revert_creds(worker->saved_creds);
151 worker->cur_creds = worker->saved_creds = NULL;
Jens Axboe181e4482019-11-25 08:52:30 -0700152 }
153
Jens Axboefcb323c2019-10-24 12:39:47 -0600154 if (current->files != worker->restore_files) {
155 __acquire(&wqe->lock);
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200156 raw_spin_unlock_irq(&wqe->lock);
Jens Axboefcb323c2019-10-24 12:39:47 -0600157 dropped_lock = true;
158
159 task_lock(current);
160 current->files = worker->restore_files;
Jens Axboe9b828492020-09-18 20:13:06 -0600161 current->nsproxy = worker->restore_nsproxy;
Jens Axboefcb323c2019-10-24 12:39:47 -0600162 task_unlock(current);
163 }
164
Jens Axboe9392a272020-02-06 21:42:51 -0700165 if (current->fs != worker->restore_fs)
166 current->fs = worker->restore_fs;
167
Jens Axboe771b53d02019-10-22 10:25:58 -0600168 /*
169 * If we have an active mm, we need to drop the wq lock before unusing
170 * it. If we do, return true and let the caller retry the idle loop.
171 */
172 if (worker->mm) {
Jens Axboefcb323c2019-10-24 12:39:47 -0600173 if (!dropped_lock) {
174 __acquire(&wqe->lock);
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200175 raw_spin_unlock_irq(&wqe->lock);
Jens Axboefcb323c2019-10-24 12:39:47 -0600176 dropped_lock = true;
177 }
Jens Axboe771b53d02019-10-22 10:25:58 -0600178 __set_current_state(TASK_RUNNING);
Christoph Hellwigf5678e72020-06-10 18:42:06 -0700179 kthread_unuse_mm(worker->mm);
Jens Axboe771b53d02019-10-22 10:25:58 -0600180 mmput(worker->mm);
181 worker->mm = NULL;
Jens Axboe771b53d02019-10-22 10:25:58 -0600182 }
183
Dennis Zhou91d8f512020-09-16 13:41:05 -0700184#ifdef CONFIG_BLK_CGROUP
185 if (worker->blkcg_css) {
186 kthread_associate_blkcg(NULL);
187 worker->blkcg_css = NULL;
188 }
189#endif
Jens Axboe69228332020-10-20 14:28:41 -0600190 if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
191 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
Jens Axboefcb323c2019-10-24 12:39:47 -0600192 return dropped_lock;
Jens Axboe771b53d02019-10-22 10:25:58 -0600193}
194
Jens Axboec5def4a2019-11-07 11:41:16 -0700195static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
196 struct io_wq_work *work)
197{
198 if (work->flags & IO_WQ_WORK_UNBOUND)
199 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
200
201 return &wqe->acct[IO_WQ_ACCT_BOUND];
202}
203
204static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
205 struct io_worker *worker)
206{
207 if (worker->flags & IO_WORKER_F_BOUND)
208 return &wqe->acct[IO_WQ_ACCT_BOUND];
209
210 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
211}
212
Jens Axboe771b53d02019-10-22 10:25:58 -0600213static void io_worker_exit(struct io_worker *worker)
214{
215 struct io_wqe *wqe = worker->wqe;
Jens Axboec5def4a2019-11-07 11:41:16 -0700216 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
Jens Axboe771b53d02019-10-22 10:25:58 -0600217
218 /*
219 * If we're not at zero, someone else is holding a brief reference
220 * to the worker. Wait for that to go away.
221 */
222 set_current_state(TASK_INTERRUPTIBLE);
223 if (!refcount_dec_and_test(&worker->ref))
224 schedule();
225 __set_current_state(TASK_RUNNING);
226
227 preempt_disable();
228 current->flags &= ~PF_IO_WORKER;
229 if (worker->flags & IO_WORKER_F_RUNNING)
Jens Axboec5def4a2019-11-07 11:41:16 -0700230 atomic_dec(&acct->nr_running);
231 if (!(worker->flags & IO_WORKER_F_BOUND))
232 atomic_dec(&wqe->wq->user->processes);
Jens Axboe771b53d02019-10-22 10:25:58 -0600233 worker->flags = 0;
234 preempt_enable();
235
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200236 raw_spin_lock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600237 hlist_nulls_del_rcu(&worker->nulls_node);
Jens Axboee61df662019-11-13 13:54:49 -0700238 list_del_rcu(&worker->all_list);
Jens Axboe771b53d02019-10-22 10:25:58 -0600239 if (__io_worker_unuse(wqe, worker)) {
240 __release(&wqe->lock);
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200241 raw_spin_lock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600242 }
Jens Axboec5def4a2019-11-07 11:41:16 -0700243 acct->nr_workers--;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200244 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600245
YueHaibing364b05f2019-11-02 15:55:01 +0800246 kfree_rcu(worker, rcu);
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800247 if (refcount_dec_and_test(&wqe->wq->refs))
248 complete(&wqe->wq->done);
Jens Axboe771b53d02019-10-22 10:25:58 -0600249}
250
Jens Axboec5def4a2019-11-07 11:41:16 -0700251static inline bool io_wqe_run_queue(struct io_wqe *wqe)
252 __must_hold(wqe->lock)
253{
Jens Axboe6206f0e2019-11-26 11:59:32 -0700254 if (!wq_list_empty(&wqe->work_list) &&
255 !(wqe->flags & IO_WQE_FLAG_STALLED))
Jens Axboec5def4a2019-11-07 11:41:16 -0700256 return true;
257 return false;
258}
259
260/*
261 * Check head of free list for an available worker. If one isn't available,
262 * caller must wake up the wq manager to create one.
263 */
264static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
265 __must_hold(RCU)
266{
267 struct hlist_nulls_node *n;
268 struct io_worker *worker;
269
Jens Axboe021d1cd2019-11-14 08:00:41 -0700270 n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
Jens Axboec5def4a2019-11-07 11:41:16 -0700271 if (is_a_nulls(n))
272 return false;
273
274 worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
275 if (io_worker_get(worker)) {
Jens Axboe506d95f2019-12-07 21:03:59 -0700276 wake_up_process(worker->task);
Jens Axboec5def4a2019-11-07 11:41:16 -0700277 io_worker_release(worker);
278 return true;
279 }
280
281 return false;
282}
283
284/*
285 * We need a worker. If we find a free one, we're good. If not, and we're
286 * below the max number of workers, wake up the manager to create one.
287 */
288static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
289{
290 bool ret;
291
292 /*
293 * Most likely an attempt to queue unbounded work on an io_wq that
294 * wasn't setup with any unbounded workers.
