xref: /freebsd/sys/kern/kern_alq.c (revision d79a9edb5ce162c1ba49e12e5c93b894e6a25ad2)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2002, Jeffrey Roberson <jeff@freebsd.org>
5  * Copyright (c) 2008-2009, Lawrence Stewart <lstewart@freebsd.org>
6  * Copyright (c) 2009-2010, The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Portions of this software were developed at the Centre for Advanced
10  * Internet Architectures, Swinburne University of Technology, Melbourne,
11  * Australia by Lawrence Stewart under sponsorship from the FreeBSD Foundation.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice unmodified, this list of conditions, and the following
18  *    disclaimer.
19  * 2. Redistributions in binary form must reproduce the above copyright
20  *    notice, this list of conditions and the following disclaimer in the
21  *    documentation and/or other materials provided with the distribution.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
24  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
25  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
26  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
27  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
28  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
32  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  */
34 
35 #include <sys/cdefs.h>
36 #include "opt_mac.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/alq.h>
41 #include <sys/eventhandler.h>
42 #include <sys/fcntl.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/mount.h>
48 #include <sys/mutex.h>
49 #include <sys/namei.h>
50 #include <sys/proc.h>
51 #include <sys/reboot.h>
52 #include <sys/unistd.h>
53 #include <sys/vnode.h>
54 
55 #include <security/mac/mac_framework.h>
56 
57 /* Async. Logging Queue */
58 struct alq {
59 	char	*aq_entbuf;		/* Buffer for stored entries */
60 	int	aq_entmax;		/* Max entries */
61 	int	aq_entlen;		/* Entry length */
62 	int	aq_freebytes;		/* Bytes available in buffer */
63 	int	aq_buflen;		/* Total length of our buffer */
64 	int	aq_writehead;		/* Location for next write */
65 	int	aq_writetail;		/* Flush starts at this location */
66 	int	aq_wrapearly;		/* # bytes left blank at end of buf */
67 	int	aq_flags;		/* Queue flags */
68 	int	aq_waiters;		/* Num threads waiting for resources
69 					 * NB: Used as a wait channel so must
70 					 * not be first field in the alq struct
71 					 */
72 	struct	ale	aq_getpost;	/* ALE for use by get/post */
73 	struct mtx	aq_mtx;		/* Queue lock */
74 	struct vnode	*aq_vp;		/* Open vnode handle */
75 	struct ucred	*aq_cred;	/* Credentials of the opening thread */
76 	LIST_ENTRY(alq)	aq_act;		/* List of active queues */
77 	LIST_ENTRY(alq)	aq_link;	/* List of all queues */
78 };
79 
80 #define	AQ_WANTED	0x0001		/* Wakeup sleeper when io is done */
81 #define	AQ_ACTIVE	0x0002		/* on the active list */
82 #define	AQ_FLUSHING	0x0004		/* doing IO */
83 #define	AQ_SHUTDOWN	0x0008		/* Queue no longer valid */
84 #define	AQ_ORDERED	0x0010		/* Queue enforces ordered writes */
85 #define	AQ_LEGACY	0x0020		/* Legacy queue (fixed length writes) */
86 
87 #define	ALQ_LOCK(alq)	mtx_lock_spin(&(alq)->aq_mtx)
88 #define	ALQ_UNLOCK(alq)	mtx_unlock_spin(&(alq)->aq_mtx)
89 
90 #define HAS_PENDING_DATA(alq) ((alq)->aq_freebytes != (alq)->aq_buflen)
91 
92 static MALLOC_DEFINE(M_ALD, "ALD", "ALD");
93 
94 /*
95  * The ald_mtx protects the ald_queues list and the ald_active list.
96  */
97 static struct mtx ald_mtx;
98 static LIST_HEAD(, alq) ald_queues;
99 static LIST_HEAD(, alq) ald_active;
100 static int ald_shutingdown = 0;
101 struct thread *ald_thread;
102 static struct proc *ald_proc;
103 static eventhandler_tag alq_eventhandler_tag = NULL;
104 
105 #define	ALD_LOCK()	mtx_lock(&ald_mtx)
106 #define	ALD_UNLOCK()	mtx_unlock(&ald_mtx)
107 
108 /* Daemon functions */
109 static int ald_add(struct alq *);
110 static int ald_rem(struct alq *);
111 static void ald_startup(void *);
112 static void ald_daemon(void);
113 static void ald_shutdown(void *, int);
114 static void ald_activate(struct alq *);
115 static void ald_deactivate(struct alq *);
116 
117 /* Internal queue functions */
118 static void alq_shutdown(struct alq *);
119 static void alq_destroy(struct alq *);
120 static int alq_doio(struct alq *);
121 
122 /*
123  * Add a new queue to the global list.  Fail if we're shutting down.
