Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Contains the core associated with submission side polling of the SQ
4 * ring, offloading submissions from the application to a kernel thread.
5 */
6#include <linux/kernel.h>
7#include <linux/errno.h>
8#include <linux/file.h>
9#include <linux/mm.h>
10#include <linux/slab.h>
11#include <linux/audit.h>
12#include <linux/security.h>
13#include <linux/cpuset.h>
14#include <linux/io_uring.h>
15
16#include <uapi/linux/io_uring.h>
17
18#include "io_uring.h"
19#include "napi.h"
20#include "sqpoll.h"
21
22#define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
23#define IORING_TW_CAP_ENTRIES_VALUE 8
24
25enum {
26 IO_SQ_THREAD_SHOULD_STOP = 0,
27 IO_SQ_THREAD_SHOULD_PARK,
28};
29
30void io_sq_thread_unpark(struct io_sq_data *sqd)
31 __releases(&sqd->lock)
32{
33 WARN_ON_ONCE(sqd->thread == current);
34
35 /*
36 * Do the dance but not conditional clear_bit() because it'd race with
37 * other threads incrementing park_pending and setting the bit.
38 */
39 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
40 if (atomic_dec_return(&sqd->park_pending))
41 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
42 mutex_unlock(&sqd->lock);
43 wake_up(&sqd->wait);
44}
45
46void io_sq_thread_park(struct io_sq_data *sqd)
47 __acquires(&sqd->lock)
48{
49 WARN_ON_ONCE(data_race(sqd->thread) == current);
50
51 atomic_inc(&sqd->park_pending);
52 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
53 mutex_lock(&sqd->lock);
54 if (sqd->thread)
55 wake_up_process(sqd->thread);
56}
57
58void io_sq_thread_stop(struct io_sq_data *sqd)
59{
60 WARN_ON_ONCE(sqd->thread == current);
61 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
62
63 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
64 mutex_lock(&sqd->lock);
65 if (sqd->thread)
66 wake_up_process(sqd->thread);
67 mutex_unlock(&sqd->lock);
68 wait_for_completion(&sqd->exited);
69}
70
71void io_put_sq_data(struct io_sq_data *sqd)
72{
73 if (refcount_dec_and_test(&sqd->refs)) {
74 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
75
76 io_sq_thread_stop(sqd);
77 kfree(sqd);
78 }
79}
80
81static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
82{
83 struct io_ring_ctx *ctx;
84 unsigned sq_thread_idle = 0;
85
86 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
87 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
88 sqd->sq_thread_idle = sq_thread_idle;
89}
90
91void io_sq_thread_finish(struct io_ring_ctx *ctx)
92{
93 struct io_sq_data *sqd = ctx->sq_data;
94
95 if (sqd) {
96 io_sq_thread_park(sqd);
97 list_del_init(&ctx->sqd_list);
98 io_sqd_update_thread_idle(sqd);
99 io_sq_thread_unpark(sqd);
100
101 io_put_sq_data(sqd);
102 ctx->sq_data = NULL;
103 }
104}
105
106static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
107{
108 struct io_ring_ctx *ctx_attach;
109 struct io_sq_data *sqd;
110 CLASS(fd, f)(p->wq_fd);
111
112 if (fd_empty(f))
113 return ERR_PTR(-ENXIO);
114 if (!io_is_uring_fops(fd_file(f)))
115 return ERR_PTR(-EINVAL);
116
117 ctx_attach = fd_file(f)->private_data;
118 sqd = ctx_attach->sq_data;
119 if (!sqd)
120 return ERR_PTR(-EINVAL);
121 if (sqd->task_tgid != current->tgid)
122 return ERR_PTR(-EPERM);
123
124 refcount_inc(&sqd->refs);
125 return sqd;
126}
127
128static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
129 bool *attached)
130{
131 struct io_sq_data *sqd;
132
133 *attached = false;
134 if (p->flags & IORING_SETUP_ATTACH_WQ) {
135 sqd = io_attach_sq_data(p);
136 if (!IS_ERR(sqd)) {
137 *attached = true;
138 return sqd;
139 }
140 /* fall through for EPERM case, setup new sqd/task */
141 if (PTR_ERR(sqd) != -EPERM)
142 return sqd;
143 }
144
145 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
146 if (!sqd)
147 return ERR_PTR(-ENOMEM);
148
149 atomic_set(&sqd->park_pending, 0);
150 refcount_set(&sqd->refs, 1);
151 INIT_LIST_HEAD(&sqd->ctx_list);
152 mutex_init(&sqd->lock);
153 init_waitqueue_head(&sqd->wait);
154 init_completion(&sqd->exited);
155 return sqd;
156}
157
158static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
