xref: /openbmc/linux/io_uring/sqpoll.c (revision 93696d8f)
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 "sqpoll.h"
19 
20 #define IORING_SQPOLL_CAP_ENTRIES_VALUE 8
21 
22 enum {
23 	IO_SQ_THREAD_SHOULD_STOP = 0,
24 	IO_SQ_THREAD_SHOULD_PARK,
25 };
26 
27 void io_sq_thread_unpark(struct io_sq_data *sqd)
28 	__releases(&sqd->lock)
29 {
30 	WARN_ON_ONCE(sqd->thread == current);
31 
32 	/*
33 	 * Do the dance but not conditional clear_bit() because it'd race with
34 	 * other threads incrementing park_pending and setting the bit.
35 	 */
36 	clear_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
37 	if (atomic_dec_return(&sqd->park_pending))
38 		set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
39 	mutex_unlock(&sqd->lock);
40 }
41 
42 void io_sq_thread_park(struct io_sq_data *sqd)
43 	__acquires(&sqd->lock)
44 {
45 	WARN_ON_ONCE(sqd->thread == current);
46 
47 	atomic_inc(&sqd->park_pending);
48 	set_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state);
49 	mutex_lock(&sqd->lock);
50 	if (sqd->thread)
51 		wake_up_process(sqd->thread);
52 }
53 
54 void io_sq_thread_stop(struct io_sq_data *sqd)
55 {
56 	WARN_ON_ONCE(sqd->thread == current);
57 	WARN_ON_ONCE(test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state));
58 
59 	set_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
60 	mutex_lock(&sqd->lock);
61 	if (sqd->thread)
62 		wake_up_process(sqd->thread);
63 	mutex_unlock(&sqd->lock);
64 	wait_for_completion(&sqd->exited);
65 }
66 
67 void io_put_sq_data(struct io_sq_data *sqd)
68 {
69 	if (refcount_dec_and_test(&sqd->refs)) {
70 		WARN_ON_ONCE(atomic_read(&sqd->park_pending));
71 
72 		io_sq_thread_stop(sqd);
73 		kfree(sqd);
74 	}
75 }
76 
77 static __cold void io_sqd_update_thread_idle(struct io_sq_data *sqd)
78 {
79 	struct io_ring_ctx *ctx;
80 	unsigned sq_thread_idle = 0;
81 
82 	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
83 		sq_thread_idle = max(sq_thread_idle, ctx->sq_thread_idle);
84 	sqd->sq_thread_idle = sq_thread_idle;
85 }
86 
87 void io_sq_thread_finish(struct io_ring_ctx *ctx)
88 {
89 	struct io_sq_data *sqd = ctx->sq_data;
90 
91 	if (sqd) {
92 		io_sq_thread_park(sqd);
93 		list_del_init(&ctx->sqd_list);
94 		io_sqd_update_thread_idle(sqd);
95 		io_sq_thread_unpark(sqd);
96 
97 		io_put_sq_data(sqd);
98 		ctx->sq_data = NULL;
99 	}
100 }
101 
102 static struct io_sq_data *io_attach_sq_data(struct io_uring_params *p)
103 {
104 	struct io_ring_ctx *ctx_attach;
105 	struct io_sq_data *sqd;
106 	struct fd f;
107 
108 	f = fdget(p->wq_fd);
109 	if (!f.file)
110 		return ERR_PTR(-ENXIO);
111 	if (!io_is_uring_fops(f.file)) {
112 		fdput(f);
113 		return ERR_PTR(-EINVAL);
114 	}
115 
116 	ctx_attach = f.file->private_data;
117 	sqd = ctx_attach->sq_data;
118 	if (!sqd) {
119 		fdput(f);
120 		return ERR_PTR(-EINVAL);
121 	}
122 	if (sqd->task_tgid != current->tgid) {
123 		fdput(f);
124 		return ERR_PTR(-EPERM);
125 	}
126 
127 	refcount_inc(&sqd->refs);
128 	fdput(f);
129 	return sqd;
130 }
131 
132 static struct io_sq_data *io_get_sq_data(struct io_uring_params *p,
133 					 bool *attached)
134 {
135 	struct io_sq_data *sqd;
136 
137 	*attached = false;
138 	if (p->flags & IORING_SETUP_ATTACH_WQ) {