124  */
125 static int
ald_add(struct alq * alq)126 ald_add(struct alq *alq)
127 {
128 	int error;
129 
130 	error = 0;
131 
132 	ALD_LOCK();
133 	if (ald_shutingdown) {
134 		error = EBUSY;
135 		goto done;
136 	}
137 	LIST_INSERT_HEAD(&ald_queues, alq, aq_link);
138 done:
139 	ALD_UNLOCK();
140 	return (error);
141 }
142 
143 /*
144  * Remove a queue from the global list unless we're shutting down.  If so,
145  * the ald will take care of cleaning up it's resources.
146  */
147 static int
ald_rem(struct alq * alq)148 ald_rem(struct alq *alq)
149 {
150 	int error;
151 
152 	error = 0;
153 
154 	ALD_LOCK();
155 	if (ald_shutingdown) {
156 		error = EBUSY;
157 		goto done;
158 	}
159 	LIST_REMOVE(alq, aq_link);
160 done:
161 	ALD_UNLOCK();
162 	return (error);
163 }
164 
165 /*
166  * Put a queue on the active list.  This will schedule it for writing.
167  */
168 static void
ald_activate(struct alq * alq)169 ald_activate(struct alq *alq)
170 {
171 	LIST_INSERT_HEAD(&ald_active, alq, aq_act);
172 	wakeup(&ald_active);
173 }
174 
175 static void
ald_deactivate(struct alq * alq)176 ald_deactivate(struct alq *alq)
177 {
178 	LIST_REMOVE(alq, aq_act);
179 	alq->aq_flags &= ~AQ_ACTIVE;
180 }
181 
182 static void
ald_startup(void * unused)183 ald_startup(void *unused)
184 {
185 	mtx_init(&ald_mtx, "ALDmtx", NULL, MTX_DEF|MTX_QUIET);
186 	LIST_INIT(&ald_queues);
187 	LIST_INIT(&ald_active);
188 }
189 
190 static void
ald_daemon(void)191 ald_daemon(void)
192 {
193 	int needwakeup;
194 	struct alq *alq;
195 
196 	ald_thread = FIRST_THREAD_IN_PROC(ald_proc);
197 
198 	alq_eventhandler_tag = EVENTHANDLER_REGISTER(shutdown_pre_sync,
199 	    ald_shutdown, NULL, SHUTDOWN_PRI_FIRST);
200 
201 	ALD_LOCK();
202 
203 	for (;;) {
204 		while ((alq = LIST_FIRST(&ald_active)) == NULL &&
205 		    !ald_shutingdown)
206 			mtx_sleep(&ald_active, &ald_mtx, PWAIT, "aldslp", 0);
207 
208 		/* Don't shutdown until all active ALQs are flushed. */
209 		if (ald_shutingdown && alq == NULL) {
210 			ALD_UNLOCK();
211 			break;
212 		}
213 
214 		ALQ_LOCK(alq);
215 		ald_deactivate(alq);
216 		ALD_UNLOCK();
217 		needwakeup = alq_doio(alq);
218 		ALQ_UNLOCK(alq);
219 		if (needwakeup)
220 			wakeup_one(alq);
221 		ALD_LOCK();
222 	}
223 
224 	kproc_exit(0);
225 }
226 
227 static void
ald_shutdown(void * arg,int howto)228 ald_shutdown(void *arg, int howto)
229 {
230 	struct alq *alq;
231 
232 	if ((howto & RB_NOSYNC) != 0 || SCHEDULER_STOPPED())
233 		return;
234 
235 	ALD_LOCK();
236 
237 	/* Ensure no new queues can be created. */
238 	ald_shutingdown = 1;
239 
240 	/* Shutdown all ALQs prior to terminating the ald_daemon. */
241 	while ((alq = LIST_FIRST(&ald_queues)) != NULL) {
242 		LIST_REMOVE(alq, aq_link);
243 		ALD_UNLOCK();
244 		alq_shutdown(alq);
245 		ALD_LOCK();
246 	}
247 
248 	/* At this point, all ALQs are flushed and shutdown. */
249 
250 	/*
251 	 * Wake ald_daemon so that it exits. It won't be able to do
252 	 * anything until we mtx_sleep because we hold the ald_mtx.
253 	 */
254 	wakeup(&ald_active);
255 
256 	/* Wait for ald_daemon to exit. */
257 	mtx_sleep(ald_proc, &ald_mtx, PWAIT, "aldslp", 0);
258 
259 	ALD_UNLOCK();
260 }
261 
262 static void
alq_shutdown(struct alq * alq)263 alq_shutdown(struct alq *alq)
264 {
265 	ALQ_LOCK(alq);
266 
267 	/* Stop any new writers. */
268 	alq->aq_flags |= AQ_SHUTDOWN;
269 
270 	/*
271 	 * If the ALQ isn't active but has unwritten data (possible if
272 	 * the ALQ_NOACTIVATE flag has been used), explicitly activate the
273 	 * ALQ here so that the pending data gets flushed by the ald_daemon.
