xref: /netbsd-src/sys/kern/kern_physio.c (revision d48f14661dda8638fee055ba15d35bdfb29b9fa8)
1 /*	$NetBSD: kern_physio.c,v 1.73 2006/04/18 09:54:32 yamt 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 1.73 2006/04/18 09:54:32 yamt 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 
83 #include <uvm/uvm_extern.h>
84 
85 ONCE_DECL(physio_initialized);
86 struct workqueue *physio_workqueue;
87 
88 /*
89  * The routines implemented in this file are described in:
90  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
91  *	    UNIX Operating System (Addison Welley, 1989)
92  * on pages 231-233.
93  *
94  * The routines "getphysbuf" and "putphysbuf" steal and return a swap
95  * buffer.  Leffler, et al., says that swap buffers are used to do the
96  * I/O, so raw I/O requests don't have to be single-threaded.  Of course,
97  * NetBSD doesn't use "swap buffers" -- we have our own memory pool for
98  * buffer descriptors.
99  */
100 
101 /* #define	PHYSIO_DEBUG */
102 #if defined(PHYSIO_DEBUG)
103 #define	DPRINTF(a)	printf a
104 #else /* defined(PHYSIO_DEBUG) */
105 #define	DPRINTF(a)	/* nothing */
106 #endif /* defined(PHYSIO_DEBUG) */
107 
108 /* abuse these members/flags of struct buf */
109 #define	b_running	b_freelistindex
110 #define	b_endoffset	b_lblkno
111 #define	B_DONTFREE	B_AGE
112 
113 /*
114  * allocate a buffer structure for use in physical I/O.
115  */
116 static struct buf *
117 getphysbuf(void)
118 {
119 	struct buf *bp;
120 
121 	bp = getiobuf();
122 	bp->b_error = 0;
123 	bp->b_flags = B_BUSY;
124 	return(bp);
125 }
126 
127 /*
128  * get rid of a swap buffer structure which has been used in physical I/O.
129  */
130 static void
131 putphysbuf(struct buf *bp)
132 {
133 
134 	if ((bp->b_flags & B_DONTFREE) != 0) {
135 		return;
136 	}
137 
138 	if (__predict_false(bp->b_flags & B_WANTED))
139 		panic("putphysbuf: private buf B_WANTED");
140 	putiobuf(bp);
141 }
142 
143 static void
144 physio_done(struct work *wk, void *dummy)
145 {
146 	struct buf *bp = (void *)wk;
147 	size_t todo = bp->b_bufsize;
148 	size_t done = bp->b_bcount - bp->b_resid;
149 	struct buf *mbp = bp->b_private;
150 
151 	KASSERT(&bp->b_work == wk);
152 	KASSERT(bp->b_bcount <= todo);
153 	KASSERT(bp->b_resid <= bp->b_bcount);
154 	KASSERT((bp->b_flags & B_PHYS) != 0);
155 	KASSERT(dummy == NULL);
156 
157 	vunmapbuf(bp, todo);
158 	uvm_vsunlock(bp->b_proc, bp->b_data, todo);
159 
160 	simple_lock(&mbp->b_interlock);
161 	if (__predict_false(done != todo)) {
162 		off_t endoffset = dbtob(bp->b_blkno) + done;
163 
164 		/*
165 		 * we got an error or hit EOM.
166 		 *
167 		 * we only care about the first one.
168 		 * ie. the one at the lowest offset.
169 		 */
170 
171 		KASSERT(mbp->b_endoffset != endoffset);
172 		DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64
173 		    ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n",
174 		    __func__, bp->b_error, dbtob(bp->b_blkno), endoffset,
175 		    bp->b_blkno, bp->b_bcount, bp->b_flags));
176 
177 		if (mbp->b_endoffset == -1 || endoffset < mbp->b_endoffset) {
178 			int error;
179 
180 			if ((bp->b_flags & B_ERROR) != 0) {
181 				if (bp->b_error == 0) {
182 					error = EIO; /* XXX */
183 				} else {
184 					error = bp->b_error;
185 				}
186 			} else {
187 				error = 0; /* EOM */
188 			}
189 
190 			DPRINTF(("%s: mbp=%p, error %d -> %d, endoff %" PRIu64
191 			    " -> %" PRIu64 "\n",
192 			    __func__, mbp,
193 			    mbp->b_error, error,
194 			    mbp->b_endoffset, endoffset));
195 
196 			mbp->b_endoffset = endoffset;
197 			mbp->b_error = error;
198 		}
199 		mbp->b_flags |= B_ERROR;
200 	} else {
201 		KASSERT((bp->b_flags & B_ERROR) == 0);
202 	}
203 
204 	mbp->b_running--;
205 	if ((mbp->b_flags & B_WANTED) != 0) {
206 		mbp->b_flags &= ~B_WANTED;
207 		wakeup(mbp);
208 	}
209 	simple_unlock(&mbp->b_interlock);
210 
211 	putphysbuf(bp);
212 }
213 
214 static void