159{
160 return READ_ONCE(sqd->state);
161}
162
163static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
164{
165 unsigned int to_submit;
166 int ret = 0;
167
168 to_submit = io_sqring_entries(ctx);
169 /* if we're handling multiple rings, cap submit size for fairness */
170 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
171 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
172
173 if (to_submit || !wq_list_empty(&ctx->iopoll_list)) {
174 const struct cred *creds = NULL;
175
176 if (ctx->sq_creds != current_cred())
177 creds = override_creds(ctx->sq_creds);
178
179 mutex_lock(&ctx->uring_lock);
180 if (!wq_list_empty(&ctx->iopoll_list))
181 io_do_iopoll(ctx, true);
182
183 /*
184 * Don't submit if refs are dying, good for io_uring_register(),
185 * but also it is relied upon by io_ring_exit_work()
186 */
187 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
188 !(ctx->flags & IORING_SETUP_R_DISABLED))
189 ret = io_submit_sqes(ctx, to_submit);
190 mutex_unlock(&ctx->uring_lock);
191
192 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
193 wake_up(&ctx->sqo_sq_wait);
194 if (creds)
195 revert_creds(creds);
196 }
197
198 return ret;
199}
200
201static bool io_sqd_handle_event(struct io_sq_data *sqd)
202{
203 bool did_sig = false;
204 struct ksignal ksig;
205
206 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
207 signal_pending(current)) {
208 mutex_unlock(&sqd->lock);
209 if (signal_pending(current))
210 did_sig = get_signal(&ksig);
211 wait_event(sqd->wait, !atomic_read(&sqd->park_pending));
212 mutex_lock(&sqd->lock);
213 sqd->sq_cpu = raw_smp_processor_id();
214 }
215 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
216}
217
218/*
219 * Run task_work, processing the retry_list first. The retry_list holds
220 * entries that we passed on in the previous run, if we had more task_work
221 * than we were asked to process. Newly queued task_work isn't run until the
222 * retry list has been fully processed.
223 */
224static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
225{
226 struct io_uring_task *tctx = current->io_uring;
227 unsigned int count = 0;
228
229 if (*retry_list) {
230 *retry_list = io_handle_tw_list(*retry_list, &count, max_entries);
231 if (count >= max_entries)
232 goto out;
233 max_entries -= count;
234 }
235 *retry_list = tctx_task_work_run(tctx, max_entries, &count);
236out:
237 if (task_work_pending(current))
238 task_work_run();
239 return count;
240}
241
242static bool io_sq_tw_pending(struct llist_node *retry_list)
243{
244 struct io_uring_task *tctx = current->io_uring;
245
246 return retry_list || !llist_empty(&tctx->task_list);
247}
248
249static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
250{
251 struct rusage end;
252
253 getrusage(current, RUSAGE_SELF, &end);
254 end.ru_stime.tv_sec -= start->ru_stime.tv_sec;
255 end.ru_stime.tv_usec -= start->ru_stime.tv_usec;
256
257 sqd->work_time += end.ru_stime.tv_usec + end.ru_stime.tv_sec * 1000000;
258}
259
260static int io_sq_thread(void *data)
261{
262 struct llist_node *retry_list = NULL;
263 struct io_sq_data *sqd = data;
264 struct io_ring_ctx *ctx;
265 struct rusage start;
266 unsigned long timeout = 0;
267 char buf[TASK_COMM_LEN];
268 DEFINE_WAIT(wait);
269
270 /* offload context creation failed, just exit */
271 if (!current->io_uring) {
272 mutex_lock(&sqd->lock);
273 sqd->thread = NULL;
274 mutex_unlock(&sqd->lock);
275 goto err_out;
276 }
277
278 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
279 set_task_comm(current, buf);
280
281 /* reset to our pid after we've set task_comm, for fdinfo */
282 sqd->task_pid = current->pid;
283
284 if (sqd->sq_cpu != -1) {
285 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
286 } else {
287 set_cpus_allowed_ptr(current, cpu_online_mask);
288 sqd->sq_cpu = raw_smp_processor_id();
289 }
290
291 /*
292 * Force audit context to get setup, in case we do prep side async
293 * operations that would trigger an audit call before any issue side
294 * audit has been done.
