xref: /netbsd-src/sys/kern/kern_physio.c (revision b5677b36047b601b9addaaa494a58ceae82c2a6c)
1 /*	$NetBSD: kern_physio.c,v 1.89 2008/11/09 12:18:07 bouyer 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 1.89 2008/11/09 12:18:07 bouyer 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 
96 /* #define	PHYSIO_DEBUG */
97 #if defined(PHYSIO_DEBUG)
98 #define	DPRINTF(a)	printf a
99 #else /* defined(PHYSIO_DEBUG) */
100 #define	DPRINTF(a)	/* nothing */
101 #endif /* defined(PHYSIO_DEBUG) */
102 
103 struct physio_stat {
104 	int ps_running;
105 	int ps_error;
106 	int ps_failed;
107 	off_t ps_endoffset;
108 	buf_t *ps_orig_bp;
109 	kmutex_t ps_lock;
110 	kcondvar_t ps_cv;
111 };
112 
113 static void
114 physio_done(struct work *wk, void *dummy)
115 {
116 	struct buf *bp = (void *)wk;
117 	size_t todo = bp->b_bufsize;
118 	size_t done = bp->b_bcount - bp->b_resid;
119 	struct physio_stat *ps = bp->b_private;
120 
121 	KASSERT(&bp->b_work == wk);
122 	KASSERT(bp->b_bcount <= todo);
123 	KASSERT(bp->b_resid <= bp->b_bcount);
124 	KASSERT((bp->b_flags & B_PHYS) != 0);
125 	KASSERT(dummy == NULL);
126 
127 	vunmapbuf(bp, todo);
128 	uvm_vsunlock(bp->b_proc->p_vmspace, bp->b_data, todo);
129 
130 	mutex_enter(&ps->ps_lock);
131 	if (__predict_false(done != todo)) {
132 		off_t endoffset = dbtob(bp->b_blkno) + done;
133 
134 		/*
135 		 * we got an error or hit EOM.
136 		 *
137 		 * we only care about the first one.
138 		 * ie. the one at the lowest offset.
139 		 */
140 
141 		KASSERT(ps->ps_endoffset != endoffset);
142 		DPRINTF(("%s: error=%d at %" PRIu64 " - %" PRIu64
143 		    ", blkno=%" PRIu64 ", bcount=%d, flags=0x%x\n",
144 		    __func__, bp->b_error, dbtob(bp->b_blkno), endoffset,
145 		    bp->b_blkno, bp->b_bcount, bp->b_flags));
146 
147 		if (ps->ps_endoffset == -1 || endoffset < ps->ps_endoffset) {
148 			DPRINTF(("%s: ps=%p, error %d -> %d, endoff %" PRIu64
149 			    " -> %" PRIu64 "\n",
150 			    __func__, ps,
151 			    ps->ps_error, bp->b_error,
152 			    ps->ps_endoffset, endoffset));
153 
154 			ps->ps_endoffset = endoffset;
155 			ps->ps_error = bp->b_error;
156 		}
157 		ps->ps_failed++;
158 	} else {
159 		KASSERT(bp->b_error == 0);
160 	}
161 
162 	ps->ps_running--;
163 	cv_signal(&ps->ps_cv);
164 	mutex_exit(&ps->ps_lock);
165 
166 	if (bp != ps->ps_orig_bp)
167 		putiobuf(bp);
168 }
169 
170 static void
171 physio_biodone(struct buf *bp)
172 {
173 #if defined(DIAGNOSTIC)
174 	struct physio_stat *ps = bp->b_private;
175 	size_t todo = bp->b_bufsize;
176 	size_t done = bp->b_bcount - bp->b_resid;
177 
178 	KASSERT(ps->ps_running > 0);
179 	KASSERT(bp->b_bcount <= todo);
180 	KASSERT(bp->b_resid <= bp->b_bcount);
181 	if (done == todo)
182 		KASSERT(bp->b_error == 0);
183 #endif /* defined(DIAGNOSTIC) */
184 
185 	workqueue_enqueue(physio_workqueue, &bp->b_work, NULL);
186 }
187 
188 static void
189 physio_wait(struct physio_stat *ps, int n)
190 {
191 
192 	KASSERT(mutex_owned(&ps->ps_lock));
193 
194 	while (ps->ps_running > n)
195 		cv_wait(&ps->ps_cv, &ps->ps_lock);
196 }
197 
198 static int
199 physio_init(void)
200 {
201 	int error;
202 
203 	KASSERT(physio_workqueue == NULL);
204 
205 	error = workqueue_create(&physio_workqueue, "physiod",
206 	    physio_done, NULL, PRI_BIO, IPL_BIO, WQ_MPSAFE);
207 
208 	return error;
209 }
210 
211 #define	PHYSIO_CONCURRENCY	16	/* XXX tune */
212 
213 /*
214  * Do "physical I/O" on behalf of a user.  "Physical I/O" is I/O directly
215  * from the raw device to user buffers, and bypasses the buffer cache.