274 	 */
275 	if (!(alq->aq_flags & AQ_ACTIVE) && HAS_PENDING_DATA(alq)) {
276 		alq->aq_flags |= AQ_ACTIVE;
277 		ALQ_UNLOCK(alq);
278 		ALD_LOCK();
279 		ald_activate(alq);
280 		ALD_UNLOCK();
281 		ALQ_LOCK(alq);
282 	}
283 
284 	/* Drain IO */
285 	while (alq->aq_flags & AQ_ACTIVE) {
286 		alq->aq_flags |= AQ_WANTED;
287 		msleep_spin(alq, &alq->aq_mtx, "aldclose", 0);
288 	}
289 	ALQ_UNLOCK(alq);
290 
291 	vn_close(alq->aq_vp, FWRITE, alq->aq_cred,
292 	    curthread);
293 	crfree(alq->aq_cred);
294 }
295 
296 void
alq_destroy(struct alq * alq)297 alq_destroy(struct alq *alq)
298 {
299 	/* Drain all pending IO. */
300 	alq_shutdown(alq);
301 
302 	mtx_destroy(&alq->aq_mtx);
303 	free(alq->aq_entbuf, M_ALD);
304 	free(alq, M_ALD);
305 }
306 
307 /*
308  * Flush all pending data to disk.  This operation will block.
309  */
310 static int
alq_doio(struct alq * alq)311 alq_doio(struct alq *alq)
312 {
313 	struct thread *td;
314 	struct mount *mp;
315 	struct vnode *vp;
316 	struct uio auio;
317 	struct iovec aiov[2];
318 	int totlen;
319 	int iov;
320 	int wrapearly;
321 
322 	KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__));
323 
324 	vp = alq->aq_vp;
325 	td = curthread;
326 	totlen = 0;
327 	iov = 1;
328 	wrapearly = alq->aq_wrapearly;
329 
330 	bzero(&aiov, sizeof(aiov));
331 	bzero(&auio, sizeof(auio));
332 
333 	/* Start the write from the location of our buffer tail pointer. */
334 	aiov[0].iov_base = alq->aq_entbuf + alq->aq_writetail;
335 
336 	if (alq->aq_writetail < alq->aq_writehead) {
337 		/* Buffer not wrapped. */
338 		totlen = aiov[0].iov_len = alq->aq_writehead - alq->aq_writetail;
339 	} else if (alq->aq_writehead == 0) {
340 		/* Buffer not wrapped (special case to avoid an empty iov). */
341 		totlen = aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail -
342 		    wrapearly;
343 	} else {
344 		/*
345 		 * Buffer wrapped, requires 2 aiov entries:
346 		 * - first is from writetail to end of buffer
347 		 * - second is from start of buffer to writehead
348 		 */
349 		aiov[0].iov_len = alq->aq_buflen - alq->aq_writetail -
350 		    wrapearly;
351 		iov++;
352 		aiov[1].iov_base = alq->aq_entbuf;
353 		aiov[1].iov_len =  alq->aq_writehead;
354 		totlen = aiov[0].iov_len + aiov[1].iov_len;
355 	}
356 
357 	alq->aq_flags |= AQ_FLUSHING;
358 	ALQ_UNLOCK(alq);
359 
360 	auio.uio_iov = &aiov[0];
361 	auio.uio_offset = 0;
362 	auio.uio_segflg = UIO_SYSSPACE;
363 	auio.uio_rw = UIO_WRITE;
364 	auio.uio_iovcnt = iov;
365 	auio.uio_resid = totlen;
366 	auio.uio_td = td;
367 
368 	/*
369 	 * Do all of the junk required to write now.
370 	 */
371 	vn_start_write(vp, &mp, V_WAIT);
372 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
373 	/*
374 	 * XXX: VOP_WRITE error checks are ignored.
375 	 */
376 #ifdef MAC
377 	if (mac_vnode_check_write(alq->aq_cred, NOCRED, vp) == 0)
378 #endif
379 		VOP_WRITE(vp, &auio, IO_UNIT | IO_APPEND, alq->aq_cred);
380 	VOP_UNLOCK(vp);
381 	vn_finished_write(mp);
382 
383 	ALQ_LOCK(alq);
384 	alq->aq_flags &= ~AQ_FLUSHING;
385 
386 	/* Adjust writetail as required, taking into account wrapping. */
387 	alq->aq_writetail = (alq->aq_writetail + totlen + wrapearly) %
388 	    alq->aq_buflen;
389 	alq->aq_freebytes += totlen + wrapearly;
390 
391 	/*
392 	 * If we just flushed part of the buffer which wrapped, reset the
393 	 * wrapearly indicator.