139 		sqd = io_attach_sq_data(p);
140 		if (!IS_ERR(sqd)) {
141 			*attached = true;
142 			return sqd;
143 		}
144 		/* fall through for EPERM case, setup new sqd/task */
145 		if (PTR_ERR(sqd) != -EPERM)
146 			return sqd;
147 	}
148 
149 	sqd = kzalloc(sizeof(*sqd), GFP_KERNEL);
150 	if (!sqd)
151 		return ERR_PTR(-ENOMEM);
152 
153 	atomic_set(&sqd->park_pending, 0);
154 	refcount_set(&sqd->refs, 1);
155 	INIT_LIST_HEAD(&sqd->ctx_list);
156 	mutex_init(&sqd->lock);
157 	init_waitqueue_head(&sqd->wait);
158 	init_completion(&sqd->exited);
159 	return sqd;
160 }
161 
162 static inline bool io_sqd_events_pending(struct io_sq_data *sqd)
163 {
164 	return READ_ONCE(sqd->state);
165 }
166 
167 static int __io_sq_thread(struct io_ring_ctx *ctx, bool cap_entries)
168 {
169 	unsigned int to_submit;
170 	int ret = 0;
171 
172 	to_submit = io_sqring_entries(ctx);
173 	/* if we're handling multiple rings, cap submit size for fairness */
174 	if (cap_entries && to_submit > IORING_SQPOLL_CAP_ENTRIES_VALUE)
175 		to_submit = IORING_SQPOLL_CAP_ENTRIES_VALUE;
176 
177 	if (!wq_list_empty(&ctx->iopoll_list) || to_submit) {
178 		const struct cred *creds = NULL;
179 
180 		if (ctx->sq_creds != current_cred())
181 			creds = override_creds(ctx->sq_creds);
182 
183 		mutex_lock(&ctx->uring_lock);
184 		if (!wq_list_empty(&ctx->iopoll_list))
185 			io_do_iopoll(ctx, true);
186 
187 		/*
188 		 * Don't submit if refs are dying, good for io_uring_register(),
189 		 * but also it is relied upon by io_ring_exit_work()
190 		 */
191 		if (to_submit && likely(!percpu_ref_is_dying(&ctx->refs)) &&
192 		    !(ctx->flags & IORING_SETUP_R_DISABLED))
193 			ret = io_submit_sqes(ctx, to_submit);
194 		mutex_unlock(&ctx->uring_lock);
195 
196 		if (to_submit && wq_has_sleeper(&ctx->sqo_sq_wait))
197 			wake_up(&ctx->sqo_sq_wait);
198 		if (creds)
199 			revert_creds(creds);
200 	}
201 
202 	return ret;
203 }
204 
205 static bool io_sqd_handle_event(struct io_sq_data *sqd)
206 {
207 	bool did_sig = false;
208 	struct ksignal ksig;
209 
210 	if (test_bit(IO_SQ_THREAD_SHOULD_PARK, &sqd->state) ||
211 	    signal_pending(current)) {
212 		mutex_unlock(&sqd->lock);
213 		if (signal_pending(current))
214 			did_sig = get_signal(&ksig);
215 		cond_resched();
216 		mutex_lock(&sqd->lock);
217 		sqd->sq_cpu = raw_smp_processor_id();
218 	}
219 	return did_sig || test_bit(IO_SQ_THREAD_SHOULD_STOP, &sqd->state);
220 }
221 
222 static int io_sq_thread(void *data)
223 {
224 	struct io_sq_data *sqd = data;
225 	struct io_ring_ctx *ctx;
226 	unsigned long timeout = 0;
227 	char buf[TASK_COMM_LEN];
228 	DEFINE_WAIT(wait);
229 
230 	snprintf(buf, sizeof(buf), "iou-sqp-%d", sqd->task_pid);
231 	set_task_comm(current, buf);
232 
233 	/* reset to our pid after we've set task_comm, for fdinfo */
234 	sqd->task_pid = current->pid;
235 
236 	if (sqd->sq_cpu != -1) {
237 		set_cpus_allowed_ptr(current, cpumask_of(sqd->sq_cpu));
238 	} else {
239 		set_cpus_allowed_ptr(current, cpu_online_mask);
240 		sqd->sq_cpu = raw_smp_processor_id();
241 	}
242 
243 	/*
244 	 * Force audit context to get setup, in case we do prep side async
245 	 * operations that would trigger an audit call before any issue side
246 	 * audit has been done.