295 */
296 audit_uring_entry(IORING_OP_NOP);
297 audit_uring_exit(true, 0);
298
299 mutex_lock(&sqd->lock);
300 while (1) {
301 bool cap_entries, sqt_spin = false;
302
303 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
304 if (io_sqd_handle_event(sqd))
305 break;
306 timeout = jiffies + sqd->sq_thread_idle;
307 }
308
309 cap_entries = !list_is_singular(&sqd->ctx_list);
310 getrusage(current, RUSAGE_SELF, &start);
311 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
312 int ret = __io_sq_thread(ctx, cap_entries);
313
314 if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
315 sqt_spin = true;
316 }
317 if (io_sq_tw(&retry_list, IORING_TW_CAP_ENTRIES_VALUE))
318 sqt_spin = true;
319
320 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
321 if (io_napi(ctx))
322 io_napi_sqpoll_busy_poll(ctx);
323
324 if (sqt_spin || !time_after(jiffies, timeout)) {
325 if (sqt_spin) {
326 io_sq_update_worktime(sqd, &start);
327 timeout = jiffies + sqd->sq_thread_idle;
328 }
329 if (unlikely(need_resched())) {
330 mutex_unlock(&sqd->lock);
331 cond_resched();
332 mutex_lock(&sqd->lock);
333 sqd->sq_cpu = raw_smp_processor_id();
334 }
335 continue;
336 }
337
338 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
339 if (!io_sqd_events_pending(sqd) && !io_sq_tw_pending(retry_list)) {
340 bool needs_sched = true;
341
342 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
343 atomic_or(IORING_SQ_NEED_WAKEUP,
344 &ctx->rings->sq_flags);
345 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
346 !wq_list_empty(&ctx->iopoll_list)) {
347 needs_sched = false;
348 break;
349 }
350
351 /*
352 * Ensure the store of the wakeup flag is not
353 * reordered with the load of the SQ tail
354 */
355 smp_mb__after_atomic();
356
357 if (io_sqring_entries(ctx)) {
358 needs_sched = false;
359 break;
360 }
361 }
362
363 if (needs_sched) {
364 mutex_unlock(&sqd->lock);
365 schedule();
366 mutex_lock(&sqd->lock);
367 sqd->sq_cpu = raw_smp_processor_id();
368 }
369 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
370 atomic_andnot(IORING_SQ_NEED_WAKEUP,
371 &ctx->rings->sq_flags);
372 }
373
374 finish_wait(&sqd->wait, &wait);
375 timeout = jiffies + sqd->sq_thread_idle;
376 }
377
378 if (retry_list)
379 io_sq_tw(&retry_list, UINT_MAX);
380
381 io_uring_cancel_generic(true, sqd);
382 sqd->thread = NULL;
383 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
384 atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
385 io_run_task_work();
386 mutex_unlock(&sqd->lock);
387err_out:
388 complete(&sqd->exited);
389 do_exit(0);
390}
391
392void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
393{
394 DEFINE_WAIT(wait);
395
396 do {
397 if (!io_sqring_full(ctx))
398 break;
399 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
400
401 if (!io_sqring_full(ctx))
402 break;
403 schedule();
404 } while (!signal_pending(current));
405
406 finish_wait(&ctx->sqo_sq_wait, &wait);
407}
408
409__cold int io_sq_offload_create(struct io_ring_ctx *ctx,
410 struct io_uring_params *p)
411{
412 struct task_struct *task_to_put = NULL;
