xref: /netbsd-src/sys/uvm/uvm_swap.c (revision b1c86f5f087524e68db12794ee9c3e3da1ab17a0)
1 /*	$NetBSD: uvm_swap.c,v 1.152 2010/07/09 08:13:33 hannken Exp $	*/
2 
3 /*
4  * Copyright (c) 1995, 1996, 1997, 2009 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 1.152 2010/07/09 08:13:33 hannken Exp $");
34 
35 #include "opt_uvmhist.h"
36 #include "opt_compat_netbsd.h"
37 #include "opt_ddb.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/buf.h>
42 #include <sys/bufq.h>
43 #include <sys/conf.h>
44 #include <sys/proc.h>
45 #include <sys/namei.h>
46 #include <sys/disklabel.h>
47 #include <sys/errno.h>
48 #include <sys/kernel.h>
49 #include <sys/malloc.h>
50 #include <sys/vnode.h>
51 #include <sys/file.h>
52 #include <sys/vmem.h>
53 #include <sys/blist.h>
54 #include <sys/mount.h>
55 #include <sys/pool.h>
56 #include <sys/syscallargs.h>
57 #include <sys/swap.h>
58 #include <sys/kauth.h>
59 #include <sys/sysctl.h>
60 #include <sys/workqueue.h>
61 
62 #include <uvm/uvm.h>
63 
64 #include <miscfs/specfs/specdev.h>
65 
66 /*
67  * uvm_swap.c: manage configuration and i/o to swap space.
68  */
69 
70 /*
71  * swap space is managed in the following way:
72  *
73  * each swap partition or file is described by a "swapdev" structure.
74  * each "swapdev" structure contains a "swapent" structure which contains
75  * information that is passed up to the user (via system calls).
76  *
77  * each swap partition is assigned a "priority" (int) which controls
78  * swap parition usage.
79  *
80  * the system maintains a global data structure describing all swap
81  * partitions/files.   there is a sorted LIST of "swappri" structures
82  * which describe "swapdev"'s at that priority.   this LIST is headed
83  * by the "swap_priority" global var.    each "swappri" contains a
84  * CIRCLEQ of "swapdev" structures at that priority.
85  *
86  * locking:
87  *  - swap_syscall_lock (krwlock_t): this lock serializes the swapctl
88  *    system call and prevents the swap priority list from changing
89  *    while we are in the middle of a system call (e.g. SWAP_STATS).
90  *  - uvm_swap_data_lock (kmutex_t): this lock protects all swap data
91  *    structures including the priority list, the swapdev structures,
92  *    and the swapmap arena.
93  *
94  * each swap device has the following info:
95  *  - swap device in use (could be disabled, preventing future use)
96  *  - swap enabled (allows new allocations on swap)
97  *  - map info in /dev/drum
98  *  - vnode pointer
99  * for swap files only:
100  *  - block size
101  *  - max byte count in buffer
102  *  - buffer
103  *
104  * userland controls and configures swap with the swapctl(2) system call.
105  * the sys_swapctl performs the following operations:
106  *  [1] SWAP_NSWAP: returns the number of swap devices currently configured
107  *  [2] SWAP_STATS: given a pointer to an array of swapent structures
108  *	(passed in via "arg") of a size passed in via "misc" ... we load
109  *	the current swap config into the array. The actual work is done
110  *	in the uvm_swap_stats(9) function.
111  *  [3] SWAP_ON: given a pathname in arg (could be device or file) and a
112  *	priority in "misc", start swapping on it.
113  *  [4] SWAP_OFF: as SWAP_ON, but stops swapping to a device
114  *  [5] SWAP_CTL: changes the priority of a swap device (new priority in
115  *	"misc")
116  */
117 
118 /*
119  * swapdev: describes a single swap partition/file
120  *
121  * note the following should be true:
122  * swd_inuse <= swd_nblks  [number of blocks in use is <= total blocks]
123  * swd_nblks <= swd_mapsize [because mapsize includes miniroot+disklabel]
124  */
125 struct swapdev {
126 	dev_t			swd_dev;	/* device id */
127 	int			swd_flags;	/* flags:inuse/enable/fake */
128 	int			swd_priority;	/* our priority */
129 	int			swd_nblks;	/* blocks in this device */
130 	char			*swd_path;	/* saved pathname of device */
131 	int			swd_pathlen;	/* length of pathname */
132 	int			swd_npages;	/* #pages we can use */
133 	int			swd_npginuse;	/* #pages in use */
134 	int			swd_npgbad;	/* #pages bad */
135 	int			swd_drumoffset;	/* page0 offset in drum */
136 	int			swd_drumsize;	/* #pages in drum */
137 	blist_t			swd_blist;	/* blist for this swapdev */
138 	struct vnode		*swd_vp;	/* backing vnode */
139 	CIRCLEQ_ENTRY(swapdev)	swd_next;	/* priority circleq */
140 
141 	int			swd_bsize;	/* blocksize (bytes) */
142 	int			swd_maxactive;	/* max active i/o reqs */
143 	struct bufq_state	*swd_tab;	/* buffer list */
144 	int			swd_active;	/* number of active buffers */
145 };
146 
147 /*
148  * swap device priority entry; the list is kept sorted on `spi_priority'.
149  */
150 struct swappri {
151 	int			spi_priority;     /* priority */
152 	CIRCLEQ_HEAD(spi_swapdev, swapdev)	spi_swapdev;
153 	/* circleq of swapdevs at this priority */
154 	LIST_ENTRY(swappri)	spi_swappri;      /* global list of pri's */
155 };
156 
157 /*
158  * The following two structures are used to keep track of data transfers
159  * on swap devices associated with regular files.
160  * NOTE: this code is more or less a copy of vnd.c; we use the same
161  * structure names here to ease porting..
162  */
163 struct vndxfer {
164 	struct buf	*vx_bp;		/* Pointer to parent buffer */
165 	struct swapdev	*vx_sdp;
166 	int		vx_error;
167 	int		vx_pending;	/* # of pending aux buffers */
168 	int		vx_flags;
169 #define VX_BUSY		1
170 #define VX_DEAD		2
171 };
172 
173 struct vndbuf {
174 	struct buf	vb_buf;
175 	struct vndxfer	*vb_xfer;
176 };
177 
178 /*
179  * NetBSD 1.3 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
180  * dev_t and has no se_path[] member.
181  */
182 struct swapent13 {
183 	int32_t	se13_dev;		/* device id */
184 	int	se13_flags;		/* flags */
185 	int	se13_nblks;		/* total blocks */
186 	int	se13_inuse;		/* blocks in use */
187 	int	se13_priority;		/* priority of this device */
188 };
189 
190 /*
191  * NetBSD 5.0 swapctl(SWAP_STATS, ...) swapent structure; uses 32 bit
192  * dev_t.
193  */
194 struct swapent50 {
195 	int32_t	se50_dev;		/* device id */
196 	int	se50_flags;		/* flags */
197 	int	se50_nblks;		/* total blocks */
198 	int	se50_inuse;		/* blocks in use */
199 	int	se50_priority;		/* priority of this device */
200 	char	se50_path[PATH_MAX+1];	/* path name */
201 };
202 
203 /*
204  * We keep a of pool vndbuf's and vndxfer structures.
205  */
206 static struct pool vndxfer_pool, vndbuf_pool;
207 
208 /*
209  * local variables
210  */
211 MALLOC_DEFINE(M_VMSWAP, "VM swap", "VM swap structures");
212 static vmem_t *swapmap;	/* controls the mapping of /dev/drum */
213 
214 /* list of all active swap devices [by priority] */
215 LIST_HEAD(swap_priority, swappri);
216 static struct swap_priority swap_priority;
217 
218 /* locks */
219 static krwlock_t swap_syscall_lock;
220 
221 /* workqueue and use counter for swap to regular files */
222 static int sw_reg_count = 0;
223 static struct workqueue *sw_reg_workqueue;
224 
225 /* tuneables */
226 u_int uvm_swapisfull_factor = 99;
227 
228 /*
229  * prototypes
230  */
231 static struct swapdev	*swapdrum_getsdp(int);
232 
233 static struct swapdev	*swaplist_find(struct vnode *, bool);
234 static void		 swaplist_insert(struct swapdev *,
235 					 struct swappri *, int);
236 static void		 swaplist_trim(void);
237 
238 static int swap_on(struct lwp *, struct swapdev *);
239 static int swap_off(struct lwp *, struct swapdev *);
240 
241 static void uvm_swap_stats_locked(int, struct swapent *, int, register_t *);
242 
243 static void sw_reg_strategy(struct swapdev *, struct buf *, int);
244 static void sw_reg_biodone(struct buf *);
245 static void sw_reg_iodone(struct work *wk, void *dummy);
246 static void sw_reg_start(struct swapdev *);
247 
248 static int uvm_swap_io(struct vm_page **, int, int, int);
249 
250 /*
251  * uvm_swap_init: init the swap system data structures and locks
252  *
253  * => called at boot time from init_main.c after the filesystems
254  *	are brought up (which happens after uvm_init())
255  */
256 void
257 uvm_swap_init(void)
258 {
259 	UVMHIST_FUNC("uvm_swap_init");
260 
261 	UVMHIST_CALLED(pdhist);
262 	/*
263 	 * first, init the swap list, its counter, and its lock.