394 	 */
395 	if (wrapearly)
396 		alq->aq_wrapearly = 0;
397 
398 	/*
399 	 * If we just flushed the buffer completely, reset indexes to 0 to
400 	 * minimise buffer wraps.
401 	 * This is also required to ensure alq_getn() can't wedge itself.
402 	 */
403 	if (!HAS_PENDING_DATA(alq))
404 		alq->aq_writehead = alq->aq_writetail = 0;
405 
406 	KASSERT((alq->aq_writetail >= 0 && alq->aq_writetail < alq->aq_buflen),
407 	    ("%s: aq_writetail < 0 || aq_writetail >= aq_buflen", __func__));
408 
409 	if (alq->aq_flags & AQ_WANTED) {
410 		alq->aq_flags &= ~AQ_WANTED;
411 		return (1);
412 	}
413 
414 	return(0);
415 }
416 
417 static struct kproc_desc ald_kp = {
418         "ALQ Daemon",
419         ald_daemon,
420         &ald_proc
421 };
422 
423 SYSINIT(aldthread, SI_SUB_KTHREAD_IDLE, SI_ORDER_ANY, kproc_start, &ald_kp);
424 SYSINIT(ald, SI_SUB_LOCK, SI_ORDER_ANY, ald_startup, NULL);
425 
426 /* User visible queue functions */
427 
428 /*
429  * Create the queue data structure, allocate the buffer, and open the file.
430  */
431 
432 int
alq_open_flags(struct alq ** alqp,const char * file,struct ucred * cred,int cmode,int size,int flags)433 alq_open_flags(struct alq **alqp, const char *file, struct ucred *cred, int cmode,
434     int size, int flags)
435 {
436 	struct nameidata nd;
437 	struct alq *alq;
438 	int oflags;
439 	int error;
440 
441 	KASSERT((size > 0), ("%s: size <= 0", __func__));
442 
443 	*alqp = NULL;
444 
445 	NDINIT(&nd, LOOKUP, NOFOLLOW, UIO_SYSSPACE, file);
446 	oflags = FWRITE | O_NOFOLLOW | O_CREAT;
447 
448 	error = vn_open_cred(&nd, &oflags, cmode, 0, cred, NULL);
449 	if (error)
450 		return (error);
451 
452 	NDFREE_PNBUF(&nd);
453 	/* We just unlock so we hold a reference */
454 	VOP_UNLOCK(nd.ni_vp);
455 
456 	alq = malloc(sizeof(*alq), M_ALD, M_WAITOK|M_ZERO);
457 	alq->aq_vp = nd.ni_vp;
458 	alq->aq_cred = crhold(cred);
459 
460 	mtx_init(&alq->aq_mtx, "ALD Queue", NULL, MTX_SPIN|MTX_QUIET);
461 
462 	alq->aq_buflen = size;
463 	alq->aq_entmax = 0;
464 	alq->aq_entlen = 0;
465 
466 	alq->aq_freebytes = alq->aq_buflen;
467 	alq->aq_entbuf = malloc(alq->aq_buflen, M_ALD, M_WAITOK|M_ZERO);
468 	alq->aq_writehead = alq->aq_writetail = 0;
469 	if (flags & ALQ_ORDERED)
470 		alq->aq_flags |= AQ_ORDERED;
471 
472 	if ((error = ald_add(alq)) != 0) {
473 		alq_destroy(alq);
474 		return (error);
475 	}
476 
477 	*alqp = alq;
478 
479 	return (0);
480 }
481 
482 int
alq_open(struct alq ** alqp,const char * file,struct ucred * cred,int cmode,int size,int count)483 alq_open(struct alq **alqp, const char *file, struct ucred *cred, int cmode,
484     int size, int count)
485 {
486 	int ret;
487 
488 	KASSERT((count >= 0), ("%s: count < 0", __func__));
489 
490 	if (count > 0) {
491 		if ((ret = alq_open_flags(alqp, file, cred, cmode,
492 		    size*count, 0)) == 0) {
493 			(*alqp)->aq_flags |= AQ_LEGACY;
494 			(*alqp)->aq_entmax = count;
495 			(*alqp)->aq_entlen = size;
496 		}
497 	} else
498 		ret = alq_open_flags(alqp, file, cred, cmode, size, 0);
499 
500 	return (ret);
501 }
502 
503 /*
504  * Copy a new entry into the queue.  If the operation would block either
505  * wait or return an error depending on the value of waitok.
