xref: /netbsd-src/sys/kern/sys_aio.c (revision 5dd36a3bc8bf2a9dec29ceb6349550414570c447)
1 /*	$NetBSD: sys_aio.c,v 1.46 2020/02/01 02:23:04 riastradh Exp $	*/
2 
3 /*
4  * Copyright (c) 2007 Mindaugas Rasiukevicius <rmind at NetBSD org>
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * Implementation of POSIX asynchronous I/O.
31  * Defined in the Base Definitions volume of IEEE Std 1003.1-2001.
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: sys_aio.c,v 1.46 2020/02/01 02:23:04 riastradh Exp $");
36 
37 #ifdef _KERNEL_OPT
38 #include "opt_ddb.h"
39 #endif
40 
41 #include <sys/param.h>
42 #include <sys/condvar.h>
43 #include <sys/file.h>
44 #include <sys/filedesc.h>
45 #include <sys/kernel.h>
46 #include <sys/kmem.h>
47 #include <sys/lwp.h>
48 #include <sys/mutex.h>
49 #include <sys/pool.h>
50 #include <sys/proc.h>
51 #include <sys/queue.h>
52 #include <sys/signal.h>
53 #include <sys/signalvar.h>
54 #include <sys/syscall.h>
55 #include <sys/syscallargs.h>
56 #include <sys/syscallvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/types.h>
60 #include <sys/vnode.h>
61 #include <sys/atomic.h>
62 #include <sys/module.h>
63 #include <sys/buf.h>
64 
65 #include <uvm/uvm_extern.h>
66 
67 MODULE(MODULE_CLASS_MISC, aio, NULL);
68 
69 /*
70  * System-wide limits and counter of AIO operations.
71  */
72 u_int			aio_listio_max = AIO_LISTIO_MAX;
73 static u_int		aio_max = AIO_MAX;
74 static u_int		aio_jobs_count;
75 
76 static struct sysctllog	*aio_sysctl;
77 static struct pool	aio_job_pool;
78 static struct pool	aio_lio_pool;
79 static void *		aio_ehook;
80 
81 static void		aio_worker(void *);
82 static void		aio_process(struct aio_job *);
83 static void		aio_sendsig(struct proc *, struct sigevent *);
84 static int		aio_enqueue_job(int, void *, struct lio_req *);
85 static void		aio_exit(proc_t *, void *);
86 
87 static int		sysctl_aio_listio_max(SYSCTLFN_PROTO);
88 static int		sysctl_aio_max(SYSCTLFN_PROTO);
89 static int		sysctl_aio_init(void);
90 
91 static const struct syscall_package aio_syscalls[] = {
92 	{ SYS_aio_cancel, 0, (sy_call_t *)sys_aio_cancel },
93 	{ SYS_aio_error, 0, (sy_call_t *)sys_aio_error },
94 	{ SYS_aio_fsync, 0, (sy_call_t *)sys_aio_fsync },
95 	{ SYS_aio_read, 0, (sy_call_t *)sys_aio_read },
96 	{ SYS_aio_return, 0, (sy_call_t *)sys_aio_return },
97 	{ SYS___aio_suspend50, 0, (sy_call_t *)sys___aio_suspend50 },
98 	{ SYS_aio_write, 0, (sy_call_t *)sys_aio_write },
99 	{ SYS_lio_listio, 0, (sy_call_t *)sys_lio_listio },
100 	{ 0, 0, NULL },
101 };
102 
103 /*
104  * Tear down all AIO state.
105  */
106 static int
107 aio_fini(bool interface)
108 {
109 	int error;
110 	proc_t *p;
111 
112 	if (interface) {
113 		/* Stop syscall activity. */
114 		error = syscall_disestablish(NULL, aio_syscalls);
115 		if (error != 0)
116 			return error;
117 		/* Abort if any processes are using AIO. */
118 		mutex_enter(proc_lock);
119 		PROCLIST_FOREACH(p, &allproc) {
120 			if (p->p_aio != NULL)
121 				break;
122 		}
123 		mutex_exit(proc_lock);
124 		if (p != NULL) {
125 			error = syscall_establish(NULL, aio_syscalls);
126 			KASSERT(error == 0);
127 			return EBUSY;
128 		}
129 	}
130 	if (aio_sysctl != NULL)
131 		sysctl_teardown(&aio_sysctl);
132 
133 	KASSERT(aio_jobs_count == 0);
134 	exithook_disestablish(aio_ehook);
135 	pool_destroy(&aio_job_pool);
136 	pool_destroy(&aio_lio_pool);
137 	return 0;
138 }
139 
140 /*
141  * Initialize global AIO state.
142  */
143 static int
144 aio_init(void)
145 {
146 	int error;
147 
148 	pool_init(&aio_job_pool, sizeof(struct aio_job), 0, 0, 0,
149 	    "aio_jobs_pool", &pool_allocator_nointr, IPL_NONE);
150 	pool_init(&aio_lio_pool, sizeof(struct lio_req), 0, 0, 0,
151 	    "aio_lio_pool", &pool_allocator_nointr, IPL_NONE);
152 	aio_ehook = exithook_establish(aio_exit, NULL);
153 
154 	error = sysctl_aio_init();
155 	if (error != 0) {
156 		(void)aio_fini(false);
157 		return error;
158 	}
159 	error = syscall_establish(NULL, aio_syscalls);
160 	if (error != 0)
161 		(void)aio_fini(false);
162 	return error;
163 }
164 
165 /*
166  * Module interface.
