xref: /linux/io_uring/sqpoll.c (revision 3f1c07fc21c68bd3bd2df9d2c9441f6485e934d9)
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/sched/cputime.h>
15 #include <linux/io_uring.h>
16 
17 #include <uapi/linux/io_uring.h>
18 
19 #include "io_uring.h"
20 #include "tctx.h"
21 #include "napi.h"
22 #include "cancel.h"
23 #include "sqpoll.h"
24 
25 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
26 #define IORING_TW_CAP_ENTRIES_VALUE	32
27 
28 enum {
29 	IO_SQ_THREAD_SHOULD_STOP = 0,
30 	IO_SQ_THREAD_SHOULD_PARK,
31 };
32 
io_sq_thread_unpark(struct io_sq_data * sqd)33 void io_sq_thread_unpark(struct io_sq_data *sqd)
34 	__releases(&sqd->lock)
35 {
36 	WARN_ON_ONCE(sqpoll_task_locked(sqd) == current);
37 
38 	/*
39 	 * Do the dance but not conditional clear_bit() because it'd race with
40 	 * other threads incrementing park_pending and setting the bit.
41 	 */
42 	clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
43 	if (atomic_dec_return(&sqd->park_pending))
44 		set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
45 	mutex_unlock(&sqd->lock);
46 	wake_up(&sqd->wait);
47 }
48 
io_sq_thread_park(struct io_sq_data * sqd)49 void io_sq_thread_park(struct io_sq_data *sqd)
50 	__acquires(&sqd->lock)
51 {
52 	struct task_struct *tsk;
53 
54 	atomic_inc(&sqd->park_pending);
55 	set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
56 	mutex_lock(&sqd->lock);
57 
58 	tsk = sqpoll_task_locked(sqd);
59 	if (tsk) {
60 		WARN_ON_ONCE(tsk == current);
61 		wake_up_process(tsk);
62 	}
63 }
64 
io_sq_thread_stop(struct io_sq_data * sqd)65 void io_sq_thread_stop(struct io_sq_data *sqd)
66 {
67 	struct task_struct *tsk;
68 
69 	WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
70 
71 	set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
72 	mutex_lock(&sqd->lock);
73 	tsk = sqpoll_task_locked(sqd);
74 	if (tsk) {
75 		WARN_ON_ONCE(tsk == current);
76 		wake_up_process(tsk);
77 	}
78 	mutex_unlock(&sqd->lock);
79 	wait_for_completion(&sqd->exited);
80 }
81 
io_put_sq_data(struct io_sq_data * sqd)82 void io_put_sq_data(struct io_sq_data *sqd)
83 {
84 	if (refcount_dec_and_test(&sqd->refs)) {
85 		WARN_ON_ONCE(atomic_read(&sqd->park_pending));
86 
87 		io_sq_thread_stop(sqd);
88 		kfree(sqd);
89 	}
90 }
91 
io_sqd_update_thread_idle(struct io_sq_data * sqd)92 static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
93 {
94 	struct io_ring_ctx *ctx;
95 	unsigned sq_thread_idle = 0;
96 
97 	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
98 		sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
99 	sqd->sq_thread_idle = sq_thread_idle;
100 }
101 
io_sq_thread_finish(struct io_ring_ctx * ctx)102 void io_sq_thread_finish(struct io_ring_ctx *ctx)
103 {
104 	struct io_sq_data *sqd = ctx->sq_data;
105 
106 	if (sqd) {
107 		io_sq_thread_park(sqd);
108 		list_del_init(&ctx->sqd_list);
109 		io_sqd_update_thread_idle(sqd);
110 		io_sq_thread_unpark(sqd);
111 
112 		io_put_sq_data(sqd);
113 		ctx->sq_data = NULL;
114 	}
115 }
116 
io_attach_sq_data(struct io_uring_params * p)117 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
118 {
119 	struct io_ring_ctx *ctx_attach;
120 	struct io_sq_data *sqd;
121 	CLASS(fd, f)(p->wq_fd);
122 
123 	if (fd_empty(f))
