xref: /netbsd-src/sys/uvm/uvm_swap.c (revision 9fbd88883c38d0c0fbfcbe66d76fe6b0fab3f9de)
1 /*	$NetBSD: uvm_swap.c,v 1.58 2001/12/16 04:51:34 enami 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  * 3. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
23  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
24  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
25  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * from: NetBSD: vm_swap.c,v 1.52 1997/12/02 13:47:37 pk Exp
31  * from: Id: uvm_swap.c,v 1.1.2.42 1998/02/02 20:38:06 chuck Exp
32  */
33 
34 #include <sys/cdefs.h>
35 __KERNEL_RCSID(0, "$NetBSD: uvm_swap.c,v 1.58 2001/12/16 04:51:34 enami Exp $");
36 
37 #include "fs_nfs.h"
38 #include "opt_uvmhist.h"
39 #include "opt_compat_netbsd.h"
40 #include "opt_ddb.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/buf.h>
45 #include <sys/conf.h>
46 #include <sys/proc.h>
47 #include <sys/namei.h>
48 #include <sys/disklabel.h>
49 #include <sys/errno.h>
50 #include <sys/kernel.h>
51 #include <sys/malloc.h>
52 #include <sys/vnode.h>
53 #include <sys/file.h>
54 #include <sys/extent.h>
55 #include <sys/mount.h>
56 #include <sys/pool.h>
57 #include <sys/syscallargs.h>
58 #include <sys/swap.h>
59 
60 #include <uvm/uvm.h>
61 
62 #include <miscfs/specfs/specdev.h>
63 
64 /*
65  * uvm_swap.c: manage configuration and i/o to swap space.
66  */
67 
68 /*
69  * swap space is managed in the following way:
70  *
71  * each swap partition or file is described by a "swapdev" structure.
72  * each "swapdev" structure contains a "swapent" structure which contains
73  * information that is passed up to the user (via system calls).
74  *
75  * each swap partition is assigned a "priority" (int) which controls
76  * swap parition usage.
77  *
78  * the system maintains a global data structure describing all swap
79  * partitions/files.   there is a sorted LIST of "swappri" structures
80  * which describe "swapdev"'s at that priority.   this LIST is headed
81  * by the "swap_priority" global var.    each "swappri" contains a
82  * CIRCLEQ of "swapdev" structures at that priority.
83  *
84  * locking:
85  *  - swap_syscall_lock (sleep lock): this lock serializes the swapctl
86  *    system call and prevents the swap priority list from changing
87  *    while we are in the middle of a system call (e.g. SWAP_STATS).
88  *  - uvm.swap_data_lock (simple_lock): this lock protects all swap data
89  *    structures including the priority list, the swapdev structures,
90  *    and the swapmap extent.
91  *
92  * each swap device has the following info:
93  *  - swap device in use (could be disabled, preventing future use)
94  *  - swap enabled (allows new allocations on swap)
95  *  - map info in /dev/drum
96  *  - vnode pointer
97  * for swap files only:
98  *  - block size
99  *  - max byte count in buffer
100  *  - buffer
101  *
102  * userland controls and configures swap with the swapctl(2) system call.
103  * the sys_swapctl performs the following operations:
104  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
105  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
106  *	(passed in via "arg") of a size passed in via "misc" ... we load
107  *	the current swap config into the array.
108  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
109  *	priority in "misc", start swapping on it.
110  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
111  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
112  *	"misc")
113  */
114 
115 /*
116  * swapdev: describes a single swap partition/file
117  *
118  * note the following should be true:
119  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
120  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
121  */
122 struct swapdev {
123 	struct oswapent swd_ose;
124 #define	swd_dev		swd_ose.ose_dev		/* device id */
125 #define	swd_flags	swd_ose.ose_flags	/* flags:inuse/enable/fake */
126 #define	swd_priority	swd_ose.ose_priority	/* our priority */
127 	/* also: swd_ose.ose_nblks, swd_ose.ose_inuse */
128 	char			*swd_path;	/* saved pathname of device */
129 	int			swd_pathlen;	/* length of pathname */
130 	int			swd_npages;	/* #pages we can use */
131 	int			swd_npginuse;	/* #pages in use */
132 	int			swd_npgbad;	/* #pages bad */
133 	int			swd_drumoffset;	/* page0 offset in drum */
134 	int			swd_drumsize;	/* #pages in drum */
135 	struct extent		*swd_ex;	/* extent for this swapdev */
136 	char			swd_exname[12];	/* name of extent above */
137 	struct vnode		*swd_vp;	/* backing vnode */
138 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
139 
140 	int			swd_bsize;	/* blocksize (bytes) */
141 	int			swd_maxactive;	/* max active i/o reqs */
142 	struct buf_queue	swd_tab;	/* buffer list */
143 	int			swd_active;	/* number of active buffers */
144 };
145 
146 /*
147  * swap device priority entry; the list is kept sorted on `spi_priority'.
148  */
149 struct swappri {
150 	int			spi_priority;     /* priority */
151 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
152 	/* circleq of swapdevs at this priority */
153 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
154 };
155 
156 /*
157  * The following two structures are used to keep track of data transfers
158  * on swap devices associated with regular files.
159  * NOTE: this code is more or less a copy of vnd.c; we use the same
160  * structure names here to ease porting..
161  */
162 struct vndxfer {
163 	struct buf	*vx_bp;		/* Pointer to parent buffer */
164 	struct swapdev	*vx_sdp;
165 	int		vx_error;
166 	int		vx_pending;	/* # of pending aux buffers */
167 	int		vx_flags;
168 #define VX_BUSY		1
169 #define VX_DEAD		2
170 };
171 
172 struct vndbuf {
173 	struct buf	vb_buf;
174 	struct vndxfer	*vb_xfer;
175 };
176 
177 
178 /*
179  * We keep a of pool vndbuf's and vndxfer structures.