167  */
168 static int
169 aio_modcmd(modcmd_t cmd, void *arg)
170 {
171 
172 	switch (cmd) {
173 	case MODULE_CMD_INIT:
174 		return aio_init();
175 	case MODULE_CMD_FINI:
176 		return aio_fini(true);
177 	default:
178 		return ENOTTY;
179 	}
180 }
181 
182 /*
183  * Initialize Asynchronous I/O data structures for the process.
184  */
185 static int
186 aio_procinit(struct proc *p)
187 {
188 	struct aioproc *aio;
189 	struct lwp *l;
190 	int error;
191 	vaddr_t uaddr;
192 
193 	/* Allocate and initialize AIO structure */
194 	aio = kmem_zalloc(sizeof(struct aioproc), KM_SLEEP);
195 
196 	/* Initialize queue and their synchronization structures */
197 	mutex_init(&aio->aio_mtx, MUTEX_DEFAULT, IPL_NONE);
198 	cv_init(&aio->aio_worker_cv, "aiowork");
199 	cv_init(&aio->done_cv, "aiodone");
200 	TAILQ_INIT(&aio->jobs_queue);
201 
202 	/*
203 	 * Create an AIO worker thread.
204 	 * XXX: Currently, AIO thread is not protected against user's actions.
205 	 */
206 	uaddr = uvm_uarea_alloc();
207 	if (uaddr == 0) {
208 		aio_exit(p, aio);
209 		return EAGAIN;
210 	}
211 	error = lwp_create(curlwp, p, uaddr, 0, NULL, 0, aio_worker,
212 	    NULL, &l, curlwp->l_class, &curlwp->l_sigmask, &curlwp->l_sigstk);
213 	if (error != 0) {
214 		uvm_uarea_free(uaddr);
215 		aio_exit(p, aio);
216 		return error;
217 	}
218 
219 	/* Recheck if we are really first */
220 	mutex_enter(p->p_lock);
221 	if (p->p_aio) {
222 		mutex_exit(p->p_lock);
223 		aio_exit(p, aio);
224 		lwp_exit(l);
225 		return 0;
226 	}
227 	p->p_aio = aio;
228 
229 	/* Complete the initialization of thread, and run it */
230 	aio->aio_worker = l;
231 	lwp_lock(l);
232 	lwp_changepri(l, MAXPRI_USER);
233 	setrunnable(l);
234 	/* LWP now unlocked */
235 	mutex_exit(p->p_lock);
236 
237 	return 0;
238 }
239 
240 /*
241  * Exit of Asynchronous I/O subsystem of process.
242  */
243 static void
244 aio_exit(struct proc *p, void *cookie)
245 {
246 	struct aio_job *a_job;
247 	struct aioproc *aio;
248 
249 	if (cookie != NULL)
250 		aio = cookie;
251 	else if ((aio = p->p_aio) == NULL)
252 		return;
253 
254 	/* Free AIO queue */
255 	while (!TAILQ_EMPTY(&aio->jobs_queue)) {
256 		a_job = TAILQ_FIRST(&aio->jobs_queue);
257 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
258 		pool_put(&aio_job_pool, a_job);
259 		atomic_dec_uint(&aio_jobs_count);
260 	}
261 
262 	/* Destroy and free the entire AIO data structure */
263 	cv_destroy(&aio->aio_worker_cv);
264 	cv_destroy(&aio->done_cv);
265 	mutex_destroy(&aio->aio_mtx);
266 	kmem_free(aio, sizeof(struct aioproc));
267 }
268 
269 /*
270  * AIO worker thread and processor.
271  */
272 static void
273 aio_worker(void *arg)
274 {
275 	struct proc *p = curlwp->l_proc;
276 	struct aioproc *aio = p->p_aio;
277 	struct aio_job *a_job;
278 	struct lio_req *lio;
279 	sigset_t oss, nss;
280 	int error __diagused, refcnt;
281 
282 	/*
283 	 * Make an empty signal mask, so it
284 	 * handles only SIGKILL and SIGSTOP.
285 	 */
286 	sigfillset(&nss);
287 	mutex_enter(p->p_lock);
288 	error = sigprocmask1(curlwp, SIG_SETMASK, &nss, &oss);
289 	mutex_exit(p->p_lock);
290 	KASSERT(error == 0);
291 
292 	for (;;) {
293 		/*
294 		 * Loop for each job in the queue.  If there
295 		 * are no jobs then sleep.
296 		 */
297 		mutex_enter(&aio->aio_mtx);
298 		while ((a_job = TAILQ_FIRST(&aio->jobs_queue)) == NULL) {
299 			if (cv_wait_sig(&aio->aio_worker_cv, &aio->aio_mtx)) {
300 				/*
301 				 * Thread was interrupted - check for
302 				 * pending exit or suspend.