295 */
296 WARN_ON_ONCE(!acct->max_workers);
297
298 rcu_read_lock();
299 ret = io_wqe_activate_free_worker(wqe);
300 rcu_read_unlock();
301
302 if (!ret && acct->nr_workers < acct->max_workers)
303 wake_up_process(wqe->wq->manager);
304}
305
306static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
307{
308 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
309
310 atomic_inc(&acct->nr_running);
311}
312
313static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
314 __must_hold(wqe->lock)
315{
316 struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
317
318 if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
319 io_wqe_wake_worker(wqe, acct);
320}
321
Jens Axboe771b53d02019-10-22 10:25:58 -0600322static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
323{
324 allow_kernel_signal(SIGINT);
325
326 current->flags |= PF_IO_WORKER;
327
328 worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
Jens Axboefcb323c2019-10-24 12:39:47 -0600329 worker->restore_files = current->files;
Jens Axboe9b828492020-09-18 20:13:06 -0600330 worker->restore_nsproxy = current->nsproxy;
Jens Axboe9392a272020-02-06 21:42:51 -0700331 worker->restore_fs = current->fs;
Jens Axboec5def4a2019-11-07 11:41:16 -0700332 io_wqe_inc_running(wqe, worker);
Jens Axboe771b53d02019-10-22 10:25:58 -0600333}
334
335/*
336 * Worker will start processing some work. Move it to the busy list, if
337 * it's currently on the freelist
338 */
339static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
340 struct io_wq_work *work)
341 __must_hold(wqe->lock)
342{
Jens Axboec5def4a2019-11-07 11:41:16 -0700343 bool worker_bound, work_bound;
344
Jens Axboe771b53d02019-10-22 10:25:58 -0600345 if (worker->flags & IO_WORKER_F_FREE) {
346 worker->flags &= ~IO_WORKER_F_FREE;
347 hlist_nulls_del_init_rcu(&worker->nulls_node);
Jens Axboe771b53d02019-10-22 10:25:58 -0600348 }
Jens Axboec5def4a2019-11-07 11:41:16 -0700349
350 /*
351 * If worker is moving from bound to unbound (or vice versa), then
352 * ensure we update the running accounting.
353 */
Dan Carpenterb2e9c7d62019-11-19 09:22:16 +0300354 worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
355 work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
356 if (worker_bound != work_bound) {
Jens Axboec5def4a2019-11-07 11:41:16 -0700357 io_wqe_dec_running(wqe, worker);
358 if (work_bound) {
359 worker->flags |= IO_WORKER_F_BOUND;
360 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
361 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
362 atomic_dec(&wqe->wq->user->processes);
363 } else {
364 worker->flags &= ~IO_WORKER_F_BOUND;
365 wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
366 wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
367 atomic_inc(&wqe->wq->user->processes);
368 }
369 io_wqe_inc_running(wqe, worker);
370 }
Jens Axboe771b53d02019-10-22 10:25:58 -0600371}
372
373/*
374 * No work, worker going to sleep. Move to freelist, and unuse mm if we
375 * have one attached. Dropping the mm may potentially sleep, so we drop
376 * the lock in that case and return success. Since the caller has to
377 * retry the loop in that case (we changed task state), we don't regrab
378 * the lock if we return success.
379 */
380static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
381 __must_hold(wqe->lock)
382{
383 if (!(worker->flags & IO_WORKER_F_FREE)) {
384 worker->flags |= IO_WORKER_F_FREE;
Jens Axboe021d1cd2019-11-14 08:00:41 -0700385 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
Jens Axboe771b53d02019-10-22 10:25:58 -0600386 }
387
388 return __io_worker_unuse(wqe, worker);
389}
390
Pavel Begunkov60cf46a2020-03-14 00:31:05 +0300391static inline unsigned int io_get_work_hash(struct io_wq_work *work)
392{
393 return work->flags >> IO_WQ_HASH_SHIFT;
394}
395
396static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
Jens Axboe771b53d02019-10-22 10:25:58 -0600397 __must_hold(wqe->lock)
398{
Jens Axboe6206f0e2019-11-26 11:59:32 -0700399 struct io_wq_work_node *node, *prev;
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300400 struct io_wq_work *work, *tail;
Pavel Begunkov60cf46a2020-03-14 00:31:05 +0300401 unsigned int hash;
Jens Axboe771b53d02019-10-22 10:25:58 -0600402
Jens Axboe6206f0e2019-11-26 11:59:32 -0700403 wq_list_for_each(node, prev, &wqe->work_list) {
404 work = container_of(node, struct io_wq_work, list);
405
Jens Axboe771b53d02019-10-22 10:25:58 -0600406 /* not hashed, can run anytime */
Pavel Begunkov8766dd52020-03-14 00:31:04 +0300407 if (!io_wq_is_hashed(work)) {
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300408 wq_list_del(&wqe->work_list, node, prev);
Jens Axboe771b53d02019-10-22 10:25:58 -0600409 return work;
410 }
411
412 /* hashed, can run if not already running */
Pavel Begunkov60cf46a2020-03-14 00:31:05 +0300413 hash = io_get_work_hash(work);
414 if (!(wqe->hash_map & BIT(hash))) {
415 wqe->hash_map |= BIT(hash);
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300416 /* all items with this hash lie in [work, tail] */
417 tail = wqe->hash_tail[hash];
418 wqe->hash_tail[hash] = NULL;
419 wq_list_cut(&wqe->work_list, &tail->list, prev);
Jens Axboe771b53d02019-10-22 10:25:58 -0600420 return work;
421 }
422 }
423
424 return NULL;
425}
426
Jens Axboecccf0ee2020-01-27 16:34:48 -0700427static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
428{
429 if (worker->mm) {
Christoph Hellwigf5678e72020-06-10 18:42:06 -0700430 kthread_unuse_mm(worker->mm);
Jens Axboecccf0ee2020-01-27 16:34:48 -0700431 mmput(worker->mm);
432 worker->mm = NULL;
433 }
Christoph Hellwig37c54f92020-06-10 18:42:10 -0700434
Jens Axboe98447d62020-10-14 10:48:51 -0600435 if (mmget_not_zero(work->identity->mm)) {
436 kthread_use_mm(work->identity->mm);
437 worker->mm = work->identity->mm;
Jens Axboecccf0ee2020-01-27 16:34:48 -0700438 return;
439 }
440
441 /* failed grabbing mm, ensure work gets cancelled */