264 	 * then get a handle on the vnode for /dev/drum by using
265 	 * the its dev_t number ("swapdev", from MD conf.c).
266 	 */
267 
268 	LIST_INIT(&swap_priority);
269 	uvmexp.nswapdev = 0;
270 	rw_init(&swap_syscall_lock);
271 	mutex_init(&uvm_swap_data_lock, MUTEX_DEFAULT, IPL_NONE);
272 
273 	if (bdevvp(swapdev, &swapdev_vp))
274 		panic("%s: can't get vnode for swap device", __func__);
275 	if (vn_lock(swapdev_vp, LK_EXCLUSIVE | LK_RETRY))
276 		panic("%s: can't lock swap device", __func__);
277 	if (VOP_OPEN(swapdev_vp, FREAD | FWRITE, NOCRED))
278 		panic("%s: can't open swap device", __func__);
279 	VOP_UNLOCK(swapdev_vp);
280 
281 	/*
282 	 * create swap block resource map to map /dev/drum.   the range
283 	 * from 1 to INT_MAX allows 2 gigablocks of swap space.  note
284 	 * that block 0 is reserved (used to indicate an allocation
285 	 * failure, or no allocation).
286 	 */
287 	swapmap = vmem_create("swapmap", 1, INT_MAX - 1, 1, NULL, NULL, NULL, 0,
288 	    VM_NOSLEEP, IPL_NONE);
289 	if (swapmap == 0) {
290 		panic("%s: vmem_create failed", __func__);
291 	}
292 
293 	pool_init(&vndxfer_pool, sizeof(struct vndxfer), 0, 0, 0, "swp vnx",
294 	    NULL, IPL_BIO);
295 	pool_init(&vndbuf_pool, sizeof(struct vndbuf), 0, 0, 0, "swp vnd",
296 	    NULL, IPL_BIO);
297 
298 	UVMHIST_LOG(pdhist, "<- done", 0, 0, 0, 0);
299 }
300 
301 /*
302  * swaplist functions: functions that operate on the list of swap
303  * devices on the system.
304  */
305 
306 /*
307  * swaplist_insert: insert swap device "sdp" into the global list
308  *
309  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
310  * => caller must provide a newly malloc'd swappri structure (we will
311  *	FREE it if we don't need it... this it to prevent malloc blocking
312  *	here while adding swap)
313  */
314 static void
315 swaplist_insert(struct swapdev *sdp, struct swappri *newspp, int priority)
316 {
317 	struct swappri *spp, *pspp;
318 	UVMHIST_FUNC("swaplist_insert"); UVMHIST_CALLED(pdhist);
319 
320 	/*
321 	 * find entry at or after which to insert the new device.
322 	 */
323 	pspp = NULL;
324 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
325 		if (priority <= spp->spi_priority)
326 			break;
327 		pspp = spp;
328 	}
329 
330 	/*
331 	 * new priority?
332 	 */
333 	if (spp == NULL || spp->spi_priority != priority) {
334 		spp = newspp;  /* use newspp! */
335 		UVMHIST_LOG(pdhist, "created new swappri = %d",
336 			    priority, 0, 0, 0);
337 
338 		spp->spi_priority = priority;
339 		CIRCLEQ_INIT(&spp->spi_swapdev);
340 
341 		if (pspp)
342 			LIST_INSERT_AFTER(pspp, spp, spi_swappri);
343 		else
344 			LIST_INSERT_HEAD(&swap_priority, spp, spi_swappri);
345 	} else {
346 	  	/* we don't need a new priority structure, free it */
347 		free(newspp, M_VMSWAP);
348 	}
349 
350 	/*
351 	 * priority found (or created).   now insert on the priority's
352 	 * circleq list and bump the total number of swapdevs.
353 	 */
354 	sdp->swd_priority = priority;
355 	CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
356 	uvmexp.nswapdev++;
357 }
358 
359 /*
360  * swaplist_find: find and optionally remove a swap device from the
361  *	global list.
362  *
363  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
364  * => we return the swapdev we found (and removed)
365  */
366 static struct swapdev *
367 swaplist_find(struct vnode *vp, bool remove)
368 {
369 	struct swapdev *sdp;
370 	struct swappri *spp;
371 
372 	/*
373 	 * search the lists for the requested vp
374 	 */
375 
376 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
377 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
378 			if (sdp->swd_vp == vp) {
379 				if (remove) {
380 					CIRCLEQ_REMOVE(&spp->spi_swapdev,
381 					    sdp, swd_next);
382 					uvmexp.nswapdev--;
383 				}
384 				return(sdp);
385 			}
386 		}
387 	}
388 	return (NULL);
389 }
390 
391 /*
392  * swaplist_trim: scan priority list for empty priority entries and kill
393  *	them.
394  *
395  * => caller must hold both swap_syscall_lock and uvm_swap_data_lock
396  */
397 static void
398 swaplist_trim(void)
399 {
400 	struct swappri *spp, *nextspp;
401 
402 	for (spp = LIST_FIRST(&swap_priority); spp != NULL; spp = nextspp) {
403 		nextspp = LIST_NEXT(spp, spi_swappri);
404 		if (CIRCLEQ_FIRST(&spp->spi_swapdev) !=
405 		    (void *)&spp->spi_swapdev)
406 			continue;
407 		LIST_REMOVE(spp, spi_swappri);
408 		free(spp, M_VMSWAP);
409 	}
410 }
411 
412 /*
413  * swapdrum_getsdp: given a page offset in /dev/drum, convert it back
414  *	to the "swapdev" that maps that section of the drum.
415  *
416  * => each swapdev takes one big contig chunk of the drum
417  * => caller must hold uvm_swap_data_lock
418  */
419 static struct swapdev *
420 swapdrum_getsdp(int pgno)
421 {
422 	struct swapdev *sdp;
423 	struct swappri *spp;
424 
425 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
426 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
427 			if (sdp->swd_flags & SWF_FAKE)
428 				continue;
429 			if (pgno >= sdp->swd_drumoffset &&
430 			    pgno < (sdp->swd_drumoffset + sdp->swd_drumsize)) {
431 				return sdp;
432 			}
433 		}
434 	}
435 	return NULL;
436 }
437 
438 
439 /*
440  * sys_swapctl: main entry point for swapctl(2) system call
441  * 	[with two helper functions: swap_on and swap_off]
442  */
443 int
444 sys_swapctl(struct lwp *l, const struct sys_swapctl_args *uap, register_t *retval)
445 {
446 	/* {
447 		syscallarg(int) cmd;
448 		syscallarg(void *) arg;
449 		syscallarg(int) misc;
450 	} */
451 	struct vnode *vp;
452 	struct nameidata nd;
453 	struct swappri *spp;
454 	struct swapdev *sdp;
455 	struct swapent *sep;
456 #define SWAP_PATH_MAX (PATH_MAX + 1)
457 	char	*userpath;
458 	size_t	len;
459 	int	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 	rw_enter(&swap_syscall_lock, RW_WRITER);
469 
470 	userpath = malloc(SWAP_PATH_MAX, M_TEMP, M_WAITOK);
471 	/*
472 	 * we handle the non-priv NSWAP and STATS request first.
473 	 *
474 	 * SWAP_NSWAP: return number of config'd swap devices
475 	 * [can also be obtained with uvmexp sysctl]
476 	 */
477 	if (SCARG(uap, cmd) == SWAP_NSWAP) {
478 		UVMHIST_LOG(pdhist, "<- done SWAP_NSWAP=%d", uvmexp.nswapdev,
479 		    0, 0, 0);
480 		*retval = uvmexp.nswapdev;
481 		error = 0;
482 		goto out;
483 	}
484 
485 	/*
486 	 * SWAP_STATS: get stats on current # of configured swap devs
487 	 *
488 	 * note that the swap_priority list can't change as long
489 	 * as we are holding the swap_syscall_lock.  we don't want
490 	 * to grab the uvm_swap_data_lock because we may fault&sleep during
491 	 * copyout() and we don't want to be holding that lock then!
