xref: /netbsd-src/sys/kern/kern_physio.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: kern_physio.c,v 1.85 2007/11/06 00:42:42 ad Exp $	*/
2 
3 /*-
4  * Copyright (c) 1982, 1986, 1990, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  *
36  *	@(#)kern_physio.c	8.1 (Berkeley) 6/10/93
37  */
38 
39 /*-
40  * Copyright (c) 1994 Christopher G. Demetriou
41  *
42  * Redistribution and use in source and binary forms, with or without
43  * modification, are permitted provided that the following conditions
44  * are met:
45  * 1. Redistributions of source code must retain the above copyright
46  *    notice, this list of conditions and the following disclaimer.
47  * 2. Redistributions in binary form must reproduce the above copyright
48  *    notice, this list of conditions and the following disclaimer in the
49  *    documentation and/or other materials provided with the distribution.
50  * 3. All advertising materials mentioning features or use of this software
51  *    must display the following acknowledgement:
52  *	This product includes software developed by the University of
53  *	California, Berkeley and its contributors.
54  * 4. Neither the name of the University nor the names of its contributors
55  *    may be used to endorse or promote products derived from this software
56  *    without specific prior written permission.
57  *
58  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
59  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
60  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
61  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
62  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
63  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
64  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
65  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
66  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
67  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
68  * SUCH DAMAGE.
69  *
70  *	@(#)kern_physio.c	8.1 (Berkeley) 6/10/93
71  */
72 
73 #include <sys/cdefs.h>
74 __KERNEL_RCSID(0, "$NetBSD: kern_physio.c,v 1.85 2007/11/06 00:42:42 ad Exp $");
75 
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/buf.h>
79 #include <sys/proc.h>
80 #include <sys/once.h>
81 #include <sys/workqueue.h>
82 #include <sys/kmem.h>
83 
84 #include <uvm/uvm_extern.h>
85 
86 ONCE_DECL(physio_initialized);
87 struct workqueue *physio_workqueue;
88 
89 /*
90  * The routines implemented in this file are described in:
91  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
92  *	    UNIX Operating System (Addison Welley, 1989)
93  * on pages 231-233.
94  *
95  * The routines "getphysbuf" and "putphysbuf" steal and return a swap
96  * buffer.  Leffler, et al., says that swap buffers are used to do the
97  * I/O, so raw I/O requests don't have to be single-threaded.  Of course,
98  * NetBSD doesn't use "swap buffers" -- we have our own memory pool for
99  * buffer descriptors.
100  */
101 
102 /* #define	PHYSIO_DEBUG */
103 #if defined(PHYSIO_DEBUG)
104 #define	DPRINTF(a)	printf a
105 #else /* defined(PHYSIO_DEBUG) */
106 #define	DPRINTF(a)	/* nothing */
107 #endif /* defined(PHYSIO_DEBUG) */
108 
109 struct physio_stat {
110 	int ps_running;
111 	int ps_error;
112 	int ps_failed;
113 	off_t ps_endoffset;
114 	kmutex_t ps_lock;
115 	kcondvar_t ps_cv;
116 };
117 
118 /* abuse these flags of struct buf */
119 #define	B_DONTFREE	B_AGE
120 
121 /*
122  * allocate a buffer structure for use in physical I/O.
123  */
124 static struct buf *
125 getphysbuf(void)
126 {
127 	struct buf *bp;
128 
129 	bp = getiobuf();
130 	bp->b_error = 0;
131 	bp->b_flags = B_BUSY;
132 	return(bp);
133 }
134 
135 /*
136  * get rid of a swap buffer structure which has been used in physical I/O.