247 	 */
248 	audit_uring_entry(IORING_OP_NOP);
249 	audit_uring_exit(true, 0);
250 
251 	mutex_lock(&sqd->lock);
252 	while (1) {
253 		bool cap_entries, sqt_spin = false;
254 
255 		if (io_sqd_events_pending(sqd) || signal_pending(current)) {
256 			if (io_sqd_handle_event(sqd))
257 				break;
258 			timeout = jiffies + sqd->sq_thread_idle;
259 		}
260 
261 		cap_entries = !list_is_singular(&sqd->ctx_list);
262 		list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
263 			int ret = __io_sq_thread(ctx, cap_entries);
264 
265 			if (!sqt_spin && (ret > 0 || !wq_list_empty(&ctx->iopoll_list)))
266 				sqt_spin = true;
267 		}
268 		if (io_run_task_work())
269 			sqt_spin = true;
270 
271 		if (sqt_spin || !time_after(jiffies, timeout)) {
272 			if (sqt_spin)
273 				timeout = jiffies + sqd->sq_thread_idle;
274 			if (unlikely(need_resched())) {
275 				mutex_unlock(&sqd->lock);
276 				cond_resched();
277 				mutex_lock(&sqd->lock);
278 				sqd->sq_cpu = raw_smp_processor_id();
279 			}
280 			continue;
281 		}
282 
283 		prepare_to_wait(&sqd->wait, &wait, TASK_INTERRUPTIBLE);
284 		if (!io_sqd_events_pending(sqd) && !task_work_pending(current)) {
285 			bool needs_sched = true;
286 
287 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list) {
288 				atomic_or(IORING_SQ_NEED_WAKEUP,
289 						&ctx->rings->sq_flags);
290 				if ((ctx->flags & IORING_SETUP_IOPOLL) &&
291 				    !wq_list_empty(&ctx->iopoll_list)) {
292 					needs_sched = false;
293 					break;
294 				}
295 
296 				/*
297 				 * Ensure the store of the wakeup flag is not
298 				 * reordered with the load of the SQ tail
299 				 */
300 				smp_mb__after_atomic();
301 
302 				if (io_sqring_entries(ctx)) {
303 					needs_sched = false;
304 					break;
305 				}
306 			}
307 
308 			if (needs_sched) {
309 				mutex_unlock(&sqd->lock);
310 				schedule();
311 				mutex_lock(&sqd->lock);
312 				sqd->sq_cpu = raw_smp_processor_id();
313 			}
314 			list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
315 				atomic_andnot(IORING_SQ_NEED_WAKEUP,
316 						&ctx->rings->sq_flags);
317 		}
318 
319 		finish_wait(&sqd->wait, &wait);
320 		timeout = jiffies + sqd->sq_thread_idle;
321 	}
322 
323 	io_uring_cancel_generic(true, sqd);
324 	sqd->thread = NULL;
325 	list_for_each_entry(ctx, &sqd->ctx_list, sqd_list)
326 		atomic_or(IORING_SQ_NEED_WAKEUP, &ctx->rings->sq_flags);
327 	io_run_task_work();
328 	mutex_unlock(&sqd->lock);
329 
330 	complete(&sqd->exited);
331 	do_exit(0);
332 }
333 
334 void io_sqpoll_wait_sq(struct io_ring_ctx *ctx)
335 {
336 	DEFINE_WAIT(wait);
337 
338 	do {
339 		if (!io_sqring_full(ctx))
340 			break;
341 		prepare_to_wait(&ctx->sqo_sq_wait, &wait, TASK_INTERRUPTIBLE);
342 
343 		if (!io_sqring_full(ctx))
344 			break;
345 		schedule();
