xref: /netbsd-src/sys/kern/sys_aio.c (revision da9817918ec7e88db2912a2882967c7570a83f47)
1 /*	$NetBSD: sys_aio.c,v 1.24 2009/05/24 21:41:26 ad 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 1.24 2009/05/24 21:41:26 ad 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 pool aio_job_pool;
77 static struct pool aio_lio_pool;
78 static void *aio_ehook;
79 
80 /* Prototypes */
81 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 const struct syscall_package aio_syscalls[] = {
88 	{ SYS_aio_cancel, 0, (sy_call_t *)sys_aio_cancel },
89 	{ SYS_aio_error, 0, (sy_call_t *)sys_aio_error },
90 	{ SYS_aio_fsync, 0, (sy_call_t *)sys_aio_fsync },
91 	{ SYS_aio_read, 0, (sy_call_t *)sys_aio_read },
92 	{ SYS_aio_return, 0, (sy_call_t *)sys_aio_return },
93 	{ SYS___aio_suspend50, 0, (sy_call_t *)sys___aio_suspend50 },
94 	{ SYS_aio_write, 0, (sy_call_t *)sys_aio_write },
95 	{ SYS_lio_listio, 0, (sy_call_t *)sys_lio_listio },
96 	{ 0, 0, NULL },
97 };
98 
99 /*
100  * Tear down all AIO state.
101  */
102 static int
103 aio_fini(bool interface)
104 {
105 	int error;
106 	proc_t *p;
107 
108 	if (interface) {
109 		/* Stop syscall activity. */
110 		error = syscall_disestablish(NULL, aio_syscalls);
111 		if (error != 0)
112 			return error;
113 		/* Abort if any processes are using AIO. */
114 		mutex_enter(proc_lock);
115 		PROCLIST_FOREACH(p, &allproc) {
116 			if (p->p_aio != NULL)
117 				break;
118 		}
119 		mutex_exit(proc_lock);
120 		if (p != NULL) {
121 			error = syscall_establish(NULL, aio_syscalls);
122 			KASSERT(error == 0);
123 			return EBUSY;
124 		}
125 	}
126 	KASSERT(aio_jobs_count == 0);
127 	exithook_disestablish(aio_ehook);
128 	pool_destroy(&aio_job_pool);
129 	pool_destroy(&aio_lio_pool);
130 	return 0;
131 }
132 
133 /*
134  * Initialize global AIO state.
135  */
136 static int
137 aio_init(void)
138 {
139 	int error;
140 
141 	pool_init(&aio_job_pool, sizeof(struct aio_job), 0, 0, 0,
142 	    "aio_jobs_pool", &pool_allocator_nointr, IPL_NONE);
143 	pool_init(&aio_lio_pool, sizeof(struct lio_req), 0, 0, 0,
144 	    "aio_lio_pool", &pool_allocator_nointr, IPL_NONE);
145 	aio_ehook = exithook_establish(aio_exit, NULL);
146 	error = syscall_establish(NULL, aio_syscalls);
147 	if (error != 0)
148 		aio_fini(false);
149 	return error;
150 }
151 
152 /*
153  * Module interface.
154  */
155 static int
156 aio_modcmd(modcmd_t cmd, void *arg)
157 {
158 
159 	switch (cmd) {
160 	case MODULE_CMD_INIT:
161 		return aio_init();
162 	case MODULE_CMD_FINI:
163 		return aio_fini(true);
164 	default:
165 		return ENOTTY;
166 	}
167 }
168 
169 /*
170  * Initialize Asynchronous I/O data structures for the process.
171  */
172 static int
173 aio_procinit(struct proc *p)
174 {
175 	struct aioproc *aio;
176 	struct lwp *l;
177 	int error;
178 	bool inmem;
179 	vaddr_t uaddr;
180 
181 	/* Allocate and initialize AIO structure */
182 	aio = kmem_zalloc(sizeof(struct aioproc), KM_SLEEP);
183 	if (aio == NULL)
184 		return EAGAIN;
185 
186 	/* Initialize queue and their synchronization structures */
187 	mutex_init(&aio->aio_mtx, MUTEX_DEFAULT, IPL_NONE);
188 	cv_init(&aio->aio_worker_cv, "aiowork");
189 	cv_init(&aio->done_cv, "aiodone");
190 	TAILQ_INIT(&aio->jobs_queue);
191 
192 	/*
193 	 * Create an AIO worker thread.
194 	 * XXX: Currently, AIO thread is not protected against user's actions.