303 				 */
304 				mutex_exit(&aio->aio_mtx);
305 				lwp_userret(curlwp);
306 				mutex_enter(&aio->aio_mtx);
307 			}
308 		}
309 
310 		/* Take the job from the queue */
311 		aio->curjob = a_job;
312 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
313 
314 		atomic_dec_uint(&aio_jobs_count);
315 		aio->jobs_count--;
316 
317 		mutex_exit(&aio->aio_mtx);
318 
319 		/* Process an AIO operation */
320 		aio_process(a_job);
321 
322 		/* Copy data structure back to the user-space */
323 		(void)copyout(&a_job->aiocbp, a_job->aiocb_uptr,
324 		    sizeof(struct aiocb));
325 
326 		mutex_enter(&aio->aio_mtx);
327 		KASSERT(aio->curjob == a_job);
328 		aio->curjob = NULL;
329 
330 		/* Decrease a reference counter, if there is a LIO structure */
331 		lio = a_job->lio;
332 		refcnt = (lio != NULL ? --lio->refcnt : -1);
333 
334 		/* Notify all suspenders */
335 		cv_broadcast(&aio->done_cv);
336 		mutex_exit(&aio->aio_mtx);
337 
338 		/* Send a signal, if any */
339 		aio_sendsig(p, &a_job->aiocbp.aio_sigevent);
340 
341 		/* Destroy the LIO structure */
342 		if (refcnt == 0) {
343 			aio_sendsig(p, &lio->sig);
344 			pool_put(&aio_lio_pool, lio);
345 		}
346 
347 		/* Destroy the job */
348 		pool_put(&aio_job_pool, a_job);
349 	}
350 
351 	/* NOTREACHED */
352 }
353 
354 static void
355 aio_process(struct aio_job *a_job)
356 {
357 	struct proc *p = curlwp->l_proc;
358 	struct aiocb *aiocbp = &a_job->aiocbp;
359 	struct file *fp;
360 	int fd = aiocbp->aio_fildes;
361 	int error = 0;
362 
363 	KASSERT(a_job->aio_op != 0);
364 
365 	if ((a_job->aio_op & (AIO_READ | AIO_WRITE)) != 0) {
366 		struct iovec aiov;
367 		struct uio auio;
368 
369 		if (aiocbp->aio_nbytes > SSIZE_MAX) {
370 			error = EINVAL;
371 			goto done;
372 		}
373 
374 		fp = fd_getfile(fd);
375 		if (fp == NULL) {
376 			error = EBADF;
377 			goto done;
378 		}
379 
380 		aiov.iov_base = (void *)(uintptr_t)aiocbp->aio_buf;
381 		aiov.iov_len = aiocbp->aio_nbytes;
382 		auio.uio_iov = &aiov;
383 		auio.uio_iovcnt = 1;
384 		auio.uio_resid = aiocbp->aio_nbytes;
385 		auio.uio_vmspace = p->p_vmspace;
386 
387 		if (a_job->aio_op & AIO_READ) {
388 			/*
389 			 * Perform a Read operation
390 			 */
391 			KASSERT((a_job->aio_op & AIO_WRITE) == 0);
392 
393 			if ((fp->f_flag & FREAD) == 0) {
394 				fd_putfile(fd);
395 				error = EBADF;
396 				goto done;
397 			}
398 			auio.uio_rw = UIO_READ;
399 			error = (*fp->f_ops->fo_read)(fp, &aiocbp->aio_offset,
400 			    &auio, fp->f_cred, FOF_UPDATE_OFFSET);
401 		} else {
402 			/*
403 			 * Perform a Write operation
404 			 */
405 			KASSERT(a_job->aio_op & AIO_WRITE);
406 
407 			if ((fp->f_flag & FWRITE) == 0) {
408 				fd_putfile(fd);
409 				error = EBADF;
410 				goto done;
411 			}
412 			auio.uio_rw = UIO_WRITE;
413 			error = (*fp->f_ops->fo_write)(fp, &aiocbp->aio_offset,
414 			    &auio, fp->f_cred, FOF_UPDATE_OFFSET);
415 		}
416 		fd_putfile(fd);
417 
418 		/* Store the result value */
419 		a_job->aiocbp.aio_nbytes -= auio.uio_resid;
420 		a_job->aiocbp._retval = (error == 0) ?
421 		    a_job->aiocbp.aio_nbytes : -1;
422 
423 	} else if ((a_job->aio_op & (AIO_SYNC | AIO_DSYNC)) != 0) {
424 		/*
425 		 * Perform a file Sync operation
426 		 */
427 		struct vnode *vp;
428 
429 		if ((error = fd_getvnode(fd, &fp)) != 0)
430 			goto done;
431 
432 		if ((fp->f_flag & FWRITE) == 0) {
433 			fd_putfile(fd);
434 			error = EBADF;
435 			goto done;
436 		}
437 
438 		vp = fp->f_vnode;
439 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
440 		if (a_job->aio_op & AIO_DSYNC) {
441 			error = VOP_FSYNC(vp, fp->f_cred,
442 			    FSYNC_WAIT | FSYNC_DATAONLY, 0, 0);
443 		} else if (a_job->aio_op & AIO_SYNC) {
444 			error = VOP_FSYNC(vp, fp->f_cred,
445 			    FSYNC_WAIT, 0, 0);
446 		}
447 		VOP_UNLOCK(vp);
448 		fd_putfile(fd);
449 
450 		/* Store the result value */
451 		a_job->aiocbp._retval = (error == 0) ? 0 : -1;
452 
453 	} else
454 		panic("aio_process: invalid operation code\n");
455 
456 done:
457 	/* Job is done, set the error, if any */
458 	a_job->aiocbp._errno = error;
459 	a_job->aiocbp._state = JOB_DONE;
460 }
461 
462 /*
463  * Send AIO signal.