137  */
138 static void
139 putphysbuf(struct buf *bp)
140 {
141 
142 	if ((bp->b_flags & B_DONTFREE) != 0) {
143 		return;
144 	}
145 
146 	if (__predict_false(bp->b_flags & B_WANTED))
147 		panic("putphysbuf: private buf B_WANTED");
148 	putiobuf(bp);
149 }
150 
151 static void
152 physio_done(struct work *wk, void *dummy)
153 {
154 	struct buf *bp = (void *)wk;
155 	size_t todo = bp->b_bufsize;
156 	size_t done = bp->b_bcount - bp->b_resid;
157 	struct physio_stat *ps = bp->b_private;
158 
159 	KASSERT(&bp->b_work == wk);
160 	KASSERT(bp->b_bcount <= todo);
161 	KASSERT(bp->b_resid <= bp->b_bcount);
162 	KASSERT((bp->b_flags & B_PHYS) != 0);
163 	KASSERT(dummy == NULL);
164 
165 	vunmapbuf(bp, todo);
166 	uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo);
167 
168 	mutex_enter(&ps->ps_lock);
169 	if (__predict_false(done != todo)) {
170 		off_t endoffset = dbtob(bp->b_blkno) + done;
171 
172 		/*
173 		 * we got an error or hit EOM.
174 		 *
175 		 * we only care about the first one.
176 		 * ie. the one at the lowest offset.
177 		 */
178 
179 		KASSERT(ps->ps_endoffset != endoffset);
180 		DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64
181 		    ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n",
182 		    __func__, bp->b_error, dbtob(bp->b_blkno), endoffset,
183 		    bp->b_blkno, bp->b_bcount, bp->b_flags));
184 
185 		if (ps->ps_endoffset == -1 || endoffset < ps->ps_endoffset) {
186 			DPRINTF(("%s: ps=%p, error %d -> %d, endoff %" PRIu64
187 			    " -> %" PRIu64 "\n",
188 			    __func__, ps,
189 			    ps->ps_error, bp->b_error,
190 			    ps->ps_endoffset, endoffset));
191 
192 			ps->ps_endoffset = endoffset;
193 			ps->ps_error = bp->b_error;
194 		}
195 		ps->ps_failed++;
196 	} else {
197 		KASSERT(bp->b_error == 0);
198 	}
199 
200 	ps->ps_running--;
201 	cv_signal(&ps->ps_cv);
202 	mutex_exit(&ps->ps_lock);
203 
204 	putphysbuf(bp);
205 }
206 
207 static void
208 physio_biodone(struct buf *bp)
209 {
210 #if defined(DIAGNOSTIC)
211 	struct physio_stat *ps = bp->b_private;
212 	size_t todo = bp->b_bufsize;
213 
214 	KASSERT(ps->ps_running > 0);
215 	KASSERT(bp->b_bcount <= todo);
216 	KASSERT(bp->b_resid <= bp->b_bcount);
217 #endif /* defined(DIAGNOSTIC) */
218 
219 	workqueue_enqueue(physio_workqueue, &bp->b_work, NULL);
220 }
221 
222 static void
223 physio_wait(struct physio_stat *ps, int n)
224 {
225 
226 	KASSERT(mutex_owned(&ps->ps_lock));
227 
228 	while (ps->ps_running > n)
229 		cv_wait(&ps->ps_cv, &ps->ps_lock);
230 }
231 
232 static int
233 physio_init(void)
234 {
235 	int error;
236 
237 	KASSERT(physio_workqueue == NULL);
238 
239 	error = workqueue_create(&physio_workqueue, "physiod",
240 	    physio_done, NULL, PRI_BIO, IPL_BIO, 0);
241 
242 	return error;
243 }
244 
245 #define	PHYSIO_CONCURRENCY	16	/* XXX tune */
246 
247 /*
248  * Do "physical I/O" on behalf of a user.  "Physical I/O" is I/O directly
249  * from the raw device to user buffers, and bypasses the buffer cache.
250  *
251  * Comments in brackets are from Leffler, et al.'s pseudo-code implementation.