195 	 */
196 	inmem = uvm_uarea_alloc(&uaddr);
197 	if (uaddr == 0) {
198 		aio_exit(p, aio);
199 		return EAGAIN;
200 	}
201 	error = lwp_create(curlwp, p, uaddr, inmem, 0, NULL, 0, aio_worker,
202 	    NULL, &l, curlwp->l_class);
203 	if (error != 0) {
204 		uvm_uarea_free(uaddr, curcpu());
205 		aio_exit(p, aio);
206 		return error;
207 	}
208 
209 	/* Recheck if we are really first */
210 	mutex_enter(p->p_lock);
211 	if (p->p_aio) {
212 		mutex_exit(p->p_lock);
213 		aio_exit(p, aio);
214 		lwp_exit(l);
215 		return 0;
216 	}
217 	p->p_aio = aio;
218 
219 	/* Complete the initialization of thread, and run it */
220 	aio->aio_worker = l;
221 	p->p_nrlwps++;
222 	lwp_lock(l);
223 	l->l_stat = LSRUN;
224 	l->l_priority = MAXPRI_USER;
225 	sched_enqueue(l, false);
226 	lwp_unlock(l);
227 	mutex_exit(p->p_lock);
228 
229 	return 0;
230 }
231 
232 /*
233  * Exit of Asynchronous I/O subsystem of process.
234  */
235 static void
236 aio_exit(struct proc *p, void *cookie)
237 {
238 	struct aio_job *a_job;
239 	struct aioproc *aio;
240 
241 	if (cookie != NULL)
242 		aio = cookie;
243 	else if ((aio = p->p_aio) == NULL)
244 		return;
245 
246 	/* Free AIO queue */
247 	while (!TAILQ_EMPTY(&aio->jobs_queue)) {
248 		a_job = TAILQ_FIRST(&aio->jobs_queue);
249 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
250 		pool_put(&aio_job_pool, a_job);
251 		atomic_dec_uint(&aio_jobs_count);
252 	}
253 
254 	/* Destroy and free the entire AIO data structure */
255 	cv_destroy(&aio->aio_worker_cv);
256 	cv_destroy(&aio->done_cv);
257 	mutex_destroy(&aio->aio_mtx);
258 	kmem_free(aio, sizeof(struct aioproc));
259 }
260 
261 /*
262  * AIO worker thread and processor.
263  */
264 void
265 aio_worker(void *arg)
266 {
267 	struct proc *p = curlwp->l_proc;
268 	struct aioproc *aio = p->p_aio;
269 	struct aio_job *a_job;
270 	struct lio_req *lio;
271 	sigset_t oss, nss;
272 	int error, refcnt;
273 
274 	/*
275 	 * Make an empty signal mask, so it
276 	 * handles only SIGKILL and SIGSTOP.
277 	 */
278 	sigfillset(&nss);
279 	mutex_enter(p->p_lock);
280 	error = sigprocmask1(curlwp, SIG_SETMASK, &nss, &oss);
281 	mutex_exit(p->p_lock);
282 	KASSERT(error == 0);
283 
284 	for (;;) {
285 		/*
286 		 * Loop for each job in the queue.  If there
287 		 * are no jobs then sleep.
288 		 */
289 		mutex_enter(&aio->aio_mtx);
290 		while ((a_job = TAILQ_FIRST(&aio->jobs_queue)) == NULL) {
291 			if (cv_wait_sig(&aio->aio_worker_cv, &aio->aio_mtx)) {
292 				/*
293 				 * Thread was interrupted - check for
294 				 * pending exit or suspend.
295 				 */
296 				mutex_exit(&aio->aio_mtx);
297 				lwp_userret(curlwp);
298 				mutex_enter(&aio->aio_mtx);
299 			}
300 		}
301 
302 		/* Take the job from the queue */
303 		aio->curjob = a_job;
304 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
305 
306 		atomic_dec_uint(&aio_jobs_count);
307 		aio->jobs_count--;
308 
309 		mutex_exit(&aio->aio_mtx);
310 
311 		/* Process an AIO operation */
312 		aio_process(a_job);
313 
314 		/* Copy data structure back to the user-space */
315 		(void)copyout(&a_job->aiocbp, a_job->aiocb_uptr,
316 		    sizeof(struct aiocb));
317 
318 		mutex_enter(&aio->aio_mtx);
319 		aio->curjob = NULL;
320 
321 		/* Decrease a reference counter, if there is a LIO structure */
322 		lio = a_job->lio;
323 		refcnt = (lio != NULL ? --lio->refcnt : -1);
324 
325 		/* Notify all suspenders */
326 		cv_broadcast(&aio->done_cv);
327 		mutex_exit(&aio->aio_mtx);
328 
329 		/* Send a signal, if any */
330 		aio_sendsig(p, &a_job->aiocbp.aio_sigevent);
331 
332 		/* Destroy the LIO structure */