442 work->flags |= IO_WQ_WORK_CANCEL;
443}
444
Dennis Zhou91d8f512020-09-16 13:41:05 -0700445static inline void io_wq_switch_blkcg(struct io_worker *worker,
446 struct io_wq_work *work)
447{
448#ifdef CONFIG_BLK_CGROUP
Jens Axboe0f203762020-10-14 09:23:55 -0600449 if (!(work->flags & IO_WQ_WORK_BLKCG))
450 return;
Jens Axboe98447d62020-10-14 10:48:51 -0600451 if (work->identity->blkcg_css != worker->blkcg_css) {
452 kthread_associate_blkcg(work->identity->blkcg_css);
453 worker->blkcg_css = work->identity->blkcg_css;
Dennis Zhou91d8f512020-09-16 13:41:05 -0700454 }
455#endif
456}
457
Jens Axboecccf0ee2020-01-27 16:34:48 -0700458static void io_wq_switch_creds(struct io_worker *worker,
459 struct io_wq_work *work)
460{
Jens Axboe98447d62020-10-14 10:48:51 -0600461 const struct cred *old_creds = override_creds(work->identity->creds);
Jens Axboecccf0ee2020-01-27 16:34:48 -0700462
Jens Axboe98447d62020-10-14 10:48:51 -0600463 worker->cur_creds = work->identity->creds;
Jens Axboecccf0ee2020-01-27 16:34:48 -0700464 if (worker->saved_creds)
465 put_cred(old_creds); /* creds set by previous switch */
466 else
467 worker->saved_creds = old_creds;
468}
469
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300470static void io_impersonate_work(struct io_worker *worker,
471 struct io_wq_work *work)
472{
Jens Axboe98447d62020-10-14 10:48:51 -0600473 if ((work->flags & IO_WQ_WORK_FILES) &&
474 current->files != work->identity->files) {
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300475 task_lock(current);
Jens Axboe98447d62020-10-14 10:48:51 -0600476 current->files = work->identity->files;
477 current->nsproxy = work->identity->nsproxy;
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300478 task_unlock(current);
479 }
Jens Axboe98447d62020-10-14 10:48:51 -0600480 if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs)
481 current->fs = work->identity->fs;
482 if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm)
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300483 io_wq_switch_mm(worker, work);
Jens Axboe98447d62020-10-14 10:48:51 -0600484 if ((work->flags & IO_WQ_WORK_CREDS) &&
485 worker->cur_creds != work->identity->creds)
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300486 io_wq_switch_creds(worker, work);
Jens Axboe69228332020-10-20 14:28:41 -0600487 if (work->flags & IO_WQ_WORK_FSIZE)
488 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize;
489 else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
490 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
Dennis Zhou91d8f512020-09-16 13:41:05 -0700491 io_wq_switch_blkcg(worker, work);
Jens Axboe4ea33a92020-10-15 13:46:44 -0600492#ifdef CONFIG_AUDIT
493 current->loginuid = work->identity->loginuid;
494 current->sessionid = work->identity->sessionid;
495#endif
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300496}
497
498static void io_assign_current_work(struct io_worker *worker,
499 struct io_wq_work *work)
500{
Pavel Begunkovd78298e2020-03-14 00:31:03 +0300501 if (work) {
502 /* flush pending signals before assigning new work */
503 if (signal_pending(current))
504 flush_signals(current);
505 cond_resched();
506 }
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300507
Jens Axboe4ea33a92020-10-15 13:46:44 -0600508#ifdef CONFIG_AUDIT
509 current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET);
510 current->sessionid = AUDIT_SID_UNSET;
511#endif
512
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300513 spin_lock_irq(&worker->lock);
514 worker->cur_work = work;
515 spin_unlock_irq(&worker->lock);
516}
517
Pavel Begunkov60cf46a2020-03-14 00:31:05 +0300518static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
519
Jens Axboe771b53d02019-10-22 10:25:58 -0600520static void io_worker_handle_work(struct io_worker *worker)
521 __releases(wqe->lock)
522{
Jens Axboe771b53d02019-10-22 10:25:58 -0600523 struct io_wqe *wqe = worker->wqe;
524 struct io_wq *wq = wqe->wq;
525
526 do {
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300527 struct io_wq_work *work;
Pavel Begunkovf462fd32020-03-04 16:14:11 +0300528get_next:
Jens Axboe771b53d02019-10-22 10:25:58 -0600529 /*
Jens Axboe771b53d02019-10-22 10:25:58 -0600530 * If we got some work, mark us as busy. If we didn't, but
531 * the list isn't empty, it means we stalled on hashed work.
532 * Mark us stalled so we don't keep looking for work when we
533 * can't make progress, any work completion or insertion will
534 * clear the stalled flag.
535 */
Pavel Begunkov60cf46a2020-03-14 00:31:05 +0300536 work = io_get_next_work(wqe);
Jens Axboe771b53d02019-10-22 10:25:58 -0600537 if (work)
538 __io_worker_busy(wqe, worker, work);
Jens Axboe6206f0e2019-11-26 11:59:32 -0700539 else if (!wq_list_empty(&wqe->work_list))
Jens Axboe771b53d02019-10-22 10:25:58 -0600540 wqe->flags |= IO_WQE_FLAG_STALLED;
541
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200542 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600543 if (!work)
544 break;
Pavel Begunkov58e39312020-03-04 16:14:10 +0300545 io_assign_current_work(worker, work);
Jens Axboe36c2f922019-11-13 09:43:34 -0700546
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300547 /* handle a whole dependent link */
548 do {
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300549 struct io_wq_work *old_work, *next_hashed, *linked;
Pavel Begunkovb089ed392020-07-25 14:42:00 +0300550 unsigned int hash = io_get_work_hash(work);
Hillf Dantonfd1c4bc2019-12-24 09:14:29 -0700551
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300552 next_hashed = wq_next_work(work);
Pavel Begunkov58e39312020-03-04 16:14:10 +0300553 io_impersonate_work(worker, work);
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300554 /*
555 * OK to set IO_WQ_WORK_CANCEL even for uncancellable
556 * work, the worker function will do the right thing.