464  */
465 static void
466 aio_sendsig(struct proc *p, struct sigevent *sig)
467 {
468 	ksiginfo_t ksi;
469 
470 	if (sig->sigev_signo == 0 || sig->sigev_notify == SIGEV_NONE)
471 		return;
472 
473 	KSI_INIT(&ksi);
474 	ksi.ksi_signo = sig->sigev_signo;
475 	ksi.ksi_code = SI_ASYNCIO;
476 	ksi.ksi_value = sig->sigev_value;
477 	mutex_enter(proc_lock);
478 	kpsignal(p, &ksi, NULL);
479 	mutex_exit(proc_lock);
480 }
481 
482 /*
483  * Enqueue the job.
484  */
485 static int
486 aio_enqueue_job(int op, void *aiocb_uptr, struct lio_req *lio)
487 {
488 	struct proc *p = curlwp->l_proc;
489 	struct aioproc *aio;
490 	struct aio_job *a_job;
491 	struct aiocb aiocbp;
492 	struct sigevent *sig;
493 	int error;
494 
495 	/* Non-accurate check for the limit */
496 	if (aio_jobs_count + 1 > aio_max)
497 		return EAGAIN;
498 
499 	/* Get the data structure from user-space */
500 	error = copyin(aiocb_uptr, &aiocbp, sizeof(struct aiocb));
501 	if (error)
502 		return error;
503 
504 	/* Check if signal is set, and validate it */
505 	sig = &aiocbp.aio_sigevent;
506 	if (sig->sigev_signo < 0 || sig->sigev_signo >= NSIG ||
507 	    sig->sigev_notify < SIGEV_NONE || sig->sigev_notify > SIGEV_SA)
508 		return EINVAL;
509 
510 	/* Buffer and byte count */
511 	if (((AIO_SYNC | AIO_DSYNC) & op) == 0)
512 		if (aiocbp.aio_buf == NULL || aiocbp.aio_nbytes > SSIZE_MAX)
513 			return EINVAL;
514 
515 	/* Check the opcode, if LIO_NOP - simply ignore */
516 	if (op == AIO_LIO) {
517 		KASSERT(lio != NULL);
518 		if (aiocbp.aio_lio_opcode == LIO_WRITE)
519 			op = AIO_WRITE;
520 		else if (aiocbp.aio_lio_opcode == LIO_READ)
521 			op = AIO_READ;
522 		else
523 			return (aiocbp.aio_lio_opcode == LIO_NOP) ? 0 : EINVAL;
524 	} else {
525 		KASSERT(lio == NULL);
526 	}
527 
528 	/*
529 	 * Look for already existing job.  If found - the job is in-progress.
530 	 * According to POSIX this is invalid, so return the error.
531 	 */
532 	aio = p->p_aio;
533 	if (aio) {
534 		mutex_enter(&aio->aio_mtx);
535 		TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
536 			if (a_job->aiocb_uptr != aiocb_uptr)
537 				continue;
538 			mutex_exit(&aio->aio_mtx);
539 			return EINVAL;
540 		}
541 		mutex_exit(&aio->aio_mtx);
542 	}
543 
544 	/*
545 	 * Check if AIO structure is initialized, if not - initialize it.
546 	 * In LIO case, we did that already.  We will recheck this with
547 	 * the lock in aio_procinit().
548 	 */
549 	if (lio == NULL && p->p_aio == NULL)
550 		if (aio_procinit(p))
551 			return EAGAIN;
552 	aio = p->p_aio;
553 
554 	/*
555 	 * Set the state with errno, and copy data
556 	 * structure back to the user-space.
557 	 */
558 	aiocbp._state = JOB_WIP;
559 	aiocbp._errno = EINPROGRESS;
560 	aiocbp._retval = -1;
561 	error = copyout(&aiocbp, aiocb_uptr, sizeof(struct aiocb));
562 	if (error)
563 		return error;
564 
565 	/* Allocate and initialize a new AIO job */
566 	a_job = pool_get(&aio_job_pool, PR_WAITOK | PR_ZERO);
567 
568 	/*
569 	 * Set the data.
570 	 * Store the user-space pointer for searching.  Since we
571 	 * are storing only per proc pointers - it is safe.
572 	 */
573 	memcpy(&a_job->aiocbp, &aiocbp, sizeof(struct aiocb));
574 	a_job->aiocb_uptr = aiocb_uptr;
575 	a_job->aio_op |= op;
576 	a_job->lio = lio;
577 
578 	/*
579 	 * Add the job to the queue, update the counters, and
580 	 * notify the AIO worker thread to handle the job.
581 	 */
582 	mutex_enter(&aio->aio_mtx);
583 
584 	/* Fail, if the limit was reached */
585 	if (atomic_inc_uint_nv(&aio_jobs_count) > aio_max ||
586 	    aio->jobs_count >= aio_listio_max) {
587 		atomic_dec_uint(&aio_jobs_count);
588 		mutex_exit(&aio->aio_mtx);
589 		pool_put(&aio_job_pool, a_job);
590 		return EAGAIN;
591 	}
592 
593 	TAILQ_INSERT_TAIL(&aio->jobs_queue, a_job, list);
594 	aio->jobs_count++;
595 	if (lio)
596 		lio->refcnt++;
597 	cv_signal(&aio->aio_worker_cv);
598 
599 	mutex_exit(&aio->aio_mtx);
600 
601 	/*
602 	 * One would handle the errors only with aio_error() function.