333 		if (refcnt == 0) {
334 			aio_sendsig(p, &lio->sig);
335 			pool_put(&aio_lio_pool, lio);
336 		}
337 
338 		/* Destroy the the job */
339 		pool_put(&aio_job_pool, a_job);
340 	}
341 
342 	/* NOTREACHED */
343 }
344 
345 static void
346 aio_process(struct aio_job *a_job)
347 {
348 	struct proc *p = curlwp->l_proc;
349 	struct aiocb *aiocbp = &a_job->aiocbp;
350 	struct file *fp;
351 	int fd = aiocbp->aio_fildes;
352 	int error = 0;
353 
354 	KASSERT(a_job->aio_op != 0);
355 
356 	if ((a_job->aio_op & (AIO_READ | AIO_WRITE)) != 0) {
357 		struct iovec aiov;
358 		struct uio auio;
359 
360 		if (aiocbp->aio_nbytes > SSIZE_MAX) {
361 			error = EINVAL;
362 			goto done;
363 		}
364 
365 		fp = fd_getfile(fd);
366 		if (fp == NULL) {
367 			error = EBADF;
368 			goto done;
369 		}
370 
371 		aiov.iov_base = (void *)(uintptr_t)aiocbp->aio_buf;
372 		aiov.iov_len = aiocbp->aio_nbytes;
373 		auio.uio_iov = &aiov;
374 		auio.uio_iovcnt = 1;
375 		auio.uio_resid = aiocbp->aio_nbytes;
376 		auio.uio_vmspace = p->p_vmspace;
377 
378 		if (a_job->aio_op & AIO_READ) {
379 			/*
380 			 * Perform a Read operation
381 			 */
382 			KASSERT((a_job->aio_op & AIO_WRITE) == 0);
383 
384 			if ((fp->f_flag & FREAD) == 0) {
385 				fd_putfile(fd);
386 				error = EBADF;
387 				goto done;
388 			}
389 			auio.uio_rw = UIO_READ;
390 			error = (*fp->f_ops->fo_read)(fp, &aiocbp->aio_offset,
391 			    &auio, fp->f_cred, FOF_UPDATE_OFFSET);
392 		} else {
393 			/*
394 			 * Perform a Write operation
395 			 */
396 			KASSERT(a_job->aio_op & AIO_WRITE);
397 
398 			if ((fp->f_flag & FWRITE) == 0) {
399 				fd_putfile(fd);
400 				error = EBADF;
401 				goto done;
402 			}
403 			auio.uio_rw = UIO_WRITE;
404 			error = (*fp->f_ops->fo_write)(fp, &aiocbp->aio_offset,
405 			    &auio, fp->f_cred, FOF_UPDATE_OFFSET);
406 		}
407 		fd_putfile(fd);
408 
409 		/* Store the result value */
410 		a_job->aiocbp.aio_nbytes -= auio.uio_resid;
411 		a_job->aiocbp._retval = (error == 0) ?
412 		    a_job->aiocbp.aio_nbytes : -1;
413 
414 	} else if ((a_job->aio_op & (AIO_SYNC | AIO_DSYNC)) != 0) {
415 		/*
416 		 * Perform a file Sync operation
417 		 */
418 		struct vnode *vp;
419 
420 		if ((error = fd_getvnode(fd, &fp)) != 0)
421 			goto done;
422 
423 		if ((fp->f_flag & FWRITE) == 0) {
424 			fd_putfile(fd);
425 			error = EBADF;
426 			goto done;
427 		}
428 
429 		vp = (struct vnode *)fp->f_data;
430 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
431 		if (a_job->aio_op & AIO_DSYNC) {
432 			error = VOP_FSYNC(vp, fp->f_cred,
433 			    FSYNC_WAIT | FSYNC_DATAONLY, 0, 0);
434 		} else if (a_job->aio_op & AIO_SYNC) {
435 			error = VOP_FSYNC(vp, fp->f_cred,
436 			    FSYNC_WAIT, 0, 0);
437 		}
438 		VOP_UNLOCK(vp, 0);
439 		fd_putfile(fd);
440 
441 		/* Store the result value */
442 		a_job->aiocbp._retval = (error == 0) ? 0 : -1;
443 
444 	} else
445 		panic("aio_process: invalid operation code\n");
446 
447 done:
448 	/* Job is done, set the error, if any */
449 	a_job->aiocbp._errno = error;
450 	a_job->aiocbp._state = JOB_DONE;
451 }
452 
453 /*
454  * Send AIO signal.
455  */
456 static void
457 aio_sendsig(struct proc *p, struct sigevent *sig)
458 {
459 	ksiginfo_t ksi;
460 
461 	if (sig->sigev_signo == 0 || sig->sigev_notify == SIGEV_NONE)
462 		return;
463 
464 	KSI_INIT(&ksi);
465 	ksi.ksi_signo = sig->sigev_signo;
466 	ksi.ksi_code = SI_ASYNCIO;
467 	ksi.ksi_value = sig->sigev_value;
468 	mutex_enter(proc_lock);
469 	kpsignal(p, &ksi, NULL);
470 	mutex_exit(proc_lock);
471 }
472 
473 /*
474  * Enqueue the job.