413 int ret;
414
415 /* Retain compatibility with failing for an invalid attach attempt */
416 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
417 IORING_SETUP_ATTACH_WQ) {
418 CLASS(fd, f)(p->wq_fd);
419 if (fd_empty(f))
420 return -ENXIO;
421 if (!io_is_uring_fops(fd_file(f)))
422 return -EINVAL;
423 }
424 if (ctx->flags & IORING_SETUP_SQPOLL) {
425 struct task_struct *tsk;
426 struct io_sq_data *sqd;
427 bool attached;
428
429 ret = security_uring_sqpoll();
430 if (ret)
431 return ret;
432
433 sqd = io_get_sq_data(p, &attached);
434 if (IS_ERR(sqd)) {
435 ret = PTR_ERR(sqd);
436 goto err;
437 }
438
439 ctx->sq_creds = get_current_cred();
440 ctx->sq_data = sqd;
441 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
442 if (!ctx->sq_thread_idle)
443 ctx->sq_thread_idle = HZ;
444
445 io_sq_thread_park(sqd);
446 list_add(&ctx->sqd_list, &sqd->ctx_list);
447 io_sqd_update_thread_idle(sqd);
448 /* don't attach to a dying SQPOLL thread, would be racy */
449 ret = (attached && !sqd->thread) ? -ENXIO : 0;
450 io_sq_thread_unpark(sqd);
451
452 if (ret < 0)
453 goto err;
454 if (attached)
455 return 0;
456
457 if (p->flags & IORING_SETUP_SQ_AFF) {
458 cpumask_var_t allowed_mask;
459 int cpu = p->sq_thread_cpu;
460
461 ret = -EINVAL;
462 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
463 goto err_sqpoll;
464 ret = -ENOMEM;
465 if (!alloc_cpumask_var(&allowed_mask, GFP_KERNEL))
466 goto err_sqpoll;
467 ret = -EINVAL;
468 cpuset_cpus_allowed(current, allowed_mask);
469 if (!cpumask_test_cpu(cpu, allowed_mask)) {
470 free_cpumask_var(allowed_mask);
471 goto err_sqpoll;
472 }
473 free_cpumask_var(allowed_mask);
474 sqd->sq_cpu = cpu;
475 } else {
476 sqd->sq_cpu = -1;
477 }
478
479 sqd->task_pid = current->pid;
480 sqd->task_tgid = current->tgid;
481 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
482 if (IS_ERR(tsk)) {
483 ret = PTR_ERR(tsk);
484 goto err_sqpoll;
485 }
486
487 sqd->thread = tsk;
488 task_to_put = get_task_struct(tsk);
489 ret = io_uring_alloc_task_context(tsk, ctx);
490 wake_up_new_task(tsk);
491 if (ret)
492 goto err;
493 } else if (p->flags & IORING_SETUP_SQ_AFF) {
494 /* Can't have SQ_AFF without SQPOLL */
495 ret = -EINVAL;
496 goto err;
497 }
498
499 if (task_to_put)
500 put_task_struct(task_to_put);
501 return 0;
502err_sqpoll:
503 complete(&ctx->sq_data->exited);
504err:
505 io_sq_thread_finish(ctx);
506 if (task_to_put)
507 put_task_struct(task_to_put);
508 return ret;
509}
510
511__cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
512 cpumask_var_t mask)
513{
514 struct io_sq_data *sqd = ctx->sq_data;
515 int ret = -EINVAL;
516
517 if (sqd) {
518 io_sq_thread_park(sqd);
519 /* Don't set affinity for a dying thread */
520 if (sqd->thread)
521 ret = io_wq_cpu_affinity(sqd->thread->io_uring, mask);
522 io_sq_thread_unpark(sqd);
523 }
524
525 return ret;
526}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Contains the core associated with submission side polling of the SQ
4 * ring, offloading submissions from the application to a kernel thread.