603 	 * This way is appropriate according to POSIX.
604 	 */
605 	return 0;
606 }
607 
608 /*
609  * Syscall functions.
610  */
611 
612 int
613 sys_aio_cancel(struct lwp *l, const struct sys_aio_cancel_args *uap,
614     register_t *retval)
615 {
616 	/* {
617 		syscallarg(int) fildes;
618 		syscallarg(struct aiocb *) aiocbp;
619 	} */
620 	struct proc *p = l->l_proc;
621 	struct aioproc *aio;
622 	struct aio_job *a_job;
623 	struct aiocb *aiocbp_ptr;
624 	struct lio_req *lio;
625 	struct filedesc	*fdp = p->p_fd;
626 	unsigned int cn, errcnt, fildes;
627 	fdtab_t *dt;
628 
629 	TAILQ_HEAD(, aio_job) tmp_jobs_list;
630 
631 	/* Check for invalid file descriptor */
632 	fildes = (unsigned int)SCARG(uap, fildes);
633 	dt = atomic_load_consume(&fdp->fd_dt);
634 	if (fildes >= dt->dt_nfiles)
635 		return EBADF;
636 	if (dt->dt_ff[fildes] == NULL || dt->dt_ff[fildes]->ff_file == NULL)
637 		return EBADF;
638 
639 	/* Check if AIO structure is initialized */
640 	if (p->p_aio == NULL) {
641 		*retval = AIO_NOTCANCELED;
642 		return 0;
643 	}
644 
645 	aio = p->p_aio;
646 	aiocbp_ptr = (struct aiocb *)SCARG(uap, aiocbp);
647 
648 	mutex_enter(&aio->aio_mtx);
649 
650 	/* Cancel the jobs, and remove them from the queue */
651 	cn = 0;
652 	TAILQ_INIT(&tmp_jobs_list);
653 	TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
654 		if (aiocbp_ptr) {
655 			if (aiocbp_ptr != a_job->aiocb_uptr)
656 				continue;
657 			if (fildes != a_job->aiocbp.aio_fildes) {
658 				mutex_exit(&aio->aio_mtx);
659 				return EBADF;
660 			}
661 		} else if (a_job->aiocbp.aio_fildes != fildes)
662 			continue;
663 
664 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
665 		TAILQ_INSERT_TAIL(&tmp_jobs_list, a_job, list);
666 
667 		/* Decrease the counters */
668 		atomic_dec_uint(&aio_jobs_count);
669 		aio->jobs_count--;
670 		lio = a_job->lio;
671 		if (lio != NULL && --lio->refcnt != 0)
672 			a_job->lio = NULL;
673 
674 		cn++;
675 		if (aiocbp_ptr)
676 			break;
677 	}
678 
679 	/* There are canceled jobs */
680 	if (cn)
681 		*retval = AIO_CANCELED;
682 
683 	/* We cannot cancel current job */
684 	a_job = aio->curjob;
685 	if (a_job && ((a_job->aiocbp.aio_fildes == fildes) ||
686 	    (a_job->aiocb_uptr == aiocbp_ptr)))
687 		*retval = AIO_NOTCANCELED;
688 
689 	mutex_exit(&aio->aio_mtx);
690 
691 	/* Free the jobs after the lock */
692 	errcnt = 0;
693 	while (!TAILQ_EMPTY(&tmp_jobs_list)) {
694 		a_job = TAILQ_FIRST(&tmp_jobs_list);
695 		TAILQ_REMOVE(&tmp_jobs_list, a_job, list);
696 		/* Set the errno and copy structures back to the user-space */
697 		a_job->aiocbp._errno = ECANCELED;
698 		a_job->aiocbp._state = JOB_DONE;
699 		if (copyout(&a_job->aiocbp, a_job->aiocb_uptr,
700 		    sizeof(struct aiocb)))
701 			errcnt++;
702 		/* Send a signal if any */
703 		aio_sendsig(p, &a_job->aiocbp.aio_sigevent);
704 		if (a_job->lio) {
705 			lio = a_job->lio;
706 			aio_sendsig(p, &lio->sig);
707 			pool_put(&aio_lio_pool, lio);
708 		}
709 		pool_put(&aio_job_pool, a_job);
710 	}
711 
712 	if (errcnt)
713 		return EFAULT;
714 
715 	/* Set a correct return value */
716 	if (*retval == 0)
717 		*retval = AIO_ALLDONE;
718 
719 	return 0;
720 }
721 
722 int
723 sys_aio_error(struct lwp *l, const struct sys_aio_error_args *uap,
724     register_t *retval)
725 {
726 	/* {
727 		syscallarg(const struct aiocb *) aiocbp;
728 	} */
729 	struct proc *p = l->l_proc;
730 	struct aioproc *aio = p->p_aio;
731 	struct aiocb aiocbp;
732 	int error;
733 
734 	if (aio == NULL)
735 		return EINVAL;
736 
737 	error = copyin(SCARG(uap, aiocbp), &aiocbp, sizeof(struct aiocb));
738 	if (error)