475  */
476 static int
477 aio_enqueue_job(int op, void *aiocb_uptr, struct lio_req *lio)
478 {
479 	struct proc *p = curlwp->l_proc;
480 	struct aioproc *aio;
481 	struct aio_job *a_job;
482 	struct aiocb aiocbp;
483 	struct sigevent *sig;
484 	int error;
485 
486 	/* Non-accurate check for the limit */
487 	if (aio_jobs_count + 1 > aio_max)
488 		return EAGAIN;
489 
490 	/* Get the data structure from user-space */
491 	error = copyin(aiocb_uptr, &aiocbp, sizeof(struct aiocb));
492 	if (error)
493 		return error;
494 
495 	/* Check if signal is set, and validate it */
496 	sig = &aiocbp.aio_sigevent;
497 	if (sig->sigev_signo < 0 || sig->sigev_signo >= NSIG ||
498 	    sig->sigev_notify < SIGEV_NONE || sig->sigev_notify > SIGEV_SA)
499 		return EINVAL;
500 
501 	/* Buffer and byte count */
502 	if (((AIO_SYNC | AIO_DSYNC) & op) == 0)
503 		if (aiocbp.aio_buf == NULL || aiocbp.aio_nbytes > SSIZE_MAX)
504 			return EINVAL;
505 
506 	/* Check the opcode, if LIO_NOP - simply ignore */
507 	if (op == AIO_LIO) {
508 		KASSERT(lio != NULL);
509 		if (aiocbp.aio_lio_opcode == LIO_WRITE)
510 			op = AIO_WRITE;
511 		else if (aiocbp.aio_lio_opcode == LIO_READ)
512 			op = AIO_READ;
513 		else
514 			return (aiocbp.aio_lio_opcode == LIO_NOP) ? 0 : EINVAL;
515 	} else {
516 		KASSERT(lio == NULL);
517 	}
518 
519 	/*
520 	 * Look for already existing job.  If found - the job is in-progress.
521 	 * According to POSIX this is invalid, so return the error.
522 	 */
523 	aio = p->p_aio;
524 	if (aio) {
525 		mutex_enter(&aio->aio_mtx);
526 		if (aio->curjob) {
527 			a_job = aio->curjob;
528 			if (a_job->aiocb_uptr == aiocb_uptr) {
529 				mutex_exit(&aio->aio_mtx);
530 				return EINVAL;
531 			}
532 		}
533 		TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
534 			if (a_job->aiocb_uptr != aiocb_uptr)
535 				continue;
536 			mutex_exit(&aio->aio_mtx);
537 			return EINVAL;
538 		}
539 		mutex_exit(&aio->aio_mtx);
540 	}
541 
542 	/*
543 	 * Check if AIO structure is initialized, if not - initialize it.
544 	 * In LIO case, we did that already.  We will recheck this with
545 	 * the lock in aio_procinit().
546 	 */
547 	if (lio == NULL && p->p_aio == NULL)
548 		if (aio_procinit(p))
549 			return EAGAIN;
550 	aio = p->p_aio;
551 
552 	/*
553 	 * Set the state with errno, and copy data
554 	 * structure back to the user-space.
555 	 */
556 	aiocbp._state = JOB_WIP;
557 	aiocbp._errno = EINPROGRESS;
558 	aiocbp._retval = -1;
559 	error = copyout(&aiocbp, aiocb_uptr, sizeof(struct aiocb));
560 	if (error)
561 		return error;
562 
563 	/* Allocate and initialize a new AIO job */
564 	a_job = pool_get(&aio_job_pool, PR_WAITOK);
565 	memset(a_job, 0, sizeof(struct aio_job));
566 
567 	/*
568 	 * Set the data.
569 	 * Store the user-space pointer for searching.  Since we
570 	 * are storing only per proc pointers - it is safe.
571 	 */
572 	memcpy(&a_job->aiocbp, &aiocbp, sizeof(struct aiocb));
573 	a_job->aiocb_uptr = aiocb_uptr;
574 	a_job->aio_op |= op;
575 	a_job->lio = lio;
576 
577 	/*
578 	 * Add the job to the queue, update the counters, and
579 	 * notify the AIO worker thread to handle the job.
580 	 */
581 	mutex_enter(&aio->aio_mtx);
582 
583 	/* Fail, if the limit was reached */
584 	if (atomic_inc_uint_nv(&aio_jobs_count) > aio_max ||
585 	    aio->jobs_count >= aio_listio_max) {
586 		atomic_dec_uint(&aio_jobs_count);
587 		mutex_exit(&aio->aio_mtx);
588 		pool_put(&aio_job_pool, a_job);
589 		return EAGAIN;
590 	}
591 
592 	TAILQ_INSERT_TAIL(&aio->jobs_queue, a_job, list);
593 	aio->jobs_count++;
594 	if (lio)
595 		lio->refcnt++;
596 	cv_signal(&aio->aio_worker_cv);
597 
598 	mutex_exit(&aio->aio_mtx);
599 
600 	/*
601 	 * One would handle the errors only with aio_error() function.
602 	 * This way is appropriate according to POSIX.
603 	 */
604 	return 0;
605 }
606 
607 /*
608  * Syscall functions.