557 */
558 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
559 work->flags |= IO_WQ_WORK_CANCEL;
Jens Axboe36c2f922019-11-13 09:43:34 -0700560
Pavel Begunkovf4db7182020-06-25 18:20:54 +0300561 old_work = work;
562 linked = wq->do_work(work);
Pavel Begunkovf2cf1142020-03-22 19:14:26 +0300563
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300564 work = next_hashed;
565 if (!work && linked && !io_wq_is_hashed(linked)) {
566 work = linked;
567 linked = NULL;
568 }
569 io_assign_current_work(worker, work);
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300570 wq->free_work(old_work);
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300571
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300572 if (linked)
573 io_wqe_enqueue(wqe, linked);
574
575 if (hash != -1U && !next_hashed) {
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200576 raw_spin_lock_irq(&wqe->lock);
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300577 wqe->hash_map &= ~BIT_ULL(hash);
578 wqe->flags &= ~IO_WQE_FLAG_STALLED;
Pavel Begunkovf462fd32020-03-04 16:14:11 +0300579 /* skip unnecessary unlock-lock wqe->lock */
580 if (!work)
581 goto get_next;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200582 raw_spin_unlock_irq(&wqe->lock);
Pavel Begunkovdc026a72020-03-04 16:14:09 +0300583 }
Pavel Begunkov58e39312020-03-04 16:14:10 +0300584 } while (work);
Jens Axboe36c2f922019-11-13 09:43:34 -0700585
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200586 raw_spin_lock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600587 } while (1);
588}
589
Jens Axboe771b53d02019-10-22 10:25:58 -0600590static int io_wqe_worker(void *data)
591{
592 struct io_worker *worker = data;
593 struct io_wqe *wqe = worker->wqe;
594 struct io_wq *wq = wqe->wq;
Jens Axboe771b53d02019-10-22 10:25:58 -0600595
596 io_worker_start(wqe, worker);
597
598 while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
Jens Axboe506d95f2019-12-07 21:03:59 -0700599 set_current_state(TASK_INTERRUPTIBLE);
Jens Axboee995d512019-12-07 21:06:46 -0700600loop:
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200601 raw_spin_lock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600602 if (io_wqe_run_queue(wqe)) {
603 __set_current_state(TASK_RUNNING);
604 io_worker_handle_work(worker);
Jens Axboee995d512019-12-07 21:06:46 -0700605 goto loop;
Jens Axboe771b53d02019-10-22 10:25:58 -0600606 }
607 /* drops the lock on success, retry */
608 if (__io_worker_idle(wqe, worker)) {
609 __release(&wqe->lock);
Jens Axboee995d512019-12-07 21:06:46 -0700610 goto loop;
Jens Axboe771b53d02019-10-22 10:25:58 -0600611 }
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200612 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600613 if (signal_pending(current))
614 flush_signals(current);
615 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
616 continue;
617 /* timed out, exit unless we're the fixed worker */
618 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
619 !(worker->flags & IO_WORKER_F_FIXED))
620 break;
621 }
622
Jens Axboe771b53d02019-10-22 10:25:58 -0600623 if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200624 raw_spin_lock_irq(&wqe->lock);
Jens Axboe6206f0e2019-11-26 11:59:32 -0700625 if (!wq_list_empty(&wqe->work_list))
Jens Axboe771b53d02019-10-22 10:25:58 -0600626 io_worker_handle_work(worker);
627 else
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200628 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600629 }
630
631 io_worker_exit(worker);
632 return 0;
633}
634
635/*
Jens Axboe771b53d02019-10-22 10:25:58 -0600636 * Called when a worker is scheduled in. Mark us as currently running.
637 */
638void io_wq_worker_running(struct task_struct *tsk)
639{
640 struct io_worker *worker = kthread_data(tsk);
641 struct io_wqe *wqe = worker->wqe;
642
643 if (!(worker->flags & IO_WORKER_F_UP))
644 return;
645 if (worker->flags & IO_WORKER_F_RUNNING)
646 return;
647 worker->flags |= IO_WORKER_F_RUNNING;
Jens Axboec5def4a2019-11-07 11:41:16 -0700648 io_wqe_inc_running(wqe, worker);
Jens Axboe771b53d02019-10-22 10:25:58 -0600649}
650
651/*
652 * Called when worker is going to sleep. If there are no workers currently
653 * running and we have work pending, wake up a free one or have the manager
654 * set one up.
655 */
656void io_wq_worker_sleeping(struct task_struct *tsk)
657{
658 struct io_worker *worker = kthread_data(tsk);
659 struct io_wqe *wqe = worker->wqe;
660
661 if (!(worker->flags & IO_WORKER_F_UP))
662 return;
663 if (!(worker->flags & IO_WORKER_F_RUNNING))
664 return;
665
666 worker->flags &= ~IO_WORKER_F_RUNNING;
667
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200668 raw_spin_lock_irq(&wqe->lock);
Jens Axboec5def4a2019-11-07 11:41:16 -0700669 io_wqe_dec_running(wqe, worker);
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200670 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600671}
672
Jens Axboeb60fda62019-11-19 08:37:07 -0700673static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
Jens Axboe771b53d02019-10-22 10:25:58 -0600674{
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800675 struct io_wqe_acct *acct = &wqe->acct[index];
Jens Axboe771b53d02019-10-22 10:25:58 -0600676 struct io_worker *worker;
677
Jann Hornad6e0052019-11-26 17:39:45 +0100678 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
Jens Axboe771b53d02019-10-22 10:25:58 -0600679 if (!worker)
Jens Axboeb60fda62019-11-19 08:37:07 -0700680 return false;
Jens Axboe771b53d02019-10-22 10:25:58 -0600681
682 refcount_set(&worker->ref, 1);
683 worker->nulls_node.pprev = NULL;
Jens Axboe771b53d02019-10-22 10:25:58 -0600684 worker->wqe = wqe;
Jens Axboe36c2f922019-11-13 09:43:34 -0700685 spin_lock_init(&worker->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600686
687 worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
Jens Axboec5def4a2019-11-07 11:41:16 -0700688 "io_wqe_worker-%d/%d", index, wqe->node);
Jens Axboe771b53d02019-10-22 10:25:58 -0600689 if (IS_ERR(worker->task)) {
690 kfree(worker);
Jens Axboeb60fda62019-11-19 08:37:07 -0700691 return false;
Jens Axboe771b53d02019-10-22 10:25:58 -0600692 }
Jens Axboea8b595b2020-10-15 10:13:07 -0600693 kthread_bind_mask(worker->task, cpumask_of_node(wqe->node));
Jens Axboe771b53d02019-10-22 10:25:58 -0600694
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200695 raw_spin_lock_irq(&wqe->lock);
Jens Axboe021d1cd2019-11-14 08:00:41 -0700696 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
Jens Axboee61df662019-11-13 13:54:49 -0700697 list_add_tail_rcu(&worker->all_list, &wqe->all_list);
Jens Axboe771b53d02019-10-22 10:25:58 -0600698 worker->flags |= IO_WORKER_F_FREE;
Jens Axboec5def4a2019-11-07 11:41:16 -0700699 if (index == IO_WQ_ACCT_BOUND)
700 worker->flags |= IO_WORKER_F_BOUND;
701 if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
Jens Axboe771b53d02019-10-22 10:25:58 -0600702 worker->flags |= IO_WORKER_F_FIXED;
Jens Axboec5def4a2019-11-07 11:41:16 -0700703 acct->nr_workers++;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200704 raw_spin_unlock_irq(&wqe->lock);
Jens Axboe771b53d02019-10-22 10:25:58 -0600705
Jens Axboec5def4a2019-11-07 11:41:16 -0700706 if (index == IO_WQ_ACCT_UNBOUND)