492 	 */
493 	if (SCARG(uap, cmd) == SWAP_STATS
494 #if defined(COMPAT_50)
495 	    || SCARG(uap, cmd) == SWAP_STATS50
496 #endif
497 #if defined(COMPAT_13)
498 	    || SCARG(uap, cmd) == SWAP_STATS13
499 #endif
500 	    ) {
501 		if ((size_t)misc > (size_t)uvmexp.nswapdev)
502 			misc = uvmexp.nswapdev;
503 #if defined(COMPAT_13)
504 		if (SCARG(uap, cmd) == SWAP_STATS13)
505 			len = sizeof(struct swapent13) * misc;
506 		else
507 #endif
508 #if defined(COMPAT_50)
509 		if (SCARG(uap, cmd) == SWAP_STATS50)
510 			len = sizeof(struct swapent50) * misc;
511 		else
512 #endif
513 			len = sizeof(struct swapent) * misc;
514 		sep = (struct swapent *)malloc(len, M_TEMP, M_WAITOK);
515 
516 		uvm_swap_stats_locked(SCARG(uap, cmd), sep, misc, retval);
517 		error = copyout(sep, SCARG(uap, arg), len);
518 
519 		free(sep, M_TEMP);
520 		UVMHIST_LOG(pdhist, "<- done SWAP_STATS", 0, 0, 0, 0);
521 		goto out;
522 	}
523 	if (SCARG(uap, cmd) == SWAP_GETDUMPDEV) {
524 		dev_t	*devp = (dev_t *)SCARG(uap, arg);
525 
526 		error = copyout(&dumpdev, devp, sizeof(dumpdev));
527 		goto out;
528 	}
529 
530 	/*
531 	 * all other requests require superuser privs.   verify.
532 	 */
533 	if ((error = kauth_authorize_system(l->l_cred, KAUTH_SYSTEM_SWAPCTL,
534 	    0, NULL, NULL, NULL)))
535 		goto out;
536 
537 	if (SCARG(uap, cmd) == SWAP_DUMPOFF) {
538 		/* drop the current dump device */
539 		dumpdev = NODEV;
540 		dumpcdev = NODEV;
541 		cpu_dumpconf();
542 		goto out;
543 	}
544 
545 	/*
546 	 * at this point we expect a path name in arg.   we will
547 	 * use namei() to gain a vnode reference (vref), and lock
548 	 * the vnode (VOP_LOCK).
549 	 *
550 	 * XXX: a NULL arg means use the root vnode pointer (e.g. for
551 	 * miniroot)
552 	 */
553 	if (SCARG(uap, arg) == NULL) {
554 		vp = rootvp;		/* miniroot */
555 		vref(vp);
556 		if (vn_lock(vp, LK_EXCLUSIVE)) {
557 			vrele(vp);
558 			error = EBUSY;
559 			goto out;
560 		}
561 		if (SCARG(uap, cmd) == SWAP_ON &&
562 		    copystr("miniroot", userpath, SWAP_PATH_MAX, &len))
563 			panic("swapctl: miniroot copy failed");
564 	} else {
565 		int	space;
566 		char	*where;
567 
568 		if (SCARG(uap, cmd) == SWAP_ON) {
569 			if ((error = copyinstr(SCARG(uap, arg), userpath,
570 			    SWAP_PATH_MAX, &len)))
571 				goto out;
572 			space = UIO_SYSSPACE;
573 			where = userpath;
574 		} else {
575 			space = UIO_USERSPACE;
576 			where = (char *)SCARG(uap, arg);
577 		}
578 		NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | TRYEMULROOT,
579 		    space, where);
580 		if ((error = namei(&nd)))
581 			goto out;
582 		vp = nd.ni_vp;
583 	}
584 	/* note: "vp" is referenced and locked */
585 
586 	error = 0;		/* assume no error */
587 	switch(SCARG(uap, cmd)) {
588 
589 	case SWAP_DUMPDEV:
590 		if (vp->v_type != VBLK) {
591 			error = ENOTBLK;
592 			break;
593 		}
594 		if (bdevsw_lookup(vp->v_rdev)) {
595 			dumpdev = vp->v_rdev;
596 			dumpcdev = devsw_blk2chr(dumpdev);
597 		} else
598 			dumpdev = NODEV;
599 		cpu_dumpconf();
600 		break;
601 
602 	case SWAP_CTL:
603 		/*
604 		 * get new priority, remove old entry (if any) and then
605 		 * reinsert it in the correct place.  finally, prune out
606 		 * any empty priority structures.
607 		 */
608 		priority = SCARG(uap, misc);
609 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
610 		mutex_enter(&uvm_swap_data_lock);
611 		if ((sdp = swaplist_find(vp, true)) == NULL) {
612 			error = ENOENT;
613 		} else {
614 			swaplist_insert(sdp, spp, priority);
615 			swaplist_trim();
616 		}
617 		mutex_exit(&uvm_swap_data_lock);
618 		if (error)
619 			free(spp, M_VMSWAP);
620 		break;
621 
622 	case SWAP_ON:
623 
624 		/*
625 		 * check for duplicates.   if none found, then insert a
626 		 * dummy entry on the list to prevent someone else from
627 		 * trying to enable this device while we are working on
628 		 * it.
629 		 */
630 
631 		priority = SCARG(uap, misc);
632 		sdp = malloc(sizeof *sdp, M_VMSWAP, M_WAITOK);
633 		spp = malloc(sizeof *spp, M_VMSWAP, M_WAITOK);
634 		memset(sdp, 0, sizeof(*sdp));
635 		sdp->swd_flags = SWF_FAKE;
636 		sdp->swd_vp = vp;
637 		sdp->swd_dev = (vp->v_type == VBLK) ? vp->v_rdev : NODEV;
638 		bufq_alloc(&sdp->swd_tab, "disksort", BUFQ_SORT_RAWBLOCK);
639 		mutex_enter(&uvm_swap_data_lock);
640 		if (swaplist_find(vp, false) != NULL) {
641 			error = EBUSY;
642 			mutex_exit(&uvm_swap_data_lock);
643 			bufq_free(sdp->swd_tab);
644 			free(sdp, M_VMSWAP);
645 			free(spp, M_VMSWAP);
646 			break;
647 		}
648 		swaplist_insert(sdp, spp, priority);
649 		mutex_exit(&uvm_swap_data_lock);
650 
651 		sdp->swd_pathlen = len;
652 		sdp->swd_path = malloc(sdp->swd_pathlen, M_VMSWAP, M_WAITOK);
653 		if (copystr(userpath, sdp->swd_path, sdp->swd_pathlen, 0) != 0)
654 			panic("swapctl: copystr");
655 
656 		/*
657 		 * we've now got a FAKE placeholder in the swap list.
658 		 * now attempt to enable swap on it.  if we fail, undo
659 		 * what we've done and kill the fake entry we just inserted.
660 		 * if swap_on is a success, it will clear the SWF_FAKE flag
661 		 */
662 
663 		if ((error = swap_on(l, sdp)) != 0) {
664 			mutex_enter(&uvm_swap_data_lock);
665 			(void) swaplist_find(vp, true);  /* kill fake entry */
666 			swaplist_trim();
667 			mutex_exit(&uvm_swap_data_lock);
668 			bufq_free(sdp->swd_tab);
669 			free(sdp->swd_path, M_VMSWAP);
670 			free(sdp, M_VMSWAP);
671 			break;
672 		}
673 		break;
674 
675 	case SWAP_OFF:
676 		mutex_enter(&uvm_swap_data_lock);
677 		if ((sdp = swaplist_find(vp, false)) == NULL) {
678 			mutex_exit(&uvm_swap_data_lock);
679 			error = ENXIO;
680 			break;
681 		}
682 
683 		/*
684 		 * If a device isn't in use or enabled, we
685 		 * can't stop swapping from it (again).
686 		 */
687 		if ((sdp->swd_flags & (SWF_INUSE|SWF_ENABLE)) == 0) {
688 			mutex_exit(&uvm_swap_data_lock);
689 			error = EBUSY;
690 			break;
691 		}
692 
693 		/*
694 		 * do the real work.
695 		 */
696 		error = swap_off(l, sdp);
697 		break;
698 
699 	default:
700 		error = EINVAL;
701 	}
702 
703 	/*
704 	 * done!  release the ref gained by namei() and unlock.
705 	 */
706 	vput(vp);
707 
708 out:
709 	free(userpath, M_TEMP);
710 	rw_exit(&swap_syscall_lock);
711 
712 	UVMHIST_LOG(pdhist, "<- done!  error=%d", error, 0, 0, 0);
713 	return (error);
714 }
715 
716 /*
717  * swap_stats: implements swapctl(SWAP_STATS). The function is kept
718  * away from sys_swapctl() in order to allow COMPAT_* swapctl()
719  * emulation to use it directly without going through sys_swapctl().
720  * The problem with using sys_swapctl() there is that it involves
721  * copying the swapent array to the stackgap, and this array's size
722  * is not known at build time. Hence it would not be possible to
723  * ensure it would fit in the stackgap in any case.