609  */
610 
611 int
612 sys_aio_cancel(struct lwp *l, const struct sys_aio_cancel_args *uap, register_t *retval)
613 {
614 	/* {
615 		syscallarg(int) fildes;
616 		syscallarg(struct aiocb *) aiocbp;
617 	} */
618 	struct proc *p = l->l_proc;
619 	struct aioproc *aio;
620 	struct aio_job *a_job;
621 	struct aiocb *aiocbp_ptr;
622 	struct lio_req *lio;
623 	struct filedesc	*fdp = p->p_fd;
624 	unsigned int cn, errcnt, fildes;
625 	fdtab_t *dt;
626 
627 	TAILQ_HEAD(, aio_job) tmp_jobs_list;
628 
629 	/* Check for invalid file descriptor */
630 	fildes = (unsigned int)SCARG(uap, fildes);
631 	dt = fdp->fd_dt;
632 	if (fildes >= dt->dt_nfiles)
633 		return EBADF;
634 	if (dt->dt_ff[fildes] == NULL || dt->dt_ff[fildes]->ff_file == NULL)
635 		return EBADF;
636 
637 	/* Check if AIO structure is initialized */
638 	if (p->p_aio == NULL) {
639 		*retval = AIO_NOTCANCELED;
640 		return 0;
641 	}
642 
643 	aio = p->p_aio;
644 	aiocbp_ptr = (struct aiocb *)SCARG(uap, aiocbp);
645 
646 	mutex_enter(&aio->aio_mtx);
647 
648 	/* Cancel the jobs, and remove them from the queue */
649 	cn = 0;
650 	TAILQ_INIT(&tmp_jobs_list);
651 	TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
652 		if (aiocbp_ptr) {
653 			if (aiocbp_ptr != a_job->aiocb_uptr)
654 				continue;
655 			if (fildes != a_job->aiocbp.aio_fildes) {
656 				mutex_exit(&aio->aio_mtx);
657 				return EBADF;
658 			}
659 		} else if (a_job->aiocbp.aio_fildes != fildes)
660 			continue;
661 
662 		TAILQ_REMOVE(&aio->jobs_queue, a_job, list);
663 		TAILQ_INSERT_TAIL(&tmp_jobs_list, a_job, list);
664 
665 		/* Decrease the counters */
666 		atomic_dec_uint(&aio_jobs_count);
667 		aio->jobs_count--;
668 		lio = a_job->lio;
669 		if (lio != NULL && --lio->refcnt != 0)
670 			a_job->lio = NULL;
671 
672 		cn++;
673 		if (aiocbp_ptr)
674 			break;
675 	}
676 
677 	/* There are canceled jobs */
678 	if (cn)
679 		*retval = AIO_CANCELED;
680 
681 	/* We cannot cancel current job */
682 	a_job = aio->curjob;
683 	if (a_job && ((a_job->aiocbp.aio_fildes == fildes) ||
684 	    (a_job->aiocb_uptr == aiocbp_ptr)))
685 		*retval = AIO_NOTCANCELED;
686 
687 	mutex_exit(&aio->aio_mtx);
688 
689 	/* Free the jobs after the lock */
690 	errcnt = 0;
691 	while (!TAILQ_EMPTY(&tmp_jobs_list)) {
692 		a_job = TAILQ_FIRST(&tmp_jobs_list);
693 		TAILQ_REMOVE(&tmp_jobs_list, a_job, list);
694 		/* Set the errno and copy structures back to the user-space */
695 		a_job->aiocbp._errno = ECANCELED;
696 		a_job->aiocbp._state = JOB_DONE;
697 		if (copyout(&a_job->aiocbp, a_job->aiocb_uptr,
698 		    sizeof(struct aiocb)))
699 			errcnt++;
700 		/* Send a signal if any */
701 		aio_sendsig(p, &a_job->aiocbp.aio_sigevent);
702 		if (a_job->lio) {
703 			lio = a_job->lio;
704 			aio_sendsig(p, &lio->sig);
705 			pool_put(&aio_lio_pool, lio);
706 		}
707 		pool_put(&aio_job_pool, a_job);
708 	}
709 
710 	if (errcnt)
711 		return EFAULT;
712 
713 	/* Set a correct return value */
714 	if (*retval == 0)
715 		*retval = AIO_ALLDONE;
716 
717 	return 0;
718 }
719 
720 int
721 sys_aio_error(struct lwp *l, const struct sys_aio_error_args *uap, register_t *retval)
722 {
723 	/* {
724 		syscallarg(const struct aiocb *) aiocbp;
725 	} */
726 	struct proc *p = l->l_proc;
727 	struct aioproc *aio = p->p_aio;
728 	struct aiocb aiocbp;
729 	int error;
730 
731 	if (aio == NULL)
732 		return EINVAL;
733 
734 	error = copyin(SCARG(uap, aiocbp), &aiocbp, sizeof(struct aiocb));
735 	if (error)
736 		return error;
737 
738 	if (aiocbp._state == JOB_NONE)
739 		return EINVAL;
740 
741 	*retval = aiocbp._errno;
742 
743 	return 0;