5 */
6#include <linux/kernel.h>
7#include <linux/errno.h>
8#include <linux/file.h>
9#include <linux/mm.h>
10#include <linux/slab.h>
11#include <linux/audit.h>
12#include <linux/security.h>
13#include <linux/io_uring.h>
14
15#include <uapi/linux/io_uring.h>
16
17#include "io_uring.h"
18#include "napi.h"
19#include "sqpoll.h"
20
21#define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
22#define IORING_TW_CAP_ENTRIES_VALUE 8
23
24enum {
25 IO_SQ_THREAD_SHOULD_STOP = 0,
26 IO_SQ_THREAD_SHOULD_PARK,
27};
28
29void io_sq_thread_unpark(struct io_sq_data *sqd)
30 __releases(&sqd->lock)
31{
32 WARN_ON_ONCE(sqd->thread == current);
33
34 /*
35 * Do the dance but not conditional clear_bit() because it'd race with
36 * other threads incrementing park_pending and setting the bit.
37 */
38 clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
39 if (atomic_dec_return(&sqd->park_pending))
40 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
41 mutex_unlock(&sqd->lock);
42}
43
44void io_sq_thread_park(struct io_sq_data *sqd)
45 __acquires(&sqd->lock)
46{
47 WARN_ON_ONCE(sqd->thread == current);
48
49 atomic_inc(&sqd->park_pending);
50 set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
51 mutex_lock(&sqd->lock);
52 if (sqd->thread)
53 wake_up_process(sqd->thread);
54}
55
56void io_sq_thread_stop(struct io_sq_data *sqd)
57{
58 WARN_ON_ONCE(sqd->thread == current);
59 WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
60
61 set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
62 mutex_lock(&sqd->lock);
63 if (sqd->thread)
64 wake_up_process(sqd->thread);
65 mutex_unlock(&sqd->lock);
66 wait_for_completion(&sqd->exited);
67}
68
69void io_put_sq_data(struct io_sq_data *sqd)
70{
71 if (refcount_dec_and_test(&sqd->refs)) {
72 WARN_ON_ONCE(atomic_read(&sqd->park_pending));
73
74 io_sq_thread_stop(sqd);
75 kfree(sqd);
76 }
77}
78
79static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
80{
81 struct io_ring_ctx *ctx;
82 unsigned sq_thread_idle = 0;
83
84 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
85 sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
86 sqd->sq_thread_idle = sq_thread_idle;
87}
88
89void io_sq_thread_finish(struct io_ring_ctx *ctx)
90{
91 struct io_sq_data *sqd = ctx->sq_data;
92
93 if (sqd) {
94 io_sq_thread_park(sqd);
95 list_del_init(&ctx->sqd_list);
96 io_sqd_update_thread_idle(sqd);
97 io_sq_thread_unpark(sqd);
98
99 io_put_sq_data(sqd);
100 ctx->sq_data = NULL;
101 }
102}
103
104static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
105{
106 struct io_ring_ctx *ctx_attach;
107 struct io_sq_data *sqd;
108 struct fd f;
109
110 f = fdget(p->wq_fd);
111 if (!f.file)
112 return ERR_PTR(-ENXIO);
113 if (!io_is_uring_fops(f.file)) {
114 fdput(f);
115 return ERR_PTR(-EINVAL);
116 }
117
118 ctx_attach = f.file->private_data;
119 sqd = ctx_attach->sq_data;
120 if (!sqd) {
121 fdput(f);
122 return ERR_PTR(-EINVAL);
123 }
124 if (sqd->task_tgid != current->tgid) {
