xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 2980e352a13e8f0b545a366830c411e7a542ada8)
1 /*	$NetBSD: uvm_swap.c,v 1.139 2008/05/29 14:51:27 mrg Exp $	*/
2 
3 /*
4  * Copyright (c) 1995, 1996, 1997 Matthew R. Green
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 ``AS IS'' AND ANY EXPRESS OR
17  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
18  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
19  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
20  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
21  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
22  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
23  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24  * 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  * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp
29  * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.139 2008/05/29 14:51:27 mrg Exp $");
34 
35 #include "fs_nfs.h"
36 #include "opt_uvmhist.h"
37 #include "opt_compat_netbsd.h"
38 #include "opt_ddb.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/buf.h>
43 #include <sys/bufq.h>
44 #include <sys/conf.h>
45 #include <sys/proc.h>
46 #include <sys/namei.h>
47 #include <sys/disklabel.h>
48 #include <sys/errno.h>
49 #include <sys/kernel.h>
50 #include <sys/malloc.h>
51 #include <sys/vnode.h>
52 #include <sys/file.h>
53 #include <sys/vmem.h>
54 #include <sys/blist.h>
55 #include <sys/mount.h>
56 #include <sys/pool.h>
57 #include <sys/syscallargs.h>
58 #include <sys/swap.h>
59 #include <sys/kauth.h>
60 #include <sys/sysctl.h>
61 #include <sys/workqueue.h>
62 
63 #include <uvm/uvm.h>
64 
65 #include <miscfs/specfs/specdev.h>
66 
67 /*
68  * uvm_swap.c: manage configuration and i/o to swap space.
69  */
70 
71 /*
72  * swap space is managed in the following way:
73  *
74  * each swap partition or file is described by a "swapdev" structure.
75  * each "swapdev" structure contains a "swapent" structure which contains
76  * information that is passed up to the user (via system calls).
77  *
78  * each swap partition is assigned a "priority" (int) which controls
79  * swap parition usage.
80  *
81  * the system maintains a global data structure describing all swap
82  * partitions/files.   there is a sorted LIST of "swappri" structures
83  * which describe "swapdev"'s at that priority.   this LIST is headed
84  * by the "swap_priority" global var.    each "swappri" contains a
85  * CIRCLEQ of "swapdev" structures at that priority.
86  *
87  * locking:
88  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
89  *    system call and prevents the swap priority list from changing
90  *    while we are in the middle of a system call (e.g. SWAP_STATS).
91  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
92  *    structures including the priority list, the swapdev structures,
93  *    and the swapmap arena.
94  *
95  * each swap device has the following info:
96  *  - swap device in use (could be disabled, preventing future use)
97  *  - swap enabled (allows new allocations on swap)
98  *  - map info in /dev/drum
99  *  - vnode pointer
100  * for swap files only:
101  *  - block size
102  *  - max byte count in buffer
103  *  - buffer
104  *
105  * userland controls and configures swap with the swapctl(2) system call.
106  * the sys_swapctl performs the following operations:
107  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
108  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
109  *	(passed in via "arg") of a size passed in via "misc" ... we load
110  *	the current swap config into the array. The actual work is done
111  *	in the uvm_swap_stats(9) function.
112  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
113  *	priority in "misc", start swapping on it.
114  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
115  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
116  *	"misc")
117  */
118 
119 /*
120  * swapdev: describes a single swap partition/file
121  *
122  * note the following should be true:
123  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
124  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
125  */
126 struct swapdev {
127 	struct oswapent swd_ose;
128 #define	swd_dev		swd_ose.ose_dev		/* device id */
129 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
130 #define	swd_priority	swd_ose.ose_priority	/* our priority */
131 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
132 	char			*swd_path;	/* saved pathname of device */
133 	int			swd_pathlen;	/* length of pathname */
134 	int			swd_npages;	/* #pages we can use */
135 	int			swd_npginuse;	/* #pages in use */
136 	int			swd_npgbad;	/* #pages bad */
137 	int			swd_drumoffset;	/* page0 offset in drum */
138 	int			swd_drumsize;	/* #pages in drum */
139 	blist_t			swd_blist;	/* blist for this swapdev */
140 	struct vnode		*swd_vp;	/* backing vnode */
141 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
142 
143 	int			swd_bsize;	/* blocksize (bytes) */
144 	int			swd_maxactive;	/* max active i/o reqs */
145 	struct bufq_state	*swd_tab;	/* buffer list */
146 	int			swd_active;	/* number of active buffers */
147 };
148 
149 /*
150  * swap device priority entry; the list is kept sorted on `spi_priority'.
151  */
152 struct swappri {
153 	int			spi_priority;     /* priority */
154 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
155 	/* circleq of swapdevs at this priority */
156 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
157 };
158 
159 /*
160  * The following two structures are used to keep track of data transfers
161  * on swap devices associated with regular files.
162  * NOTE: this code is more or less a copy of vnd.c; we use the same
163  * structure names here to ease porting..
164  */
165 struct vndxfer {
166 	struct buf	*vx_bp;		/* Pointer to parent buffer */
167 	struct swapdev	*vx_sdp;
168 	int		vx_error;
169 	int		vx_pending;	/* # of pending aux buffers */
170 	int		vx_flags;
171 #define VX_BUSY		1
172 #define VX_DEAD		2
173 };
174 
175 struct vndbuf {
176 	struct buf	vb_buf;
177 	struct vndxfer	*vb_xfer;
178 };
179 
180 
181 /*
182  * We keep a of pool vndbuf's and vndxfer structures.
183  */
184 POOL_INIT(vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx", NULL,
185     IPL_BIO);
186 POOL_INIT(vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd", NULL,
187     IPL_BIO);
188 
189 /*
190  * local variables
191  */
192 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
193 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
194 
195 /* list of all active swap devices [by priority] */
196 LIST_HEAD(swap_priority, swappri);
197 static struct swap_priority swap_priority;
198 
199 /* locks */
200 static krwlock_t swap_syscall_lock;
201 
202 /* workqueue and use counter for swap to regular files */
203 static int sw_reg_count = 0;
204 static struct workqueue *sw_reg_workqueue;
205 
206 /*
207  * prototypes
208  */
209 static struct swapdev	*swapdrum_getsdp(int);
210 
211 static struct swapdev	*swaplist_find(struct vnode *, bool);
212 static void		 swaplist_insert(struct swapdev *,
213 					 struct swappri *, int);
214 static void		 swaplist_trim(void);
215 
216 static int swap_on(struct lwp *, struct swapdev *);
217 static int swap_off(struct lwp *, struct swapdev *);
218 
219 static void uvm_swap_stats_locked(int, struct swapent *, int, register_t *);
220 
221 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
222 static void sw_reg_biodone(struct buf *);
223 static void sw_reg_iodone(struct work *wk, void *dummy);
224 static void sw_reg_start(struct swapdev *);
225 
226 static int uvm_swap_io(struct vm_page **, int, int, int);
227 
228 /*
229  * uvm_swap_init: init the swap system data structures and locks
230  *
231  * => called at boot time from init_main.c after the filesystems
232  *	are brought up (which happens after uvm_init())
233  */
234 void
235 uvm_swap_init(void)
236 {
237 	UVMHIST_FUNC("uvm_swap_init");
238 
239 	UVMHIST_CALLED(pdhist);
240 	/*
241 	 * first, init the swap list, its counter, and its lock.