506  */
507 int
alq_writen(struct alq * alq,void * data,int len,int flags)508 alq_writen(struct alq *alq, void *data, int len, int flags)
509 {
510 	int activate, copy, ret;
511 	void *waitchan;
512 
513 	KASSERT((len > 0 && len <= alq->aq_buflen),
514 	    ("%s: len <= 0 || len > aq_buflen", __func__));
515 
516 	activate = ret = 0;
517 	copy = len;
518 	waitchan = NULL;
519 
520 	ALQ_LOCK(alq);
521 
522 	/*
523 	 * Fail to perform the write and return EWOULDBLOCK if:
524 	 * - The message is larger than our underlying buffer.
525 	 * - The ALQ is being shutdown.
526 	 * - There is insufficient free space in our underlying buffer
527 	 *   to accept the message and the user can't wait for space.
528 	 * - There is insufficient free space in our underlying buffer
529 	 *   to accept the message and the alq is inactive due to prior
530 	 *   use of the ALQ_NOACTIVATE flag (which would lead to deadlock).
531 	 */
532 	if (len > alq->aq_buflen ||
533 	    alq->aq_flags & AQ_SHUTDOWN ||
534 	    (((flags & ALQ_NOWAIT) || (!(alq->aq_flags & AQ_ACTIVE) &&
535 	    HAS_PENDING_DATA(alq))) && alq->aq_freebytes < len)) {
536 		ALQ_UNLOCK(alq);
537 		return (EWOULDBLOCK);
538 	}
539 
540 	/*
541 	 * If we want ordered writes and there is already at least one thread
542 	 * waiting for resources to become available, sleep until we're woken.
543 	 */
544 	if (alq->aq_flags & AQ_ORDERED && alq->aq_waiters > 0) {
545 		KASSERT(!(flags & ALQ_NOWAIT),
546 		    ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__));
547 		alq->aq_waiters++;
548 		msleep_spin(&alq->aq_waiters, &alq->aq_mtx, "alqwnord", 0);
549 		alq->aq_waiters--;
550 	}
551 
552 	/*
553 	 * (ALQ_WAITOK && aq_freebytes < len) or aq_freebytes >= len, either
554 	 * enter while loop and sleep until we have enough free bytes (former)
555 	 * or skip (latter). If AQ_ORDERED is set, only 1 thread at a time will
556 	 * be in this loop. Otherwise, multiple threads may be sleeping here
557 	 * competing for ALQ resources.
558 	 */
559 	while (alq->aq_freebytes < len && !(alq->aq_flags & AQ_SHUTDOWN)) {
560 		KASSERT(!(flags & ALQ_NOWAIT),
561 		    ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__));
562 		alq->aq_flags |= AQ_WANTED;
563 		alq->aq_waiters++;
564 		if (waitchan)
565 			wakeup(waitchan);
566 		msleep_spin(alq, &alq->aq_mtx, "alqwnres", 0);
567 		alq->aq_waiters--;
568 
569 		/*
570 		 * If we're the first thread to wake after an AQ_WANTED wakeup
571 		 * but there isn't enough free space for us, we're going to loop
572 		 * and sleep again. If there are other threads waiting in this
573 		 * loop, schedule a wakeup so that they can see if the space
574 		 * they require is available.
575 		 */
576 		if (alq->aq_waiters > 0 && !(alq->aq_flags & AQ_ORDERED) &&
577 		    alq->aq_freebytes < len && !(alq->aq_flags & AQ_WANTED))
578 			waitchan = alq;
579 		else
580 			waitchan = NULL;
581 	}
582 
583 	/*
584 	 * If there are waiters, we need to signal the waiting threads after we
585 	 * complete our work. The alq ptr is used as a wait channel for threads
586 	 * requiring resources to be freed up. In the AQ_ORDERED case, threads
587 	 * are not allowed to concurrently compete for resources in the above
588 	 * while loop, so we use a different wait channel in this case.
589 	 */
590 	if (alq->aq_waiters > 0) {
591 		if (alq->aq_flags & AQ_ORDERED)
592 			waitchan = &alq->aq_waiters;
593 		else
594 			waitchan = alq;
595 	} else
596 		waitchan = NULL;
597 
598 	/* Bail if we're shutting down. */
599 	if (alq->aq_flags & AQ_SHUTDOWN) {
600 		ret = EWOULDBLOCK;
601 		goto unlock;
602 	}
603 
604 	/*
605 	 * If we need to wrap the buffer to accommodate the write,
606 	 * we'll need 2 calls to bcopy.