739 		return error;
740 
741 	if (aiocbp._state == JOB_NONE)
742 		return EINVAL;
743 
744 	*retval = aiocbp._errno;
745 
746 	return 0;
747 }
748 
749 int
750 sys_aio_fsync(struct lwp *l, const struct sys_aio_fsync_args *uap,
751     register_t *retval)
752 {
753 	/* {
754 		syscallarg(int) op;
755 		syscallarg(struct aiocb *) aiocbp;
756 	} */
757 	int op = SCARG(uap, op);
758 
759 	if ((op != O_DSYNC) && (op != O_SYNC))
760 		return EINVAL;
761 
762 	op = O_DSYNC ? AIO_DSYNC : AIO_SYNC;
763 
764 	return aio_enqueue_job(op, SCARG(uap, aiocbp), NULL);
765 }
766 
767 int
768 sys_aio_read(struct lwp *l, const struct sys_aio_read_args *uap,
769     register_t *retval)
770 {
771 	/* {
772 		syscallarg(struct aiocb *) aiocbp;
773 	} */
774 
775 	return aio_enqueue_job(AIO_READ, SCARG(uap, aiocbp), NULL);
776 }
777 
778 int
779 sys_aio_return(struct lwp *l, const struct sys_aio_return_args *uap,
780     register_t *retval)
781 {
782 	/* {
783 		syscallarg(struct aiocb *) aiocbp;
784 	} */
785 	struct proc *p = l->l_proc;
786 	struct aioproc *aio = p->p_aio;
787 	struct aiocb aiocbp;
788 	int error;
789 
790 	if (aio == NULL)
791 		return EINVAL;
792 
793 	error = copyin(SCARG(uap, aiocbp), &aiocbp, sizeof(struct aiocb));
794 	if (error)
795 		return error;
796 
797 	if (aiocbp._errno == EINPROGRESS || aiocbp._state != JOB_DONE)
798 		return EINVAL;
799 
800 	*retval = aiocbp._retval;
801 
802 	/* Reset the internal variables */
803 	aiocbp._errno = 0;
804 	aiocbp._retval = -1;
805 	aiocbp._state = JOB_NONE;
806 	error = copyout(&aiocbp, SCARG(uap, aiocbp), sizeof(struct aiocb));
807 
808 	return error;
809 }
810 
811 int
812 sys___aio_suspend50(struct lwp *l, const struct sys___aio_suspend50_args *uap,
813     register_t *retval)
814 {
815 	/* {
816 		syscallarg(const struct aiocb *const[]) list;
817 		syscallarg(int) nent;
818 		syscallarg(const struct timespec *) timeout;
819 	} */
820 	struct aiocb **list;
821 	struct timespec ts;
822 	int error, nent;
823 
824 	nent = SCARG(uap, nent);
825 	if (nent <= 0 || nent > aio_listio_max)
826 		return EAGAIN;
827 
828 	if (SCARG(uap, timeout)) {
829 		/* Convert timespec to ticks */
830 		error = copyin(SCARG(uap, timeout), &ts,
831 		    sizeof(struct timespec));
832 		if (error)
833 			return error;
834 	}
835 
836 	list = kmem_alloc(nent * sizeof(*list), KM_SLEEP);
837 	error = copyin(SCARG(uap, list), list, nent * sizeof(*list));
838 	if (error)
839 		goto out;
840 	error = aio_suspend1(l, list, nent, SCARG(uap, timeout) ? &ts : NULL);
841 out:
842 	kmem_free(list, nent * sizeof(*list));
843 	return error;
844 }
845 
846 int
847 aio_suspend1(struct lwp *l, struct aiocb **aiocbp_list, int nent,
848     struct timespec *ts)
849 {
850 	struct proc *p = l->l_proc;
851 	struct aioproc *aio;
852 	struct aio_job *a_job;
853 	int i, error, timo;
854 
855 	if (p->p_aio == NULL)
856 		return EAGAIN;
857 	aio = p->p_aio;
858 
859 	if (ts) {
860 		timo = mstohz((ts->tv_sec * 1000) + (ts->tv_nsec / 1000000));
861 		if (timo == 0 && ts->tv_sec == 0 && ts->tv_nsec > 0)
862 			timo = 1;
863 		if (timo <= 0)
864 			return EAGAIN;
865 	} else
866 		timo = 0;
867 
868 	mutex_enter(&aio->aio_mtx);
869 	for (;;) {
870 		for (i = 0; i < nent; i++) {
871 
872 			/* Skip NULL entries */
873 			if (aiocbp_list[i] == NULL)
874 				continue;
875 
876 			/* Skip current job */
877 			if (aio->curjob) {
878 				a_job = aio->curjob;
879 				if (a_job->aiocb_uptr == aiocbp_list[i])
880 					continue;
881 			}
882 
883 			/* Look for a job in the queue */
884 			TAILQ_FOREACH(a_job, &aio->jobs_queue, list)
885 				if (a_job->aiocb_uptr == aiocbp_list[i])