707 atomic_inc(&wq->user->processes);
708
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800709 refcount_inc(&wq->refs);
Jens Axboe771b53d02019-10-22 10:25:58 -0600710 wake_up_process(worker->task);
Jens Axboeb60fda62019-11-19 08:37:07 -0700711 return true;
Jens Axboe771b53d02019-10-22 10:25:58 -0600712}
713
Jens Axboec5def4a2019-11-07 11:41:16 -0700714static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
Jens Axboe771b53d02019-10-22 10:25:58 -0600715 __must_hold(wqe->lock)
716{
Jens Axboec5def4a2019-11-07 11:41:16 -0700717 struct io_wqe_acct *acct = &wqe->acct[index];
Jens Axboe771b53d02019-10-22 10:25:58 -0600718
Jens Axboec5def4a2019-11-07 11:41:16 -0700719 /* if we have available workers or no work, no need */
Jens Axboe021d1cd2019-11-14 08:00:41 -0700720 if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
Jens Axboec5def4a2019-11-07 11:41:16 -0700721 return false;
722 return acct->nr_workers < acct->max_workers;
Jens Axboe771b53d02019-10-22 10:25:58 -0600723}
724
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800725static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
726{
727 send_sig(SIGINT, worker->task, 1);
728 return false;
729}
730
731/*
732 * Iterate the passed in list and call the specific function for each
733 * worker that isn't exiting
734 */
735static bool io_wq_for_each_worker(struct io_wqe *wqe,
736 bool (*func)(struct io_worker *, void *),
737 void *data)
738{
739 struct io_worker *worker;
740 bool ret = false;
741
742 list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
743 if (io_worker_get(worker)) {
744 /* no task if node is/was offline */
745 if (worker->task)
746 ret = func(worker, data);
747 io_worker_release(worker);
748 if (ret)
749 break;
750 }
751 }
752
753 return ret;
754}
755
756static bool io_wq_worker_wake(struct io_worker *worker, void *data)
757{
758 wake_up_process(worker->task);
759 return false;
760}
761
Jens Axboe771b53d02019-10-22 10:25:58 -0600762/*
763 * Manager thread. Tasked with creating new workers, if we need them.
764 */
765static int io_wq_manager(void *data)
766{
767 struct io_wq *wq = data;
Jann Horn3fc50ab2019-11-26 19:10:20 +0100768 int node;
Jens Axboeb60fda62019-11-19 08:37:07 -0700769
770 /* create fixed workers */
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800771 refcount_set(&wq->refs, 1);
Jann Horn3fc50ab2019-11-26 19:10:20 +0100772 for_each_node(node) {
Jens Axboe75634392020-02-11 06:30:06 -0700773 if (!node_online(node))
774 continue;
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800775 if (create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
776 continue;
777 set_bit(IO_WQ_BIT_ERROR, &wq->state);
778 set_bit(IO_WQ_BIT_EXIT, &wq->state);
779 goto out;
Jens Axboeb60fda62019-11-19 08:37:07 -0700780 }
781
782 complete(&wq->done);
Jens Axboe771b53d02019-10-22 10:25:58 -0600783
784 while (!kthread_should_stop()) {
Jens Axboeaa96bf82020-04-03 11:26:26 -0600785 if (current->task_works)
786 task_work_run();
787
Jann Horn3fc50ab2019-11-26 19:10:20 +0100788 for_each_node(node) {
789 struct io_wqe *wqe = wq->wqes[node];
Jens Axboec5def4a2019-11-07 11:41:16 -0700790 bool fork_worker[2] = { false, false };
Jens Axboe771b53d02019-10-22 10:25:58 -0600791
Jens Axboe75634392020-02-11 06:30:06 -0700792 if (!node_online(node))
793 continue;
794
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200795 raw_spin_lock_irq(&wqe->lock);
Jens Axboec5def4a2019-11-07 11:41:16 -0700796 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
797 fork_worker[IO_WQ_ACCT_BOUND] = true;
798 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
799 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200800 raw_spin_unlock_irq(&wqe->lock);
Jens Axboec5def4a2019-11-07 11:41:16 -0700801 if (fork_worker[IO_WQ_ACCT_BOUND])
802 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
803 if (fork_worker[IO_WQ_ACCT_UNBOUND])
804 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
Jens Axboe771b53d02019-10-22 10:25:58 -0600805 }
806 set_current_state(TASK_INTERRUPTIBLE);
807 schedule_timeout(HZ);
808 }
809
Jens Axboeaa96bf82020-04-03 11:26:26 -0600810 if (current->task_works)
811 task_work_run();
812
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800813out:
814 if (refcount_dec_and_test(&wq->refs)) {
Jens Axboeb60fda62019-11-19 08:37:07 -0700815 complete(&wq->done);
Hillf Dantonc4068bf2020-09-26 21:26:55 +0800816 return 0;
817 }
818 /* if ERROR is set and we get here, we have workers to wake */
819 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
820 rcu_read_lock();
821 for_each_node(node)
822 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
823 rcu_read_unlock();
824 }
Jens Axboeb60fda62019-11-19 08:37:07 -0700825 return 0;
Jens Axboe771b53d02019-10-22 10:25:58 -0600826}
827
Jens Axboec5def4a2019-11-07 11:41:16 -0700828static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
829 struct io_wq_work *work)
830{
831 bool free_worker;
832
833 if (!(work->flags & IO_WQ_WORK_UNBOUND))
834 return true;
835 if (atomic_read(&acct->nr_running))
836 return true;
837
838 rcu_read_lock();
Jens Axboe021d1cd2019-11-14 08:00:41 -0700839 free_worker = !hlist_nulls_empty(&wqe->free_list);
Jens Axboec5def4a2019-11-07 11:41:16 -0700840 rcu_read_unlock();
841 if (free_worker)
842 return true;
843
844 if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
845 !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
846 return false;
847
848 return true;
849}
850
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300851static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
Pavel Begunkovfc04c392020-03-01 19:18:19 +0300852{
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300853 struct io_wq *wq = wqe->wq;
854
Pavel Begunkovfc04c392020-03-01 19:18:19 +0300855 do {
856 struct io_wq_work *old_work = work;
857
858 work->flags |= IO_WQ_WORK_CANCEL;
Pavel Begunkovf4db7182020-06-25 18:20:54 +0300859 work = wq->do_work(work);
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300860 wq->free_work(old_work);
Pavel Begunkovfc04c392020-03-01 19:18:19 +0300861 } while (work);
862}
863
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300864static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
865{
866 unsigned int hash;
867 struct io_wq_work *tail;
868
869 if (!io_wq_is_hashed(work)) {
870append:
871 wq_list_add_tail(&work->list, &wqe->work_list);
872 return;
873 }
874
875 hash = io_get_work_hash(work);
876 tail = wqe->hash_tail[hash];
877 wqe->hash_tail[hash] = work;
878 if (!tail)
879 goto append;
880
881 wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
882}
883
Jens Axboe771b53d02019-10-22 10:25:58 -0600884static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
885{
Jens Axboec5def4a2019-11-07 11:41:16 -0700886 struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
Jens Axboe895e2ca2019-12-17 08:46:33 -0700887 int work_flags;
Jens Axboe771b53d02019-10-22 10:25:58 -0600888 unsigned long flags;
889
Jens Axboec5def4a2019-11-07 11:41:16 -0700890 /*
891 * Do early check to see if we need a new unbound worker, and if we do,
892 * if we're allowed to do so. This isn't 100% accurate as there's a
893 * gap between this check and incrementing the value, but that's OK.