180  */
181 static struct pool vndxfer_pool;
182 static struct pool vndbuf_pool;
183 
184 #define	getvndxfer(vnx)	do {						\
185 	int s = splbio();						\
186 	vnx = pool_get(&vndxfer_pool, PR_MALLOCOK|PR_WAITOK);		\
187 	splx(s);							\
188 } while (0)
189 
190 #define putvndxfer(vnx) {						\
191 	pool_put(&vndxfer_pool, (void *)(vnx));				\
192 }
193 
194 #define	getvndbuf(vbp)	do {						\
195 	int s = splbio();						\
196 	vbp = pool_get(&vndbuf_pool, PR_MALLOCOK|PR_WAITOK);		\
197 	splx(s);							\
198 } while (0)
199 
200 #define putvndbuf(vbp) {						\
201 	pool_put(&vndbuf_pool, (void *)(vbp));				\
202 }
203 
204 /* /dev/drum */
205 bdev_decl(sw);
206 cdev_decl(sw);
207 
208 /*
209  * local variables
210  */
211 static struct extent *swapmap;		/* controls the mapping of /dev/drum */
212 
213 /* list of all active swap devices [by priority] */
214 LIST_HEAD(swap_priority, swappri);
215 static struct swap_priority swap_priority;
216 
217 /* locks */
218 struct lock swap_syscall_lock;
219 
220 /*
221  * prototypes
222  */
223 static struct swapdev	*swapdrum_getsdp __P((int));
224 
225 static struct swapdev	*swaplist_find __P((struct vnode *, int));
226 static void		 swaplist_insert __P((struct swapdev *,
227 					     struct swappri *, int));
228 static void		 swaplist_trim __P((void));
229 
230 static int swap_on __P((struct proc *, struct swapdev *));
231 static int swap_off __P((struct proc *, struct swapdev *));
232 
233 static void sw_reg_strategy __P((struct swapdev *, struct buf *, int));
234 static void sw_reg_iodone __P((struct buf *));
235 static void sw_reg_start __P((struct swapdev *));
236 
237 static int uvm_swap_io __P((struct vm_page **, int, int, int));
238 
239 /*
240  * uvm_swap_init: init the swap system data structures and locks
241  *
242  * => called at boot time from init_main.c after the filesystems
243  *	are brought up (which happens after uvm_init())
244  */
245 void
246 uvm_swap_init()
247 {
248 	UVMHIST_FUNC("uvm_swap_init");
249 
250 	UVMHIST_CALLED(pdhist);
251 	/*
252 	 * first, init the swap list, its counter, and its lock.
253 	 * then get a handle on the vnode for /dev/drum by using
254 	 * the its dev_t number ("swapdev", from MD conf.c).
255 	 */
256 
257 	LIST_INIT(&swap_priority);
258 	uvmexp.nswapdev = 0;
259 	lockinit(&swap_syscall_lock, PVM, "swapsys", 0, 0);
260 	simple_lock_init(&uvm.swap_data_lock);
261 
262 	if (bdevvp(swapdev, &swapdev_vp))
263 		panic("uvm_swap_init: can't get vnode for swap device");
264 
265 	/*
266 	 * create swap block resource map to map /dev/drum.   the range
267 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
268 	 * that block 0 is reserved (used to indicate an allocation
269 	 * failure, or no allocation).
270 	 */
271 	swapmap = extent_create("swapmap", 1, INT_MAX,
272 				M_VMSWAP, 0, 0, EX_NOWAIT);
273 	if (swapmap == 0)
274 		panic("uvm_swap_init: extent_create failed");
275 
276 	/*
277 	 * allocate pools for structures used for swapping to files.
278 	 */
279 
280 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0,
281 	    "swp vnx", 0, NULL, NULL, 0);
282 
283 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0,
284 	    "swp vnd", 0, NULL, NULL, 0);
285 
286 	/*
287 	 * done!
288 	 */
289 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
290 }
291 
292 /*
293  * swaplist functions: functions that operate on the list of swap
294  * devices on the system.
295  */
296 
297 /*
298  * swaplist_insert: insert swap device "sdp" into the global list
299  *
300  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
301  * => caller must provide a newly malloc'd swappri structure (we will
302  *	FREE it if we don't need it... this it to prevent malloc blocking
303  *	here while adding swap)
304  */
305 static void
306 swaplist_insert(sdp, newspp, priority)
307 	struct swapdev *sdp;
308 	struct swappri *newspp;
309 	int priority;
310 {
311 	struct swappri *spp, *pspp;
312 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
313 
314 	/*
315 	 * find entry at or after which to insert the new device.
316 	 */
317 	pspp = NULL;
318 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
319 		if (priority <= spp->spi_priority)
320 			break;
321 		pspp = spp;
322 	}
323 
324 	/*
325 	 * new priority?
326 	 */
327 	if (spp == NULL || spp->spi_priority != priority) {
328 		spp = newspp;  /* use newspp! */
329 		UVMHIST_LOG(pdhist, "created new swappri = %d",
330 			    priority, 0, 0, 0);
331 
332 		spp->spi_priority = priority;
333 		CIRCLEQ_INIT(&spp->spi_swapdev);
334 
335 		if (pspp)
336 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
337 		else
338 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
339 	} else {
340 	  	/* we don't need a new priority structure, free it */
341 		FREE(newspp, M_VMSWAP);
342 	}
343 
344 	/*
345 	 * priority found (or created).   now insert on the priority's
346 	 * circleq list and bump the total number of swapdevs.
347 	 */
348 	sdp->swd_priority = priority;
349 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
350 	uvmexp.nswapdev++;
351 }
352 
353 /*
354  * swaplist_find: find and optionally remove a swap device from the
355  *	global list.
356  *
357  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
358  * => we return the swapdev we found (and removed)
359  */
360 static struct swapdev *
361 swaplist_find(vp, remove)
362 	struct vnode *vp;
363 	boolean_t remove;
364 {
365 	struct swapdev *sdp;
366 	struct swappri *spp;
367 
368 	/*
369 	 * search the lists for the requested vp
370 	 */
371 
372 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
373 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
374 			if (sdp->swd_vp == vp) {
375 				if (remove) {
376 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
377 					    sdp, swd_next);
378 					uvmexp.nswapdev--;
379 				}
380 				return(sdp);
381 			}
382 		}
383 	}
384 	return (NULL);
385 }
386 
387 
388 /*
389  * swaplist_trim: scan priority list for empty priority entries and kill
390  *	them.
391  *
392  * => caller must hold both swap_syscall_lock and uvm.swap_data_lock
393  */
394 static void
395 swaplist_trim()
396 {
397 	struct swappri *spp, *nextspp;
398 
399 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
400 		nextspp = LIST_NEXT(spp, spi_swappri);
401 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
402 		    (void *)&spp->spi_swapdev)
403 			continue;
404 		LIST_REMOVE(spp, spi_swappri);
405 		free(spp, M_VMSWAP);
406 	}
407 }
408 
409 /*
410  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
411  *	to the "swapdev" that maps that section of the drum.