242 	 * then get a handle on the vnode for /dev/drum by using
243 	 * the its dev_t number ("swapdev", from MD conf.c).
244 	 */
245 
246 	LIST_INIT(&swap_priority);
247 	uvmexp.nswapdev = 0;
248 	rw_init(&swap_syscall_lock);
249 	cv_init(&uvm.scheduler_cv, "schedule");
250 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
251 
252 	/* XXXSMP should be at IPL_VM, but for audio interrupt handlers. */
253 	mutex_init(&uvm_scheduler_mutex, MUTEX_SPIN, IPL_SCHED);
254 
255 	if (bdevvp(swapdev, &swapdev_vp))
256 		panic("uvm_swap_init: can't get vnode for swap device");
257 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
258 		panic("uvm_swap_init: can't lock swap device");
259 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
260 		panic("uvm_swap_init: can't open swap device");
261 	VOP_UNLOCK(swapdev_vp, 0);
262 
263 	/*
264 	 * create swap block resource map to map /dev/drum.   the range
265 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
266 	 * that block 0 is reserved (used to indicate an allocation
267 	 * failure, or no allocation).
268 	 */
269 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
270 	    VM_NOSLEEP, IPL_NONE);
271 	if (swapmap == 0)
272 		panic("uvm_swap_init: extent_create failed");
273 
274 	/*
275 	 * done!
276 	 */
277 	uvm.swap_running = true;
278 	uvm.swapout_enabled = 1;
279 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
280 
281         sysctl_createv(NULL, 0, NULL, NULL,
282             CTLFLAG_READWRITE,
283             CTLTYPE_INT, "swapout",
284             SYSCTL_DESCR("Set 0 to disable swapout of kernel stacks"),
285             NULL, 0, &uvm.swapout_enabled, 0, CTL_VM, CTL_CREATE, CTL_EOL);
286 }
287 
288 /*
289  * swaplist functions: functions that operate on the list of swap
290  * devices on the system.
291  */
292 
293 /*
294  * swaplist_insert: insert swap device "sdp" into the global list
295  *
296  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
297  * => caller must provide a newly malloc'd swappri structure (we will
298  *	FREE it if we don't need it... this it to prevent malloc blocking
299  *	here while adding swap)
300  */
301 static void
302 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
303 {
304 	struct swappri *spp, *pspp;
305 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
306 
307 	/*
308 	 * find entry at or after which to insert the new device.
309 	 */
310 	pspp = NULL;
311 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
312 		if (priority <= spp->spi_priority)
313 			break;
314 		pspp = spp;
315 	}
316 
317 	/*
318 	 * new priority?
319 	 */
320 	if (spp == NULL || spp->spi_priority != priority) {
321 		spp = newspp;  /* use newspp! */
322 		UVMHIST_LOG(pdhist, "created new swappri = %d",
323 			    priority, 0, 0, 0);
324 
325 		spp->spi_priority = priority;
326 		CIRCLEQ_INIT(&spp->spi_swapdev);
327 
328 		if (pspp)
329 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
330 		else
331 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
332 	} else {
333 	  	/* we don't need a new priority structure, free it */
334 		FREE(newspp, M_VMSWAP);
335 	}
336 
337 	/*
338 	 * priority found (or created).   now insert on the priority's
339 	 * circleq list and bump the total number of swapdevs.
340 	 */
341 	sdp->swd_priority = priority;
342 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
343 	uvmexp.nswapdev++;
344 }
345 
346 /*
347  * swaplist_find: find and optionally remove a swap device from the
348  *	global list.
349  *
350  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
351  * => we return the swapdev we found (and removed)
352  */
353 static struct swapdev *
354 swaplist_find(struct vnode *vp, bool remove)
355 {
356 	struct swapdev *sdp;
357 	struct swappri *spp;
358 
359 	/*
360 	 * search the lists for the requested vp
361 	 */
362 
363 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
364 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
365 			if (sdp->swd_vp == vp) {
366 				if (remove) {
367 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
368 					    sdp, swd_next);
369 					uvmexp.nswapdev--;
370 				}
371 				return(sdp);
372 			}
373 		}
374 	}
375 	return (NULL);
376 }
377 
378 /*
379  * swaplist_trim: scan priority list for empty priority entries and kill
380  *	them.
381  *
382  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
383  */
384 static void
385 swaplist_trim(void)
386 {
387 	struct swappri *spp, *nextspp;
388 
389 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
390 		nextspp = LIST_NEXT(spp, spi_swappri);
391 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
392 		    (void *)&spp->spi_swapdev)
393 			continue;
394 		LIST_REMOVE(spp, spi_swappri);
395 		free(spp, M_VMSWAP);
396 	}
397 }
398 
399 /*
400  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
401  *	to the "swapdev" that maps that section of the drum.
402  *
403  * => each swapdev takes one big contig chunk of the drum
404  * => caller must hold uvm_swap_data_lock
405  */
406 static struct swapdev *
407 swapdrum_getsdp(int pgno)
408 {
409 	struct swapdev *sdp;
410 	struct swappri *spp;
411 
412 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
413 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
414 			if (sdp->swd_flags & SWF_FAKE)
415 				continue;
416 			if (pgno >= sdp->swd_drumoffset &&
417 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
418 				return sdp;
419 			}
420 		}
421 	}
422 	return NULL;
423 }
424 
425 
426 /*
427  * sys_swapctl: main entry point for swapctl(2) system call
428  * 	[with two helper functions: swap_on and swap_off]
429  */
430 int
431 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
432 {
433 	/* {
434 		syscallarg(int) cmd;
435 		syscallarg(void *) arg;
436 		syscallarg(int) misc;
437 	} */
438 	struct vnode *vp;
439 	struct nameidata nd;
440 	struct swappri *spp;
441 	struct swapdev *sdp;
442 	struct swapent *sep;
443 #define SWAP_PATH_MAX (PATH_MAX + 1)
444 	char	*userpath;
445 	size_t	len;
446 	int	error, misc;
447 	int	priority;
448 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
449 
450 	misc = SCARG(uap, misc);
451 
452 	/*
453 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
454 	 */
455 	rw_enter(&swap_syscall_lock, RW_WRITER);
456 
457 	userpath = malloc(SWAP_PATH_MAX, M_TEMP, M_WAITOK);
458 	/*
459 	 * we handle the non-priv NSWAP and STATS request first.