215 physio_biodone(struct buf *bp)
216 {
217 #if defined(DIAGNOSTIC)
218 	struct buf *mbp = bp->b_private;
219 	size_t todo = bp->b_bufsize;
220 
221 	KASSERT(mbp->b_running > 0);
222 	KASSERT(bp->b_bcount <= todo);
223 	KASSERT(bp->b_resid <= bp->b_bcount);
224 #endif /* defined(DIAGNOSTIC) */
225 
226 	workqueue_enqueue(physio_workqueue, &bp->b_work);
227 }
228 
229 static int
230 physio_wait(struct buf *bp, int n, const char *wchan)
231 {
232 	int error = 0;
233 
234 	LOCK_ASSERT(simple_lock_held(&bp->b_interlock));
235 
236 	while (bp->b_running > n) {
237 		bp->b_flags |= B_WANTED;
238 		error = ltsleep(bp, PRIBIO + 1, wchan, 0, &bp->b_interlock);
239 		if (error) {
240 			break;
241 		}
242 	}
243 
244 	return error;
245 }
246 
247 static int
248 physio_init(void)
249 {
250 	int error;
251 
252 	KASSERT(physio_workqueue == NULL);
253 
254 	error = workqueue_create(&physio_workqueue, "physiod",
255 	    physio_done, NULL, PRIBIO, 0/* IPL_BIO notyet */, 0);
256 
257 	return error;
258 }
259 
260 #define	PHYSIO_CONCURRENCY	16	/* XXX tune */
261 
262 /*
263  * Do "physical I/O" on behalf of a user.  "Physical I/O" is I/O directly
264  * from the raw device to user buffers, and bypasses the buffer cache.
265  *
266  * Comments in brackets are from Leffler, et al.'s pseudo-code implementation.
267  */
268 int
269 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags,
270     void (*min_phys)(struct buf *), struct uio *uio)
271 {
272 	struct iovec *iovp;
273 	struct lwp *l = curlwp;
274 	struct proc *p = l->l_proc;
275 	int i, s;
276 	int error;
277 	int error2;
278 	struct buf *bp = NULL;
279 	struct buf *mbp;
280 	int concurrency = PHYSIO_CONCURRENCY - 1;
281 
282 	error = RUN_ONCE(&physio_initialized, physio_init);
283 	if (__predict_false(error != 0)) {
284 		return error;
285 	}
286 
287 	DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n",
288 	    __func__, uio->uio_offset, uio->uio_resid));
289 
290 	flags &= B_READ | B_WRITE;
291 
292 	/* Make sure we have a buffer, creating one if necessary. */
293 	if (obp != NULL) {
294 		/* [raise the processor priority level to splbio;] */
295 		s = splbio();
296 		simple_lock(&obp->b_interlock);
297 
298 		/* [while the buffer is marked busy] */
299 		while (obp->b_flags & B_BUSY) {
300 			/* [mark the buffer wanted] */
301 			obp->b_flags |= B_WANTED;
302 			/* [wait until the buffer is available] */
303 			ltsleep(obp, PRIBIO+1, "physbuf", 0, &obp->b_interlock);
304 		}
305 
306 		/* Mark it busy, so nobody else will use it. */
307 		obp->b_flags = B_BUSY | B_DONTFREE;
308 
309 		/* [lower the priority level] */
310 		simple_unlock(&obp->b_interlock);
311 		splx(s);
312 
313 		concurrency = 0; /* see "XXXkludge" comment below */
314 	}
315 
316 	mbp = getphysbuf();
317 	mbp->b_running = 0;
318 	mbp->b_endoffset = -1;
319 
320 	PHOLD(l);
321 
322 	for (i = 0; i < uio->uio_iovcnt; i++) {
323 		boolean_t sync = TRUE;
324 
325 		iovp = &uio->uio_iov[i];
326 		while (iovp->iov_len > 0) {
327 			size_t todo;
328 			vaddr_t endp;
329 
330 			simple_lock(&mbp->b_interlock);
331 			if ((mbp->b_flags & B_ERROR) != 0) {
332 				goto done_locked;
333 			}
334 			error = physio_wait(mbp, sync ? 0 : concurrency,
335 			    "physio1");
336 			if (error) {
337 				goto done_locked;
338 			}
339 			simple_unlock(&mbp->b_interlock);
340 			if (obp != NULL) {
341 				/*
342 				 * XXXkludge
343 				 * some drivers use "obp" as an identifier.
344 				 */
345 				bp = obp;
346 			} else {
347 				bp = getphysbuf();
348 			}
349 			bp->b_dev = dev;
350 			bp->b_proc = p;
351 			bp->b_private = mbp;
352 			bp->b_vp = NULL;
353 
354 			/*
355 			 * [mark the buffer busy for physical I/O]
356 			 * (i.e. set B_PHYS (because it's an I/O to user
357 			 * memory, and B_RAW, because B_RAW is to be
358 			 * "Set by physio for raw transfers.", in addition
359 			 * to the "busy" and read/write flag.)