894 * It's close enough to not be an issue, fork() has the same delay.
895 */
896 if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
Pavel Begunkove9fd9392020-03-04 16:14:12 +0300897 io_run_cancel(work, wqe);
Jens Axboec5def4a2019-11-07 11:41:16 -0700898 return;
899 }
900
Jens Axboe895e2ca2019-12-17 08:46:33 -0700901 work_flags = work->flags;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200902 raw_spin_lock_irqsave(&wqe->lock, flags);
Pavel Begunkov86f3cd12020-03-23 22:57:22 +0300903 io_wqe_insert_work(wqe, work);
Jens Axboe771b53d02019-10-22 10:25:58 -0600904 wqe->flags &= ~IO_WQE_FLAG_STALLED;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +0200905 raw_spin_unlock_irqrestore(&wqe->lock, flags);
Jens Axboe771b53d02019-10-22 10:25:58 -0600906
Jens Axboe895e2ca2019-12-17 08:46:33 -0700907 if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
908 !atomic_read(&acct->nr_running))
Jens Axboec5def4a2019-11-07 11:41:16 -0700909 io_wqe_wake_worker(wqe, acct);
Jens Axboe771b53d02019-10-22 10:25:58 -0600910}
911
912void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
913{
914 struct io_wqe *wqe = wq->wqes[numa_node_id()];
915
916 io_wqe_enqueue(wqe, work);
917}
918
919/*
Pavel Begunkov8766dd52020-03-14 00:31:04 +0300920 * Work items that hash to the same value will not be done in parallel.
921 * Used to limit concurrent writes, generally hashed by inode.
Jens Axboe771b53d02019-10-22 10:25:58 -0600922 */
Pavel Begunkov8766dd52020-03-14 00:31:04 +0300923void io_wq_hash_work(struct io_wq_work *work, void *val)
Jens Axboe771b53d02019-10-22 10:25:58 -0600924{
Pavel Begunkov8766dd52020-03-14 00:31:04 +0300925 unsigned int bit;
Jens Axboe771b53d02019-10-22 10:25:58 -0600926
927 bit = hash_ptr(val, IO_WQ_HASH_ORDER);
928 work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
Jens Axboe771b53d02019-10-22 10:25:58 -0600929}
930
Jens Axboe771b53d02019-10-22 10:25:58 -0600931void io_wq_cancel_all(struct io_wq *wq)
932{
Jann Horn3fc50ab2019-11-26 19:10:20 +0100933 int node;
Jens Axboe771b53d02019-10-22 10:25:58 -0600934
935 set_bit(IO_WQ_BIT_CANCEL, &wq->state);
936
Jens Axboe771b53d02019-10-22 10:25:58 -0600937 rcu_read_lock();
Jann Horn3fc50ab2019-11-26 19:10:20 +0100938 for_each_node(node) {
939 struct io_wqe *wqe = wq->wqes[node];
Jens Axboe771b53d02019-10-22 10:25:58 -0600940
Jens Axboee61df662019-11-13 13:54:49 -0700941 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
Jens Axboe771b53d02019-10-22 10:25:58 -0600942 }
943 rcu_read_unlock();
944}
945
Jens Axboe62755e32019-10-28 21:49:21 -0600946struct io_cb_cancel_data {
Pavel Begunkov2293b412020-03-07 01:15:39 +0300947 work_cancel_fn *fn;
948 void *data;
Pavel Begunkov4f26bda2020-06-15 10:24:03 +0300949 int nr_running;
950 int nr_pending;
951 bool cancel_all;
Jens Axboe62755e32019-10-28 21:49:21 -0600952};
953
Pavel Begunkov2293b412020-03-07 01:15:39 +0300954static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
Jens Axboe62755e32019-10-28 21:49:21 -0600955{
Pavel Begunkov2293b412020-03-07 01:15:39 +0300956 struct io_cb_cancel_data *match = data;
Jens Axboe6f726532019-11-05 13:51:51 -0700957 unsigned long flags;
Jens Axboe62755e32019-10-28 21:49:21 -0600958
959 /*
960 * Hold the lock to avoid ->cur_work going out of scope, caller
Jens Axboe36c2f922019-11-13 09:43:34 -0700961 * may dereference the passed in work.