460 	 *
461 	 * SWAP_NSWAP: return number of config'd swap devices
462 	 * [can also be obtained with uvmexp sysctl]
463 	 */
464 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
465 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
466 		    0, 0, 0);
467 		*retval = uvmexp.nswapdev;
468 		error = 0;
469 		goto out;
470 	}
471 
472 	/*
473 	 * SWAP_STATS: get stats on current # of configured swap devs
474 	 *
475 	 * note that the swap_priority list can't change as long
476 	 * as we are holding the swap_syscall_lock.  we don't want
477 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
478 	 * copyout() and we don't want to be holding that lock then!
479 	 */
480 	if (SCARG(uap, cmd) == SWAP_STATS
481 #if defined(COMPAT_13)
482 	    || SCARG(uap, cmd) == SWAP_OSTATS
483 #endif
484 	    ) {
485 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
486 			misc = uvmexp.nswapdev;
487 #if defined(COMPAT_13)
488 		if (SCARG(uap, cmd) == SWAP_OSTATS)
489 			len = sizeof(struct oswapent) * misc;
490 		else
491 #endif
492 			len = sizeof(struct swapent) * misc;
493 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
494 
495 		uvm_swap_stats_locked(SCARG(uap, cmd), sep, misc, retval);
496 		error = copyout(sep, SCARG(uap, arg), len);
497 
498 		free(sep, M_TEMP);
499 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
500 		goto out;
501 	}
502 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
503 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
504 
505 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
506 		goto out;
507 	}
508 
509 	/*
510 	 * all other requests require superuser privs.   verify.
511 	 */
512 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
513 	    0, NULL, NULL, NULL)))
514 		goto out;
515 
516 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
517 		/* drop the current dump device */
518 		dumpdev = NODEV;
519 		dumpcdev = NODEV;
520 		cpu_dumpconf();
521 		goto out;
522 	}
523 
524 	/*
525 	 * at this point we expect a path name in arg.   we will
526 	 * use namei() to gain a vnode reference (vref), and lock
527 	 * the vnode (VOP_LOCK).
528 	 *
529 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
530 	 * miniroot)
531 	 */
532 	if (SCARG(uap, arg) == NULL) {
533 		vp = rootvp;		/* miniroot */
534 		if (vget(vp, LK_EXCLUSIVE)) {
535 			error = EBUSY;
536 			goto out;
537 		}
538 		if (SCARG(uap, cmd) == SWAP_ON &&
539 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
540 			panic("swapctl: miniroot copy failed");
541 	} else {
542 		int	space;
543 		char	*where;
544 
545 		if (SCARG(uap, cmd) == SWAP_ON) {
546 			if ((error = copyinstr(SCARG(uap, arg), userpath,
547 			    SWAP_PATH_MAX, &len)))
548 				goto out;
549 			space = UIO_SYSSPACE;
550 			where = userpath;
551 		} else {
552 			space = UIO_USERSPACE;
553 			where = (char *)SCARG(uap, arg);
554 		}
555 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT,
556 		    space, where);
557 		if ((error = namei(&nd)))
558 			goto out;
559 		vp = nd.ni_vp;
560 	}
561 	/* note: "vp" is referenced and locked */
562 
563 	error = 0;		/* assume no error */
564 	switch(SCARG(uap, cmd)) {
565 
566 	case SWAP_DUMPDEV:
567 		if (vp->v_type != VBLK) {
568 			error = ENOTBLK;
569 			break;
570 		}
571 		if (bdevsw_lookup(vp->v_rdev)) {
572 			dumpdev = vp->v_rdev;
573 			dumpcdev = devsw_blk2chr(dumpdev);
574 		} else
575 			dumpdev = NODEV;
576 		cpu_dumpconf();
577 		break;
578 
579 	case SWAP_CTL:
580 		/*
581 		 * get new priority, remove old entry (if any) and then
582 		 * reinsert it in the correct place.  finally, prune out
583 		 * any empty priority structures.
584 		 */
585 		priority = SCARG(uap, misc);
586 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
587 		mutex_enter(&uvm_swap_data_lock);
588 		if ((sdp = swaplist_find(vp, true)) == NULL) {
589 			error = ENOENT;
590 		} else {
591 			swaplist_insert(sdp, spp, priority);
592 			swaplist_trim();
593 		}
594 		mutex_exit(&uvm_swap_data_lock);
595 		if (error)
596 			free(spp, M_VMSWAP);
597 		break;
598 
599 	case SWAP_ON:
600 
601 		/*
602 		 * check for duplicates.   if none found, then insert a
603 		 * dummy entry on the list to prevent someone else from
604 		 * trying to enable this device while we are working on
605 		 * it.
606 		 */
607 
608 		priority = SCARG(uap, misc);
609 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
610 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
611 		memset(sdp, 0, sizeof(*sdp));
612 		sdp->swd_flags = SWF_FAKE;
613 		sdp->swd_vp = vp;
614 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
615 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
616 		mutex_enter(&uvm_swap_data_lock);
617 		if (swaplist_find(vp, false) != NULL) {
618 			error = EBUSY;
619 			mutex_exit(&uvm_swap_data_lock);
620 			bufq_free(sdp->swd_tab);
621 			free(sdp, M_VMSWAP);
622 			free(spp, M_VMSWAP);
623 			break;
624 		}
625 		swaplist_insert(sdp, spp, priority);
626 		mutex_exit(&uvm_swap_data_lock);
627 
628 		sdp->swd_pathlen = len;
629 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
630 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
631 			panic("swapctl: copystr");
632 
633 		/*
634 		 * we've now got a FAKE placeholder in the swap list.
635 		 * now attempt to enable swap on it.  if we fail, undo
636 		 * what we've done and kill the fake entry we just inserted.
637 		 * if swap_on is a success, it will clear the SWF_FAKE flag
638 		 */
639 
640 		if ((error = swap_on(l, sdp)) != 0) {
641 			mutex_enter(&uvm_swap_data_lock);
642 			(void) swaplist_find(vp, true);  /* kill fake entry */
643 			swaplist_trim();
644 			mutex_exit(&uvm_swap_data_lock);
645 			bufq_free(sdp->swd_tab);
646 			free(sdp->swd_path, M_VMSWAP);
647 			free(sdp, M_VMSWAP);
648 			break;
649 		}
650 		break;
651 
652 	case SWAP_OFF:
653 		mutex_enter(&uvm_swap_data_lock);
654 		if ((sdp = swaplist_find(vp, false)) == NULL) {
655 			mutex_exit(&uvm_swap_data_lock);
656 			error = ENXIO;
657 			break;
658 		}
659 
660 		/*
661 		 * If a device isn't in use or enabled, we
662 		 * can't stop swapping from it (again).
663 		 */
664 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
665 			mutex_exit(&uvm_swap_data_lock);
666 			error = EBUSY;
667 			break;
668 		}
669 
670 		/*
671 		 * do the real work.