124 		return ERR_PTR(-ENXIO);
125 	if (!io_is_uring_fops(fd_file(f)))
126 		return ERR_PTR(-EINVAL);
127 
128 	ctx_attach = fd_file(f)->private_data;
129 	sqd = ctx_attach->sq_data;
130 	if (!sqd)
131 		return ERR_PTR(-EINVAL);
132 	if (sqd->task_tgid != current->tgid)
133 		return ERR_PTR(-EPERM);
134 
135 	refcount_inc(&sqd->refs);
136 	return sqd;
137 }
138 
io_get_sq_data(struct io_uring_params * p,bool * attached)139 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
140 					 bool *attached)
141 {
142 	struct io_sq_data *sqd;
143 
144 	*attached = false;
145 	if (p->flags & IORING_SETUP_ATTACH_WQ) {
146 		sqd = io_attach_sq_data(p);
147 		if (!IS_ERR(sqd)) {
148 			*attached = true;
149 			return sqd;
150 		}
151 		/* fall through for EPERM case, setup new sqd/task */
152 		if (PTR_ERR(sqd) != -EPERM)
153 			return sqd;
154 	}
155 
156 	sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
157 	if (!sqd)
158 		return ERR_PTR(-ENOMEM);
159 
160 	atomic_set(&sqd->park_pending, 0);
161 	refcount_set(&sqd->refs, 1);
162 	INIT_LIST_HEAD(&sqd->ctx_list);
163 	mutex_init(&sqd->lock);
164 	init_waitqueue_head(&sqd->wait);
165 	init_completion(&sqd->exited);
166 	return sqd;
167 }
168 
io_sqd_events_pending(struct io_sq_data * sqd)169 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
170 {
171 	return READ_ONCE(sqd->state);
172 }
173 
174 struct io_sq_time {
175 	bool started;
176 	u64 usec;
177 };
178 
io_sq_cpu_usec(struct task_struct * tsk)179 u64 io_sq_cpu_usec(struct task_struct *tsk)
180 {
181 	u64 utime, stime;
182 
183 	task_cputime_adjusted(tsk, &utime, &stime);
184 	do_div(stime, 1000);
185 	return stime;
186 }
187 
io_sq_update_worktime(struct io_sq_data * sqd,struct io_sq_time * ist)188 static void io_sq_update_worktime(struct io_sq_data *sqd, struct io_sq_time *ist)
189 {
190 	if (!ist->started)
191 		return;
192 	ist->started = false;
193 	sqd->work_time += io_sq_cpu_usec(current) - ist->usec;
194 }
195 
io_sq_start_worktime(struct io_sq_time * ist)196 static void io_sq_start_worktime(struct io_sq_time *ist)
197 {
198 	if (ist->started)
199 		return;
200 	ist->started = true;
201 	ist->usec = io_sq_cpu_usec(current);
202 }
203 
__io_sq_thread(struct io_ring_ctx * ctx,struct io_sq_data * sqd,bool cap_entries,struct io_sq_time * ist)204 static int __io_sq_thread(struct io_ring_ctx *ctx, struct io_sq_data *sqd,
205 			  bool cap_entries, struct io_sq_time *ist)
206 {
207 	unsigned int to_submit;
208 	int ret = 0;
209 
210 	to_submit = io_sqring_entries(ctx);
211 	/* if we're handling multiple rings, cap submit size for fairness */
212 	if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
213 		to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
214 
215 	if (to_submit || !wq_list_empty(&ctx->iopoll_list)) {
216 		const struct cred *creds = NULL;
217 
218 		io_sq_start_worktime(ist);
219 
220 		if (ctx->sq_creds != current_cred())
221 			creds = override_creds(ctx->sq_creds);
222 
223 		mutex_lock(&ctx->uring_lock);
224 		if (!wq_list_empty(&ctx->iopoll_list))
225 			io_do_iopoll(ctx, true);
226 
227 		/*
228 		 * Don't submit if refs are dying, good for io_uring_register(),
229 		 * but also it is relied upon by io_ring_exit_work()
230 		 */