125 fdput(f);
126 return ERR_PTR(-EPERM);
127 }
128
129 refcount_inc(&sqd->refs);
130 fdput(f);
131 return sqd;
132}
133
134static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
135 bool *attached)
136{
137 struct io_sq_data *sqd;
138
139 *attached = false;
140 if (p->flags & IORING_SETUP_ATTACH_WQ) {
141 sqd = io_attach_sq_data(p);
142 if (!IS_ERR(sqd)) {
143 *attached = true;
144 return sqd;
145 }
146 /* fall through for EPERM case, setup new sqd/task */
147 if (PTR_ERR(sqd) != -EPERM)
148 return sqd;
149 }
150
151 sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
152 if (!sqd)
153 return ERR_PTR(-ENOMEM);
154
155 atomic_set(&sqd->park_pending, 0);
156 refcount_set(&sqd->refs, 1);
157 INIT_LIST_HEAD(&sqd->ctx_list);
158 mutex_init(&sqd->lock);
159 init_waitqueue_head(&sqd->wait);
160 init_completion(&sqd->exited);
161 return sqd;
162}
163
164static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
165{
166 return READ_ONCE(sqd->state);
167}
168
169static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
170{
171 unsigned int to_submit;
172 int ret = 0;
173
174 to_submit = io_sqring_entries(ctx);
175 /* if we're handling multiple rings, cap submit size for fairness */
176 if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
177 to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
178
179 if (!wq_list_empty(&ctx->iopoll_list) || to_submit) {
180 const struct cred *creds = NULL;
181
182 if (ctx->sq_creds != current_cred())
183 creds = override_creds(ctx->sq_creds);
184
185 mutex_lock(&ctx->uring_lock);
186 if (!wq_list_empty(&ctx->iopoll_list))
187 io_do_iopoll(ctx, true);
188
189 /*
190 * Don't submit if refs are dying, good for io_uring_register(),
191 * but also it is relied upon by io_ring_exit_work()
192 */
193 if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
194 !(ctx->flags & IORING_SETUP_R_DISABLED))
195 ret = io_submit_sqes(ctx, to_submit);
196 mutex_unlock(&ctx->uring_lock);
197
198 if (io_napi(ctx))
199 ret += io_napi_sqpoll_busy_poll(ctx);
200
201 if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
202 wake_up(&ctx->sqo_sq_wait);
203 if (creds)
204 revert_creds(creds);
205 }
206
207 return ret;
208}
209
210static bool io_sqd_handle_event(struct io_sq_data *sqd)
211{
212 bool did_sig = false;
213 struct ksignal ksig;
214
215 if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
216 signal_pending(current)) {
217 mutex_unlock(&sqd->lock);
218 if (signal_pending(current))
219 did_sig = get_signal(&ksig);
220 cond_resched();
221 mutex_lock(&sqd->lock);
222 sqd->sq_cpu = raw_smp_processor_id();
223 }
224 return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
225}
226
227/*
228 * Run task_work, processing the retry_list first. The retry_list holds
229 * entries that we passed on in the previous run, if we had more task_work
230 * than we were asked to process. Newly queued task_work isn't run until the
231 * retry list has been fully processed.