Jens Axboe62755e32019-10-28 21:49:21 -0600962 */
Jens Axboe36c2f922019-11-13 09:43:34 -0700963 spin_lock_irqsave(&worker->lock, flags);
Jens Axboe62755e32019-10-28 21:49:21 -0600964 if (worker->cur_work &&
Jens Axboe0c9d5cc2019-12-11 19:29:43 -0700965 !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
Pavel Begunkov2293b412020-03-07 01:15:39 +0300966 match->fn(worker->cur_work, match->data)) {
Jens Axboe771b53d02019-10-22 10:25:58 -0600967 send_sig(SIGINT, worker->task, 1);
Pavel Begunkov4f26bda2020-06-15 10:24:03 +0300968 match->nr_running++;
Jens Axboe771b53d02019-10-22 10:25:58 -0600969 }
Jens Axboe36c2f922019-11-13 09:43:34 -0700970 spin_unlock_irqrestore(&worker->lock, flags);
Jens Axboe771b53d02019-10-22 10:25:58 -0600971
Pavel Begunkov4f26bda2020-06-15 10:24:03 +0300972 return match->nr_running && !match->cancel_all;
Jens Axboe771b53d02019-10-22 10:25:58 -0600973}
974
Pavel Begunkov204361a2020-08-23 20:33:10 +0300975static inline void io_wqe_remove_pending(struct io_wqe *wqe,
976 struct io_wq_work *work,
977 struct io_wq_work_node *prev)
978{
979 unsigned int hash = io_get_work_hash(work);
980 struct io_wq_work *prev_work = NULL;
981
982 if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
983 if (prev)
984 prev_work = container_of(prev, struct io_wq_work, list);
985 if (prev_work && io_get_work_hash(prev_work) == hash)
986 wqe->hash_tail[hash] = prev_work;
987 else
988 wqe->hash_tail[hash] = NULL;
989 }
990 wq_list_del(&wqe->work_list, &work->list, prev);
991}
992
Pavel Begunkov4f26bda2020-06-15 10:24:03 +0300993static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
Pavel Begunkovf4c26652020-06-15 10:24:02 +0300994 struct io_cb_cancel_data *match)
Jens Axboe771b53d02019-10-22 10:25:58 -0600995{
Jens Axboe6206f0e2019-11-26 11:59:32 -0700996 struct io_wq_work_node *node, *prev;
Jens Axboe771b53d02019-10-22 10:25:58 -0600997 struct io_wq_work *work;
Jens Axboe6f726532019-11-05 13:51:51 -0700998 unsigned long flags;
Jens Axboe771b53d02019-10-22 10:25:58 -0600999
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001000retry:
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +02001001 raw_spin_lock_irqsave(&wqe->lock, flags);
Jens Axboe6206f0e2019-11-26 11:59:32 -07001002 wq_list_for_each(node, prev, &wqe->work_list) {
1003 work = container_of(node, struct io_wq_work, list);
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001004 if (!match->fn(work, match->data))
1005 continue;
Pavel Begunkov204361a2020-08-23 20:33:10 +03001006 io_wqe_remove_pending(wqe, work, prev);
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +02001007 raw_spin_unlock_irqrestore(&wqe->lock, flags);
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001008 io_run_cancel(work, wqe);
1009 match->nr_pending++;
1010 if (!match->cancel_all)
1011 return;
1012
1013 /* not safe to continue after unlock */
1014 goto retry;
Jens Axboe771b53d02019-10-22 10:25:58 -06001015 }
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +02001016 raw_spin_unlock_irqrestore(&wqe->lock, flags);
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001017}
Jens Axboe771b53d02019-10-22 10:25:58 -06001018
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001019static void io_wqe_cancel_running_work(struct io_wqe *wqe,
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001020 struct io_cb_cancel_data *match)
1021{
Jens Axboe771b53d02019-10-22 10:25:58 -06001022 rcu_read_lock();
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001023 io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
Jens Axboe771b53d02019-10-22 10:25:58 -06001024 rcu_read_unlock();
Jens Axboe771b53d02019-10-22 10:25:58 -06001025}
1026
Pavel Begunkov2293b412020-03-07 01:15:39 +03001027enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001028 void *data, bool cancel_all)
Pavel Begunkov2293b412020-03-07 01:15:39 +03001029{
1030 struct io_cb_cancel_data match = {
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001031 .fn = cancel,
1032 .data = data,
1033 .cancel_all = cancel_all,
Pavel Begunkov2293b412020-03-07 01:15:39 +03001034 };
Pavel Begunkov2293b412020-03-07 01:15:39 +03001035 int node;
1036
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001037 /*
1038 * First check pending list, if we're lucky we can just remove it
1039 * from there. CANCEL_OK means that the work is returned as-new,
1040 * no completion will be posted for it.
1041 */
Pavel Begunkov2293b412020-03-07 01:15:39 +03001042 for_each_node(node) {
1043 struct io_wqe *wqe = wq->wqes[node];
1044
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001045 io_wqe_cancel_pending_work(wqe, &match);
1046 if (match.nr_pending && !match.cancel_all)
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001047 return IO_WQ_CANCEL_OK;
Pavel Begunkov2293b412020-03-07 01:15:39 +03001048 }
1049
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001050 /*
1051 * Now check if a free (going busy) or busy worker has the work
1052 * currently running. If we find it there, we'll return CANCEL_RUNNING
1053 * as an indication that we attempt to signal cancellation. The
1054 * completion will run normally in this case.