412  *
413  * => each swapdev takes one big contig chunk of the drum
414  * => caller must hold uvm.swap_data_lock
415  */
416 static struct swapdev *
417 swapdrum_getsdp(pgno)
418 	int pgno;
419 {
420 	struct swapdev *sdp;
421 	struct swappri *spp;
422 
423 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
424 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
425 			if (sdp->swd_flags & SWF_FAKE)
426 				continue;
427 			if (pgno >= sdp->swd_drumoffset &&
428 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
429 				return sdp;
430 			}
431 		}
432 	}
433 	return NULL;
434 }
435 
436 
437 /*
438  * sys_swapctl: main entry point for swapctl(2) system call
439  * 	[with two helper functions: swap_on and swap_off]
440  */
441 int
442 sys_swapctl(p, v, retval)
443 	struct proc *p;
444 	void *v;
445 	register_t *retval;
446 {
447 	struct sys_swapctl_args /* {
448 		syscallarg(int) cmd;
449 		syscallarg(void *) arg;
450 		syscallarg(int) misc;
451 	} */ *uap = (struct sys_swapctl_args *)v;
452 	struct vnode *vp;
453 	struct nameidata nd;
454 	struct swappri *spp;
455 	struct swapdev *sdp;
456 	struct swapent *sep;
457 	char	userpath[PATH_MAX + 1];
458 	size_t	len;
459 	int	count, error, misc;
460 	int	priority;
461 	UVMHIST_FUNC("sys_swapctl"); UVMHIST_CALLED(pdhist);
462 
463 	misc = SCARG(uap, misc);
464 
465 	/*
466 	 * ensure serialized syscall access by grabbing the swap_syscall_lock
467 	 */
468 	lockmgr(&swap_syscall_lock, LK_EXCLUSIVE, NULL);
469 
470 	/*
471 	 * we handle the non-priv NSWAP and STATS request first.
472 	 *
473 	 * SWAP_NSWAP: return number of config'd swap devices
474 	 * [can also be obtained with uvmexp sysctl]
475 	 */
476 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
477 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
478 		    0, 0, 0);
479 		*retval = uvmexp.nswapdev;
480 		error = 0;
481 		goto out;
482 	}
483 
484 	/*
485 	 * SWAP_STATS: get stats on current # of configured swap devs
486 	 *
487 	 * note that the swap_priority list can't change as long
488 	 * as we are holding the swap_syscall_lock.  we don't want
489 	 * to grab the uvm.swap_data_lock because we may fault&sleep during
490 	 * copyout() and we don't want to be holding that lock then!
491 	 */
492 	if (SCARG(uap, cmd) == SWAP_STATS
493 #if defined(COMPAT_13)
494 	    || SCARG(uap, cmd) == SWAP_OSTATS
495 #endif
496 	    ) {
497 		sep = (struct swapent *)SCARG(uap, arg);
498 		count = 0;
499 
500 		LIST_FOREACH(spp, &swap_priority, spi_swappri) {
501 			for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
502 			     sdp != (void *)&spp->spi_swapdev && misc-- > 0;
503 			     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
504 			  	/*
505 				 * backwards compatibility for system call.
506 				 * note that we use 'struct oswapent' as an
507 				 * overlay into both 'struct swapdev' and
508 				 * the userland 'struct swapent', as we
509 				 * want to retain backwards compatibility
510 				 * with NetBSD 1.3.
511 				 */
512 				sdp->swd_ose.ose_inuse =
513 				    btodb((u_int64_t)sdp->swd_npginuse <<
514 				    PAGE_SHIFT);
515 				error = copyout(&sdp->swd_ose, sep,
516 						sizeof(struct oswapent));
517 
518 				/* now copy out the path if necessary */
519 #if defined(COMPAT_13)
520 				if (error == 0 && SCARG(uap, cmd) == SWAP_STATS)
521 #else
522 				if (error == 0)
523 #endif
524 					error = copyout(sdp->swd_path,
525 					    &sep->se_path, sdp->swd_pathlen);
526 
527 				if (error)
528 					goto out;
529 				count++;
530 #if defined(COMPAT_13)
531 				if (SCARG(uap, cmd) == SWAP_OSTATS)
532 					sep = (struct swapent *)
533 					    ((struct oswapent *)sep + 1);
534 				else
535 #endif
536 					sep++;
537 			}
538 		}
539 
540 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
541 
542 		*retval = count;
543 		error = 0;
544 		goto out;
545 	}
546 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
547 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
548 
549 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
550 		goto out;
551 	}
552 
553 	/*
554 	 * all other requests require superuser privs.   verify.
555 	 */
556 	if ((error = suser(p->p_ucred, &p->p_acflag)))
557 		goto out;
558 
559 	/*
560 	 * at this point we expect a path name in arg.   we will
561 	 * use namei() to gain a vnode reference (vref), and lock
562 	 * the vnode (VOP_LOCK).
563 	 *
564 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
565 	 * miniroot)
566 	 */
567 	if (SCARG(uap, arg) == NULL) {
568 		vp = rootvp;		/* miniroot */
569 		if (vget(vp, LK_EXCLUSIVE)) {
570 			error = EBUSY;
571 			goto out;
572 		}
573 		if (SCARG(uap, cmd) == SWAP_ON &&
574 		    copystr("miniroot", userpath, sizeof userpath, &len))
575 			panic("swapctl: miniroot copy failed");
576 	} else {
577 		int	space;
578 		char	*where;
579 
580 		if (SCARG(uap, cmd) == SWAP_ON) {
581 			if ((error = copyinstr(SCARG(uap, arg), userpath,
582 			    sizeof userpath, &len)))
583 				goto out;
584 			space = UIO_SYSSPACE;
585 			where = userpath;
586 		} else {
587 			space = UIO_USERSPACE;
588 			where = (char *)SCARG(uap, arg);
589 		}
590 		NDINIT(&nd, LOOKUP, FOLLOW|LOCKLEAF, space, where, p);
591 		if ((error = namei(&nd)))
592 			goto out;
593 		vp = nd.ni_vp;
594 	}
595 	/* note: "vp" is referenced and locked */
596 
597 	error = 0;		/* assume no error */
598 	switch(SCARG(uap, cmd)) {
599 
600 	case SWAP_DUMPDEV:
601 		if (vp->v_type != VBLK) {
602 			error = ENOTBLK;
603 			break;
604 		}
605 		dumpdev = vp->v_rdev;
606 		break;
607 
608 	case SWAP_CTL:
609 		/*
610 		 * get new priority, remove old entry (if any) and then
611 		 * reinsert it in the correct place.  finally, prune out
612 		 * any empty priority structures.
613 		 */
614 		priority = SCARG(uap, misc);
615 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
616 		simple_lock(&uvm.swap_data_lock);
617 		if ((sdp = swaplist_find(vp, 1)) == NULL) {
618 			error = ENOENT;
619 		} else {
620 			swaplist_insert(sdp, spp, priority);
621 			swaplist_trim();
622 		}
623 		simple_unlock(&uvm.swap_data_lock);
624 		if (error)
625 			free(spp, M_VMSWAP);
626 		break;
627 
628 	case SWAP_ON:
629 
630 		/*
631 		 * check for duplicates.   if none found, then insert a
632 		 * dummy entry on the list to prevent someone else from
633 		 * trying to enable this device while we are working on
634 		 * it.