672 		 */
673 		error = swap_off(l, sdp);
674 		break;
675 
676 	default:
677 		error = EINVAL;
678 	}
679 
680 	/*
681 	 * done!  release the ref gained by namei() and unlock.
682 	 */
683 	vput(vp);
684 
685 out:
686 	free(userpath, M_TEMP);
687 	rw_exit(&swap_syscall_lock);
688 
689 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
690 	return (error);
691 }
692 
693 /*
694  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
695  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
696  * emulation to use it directly without going through sys_swapctl().
697  * The problem with using sys_swapctl() there is that it involves
698  * copying the swapent array to the stackgap, and this array's size
699  * is not known at build time. Hence it would not be possible to
700  * ensure it would fit in the stackgap in any case.
701  */
702 void
703 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
704 {
705 
706 	rw_enter(&swap_syscall_lock, RW_READER);
707 	uvm_swap_stats_locked(cmd, sep, sec, retval);
708 	rw_exit(&swap_syscall_lock);
709 }
710 
711 static void
712 uvm_swap_stats_locked(int cmd, struct swapent *sep, int sec, register_t *retval)
713 {
714 	struct swappri *spp;
715 	struct swapdev *sdp;
716 	int count = 0;
717 
718 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
719 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
720 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
721 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
722 		  	/*
723 			 * backwards compatibility for system call.
724 			 * note that we use 'struct oswapent' as an
725 			 * overlay into both 'struct swapdev' and
726 			 * the userland 'struct swapent', as we
727 			 * want to retain backwards compatibility
728 			 * with NetBSD 1.3.
729 			 */
730 			sdp->swd_ose.ose_inuse =
731 			    btodb((uint64_t)sdp->swd_npginuse <<
732 			    PAGE_SHIFT);
733 			(void)memcpy(sep, &sdp->swd_ose,
734 			    sizeof(struct oswapent));
735 
736 			/* now copy out the path if necessary */
737 #if !defined(COMPAT_13)
738 			(void) cmd;
739 #endif
740 #if defined(COMPAT_13)
741 			if (cmd == SWAP_STATS)
742 #endif
743 				(void)memcpy(&sep->se_path, sdp->swd_path,
744 				    sdp->swd_pathlen);
745 
746 			count++;
747 #if defined(COMPAT_13)
748 			if (cmd == SWAP_OSTATS)
749 				sep = (struct swapent *)
750 				    ((struct oswapent *)sep + 1);
751 			else
752 #endif
753 				sep++;
754 		}
755 	}
756 
757 	*retval = count;
758 	return;
759 }
760 
761 /*
762  * swap_on: attempt to enable a swapdev for swapping.   note that the
763  *	swapdev is already on the global list, but disabled (marked
764  *	SWF_FAKE).
765  *
766  * => we avoid the start of the disk (to protect disk labels)
767  * => we also avoid the miniroot, if we are swapping to root.
768  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
769  *	if needed.
770  */
771 static int
772 swap_on(struct lwp *l, struct swapdev *sdp)
773 {
774 	struct vnode *vp;
775 	int error, npages, nblocks, size;
776 	long addr;
777 	u_long result;
778 	struct vattr va;
779 #ifdef NFS
780 	extern int (**nfsv2_vnodeop_p)(void *);
781 #endif /* NFS */
782 	const struct bdevsw *bdev;
783 	dev_t dev;
784 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
785 
786 	/*
787 	 * we want to enable swapping on sdp.   the swd_vp contains
788 	 * the vnode we want (locked and ref'd), and the swd_dev
789 	 * contains the dev_t of the file, if it a block device.
790 	 */
791 
792 	vp = sdp->swd_vp;
793 	dev = sdp->swd_dev;
794 
795 	/*
796 	 * open the swap file (mostly useful for block device files to
797 	 * let device driver know what is up).
798 	 *
799 	 * we skip the open/close for root on swap because the root
800 	 * has already been opened when root was mounted (mountroot).
801 	 */
802 	if (vp != rootvp) {
803 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
804 			return (error);
805 	}
806 
807 	/* XXX this only works for block devices */
808 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
809 
810 	/*
811 	 * we now need to determine the size of the swap area.   for
812 	 * block specials we can call the d_psize function.
813 	 * for normal files, we must stat [get attrs].
814 	 *
815 	 * we put the result in nblks.
816 	 * for normal files, we also want the filesystem block size
817 	 * (which we get with statfs).
818 	 */
819 	switch (vp->v_type) {
820 	case VBLK:
821 		bdev = bdevsw_lookup(dev);
822 		if (bdev == NULL || bdev->d_psize == NULL ||
823 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
824 			error = ENXIO;
825 			goto bad;
826 		}
827 		break;
828 
829 	case VREG:
830 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
831 			goto bad;
832 		nblocks = (int)btodb(va.va_size);
833 		if ((error =
834 		     VFS_STATVFS(vp->v_mount, &vp->v_mount->mnt_stat)) != 0)
835 			goto bad;
836 
837 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
838 		/*
839 		 * limit the max # of outstanding I/O requests we issue
840 		 * at any one time.   take it easy on NFS servers.
841 		 */
842 #ifdef NFS
843 		if (vp->v_op == nfsv2_vnodeop_p)
844 			sdp->swd_maxactive = 2; /* XXX */
845 		else
846 #endif /* NFS */
847 			sdp->swd_maxactive = 8; /* XXX */
848 		break;
849 
850 	default:
851 		error = ENXIO;
852 		goto bad;
853 	}
854 
855 	/*
856 	 * save nblocks in a safe place and convert to pages.
857 	 */
858 
859 	sdp->swd_ose.ose_nblks = nblocks;
860 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
861 
862 	/*
863 	 * for block special files, we want to make sure that leave
864 	 * the disklabel and bootblocks alone, so we arrange to skip
865 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
866 	 * note that because of this the "size" can be less than the
867 	 * actual number of blocks on the device.
868 	 */
869 	if (vp->v_type == VBLK) {
870 		/* we use pages 1 to (size - 1) [inclusive] */
871 		size = npages - 1;
872 		addr = 1;
873 	} else {
874 		/* we use pages 0 to (size - 1) [inclusive] */
875 		size = npages;
876 		addr = 0;
877 	}
878 
879 	/*
880 	 * make sure we have enough blocks for a reasonable sized swap
881 	 * area.   we want at least one page.
882 	 */
883 
884 	if (size < 1) {
885 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
886 		error = EINVAL;
887 		goto bad;
888 	}
889 
890 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
891 
892 	/*
893 	 * now we need to allocate an extent to manage this swap device
894 	 */
895 
896 	sdp->swd_blist = blist_create(npages);
897 	/* mark all expect the `saved' region free. */
898 	blist_free(sdp->swd_blist, addr, size);
899 
900 	/*
901 	 * if the vnode we are swapping to is the root vnode
902 	 * (i.e. we are swapping to the miniroot) then we want
903 	 * to make sure we don't overwrite it.   do a statfs to
904 	 * find its size and skip over it.