232 */
233static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
234{
235 struct io_uring_task *tctx = current->io_uring;
236 unsigned int count = 0;
237
238 if (*retry_list) {
239 *retry_list = io_handle_tw_list(*retry_list, &count, max_entries);
240 if (count >= max_entries)
241 goto out;
242 max_entries -= count;
243 }
244 *retry_list = tctx_task_work_run(tctx, max_entries, &count);
245out:
246 if (task_work_pending(current))
247 task_work_run();
248 return count;
249}
250
251static bool io_sq_tw_pending(struct llist_node *retry_list)
252{
253 struct io_uring_task *tctx = current->io_uring;
254
255 return retry_list || !llist_empty(&tctx->task_list);
256}
257
258static void io_sq_update_worktime(struct io_sq_data *sqd, struct rusage *start)
259{
260 struct rusage end;
261
262 getrusage(current, RUSAGE_SELF, &end);
263 end.ru_stime.tv_sec -= start->ru_stime.tv_sec;
264 end.ru_stime.tv_usec -= start->ru_stime.tv_usec;
265
266 sqd->work_time += end.ru_stime.tv_usec + end.ru_stime.tv_sec * 1000000;
267}
268
269static int io_sq_thread(void *data)
270{
271 struct llist_node *retry_list = NULL;
272 struct io_sq_data *sqd = data;
273 struct io_ring_ctx *ctx;
274 struct rusage start;
275 unsigned long timeout = 0;
276 char buf[TASK_COMM_LEN];
277 DEFINE_WAIT(wait);
278
279 /* offload context creation failed, just exit */
280 if (!current->io_uring)
281 goto err_out;
282
283 snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
284 set_task_comm(current, buf);
285
286 /* reset to our pid after we've set task_comm, for fdinfo */
287 sqd->task_pid = current->pid;
288
289 if (sqd->sq_cpu != -1) {
290 set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
291 } else {
292 set_cpus_allowed_ptr(current, cpu_online_mask);
293 sqd->sq_cpu = raw_smp_processor_id();
294 }
295
296 mutex_lock(&sqd->lock);
297 while (1) {
298 bool cap_entries, sqt_spin = false;
299
300 if (io_sqd_events_pending(sqd) || signal_pending(current)) {
301 if (io_sqd_handle_event(sqd))
302 break;
303 timeout = jiffies + sqd->sq_thread_idle;
304 }
305
306 cap_entries = !list_is_singular(&sqd->ctx_list);
307 getrusage(current, RUSAGE_SELF, &start);
308 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
309 int ret = __io_sq_thread(ctx, cap_entries);
310
311 if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
312 sqt_spin = true;
313 }
314 if (io_sq_tw(&retry_list, IORING_TW_CAP_ENTRIES_VALUE))
315 sqt_spin = true;
316
317 if (sqt_spin || !time_after(jiffies, timeout)) {
318 if (sqt_spin) {
319 io_sq_update_worktime(sqd, &start);
320 timeout = jiffies + sqd->sq_thread_idle;
321 }
322 if (unlikely(need_resched())) {
323 mutex_unlock(&sqd->lock);
324 cond_resched();
325 mutex_lock(&sqd->lock);
326 sqd->sq_cpu = raw_smp_processor_id();
327 }
328 continue;
329 }
330
331 prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
332 if (!io_sqd_events_pending(sqd) && !io_sq_tw_pending(retry_list)) {
333 bool needs_sched = true;
334
335 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
336 atomic_or(IORING_SQ_NEED_WAKEUP,
337 &ctx->rings->sq_flags);
338 if ((ctx->flags & IORING_SETUP_IOPOLL) &&
339 !wq_list_empty(&ctx->iopoll_list)) {
340 needs_sched = false;
341 break;
342 }
343
344 /*
345 * Ensure the store of the wakeup flag is not
346 * reordered with the load of the SQ tail
347 */
348 smp_mb__after_atomic();
349
350 if (io_sqring_entries(ctx)) {
351 needs_sched = false;
352 break;
353 }
354 }
355
356 if (needs_sched) {
357 mutex_unlock(&sqd->lock);
358 schedule();
