Home | History | Annotate | Line # | Download | only in kern
kern_physio.c revision 1.30
      1 /*	$NetBSD: kern_physio.c,v 1.30 1998/02/10 14:09:35 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 "opt_uvm.h"
     45 
     46 #include <sys/param.h>
     47 #include <sys/systm.h>
     48 #include <sys/buf.h>
     49 #include <sys/conf.h>
     50 #include <sys/malloc.h>
     51 #include <sys/proc.h>
     52 
     53 #include <vm/vm.h>
     54 
     55 #if defined(UVM)
     56 #include <uvm/uvm_extern.h>
     57 #endif
     58 
     59 /*
     60  * The routines implemented in this file are described in:
     61  *	Leffler, et al.: The Design and Implementation of the 4.3BSD
     62  *	    UNIX Operating System (Addison Welley, 1989)
     63  * on pages 231-233.
     64  *
     65  * The routines "getphysbuf" and "putphysbuf" steal and return a swap
     66  * buffer.  Leffler, et al., says that swap buffers are used to do the
     67  * I/O, so raw I/O requests don't have to be single-threaded.
     68  */
     69 
     70 struct buf *getphysbuf __P((void));
     71 void putphysbuf __P((struct buf *bp));
     72 
     73 /*
     74  * Do "physical I/O" on behalf of a user.  "Physical I/O" is I/O directly
     75  * from the raw device to user buffers, and bypasses the buffer cache.
     76  *
     77  * Comments in brackets are from Leffler, et al.'s pseudo-code implementation.
     78  */
     79 int
     80 physio(strategy, bp, dev, flags, minphys, uio)
     81 	void (*strategy) __P((struct buf *));
     82 	struct buf *bp;
     83 	dev_t dev;
     84 	int flags;
     85 	void (*minphys) __P((struct buf *));
     86 	struct uio *uio;
     87 {
     88 	struct iovec *iovp;
     89 	struct proc *p = curproc;
     90 	int error, done, i, nobuf, s, todo;
     91 
     92 	error = 0;
     93 	flags &= B_READ | B_WRITE;
     94 
     95 	/*
     96 	 * [check user read/write access to the data buffer]
     97 	 *
     98 	 * Check each iov one by one.  Note that we know if we're reading or
     99 	 * writing, so we ignore the uio's rw parameter.  Also note that if
    100 	 * we're doing a read, that's a *write* to user-space.
    101 	 */
    102 	if (uio->uio_segflg == UIO_USERSPACE)
    103 		for (i = 0; i < uio->uio_iovcnt; i++)
    104 #if defined(UVM) /* XXXCDC: map not locked, rethink */
    105 			if (!uvm_useracc(uio->uio_iov[i].iov_base,
    106 				     uio->uio_iov[i].iov_len,
    107 				     (flags == B_READ) ? B_WRITE : B_READ))
    108 				return (EFAULT);
    109 #else
    110 			if (!useracc(uio->uio_iov[i].iov_base,
    111 				     uio->uio_iov[i].iov_len,
    112 				     (flags == B_READ) ? B_WRITE : B_READ))
    113 				return (EFAULT);
    114 #endif
    115 
    116 	/* Make sure we have a buffer, creating one if necessary. */
    117 	if ((nobuf = (bp == NULL)) != 0) {
    118 
    119 		bp = getphysbuf();
    120 		/* bp was just malloc'd so can't already be busy */
    121 		bp->b_flags |= B_BUSY;
    122 
    123 	} else {
    124 
    125 		/* [raise the processor priority level to splbio;] */
    126 		s = splbio();
    127 
    128 		/* [while the buffer is marked busy] */
    129 		while (bp->b_flags & B_BUSY) {
    130 			/* [mark the buffer wanted] */
    131 			bp->b_flags |= B_WANTED;
    132 			/* [wait until the buffer is available] */
    133 			tsleep((caddr_t)bp, PRIBIO+1, "physbuf", 0);
    134 		}
    135 
    136 		/* Mark it busy, so nobody else will use it. */
    137 		bp->b_flags |= B_BUSY;
    138 
    139 		/* [lower the priority level] */
    140 		splx(s);
    141 
    142 	}
    143 
    144 	/* [set up the fixed part of the buffer for a transfer] */
    145 	bp->b_dev = dev;
    146 	bp->b_error = 0;
    147 	bp->b_proc = p;
    148 
    149 	/*
    150 	 * [while there are data to transfer and no I/O error]
    151 	 * Note that I/O errors are handled with a 'goto' at the bottom
    152 	 * of the 'while' loop.
    153 	 */
    154 	for (i = 0; i < uio->uio_iovcnt; i++) {
    155 		iovp = &uio->uio_iov[i];
    156 		while (iovp->iov_len > 0) {
    157 			/*
    158 			 * [mark the buffer busy for physical I/O]
    159 			 * (i.e. set B_PHYS (because it's an I/O to user
    160 			 * memory, and B_RAW, because B_RAW is to be
    161 			 * "Set by physio for raw transfers.", in addition
    162 			 * to the "busy" and read/write flag.)
    163 			 */
    164 			bp->b_flags = B_BUSY | B_PHYS | B_RAW | flags;
    165 
    166 			/* [set up the buffer for a maximum-sized transfer] */
    167 			bp->b_blkno = btodb(uio->uio_offset);
    168 			bp->b_bcount = iovp->iov_len;
    169 			bp->b_data = iovp->iov_base;
    170 
    171 			/*
    172 			 * [call minphys to bound the tranfer size]
    173 			 * and remember the amount of data to transfer,
    174 			 * for later comparison.