635 		 */
636 
637 		priority = SCARG(uap, misc);
638 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
639 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
640 		simple_lock(&uvm.swap_data_lock);
641 		if (swaplist_find(vp, 0) != NULL) {
642 			error = EBUSY;
643 			simple_unlock(&uvm.swap_data_lock);
644 			free(sdp, M_VMSWAP);
645 			free(spp, M_VMSWAP);
646 			break;
647 		}
648 		memset(sdp, 0, sizeof(*sdp));
649 		sdp->swd_flags = SWF_FAKE;	/* placeholder only */
650 		sdp->swd_vp = vp;
651 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
652 		BUFQ_INIT(&sdp->swd_tab);
653 
654 		swaplist_insert(sdp, spp, priority);
655 		simple_unlock(&uvm.swap_data_lock);
656 
657 		sdp->swd_pathlen = len;
658 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
659 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
660 			panic("swapctl: copystr");
661 
662 		/*
663 		 * we've now got a FAKE placeholder in the swap list.
664 		 * now attempt to enable swap on it.  if we fail, undo
665 		 * what we've done and kill the fake entry we just inserted.
666 		 * if swap_on is a success, it will clear the SWF_FAKE flag
667 		 */
668 
669 		if ((error = swap_on(p, sdp)) != 0) {
670 			simple_lock(&uvm.swap_data_lock);
671 			(void) swaplist_find(vp, 1);  /* kill fake entry */
672 			swaplist_trim();
673 			simple_unlock(&uvm.swap_data_lock);
674 			free(sdp->swd_path, M_VMSWAP);
675 			free(sdp, M_VMSWAP);
676 			break;
677 		}
678 		break;
679 
680 	case SWAP_OFF:
681 		simple_lock(&uvm.swap_data_lock);
682 		if ((sdp = swaplist_find(vp, 0)) == NULL) {
683 			simple_unlock(&uvm.swap_data_lock);
684 			error = ENXIO;
685 			break;
686 		}
687 
688 		/*
689 		 * If a device isn't in use or enabled, we
690 		 * can't stop swapping from it (again).
691 		 */
692 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
693 			simple_unlock(&uvm.swap_data_lock);
694 			error = EBUSY;
695 			break;
696 		}
697 
698 		/*
699 		 * do the real work.
700 		 */
701 		error = swap_off(p, sdp);
702 		break;
703 
704 	default:
705 		error = EINVAL;
706 	}
707 
708 	/*
709 	 * done!  release the ref gained by namei() and unlock.
710 	 */
711 	vput(vp);
712 
713 out:
714 	lockmgr(&swap_syscall_lock, LK_RELEASE, NULL);
715 
716 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
717 	return (error);
718 }
719 
720 /*
721  * swap_on: attempt to enable a swapdev for swapping.   note that the
722  *	swapdev is already on the global list, but disabled (marked
723  *	SWF_FAKE).
724  *
725  * => we avoid the start of the disk (to protect disk labels)
726  * => we also avoid the miniroot, if we are swapping to root.
727  * => caller should leave uvm.swap_data_lock unlocked, we may lock it
728  *	if needed.
729  */
730 static int
731 swap_on(p, sdp)
732 	struct proc *p;
733 	struct swapdev *sdp;
734 {
735 	static int count = 0;	/* static */
736 	struct vnode *vp;
737 	int error, npages, nblocks, size;
738 	long addr;
739 	u_long result;
740 	struct vattr va;
741 #ifdef NFS
742 	extern int (**nfsv2_vnodeop_p) __P((void *));
743 #endif /* NFS */
744 	dev_t dev;
745 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
746 
747 	/*
748 	 * we want to enable swapping on sdp.   the swd_vp contains
749 	 * the vnode we want (locked and ref'd), and the swd_dev
750 	 * contains the dev_t of the file, if it a block device.
751 	 */
752 
753 	vp = sdp->swd_vp;
754 	dev = sdp->swd_dev;
755 
756 	/*
757 	 * open the swap file (mostly useful for block device files to
758 	 * let device driver know what is up).
759 	 *
760 	 * we skip the open/close for root on swap because the root
761 	 * has already been opened when root was mounted (mountroot).
762 	 */
763 	if (vp != rootvp) {
764 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, p->p_ucred, p)))
765 			return (error);
766 	}
767 
768 	/* XXX this only works for block devices */
769 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
770 
771 	/*
772 	 * we now need to determine the size of the swap area.   for
773 	 * block specials we can call the d_psize function.
774 	 * for normal files, we must stat [get attrs].
775 	 *
776 	 * we put the result in nblks.
777 	 * for normal files, we also want the filesystem block size
778 	 * (which we get with statfs).
779 	 */
780 	switch (vp->v_type) {
781 	case VBLK:
782 		if (bdevsw[major(dev)].d_psize == 0 ||
783 		    (nblocks = (*bdevsw[major(dev)].d_psize)(dev)) == -1) {
784 			error = ENXIO;
785 			goto bad;
786 		}
787 		break;
788 
789 	case VREG:
790 		if ((error = VOP_GETATTR(vp, &va, p->p_ucred, p)))
791 			goto bad;
792 		nblocks = (int)btodb(va.va_size);
793 		if ((error =
794 		     VFS_STATFS(vp->v_mount, &vp->v_mount->mnt_stat, p)) != 0)
795 			goto bad;
796 
797 		sdp->swd_bsize = vp->v_mount->mnt_stat.f_iosize;
798 		/*
799 		 * limit the max # of outstanding I/O requests we issue
800 		 * at any one time.   take it easy on NFS servers.
801 		 */
802 #ifdef NFS
803 		if (vp->v_op == nfsv2_vnodeop_p)
804 			sdp->swd_maxactive = 2; /* XXX */
805 		else
806 #endif /* NFS */
807 			sdp->swd_maxactive = 8; /* XXX */
808 		break;
809 
810 	default:
811 		error = ENXIO;
812 		goto bad;
813 	}
814 
815 	/*
816 	 * save nblocks in a safe place and convert to pages.
817 	 */
818 
819 	sdp->swd_ose.ose_nblks = nblocks;
820 	npages = dbtob((u_int64_t)nblocks) >> PAGE_SHIFT;
821 
822 	/*
823 	 * for block special files, we want to make sure that leave
824 	 * the disklabel and bootblocks alone, so we arrange to skip
825 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
826 	 * note that because of this the "size" can be less than the
827 	 * actual number of blocks on the device.
828 	 */
829 	if (vp->v_type == VBLK) {
830 		/* we use pages 1 to (size - 1) [inclusive] */
831 		size = npages - 1;
832 		addr = 1;
833 	} else {
834 		/* we use pages 0 to (size - 1) [inclusive] */
835 		size = npages;
836 		addr = 0;
837 	}
838 
839 	/*
840 	 * make sure we have enough blocks for a reasonable sized swap
841 	 * area.   we want at least one page.