607 	 */
608 	if ((alq->aq_buflen - alq->aq_writehead) < len)
609 		copy = alq->aq_buflen - alq->aq_writehead;
610 
611 	/* Copy message (or part thereof if wrap required) to the buffer. */
612 	bcopy(data, alq->aq_entbuf + alq->aq_writehead, copy);
613 	alq->aq_writehead += copy;
614 
615 	if (alq->aq_writehead >= alq->aq_buflen) {
616 		KASSERT((alq->aq_writehead == alq->aq_buflen),
617 		    ("%s: alq->aq_writehead (%d) > alq->aq_buflen (%d)",
618 		    __func__,
619 		    alq->aq_writehead,
620 		    alq->aq_buflen));
621 		alq->aq_writehead = 0;
622 	}
623 
624 	if (copy != len) {
625 		/*
626 		 * Wrap the buffer by copying the remainder of our message
627 		 * to the start of the buffer and resetting aq_writehead.
628 		 */
629 		bcopy(((uint8_t *)data)+copy, alq->aq_entbuf, len - copy);
630 		alq->aq_writehead = len - copy;
631 	}
632 
633 	KASSERT((alq->aq_writehead >= 0 && alq->aq_writehead < alq->aq_buflen),
634 	    ("%s: aq_writehead < 0 || aq_writehead >= aq_buflen", __func__));
635 
636 	alq->aq_freebytes -= len;
637 
638 	if (!(alq->aq_flags & AQ_ACTIVE) && !(flags & ALQ_NOACTIVATE)) {
639 		alq->aq_flags |= AQ_ACTIVE;
640 		activate = 1;
641 	}
642 
643 	KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__));
644 
645 unlock:
646 	ALQ_UNLOCK(alq);
647 
648 	if (activate) {
649 		ALD_LOCK();
650 		ald_activate(alq);
651 		ALD_UNLOCK();
652 	}
653 
654 	/* NB: We rely on wakeup_one waking threads in a FIFO manner. */
655 	if (waitchan != NULL)
656 		wakeup_one(waitchan);
657 
658 	return (ret);
659 }
660 
661 int
alq_write(struct alq * alq,void * data,int flags)662 alq_write(struct alq *alq, void *data, int flags)
663 {
664 	/* Should only be called in fixed length message (legacy) mode. */
665 	KASSERT((alq->aq_flags & AQ_LEGACY),
666 	    ("%s: fixed length write on variable length queue", __func__));
667 	return (alq_writen(alq, data, alq->aq_entlen, flags));
668 }
669 
670 /*
671  * Retrieve a pointer for the ALQ to write directly into, avoiding bcopy.
672  */
673 struct ale *
alq_getn(struct alq * alq,int len,int flags)674 alq_getn(struct alq *alq, int len, int flags)
675 {
676 	int contigbytes;
677 	void *waitchan;
678 
679 	KASSERT((len > 0 && len <= alq->aq_buflen),
680 	    ("%s: len <= 0 || len > alq->aq_buflen", __func__));
681 
682 	waitchan = NULL;
683 
684 	ALQ_LOCK(alq);
685 
686 	/*
687 	 * Determine the number of free contiguous bytes.
688 	 * We ensure elsewhere that if aq_writehead == aq_writetail because
689 	 * the buffer is empty, they will both be set to 0 and therefore
690 	 * aq_freebytes == aq_buflen and is fully contiguous.
691 	 * If they are equal and the buffer is not empty, aq_freebytes will
692 	 * be 0 indicating the buffer is full.
693 	 */
694 	if (alq->aq_writehead <= alq->aq_writetail)
695 		contigbytes = alq->aq_freebytes;
696 	else {
697 		contigbytes = alq->aq_buflen - alq->aq_writehead;
698 
699 		if (contigbytes < len) {
700 			/*
701 			 * Insufficient space at end of buffer to handle a
702 			 * contiguous write. Wrap early if there's space at
703 			 * the beginning. This will leave a hole at the end
704 			 * of the buffer which we will have to skip over when
705 			 * flushing the buffer to disk.
706 			 */
707 			if (alq->aq_writetail >= len || flags & ALQ_WAITOK) {
708 				/* Keep track of # bytes left blank. */
709 				alq->aq_wrapearly = contigbytes;
710 				/* Do the wrap and adjust counters. */
711 				contigbytes = alq->aq_freebytes =
712 				    alq->aq_writetail;
713 				alq->aq_writehead = 0;
714 			}
715 		}
716 	}
717 
718 	/*
719 	 * Return a NULL ALE if:
720 	 * - The message is larger than our underlying buffer.
721 	 * - The ALQ is being shutdown.
722 	 * - There is insufficient free space in our underlying buffer
723 	 *   to accept the message and the user can't wait for space.
724 	 * - There is insufficient free space in our underlying buffer
725 	 *   to accept the message and the alq is inactive due to prior
726 	 *   use of the ALQ_NOACTIVATE flag (which would lead to deadlock).