886 					break;
887 
888 			if (a_job == NULL) {
889 				struct aiocb aiocbp;
890 
891 				mutex_exit(&aio->aio_mtx);
892 
893 				/* Check if the job is done. */
894 				error = copyin(aiocbp_list[i], &aiocbp,
895 				    sizeof(struct aiocb));
896 				if (error == 0 && aiocbp._state != JOB_DONE) {
897 					mutex_enter(&aio->aio_mtx);
898 					continue;
899 				}
900 				return error;
901 			}
902 		}
903 
904 		/* Wait for a signal or when timeout occurs */
905 		error = cv_timedwait_sig(&aio->done_cv, &aio->aio_mtx, timo);
906 		if (error) {
907 			if (error == EWOULDBLOCK)
908 				error = EAGAIN;
909 			break;
910 		}
911 	}
912 	mutex_exit(&aio->aio_mtx);
913 	return error;
914 }
915 
916 int
917 sys_aio_write(struct lwp *l, const struct sys_aio_write_args *uap,
918     register_t *retval)
919 {
920 	/* {
921 		syscallarg(struct aiocb *) aiocbp;
922 	} */
923 
924 	return aio_enqueue_job(AIO_WRITE, SCARG(uap, aiocbp), NULL);
925 }
926 
927 int
928 sys_lio_listio(struct lwp *l, const struct sys_lio_listio_args *uap,
929     register_t *retval)
930 {
931 	/* {
932 		syscallarg(int) mode;
933 		syscallarg(struct aiocb *const[]) list;
934 		syscallarg(int) nent;
935 		syscallarg(struct sigevent *) sig;
936 	} */
937 	struct proc *p = l->l_proc;
938 	struct aioproc *aio;
939 	struct aiocb **aiocbp_list;
940 	struct lio_req *lio;
941 	int i, error, errcnt, mode, nent;
942 
943 	mode = SCARG(uap, mode);
944 	nent = SCARG(uap, nent);
945 
946 	/* Non-accurate checks for the limit and invalid values */
947 	if (nent < 1 || nent > aio_listio_max)
948 		return EINVAL;
949 	if (aio_jobs_count + nent > aio_max)
950 		return EAGAIN;
951 
952 	/* Check if AIO structure is initialized, if not - initialize it */
953 	if (p->p_aio == NULL)
954 		if (aio_procinit(p))
955 			return EAGAIN;
956 	aio = p->p_aio;
957 
958 	/* Create a LIO structure */
959 	lio = pool_get(&aio_lio_pool, PR_WAITOK);
960 	lio->refcnt = 1;
961 	error = 0;
962 
963 	switch (mode) {
964 	case LIO_WAIT:
965 		memset(&lio->sig, 0, sizeof(struct sigevent));
966 		break;
967 	case LIO_NOWAIT:
968 		/* Check for signal, validate it */
969 		if (SCARG(uap, sig)) {
970 			struct sigevent *sig = &lio->sig;
971 
972 			error = copyin(SCARG(uap, sig), &lio->sig,
973 			    sizeof(struct sigevent));
974 			if (error == 0 &&
975 			    (sig->sigev_signo < 0 ||
976 			    sig->sigev_signo >= NSIG ||
977 			    sig->sigev_notify < SIGEV_NONE ||
978 			    sig->sigev_notify > SIGEV_SA))
979 				error = EINVAL;
980 		} else
981 			memset(&lio->sig, 0, sizeof(struct sigevent));
982 		break;
983 	default:
984 		error = EINVAL;
985 		break;
986 	}
987 
988 	if (error != 0) {
989 		pool_put(&aio_lio_pool, lio);
990 		return error;
991 	}
992 
993 	/* Get the list from user-space */
994 	aiocbp_list = kmem_alloc(nent * sizeof(*aiocbp_list), KM_SLEEP);
995 	error = copyin(SCARG(uap, list), aiocbp_list,
996 	    nent * sizeof(*aiocbp_list));
997 	if (error) {
998 		mutex_enter(&aio->aio_mtx);
999 		goto err;
1000 	}
1001 
1002 	/* Enqueue all jobs */
1003 	errcnt = 0;
1004 	for (i = 0; i < nent; i++) {
1005 		error = aio_enqueue_job(AIO_LIO, aiocbp_list[i], lio);
1006 		/*
1007 		 * According to POSIX, in such error case it may
1008 		 * fail with other I/O operations initiated.
1009 		 */
1010 		if (error)
1011 			errcnt++;
1012 	}
1013 
1014 	mutex_enter(&aio->aio_mtx);
1015 
1016 	/* Return an error, if any */
1017 	if (errcnt) {
1018 		error = EIO;
1019 		goto err;
1020 	}
1021 
1022 	if (mode == LIO_WAIT) {
1023 		/*
1024 		 * Wait for AIO completion.  In such case,
1025 		 * the LIO structure will be freed here.