905 	 */
906 	if (vp == rootvp) {
907 		struct mount *mp;
908 		struct statvfs *sp;
909 		int rootblocks, rootpages;
910 
911 		mp = rootvnode->v_mount;
912 		sp = &mp->mnt_stat;
913 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
914 		/*
915 		 * XXX: sp->f_blocks isn't the total number of
916 		 * blocks in the filesystem, it's the number of
917 		 * data blocks.  so, our rootblocks almost
918 		 * definitely underestimates the total size
919 		 * of the filesystem - how badly depends on the
920 		 * details of the filesystem type.  there isn't
921 		 * an obvious way to deal with this cleanly
922 		 * and perfectly, so for now we just pad our
923 		 * rootblocks estimate with an extra 5 percent.
924 		 */
925 		rootblocks += (rootblocks >> 5) +
926 			(rootblocks >> 6) +
927 			(rootblocks >> 7);
928 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
929 		if (rootpages > size)
930 			panic("swap_on: miniroot larger than swap?");
931 
932 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
933 			panic("swap_on: unable to preserve miniroot");
934 		}
935 
936 		size -= rootpages;
937 		printf("Preserved %d pages of miniroot ", rootpages);
938 		printf("leaving %d pages of swap\n", size);
939 	}
940 
941 	/*
942 	 * add a ref to vp to reflect usage as a swap device.
943 	 */
944 	vref(vp);
945 
946 	/*
947 	 * now add the new swapdev to the drum and enable.
948 	 */
949 	result = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP);
950 	if (result == 0)
951 		panic("swapdrum_add");
952 	/*
953 	 * If this is the first regular swap create the workqueue.
954 	 * => Protected by swap_syscall_lock.
955 	 */
956 	if (vp->v_type != VBLK) {
957 		if (sw_reg_count++ == 0) {
958 			KASSERT(sw_reg_workqueue == NULL);
959 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
960 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
961 				panic("swap_add: workqueue_create failed");
962 		}
963 	}
964 
965 	sdp->swd_drumoffset = (int)result;
966 	sdp->swd_drumsize = npages;
967 	sdp->swd_npages = size;
968 	mutex_enter(&uvm_swap_data_lock);
969 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
970 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
971 	uvmexp.swpages += size;
972 	uvmexp.swpgavail += size;
973 	mutex_exit(&uvm_swap_data_lock);
974 	return (0);
975 
976 	/*
977 	 * failure: clean up and return error.
978 	 */
979 
980 bad:
981 	if (sdp->swd_blist) {
982 		blist_destroy(sdp->swd_blist);
983 	}
984 	if (vp != rootvp) {
985 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
986 	}
987 	return (error);
988 }
989 
990 /*
991  * swap_off: stop swapping on swapdev
992  *
993  * => swap data should be locked, we will unlock.
994  */
995 static int
996 swap_off(struct lwp *l, struct swapdev *sdp)
997 {
998 	int npages = sdp->swd_npages;
999 	int error = 0;
1000 
1001 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
1002 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
1003 
1004 	/* disable the swap area being removed */
1005 	sdp->swd_flags &= ~SWF_ENABLE;
1006 	uvmexp.swpgavail -= npages;
1007 	mutex_exit(&uvm_swap_data_lock);
1008 
1009 	/*
1010 	 * the idea is to find all the pages that are paged out to this
1011 	 * device, and page them all in.  in uvm, swap-backed pageable
1012 	 * memory can take two forms: aobjs and anons.  call the
1013 	 * swapoff hook for each subsystem to bring in pages.
1014 	 */
1015 
1016 	if (uao_swap_off(sdp->swd_drumoffset,
1017 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1018 	    amap_swap_off(sdp->swd_drumoffset,
1019 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1020 		error = ENOMEM;
1021 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1022 		error = EBUSY;
1023 	}
1024 
1025 	if (error) {
1026 		mutex_enter(&uvm_swap_data_lock);
1027 		sdp->swd_flags |= SWF_ENABLE;
1028 		uvmexp.swpgavail += npages;
1029 		mutex_exit(&uvm_swap_data_lock);
1030 
1031 		return error;
1032 	}
1033 
1034 	/*
1035 	 * If this is the last regular swap destroy the workqueue.
1036 	 * => Protected by swap_syscall_lock.
1037 	 */
1038 	if (sdp->swd_vp->v_type != VBLK) {
1039 		KASSERT(sw_reg_count > 0);
1040 		KASSERT(sw_reg_workqueue != NULL);
1041 		if (--sw_reg_count == 0) {
1042 			workqueue_destroy(sw_reg_workqueue);
1043 			sw_reg_workqueue = NULL;
1044 		}
1045 	}
1046 
1047 	/*
1048 	 * done with the vnode.
1049 	 * drop our ref on the vnode before calling VOP_CLOSE()
1050 	 * so that spec_close() can tell if this is the last close.
1051 	 */
1052 	vrele(sdp->swd_vp);
1053 	if (sdp->swd_vp != rootvp) {
1054 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1055 	}
1056 
1057 	mutex_enter(&uvm_swap_data_lock);
1058 	uvmexp.swpages -= npages;
1059 	uvmexp.swpginuse -= sdp->swd_npgbad;
1060 
1061 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1062 		panic("swap_off: swapdev not in list");
1063 	swaplist_trim();
1064 	mutex_exit(&uvm_swap_data_lock);
1065 
1066 	/*
1067 	 * free all resources!
1068 	 */
1069 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1070 	blist_destroy(sdp->swd_blist);
1071 	bufq_free(sdp->swd_tab);
1072 	free(sdp, M_VMSWAP);
1073 	return (0);
1074 }
1075 
1076 /*
1077  * /dev/drum interface and i/o functions
1078  */
1079 
1080 /*
1081  * swstrategy: perform I/O on the drum
1082  *
1083  * => we must map the i/o request from the drum to the correct swapdev.
1084  */
1085 static void
1086 swstrategy(struct buf *bp)
1087 {
1088 	struct swapdev *sdp;
1089 	struct vnode *vp;
1090 	int pageno, bn;
1091 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1092 
1093 	/*
1094 	 * convert block number to swapdev.   note that swapdev can't
1095 	 * be yanked out from under us because we are holding resources
1096 	 * in it (i.e. the blocks we are doing I/O on).
1097 	 */
1098 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1099 	mutex_enter(&uvm_swap_data_lock);
1100 	sdp = swapdrum_getsdp(pageno);
1101 	mutex_exit(&uvm_swap_data_lock);
1102 	if (sdp == NULL) {
1103 		bp->b_error = EINVAL;
1104 		biodone(bp);
1105 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1106 		return;
1107 	}
1108 
1109 	/*
1110 	 * convert drum page number to block number on this swapdev.
1111 	 */
1112 
1113 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1114 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1115 
1116 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1117 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1118 		sdp->swd_drumoffset, bn, bp->b_bcount);
1119 
1120 	/*
1121 	 * for block devices we finish up here.
