sqpoll.c

  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
 28enum {
 29	IO_SQ_THREAD_SHOULD_STOP = 0,
 30	IO_SQ_THREAD_SHOULD_PARK,
 31};
 32
 33void 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
 49void 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
 65void 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
 82void 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
 92static __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
102void 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
117static 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
139static 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
169static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
170{
171	return READ_ONCE(sqd->state);
172}
173
174struct io_sq_time {
175	bool started;
176	u64 usec;
177};
178
179u64 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
188static 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
196static 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
204static 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
245static 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 */
268static 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);
280out:
281	if (task_work_pending(current))
282		task_work_run();
283	return count;
284}
285
286static 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
293static 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);
424err_out:
425	complete(&sqd->exited);
426	do_exit(0);
427}
428
429void 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
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;
538err_sqpoll:
539	complete(&ctx->sq_data->exited);
540err:
541	io_sq_thread_finish(ctx);
542	return ret;
543}
544
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}