1026 		 */
1027 		while (lio->refcnt > 1 && error == 0)
1028 			error = cv_wait_sig(&aio->done_cv, &aio->aio_mtx);
1029 		if (error)
1030 			error = EINTR;
1031 	}
1032 
1033 err:
1034 	if (--lio->refcnt != 0)
1035 		lio = NULL;
1036 	mutex_exit(&aio->aio_mtx);
1037 	if (lio != NULL) {
1038 		aio_sendsig(p, &lio->sig);
1039 		pool_put(&aio_lio_pool, lio);
1040 	}
1041 	kmem_free(aiocbp_list, nent * sizeof(*aiocbp_list));
1042 	return error;
1043 }
1044 
1045 /*
1046  * SysCtl
1047  */
1048 
1049 static int
1050 sysctl_aio_listio_max(SYSCTLFN_ARGS)
1051 {
1052 	struct sysctlnode node;
1053 	int error, newsize;
1054 
1055 	node = *rnode;
1056 	node.sysctl_data = &newsize;
1057 
1058 	newsize = aio_listio_max;
1059 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1060 	if (error || newp == NULL)
1061 		return error;
1062 
1063 	if (newsize < 1 || newsize > aio_max)
1064 		return EINVAL;
1065 	aio_listio_max = newsize;
1066 
1067 	return 0;
1068 }
1069 
1070 static int
1071 sysctl_aio_max(SYSCTLFN_ARGS)
1072 {
1073 	struct sysctlnode node;
1074 	int error, newsize;
1075 
1076 	node = *rnode;
1077 	node.sysctl_data = &newsize;
1078 
1079 	newsize = aio_max;
1080 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1081 	if (error || newp == NULL)
1082 		return error;
1083 
1084 	if (newsize < 1 || newsize < aio_listio_max)
1085 		return EINVAL;
1086 	aio_max = newsize;
1087 
1088 	return 0;
1089 }
1090 
1091 static int
1092 sysctl_aio_init(void)
1093 {
1094 	int rv;
1095 
1096 	aio_sysctl = NULL;
1097 
1098 	rv = sysctl_createv(&aio_sysctl, 0, NULL, NULL,
1099 		CTLFLAG_PERMANENT | CTLFLAG_IMMEDIATE,
1100 		CTLTYPE_INT, "posix_aio",
1101 		SYSCTL_DESCR("Version of IEEE Std 1003.1 and its "
1102 			     "Asynchronous I/O option to which the "
1103 			     "system attempts to conform"),
1104 		NULL, _POSIX_ASYNCHRONOUS_IO, NULL, 0,
1105 		CTL_KERN, CTL_CREATE, CTL_EOL);
1106 
1107 	if (rv != 0)
1108 		return rv;
1109 
1110 	rv = sysctl_createv(&aio_sysctl, 0, NULL, NULL,
1111 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
1112 		CTLTYPE_INT, "aio_listio_max",
1113 		SYSCTL_DESCR("Maximum number of asynchronous I/O "
1114 			     "operations in a single list I/O call"),
1115 		sysctl_aio_listio_max, 0, &aio_listio_max, 0,
1116 		CTL_KERN, CTL_CREATE, CTL_EOL);
1117 
1118 	if (rv != 0)
1119 		return rv;
1120 
1121 	rv = sysctl_createv(&aio_sysctl, 0, NULL, NULL,
1122 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
1123 		CTLTYPE_INT, "aio_max",
1124 		SYSCTL_DESCR("Maximum number of asynchronous I/O "
1125 			     "operations"),
1126 		sysctl_aio_max, 0, &aio_max, 0,
1127 		CTL_KERN, CTL_CREATE, CTL_EOL);
1128 
1129 	return rv;
1130 }
1131 
1132 /*
1133  * Debugging
1134  */
1135 #if defined(DDB)
1136 void
1137 aio_print_jobs(void (*pr)(const char *, ...))
1138 {
1139 	struct proc *p = curlwp->l_proc;
1140 	struct aioproc *aio;
1141 	struct aio_job *a_job;
1142 	struct aiocb *aiocbp;
1143 
1144 	if (p == NULL) {
1145 		(*pr)("AIO: We are not in the processes right now.\n");
1146 		return;
1147 	}
1148 
1149 	aio = p->p_aio;
1150 	if (aio == NULL) {
1151 		(*pr)("AIO data is not initialized (PID = %d).\n", p->p_pid);
1152 		return;
1153 	}
1154 
1155 	(*pr)("AIO: PID = %d\n", p->p_pid);
1156 	(*pr)("AIO: Global count of the jobs = %u\n", aio_jobs_count);
1157 	(*pr)("AIO: Count of the jobs = %u\n", aio->jobs_count);
1158 
1159 	if (aio->curjob) {
1160 		a_job = aio->curjob;
1161 		(*pr)("\nAIO current job:\n");
1162 		(*pr)(" opcode = %d, errno = %d, state = %d, aiocb_ptr = %p\n",
1163 		    a_job->aio_op, a_job->aiocbp._errno,
1164 		    a_job->aiocbp._state, a_job->aiocb_uptr);
1165 		aiocbp = &a_job->aiocbp;
1166 		(*pr)("   fd = %d, offset = %u, buf = %p, nbytes = %u\n",
1167 		    aiocbp->aio_fildes, aiocbp->aio_offset,
1168 		    aiocbp->aio_buf, aiocbp->aio_nbytes);
1169 	}
1170 
1171 	(*pr)("\nAIO queue:\n");
1172 	TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
1173 		(*pr)(" opcode = %d, errno = %d, state = %d, aiocb_ptr = %p\n",
1174 		    a_job->aio_op, a_job->aiocbp._errno,
1175 		    a_job->aiocbp._state, a_job->aiocb_uptr);
1176 		aiocbp = &a_job->aiocbp;
1177 		(*pr)("   fd = %d, offset = %u, buf = %p, nbytes = %u\n",
1178 		    aiocbp->aio_fildes, aiocbp->aio_offset,
1179 		    aiocbp->aio_buf, aiocbp->aio_nbytes);
1180 	}
1181 }
1182 #endif /* defined(DDB) */
1183