346 	} while (!signal_pending(current));
347 
348 	finish_wait(&ctx->sqo_sq_wait, &wait);
349 }
350 
351 __cold int io_sq_offload_create(struct io_ring_ctx *ctx,
352 				struct io_uring_params *p)
353 {
354 	int ret;
355 
356 	/* Retain compatibility with failing for an invalid attach attempt */
357 	if ((ctx->flags & (IORING_SETUP_ATTACH_WQ | IORING_SETUP_SQPOLL)) ==
358 				IORING_SETUP_ATTACH_WQ) {
359 		struct fd f;
360 
361 		f = fdget(p->wq_fd);
362 		if (!f.file)
363 			return -ENXIO;
364 		if (!io_is_uring_fops(f.file)) {
365 			fdput(f);
366 			return -EINVAL;
367 		}
368 		fdput(f);
369 	}
370 	if (ctx->flags & IORING_SETUP_SQPOLL) {
371 		struct task_struct *tsk;
372 		struct io_sq_data *sqd;
373 		bool attached;
374 
375 		ret = security_uring_sqpoll();
376 		if (ret)
377 			return ret;
378 
379 		sqd = io_get_sq_data(p, &attached);
380 		if (IS_ERR(sqd)) {
381 			ret = PTR_ERR(sqd);
382 			goto err;
383 		}
384 
385 		ctx->sq_creds = get_current_cred();
386 		ctx->sq_data = sqd;
387 		ctx->sq_thread_idle = msecs_to_jiffies(p->sq_thread_idle);
388 		if (!ctx->sq_thread_idle)
389 			ctx->sq_thread_idle = HZ;
390 
391 		io_sq_thread_park(sqd);
392 		list_add(&ctx->sqd_list, &sqd->ctx_list);
393 		io_sqd_update_thread_idle(sqd);
394 		/* don't attach to a dying SQPOLL thread, would be racy */
395 		ret = (attached && !sqd->thread) ? -ENXIO : 0;
396 		io_sq_thread_unpark(sqd);
397 
398 		if (ret < 0)
399 			goto err;
400 		if (attached)
401 			return 0;
402 
403 		if (p->flags & IORING_SETUP_SQ_AFF) {
404 			int cpu = p->sq_thread_cpu;
405 
406 			ret = -EINVAL;
407 			if (cpu >= nr_cpu_ids || !cpu_online(cpu))
408 				goto err_sqpoll;
409 			sqd->sq_cpu = cpu;
410 		} else {
411 			sqd->sq_cpu = -1;
412 		}
413 
414 		sqd->task_pid = current->pid;
415 		sqd->task_tgid = current->tgid;
416 		tsk = create_io_thread(io_sq_thread, sqd, NUMA_NO_NODE);
417 		if (IS_ERR(tsk)) {
418 			ret = PTR_ERR(tsk);
419 			goto err_sqpoll;
420 		}
421 
422 		sqd->thread = tsk;
423 		ret = io_uring_alloc_task_context(tsk, ctx);
424 		wake_up_new_task(tsk);
425 		if (ret)
426 			goto err;
427 	} else if (p->flags & IORING_SETUP_SQ_AFF) {
428 		/* Can't have SQ_AFF without SQPOLL */
429 		ret = -EINVAL;
430 		goto err;
431 	}
432 
433 	return 0;
434 err_sqpoll:
435 	complete(&ctx->sq_data->exited);
436 err:
437 	io_sq_thread_finish(ctx);
438 	return ret;
439 }
440 
441 __cold int io_sqpoll_wq_cpu_affinity(struct io_ring_ctx *ctx,
442 				     cpumask_var_t mask)
443 {
444 	struct io_sq_data *sqd = ctx->sq_data;
445 	int ret = -EINVAL;
446 
447 	if (sqd) {
448 		io_sq_thread_park(sqd);
449 		/* Don't set affinity for a dying thread */
450 		if (sqd->thread)
451 			ret = io_wq_cpu_affinity(sqd->thread->io_uring, mask);
452 		io_sq_thread_unpark(sqd);
453 	}
454 
455 	return ret;
456 }
457