231 		if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
232 		    !(ctx->flags & IORING_SETUP_R_DISABLED))
233 			ret = io_submit_sqes(ctx, to_submit);
234 		mutex_unlock(&ctx->uring_lock);
235 
236 		if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
237 			wake_up(&ctx->sqo_sq_wait);
238 		if (creds)
239 			revert_creds(creds);
240 	}
241 
242 	return ret;
243 }
244 
io_sqd_handle_event(struct io_sq_data * sqd)245 static bool io_sqd_handle_event(struct io_sq_data *sqd)
246 {
247 	bool did_sig = false;
248 	struct ksignal ksig;
249 
250 	if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
251 	    signal_pending(current)) {
252 		mutex_unlock(&sqd->lock);
253 		if (signal_pending(current))
254 			did_sig = get_signal(&ksig);
255 		wait_event(sqd->wait, !atomic_read(&sqd->park_pending));
256 		mutex_lock(&sqd->lock);
257 		sqd->sq_cpu = raw_smp_processor_id();
258 	}
259 	return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
260 }
261 
262 /*
263  * Run task_work, processing the retry_list first. The retry_list holds
264  * entries that we passed on in the previous run, if we had more task_work
265  * than we were asked to process. Newly queued task_work isn't run until the
266  * retry list has been fully processed.
267  */
io_sq_tw(struct llist_node ** retry_list,int max_entries)268 static unsigned int io_sq_tw(struct llist_node **retry_list, int max_entries)
269 {
270 	struct io_uring_task *tctx = current->io_uring;
271 	unsigned int count = 0;
272 
273 	if (*retry_list) {
274 		*retry_list = io_handle_tw_list(*retry_list, &count, max_entries);
275 		if (count >= max_entries)
276 			goto out;
277 		max_entries -= count;
278 	}
279 	*retry_list = tctx_task_work_run(tctx, max_entries, &count);
280 out:
281 	if (task_work_pending(current))
282 		task_work_run();
283 	return count;
284 }
285 
io_sq_tw_pending(struct llist_node * retry_list)286 static bool io_sq_tw_pending(struct llist_node *retry_list)
287 {
288 	struct io_uring_task *tctx = current->io_uring;
289 
290 	return retry_list || !llist_empty(&tctx->task_list);
291 }
292 
io_sq_thread(void * data)293 static int io_sq_thread(void *data)
294 {
295 	struct llist_node *retry_list = NULL;
296 	struct io_sq_data *sqd = data;
297 	struct io_ring_ctx *ctx;
298 	unsigned long timeout = 0;
299 	char buf[TASK_COMM_LEN] = {};
300 	DEFINE_WAIT(wait);
301 
302 	/* offload context creation failed, just exit */
303 	if (!current->io_uring) {
304 		mutex_lock(&sqd->lock);
305 		rcu_assign_pointer(sqd->thread, NULL);
306 		put_task_struct(current);
307 		mutex_unlock(&sqd->lock);
308 		goto err_out;
309 	}
310 
311 	snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
312 	set_task_comm(current, buf);
313 
314 	/* reset to our pid after we've set task_comm, for fdinfo */
315 	sqd->task_pid = current->pid;
316 
317 	if (sqd->sq_cpu != -1) {
318 		set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
319 	} else {
320 		set_cpus_allowed_ptr(current, cpu_online_mask);
321 		sqd->sq_cpu = raw_smp_processor_id();
322 	}
323 
324 	/*
325 	 * Force audit context to get setup, in case we do prep side async
326 	 * operations that would trigger an audit call before any issue side
327 	 * audit has been done.