252  */
253 int
254 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags,
255     void (*min_phys)(struct buf *), struct uio *uio)
256 {
257 	struct iovec *iovp;
258 	struct lwp *l = curlwp;
259 	struct proc *p = l->l_proc;
260 	int i, s;
261 	int error;
262 	struct buf *bp = NULL;
263 	struct physio_stat *ps;
264 	int concurrency = PHYSIO_CONCURRENCY - 1;
265 
266 	error = RUN_ONCE(&physio_initialized, physio_init);
267 	if (__predict_false(error != 0)) {
268 		return error;
269 	}
270 
271 	DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n",
272 	    __func__, uio->uio_offset, uio->uio_resid));
273 
274 	flags &= B_READ | B_WRITE;
275 
276 	if ((ps = kmem_zalloc(sizeof(*ps), KM_SLEEP)) == NULL)
277 		return ENOMEM;
278 	/* ps->ps_running = 0; */
279 	/* ps->ps_error = 0; */
280 	/* ps->ps_failed = 0; */
281 	ps->ps_endoffset = -1;
282 	mutex_init(&ps->ps_lock, MUTEX_DEFAULT, IPL_NONE);
283 	cv_init(&ps->ps_cv, "physio");
284 
285 	/* Make sure we have a buffer, creating one if necessary. */
286 	if (obp != NULL) {
287 		/* [raise the processor priority level to splbio;] */
288 		s = splbio();
289 		simple_lock(&obp->b_interlock);
290 
291 		/* [while the buffer is marked busy] */
292 		while (obp->b_flags & B_BUSY) {
293 			/* [mark the buffer wanted] */
294 			obp->b_flags |= B_WANTED;
295 			/* [wait until the buffer is available] */
296 			ltsleep(obp, PRIBIO+1, "physbuf", 0, &obp->b_interlock);
297 		}
298 
299 		/* Mark it busy, so nobody else will use it. */
300 		obp->b_flags = B_BUSY | B_DONTFREE;
301 
302 		/* [lower the priority level] */
303 		simple_unlock(&obp->b_interlock);
304 		splx(s);
305 
306 		concurrency = 0; /* see "XXXkludge" comment below */
307 	}
308 
309 	uvm_lwp_hold(l);
310 
311 	for (i = 0; i < uio->uio_iovcnt; i++) {
312 		bool sync = true;
313 
314 		iovp = &uio->uio_iov[i];
315 		while (iovp->iov_len > 0) {
316 			size_t todo;
317 			vaddr_t endp;
318 
319 			mutex_enter(&ps->ps_lock);
320 			if (ps->ps_failed != 0) {
321 				goto done_locked;
322 			}
323 			physio_wait(ps, sync ? 0 : concurrency);
324 			mutex_exit(&ps->ps_lock);
325 			if (obp != NULL) {
326 				/*
327 				 * XXXkludge
328 				 * some drivers use "obp" as an identifier.
329 				 */
330 				bp = obp;
331 			} else {
332 				bp = getphysbuf();
333 			}
334 			bp->b_dev = dev;
335 			bp->b_proc = p;
336 			bp->b_private = ps;
337 			bp->b_vp = NULL;
338 
339 			/*
340 			 * [mark the buffer busy for physical I/O]
341 			 * (i.e. set B_PHYS (because it's an I/O to user
342 			 * memory, and B_RAW, because B_RAW is to be
343 			 * "Set by physio for raw transfers.", in addition
344 			 * to the "busy" and read/write flag.)
345 			 */
346 			bp->b_flags = (bp->b_flags & B_DONTFREE) |
347 			    B_BUSY | B_PHYS | B_RAW | B_CALL | flags;
348 			bp->b_iodone = physio_biodone;
349 
350 			/* [set up the buffer for a maximum-sized transfer] */
351 			bp->b_blkno = btodb(uio->uio_offset);
352 			if (dbtob(bp->b_blkno) != uio->uio_offset) {
353 				error = EINVAL;
354 				goto done;
355 			}
356 			bp->b_bcount = MIN(MAXPHYS, iovp->iov_len);
357 			bp->b_data = iovp->iov_base;
358 
359 			/*
360 			 * [call minphys to bound the transfer size]
361 			 * and remember the amount of data to transfer,
362 			 * for later comparison.