724  */
725 void
726 uvm_swap_stats(int cmd, struct swapent *sep, int sec, register_t *retval)
727 {
728 
729 	rw_enter(&swap_syscall_lock, RW_READER);
730 	uvm_swap_stats_locked(cmd, sep, sec, retval);
731 	rw_exit(&swap_syscall_lock);
732 }
733 
734 static void
735 uvm_swap_stats_locked(int cmd, struct swapent *sep, int sec, register_t *retval)
736 {
737 	struct swappri *spp;
738 	struct swapdev *sdp;
739 	int count = 0;
740 
741 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
742 		for (sdp = CIRCLEQ_FIRST(&spp->spi_swapdev);
743 		     sdp != (void *)&spp->spi_swapdev && sec-- > 0;
744 		     sdp = CIRCLEQ_NEXT(sdp, swd_next)) {
745 			int inuse;
746 
747 		  	/*
748 			 * backwards compatibility for system call.
749 			 * For NetBSD 1.3 and 5.0, we have to use
750 			 * the 32 bit dev_t.  For 5.0 and -current
751 			 * we have to add the path.
752 			 */
753 			inuse = btodb((uint64_t)sdp->swd_npginuse <<
754 			    PAGE_SHIFT);
755 
756 #if defined(COMPAT_13) || defined(COMPAT_50)
757 			if (cmd == SWAP_STATS) {
758 #endif
759 				sep->se_dev = sdp->swd_dev;
760 				sep->se_flags = sdp->swd_flags;
761 				sep->se_nblks = sdp->swd_nblks;
762 				sep->se_inuse = inuse;
763 				sep->se_priority = sdp->swd_priority;
764 				memcpy(&sep->se_path, sdp->swd_path,
765 				       sizeof sep->se_path);
766 				sep++;
767 #if defined(COMPAT_13)
768 			} else if (cmd == SWAP_STATS13) {
769 				struct swapent13 *sep13 =
770 				    (struct swapent13 *)sep;
771 
772 				sep13->se13_dev = sdp->swd_dev;
773 				sep13->se13_flags = sdp->swd_flags;
774 				sep13->se13_nblks = sdp->swd_nblks;
775 				sep13->se13_inuse = inuse;
776 				sep13->se13_priority = sdp->swd_priority;
777 				sep = (struct swapent *)(sep13 + 1);
778 #endif
779 #if defined(COMPAT_50)
780 			} else if (cmd == SWAP_STATS50) {
781 				struct swapent50 *sep50 =
782 				    (struct swapent50 *)sep;
783 
784 				sep50->se50_dev = sdp->swd_dev;
785 				sep50->se50_flags = sdp->swd_flags;
786 				sep50->se50_nblks = sdp->swd_nblks;
787 				sep50->se50_inuse = inuse;
788 				sep50->se50_priority = sdp->swd_priority;
789 				memcpy(&sep50->se50_path, sdp->swd_path,
790 				       sizeof sep50->se50_path);
791 				sep = (struct swapent *)(sep50 + 1);
792 #endif
793 #if defined(COMPAT_13) || defined(COMPAT_50)
794 			}
795 #endif
796 			count++;
797 		}
798 	}
799 
800 	*retval = count;
801 	return;
802 }
803 
804 /*
805  * swap_on: attempt to enable a swapdev for swapping.   note that the
806  *	swapdev is already on the global list, but disabled (marked
807  *	SWF_FAKE).
808  *
809  * => we avoid the start of the disk (to protect disk labels)
810  * => we also avoid the miniroot, if we are swapping to root.
811  * => caller should leave uvm_swap_data_lock unlocked, we may lock it
812  *	if needed.
813  */
814 static int
815 swap_on(struct lwp *l, struct swapdev *sdp)
816 {
817 	struct vnode *vp;
818 	int error, npages, nblocks, size;
819 	long addr;
820 	u_long result;
821 	struct vattr va;
822 	const struct bdevsw *bdev;
823 	dev_t dev;
824 	UVMHIST_FUNC("swap_on"); UVMHIST_CALLED(pdhist);
825 
826 	/*
827 	 * we want to enable swapping on sdp.   the swd_vp contains
828 	 * the vnode we want (locked and ref'd), and the swd_dev
829 	 * contains the dev_t of the file, if it a block device.
830 	 */
831 
832 	vp = sdp->swd_vp;
833 	dev = sdp->swd_dev;
834 
835 	/*
836 	 * open the swap file (mostly useful for block device files to
837 	 * let device driver know what is up).
838 	 *
839 	 * we skip the open/close for root on swap because the root
840 	 * has already been opened when root was mounted (mountroot).
841 	 */
842 	if (vp != rootvp) {
843 		if ((error = VOP_OPEN(vp, FREAD|FWRITE, l->l_cred)))
844 			return (error);
845 	}
846 
847 	/* XXX this only works for block devices */
848 	UVMHIST_LOG(pdhist, "  dev=%d, major(dev)=%d", dev, major(dev), 0,0);
849 
850 	/*
851 	 * we now need to determine the size of the swap area.   for
852 	 * block specials we can call the d_psize function.
853 	 * for normal files, we must stat [get attrs].
854 	 *
855 	 * we put the result in nblks.
856 	 * for normal files, we also want the filesystem block size
857 	 * (which we get with statfs).
858 	 */
859 	switch (vp->v_type) {
860 	case VBLK:
861 		bdev = bdevsw_lookup(dev);
862 		if (bdev == NULL || bdev->d_psize == NULL ||
863 		    (nblocks = (*bdev->d_psize)(dev)) == -1) {
864 			error = ENXIO;
865 			goto bad;
866 		}
867 		break;
868 
869 	case VREG:
870 		if ((error = VOP_GETATTR(vp, &va, l->l_cred)))
871 			goto bad;
872 		nblocks = (int)btodb(va.va_size);
873 		sdp->swd_bsize = 1 << vp->v_mount->mnt_fs_bshift;
874 		/*
875 		 * limit the max # of outstanding I/O requests we issue
876 		 * at any one time.   take it easy on NFS servers.
877 		 */
878 		if (vp->v_tag == VT_NFS)
879 			sdp->swd_maxactive = 2; /* XXX */
880 		else
881 			sdp->swd_maxactive = 8; /* XXX */
882 		break;
883 
884 	default:
885 		error = ENXIO;
886 		goto bad;
887 	}
888 
889 	/*
890 	 * save nblocks in a safe place and convert to pages.
891 	 */
892 
893 	sdp->swd_nblks = nblocks;
894 	npages = dbtob((uint64_t)nblocks) >> PAGE_SHIFT;
895 
896 	/*
897 	 * for block special files, we want to make sure that leave
898 	 * the disklabel and bootblocks alone, so we arrange to skip
899 	 * over them (arbitrarily choosing to skip PAGE_SIZE bytes).
900 	 * note that because of this the "size" can be less than the
901 	 * actual number of blocks on the device.
902 	 */
903 	if (vp->v_type == VBLK) {
904 		/* we use pages 1 to (size - 1) [inclusive] */
905 		size = npages - 1;
906 		addr = 1;
907 	} else {
908 		/* we use pages 0 to (size - 1) [inclusive] */
909 		size = npages;
910 		addr = 0;
911 	}
912 
913 	/*
914 	 * make sure we have enough blocks for a reasonable sized swap
915 	 * area.   we want at least one page.
916 	 */
917 
918 	if (size < 1) {
919 		UVMHIST_LOG(pdhist, "  size <= 1!!", 0, 0, 0, 0);
920 		error = EINVAL;
921 		goto bad;
922 	}
923 
924 	UVMHIST_LOG(pdhist, "  dev=%x: size=%d addr=%ld\n", dev, size, addr, 0);
925 
926 	/*
927 	 * now we need to allocate an extent to manage this swap device
928 	 */
929 
930 	sdp->swd_blist = blist_create(npages);
931 	/* mark all expect the `saved' region free. */
932 	blist_free(sdp->swd_blist, addr, size);
933 
934 	/*
935 	 * if the vnode we are swapping to is the root vnode
936 	 * (i.e. we are swapping to the miniroot) then we want
937 	 * to make sure we don't overwrite it.   do a statfs to
938 	 * find its size and skip over it.
939 	 */
940 	if (vp == rootvp) {
941 		struct mount *mp;
942 		struct statvfs *sp;
943 		int rootblocks, rootpages;
944 
945 		mp = rootvnode->v_mount;
946 		sp = &mp->mnt_stat;
947 		rootblocks = sp->f_blocks * btodb(sp->f_frsize);
948 		/*
949 		 * XXX: sp->f_blocks isn't the total number of
950 		 * blocks in the filesystem, it's the number of
951 		 * data blocks.  so, our rootblocks almost
952 		 * definitely underestimates the total size
953 		 * of the filesystem - how badly depends on the
954 		 * details of the filesystem type.  there isn't
955 		 * an obvious way to deal with this cleanly
956 		 * and perfectly, so for now we just pad our
957 		 * rootblocks estimate with an extra 5 percent.