1122 	 * for regular files we have to do more work which we delegate
1123 	 * to sw_reg_strategy().
1124 	 */
1125 
1126 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1127 	switch (vp->v_type) {
1128 	default:
1129 		panic("swstrategy: vnode type 0x%x", vp->v_type);
1130 
1131 	case VBLK:
1132 
1133 		/*
1134 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1135 		 * on the swapdev (sdp).
1136 		 */
1137 		bp->b_blkno = bn;		/* swapdev block number */
1138 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1139 
1140 		/*
1141 		 * if we are doing a write, we have to redirect the i/o on
1142 		 * drum's v_numoutput counter to the swapdevs.
1143 		 */
1144 		if ((bp->b_flags & B_READ) == 0) {
1145 			mutex_enter(bp->b_objlock);
1146 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1147 			mutex_exit(bp->b_objlock);
1148 			mutex_enter(&vp->v_interlock);
1149 			vp->v_numoutput++;	/* put it on swapdev */
1150 			mutex_exit(&vp->v_interlock);
1151 		}
1152 
1153 		/*
1154 		 * finally plug in swapdev vnode and start I/O
1155 		 */
1156 		bp->b_vp = vp;
1157 		bp->b_objlock = &vp->v_interlock;
1158 		VOP_STRATEGY(vp, bp);
1159 		return;
1160 
1161 	case VREG:
1162 		/*
1163 		 * delegate to sw_reg_strategy function.
1164 		 */
1165 		sw_reg_strategy(sdp, bp, bn);
1166 		return;
1167 	}
1168 	/* NOTREACHED */
1169 }
1170 
1171 /*
1172  * swread: the read function for the drum (just a call to physio)
1173  */
1174 /*ARGSUSED*/
1175 static int
1176 swread(dev_t dev, struct uio *uio, int ioflag)
1177 {
1178 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1179 
1180 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1181 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1182 }
1183 
1184 /*
1185  * swwrite: the write function for the drum (just a call to physio)
1186  */
1187 /*ARGSUSED*/
1188 static int
1189 swwrite(dev_t dev, struct uio *uio, int ioflag)
1190 {
1191 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1192 
1193 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1194 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1195 }
1196 
1197 const struct bdevsw swap_bdevsw = {
1198 	nullopen, nullclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1199 };
1200 
1201 const struct cdevsw swap_cdevsw = {
1202 	nullopen, nullclose, swread, swwrite, noioctl,
1203 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1204 };
1205 
1206 /*
1207  * sw_reg_strategy: handle swap i/o to regular files
1208  */
1209 static void
1210 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1211 {
1212 	struct vnode	*vp;
1213 	struct vndxfer	*vnx;
1214 	daddr_t		nbn;
1215 	char 		*addr;
1216 	off_t		byteoff;
1217 	int		s, off, nra, error, sz, resid;
1218 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1219 
1220 	/*
1221 	 * allocate a vndxfer head for this transfer and point it to
1222 	 * our buffer.
1223 	 */
1224 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1225 	vnx->vx_flags = VX_BUSY;
1226 	vnx->vx_error = 0;
1227 	vnx->vx_pending = 0;
1228 	vnx->vx_bp = bp;
1229 	vnx->vx_sdp = sdp;
1230 
1231 	/*
1232 	 * setup for main loop where we read filesystem blocks into
1233 	 * our buffer.
1234 	 */
1235 	error = 0;
1236 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1237 	addr = bp->b_data;		/* current position in buffer */
1238 	byteoff = dbtob((uint64_t)bn);
1239 
1240 	for (resid = bp->b_resid; resid; resid -= sz) {
1241 		struct vndbuf	*nbp;
1242 
1243 		/*
1244 		 * translate byteoffset into block number.  return values:
1245 		 *   vp = vnode of underlying device
1246 		 *  nbn = new block number (on underlying vnode dev)
1247 		 *  nra = num blocks we can read-ahead (excludes requested
1248 		 *	block)
1249 		 */
1250 		nra = 0;
1251 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1252 				 	&vp, &nbn, &nra);
1253 
1254 		if (error == 0 && nbn == (daddr_t)-1) {
1255 			/*
1256 			 * this used to just set error, but that doesn't
1257 			 * do the right thing.  Instead, it causes random
1258 			 * memory errors.  The panic() should remain until
1259 			 * this condition doesn't destabilize the system.
1260 			 */
1261 #if 1
1262 			panic("sw_reg_strategy: swap to sparse file");
1263 #else
1264 			error = EIO;	/* failure */
1265 #endif
1266 		}
1267 
1268 		/*
1269 		 * punt if there was an error or a hole in the file.
1270 		 * we must wait for any i/o ops we have already started
1271 		 * to finish before returning.
1272 		 *
1273 		 * XXX we could deal with holes here but it would be
1274 		 * a hassle (in the write case).
1275 		 */
1276 		if (error) {
1277 			s = splbio();
1278 			vnx->vx_error = error;	/* pass error up */
1279 			goto out;
1280 		}
1281 
1282 		/*
1283 		 * compute the size ("sz") of this transfer (in bytes).
1284 		 */
1285 		off = byteoff % sdp->swd_bsize;
1286 		sz = (1 + nra) * sdp->swd_bsize - off;
1287 		if (sz > resid)
1288 			sz = resid;
1289 
1290 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1291 			    "vp %p/%p offset 0x%x/0x%x",
1292 			    sdp->swd_vp, vp, byteoff, nbn);
1293 
1294 		/*
1295 		 * now get a buf structure.   note that the vb_buf is
1296 		 * at the front of the nbp structure so that you can
1297 		 * cast pointers between the two structure easily.