842 	 */
843 
844 	if (size < 1) {
845 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
846 		error = EINVAL;
847 		goto bad;
848 	}
849 
850 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
851 
852 	/*
853 	 * now we need to allocate an extent to manage this swap device
854 	 */
855 	snprintf(sdp->swd_exname, sizeof(sdp->swd_exname), "swap0x%04x",
856 	    count++);
857 
858 	/* note that extent_create's 3rd arg is inclusive, thus "- 1" */
859 	sdp->swd_ex = extent_create(sdp->swd_exname, 0, npages - 1, M_VMSWAP,
860 				    0, 0, EX_WAITOK);
861 	/* allocate the `saved' region from the extent so it won't be used */
862 	if (addr) {
863 		if (extent_alloc_region(sdp->swd_ex, 0, addr, EX_WAITOK))
864 			panic("disklabel region");
865 	}
866 
867 	/*
868 	 * if the vnode we are swapping to is the root vnode
869 	 * (i.e. we are swapping to the miniroot) then we want
870 	 * to make sure we don't overwrite it.   do a statfs to
871 	 * find its size and skip over it.
872 	 */
873 	if (vp == rootvp) {
874 		struct mount *mp;
875 		struct statfs *sp;
876 		int rootblocks, rootpages;
877 
878 		mp = rootvnode->v_mount;
879 		sp = &mp->mnt_stat;
880 		rootblocks = sp->f_blocks * btodb(sp->f_bsize);
881 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
882 		if (rootpages > size)
883 			panic("swap_on: miniroot larger than swap?");
884 
885 		if (extent_alloc_region(sdp->swd_ex, addr,
886 					rootpages, EX_WAITOK))
887 			panic("swap_on: unable to preserve miniroot");
888 
889 		size -= rootpages;
890 		printf("Preserved %d pages of miniroot ", rootpages);
891 		printf("leaving %d pages of swap\n", size);
892 	}
893 
894   	/*
895 	 * try to add anons to reflect the new swap space.
896 	 */
897 
898 	error = uvm_anon_add(size);
899 	if (error) {
900 		goto bad;
901 	}
902 
903 	/*
904 	 * add a ref to vp to reflect usage as a swap device.
905 	 */
906 	vref(vp);
907 
908 	/*
909 	 * now add the new swapdev to the drum and enable.
910 	 */
911 	if (extent_alloc(swapmap, npages, EX_NOALIGN, EX_NOBOUNDARY,
912 	    EX_WAITOK, &result))
913 		panic("swapdrum_add");
914 
915 	sdp->swd_drumoffset = (int)result;
916 	sdp->swd_drumsize = npages;
917 	sdp->swd_npages = size;
918 	simple_lock(&uvm.swap_data_lock);
919 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
920 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
921 	uvmexp.swpages += size;
922 	simple_unlock(&uvm.swap_data_lock);
923 	return (0);
924 
925 	/*
926 	 * failure: clean up and return error.
927 	 */
928 
929 bad:
930 	if (sdp->swd_ex) {
931 		extent_destroy(sdp->swd_ex);
932 	}
933 	if (vp != rootvp) {
934 		(void)VOP_CLOSE(vp, FREAD|FWRITE, p->p_ucred, p);
935 	}
936 	return (error);
937 }
938 
939 /*
940  * swap_off: stop swapping on swapdev
941  *
942  * => swap data should be locked, we will unlock.
943  */
944 static int
945 swap_off(p, sdp)
946 	struct proc *p;
947 	struct swapdev *sdp;
948 {
949 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
950 	UVMHIST_LOG(pdhist, "  dev=%x", sdp->swd_dev,0,0,0);
951 
952 	/* disable the swap area being removed */
953 	sdp->swd_flags &= ~SWF_ENABLE;
954 	simple_unlock(&uvm.swap_data_lock);
955 
956 	/*
957 	 * the idea is to find all the pages that are paged out to this
958 	 * device, and page them all in.  in uvm, swap-backed pageable
959 	 * memory can take two forms: aobjs and anons.  call the
960 	 * swapoff hook for each subsystem to bring in pages.
961 	 */
962 
963 	if (uao_swap_off(sdp->swd_drumoffset,
964 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
965 	    anon_swap_off(sdp->swd_drumoffset,
966 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
967 
968 		simple_lock(&uvm.swap_data_lock);
969 		sdp->swd_flags |= SWF_ENABLE;
970 		simple_unlock(&uvm.swap_data_lock);
971 		return ENOMEM;
972 	}
973 	KASSERT(sdp->swd_npginuse == sdp->swd_npgbad);
974 
975 	/*
976 	 * done with the vnode.
977 	 * drop our ref on the vnode before calling VOP_CLOSE()
978 	 * so that spec_close() can tell if this is the last close.
979 	 */
980 	vrele(sdp->swd_vp);
981 	if (sdp->swd_vp != rootvp) {
982 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, p->p_ucred, p);
983 	}
984 
985 	/* remove anons from the system */
986 	uvm_anon_remove(sdp->swd_npages);
987 
988 	simple_lock(&uvm.swap_data_lock);
989 	uvmexp.swpages -= sdp->swd_npages;
990 
991 	if (swaplist_find(sdp->swd_vp, 1) == NULL)
992 		panic("swap_off: swapdev not in list\n");
993 	swaplist_trim();
994 	simple_unlock(&uvm.swap_data_lock);
995 
996 	/*
997 	 * free all resources!
998 	 */
999 	extent_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize,
1000 		    EX_WAITOK);
1001 	extent_destroy(sdp->swd_ex);
1002 	free(sdp, M_VMSWAP);
1003 	return (0);
1004 }
1005 
1006 /*
1007  * /dev/drum interface and i/o functions
1008  */
1009 
1010 /*
1011  * swread: the read function for the drum (just a call to physio)
1012  */
1013 /*ARGSUSED*/
1014 int
1015 swread(dev, uio, ioflag)
1016 	dev_t dev;
1017 	struct uio *uio;
1018 	int ioflag;
1019 {
1020 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1021 
1022 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1023 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1024 }
1025 
1026 /*
1027  * swwrite: the write function for the drum (just a call to physio)
1028  */
1029 /*ARGSUSED*/
1030 int
1031 swwrite(dev, uio, ioflag)
1032 	dev_t dev;
1033 	struct uio *uio;
1034 	int ioflag;
1035 {
1036 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1037 
1038 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1039 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1040 }
1041 
1042 /*
1043  * swstrategy: perform I/O on the drum
1044  *
1045  * => we must map the i/o request from the drum to the correct swapdev.
1046  */
1047 void
1048 swstrategy(bp)
1049 	struct buf *bp;
1050 {
1051 	struct swapdev *sdp;
1052 	struct vnode *vp;
1053 	int s, pageno, bn;
1054 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1055 
1056 	/*
1057 	 * convert block number to swapdev.   note that swapdev can't
1058 	 * be yanked out from under us because we are holding resources
1059 	 * in it (i.e. the blocks we are doing I/O on).