216  *
217  * Comments in brackets are from Leffler, et al.'s pseudo-code implementation.
218  */
219 int
220 physio(void (*strategy)(struct buf *), struct buf *obp, dev_t dev, int flags,
221     void (*min_phys)(struct buf *), struct uio *uio)
222 {
223 	struct iovec *iovp;
224 	struct lwp *l = curlwp;
225 	struct proc *p = l->l_proc;
226 	int i, error;
227 	struct buf *bp = NULL;
228 	struct physio_stat *ps;
229 	int concurrency = PHYSIO_CONCURRENCY - 1;
230 
231 	error = RUN_ONCE(&physio_initialized, physio_init);
232 	if (__predict_false(error != 0)) {
233 		return error;
234 	}
235 
236 	DPRINTF(("%s: called: off=%" PRIu64 ", resid=%zu\n",
237 	    __func__, uio->uio_offset, uio->uio_resid));
238 
239 	flags &= B_READ | B_WRITE;
240 
241 	if ((ps = kmem_zalloc(sizeof(*ps), KM_SLEEP)) == NULL)
242 		return ENOMEM;
243 	/* ps->ps_running = 0; */
244 	/* ps->ps_error = 0; */
245 	/* ps->ps_failed = 0; */
246 	ps->ps_orig_bp = obp;
247 	ps->ps_endoffset = -1;
248 	mutex_init(&ps->ps_lock, MUTEX_DEFAULT, IPL_NONE);
249 	cv_init(&ps->ps_cv, "physio");
250 
251 	/* Make sure we have a buffer, creating one if necessary. */
252 	if (obp != NULL) {
253 		/* [raise the processor priority level to splbio;] */
254 		mutex_enter(&bufcache_lock);
255 		/* Mark it busy, so nobody else will use it. */
256 		while (bbusy(obp, false, 0, NULL) == EPASSTHROUGH)
257 			;
258 		mutex_exit(&bufcache_lock);
259 		concurrency = 0; /* see "XXXkludge" comment below */
260 	}
261 
262 	uvm_lwp_hold(l);
263 
264 	for (i = 0; i < uio->uio_iovcnt; i++) {
265 		bool sync = true;
266 
267 		iovp = &uio->uio_iov[i];
268 		while (iovp->iov_len > 0) {
269 			size_t todo;
270 			vaddr_t endp;
271 
272 			mutex_enter(&ps->ps_lock);
273 			if (ps->ps_failed != 0) {
274 				goto done_locked;
275 			}
276 			physio_wait(ps, sync ? 0 : concurrency);
277 			mutex_exit(&ps->ps_lock);
278 			if (obp != NULL) {
279 				/*
280 				 * XXXkludge
281 				 * some drivers use "obp" as an identifier.
282 				 */
283 				bp = obp;
284 			} else {
285 				bp = getiobuf(NULL, true);
286 				bp->b_cflags = BC_BUSY;
287 			}
288 			bp->b_dev = dev;
289 			bp->b_proc = p;
290 			bp->b_private = ps;
291 
292 			/*
293 			 * [mark the buffer busy for physical I/O]
294 			 * (i.e. set B_PHYS (because it's an I/O to user
295 			 * memory, and B_RAW, because B_RAW is to be
296 			 * "Set by physio for raw transfers.", in addition
297 			 * to the "busy" and read/write flag.)
298 			 */
299 			bp->b_oflags = 0;
300 			bp->b_cflags = BC_BUSY;
301 			bp->b_flags = flags | B_PHYS | B_RAW;
302 			bp->b_iodone = physio_biodone;
303 
304 			/* [set up the buffer for a maximum-sized transfer] */
305 			bp->b_blkno = btodb(uio->uio_offset);
306 			if (dbtob(bp->b_blkno) != uio->uio_offset) {
307 				error = EINVAL;
308 				goto done;
309 			}
310 			bp->b_bcount = MIN(MAXPHYS, iovp->iov_len);
311 			bp->b_data = iovp->iov_base;
312 
313 			/*
314 			 * [call minphys to bound the transfer size]
315 			 * and remember the amount of data to transfer,
316 			 * for later comparison.