958 		 */
959 		rootblocks += (rootblocks >> 5) +
960 			(rootblocks >> 6) +
961 			(rootblocks >> 7);
962 		rootpages = round_page(dbtob(rootblocks)) >> PAGE_SHIFT;
963 		if (rootpages > size)
964 			panic("swap_on: miniroot larger than swap?");
965 
966 		if (rootpages != blist_fill(sdp->swd_blist, addr, rootpages)) {
967 			panic("swap_on: unable to preserve miniroot");
968 		}
969 
970 		size -= rootpages;
971 		printf("Preserved %d pages of miniroot ", rootpages);
972 		printf("leaving %d pages of swap\n", size);
973 	}
974 
975 	/*
976 	 * add a ref to vp to reflect usage as a swap device.
977 	 */
978 	vref(vp);
979 
980 	/*
981 	 * now add the new swapdev to the drum and enable.
982 	 */
983 	result = vmem_alloc(swapmap, npages, VM_BESTFIT | VM_SLEEP);
984 	if (result == 0)
985 		panic("swapdrum_add");
986 	/*
987 	 * If this is the first regular swap create the workqueue.
988 	 * => Protected by swap_syscall_lock.
989 	 */
990 	if (vp->v_type != VBLK) {
991 		if (sw_reg_count++ == 0) {
992 			KASSERT(sw_reg_workqueue == NULL);
993 			if (workqueue_create(&sw_reg_workqueue, "swapiod",
994 			    sw_reg_iodone, NULL, PRIBIO, IPL_BIO, 0) != 0)
995 				panic("%s: workqueue_create failed", __func__);
996 		}
997 	}
998 
999 	sdp->swd_drumoffset = (int)result;
1000 	sdp->swd_drumsize = npages;
1001 	sdp->swd_npages = size;
1002 	mutex_enter(&uvm_swap_data_lock);
1003 	sdp->swd_flags &= ~SWF_FAKE;	/* going live */
1004 	sdp->swd_flags |= (SWF_INUSE|SWF_ENABLE);
1005 	uvmexp.swpages += size;
1006 	uvmexp.swpgavail += size;
1007 	mutex_exit(&uvm_swap_data_lock);
1008 	return (0);
1009 
1010 	/*
1011 	 * failure: clean up and return error.
1012 	 */
1013 
1014 bad:
1015 	if (sdp->swd_blist) {
1016 		blist_destroy(sdp->swd_blist);
1017 	}
1018 	if (vp != rootvp) {
1019 		(void)VOP_CLOSE(vp, FREAD|FWRITE, l->l_cred);
1020 	}
1021 	return (error);
1022 }
1023 
1024 /*
1025  * swap_off: stop swapping on swapdev
1026  *
1027  * => swap data should be locked, we will unlock.
1028  */
1029 static int
1030 swap_off(struct lwp *l, struct swapdev *sdp)
1031 {
1032 	int npages = sdp->swd_npages;
1033 	int error = 0;
1034 
1035 	UVMHIST_FUNC("swap_off"); UVMHIST_CALLED(pdhist);
1036 	UVMHIST_LOG(pdhist, "  dev=%x, npages=%d", sdp->swd_dev,npages,0,0);
1037 
1038 	/* disable the swap area being removed */
1039 	sdp->swd_flags &= ~SWF_ENABLE;
1040 	uvmexp.swpgavail -= npages;
1041 	mutex_exit(&uvm_swap_data_lock);
1042 
1043 	/*
1044 	 * the idea is to find all the pages that are paged out to this
1045 	 * device, and page them all in.  in uvm, swap-backed pageable
1046 	 * memory can take two forms: aobjs and anons.  call the
1047 	 * swapoff hook for each subsystem to bring in pages.
1048 	 */
1049 
1050 	if (uao_swap_off(sdp->swd_drumoffset,
1051 			 sdp->swd_drumoffset + sdp->swd_drumsize) ||
1052 	    amap_swap_off(sdp->swd_drumoffset,
1053 			  sdp->swd_drumoffset + sdp->swd_drumsize)) {
1054 		error = ENOMEM;
1055 	} else if (sdp->swd_npginuse > sdp->swd_npgbad) {
1056 		error = EBUSY;
1057 	}
1058 
1059 	if (error) {
1060 		mutex_enter(&uvm_swap_data_lock);
1061 		sdp->swd_flags |= SWF_ENABLE;
1062 		uvmexp.swpgavail += npages;
1063 		mutex_exit(&uvm_swap_data_lock);
1064 
1065 		return error;
1066 	}
1067 
1068 	/*
1069 	 * If this is the last regular swap destroy the workqueue.
1070 	 * => Protected by swap_syscall_lock.
1071 	 */
1072 	if (sdp->swd_vp->v_type != VBLK) {
1073 		KASSERT(sw_reg_count > 0);
1074 		KASSERT(sw_reg_workqueue != NULL);
1075 		if (--sw_reg_count == 0) {
1076 			workqueue_destroy(sw_reg_workqueue);
1077 			sw_reg_workqueue = NULL;
1078 		}
1079 	}
1080 
1081 	/*
1082 	 * done with the vnode.
1083 	 * drop our ref on the vnode before calling VOP_CLOSE()
1084 	 * so that spec_close() can tell if this is the last close.
1085 	 */
1086 	vrele(sdp->swd_vp);
1087 	if (sdp->swd_vp != rootvp) {
1088 		(void) VOP_CLOSE(sdp->swd_vp, FREAD|FWRITE, l->l_cred);
1089 	}
1090 
1091 	mutex_enter(&uvm_swap_data_lock);
1092 	uvmexp.swpages -= npages;
1093 	uvmexp.swpginuse -= sdp->swd_npgbad;
1094 
1095 	if (swaplist_find(sdp->swd_vp, true) == NULL)
1096 		panic("%s: swapdev not in list", __func__);
1097 	swaplist_trim();
1098 	mutex_exit(&uvm_swap_data_lock);
1099 
1100 	/*
1101 	 * free all resources!
1102 	 */
1103 	vmem_free(swapmap, sdp->swd_drumoffset, sdp->swd_drumsize);
1104 	blist_destroy(sdp->swd_blist);
1105 	bufq_free(sdp->swd_tab);
1106 	free(sdp, M_VMSWAP);
1107 	return (0);
1108 }
1109 
1110 /*
1111  * /dev/drum interface and i/o functions
1112  */
1113 
1114 /*
1115  * swstrategy: perform I/O on the drum
1116  *
1117  * => we must map the i/o request from the drum to the correct swapdev.
1118  */
1119 static void
1120 swstrategy(struct buf *bp)
1121 {
1122 	struct swapdev *sdp;
1123 	struct vnode *vp;
1124 	int pageno, bn;
1125 	UVMHIST_FUNC("swstrategy"); UVMHIST_CALLED(pdhist);
1126 
1127 	/*
1128 	 * convert block number to swapdev.   note that swapdev can't
1129 	 * be yanked out from under us because we are holding resources
1130 	 * in it (i.e. the blocks we are doing I/O on).
1131 	 */
1132 	pageno = dbtob((int64_t)bp->b_blkno) >> PAGE_SHIFT;
1133 	mutex_enter(&uvm_swap_data_lock);
1134 	sdp = swapdrum_getsdp(pageno);
1135 	mutex_exit(&uvm_swap_data_lock);
1136 	if (sdp == NULL) {
1137 		bp->b_error = EINVAL;
1138 		biodone(bp);
1139 		UVMHIST_LOG(pdhist, "  failed to get swap device", 0, 0, 0, 0);
1140 		return;
1141 	}
1142 
1143 	/*
1144 	 * convert drum page number to block number on this swapdev.
1145 	 */
1146 
1147 	pageno -= sdp->swd_drumoffset;	/* page # on swapdev */
1148 	bn = btodb((uint64_t)pageno << PAGE_SHIFT); /* convert to diskblock */
1149 
1150 	UVMHIST_LOG(pdhist, "  %s: mapoff=%x bn=%x bcount=%ld",
1151 		((bp->b_flags & B_READ) == 0) ? "write" : "read",
1152 		sdp->swd_drumoffset, bn, bp->b_bcount);
1153 
1154 	/*
1155 	 * for block devices we finish up here.
1156 	 * for regular files we have to do more work which we delegate
1157 	 * to sw_reg_strategy().
1158 	 */
1159 
1160 	vp = sdp->swd_vp;		/* swapdev vnode pointer */
1161 	switch (vp->v_type) {
1162 	default:
1163 		panic("%s: vnode type 0x%x", __func__, vp->v_type);
1164 
1165 	case VBLK:
1166 
1167 		/*
1168 		 * must convert "bp" from an I/O on /dev/drum to an I/O
1169 		 * on the swapdev (sdp).
1170 		 */
1171 		bp->b_blkno = bn;		/* swapdev block number */
1172 		bp->b_dev = sdp->swd_dev;	/* swapdev dev_t */
1173 
1174 		/*
1175 		 * if we are doing a write, we have to redirect the i/o on
1176 		 * drum's v_numoutput counter to the swapdevs.