744 }
745 
746 int
747 sys_aio_fsync(struct lwp *l, const struct sys_aio_fsync_args *uap, register_t *retval)
748 {
749 	/* {
750 		syscallarg(int) op;
751 		syscallarg(struct aiocb *) aiocbp;
752 	} */
753 	int op = SCARG(uap, op);
754 
755 	if ((op != O_DSYNC) && (op != O_SYNC))
756 		return EINVAL;
757 
758 	op = O_DSYNC ? AIO_DSYNC : AIO_SYNC;
759 
760 	return aio_enqueue_job(op, SCARG(uap, aiocbp), NULL);
761 }
762 
763 int
764 sys_aio_read(struct lwp *l, const struct sys_aio_read_args *uap, register_t *retval)
765 {
766 	/* {
767 		syscallarg(struct aiocb *) aiocbp;
768 	} */
769 
770 	return aio_enqueue_job(AIO_READ, SCARG(uap, aiocbp), NULL);
771 }
772 
773 int
774 sys_aio_return(struct lwp *l, const struct sys_aio_return_args *uap, register_t *retval)
775 {
776 	/* {
777 		syscallarg(struct aiocb *) aiocbp;
778 	} */
779 	struct proc *p = l->l_proc;
780 	struct aioproc *aio = p->p_aio;
781 	struct aiocb aiocbp;
782 	int error;
783 
784 	if (aio == NULL)
785 		return EINVAL;
786 
787 	error = copyin(SCARG(uap, aiocbp), &aiocbp, sizeof(struct aiocb));
788 	if (error)
789 		return error;
790 
791 	if (aiocbp._errno == EINPROGRESS || aiocbp._state != JOB_DONE)
792 		return EINVAL;
793 
794 	*retval = aiocbp._retval;
795 
796 	/* Reset the internal variables */
797 	aiocbp._errno = 0;
798 	aiocbp._retval = -1;
799 	aiocbp._state = JOB_NONE;
800 	error = copyout(&aiocbp, SCARG(uap, aiocbp), sizeof(struct aiocb));
801 
802 	return error;
803 }
804 
805 int
806 sys___aio_suspend50(struct lwp *l, const struct sys___aio_suspend50_args *uap,
807     register_t *retval)
808 {
809 	/* {
810 		syscallarg(const struct aiocb *const[]) list;
811 		syscallarg(int) nent;
812 		syscallarg(const struct timespec *) timeout;
813 	} */
814 	struct aiocb **list;
815 	struct timespec ts;
816 	int error, nent;
817 
818 	nent = SCARG(uap, nent);
819 	if (nent <= 0 || nent > aio_listio_max)
820 		return EAGAIN;
821 
822 	if (SCARG(uap, timeout)) {
823 		/* Convert timespec to ticks */
824 		error = copyin(SCARG(uap, timeout), &ts,
825 		    sizeof(struct timespec));
826 		if (error)
827 			return error;
828 	}
829 	list = kmem_zalloc(nent * sizeof(struct aio_job), KM_SLEEP);
830 	error = copyin(SCARG(uap, list), list, nent * sizeof(struct aiocb));
831 	if (error)
832 		goto out;
833 	error = aio_suspend1(l, list, nent, SCARG(uap, timeout) ? &ts : NULL);
834 out:
835 	kmem_free(list, nent * sizeof(struct aio_job));
836 	return error;
837 }
838 
839 int
840 aio_suspend1(struct lwp *l, struct aiocb **aiocbp_list, int nent,
841     struct timespec *ts)
842 {
843 	struct proc *p = l->l_proc;
844 	struct aioproc *aio;
845 	struct aio_job *a_job;
846 	int i, error, timo;
847 
848 	if (p->p_aio == NULL)
849 		return EAGAIN;
850 	aio = p->p_aio;
851 
852 	if (ts) {
853 		timo = mstohz((ts->tv_sec * 1000) + (ts->tv_nsec / 1000000));
854 		if (timo == 0 && ts->tv_sec == 0 && ts->tv_nsec > 0)
855 			timo = 1;
856 		if (timo <= 0)
857 			return EAGAIN;
858 	} else
859 		timo = 0;
860 
861 	/* Get the list from user-space */
862 
863 	mutex_enter(&aio->aio_mtx);
864 	for (;;) {
865 
866 		for (i = 0; i < nent; i++) {
867 
868 			/* Skip NULL entries */
869 			if (aiocbp_list[i] == NULL)
870 				continue;
871 
872 			/* Skip current job */
873 			if (aio->curjob) {
874 				a_job = aio->curjob;
875 				if (a_job->aiocb_uptr == aiocbp_list[i])
876 					continue;
877 			}
878 
879 			/* Look for a job in the queue */
880 			TAILQ_FOREACH(a_job, &aio->jobs_queue, list)
881 				if (a_job->aiocb_uptr == aiocbp_list[i])
882 					break;
883 
884 			if (a_job == NULL) {