1298 		 */
1299 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1300 		buf_init(&nbp->vb_buf);
1301 		nbp->vb_buf.b_flags    = bp->b_flags;
1302 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1303 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1304 		nbp->vb_buf.b_bcount   = sz;
1305 		nbp->vb_buf.b_bufsize  = sz;
1306 		nbp->vb_buf.b_error    = 0;
1307 		nbp->vb_buf.b_data     = addr;
1308 		nbp->vb_buf.b_lblkno   = 0;
1309 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1310 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1311 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1312 		nbp->vb_buf.b_vp       = vp;
1313 		nbp->vb_buf.b_objlock  = &vp->v_interlock;
1314 		if (vp->v_type == VBLK) {
1315 			nbp->vb_buf.b_dev = vp->v_rdev;
1316 		}
1317 
1318 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1319 
1320 		/*
1321 		 * Just sort by block number
1322 		 */
1323 		s = splbio();
1324 		if (vnx->vx_error != 0) {
1325 			buf_destroy(&nbp->vb_buf);
1326 			pool_put(&vndbuf_pool, nbp);
1327 			goto out;
1328 		}
1329 		vnx->vx_pending++;
1330 
1331 		/* sort it in and start I/O if we are not over our limit */
1332 		/* XXXAD locking */
1333 		BUFQ_PUT(sdp->swd_tab, &nbp->vb_buf);
1334 		sw_reg_start(sdp);
1335 		splx(s);
1336 
1337 		/*
1338 		 * advance to the next I/O
1339 		 */
1340 		byteoff += sz;
1341 		addr += sz;
1342 	}
1343 
1344 	s = splbio();
1345 
1346 out: /* Arrive here at splbio */
1347 	vnx->vx_flags &= ~VX_BUSY;
1348 	if (vnx->vx_pending == 0) {
1349 		error = vnx->vx_error;
1350 		pool_put(&vndxfer_pool, vnx);
1351 		bp->b_error = error;
1352 		biodone(bp);
1353 	}
1354 	splx(s);
1355 }
1356 
1357 /*
1358  * sw_reg_start: start an I/O request on the requested swapdev
1359  *
1360  * => reqs are sorted by b_rawblkno (above)
1361  */
1362 static void
1363 sw_reg_start(struct swapdev *sdp)
1364 {
1365 	struct buf	*bp;
1366 	struct vnode	*vp;
1367 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1368 
1369 	/* recursion control */
1370 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1371 		return;
1372 
1373 	sdp->swd_flags |= SWF_BUSY;
1374 
1375 	while (sdp->swd_active < sdp->swd_maxactive) {
1376 		bp = BUFQ_GET(sdp->swd_tab);
1377 		if (bp == NULL)
1378 			break;
1379 		sdp->swd_active++;
1380 
1381 		UVMHIST_LOG(pdhist,
1382 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1383 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1384 		vp = bp->b_vp;
1385 		KASSERT(bp->b_objlock == &vp->v_interlock);
1386 		if ((bp->b_flags & B_READ) == 0) {
1387 			mutex_enter(&vp->v_interlock);
1388 			vp->v_numoutput++;
1389 			mutex_exit(&vp->v_interlock);
1390 		}
1391 		VOP_STRATEGY(vp, bp);
1392 	}
1393 	sdp->swd_flags &= ~SWF_BUSY;
1394 }
1395 
1396 /*
1397  * sw_reg_biodone: one of our i/o's has completed
1398  */
1399 static void
1400 sw_reg_biodone(struct buf *bp)
1401 {
1402 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1403 }
1404 
1405 /*
1406  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1407  *
1408  * => note that we can recover the vndbuf struct by casting the buf ptr
1409  */
1410 static void
1411 sw_reg_iodone(struct work *wk, void *dummy)
1412 {
1413 	struct vndbuf *vbp = (void *)wk;
1414 	struct vndxfer *vnx = vbp->vb_xfer;
1415 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1416 	struct swapdev	*sdp = vnx->vx_sdp;
1417 	int s, resid, error;
1418 	KASSERT(&vbp->vb_buf.b_work == wk);
1419 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1420 
1421 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1422 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1423 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1424 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1425 
1426 	/*
1427 	 * protect vbp at splbio and update.
1428 	 */
1429 
1430 	s = splbio();
1431 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1432 	pbp->b_resid -= resid;
1433 	vnx->vx_pending--;
1434 
1435 	if (vbp->vb_buf.b_error != 0) {
1436 		/* pass error upward */
1437 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1438 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1439 		vnx->vx_error = error;
1440 	}
1441 
1442 	/*
1443 	 * kill vbp structure
1444 	 */
1445 	buf_destroy(&vbp->vb_buf);
1446 	pool_put(&vndbuf_pool, vbp);
1447 
1448 	/*
1449 	 * wrap up this transaction if it has run to completion or, in
1450 	 * case of an error, when all auxiliary buffers have returned.
1451 	 */
1452 	if (vnx->vx_error != 0) {
1453 		/* pass error upward */
1454 		error = vnx->vx_error;
1455 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1456 			pbp->b_error = error;
1457 			biodone(pbp);
1458 			pool_put(&vndxfer_pool, vnx);
1459 		}
1460 	} else if (pbp->b_resid == 0) {
1461 		KASSERT(vnx->vx_pending == 0);
1462 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1463 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1464 			    pbp, vnx->vx_error, 0, 0);
1465 			biodone(pbp);
1466 			pool_put(&vndxfer_pool, vnx);
1467 		}
1468 	}
1469 
1470 	/*
1471 	 * done!   start next swapdev I/O if one is pending
1472 	 */
1473 	sdp->swd_active--;
1474 	sw_reg_start(sdp);
1475 	splx(s);
1476 }
1477 
1478 
1479 /*
1480  * uvm_swap_alloc: allocate space on swap
1481  *
1482  * => allocation is done "round robin" down the priority list, as we
1483  *	allocate in a priority we "rotate" the circle queue.
1484  * => space can be freed with uvm_swap_free
1485  * => we return the page slot number in /dev/drum (0 == invalid slot)
1486  * => we lock uvm_swap_data_lock
1487  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1488  */
1489 int
1490 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1491 {
1492 	struct swapdev *sdp;
1493 	struct swappri *spp;
1494 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1495 
1496 	/*
1497 	 * no swap devices configured yet?   definite failure.
1498 	 */
1499 	if (uvmexp.nswapdev < 1)
1500 		return 0;
1501 
1502 	/*
1503 	 * lock data lock, convert slots into blocks, and enter loop
1504 	 */
1505 	mutex_enter(&uvm_swap_data_lock);
1506 
1507 ReTry:	/* XXXMRG */
1508 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1509 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1510 			uint64_t result;
1511 
1512 			/* if it's not enabled, then we can't swap from it */
1513 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1514 				continue;
1515 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1516 				continue;
1517 			result = blist_alloc(sdp->swd_blist, *nslots);
1518 			if (result == BLIST_NONE) {
1519 				continue;
1520 			}
1521 			KASSERT(result < sdp->swd_drumsize);
1522 
1523 			/*
1524 			 * successful allocation!  now rotate the circleq.
1525 			 */
1526 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1527 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1528 			sdp->swd_npginuse += *nslots;
1529 			uvmexp.swpginuse += *nslots;
1530 			mutex_exit(&uvm_swap_data_lock);
1531 			/* done!  return drum slot number */
1532 			UVMHIST_LOG(pdhist,
1533 			    "success!  returning %d slots starting at %d",
1534 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1535 			return (result + sdp->swd_drumoffset);
1536 		}
1537 	}
1538 
1539 	/* XXXMRG: BEGIN HACK */
1540 	if (*nslots > 1 && lessok) {
1541 		*nslots = 1;
1542 		/* XXXMRG: ugh!  blist should support this for us */
1543 		goto ReTry;
1544 	}
1545 	/* XXXMRG: END HACK */
1546 
1547 	mutex_exit(&uvm_swap_data_lock);
1548 	return 0;
1549 }
1550 
1551 bool
1552 uvm_swapisfull(void)
1553 {
1554 	bool rv;
1555 
1556 	mutex_enter(&uvm_swap_data_lock);
1557 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1558 	rv = (uvmexp.swpgonly >= uvmexp.swpgavail);
1559 	mutex_exit(&uvm_swap_data_lock);
1560 
1561 	return (rv);
1562 }
1563 
1564 /*
1565  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1566  *
1567  * => we lock uvm_swap_data_lock
1568  */
1569 void
1570 uvm_swap_markbad(int startslot, int nslots)
1571 {
1572 	struct swapdev *sdp;
1573 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1574 
1575 	mutex_enter(&uvm_swap_data_lock);
1576 	sdp = swapdrum_getsdp(startslot);
1577 	KASSERT(sdp != NULL);
1578 
1579 	/*
1580 	 * we just keep track of how many pages have been marked bad
1581 	 * in this device, to make everything add up in swap_off().
