xref: /netbsd-src/sys/kern/kern_physio.c (revision 481fca6e59249d8ffcf24fef7cfbe7b131bfb080)
1 /*	$NetBSD: kern_physio.c,v 1.43 2000/06/27 17:41:24 mrg Exp $	*/
2 
3 /*-
4  * Copyright (c) 1994 Christopher G. Demetriou
5  * Copyright (c) 1982, 1986, 1990, 1993
6  *	The Regents of the University of California.  All rights reserved.
7  * (c) UNIX System Laboratories, Inc.
8  * All or some portions of this file are derived from material licensed
9  * to the University of California by American Telephone and Telegraph
10  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
11  * the permission of UNIX System Laboratories, Inc.
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  * 3. All advertising materials mentioning features or use of this software
22  *    must display the following acknowledgement:
23  *	This product includes software developed by the University of
24  *	California, Berkeley and its contributors.
25  * 4. Neither the name of the University nor the names of its contributors
26  *    may be used to endorse or promote products derived from this software
27  *    without specific prior written permission.
28  *
29  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39  * SUCH DAMAGE.
40  *
41  *	@(#)kern_physio.c	8.1 (Berkeley) 6/10/93
42  */
43 
44 #include <sys/param.h>
45 #include <sys/systm.h>
46 #include <sys/buf.h>
47 #include <sys/malloc.h>
48 #include <sys/proc.h>
49 
50 #include <uvm/uvm_extern.h>
51 
52 /*
53  * The routines implemented in this file are described in:
54  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
55  *	    UNIX Operating System (Addison Welley, 1989)
56  * on pages 231-233.
57  *
58  * The routines "getphysbuf" and "putphysbuf" steal and return a swap
59  * buffer.  Leffler, et al., says that swap buffers are used to do the
60  * I/O, so raw I/O requests don't have to be single-threaded.
61  */
62 
63 struct buf *getphysbuf __P((void));
64 void putphysbuf __P((struct buf *bp));
65 
66 /*
67  * Do "physical I/O" on behalf of a user.  "Physical I/O" is I/O directly
68  * from the raw device to user buffers, and bypasses the buffer cache.
69  *
70  * Comments in brackets are from Leffler, et al.'s pseudo-code implementation.
71  */
72 int
73 physio(strategy, bp, dev, flags, minphys, uio)
74 	void (*strategy) __P((struct buf *));
75 	struct buf *bp;
76 	dev_t dev;
77 	int flags;
78 	void (*minphys) __P((struct buf *));
79 	struct uio *uio;
80 {
81 	struct iovec *iovp;
82 	struct proc *p = curproc;
83 	int error, done, i, nobuf, s, todo;
84 
85 	error = 0;
86 	flags &= B_READ | B_WRITE | B_ORDERED;
87 
88 	/*
89 	 * [check user read/write access to the data buffer]
90 	 *
91 	 * Check each iov one by one.  Note that we know if we're reading or
92 	 * writing, so we ignore the uio's rw parameter.  Also note that if
93 	 * we're doing a read, that's a *write* to user-space.
94 	 */
95 	if (uio->uio_segflg == UIO_USERSPACE) {
96 		for (i = 0; i < uio->uio_iovcnt; i++) {
97 			/* XXXCDC: map not locked, rethink */
98 			if (__predict_false(!uvm_useracc(uio->uio_iov[i].iov_base,
99 				     uio->uio_iov[i].iov_len,
100 				     (flags == B_READ) ? B_WRITE : B_READ)))
101 				return (EFAULT);
102 		}
103 	}
104 
105 	/* Make sure we have a buffer, creating one if necessary. */
106 	if ((nobuf = (bp == NULL)) != 0) {
107 
108 		bp = getphysbuf();
109 		/* bp was just malloc'd so can't already be busy */
110 		bp->b_flags |= B_BUSY;
111 
112 	} else {
113 
114 		/* [raise the processor priority level to splbio;] */
115 		s = splbio();
116 
117 		/* [while the buffer is marked busy] */
118 		while (bp->b_flags & B_BUSY) {
119 			/* [mark the buffer wanted] */
120 			bp->b_flags |= B_WANTED;
121 			/* [wait until the buffer is available] */
122 			tsleep((caddr_t)bp, PRIBIO+1, "physbuf", 0);
123 		}
124 
125 		/* Mark it busy, so nobody else will use it. */
126 		bp->b_flags |= B_BUSY;
127 
128 		/* [lower the priority level] */
129 		splx(s);
130 
131 	}
132 
133 	/* [set up the fixed part of the buffer for a transfer] */
134 	bp->b_dev = dev;
135 	bp->b_error = 0;
136 	bp->b_proc = p;
137 	LIST_INIT(&bp->b_dep);
138 
139 	/*
140 	 * [while there are data to transfer and no I/O error]
141 	 * Note that I/O errors are handled with a 'goto' at the bottom
142 	 * of the 'while' loop.
143 	 */
144 	for (i = 0; i < uio->uio_iovcnt; i++) {
145 		iovp = &uio->uio_iov[i];
146 		while (iovp->iov_len > 0) {
147 			/*
148 			 * [mark the buffer busy for physical I/O]
149 			 * (i.e. set B_PHYS (because it's an I/O to user
150 			 * memory, and B_RAW, because B_RAW is to be
151 			 * "Set by physio for raw transfers.", in addition
152 			 * to the "busy" and read/write flag.)
153 			 */
154 			bp->b_flags = B_BUSY | B_PHYS | B_RAW | flags;
155 
156 			/* [set up the buffer for a maximum-sized transfer] */
157 			bp->b_blkno = btodb(uio->uio_offset);
158 			bp->b_bcount = iovp->iov_len;
159 			bp->b_data = iovp->iov_base;
160 
161 			/*
162 			 * [call minphys to bound the tranfer size]
163 			 * and remember the amount of data to transfer,
164 			 * for later comparison.