328 	 */
329 	audit_uring_entry(IORING_OP_NOP);
330 	audit_uring_exit(true, 0);
331 
332 	mutex_lock(&sqd->lock);
333 	while (1) {
334 		bool cap_entries, sqt_spin = false;
335 		struct io_sq_time ist = { };
336 
337 		if (io_sqd_events_pending(sqd) || signal_pending(current)) {
338 			if (io_sqd_handle_event(sqd))
339 				break;
340 			timeout = jiffies + sqd->sq_thread_idle;
341 		}
342 
343 		cap_entries = !list_is_singular(&sqd->ctx_list);
344 		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
345 			int ret = __io_sq_thread(ctx, sqd, cap_entries, &ist);
346 
347 			if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
348 				sqt_spin = true;
349 		}
350 		if (io_sq_tw(&retry_list, IORING_TW_CAP_ENTRIES_VALUE))
351 			sqt_spin = true;
352 
353 		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
354 			if (io_napi(ctx)) {
355 				io_sq_start_worktime(&ist);
356 				io_napi_sqpoll_busy_poll(ctx);
357 			}
358 		}
359 
360 		io_sq_update_worktime(sqd, &ist);
361 
362 		if (sqt_spin || !time_after(jiffies, timeout)) {
363 			if (sqt_spin)
364 				timeout = jiffies + sqd->sq_thread_idle;
365 			if (unlikely(need_resched())) {
366 				mutex_unlock(&sqd->lock);
367 				cond_resched();
368 				mutex_lock(&sqd->lock);
369 				sqd->sq_cpu = raw_smp_processor_id();
370 			}
371 			continue;
372 		}
373 
374 		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
375 		if (!io_sqd_events_pending(sqd) && !io_sq_tw_pending(retry_list)) {
376 			bool needs_sched = true;
377 
378 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
379 				atomic_or(IORING_SQ_NEED_WAKEUP,
380 						&ctx->rings->sq_flags);
381 				if ((ctx->flags & IORING_SETUP_IOPOLL) &&
382 				    !wq_list_empty(&ctx->iopoll_list)) {
383 					needs_sched = false;
384 					break;
385 				}
386 
387 				/*
388 				 * Ensure the store of the wakeup flag is not
389 				 * reordered with the load of the SQ tail
390 				 */
391 				smp_mb__after_atomic();
392 
393 				if (io_sqring_entries(ctx)) {
394 					needs_sched = false;
395 					break;
396 				}
397 			}
398 
399 			if (needs_sched) {
400 				mutex_unlock(&sqd->lock);
401 				schedule();
402 				mutex_lock(&sqd->lock);
403 				sqd->sq_cpu = raw_smp_processor_id();
404 			}
405 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
406 				atomic_andnot(IORING_SQ_NEED_WAKEUP,
407 						&ctx->rings->sq_flags);
408 		}
409 
410 		finish_wait(&sqd->wait, &wait);
411 		timeout = jiffies + sqd->sq_thread_idle;
412 	}
413 
414 	if (retry_list)
415 		io_sq_tw(&retry_list, UINT_MAX);
416 
417 	io_uring_cancel_generic(true, sqd);
418 	rcu_assign_pointer(sqd->thread, NULL);
419 	put_task_struct(current);
420 	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
421 		atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
422 	io_run_task_work();
423 	mutex_unlock(&sqd->lock);
424 err_out:
425 	complete(&sqd->exited);
426 	do_exit(0);
427 }
428 
io_sqpoll_wait_sq(struct io_ring_ctx * ctx)429 void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
430 {
431 	DEFINE_WAIT(wait);
432 
433 	do {
434 		if (!io_sqring_full(ctx))
435 			break;
436 		prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
437 
438 		if (!io_sqring_full(ctx))
439 			break;
440 		schedule();
441 	} while (!signal_pending(current));
442 
443 	finish_wait(&ctx->sqo_sq_wait, &wait);
444 }
445 