1055 */
1056 for_each_node(node) {
1057 struct io_wqe *wqe = wq->wqes[node];
1058
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001059 io_wqe_cancel_running_work(wqe, &match);
1060 if (match.nr_running && !match.cancel_all)
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001061 return IO_WQ_CANCEL_RUNNING;
1062 }
1063
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001064 if (match.nr_running)
1065 return IO_WQ_CANCEL_RUNNING;
1066 if (match.nr_pending)
1067 return IO_WQ_CANCEL_OK;
Pavel Begunkovf4c26652020-06-15 10:24:02 +03001068 return IO_WQ_CANCEL_NOTFOUND;
Pavel Begunkov2293b412020-03-07 01:15:39 +03001069}
1070
1071static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data)
Jens Axboe00bcda12020-02-08 19:13:32 -07001072{
1073 return work == data;
1074}
1075
Jens Axboe771b53d02019-10-22 10:25:58 -06001076enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
1077{
Pavel Begunkov4f26bda2020-06-15 10:24:03 +03001078 return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false);
Jens Axboe771b53d02019-10-22 10:25:58 -06001079}
1080
Jens Axboe576a3472019-11-25 08:49:20 -07001081struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
Jens Axboe771b53d02019-10-22 10:25:58 -06001082{
Jann Horn3fc50ab2019-11-26 19:10:20 +01001083 int ret = -ENOMEM, node;
Jens Axboe771b53d02019-10-22 10:25:58 -06001084 struct io_wq *wq;
1085
Pavel Begunkovf5fa38c2020-06-08 21:08:20 +03001086 if (WARN_ON_ONCE(!data->free_work || !data->do_work))
Pavel Begunkove9fd9392020-03-04 16:14:12 +03001087 return ERR_PTR(-EINVAL);
1088
Jann Hornad6e0052019-11-26 17:39:45 +01001089 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
Jens Axboe771b53d02019-10-22 10:25:58 -06001090 if (!wq)
1091 return ERR_PTR(-ENOMEM);
1092
Jann Horn3fc50ab2019-11-26 19:10:20 +01001093 wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
Jens Axboe771b53d02019-10-22 10:25:58 -06001094 if (!wq->wqes) {
1095 kfree(wq);
1096 return ERR_PTR(-ENOMEM);
1097 }
1098
Pavel Begunkove9fd9392020-03-04 16:14:12 +03001099 wq->free_work = data->free_work;
Pavel Begunkovf5fa38c2020-06-08 21:08:20 +03001100 wq->do_work = data->do_work;
Jens Axboe7d723062019-11-12 22:31:31 -07001101
Jens Axboec5def4a2019-11-07 11:41:16 -07001102 /* caller must already hold a reference to this */
Jens Axboe576a3472019-11-25 08:49:20 -07001103 wq->user = data->user;
Jens Axboec5def4a2019-11-07 11:41:16 -07001104
Jann Horn3fc50ab2019-11-26 19:10:20 +01001105 for_each_node(node) {
Jens Axboe771b53d02019-10-22 10:25:58 -06001106 struct io_wqe *wqe;
Jens Axboe75634392020-02-11 06:30:06 -07001107 int alloc_node = node;
Jens Axboe771b53d02019-10-22 10:25:58 -06001108
Jens Axboe75634392020-02-11 06:30:06 -07001109 if (!node_online(alloc_node))
1110 alloc_node = NUMA_NO_NODE;
1111 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
Jens Axboe771b53d02019-10-22 10:25:58 -06001112 if (!wqe)
Jann Horn3fc50ab2019-11-26 19:10:20 +01001113 goto err;
1114 wq->wqes[node] = wqe;
Jens Axboe75634392020-02-11 06:30:06 -07001115 wqe->node = alloc_node;
Jens Axboec5def4a2019-11-07 11:41:16 -07001116 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1117 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
Jens Axboe576a3472019-11-25 08:49:20 -07001118 if (wq->user) {
Jens Axboec5def4a2019-11-07 11:41:16 -07001119 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1120 task_rlimit(current, RLIMIT_NPROC);
1121 }
1122 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
Jens Axboe771b53d02019-10-22 10:25:58 -06001123 wqe->wq = wq;
Sebastian Andrzej Siewior95da8462020-09-01 10:41:46 +02001124 raw_spin_lock_init(&wqe->lock);
Jens Axboe6206f0e2019-11-26 11:59:32 -07001125 INIT_WQ_LIST(&wqe->work_list);
Jens Axboe021d1cd2019-11-14 08:00:41 -07001126 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
Jens Axboee61df662019-11-13 13:54:49 -07001127 INIT_LIST_HEAD(&wqe->all_list);
Jens Axboe771b53d02019-10-22 10:25:58 -06001128 }
1129
1130 init_completion(&wq->done);
1131
Jens Axboe771b53d02019-10-22 10:25:58 -06001132 wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1133 if (!IS_ERR(wq->manager)) {
1134 wake_up_process(wq->manager);
Jens Axboeb60fda62019-11-19 08:37:07 -07001135 wait_for_completion(&wq->done);
1136 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1137 ret = -ENOMEM;
1138 goto err;
1139 }
Jens Axboe848f7e12020-01-23 15:33:32 -07001140 refcount_set(&wq->use_refs, 1);
Jens Axboeb60fda62019-11-19 08:37:07 -07001141 reinit_completion(&wq->done);
Jens Axboe771b53d02019-10-22 10:25:58 -06001142 return wq;
1143 }
1144
1145 ret = PTR_ERR(wq->manager);
Jens Axboe771b53d02019-10-22 10:25:58 -06001146 complete(&wq->done);
Jens Axboeb60fda62019-11-19 08:37:07 -07001147err:
Jann Horn3fc50ab2019-11-26 19:10:20 +01001148 for_each_node(node)
1149 kfree(wq->wqes[node]);
Jens Axboeb60fda62019-11-19 08:37:07 -07001150 kfree(wq->wqes);
1151 kfree(wq);
Jens Axboe771b53d02019-10-22 10:25:58 -06001152 return ERR_PTR(ret);
1153}
1154
Pavel Begunkoveba6f5a2020-01-28 03:15:47 +03001155bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1156{
Pavel Begunkovf5fa38c2020-06-08 21:08:20 +03001157 if (data->free_work != wq->free_work || data->do_work != wq->do_work)
Pavel Begunkoveba6f5a2020-01-28 03:15:47 +03001158 return false;
1159
1160 return refcount_inc_not_zero(&wq->use_refs);
1161}
1162
Jens Axboe848f7e12020-01-23 15:33:32 -07001163static void __io_wq_destroy(struct io_wq *wq)
Jens Axboe771b53d02019-10-22 10:25:58 -06001164{
Jann Horn3fc50ab2019-11-26 19:10:20 +01001165 int node;
Jens Axboe771b53d02019-10-22 10:25:58 -06001166
Jens Axboeb60fda62019-11-19 08:37:07 -07001167 set_bit(IO_WQ_BIT_EXIT, &wq->state);
1168 if (wq->manager)
Jens Axboe771b53d02019-10-22 10:25:58 -06001169 kthread_stop(wq->manager);
Jens Axboe771b53d02019-10-22 10:25:58 -06001170
1171 rcu_read_lock();
Jann Horn3fc50ab2019-11-26 19:10:20 +01001172 for_each_node(node)
1173 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
Jens Axboe771b53d02019-10-22 10:25:58 -06001174 rcu_read_unlock();
1175
1176 wait_for_completion(&wq->done);
1177
Jann Horn3fc50ab2019-11-26 19:10:20 +01001178 for_each_node(node)
1179 kfree(wq->wqes[node]);
Jens Axboe771b53d02019-10-22 10:25:58 -06001180 kfree(wq->wqes);
1181 kfree(wq);
1182}
Jens Axboe848f7e12020-01-23 15:33:32 -07001183
1184void io_wq_destroy(struct io_wq *wq)
1185{
1186 if (refcount_dec_and_test(&wq->use_refs))
1187 __io_wq_destroy(wq);
1188}
Jens Axboeaa96bf82020-04-03 11:26:26 -06001189
1190struct task_struct *io_wq_get_task(struct io_wq *wq)
1191{
1192 return wq->manager;
1193}