165 			 */
166 			(*minphys)(bp);
167 			todo = bp->b_bcount;
168 #ifdef DIAGNOSTIC
169 			if (todo < 0)
170 				panic("todo < 0; minphys broken");
171 			if (todo > MAXPHYS)
172 				panic("todo > MAXPHYS; minphys broken");
173 #endif
174 
175 			/*
176 			 * [lock the part of the user address space involved
177 			 *    in the transfer]
178 			 * Beware vmapbuf(); it clobbers b_data and
179 			 * saves it in b_saveaddr.  However, vunmapbuf()
180 			 * restores it.
181 			 */
182 			PHOLD(p);
183 			if (__predict_false(uvm_vslock(p, bp->b_data, todo,
184 			    (flags & B_READ) ?
185 			    VM_PROT_READ | VM_PROT_WRITE : VM_PROT_READ)
186 			    != KERN_SUCCESS)) {
187 				bp->b_flags |= B_ERROR;
188 				bp->b_error = EFAULT;
189 				goto after_vsunlock;
190 			}
191 			vmapbuf(bp, todo);
192 
193 			/* [call strategy to start the transfer] */
194 			(*strategy)(bp);
195 
196 			/*
197 			 * Note that the raise/wait/lower/get error
198 			 * steps below would be done by biowait(), but
199 			 * we want to unlock the address space before
200 			 * we lower the priority.
201 			 *
202 			 * [raise the priority level to splbio]
203 			 */
204 			s = splbio();
205 
206 			/* [wait for the transfer to complete] */
207 			while ((bp->b_flags & B_DONE) == 0)
208 				tsleep((caddr_t) bp, PRIBIO + 1, "physio", 0);
209 
210 			/* Mark it busy again, so nobody else will use it. */
211 			bp->b_flags |= B_BUSY;
212 
213 			/* [lower the priority level] */
214 			splx(s);
215 
216 			/*
217 			 * [unlock the part of the address space previously
218 			 *    locked]
219 			 */
220 			vunmapbuf(bp, todo);
221 			uvm_vsunlock(p, bp->b_data, todo);
222  after_vsunlock:
223 			PRELE(p);
224 
225 			/* remember error value (save a splbio/splx pair) */
226 			if (bp->b_flags & B_ERROR)
227 				error = (bp->b_error ? bp->b_error : EIO);
228 
229 			/*
230 			 * [deduct the transfer size from the total number
231 			 *    of data to transfer]
232 			 */
233 			done = bp->b_bcount - bp->b_resid;
234 #ifdef DIAGNOSTIC
235 			if (__predict_false(done < 0))
236 				panic("done < 0; strategy broken");
237 			if (__predict_false(done > todo))
238 				panic("done > todo; strategy broken");
239 #endif
240 			iovp->iov_len -= done;
241 			iovp->iov_base = (caddr_t)iovp->iov_base + done;
242 			uio->uio_offset += done;
243 			uio->uio_resid -= done;
244 
245 			/*
246 			 * Now, check for an error.
247 			 * Also, handle weird end-of-disk semantics.
248 			 */
249 			if (error || done < todo)
250 				goto done;
251 		}
252 	}
253 
254 done:
255 	/*
256 	 * [clean up the state of the buffer]
257 	 * Remember if somebody wants it, so we can wake them up below.
258 	 * Also, if we had to steal it, give it back.
259 	 */
260 	s = splbio();
261 	bp->b_flags &= ~(B_BUSY | B_PHYS | B_RAW);
262 	if (nobuf)
263 		putphysbuf(bp);
264 	else {
265 		/*
266 		 * [if another process is waiting for the raw I/O buffer,
267 		 *    wake up processes waiting to do physical I/O;
268 		 */
269 		if (bp->b_flags & B_WANTED) {
270 			bp->b_flags &= ~B_WANTED;
271 			wakeup(bp);
272 		}
273 	}
274 	splx(s);
275 
276 	return (error);
277 }
278 
279 /*
280  * allocate a buffer structure for use in physical I/O.
281  */
282 struct buf *
283 getphysbuf()
284 {
285 	struct buf *bp;
286 	int s;
287 
288 	s = splbio();
289 	bp = pool_get(&bufpool, PR_WAITOK);
290 	splx(s);
291 	memset(bp, 0, sizeof(*bp));
292 
293 	/* XXXCDC: are the following two lines necessary? */
294 	bp->b_rcred = bp->b_wcred = NOCRED;
295 	bp->b_vnbufs.le_next = NOLIST;
296 
297 	return(bp);
298 }
299 
300 /*
301  * get rid of a swap buffer structure which has been used in physical I/O.
302  */
303 void
304 putphysbuf(bp)
305         struct buf *bp;
306 {
307 	int s;
308 
309 	/* XXXCDC: is this necesary? */
310 	if (bp->b_vp)
311 		brelvp(bp);
312 
313 	if (__predict_false(bp->b_flags & B_WANTED))
314 		panic("putphysbuf: private buf B_WANTED");
315 	s = splbio();
316 	pool_put(&bufpool, bp);
317 	splx(s);
318 }
319 
320 /*
321  * Leffler, et al., says on p. 231:
322  * "The minphys() routine is called by physio() to adjust the
323  * size of each I/O transfer before the latter is passed to
324  * the strategy routine..."
325  *
326  * so, just adjust the buffer's count accounting to MAXPHYS here,
327  * and return the new count;
328  */
329 void
330 minphys(bp)
331 	struct buf *bp;
332 {
333 
334 	if (bp->b_bcount > MAXPHYS)
335 		bp->b_bcount = MAXPHYS;
336 }
337