363 			 */
364 			(*min_phys)(bp);
365 			todo = bp->b_bufsize = bp->b_bcount;
366 #if defined(DIAGNOSTIC)
367 			if (todo > MAXPHYS)
368 				panic("todo(%zu) > MAXPHYS; minphys broken",
369 				    todo);
370 #endif /* defined(DIAGNOSTIC) */
371 
372 			sync = false;
373 			endp = (vaddr_t)bp->b_data + todo;
374 			if (trunc_page(endp) != endp) {
375 				/*
376 				 * following requests can overlap.
377 				 * note that uvm_vslock does round_page.
378 				 */
379 				sync = true;
380 			}
381 
382 			/*
383 			 * [lock the part of the user address space involved
384 			 *    in the transfer]
385 			 * Beware vmapbuf(); it clobbers b_data and
386 			 * saves it in b_saveaddr.  However, vunmapbuf()
387 			 * restores it.
388 			 */
389 			error = uvm_vslock(p->p_vmspace, bp->b_data, todo,
390 			    (flags & B_READ) ?  VM_PROT_WRITE : VM_PROT_READ);
391 			if (error) {
392 				goto done;
393 			}
394 			vmapbuf(bp, todo);
395 
396 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
397 
398 			mutex_enter(&ps->ps_lock);
399 			ps->ps_running++;
400 			mutex_exit(&ps->ps_lock);
401 
402 			/* [call strategy to start the transfer] */
403 			(*strategy)(bp);
404 			bp = NULL;
405 
406 			iovp->iov_len -= todo;
407 			iovp->iov_base = (char *)iovp->iov_base + todo;
408 			uio->uio_offset += todo;
409 			uio->uio_resid -= todo;
410 		}
411 	}
412 
413 done:
414 	mutex_enter(&ps->ps_lock);
415 done_locked:
416 	physio_wait(ps, 0);
417 	mutex_exit(&ps->ps_lock);
418 
419 	if (ps->ps_failed != 0) {
420 		off_t delta;
421 
422 		delta = uio->uio_offset - ps->ps_endoffset;
423 		KASSERT(delta > 0);
424 		uio->uio_resid += delta;
425 		/* uio->uio_offset = ps->ps_endoffset; */
426 	} else {
427 		KASSERT(ps->ps_endoffset == -1);
428 	}
429 	if (bp != NULL) {
430 		putphysbuf(bp);
431 	}
432 	if (error == 0) {
433 		error = ps->ps_error;
434 	}
435 	mutex_destroy(&ps->ps_lock);
436 	cv_destroy(&ps->ps_cv);
437 	kmem_free(ps, sizeof(*ps));
438 
439 	/*
440 	 * [clean up the state of the buffer]
441 	 * Remember if somebody wants it, so we can wake them up below.
442 	 * Also, if we had to steal it, give it back.
443 	 */
444 	if (obp != NULL) {
445 		KASSERT((obp->b_flags & B_BUSY) != 0);
446 		KASSERT((obp->b_flags & B_DONTFREE) != 0);
447 
448 		/*
449 		 * [if another process is waiting for the raw I/O buffer,
450 		 *    wake up processes waiting to do physical I/O;
451 		 */
452 		s = splbio();
453 		simple_lock(&obp->b_interlock);
454 		obp->b_flags &=
455 		    ~(B_BUSY | B_PHYS | B_RAW | B_CALL | B_DONTFREE);
456 		if ((obp->b_flags & B_WANTED) != 0) {
457 			obp->b_flags &= ~B_WANTED;
458 			wakeup(obp);
459 		}
460 		simple_unlock(&obp->b_interlock);
461 		splx(s);
462 	}
463 	uvm_lwp_rele(l);
464 
465 	DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n",
466 	    __func__, uio->uio_offset, uio->uio_resid));
467 
468 	return error;
469 }
470 
471 /*
472  * Leffler, et al., says on p. 231:
473  * "The minphys() routine is called by physio() to adjust the
474  * size of each I/O transfer before the latter is passed to
475  * the strategy routine..."
476  *
477  * so, just adjust the buffer's count accounting to MAXPHYS here,
478  * and return the new count;
479  */
480 void
481 minphys(struct buf *bp)
482 {
483 
484 	if (bp->b_bcount > MAXPHYS)
485 		bp->b_bcount = MAXPHYS;
486 }
487