360 			 */
361 			bp->b_flags = (bp->b_flags & B_DONTFREE) |
362 			    B_BUSY | B_PHYS | B_RAW | B_CALL | flags;
363 			bp->b_iodone = physio_biodone;
364 
365 			/* [set up the buffer for a maximum-sized transfer] */
366 			bp->b_blkno = btodb(uio->uio_offset);
367 			if (dbtob(bp->b_blkno) != uio->uio_offset) {
368 				error = EINVAL;
369 				goto done;
370 			}
371 			bp->b_bcount = MIN(MAXPHYS, iovp->iov_len);
372 			bp->b_data = iovp->iov_base;
373 
374 			/*
375 			 * [call minphys to bound the transfer size]
376 			 * and remember the amount of data to transfer,
377 			 * for later comparison.
378 			 */
379 			(*min_phys)(bp);
380 			todo = bp->b_bufsize = bp->b_bcount;
381 #if defined(DIAGNOSTIC)
382 			if (todo > MAXPHYS)
383 				panic("todo(%zu) > MAXPHYS; minphys broken",
384 				    todo);
385 #endif /* defined(DIAGNOSTIC) */
386 
387 			sync = FALSE;
388 			endp = (vaddr_t)bp->b_data + todo;
389 			if (trunc_page(endp) != endp) {
390 				/*
391 				 * following requests can overlap.
392 				 * note that uvm_vslock does round_page.
393 				 */
394 				sync = TRUE;
395 			}
396 
397 			/*
398 			 * [lock the part of the user address space involved
399 			 *    in the transfer]
400 			 * Beware vmapbuf(); it clobbers b_data and
401 			 * saves it in b_saveaddr.  However, vunmapbuf()
402 			 * restores it.
403 			 */
404 			error = uvm_vslock(p, bp->b_data, todo,
405 			    (flags & B_READ) ?  VM_PROT_WRITE : VM_PROT_READ);
406 			if (error) {
407 				goto done;
408 			}
409 			vmapbuf(bp, todo);
410 
411 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
412 
413 			simple_lock(&mbp->b_interlock);
414 			mbp->b_running++;
415 			simple_unlock(&mbp->b_interlock);
416 
417 			/* [call strategy to start the transfer] */
418 			(*strategy)(bp);
419 			bp = NULL;
420 
421 			iovp->iov_len -= todo;
422 			iovp->iov_base = (caddr_t)iovp->iov_base + todo;
423 			uio->uio_offset += todo;
424 			uio->uio_resid -= todo;
425 		}
426 	}
427 
428 done:
429 	simple_lock(&mbp->b_interlock);
430 done_locked:
431 	error2 = physio_wait(mbp, 0, "physio2");
432 	if (error == 0) {
433 		error = error2;
434 	}
435 	simple_unlock(&mbp->b_interlock);
436 
437 	if ((mbp->b_flags & B_ERROR) != 0) {
438 		off_t delta;
439 
440 		delta = uio->uio_offset - mbp->b_endoffset;
441 		KASSERT(delta > 0);
442 		uio->uio_resid += delta;
443 		/* uio->uio_offset = mbp->b_endoffset; */
444 	} else {
445 		KASSERT(mbp->b_endoffset == -1);
446 	}
447 	if (bp != NULL) {
448 		putphysbuf(bp);
449 	}
450 	if (error == 0) {
451 		error = mbp->b_error;
452 	}
453 	putphysbuf(mbp);
454 
455 	/*
456 	 * [clean up the state of the buffer]
457 	 * Remember if somebody wants it, so we can wake them up below.
458 	 * Also, if we had to steal it, give it back.
459 	 */
460 	if (obp != NULL) {
461 		KASSERT((obp->b_flags & B_BUSY) != 0);
462 		KASSERT((obp->b_flags & B_DONTFREE) != 0);
463 
464 		/*
465 		 * [if another process is waiting for the raw I/O buffer,
466 		 *    wake up processes waiting to do physical I/O;
467 		 */
468 		s = splbio();
469 		simple_lock(&obp->b_interlock);
470 		obp->b_flags &=
471 		    ~(B_BUSY | B_PHYS | B_RAW | B_CALL | B_DONTFREE);
472 		if ((obp->b_flags & B_WANTED) != 0) {
473 			obp->b_flags &= ~B_WANTED;
474 			wakeup(obp);
475 		}
476 		simple_unlock(&obp->b_interlock);
477 		splx(s);
478 	}
479 	PRELE(l);
480 
481 	DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n",
482 	    __func__, uio->uio_offset, uio->uio_resid));
483 
484 	return error;
485 }
486 
487 /*
488  * Leffler, et al., says on p. 231:
489  * "The minphys() routine is called by physio() to adjust the
490  * size of each I/O transfer before the latter is passed to
491  * the strategy routine..."
492  *
493  * so, just adjust the buffer's count accounting to MAXPHYS here,
494  * and return the new count;
495  */
496 void
497 minphys(struct buf *bp)
498 {
499 
500 	if (bp->b_bcount > MAXPHYS)
501 		bp->b_bcount = MAXPHYS;
502 }
503