io_sq_offload_create(struct io_ring_ctx * ctx,struct io_uring_params * p)446 __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
447 				struct io_uring_params *p)
448 {
449 	int ret;
450 
451 	/* Retain compatibility with failing for an invalid attach attempt */
452 	if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
453 				IORING_SETUP_ATTACH_WQ) {
454 		CLASS(fd, f)(p->wq_fd);
455 		if (fd_empty(f))
456 			return -ENXIO;
457 		if (!io_is_uring_fops(fd_file(f)))
458 			return -EINVAL;
459 	}
460 	if (ctx->flags & IORING_SETUP_SQPOLL) {
461 		struct task_struct *tsk;
462 		struct io_sq_data *sqd;
463 		bool attached;
464 
465 		ret = security_uring_sqpoll();
466 		if (ret)
467 			return ret;
468 
469 		sqd = io_get_sq_data(p, &attached);
470 		if (IS_ERR(sqd)) {
471 			ret = PTR_ERR(sqd);
472 			goto err;
473 		}
474 
475 		ctx->sq_creds = get_current_cred();
476 		ctx->sq_data = sqd;
477 		ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
478 		if (!ctx->sq_thread_idle)
479 			ctx->sq_thread_idle = HZ;
480 
481 		io_sq_thread_park(sqd);
482 		list_add(&ctx->sqd_list, &sqd->ctx_list);
483 		io_sqd_update_thread_idle(sqd);
484 		/* don't attach to a dying SQPOLL thread, would be racy */
485 		ret = (attached && !sqd->thread) ? -ENXIO : 0;
486 		io_sq_thread_unpark(sqd);
487 
488 		if (ret < 0)
489 			goto err;
490 		if (attached)
491 			return 0;
492 
493 		if (p->flags & IORING_SETUP_SQ_AFF) {
494 			cpumask_var_t allowed_mask;
495 			int cpu = p->sq_thread_cpu;
496 
497 			ret = -EINVAL;
498 			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
499 				goto err_sqpoll;
500 			ret = -ENOMEM;
501 			if (!alloc_cpumask_var(&allowed_mask, GFP_KERNEL))
502 				goto err_sqpoll;
503 			ret = -EINVAL;
504 			cpuset_cpus_allowed(current, allowed_mask);
505 			if (!cpumask_test_cpu(cpu, allowed_mask)) {
506 				free_cpumask_var(allowed_mask);
507 				goto err_sqpoll;
508 			}
509 			free_cpumask_var(allowed_mask);
510 			sqd->sq_cpu = cpu;
511 		} else {
512 			sqd->sq_cpu = -1;
513 		}
514 
515 		sqd->task_pid = current->pid;
516 		sqd->task_tgid = current->tgid;
517 		tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
518 		if (IS_ERR(tsk)) {
519 			ret = PTR_ERR(tsk);
520 			goto err_sqpoll;
521 		}
522 
523 		mutex_lock(&sqd->lock);
524 		rcu_assign_pointer(sqd->thread, tsk);
525 		mutex_unlock(&sqd->lock);
526 
527 		get_task_struct(tsk);
528 		ret = io_uring_alloc_task_context(tsk, ctx);
529 		wake_up_new_task(tsk);
530 		if (ret)
531 			goto err;
532 	} else if (p->flags & IORING_SETUP_SQ_AFF) {
533 		/* Can't have SQ_AFF without SQPOLL */
534 		ret = -EINVAL;
535 		goto err;
536 	}
537 	return 0;
538 err_sqpoll:
539 	complete(&ctx->sq_data->exited);
540 err:
541 	io_sq_thread_finish(ctx);
542 	return ret;
543 }
544 
io_sqpoll_wq_cpu_affinity(struct io_ring_ctx * ctx,cpumask_var_t mask)545 __cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
546 				     cpumask_var_t mask)
547 {
548 	struct io_sq_data *sqd = ctx->sq_data;
549 	int ret = -EINVAL;
550 
551 	if (sqd) {
552 		struct task_struct *tsk;
553 
554 		io_sq_thread_park(sqd);
555 		/* Don't set affinity for a dying thread */
556 		tsk = sqpoll_task_locked(sqd);
557 		if (tsk)
558 			ret = io_wq_cpu_affinity(tsk->io_uring, mask);
559 		io_sq_thread_unpark(sqd);
560 	}
561 
562 	return ret;
563 }
564