1177 		 */
1178 		if ((bp->b_flags & B_READ) == 0) {
1179 			mutex_enter(bp->b_objlock);
1180 			vwakeup(bp);	/* kills one 'v_numoutput' on drum */
1181 			mutex_exit(bp->b_objlock);
1182 			mutex_enter(&vp->v_interlock);
1183 			vp->v_numoutput++;	/* put it on swapdev */
1184 			mutex_exit(&vp->v_interlock);
1185 		}
1186 
1187 		/*
1188 		 * finally plug in swapdev vnode and start I/O
1189 		 */
1190 		bp->b_vp = vp;
1191 		bp->b_objlock = &vp->v_interlock;
1192 		VOP_STRATEGY(vp, bp);
1193 		return;
1194 
1195 	case VREG:
1196 		/*
1197 		 * delegate to sw_reg_strategy function.
1198 		 */
1199 		sw_reg_strategy(sdp, bp, bn);
1200 		return;
1201 	}
1202 	/* NOTREACHED */
1203 }
1204 
1205 /*
1206  * swread: the read function for the drum (just a call to physio)
1207  */
1208 /*ARGSUSED*/
1209 static int
1210 swread(dev_t dev, struct uio *uio, int ioflag)
1211 {
1212 	UVMHIST_FUNC("swread"); UVMHIST_CALLED(pdhist);
1213 
1214 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1215 	return (physio(swstrategy, NULL, dev, B_READ, minphys, uio));
1216 }
1217 
1218 /*
1219  * swwrite: the write function for the drum (just a call to physio)
1220  */
1221 /*ARGSUSED*/
1222 static int
1223 swwrite(dev_t dev, struct uio *uio, int ioflag)
1224 {
1225 	UVMHIST_FUNC("swwrite"); UVMHIST_CALLED(pdhist);
1226 
1227 	UVMHIST_LOG(pdhist, "  dev=%x offset=%qx", dev, uio->uio_offset, 0, 0);
1228 	return (physio(swstrategy, NULL, dev, B_WRITE, minphys, uio));
1229 }
1230 
1231 const struct bdevsw swap_bdevsw = {
1232 	nullopen, nullclose, swstrategy, noioctl, nodump, nosize, D_OTHER,
1233 };
1234 
1235 const struct cdevsw swap_cdevsw = {
1236 	nullopen, nullclose, swread, swwrite, noioctl,
1237 	nostop, notty, nopoll, nommap, nokqfilter, D_OTHER,
1238 };
1239 
1240 /*
1241  * sw_reg_strategy: handle swap i/o to regular files
1242  */
1243 static void
1244 sw_reg_strategy(struct swapdev *sdp, struct buf *bp, int bn)
1245 {
1246 	struct vnode	*vp;
1247 	struct vndxfer	*vnx;
1248 	daddr_t		nbn;
1249 	char 		*addr;
1250 	off_t		byteoff;
1251 	int		s, off, nra, error, sz, resid;
1252 	UVMHIST_FUNC("sw_reg_strategy"); UVMHIST_CALLED(pdhist);
1253 
1254 	/*
1255 	 * allocate a vndxfer head for this transfer and point it to
1256 	 * our buffer.
1257 	 */
1258 	vnx = pool_get(&vndxfer_pool, PR_WAITOK);
1259 	vnx->vx_flags = VX_BUSY;
1260 	vnx->vx_error = 0;
1261 	vnx->vx_pending = 0;
1262 	vnx->vx_bp = bp;
1263 	vnx->vx_sdp = sdp;
1264 
1265 	/*
1266 	 * setup for main loop where we read filesystem blocks into
1267 	 * our buffer.
1268 	 */
1269 	error = 0;
1270 	bp->b_resid = bp->b_bcount;	/* nothing transfered yet! */
1271 	addr = bp->b_data;		/* current position in buffer */
1272 	byteoff = dbtob((uint64_t)bn);
1273 
1274 	for (resid = bp->b_resid; resid; resid -= sz) {
1275 		struct vndbuf	*nbp;
1276 
1277 		/*
1278 		 * translate byteoffset into block number.  return values:
1279 		 *   vp = vnode of underlying device
1280 		 *  nbn = new block number (on underlying vnode dev)
1281 		 *  nra = num blocks we can read-ahead (excludes requested
1282 		 *	block)
1283 		 */
1284 		nra = 0;
1285 		error = VOP_BMAP(sdp->swd_vp, byteoff / sdp->swd_bsize,
1286 				 	&vp, &nbn, &nra);
1287 
1288 		if (error == 0 && nbn == (daddr_t)-1) {
1289 			/*
1290 			 * this used to just set error, but that doesn't
1291 			 * do the right thing.  Instead, it causes random
1292 			 * memory errors.  The panic() should remain until
1293 			 * this condition doesn't destabilize the system.
1294 			 */
1295 #if 1
1296 			panic("%s: swap to sparse file", __func__);
1297 #else
1298 			error = EIO;	/* failure */
1299 #endif
1300 		}
1301 
1302 		/*
1303 		 * punt if there was an error or a hole in the file.
1304 		 * we must wait for any i/o ops we have already started
1305 		 * to finish before returning.
1306 		 *
1307 		 * XXX we could deal with holes here but it would be
1308 		 * a hassle (in the write case).
1309 		 */
1310 		if (error) {
1311 			s = splbio();
1312 			vnx->vx_error = error;	/* pass error up */
1313 			goto out;
1314 		}
1315 
1316 		/*
1317 		 * compute the size ("sz") of this transfer (in bytes).
1318 		 */
1319 		off = byteoff % sdp->swd_bsize;
1320 		sz = (1 + nra) * sdp->swd_bsize - off;
1321 		if (sz > resid)
1322 			sz = resid;
1323 
1324 		UVMHIST_LOG(pdhist, "sw_reg_strategy: "
1325 			    "vp %p/%p offset 0x%x/0x%x",
1326 			    sdp->swd_vp, vp, byteoff, nbn);
1327 
1328 		/*
1329 		 * now get a buf structure.   note that the vb_buf is
1330 		 * at the front of the nbp structure so that you can
1331 		 * cast pointers between the two structure easily.
1332 		 */
1333 		nbp = pool_get(&vndbuf_pool, PR_WAITOK);
1334 		buf_init(&nbp->vb_buf);
1335 		nbp->vb_buf.b_flags    = bp->b_flags;
1336 		nbp->vb_buf.b_cflags   = bp->b_cflags;
1337 		nbp->vb_buf.b_oflags   = bp->b_oflags;
1338 		nbp->vb_buf.b_bcount   = sz;
1339 		nbp->vb_buf.b_bufsize  = sz;
1340 		nbp->vb_buf.b_error    = 0;
1341 		nbp->vb_buf.b_data     = addr;
1342 		nbp->vb_buf.b_lblkno   = 0;
1343 		nbp->vb_buf.b_blkno    = nbn + btodb(off);
1344 		nbp->vb_buf.b_rawblkno = nbp->vb_buf.b_blkno;
1345 		nbp->vb_buf.b_iodone   = sw_reg_biodone;
1346 		nbp->vb_buf.b_vp       = vp;
1347 		nbp->vb_buf.b_objlock  = &vp->v_interlock;
1348 		if (vp->v_type == VBLK) {
1349 			nbp->vb_buf.b_dev = vp->v_rdev;
1350 		}
1351 
1352 		nbp->vb_xfer = vnx;	/* patch it back in to vnx */
1353 
1354 		/*
1355 		 * Just sort by block number
1356 		 */
1357 		s = splbio();
1358 		if (vnx->vx_error != 0) {
1359 			buf_destroy(&nbp->vb_buf);
1360 			pool_put(&vndbuf_pool, nbp);
1361 			goto out;
1362 		}
1363 		vnx->vx_pending++;
1364 
1365 		/* sort it in and start I/O if we are not over our limit */
1366 		/* XXXAD locking */
1367 		bufq_put(sdp->swd_tab, &nbp->vb_buf);
1368 		sw_reg_start(sdp);
1369 		splx(s);
1370 
1371 		/*
1372 		 * advance to the next I/O
1373 		 */
1374 		byteoff += sz;
1375 		addr += sz;
1376 	}
1377 
1378 	s = splbio();
1379 
1380 out: /* Arrive here at splbio */
1381 	vnx->vx_flags &= ~VX_BUSY;
1382 	if (vnx->vx_pending == 0) {
1383 		error = vnx->vx_error;
1384 		pool_put(&vndxfer_pool, vnx);
1385 		bp->b_error = error;