    175 			 */
    176 			(*minphys)(bp);
    177 			todo = bp->b_bcount;
    178 #ifdef DIAGNOSTIC
    179 			if (todo < 0)
    180 				panic("todo < 0; minphys broken");
    181 			if (todo > MAXPHYS)
    182 				panic("todo > MAXPHYS; minphys broken");
    183 #endif
    184 
    185 			/*
    186 			 * [lock the part of the user address space involved
    187 			 *    in the transfer]
    188 			 * Beware vmapbuf(); it clobbers b_data and
    189 			 * saves it in b_saveaddr.  However, vunmapbuf()
    190 			 * restores it.
    191 			 */
    192 			PHOLD(p);
    193 #if defined(UVM)
    194 			uvm_vslock(bp->b_data, todo);
    195 #else
    196 			vslock(bp->b_data, todo);
    197 #endif
    198 			vmapbuf(bp, todo);
    199 
    200 			/* [call strategy to start the transfer] */
    201 			(*strategy)(bp);
    202 
    203 			/*
    204 			 * Note that the raise/wait/lower/get error
    205 			 * steps below would be done by biowait(), but
    206 			 * we want to unlock the address space before
    207 			 * we lower the priority.
    208 			 *
    209 			 * [raise the priority level to splbio]
    210 			 */
    211 			s = splbio();
    212 
    213 			/* [wait for the transfer to complete] */
    214 			while ((bp->b_flags & B_DONE) == 0)
    215 				tsleep((caddr_t) bp, PRIBIO + 1, "physio", 0);
    216 
    217 			/* Mark it busy again, so nobody else will use it. */
    218 			bp->b_flags |= B_BUSY;
    219 
    220 			/* [lower the priority level] */
    221 			splx(s);
    222 
    223 			/*
    224 			 * [unlock the part of the address space previously
    225 			 *    locked]
    226 			 */
    227 			vunmapbuf(bp, todo);
    228 #if defined(UVM)
    229 			uvm_vsunlock(bp->b_data, todo);
    230 #else
    231 			vsunlock(bp->b_data, todo);
    232 #endif
    233 			PRELE(p);
    234 
    235 			/* remember error value (save a splbio/splx pair) */
    236 			if (bp->b_flags & B_ERROR)
    237 				error = (bp->b_error ? bp->b_error : EIO);
    238 
    239 			/*
    240 			 * [deduct the transfer size from the total number
    241 			 *    of data to transfer]
    242 			 */
    243 			done = bp->b_bcount - bp->b_resid;
    244 #ifdef DIAGNOSTIC
    245 			if (done < 0)
    246 				panic("done < 0; strategy broken");
    247 			if (done > todo)
    248 				panic("done > todo; strategy broken");
    249 #endif
    250 			iovp->iov_len -= done;
    251                         iovp->iov_base += done;
    252                         uio->uio_offset += done;
    253                         uio->uio_resid -= done;
    254 
    255 			/*
    256 			 * Now, check for an error.
    257 			 * Also, handle weird end-of-disk semantics.
    258 			 */
    259 			if (error || done < todo)
    260 				goto done;
    261 		}
    262 	}
    263 
    264 done:
    265 	/*
    266 	 * [clean up the state of the buffer]
    267 	 * Remember if somebody wants it, so we can wake them up below.
    268 	 * Also, if we had to steal it, give it back.
    269 	 */
    270 	s = splbio();
    271 	bp->b_flags &= ~(B_BUSY | B_PHYS | B_RAW);
    272 	if (nobuf)
    273 		putphysbuf(bp);
    274 	else {
    275 		/*
    276 		 * [if another process is waiting for the raw I/O buffer,
    277 		 *    wake up processes waiting to do physical I/O;
    278 		 */
    279 		if (bp->b_flags & B_WANTED) {
    280 			bp->b_flags &= ~B_WANTED;
    281 			wakeup(bp);
    282 		}
    283 	}
    284 	splx(s);
    285 
    286 	return (error);
    287 }
    288 
    289 /*
    290  * allocate a buffer structure for use in physical I/O.
    291  */
    292 struct buf *
    293 getphysbuf()
    294 {
    295 	struct buf *bp;
    296 
    297 	bp = malloc(sizeof(*bp), M_TEMP, M_WAITOK);
    298 	bzero(bp, sizeof(*bp));
    299 
    300 	/* XXXCDC: are the following two lines necessary? */
    301 	bp->b_rcred = bp->b_wcred = NOCRED;
    302 	bp->b_vnbufs.le_next = NOLIST;
    303 
    304 	return(bp);
    305 }
    306 
    307 /*
    308  * get rid of a swap buffer structure which has been used in physical I/O.
    309  */
    310 void
    311 putphysbuf(bp)
    312         struct buf *bp;
    313 {
    314 
    315 	/* XXXCDC: is this necesary? */
    316 	if (bp->b_vp)
    317 		brelvp(bp);
    318 
    319 	if (bp->b_flags & B_WANTED)
    320 		panic("putphysbuf: private buf B_WANTED");
    321 	free(bp, M_TEMP);
    322 
    323 }
    324 
    325 /*
    326  * Leffler, et al., says on p. 231:
    327  * "The minphys() routine is called by physio() to adjust the
    328  * size of each I/O transfer before the latter is passed to
    329  * the strategy routine..."
    330  *
    331  * so, just adjust the buffer's count accounting to MAXPHYS here,
    332  * and return the new count;
    333  */
    334 void
    335 minphys(bp)
    336 	struct buf *bp;
    337 {
    338 
    339 	if (bp->b_bcount > MAXPHYS)
    340 		bp->b_bcount = MAXPHYS;
    341 }
    342