885 				struct aiocb aiocbp;
886 
887 				mutex_exit(&aio->aio_mtx);
888 
889 				error = copyin(aiocbp_list[i], &aiocbp,
890 				    sizeof(struct aiocb));
891 				if (error == 0 && aiocbp._state != JOB_DONE) {
892 					mutex_enter(&aio->aio_mtx);
893 					continue;
894 				}
895 
896 				kmem_free(aiocbp_list,
897 				    nent * sizeof(struct aio_job));
898 				return error;
899 			}
900 		}
901 
902 		/* Wait for a signal or when timeout occurs */
903 		error = cv_timedwait_sig(&aio->done_cv, &aio->aio_mtx, timo);
904 		if (error) {
905 			if (error == EWOULDBLOCK)
906 				error = EAGAIN;
907 			break;
908 		}
909 	}
910 	mutex_exit(&aio->aio_mtx);
911 	return error;
912 }
913 
914 int
915 sys_aio_write(struct lwp *l, const struct sys_aio_write_args *uap, register_t *retval)
916 {
917 	/* {
918 		syscallarg(struct aiocb *) aiocbp;
919 	} */
920 
921 	return aio_enqueue_job(AIO_WRITE, SCARG(uap, aiocbp), NULL);
922 }
923 
924 int
925 sys_lio_listio(struct lwp *l, const struct sys_lio_listio_args *uap, register_t *retval)
926 {
927 	/* {
928 		syscallarg(int) mode;
929 		syscallarg(struct aiocb *const[]) list;
930 		syscallarg(int) nent;
931 		syscallarg(struct sigevent *) sig;
932 	} */
933 	struct proc *p = l->l_proc;
934 	struct aioproc *aio;
935 	struct aiocb **aiocbp_list;
936 	struct lio_req *lio;
937 	int i, error, errcnt, mode, nent;
938 
939 	mode = SCARG(uap, mode);
940 	nent = SCARG(uap, nent);
941 
942 	/* Non-accurate checks for the limit and invalid values */
943 	if (nent < 1 || nent > aio_listio_max)
944 		return EINVAL;
945 	if (aio_jobs_count + nent > aio_max)
946 		return EAGAIN;
947 
948 	/* Check if AIO structure is initialized, if not - initialize it */
949 	if (p->p_aio == NULL)
950 		if (aio_procinit(p))
951 			return EAGAIN;
952 	aio = p->p_aio;
953 
954 	/* Create a LIO structure */
955 	lio = pool_get(&aio_lio_pool, PR_WAITOK);
956 	lio->refcnt = 1;
957 	error = 0;
958 
959 	switch (mode) {
960 	case LIO_WAIT:
961 		memset(&lio->sig, 0, sizeof(struct sigevent));
962 		break;
963 	case LIO_NOWAIT:
964 		/* Check for signal, validate it */
965 		if (SCARG(uap, sig)) {
966 			struct sigevent *sig = &lio->sig;
967 
968 			error = copyin(SCARG(uap, sig), &lio->sig,
969 			    sizeof(struct sigevent));
970 			if (error == 0 &&
971 			    (sig->sigev_signo < 0 ||
972 			    sig->sigev_signo >= NSIG ||
973 			    sig->sigev_notify < SIGEV_NONE ||
974 			    sig->sigev_notify > SIGEV_SA))
975 				error = EINVAL;
976 		} else
977 			memset(&lio->sig, 0, sizeof(struct sigevent));
978 		break;
979 	default:
980 		error = EINVAL;
981 		break;
982 	}
983 
984 	if (error != 0) {
985 		pool_put(&aio_lio_pool, lio);
986 		return error;
987 	}
988 
989 	/* Get the list from user-space */
990 	aiocbp_list = kmem_zalloc(nent * sizeof(struct aio_job), KM_SLEEP);
991 	error = copyin(SCARG(uap, list), aiocbp_list,
992 	    nent * sizeof(struct aiocb));
993 	if (error) {
994 		mutex_enter(&aio->aio_mtx);
995 		goto err;
996 	}
997 
998 	/* Enqueue all jobs */
999 	errcnt = 0;
1000 	for (i = 0; i < nent; i++) {
1001 		error = aio_enqueue_job(AIO_LIO, aiocbp_list[i], lio);
1002 		/*
1003 		 * According to POSIX, in such error case it may
1004 		 * fail with other I/O operations initiated.
1005 		 */
1006 		if (error)
1007 			errcnt++;
1008 	}
1009 
1010 	mutex_enter(&aio->aio_mtx);
1011 
1012 	/* Return an error, if any */
1013 	if (errcnt) {
1014 		error = EIO;
1015 		goto err;
1016 	}
1017 
1018 	if (mode == LIO_WAIT) {
1019 		/*
1020 		 * Wait for AIO completion.  In such case,
1021 		 * the LIO structure will be freed here.