1060 	 */
1061 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1062 	simple_lock(&uvm.swap_data_lock);
1063 	sdp = swapdrum_getsdp(pageno);
1064 	simple_unlock(&uvm.swap_data_lock);
1065 	if (sdp == NULL) {
1066 		bp->b_error = EINVAL;
1067 		bp->b_flags |= B_ERROR;
1068 		biodone(bp);
1069 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1070 		return;
1071 	}
1072 
1073 	/*
1074 	 * convert drum page number to block number on this swapdev.
1075 	 */
1076 
1077 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1078 	bn = btodb((u_int64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1079 
1080 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1081 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1082 		sdp->swd_drumoffset, bn, bp->b_bcount);
1083 
1084 	/*
1085 	 * for block devices we finish up here.
1086 	 * for regular files we have to do more work which we delegate
1087 	 * to sw_reg_strategy().
1088 	 */
1089 
1090 	switch (sdp->swd_vp->v_type) {
1091 	default:
1092 		panic("swstrategy: vnode type 0x%x", sdp->swd_vp->v_type);
1093 
1094 	case VBLK:
1095 
1096 		/*
1097 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1098 		 * on the swapdev (sdp).
1099 		 */
1100 		s = splbio();
1101 		bp->b_blkno = bn;		/* swapdev block number */
1102 		vp = sdp->swd_vp;		/* swapdev vnode pointer */
1103 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1104 
1105 		/*
1106 		 * if we are doing a write, we have to redirect the i/o on
1107 		 * drum's v_numoutput counter to the swapdevs.
1108 		 */
1109 		if ((bp->b_flags & B_READ) == 0) {
1110 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1111 			vp->v_numoutput++;	/* put it on swapdev */
1112 		}
1113 
1114 		/*
1115 		 * finally plug in swapdev vnode and start I/O
1116 		 */
1117 		bp->b_vp = vp;
1118 		splx(s);
1119 		VOP_STRATEGY(bp);
1120 		return;
1121 
1122 	case VREG:
1123 		/*
1124 		 * delegate to sw_reg_strategy function.
1125 		 */
1126 		sw_reg_strategy(sdp, bp, bn);
1127 		return;
1128 	}
1129 	/* NOTREACHED */
1130 }
1131 
1132 /*
1133  * sw_reg_strategy: handle swap i/o to regular files
1134  */
1135 static void
1136 sw_reg_strategy(sdp, bp, bn)
1137 	struct swapdev	*sdp;
1138 	struct buf	*bp;
1139 	int		bn;
1140 {
1141 	struct vnode	*vp;
1142 	struct vndxfer	*vnx;
1143 	daddr_t		nbn;
1144 	caddr_t		addr;
1145 	off_t		byteoff;
1146 	int		s, off, nra, error, sz, resid;
1147 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1148 
1149 	/*
1150 	 * allocate a vndxfer head for this transfer and point it to
1151 	 * our buffer.
1152 	 */
1153 	getvndxfer(vnx);
1154 	vnx->vx_flags = VX_BUSY;
1155 	vnx->vx_error = 0;
1156 	vnx->vx_pending = 0;
1157 	vnx->vx_bp = bp;
1158 	vnx->vx_sdp = sdp;
1159 
1160 	/*
1161 	 * setup for main loop where we read filesystem blocks into
1162 	 * our buffer.
1163 	 */
1164 	error = 0;
1165 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1166 	addr = bp->b_data;		/* current position in buffer */
1167 	byteoff = dbtob((u_int64_t)bn);
1168 
1169 	for (resid = bp->b_resid; resid; resid -= sz) {
1170 		struct vndbuf	*nbp;
1171 
1172 		/*
1173 		 * translate byteoffset into block number.  return values:
1174 		 *   vp = vnode of underlying device
1175 		 *  nbn = new block number (on underlying vnode dev)
1176 		 *  nra = num blocks we can read-ahead (excludes requested
1177 		 *	block)
1178 		 */
1179 		nra = 0;
1180 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1181 				 	&vp, &nbn, &nra);
1182 
1183 		if (error == 0 && nbn == (daddr_t)-1) {
1184 			/*
1185 			 * this used to just set error, but that doesn't
1186 			 * do the right thing.  Instead, it causes random
1187 			 * memory errors.  The panic() should remain until
1188 			 * this condition doesn't destabilize the system.
1189 			 */
1190 #if 1
1191 			panic("sw_reg_strategy: swap to sparse file");
1192 #else
1193 			error = EIO;	/* failure */
1194 #endif
1195 		}
1196 
1197 		/*
1198 		 * punt if there was an error or a hole in the file.
1199 		 * we must wait for any i/o ops we have already started
1200 		 * to finish before returning.
1201 		 *
1202 		 * XXX we could deal with holes here but it would be
1203 		 * a hassle (in the write case).
1204 		 */
1205 		if (error) {
1206 			s = splbio();
1207 			vnx->vx_error = error;	/* pass error up */
1208 			goto out;
1209 		}
1210 
1211 		/*
1212 		 * compute the size ("sz") of this transfer (in bytes).
1213 		 */
1214 		off = byteoff % sdp->swd_bsize;
1215 		sz = (1 + nra) * sdp->swd_bsize - off;
1216 		if (sz > resid)
1217 			sz = resid;
1218 
1219 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1220 			    "vp %p/%p offset 0x%x/0x%x",
1221 			    sdp->swd_vp, vp, byteoff, nbn);
1222 
1223 		/*
1224 		 * now get a buf structure.   note that the vb_buf is
1225 		 * at the front of the nbp structure so that you can
1226 		 * cast pointers between the two structure easily.