727 	 */
728 	if (len > alq->aq_buflen ||
729 	    alq->aq_flags & AQ_SHUTDOWN ||
730 	    (((flags & ALQ_NOWAIT) || (!(alq->aq_flags & AQ_ACTIVE) &&
731 	    HAS_PENDING_DATA(alq))) && contigbytes < len)) {
732 		ALQ_UNLOCK(alq);
733 		return (NULL);
734 	}
735 
736 	/*
737 	 * If we want ordered writes and there is already at least one thread
738 	 * waiting for resources to become available, sleep until we're woken.
739 	 */
740 	if (alq->aq_flags & AQ_ORDERED && alq->aq_waiters > 0) {
741 		KASSERT(!(flags & ALQ_NOWAIT),
742 		    ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__));
743 		alq->aq_waiters++;
744 		msleep_spin(&alq->aq_waiters, &alq->aq_mtx, "alqgnord", 0);
745 		alq->aq_waiters--;
746 	}
747 
748 	/*
749 	 * (ALQ_WAITOK && contigbytes < len) or contigbytes >= len, either enter
750 	 * while loop and sleep until we have enough contiguous free bytes
751 	 * (former) or skip (latter). If AQ_ORDERED is set, only 1 thread at a
752 	 * time will be in this loop. Otherwise, multiple threads may be
753 	 * sleeping here competing for ALQ resources.
754 	 */
755 	while (contigbytes < len && !(alq->aq_flags & AQ_SHUTDOWN)) {
756 		KASSERT(!(flags & ALQ_NOWAIT),
757 		    ("%s: ALQ_NOWAIT set but incorrectly ignored!", __func__));
758 		alq->aq_flags |= AQ_WANTED;
759 		alq->aq_waiters++;
760 		if (waitchan)
761 			wakeup(waitchan);
762 		msleep_spin(alq, &alq->aq_mtx, "alqgnres", 0);
763 		alq->aq_waiters--;
764 
765 		if (alq->aq_writehead <= alq->aq_writetail)
766 			contigbytes = alq->aq_freebytes;
767 		else
768 			contigbytes = alq->aq_buflen - alq->aq_writehead;
769 
770 		/*
771 		 * If we're the first thread to wake after an AQ_WANTED wakeup
772 		 * but there isn't enough free space for us, we're going to loop
773 		 * and sleep again. If there are other threads waiting in this
774 		 * loop, schedule a wakeup so that they can see if the space
775 		 * they require is available.
776 		 */
777 		if (alq->aq_waiters > 0 && !(alq->aq_flags & AQ_ORDERED) &&
778 		    contigbytes < len && !(alq->aq_flags & AQ_WANTED))
779 			waitchan = alq;
780 		else
781 			waitchan = NULL;
782 	}
783 
784 	/*
785 	 * If there are waiters, we need to signal the waiting threads after we
786 	 * complete our work. The alq ptr is used as a wait channel for threads
787 	 * requiring resources to be freed up. In the AQ_ORDERED case, threads
788 	 * are not allowed to concurrently compete for resources in the above
789 	 * while loop, so we use a different wait channel in this case.
790 	 */
791 	if (alq->aq_waiters > 0) {
792 		if (alq->aq_flags & AQ_ORDERED)
793 			waitchan = &alq->aq_waiters;
794 		else
795 			waitchan = alq;
796 	} else
797 		waitchan = NULL;
798 
799 	/* Bail if we're shutting down. */
800 	if (alq->aq_flags & AQ_SHUTDOWN) {
801 		ALQ_UNLOCK(alq);
802 		if (waitchan != NULL)
803 			wakeup_one(waitchan);
804 		return (NULL);
805 	}
806 
807 	/*
808 	 * If we are here, we have a contiguous number of bytes >= len
809 	 * available in our buffer starting at aq_writehead.