1582 	 * we assume here that the range of slots will all be within
1583 	 * one swap device.
1584 	 */
1585 
1586 	KASSERT(uvmexp.swpgonly >= nslots);
1587 	uvmexp.swpgonly -= nslots;
1588 	sdp->swd_npgbad += nslots;
1589 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1590 	mutex_exit(&uvm_swap_data_lock);
1591 }
1592 
1593 /*
1594  * uvm_swap_free: free swap slots
1595  *
1596  * => this can be all or part of an allocation made by uvm_swap_alloc
1597  * => we lock uvm_swap_data_lock
1598  */
1599 void
1600 uvm_swap_free(int startslot, int nslots)
1601 {
1602 	struct swapdev *sdp;
1603 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1604 
1605 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1606 	    startslot, 0, 0);
1607 
1608 	/*
1609 	 * ignore attempts to free the "bad" slot.
1610 	 */
1611 
1612 	if (startslot == SWSLOT_BAD) {
1613 		return;
1614 	}
1615 
1616 	/*
1617 	 * convert drum slot offset back to sdp, free the blocks
1618 	 * in the extent, and return.   must hold pri lock to do
1619 	 * lookup and access the extent.
1620 	 */
1621 
1622 	mutex_enter(&uvm_swap_data_lock);
1623 	sdp = swapdrum_getsdp(startslot);
1624 	KASSERT(uvmexp.nswapdev >= 1);
1625 	KASSERT(sdp != NULL);
1626 	KASSERT(sdp->swd_npginuse >= nslots);
1627 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1628 	sdp->swd_npginuse -= nslots;
1629 	uvmexp.swpginuse -= nslots;
1630 	mutex_exit(&uvm_swap_data_lock);
1631 }
1632 
1633 /*
1634  * uvm_swap_put: put any number of pages into a contig place on swap
1635  *
1636  * => can be sync or async
1637  */
1638 
1639 int
1640 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1641 {
1642 	int error;
1643 
1644 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1645 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1646 	return error;
1647 }
1648 
1649 /*
1650  * uvm_swap_get: get a single page from swap
1651  *
1652  * => usually a sync op (from fault)
1653  */
1654 
1655 int
1656 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1657 {
1658 	int error;
1659 
1660 	uvmexp.nswget++;
1661 	KASSERT(flags & PGO_SYNCIO);
1662 	if (swslot == SWSLOT_BAD) {
1663 		return EIO;
1664 	}
1665 
1666 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1667 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1668 	if (error == 0) {
1669 
1670 		/*
1671 		 * this page is no longer only in swap.
1672 		 */
1673 
1674 		mutex_enter(&uvm_swap_data_lock);
1675 		KASSERT(uvmexp.swpgonly > 0);
1676 		uvmexp.swpgonly--;
1677 		mutex_exit(&uvm_swap_data_lock);
1678 	}
1679 	return error;
1680 }
1681 
1682 /*
1683  * uvm_swap_io: do an i/o operation to swap
1684  */
1685 
1686 static int
1687 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1688 {
1689 	daddr_t startblk;
1690 	struct	buf *bp;
1691 	vaddr_t kva;
1692 	int	error, mapinflags;
1693 	bool write, async;
1694 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1695 
1696 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1697 	    startslot, npages, flags, 0);
1698 
1699 	write = (flags & B_READ) == 0;
1700 	async = (flags & B_ASYNC) != 0;
1701 
1702 	/*
1703 	 * allocate a buf for the i/o.
1704 	 */
1705 
1706 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1707 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1708 	if (bp == NULL) {
1709 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1710 		return ENOMEM;
1711 	}
1712 
1713 	/*
1714 	 * convert starting drum slot to block number
1715 	 */
1716 
1717 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1718 
1719 	/*
1720 	 * first, map the pages into the kernel.
1721 	 */
1722 
1723 	mapinflags = !write ?
1724 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1725 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1726 	kva = uvm_pagermapin(pps, npages, mapinflags);
1727 
1728 	/*
1729 	 * fill in the bp/sbp.   we currently route our i/o through
1730 	 * /dev/drum's vnode [swapdev_vp].
1731 	 */
1732 
1733 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1734 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1735 	bp->b_proc = &proc0;	/* XXX */
1736 	bp->b_vnbufs.le_next = NOLIST;
1737 	bp->b_data = (void *)kva;
1738 	bp->b_blkno = startblk;
1739 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1740 
1741 	/*
1742 	 * bump v_numoutput (counter of number of active outputs).
1743 	 */
1744 
1745 	if (write) {
1746 		mutex_enter(&swapdev_vp->v_interlock);
1747 		swapdev_vp->v_numoutput++;
1748 		mutex_exit(&swapdev_vp->v_interlock);
1749 	}
1750 
1751 	/*
1752 	 * for async ops we must set up the iodone handler.
1753 	 */
1754 
1755 	if (async) {
1756 		bp->b_iodone = uvm_aio_biodone;
1757 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1758 		if (curlwp == uvm.pagedaemon_lwp)
1759 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1760 		else
1761 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1762 	} else {
1763 		bp->b_iodone = NULL;
1764 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1765 	}
1766 	UVMHIST_LOG(pdhist,
1767 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1768 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1769 
1770 	/*
1771 	 * now we start the I/O, and if async, return.
1772 	 */
1773 
1774 	VOP_STRATEGY(swapdev_vp, bp);
1775 	if (async)
1776 		return 0;
1777 
1778 	/*
1779 	 * must be sync i/o.   wait for it to finish
1780 	 */
1781 
1782 	error = biowait(bp);
1783 
1784 	/*
1785 	 * kill the pager mapping
1786 	 */
1787 
1788 	uvm_pagermapout(kva, npages);
1789 
1790 	/*
1791 	 * now dispose of the buf and we're done.
1792 	 */
1793 
1794 	if (write) {
1795 		mutex_enter(&swapdev_vp->v_interlock);
1796 		vwakeup(bp);
1797 		mutex_exit(&swapdev_vp->v_interlock);
1798 	}
1799 	putiobuf(bp);
1800 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1801 
1802 	return (error);
1803 }
1804