317 			 */
318 			(*min_phys)(bp);
319 			todo = bp->b_bufsize = bp->b_bcount;
320 #if defined(DIAGNOSTIC)
321 			if (todo > MAXPHYS)
322 				panic("todo(%zu) > MAXPHYS; minphys broken",
323 				    todo);
324 #endif /* defined(DIAGNOSTIC) */
325 
326 			sync = false;
327 			endp = (vaddr_t)bp->b_data + todo;
328 			if (trunc_page(endp) != endp) {
329 				/*
330 				 * following requests can overlap.
331 				 * note that uvm_vslock does round_page.
332 				 */
333 				sync = true;
334 			}
335 
336 			/*
337 			 * [lock the part of the user address space involved
338 			 *    in the transfer]
339 			 * Beware vmapbuf(); it clobbers b_data and
340 			 * saves it in b_saveaddr.  However, vunmapbuf()
341 			 * restores it.
342 			 */
343 			error = uvm_vslock(p->p_vmspace, bp->b_data, todo,
344 			    (flags & B_READ) ?  VM_PROT_WRITE : VM_PROT_READ);
345 			if (error) {
346 				goto done;
347 			}
348 			vmapbuf(bp, todo);
349 
350 			BIO_SETPRIO(bp, BPRIO_TIMECRITICAL);
351 
352 			mutex_enter(&ps->ps_lock);
353 			ps->ps_running++;
354 			mutex_exit(&ps->ps_lock);
355 
356 			/* [call strategy to start the transfer] */
357 			(*strategy)(bp);
358 			bp = NULL;
359 
360 			iovp->iov_len -= todo;
361 			iovp->iov_base = (char *)iovp->iov_base + todo;
362 			uio->uio_offset += todo;
363 			uio->uio_resid -= todo;
364 		}
365 	}
366 
367 done:
368 	mutex_enter(&ps->ps_lock);
369 done_locked:
370 	physio_wait(ps, 0);
371 	mutex_exit(&ps->ps_lock);
372 
373 	if (ps->ps_failed != 0) {
374 		off_t delta;
375 
376 		delta = uio->uio_offset - ps->ps_endoffset;
377 		KASSERT(delta > 0);
378 		uio->uio_resid += delta;
379 		/* uio->uio_offset = ps->ps_endoffset; */
380 	} else {
381 		KASSERT(ps->ps_endoffset == -1);
382 	}
383 	if (bp != NULL && bp != obp) {
384 		putiobuf(bp);
385 	}
386 	if (error == 0) {
387 		error = ps->ps_error;
388 	}
389 	mutex_destroy(&ps->ps_lock);
390 	cv_destroy(&ps->ps_cv);
391 	kmem_free(ps, sizeof(*ps));
392 
393 	/*
394 	 * [clean up the state of the buffer]
395 	 * Remember if somebody wants it, so we can wake them up below.
396 	 * Also, if we had to steal it, give it back.
397 	 */
398 	if (obp != NULL) {
399 		KASSERT((obp->b_cflags & BC_BUSY) != 0);
400 
401 		/*
402 		 * [if another process is waiting for the raw I/O buffer,
403 		 *    wake up processes waiting to do physical I/O;
404 		 */
405 		mutex_enter(&bufcache_lock);
406 		obp->b_cflags &= ~(BC_BUSY | BC_WANTED);
407 		obp->b_flags &= ~(B_PHYS | B_RAW);
408 		obp->b_iodone = NULL;
409 		cv_broadcast(&obp->b_busy);
410 		mutex_exit(&bufcache_lock);
411 	}
412 	uvm_lwp_rele(l);
413 
414 	DPRINTF(("%s: done: off=%" PRIu64 ", resid=%zu\n",
415 	    __func__, uio->uio_offset, uio->uio_resid));
416 
417 	return error;
418 }
419 
420 /*
421  * Leffler, et al., says on p. 231:
422  * "The minphys() routine is called by physio() to adjust the
423  * size of each I/O transfer before the latter is passed to
424  * the strategy routine..."
425  *
426  * so, just adjust the buffer's count accounting to MAXPHYS here,
427  * and return the new count;
428  */
429 void
430 minphys(struct buf *bp)
431 {
432 
433 	if (bp->b_bcount > MAXPHYS)
434 		bp->b_bcount = MAXPHYS;
435 }
436