1386 		biodone(bp);
1387 	}
1388 	splx(s);
1389 }
1390 
1391 /*
1392  * sw_reg_start: start an I/O request on the requested swapdev
1393  *
1394  * => reqs are sorted by b_rawblkno (above)
1395  */
1396 static void
1397 sw_reg_start(struct swapdev *sdp)
1398 {
1399 	struct buf	*bp;
1400 	struct vnode	*vp;
1401 	UVMHIST_FUNC("sw_reg_start"); UVMHIST_CALLED(pdhist);
1402 
1403 	/* recursion control */
1404 	if ((sdp->swd_flags & SWF_BUSY) != 0)
1405 		return;
1406 
1407 	sdp->swd_flags |= SWF_BUSY;
1408 
1409 	while (sdp->swd_active < sdp->swd_maxactive) {
1410 		bp = bufq_get(sdp->swd_tab);
1411 		if (bp == NULL)
1412 			break;
1413 		sdp->swd_active++;
1414 
1415 		UVMHIST_LOG(pdhist,
1416 		    "sw_reg_start:  bp %p vp %p blkno %p cnt %lx",
1417 		    bp, bp->b_vp, bp->b_blkno, bp->b_bcount);
1418 		vp = bp->b_vp;
1419 		KASSERT(bp->b_objlock == &vp->v_interlock);
1420 		if ((bp->b_flags & B_READ) == 0) {
1421 			mutex_enter(&vp->v_interlock);
1422 			vp->v_numoutput++;
1423 			mutex_exit(&vp->v_interlock);
1424 		}
1425 		VOP_STRATEGY(vp, bp);
1426 	}
1427 	sdp->swd_flags &= ~SWF_BUSY;
1428 }
1429 
1430 /*
1431  * sw_reg_biodone: one of our i/o's has completed
1432  */
1433 static void
1434 sw_reg_biodone(struct buf *bp)
1435 {
1436 	workqueue_enqueue(sw_reg_workqueue, &bp->b_work, NULL);
1437 }
1438 
1439 /*
1440  * sw_reg_iodone: one of our i/o's has completed and needs post-i/o cleanup
1441  *
1442  * => note that we can recover the vndbuf struct by casting the buf ptr
1443  */
1444 static void
1445 sw_reg_iodone(struct work *wk, void *dummy)
1446 {
1447 	struct vndbuf *vbp = (void *)wk;
1448 	struct vndxfer *vnx = vbp->vb_xfer;
1449 	struct buf *pbp = vnx->vx_bp;		/* parent buffer */
1450 	struct swapdev	*sdp = vnx->vx_sdp;
1451 	int s, resid, error;
1452 	KASSERT(&vbp->vb_buf.b_work == wk);
1453 	UVMHIST_FUNC("sw_reg_iodone"); UVMHIST_CALLED(pdhist);
1454 
1455 	UVMHIST_LOG(pdhist, "  vbp=%p vp=%p blkno=%x addr=%p",
1456 	    vbp, vbp->vb_buf.b_vp, vbp->vb_buf.b_blkno, vbp->vb_buf.b_data);
1457 	UVMHIST_LOG(pdhist, "  cnt=%lx resid=%lx",
1458 	    vbp->vb_buf.b_bcount, vbp->vb_buf.b_resid, 0, 0);
1459 
1460 	/*
1461 	 * protect vbp at splbio and update.
1462 	 */
1463 
1464 	s = splbio();
1465 	resid = vbp->vb_buf.b_bcount - vbp->vb_buf.b_resid;
1466 	pbp->b_resid -= resid;
1467 	vnx->vx_pending--;
1468 
1469 	if (vbp->vb_buf.b_error != 0) {
1470 		/* pass error upward */
1471 		error = vbp->vb_buf.b_error ? vbp->vb_buf.b_error : EIO;
1472 		UVMHIST_LOG(pdhist, "  got error=%d !", error, 0, 0, 0);
1473 		vnx->vx_error = error;
1474 	}
1475 
1476 	/*
1477 	 * kill vbp structure
1478 	 */
1479 	buf_destroy(&vbp->vb_buf);
1480 	pool_put(&vndbuf_pool, vbp);
1481 
1482 	/*
1483 	 * wrap up this transaction if it has run to completion or, in
1484 	 * case of an error, when all auxiliary buffers have returned.
1485 	 */
1486 	if (vnx->vx_error != 0) {
1487 		/* pass error upward */
1488 		error = vnx->vx_error;
1489 		if ((vnx->vx_flags & VX_BUSY) == 0 && vnx->vx_pending == 0) {
1490 			pbp->b_error = error;
1491 			biodone(pbp);
1492 			pool_put(&vndxfer_pool, vnx);
1493 		}
1494 	} else if (pbp->b_resid == 0) {
1495 		KASSERT(vnx->vx_pending == 0);
1496 		if ((vnx->vx_flags & VX_BUSY) == 0) {
1497 			UVMHIST_LOG(pdhist, "  iodone error=%d !",
1498 			    pbp, vnx->vx_error, 0, 0);
1499 			biodone(pbp);
1500 			pool_put(&vndxfer_pool, vnx);
1501 		}
1502 	}
1503 
1504 	/*
1505 	 * done!   start next swapdev I/O if one is pending
1506 	 */
1507 	sdp->swd_active--;
1508 	sw_reg_start(sdp);
1509 	splx(s);
1510 }
1511 
1512 
1513 /*
1514  * uvm_swap_alloc: allocate space on swap
1515  *
1516  * => allocation is done "round robin" down the priority list, as we
1517  *	allocate in a priority we "rotate" the circle queue.
1518  * => space can be freed with uvm_swap_free
1519  * => we return the page slot number in /dev/drum (0 == invalid slot)
1520  * => we lock uvm_swap_data_lock
1521  * => XXXMRG: "LESSOK" INTERFACE NEEDED TO EXTENT SYSTEM
1522  */
1523 int
1524 uvm_swap_alloc(int *nslots /* IN/OUT */, bool lessok)
1525 {
1526 	struct swapdev *sdp;
1527 	struct swappri *spp;
1528 	UVMHIST_FUNC("uvm_swap_alloc"); UVMHIST_CALLED(pdhist);
1529 
1530 	/*
1531 	 * no swap devices configured yet?   definite failure.
1532 	 */
1533 	if (uvmexp.nswapdev < 1)
1534 		return 0;
1535 
1536 	/*
1537 	 * lock data lock, convert slots into blocks, and enter loop
1538 	 */
1539 	mutex_enter(&uvm_swap_data_lock);
1540 
1541 ReTry:	/* XXXMRG */
1542 	LIST_FOREACH(spp, &swap_priority, spi_swappri) {
1543 		CIRCLEQ_FOREACH(sdp, &spp->spi_swapdev, swd_next) {
1544 			uint64_t result;
1545 
1546 			/* if it's not enabled, then we can't swap from it */
1547 			if ((sdp->swd_flags & SWF_ENABLE) == 0)
1548 				continue;
1549 			if (sdp->swd_npginuse + *nslots > sdp->swd_npages)
1550 				continue;
1551 			result = blist_alloc(sdp->swd_blist, *nslots);
1552 			if (result == BLIST_NONE) {
1553 				continue;
1554 			}
1555 			KASSERT(result < sdp->swd_drumsize);
1556 
1557 			/*
1558 			 * successful allocation!  now rotate the circleq.
1559 			 */
1560 			CIRCLEQ_REMOVE(&spp->spi_swapdev, sdp, swd_next);
1561 			CIRCLEQ_INSERT_TAIL(&spp->spi_swapdev, sdp, swd_next);
1562 			sdp->swd_npginuse += *nslots;
1563 			uvmexp.swpginuse += *nslots;
1564 			mutex_exit(&uvm_swap_data_lock);
1565 			/* done!  return drum slot number */
1566 			UVMHIST_LOG(pdhist,
1567 			    "success!  returning %d slots starting at %d",
1568 			    *nslots, result + sdp->swd_drumoffset, 0, 0);
1569 			return (result + sdp->swd_drumoffset);
1570 		}
1571 	}
1572 
1573 	/* XXXMRG: BEGIN HACK */
1574 	if (*nslots > 1 && lessok) {
1575 		*nslots = 1;
1576 		/* XXXMRG: ugh!  blist should support this for us */
1577 		goto ReTry;
1578 	}
1579 	/* XXXMRG: END HACK */
1580 
1581 	mutex_exit(&uvm_swap_data_lock);
1582 	return 0;
1583 }
1584 
1585 /*
1586  * uvm_swapisfull: return true if most of available swap is allocated
1587  * and in use.  we don't count some small portion as it may be inaccessible
1588  * to us at any given moment, for example if there is lock contention or if
1589  * pages are busy.