359 mutex_lock(&sqd->lock);
360 sqd->sq_cpu = raw_smp_processor_id();
361 }
362 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
363 atomic_andnot(IORING_SQ_NEED_WAKEUP,
364 &ctx->rings->sq_flags);
365 }
366
367 finish_wait(&sqd->wait, &wait);
368 timeout = jiffies + sqd->sq_thread_idle;
369 }
370
371 if (retry_list)
372 io_sq_tw(&retry_list, UINT_MAX);
373
374 io_uring_cancel_generic(true, sqd);
375 sqd->thread = NULL;
376 list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
377 atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
378 io_run_task_work();
379 mutex_unlock(&sqd->lock);
380err_out:
381 complete(&sqd->exited);
382 do_exit(0);
383}
384
385void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
386{
387 DEFINE_WAIT(wait);
388
389 do {
390 if (!io_sqring_full(ctx))
391 break;
392 prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
393
394 if (!io_sqring_full(ctx))
395 break;
396 schedule();
397 } while (!signal_pending(current));
398
399 finish_wait(&ctx->sqo_sq_wait, &wait);
400}
401
402__cold int io_sq_offload_create(struct io_ring_ctx *ctx,
403 struct io_uring_params *p)
404{
405 int ret;
406
407 /* Retain compatibility with failing for an invalid attach attempt */
408 if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
409 IORING_SETUP_ATTACH_WQ) {
410 struct fd f;
411
412 f = fdget(p->wq_fd);
413 if (!f.file)
414 return -ENXIO;
415 if (!io_is_uring_fops(f.file)) {
416 fdput(f);
417 return -EINVAL;
418 }
419 fdput(f);
420 }
421 if (ctx->flags & IORING_SETUP_SQPOLL) {
422 struct task_struct *tsk;
423 struct io_sq_data *sqd;
424 bool attached;
425
426 ret = security_uring_sqpoll();
427 if (ret)
428 return ret;
429
430 sqd = io_get_sq_data(p, &attached);
431 if (IS_ERR(sqd)) {
432 ret = PTR_ERR(sqd);
433 goto err;
434 }
435
436 ctx->sq_creds = get_current_cred();
437 ctx->sq_data = sqd;
438 ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
439 if (!ctx->sq_thread_idle)
440 ctx->sq_thread_idle = HZ;
441
442 io_sq_thread_park(sqd);
443 list_add(&ctx->sqd_list, &sqd->ctx_list);
444 io_sqd_update_thread_idle(sqd);
445 /* don't attach to a dying SQPOLL thread, would be racy */
446 ret = (attached && !sqd->thread) ? -ENXIO : 0;
447 io_sq_thread_unpark(sqd);
448
449 if (ret < 0)
450 goto err;
451 if (attached)
452 return 0;
453
454 if (p->flags & IORING_SETUP_SQ_AFF) {
455 int cpu = p->sq_thread_cpu;
456
457 ret = -EINVAL;
458 if (cpu >= nr_cpu_ids || !cpu_online(cpu))
459 goto err_sqpoll;
460 sqd->sq_cpu = cpu;
461 } else {
462 sqd->sq_cpu = -1;
463 }
464
465 sqd->task_pid = current->pid;
466 sqd->task_tgid = current->tgid;
467 tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
468 if (IS_ERR(tsk)) {
469 ret = PTR_ERR(tsk);
470 goto err_sqpoll;
471 }
472
473 sqd->thread = tsk;
474 ret = io_uring_alloc_task_context(tsk, ctx);
475 wake_up_new_task(tsk);
476 if (ret)
477 goto err;
478 } else if (p->flags & IORING_SETUP_SQ_AFF) {
479 /* Can't have SQ_AFF without SQPOLL */
480 ret = -EINVAL;
481 goto err;
482 }
483
484 return 0;
485err_sqpoll:
486 complete(&ctx->sq_data->exited);
487err:
488 io_sq_thread_finish(ctx);
489 return ret;
490}
491
492__cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
493 cpumask_var_t mask)
494{
495 struct io_sq_data *sqd = ctx->sq_data;
496 int ret = -EINVAL;
497
498 if (sqd) {
499 io_sq_thread_park(sqd);
500 /* Don't set affinity for a dying thread */
501 if (sqd->thread)
502 ret = io_wq_cpu_affinity(sqd->thread->io_uring, mask);
503 io_sq_thread_unpark(sqd);
504 }
505
506 return ret;
507}