1227 		 */
1228 		getvndbuf(nbp);
1229 		nbp->vb_buf.b_flags    = bp->b_flags | B_CALL;
1230 		nbp->vb_buf.b_bcount   = sz;
1231 		nbp->vb_buf.b_bufsize  = sz;
1232 		nbp->vb_buf.b_error    = 0;
1233 		nbp->vb_buf.b_data     = addr;
1234 		nbp->vb_buf.b_lblkno   = 0;
1235 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1236 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1237 		nbp->vb_buf.b_iodone   = sw_reg_iodone;
1238 		nbp->vb_buf.b_vp       = vp;
1239 		if (vp->v_type == VBLK) {
1240 			nbp->vb_buf.b_dev = vp->v_rdev;
1241 		}
1242 		LIST_INIT(&nbp->vb_buf.b_dep);
1243 
1244 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1245 
1246 		/*
1247 		 * Just sort by block number
1248 		 */
1249 		s = splbio();
1250 		if (vnx->vx_error != 0) {
1251 			putvndbuf(nbp);
1252 			goto out;
1253 		}
1254 		vnx->vx_pending++;
1255 
1256 		/* sort it in and start I/O if we are not over our limit */
1257 		disksort_blkno(&sdp->swd_tab, &nbp->vb_buf);
1258 		sw_reg_start(sdp);
1259 		splx(s);
1260 
1261 		/*
1262 		 * advance to the next I/O
1263 		 */
1264 		byteoff += sz;
1265 		addr += sz;
1266 	}
1267 
1268 	s = splbio();
1269 
1270 out: /* Arrive here at splbio */
1271 	vnx->vx_flags &= ~VX_BUSY;
1272 	if (vnx->vx_pending == 0) {
1273 		if (vnx->vx_error != 0) {
1274 			bp->b_error = vnx->vx_error;
1275 			bp->b_flags |= B_ERROR;
1276 		}
1277 		putvndxfer(vnx);
1278 		biodone(bp);
1279 	}
1280 	splx(s);
1281 }
1282 
1283 /*
1284  * sw_reg_start: start an I/O request on the requested swapdev
1285  *
1286  * => reqs are sorted by disksort (above)
1287  */
1288 static void
1289 sw_reg_start(sdp)
1290 	struct swapdev	*sdp;
1291 {
1292 	struct buf	*bp;
1293 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1294 
1295 	/* recursion control */
1296 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1297 		return;
1298 
1299 	sdp->swd_flags |= SWF_BUSY;
1300 
1301 	while (sdp->swd_active < sdp->swd_maxactive) {
1302 		bp = BUFQ_FIRST(&sdp->swd_tab);
1303 		if (bp == NULL)
1304 			break;
1305 		BUFQ_REMOVE(&sdp->swd_tab, bp);
1306 		sdp->swd_active++;
1307 
1308 		UVMHIST_LOG(pdhist,
1309 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1310 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1311 		if ((bp->b_flags & B_READ) == 0)
1312 			bp->b_vp->v_numoutput++;
1313 
1314 		VOP_STRATEGY(bp);
1315 	}
1316 	sdp->swd_flags &= ~SWF_BUSY;
1317 }
1318 
1319 /*
1320  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1321  *
1322  * => note that we can recover the vndbuf struct by casting the buf ptr
1323  */
1324 static void
1325 sw_reg_iodone(bp)
1326 	struct buf *bp;
1327 {
1328 	struct vndbuf *vbp = (struct vndbuf *) bp;
1329 	struct vndxfer *vnx = vbp->vb_xfer;
1330 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1331 	struct swapdev	*sdp = vnx->vx_sdp;
1332 	int		s, resid;
1333 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1334 
1335 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1336 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1337 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1338 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1339 
1340 	/*
1341 	 * protect vbp at splbio and update.
1342 	 */
1343 
1344 	s = splbio();
1345 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1346 	pbp->b_resid -= resid;
1347 	vnx->vx_pending--;
1348 
1349 	if (vbp->vb_buf.b_error) {
1350 		UVMHIST_LOG(pdhist, "  got error=%d !",
1351 		    vbp->vb_buf.b_error, 0, 0, 0);
1352 
1353 		/* pass error upward */
1354 		vnx->vx_error = vbp->vb_buf.b_error;
1355 	}
1356 
1357 	/*
1358 	 * kill vbp structure
1359 	 */
1360 	putvndbuf(vbp);
1361 
1362 	/*
1363 	 * wrap up this transaction if it has run to completion or, in
1364 	 * case of an error, when all auxiliary buffers have returned.
1365 	 */
1366 	if (vnx->vx_error != 0) {
1367 		/* pass error upward */
1368 		pbp->b_flags |= B_ERROR;
1369 		pbp->b_error = vnx->vx_error;
1370 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1371 			putvndxfer(vnx);
1372 			biodone(pbp);
1373 		}
1374 	} else if (pbp->b_resid == 0) {
1375 		KASSERT(vnx->vx_pending == 0);
1376 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1377 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1378 			    pbp, vnx->vx_error, 0, 0);
1379 			putvndxfer(vnx);
1380 			biodone(pbp);
1381 		}
1382 	}
1383 
1384 	/*
1385 	 * done!   start next swapdev I/O if one is pending
1386 	 */
1387 	sdp->swd_active--;
1388 	sw_reg_start(sdp);
1389 	splx(s);
1390 }
1391 
1392 
1393 /*
1394  * uvm_swap_alloc: allocate space on swap
1395  *
1396  * => allocation is done "round robin" down the priority list, as we
1397  *	allocate in a priority we "rotate" the circle queue.
1398  * => space can be freed with uvm_swap_free
1399  * => we return the page slot number in /dev/drum (0 == invalid slot)
1400  * => we lock uvm.swap_data_lock
1401  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1402  */
1403 int
1404 uvm_swap_alloc(nslots, lessok)
1405 	int *nslots;	/* IN/OUT */
1406 	boolean_t lessok;
1407 {
1408 	struct swapdev *sdp;
1409 	struct swappri *spp;
1410 	u_long	result;
1411 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1412 
1413 	/*
1414 	 * no swap devices configured yet?   definite failure.
1415 	 */
1416 	if (uvmexp.nswapdev < 1)
1417 		return 0;
1418 
1419 	/*
1420 	 * lock data lock, convert slots into blocks, and enter loop
1421 	 */
1422 	simple_lock(&uvm.swap_data_lock);
1423 
1424 ReTry:	/* XXXMRG */
1425 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1426 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1427 			/* if it's not enabled, then we can't swap from it */
1428 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1429 				continue;
1430 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1431 				continue;
1432 			if (extent_alloc(sdp->swd_ex, *nslots, EX_NOALIGN,
1433 					 EX_NOBOUNDARY, EX_MALLOCOK|EX_NOWAIT,
1434 					 &result) != 0) {
1435 				continue;
1436 			}
1437 
1438 			/*
1439 			 * successful allocation!  now rotate the circleq.
1440 			 */
1441 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1442 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1443 			sdp->swd_npginuse += *nslots;
1444 			uvmexp.swpginuse += *nslots;
1445 			simple_unlock(&uvm.swap_data_lock);
1446 			/* done!  return drum slot number */
1447 			UVMHIST_LOG(pdhist,
1448 			    "success!  returning %d slots starting at %d",
1449 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1450 			return (result + sdp->swd_drumoffset);
1451 		}
1452 	}
1453 
1454 	/* XXXMRG: BEGIN HACK */
1455 	if (*nslots > 1 && lessok) {
1456 		*nslots = 1;
1457 		goto ReTry;	/* XXXMRG: ugh!  extent should support this for us */
1458 	}
1459 	/* XXXMRG: END HACK */
1460 
1461 	simple_unlock(&uvm.swap_data_lock);
1462 	return 0;
1463 }
1464 
1465 /*
1466  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1467  *
1468  * => we lock uvm.swap_data_lock
1469  */
1470 void
1471 uvm_swap_markbad(startslot, nslots)
1472 	int startslot;
1473 	int nslots;
1474 {
1475 	struct swapdev *sdp;
1476 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1477 
1478 	simple_lock(&uvm.swap_data_lock);
1479 	sdp = swapdrum_getsdp(startslot);
1480 
1481 	/*
1482 	 * we just keep track of how many pages have been marked bad
1483 	 * in this device, to make everything add up in swap_off().