1590  */
1591 bool
1592 uvm_swapisfull(void)
1593 {
1594 	int swpgonly;
1595 	bool rv;
1596 
1597 	mutex_enter(&uvm_swap_data_lock);
1598 	KASSERT(uvmexp.swpgonly <= uvmexp.swpages);
1599 	swpgonly = (int)((uint64_t)uvmexp.swpgonly * 100 /
1600 	    uvm_swapisfull_factor);
1601 	rv = (swpgonly >= uvmexp.swpgavail);
1602 	mutex_exit(&uvm_swap_data_lock);
1603 
1604 	return (rv);
1605 }
1606 
1607 /*
1608  * uvm_swap_markbad: keep track of swap ranges where we've had i/o errors
1609  *
1610  * => we lock uvm_swap_data_lock
1611  */
1612 void
1613 uvm_swap_markbad(int startslot, int nslots)
1614 {
1615 	struct swapdev *sdp;
1616 	UVMHIST_FUNC("uvm_swap_markbad"); UVMHIST_CALLED(pdhist);
1617 
1618 	mutex_enter(&uvm_swap_data_lock);
1619 	sdp = swapdrum_getsdp(startslot);
1620 	KASSERT(sdp != NULL);
1621 
1622 	/*
1623 	 * we just keep track of how many pages have been marked bad
1624 	 * in this device, to make everything add up in swap_off().
1625 	 * we assume here that the range of slots will all be within
1626 	 * one swap device.
1627 	 */
1628 
1629 	KASSERT(uvmexp.swpgonly >= nslots);
1630 	uvmexp.swpgonly -= nslots;
1631 	sdp->swd_npgbad += nslots;
1632 	UVMHIST_LOG(pdhist, "now %d bad", sdp->swd_npgbad, 0,0,0);
1633 	mutex_exit(&uvm_swap_data_lock);
1634 }
1635 
1636 /*
1637  * uvm_swap_free: free swap slots
1638  *
1639  * => this can be all or part of an allocation made by uvm_swap_alloc
1640  * => we lock uvm_swap_data_lock
1641  */
1642 void
1643 uvm_swap_free(int startslot, int nslots)
1644 {
1645 	struct swapdev *sdp;
1646 	UVMHIST_FUNC("uvm_swap_free"); UVMHIST_CALLED(pdhist);
1647 
1648 	UVMHIST_LOG(pdhist, "freeing %d slots starting at %d", nslots,
1649 	    startslot, 0, 0);
1650 
1651 	/*
1652 	 * ignore attempts to free the "bad" slot.
1653 	 */
1654 
1655 	if (startslot == SWSLOT_BAD) {
1656 		return;
1657 	}
1658 
1659 	/*
1660 	 * convert drum slot offset back to sdp, free the blocks
1661 	 * in the extent, and return.   must hold pri lock to do
1662 	 * lookup and access the extent.
1663 	 */
1664 
1665 	mutex_enter(&uvm_swap_data_lock);
1666 	sdp = swapdrum_getsdp(startslot);
1667 	KASSERT(uvmexp.nswapdev >= 1);
1668 	KASSERT(sdp != NULL);
1669 	KASSERT(sdp->swd_npginuse >= nslots);
1670 	blist_free(sdp->swd_blist, startslot - sdp->swd_drumoffset, nslots);
1671 	sdp->swd_npginuse -= nslots;
1672 	uvmexp.swpginuse -= nslots;
1673 	mutex_exit(&uvm_swap_data_lock);
1674 }
1675 
1676 /*
1677  * uvm_swap_put: put any number of pages into a contig place on swap
1678  *
1679  * => can be sync or async
1680  */
1681 
1682 int
1683 uvm_swap_put(int swslot, struct vm_page **ppsp, int npages, int flags)
1684 {
1685 	int error;
1686 
1687 	error = uvm_swap_io(ppsp, swslot, npages, B_WRITE |
1688 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1689 	return error;
1690 }
1691 
1692 /*
1693  * uvm_swap_get: get a single page from swap
1694  *
1695  * => usually a sync op (from fault)
1696  */
1697 
1698 int
1699 uvm_swap_get(struct vm_page *page, int swslot, int flags)
1700 {
1701 	int error;
1702 
1703 	uvmexp.nswget++;
1704 	KASSERT(flags & PGO_SYNCIO);
1705 	if (swslot == SWSLOT_BAD) {
1706 		return EIO;
1707 	}
1708 
1709 	error = uvm_swap_io(&page, swslot, 1, B_READ |
1710 	    ((flags & PGO_SYNCIO) ? 0 : B_ASYNC));
1711 	if (error == 0) {
1712 
1713 		/*
1714 		 * this page is no longer only in swap.
1715 		 */
1716 
1717 		mutex_enter(&uvm_swap_data_lock);
1718 		KASSERT(uvmexp.swpgonly > 0);
1719 		uvmexp.swpgonly--;
1720 		mutex_exit(&uvm_swap_data_lock);
1721 	}
1722 	return error;
1723 }
1724 
1725 /*
1726  * uvm_swap_io: do an i/o operation to swap
1727  */
1728 
1729 static int
1730 uvm_swap_io(struct vm_page **pps, int startslot, int npages, int flags)
1731 {
1732 	daddr_t startblk;
1733 	struct	buf *bp;
1734 	vaddr_t kva;
1735 	int	error, mapinflags;
1736 	bool write, async;
1737 	UVMHIST_FUNC("uvm_swap_io"); UVMHIST_CALLED(pdhist);
1738 
1739 	UVMHIST_LOG(pdhist, "<- called, startslot=%d, npages=%d, flags=%d",
1740 	    startslot, npages, flags, 0);
1741 
1742 	write = (flags & B_READ) == 0;
1743 	async = (flags & B_ASYNC) != 0;
1744 
1745 	/*
1746 	 * allocate a buf for the i/o.
1747 	 */
1748 
1749 	KASSERT(curlwp != uvm.pagedaemon_lwp || (write && async));
1750 	bp = getiobuf(swapdev_vp, curlwp != uvm.pagedaemon_lwp);
1751 	if (bp == NULL) {
1752 		uvm_aio_aiodone_pages(pps, npages, true, ENOMEM);
1753 		return ENOMEM;
1754 	}
1755 
1756 	/*
1757 	 * convert starting drum slot to block number
1758 	 */
1759 
1760 	startblk = btodb((uint64_t)startslot << PAGE_SHIFT);
1761 
1762 	/*
1763 	 * first, map the pages into the kernel.
1764 	 */
1765 
1766 	mapinflags = !write ?
1767 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_READ :
1768 		UVMPAGER_MAPIN_WAITOK|UVMPAGER_MAPIN_WRITE;
1769 	kva = uvm_pagermapin(pps, npages, mapinflags);
1770 
1771 	/*
1772 	 * fill in the bp/sbp.   we currently route our i/o through
1773 	 * /dev/drum's vnode [swapdev_vp].
1774 	 */
1775 
1776 	bp->b_cflags = BC_BUSY | BC_NOCACHE;
1777 	bp->b_flags = (flags & (B_READ|B_ASYNC));
1778 	bp->b_proc = &proc0;	/* XXX */
1779 	bp->b_vnbufs.le_next = NOLIST;
1780 	bp->b_data = (void *)kva;
1781 	bp->b_blkno = startblk;
1782 	bp->b_bufsize = bp->b_bcount = npages << PAGE_SHIFT;
1783 
1784 	/*
1785 	 * bump v_numoutput (counter of number of active outputs).
1786 	 */
1787 
1788 	if (write) {
1789 		mutex_enter(&swapdev_vp->v_interlock);
1790 		swapdev_vp->v_numoutput++;
1791 		mutex_exit(&swapdev_vp->v_interlock);
1792 	}
1793 
1794 	/*
1795 	 * for async ops we must set up the iodone handler.
1796 	 */
1797 
1798 	if (async) {
1799 		bp->b_iodone = uvm_aio_biodone;
1800 		UVMHIST_LOG(pdhist, "doing async!", 0, 0, 0, 0);
1801 		if (curlwp == uvm.pagedaemon_lwp)
1802 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1803 		else
1804 			BIO_SETPRIO(bp, BPRIO_TIMELIMITED);
1805 	} else {
1806 		bp->b_iodone = NULL;
1807 		BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
1808 	}
1809 	UVMHIST_LOG(pdhist,
1810 	    "about to start io: data = %p blkno = 0x%x, bcount = %ld",
1811 	    bp->b_data, bp->b_blkno, bp->b_bcount, 0);
1812 
1813 	/*
1814 	 * now we start the I/O, and if async, return.
1815 	 */
1816 
1817 	VOP_STRATEGY(swapdev_vp, bp);
1818 	if (async)
1819 		return 0;
1820 
1821 	/*
1822 	 * must be sync i/o.   wait for it to finish
1823 	 */
1824 
1825 	error = biowait(bp);
1826 
1827 	/*
1828 	 * kill the pager mapping
1829 	 */
1830 
1831 	uvm_pagermapout(kva, npages);
1832 
1833 	/*
1834 	 * now dispose of the buf and we're done.
1835 	 */
1836 
1837 	if (write) {
1838 		mutex_enter(&swapdev_vp->v_interlock);
1839 		vwakeup(bp);
1840 		mutex_exit(&swapdev_vp->v_interlock);
1841 	}
1842 	putiobuf(bp);
1843 	UVMHIST_LOG(pdhist, "<- done (sync)  error=%d", error, 0, 0, 0);
1844 
1845 	return (error);
1846 }
1847