810 	 */
811 	alq->aq_getpost.ae_data = alq->aq_entbuf + alq->aq_writehead;
812 	alq->aq_getpost.ae_bytesused = len;
813 
814 	return (&alq->aq_getpost);
815 }
816 
817 struct ale *
alq_get(struct alq * alq,int flags)818 alq_get(struct alq *alq, int flags)
819 {
820 	/* Should only be called in fixed length message (legacy) mode. */
821 	KASSERT((alq->aq_flags & AQ_LEGACY),
822 	    ("%s: fixed length get on variable length queue", __func__));
823 	return (alq_getn(alq, alq->aq_entlen, flags));
824 }
825 
826 void
alq_post_flags(struct alq * alq,struct ale * ale,int flags)827 alq_post_flags(struct alq *alq, struct ale *ale, int flags)
828 {
829 	int activate;
830 	void *waitchan;
831 
832 	activate = 0;
833 
834 	if (ale->ae_bytesused > 0) {
835 		if (!(alq->aq_flags & AQ_ACTIVE) &&
836 		    !(flags & ALQ_NOACTIVATE)) {
837 			alq->aq_flags |= AQ_ACTIVE;
838 			activate = 1;
839 		}
840 
841 		alq->aq_writehead += ale->ae_bytesused;
842 		alq->aq_freebytes -= ale->ae_bytesused;
843 
844 		/* Wrap aq_writehead if we filled to the end of the buffer. */
845 		if (alq->aq_writehead == alq->aq_buflen)
846 			alq->aq_writehead = 0;
847 
848 		KASSERT((alq->aq_writehead >= 0 &&
849 		    alq->aq_writehead < alq->aq_buflen),
850 		    ("%s: aq_writehead < 0 || aq_writehead >= aq_buflen",
851 		    __func__));
852 
853 		KASSERT((HAS_PENDING_DATA(alq)), ("%s: queue empty!", __func__));
854 	}
855 
856 	/*
857 	 * If there are waiters, we need to signal the waiting threads after we
858 	 * complete our work. The alq ptr is used as a wait channel for threads
859 	 * requiring resources to be freed up. In the AQ_ORDERED case, threads
860 	 * are not allowed to concurrently compete for resources in the
861 	 * alq_getn() while loop, so we use a different wait channel in this case.
862 	 */
863 	if (alq->aq_waiters > 0) {
864 		if (alq->aq_flags & AQ_ORDERED)
865 			waitchan = &alq->aq_waiters;
866 		else
867 			waitchan = alq;
868 	} else
869 		waitchan = NULL;
870 
871 	ALQ_UNLOCK(alq);
872 
873 	if (activate) {
874 		ALD_LOCK();
875 		ald_activate(alq);
876 		ALD_UNLOCK();
877 	}
878 
879 	/* NB: We rely on wakeup_one waking threads in a FIFO manner. */
880 	if (waitchan != NULL)
881 		wakeup_one(waitchan);
882 }
883 
884 void
alq_flush(struct alq * alq)885 alq_flush(struct alq *alq)
886 {
887 	int needwakeup = 0;
888 
889 	ALD_LOCK();
890 	ALQ_LOCK(alq);
891 
892 	/*
893 	 * Pull the lever iff there is data to flush and we're
894 	 * not already in the middle of a flush operation.
895 	 */
896 	if (HAS_PENDING_DATA(alq) && !(alq->aq_flags & AQ_FLUSHING)) {
897 		if (alq->aq_flags & AQ_ACTIVE)
898 			ald_deactivate(alq);
899 
900 		ALD_UNLOCK();
901 		needwakeup = alq_doio(alq);
902 	} else
903 		ALD_UNLOCK();
904 
905 	ALQ_UNLOCK(alq);
906 
907 	if (needwakeup)
908 		wakeup_one(alq);
909 }
910 
911 /*
912  * Flush remaining data, close the file and free all resources.
913  */
914 void
alq_close(struct alq * alq)915 alq_close(struct alq *alq)
916 {
917 	/* Only flush and destroy alq if not already shutting down. */
918 	if (ald_rem(alq) == 0)
919 		alq_destroy(alq);
920 }
921 
922 static int
alq_load_handler(module_t mod,int what,void * arg)923 alq_load_handler(module_t mod, int what, void *arg)
924 {
925 	int ret;
926 
927 	ret = 0;
928 
929 	switch (what) {
930 	case MOD_LOAD:
931 	case MOD_SHUTDOWN:
932 		break;
933 
934 	case MOD_QUIESCE:
935 		ALD_LOCK();
936 		/* Only allow unload if there are no open queues. */
937 		if (LIST_FIRST(&ald_queues) == NULL) {
938 			ald_shutingdown = 1;
939 			ALD_UNLOCK();
940 			EVENTHANDLER_DEREGISTER(shutdown_pre_sync,
941 			    alq_eventhandler_tag);
942 			ald_shutdown(NULL, 0);
943 			mtx_destroy(&ald_mtx);
944 		} else {
945 			ALD_UNLOCK();
946 			ret = EBUSY;
947 		}
948 		break;
949 
950 	case MOD_UNLOAD:
951 		/* If MOD_QUIESCE failed we must fail here too. */
952 		if (ald_shutingdown == 0)
953 			ret = EBUSY;
954 		break;
955 
956 	default:
957 		ret = EINVAL;
958 		break;
959 	}
960 
961 	return (ret);
962 }
963 
964 static moduledata_t alq_mod =
965 {
966 	"alq",
967 	alq_load_handler,
968 	NULL
969 };
970 
971 DECLARE_MODULE(alq, alq_mod, SI_SUB_LAST, SI_ORDER_ANY);
972 MODULE_VERSION(alq, 1);
973