1484 	 * we assume here that the range of slots will all be within
1485 	 * one swap device.
1486 	 */
1487 
1488 	sdp->swd_npgbad += nslots;
1489 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1490 	simple_unlock(&uvm.swap_data_lock);
1491 }
1492 
1493 /*
1494  * uvm_swap_free: free swap slots
1495  *
1496  * => this can be all or part of an allocation made by uvm_swap_alloc
1497  * => we lock uvm.swap_data_lock
1498  */
1499 void
1500 uvm_swap_free(startslot, nslots)
1501 	int startslot;
1502 	int nslots;
1503 {
1504 	struct swapdev *sdp;
1505 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1506 
1507 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1508 	    startslot, 0, 0);
1509 
1510 	/*
1511 	 * ignore attempts to free the "bad" slot.
1512 	 */
1513 
1514 	if (startslot == SWSLOT_BAD) {
1515 		return;
1516 	}
1517 
1518 	/*
1519 	 * convert drum slot offset back to sdp, free the blocks
1520 	 * in the extent, and return.   must hold pri lock to do
1521 	 * lookup and access the extent.
1522 	 */
1523 
1524 	simple_lock(&uvm.swap_data_lock);
1525 	sdp = swapdrum_getsdp(startslot);
1526 	KASSERT(uvmexp.nswapdev >= 1);
1527 	KASSERT(sdp != NULL);
1528 	KASSERT(sdp->swd_npginuse >= nslots);
1529 	if (extent_free(sdp->swd_ex, startslot - sdp->swd_drumoffset, nslots,
1530 			EX_MALLOCOK|EX_NOWAIT) != 0) {
1531 		printf("warning: resource shortage: %d pages of swap lost\n",
1532 			nslots);
1533 	}
1534 	sdp->swd_npginuse -= nslots;
1535 	uvmexp.swpginuse -= nslots;
1536 	simple_unlock(&uvm.swap_data_lock);
1537 }
1538 
1539 /*
1540  * uvm_swap_put: put any number of pages into a contig place on swap
1541  *
1542  * => can be sync or async
1543  */
1544 
1545 int
1546 uvm_swap_put(swslot, ppsp, npages, flags)
1547 	int swslot;
1548 	struct vm_page **ppsp;
1549 	int npages;
1550 	int flags;
1551 {
1552 	int error;
1553 
1554 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1555 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1556 	return error;
1557 }
1558 
1559 /*
1560  * uvm_swap_get: get a single page from swap
1561  *
1562  * => usually a sync op (from fault)
1563  */
1564 
1565 int
1566 uvm_swap_get(page, swslot, flags)
1567 	struct vm_page *page;
1568 	int swslot, flags;
1569 {
1570 	int error;
1571 
1572 	uvmexp.nswget++;
1573 	KASSERT(flags & PGO_SYNCIO);
1574 	if (swslot == SWSLOT_BAD) {
1575 		return EIO;
1576 	}
1577 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1578 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1579 	if (error == 0) {
1580 
1581 		/*
1582 		 * this page is no longer only in swap.
1583 		 */
1584 
1585 		simple_lock(&uvm.swap_data_lock);
1586 		KASSERT(uvmexp.swpgonly > 0);
1587 		uvmexp.swpgonly--;
1588 		simple_unlock(&uvm.swap_data_lock);
1589 	}
1590 	return error;
1591 }
1592 
1593 /*
1594  * uvm_swap_io: do an i/o operation to swap
1595  */
1596 
1597 static int
1598 uvm_swap_io(pps, startslot, npages, flags)
1599 	struct vm_page **pps;
1600 	int startslot, npages, flags;
1601 {
1602 	daddr_t startblk;
1603 	struct	buf *bp;
1604 	vaddr_t kva;
1605 	int	error, s, mapinflags;
1606 	boolean_t write, async;
1607 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1608 
1609 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1610 	    startslot, npages, flags, 0);
1611 
1612 	write = (flags & B_READ) == 0;
1613 	async = (flags & B_ASYNC) != 0;
1614 
1615 	/*
1616 	 * convert starting drum slot to block number
1617 	 */
1618 
1619 	startblk = btodb((u_int64_t)startslot << PAGE_SHIFT);
1620 
1621 	/*
1622 	 * first, map the pages into the kernel.
1623 	 */
1624 
1625 	mapinflags = !write ?
1626 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1627 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1628 	kva = uvm_pagermapin(pps, npages, mapinflags);
1629 
1630 	/*
1631 	 * now allocate a buf for the i/o.
1632 	 */
1633 
1634 	s = splbio();
1635 	bp = pool_get(&bufpool, PR_WAITOK);
1636 	splx(s);
1637 
1638 	/*
1639 	 * fill in the bp/sbp.   we currently route our i/o through
1640 	 * /dev/drum's vnode [swapdev_vp].
1641 	 */
1642 
1643 	bp->b_flags = B_BUSY | B_NOCACHE | (flags & (B_READ|B_ASYNC));
1644 	bp->b_proc = &proc0;	/* XXX */
1645 	bp->b_vnbufs.le_next = NOLIST;
1646 	bp->b_data = (caddr_t)kva;
1647 	bp->b_blkno = startblk;
1648 	bp->b_vp = swapdev_vp;
1649 	bp->b_dev = swapdev_vp->v_rdev;
1650 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1651 	LIST_INIT(&bp->b_dep);
1652 
1653 	/*
1654 	 * bump v_numoutput (counter of number of active outputs).
1655 	 */
1656 
1657 	if (write) {
1658 		s = splbio();
1659 		swapdev_vp->v_numoutput++;
1660 		splx(s);
1661 	}
1662 
1663 	/*
1664 	 * for async ops we must set up the iodone handler.
1665 	 */
1666 
1667 	if (async) {
1668 		bp->b_flags |= B_CALL;
1669 		bp->b_iodone = uvm_aio_biodone;
1670 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1671 	}
1672 	UVMHIST_LOG(pdhist,
1673 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1674 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1675 
1676 	/*
1677 	 * now we start the I/O, and if async, return.
1678 	 */
1679 
1680 	VOP_STRATEGY(bp);
1681 	if (async)
1682 		return 0;
1683 
1684 	/*
1685 	 * must be sync i/o.   wait for it to finish
1686 	 */
1687 
1688 	error = biowait(bp);
1689 
1690 	/*
1691 	 * kill the pager mapping
1692 	 */
1693 
1694 	uvm_pagermapout(kva, npages);
1695 
1696 	/*
1697 	 * now dispose of the buf and we're done.
1698 	 */
1699 
1700 	s = splbio();
1701 	if (write)
1702 		vwakeup(bp);
1703 	pool_put(&bufpool, bp);
1704 	splx(s);
1705 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1706 	return (error);
1707 }
1708