1022 		 */
1023 		while (lio->refcnt > 1 && error == 0)
1024 			error = cv_wait_sig(&aio->done_cv, &aio->aio_mtx);
1025 		if (error)
1026 			error = EINTR;
1027 	}
1028 
1029 err:
1030 	if (--lio->refcnt != 0)
1031 		lio = NULL;
1032 	mutex_exit(&aio->aio_mtx);
1033 	if (lio != NULL) {
1034 		aio_sendsig(p, &lio->sig);
1035 		pool_put(&aio_lio_pool, lio);
1036 	}
1037 	kmem_free(aiocbp_list, nent * sizeof(struct aio_job));
1038 	return error;
1039 }
1040 
1041 /*
1042  * SysCtl
1043  */
1044 
1045 static int
1046 sysctl_aio_listio_max(SYSCTLFN_ARGS)
1047 {
1048 	struct sysctlnode node;
1049 	int error, newsize;
1050 
1051 	node = *rnode;
1052 	node.sysctl_data = &newsize;
1053 
1054 	newsize = aio_listio_max;
1055 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1056 	if (error || newp == NULL)
1057 		return error;
1058 
1059 	if (newsize < 1 || newsize > aio_max)
1060 		return EINVAL;
1061 	aio_listio_max = newsize;
1062 
1063 	return 0;
1064 }
1065 
1066 static int
1067 sysctl_aio_max(SYSCTLFN_ARGS)
1068 {
1069 	struct sysctlnode node;
1070 	int error, newsize;
1071 
1072 	node = *rnode;
1073 	node.sysctl_data = &newsize;
1074 
1075 	newsize = aio_max;
1076 	error = sysctl_lookup(SYSCTLFN_CALL(&node));
1077 	if (error || newp == NULL)
1078 		return error;
1079 
1080 	if (newsize < 1 || newsize < aio_listio_max)
1081 		return EINVAL;
1082 	aio_max = newsize;
1083 
1084 	return 0;
1085 }
1086 
1087 SYSCTL_SETUP(sysctl_aio_setup, "sysctl aio setup")
1088 {
1089 
1090 	sysctl_createv(clog, 0, NULL, NULL,
1091 		CTLFLAG_PERMANENT,
1092 		CTLTYPE_NODE, "kern", NULL,
1093 		NULL, 0, NULL, 0,
1094 		CTL_KERN, CTL_EOL);
1095 	sysctl_createv(clog, 0, NULL, NULL,
1096 		CTLFLAG_PERMANENT | CTLFLAG_IMMEDIATE,
1097 		CTLTYPE_INT, "posix_aio",
1098 		SYSCTL_DESCR("Version of IEEE Std 1003.1 and its "
1099 			     "Asynchronous I/O option to which the "
1100 			     "system attempts to conform"),
1101 		NULL, _POSIX_ASYNCHRONOUS_IO, NULL, 0,
1102 		CTL_KERN, CTL_CREATE, CTL_EOL);
1103 	sysctl_createv(clog, 0, NULL, NULL,
1104 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
1105 		CTLTYPE_INT, "aio_listio_max",
1106 		SYSCTL_DESCR("Maximum number of asynchronous I/O "
1107 			     "operations in a single list I/O call"),
1108 		sysctl_aio_listio_max, 0, &aio_listio_max, 0,
1109 		CTL_KERN, CTL_CREATE, CTL_EOL);
1110 	sysctl_createv(clog, 0, NULL, NULL,
1111 		CTLFLAG_PERMANENT | CTLFLAG_READWRITE,
1112 		CTLTYPE_INT, "aio_max",
1113 		SYSCTL_DESCR("Maximum number of asynchronous I/O "
1114 			     "operations"),
1115 		sysctl_aio_max, 0, &aio_max, 0,
1116 		CTL_KERN, CTL_CREATE, CTL_EOL);
1117 }
1118 
1119 /*
1120  * Debugging
1121  */
1122 #if defined(DDB)
1123 void
1124 aio_print_jobs(void (*pr)(const char *, ...))
1125 {
1126 	struct proc *p = (curlwp == NULL ? NULL : curlwp->l_proc);
1127 	struct aioproc *aio;
1128 	struct aio_job *a_job;
1129 	struct aiocb *aiocbp;
1130 
1131 	if (p == NULL) {
1132 		(*pr)("AIO: We are not in the processes right now.\n");
1133 		return;
1134 	}
1135 
1136 	aio = p->p_aio;
1137 	if (aio == NULL) {
1138 		(*pr)("AIO data is not initialized (PID = %d).\n", p->p_pid);
1139 		return;
1140 	}
1141 
1142 	(*pr)("AIO: PID = %d\n", p->p_pid);
1143 	(*pr)("AIO: Global count of the jobs = %u\n", aio_jobs_count);
1144 	(*pr)("AIO: Count of the jobs = %u\n", aio->jobs_count);
1145 
1146 	if (aio->curjob) {
1147 		a_job = aio->curjob;
1148 		(*pr)("\nAIO current job:\n");
1149 		(*pr)(" opcode = %d, errno = %d, state = %d, aiocb_ptr = %p\n",
1150 		    a_job->aio_op, a_job->aiocbp._errno,
1151 		    a_job->aiocbp._state, a_job->aiocb_uptr);
1152 		aiocbp = &a_job->aiocbp;
1153 		(*pr)("   fd = %d, offset = %u, buf = %p, nbytes = %u\n",
1154 		    aiocbp->aio_fildes, aiocbp->aio_offset,
1155 		    aiocbp->aio_buf, aiocbp->aio_nbytes);
1156 	}
1157 
1158 	(*pr)("\nAIO queue:\n");
1159 	TAILQ_FOREACH(a_job, &aio->jobs_queue, list) {
1160 		(*pr)(" opcode = %d, errno = %d, state = %d, aiocb_ptr = %p\n",
1161 		    a_job->aio_op, a_job->aiocbp._errno,
1162 		    a_job->aiocbp._state, a_job->aiocb_uptr);
1163 		aiocbp = &a_job->aiocbp;
1164 		(*pr)("   fd = %d, offset = %u, buf = %p, nbytes = %u\n",
1165 		    aiocbp->aio_fildes, aiocbp->aio_offset,
1166 		    aiocbp->aio_buf, aiocbp->aio_nbytes);
1167 	}
1168 }
1169 #endif /* defined(DDB) */
1170