xref: /netbsd-src/sys/uvm/uvm_readahead.c (revision 8b0f9554ff8762542c4defc4f70e1eb76fb508fa)
1 /*	$NetBSD: uvm_readahead.c,v 1.4 2007/05/11 12:11:09 tsutsui Exp $	*/
2 
3 /*-
4  * Copyright (c)2003, 2005 YAMAMOTO Takashi,
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * uvm_object read-ahead
31  *
32  * TODO:
33  *	- tune.
34  *	- handle multiple streams.
35  *	- find a better way to deal with PGO_LOCKED pager requests.
36  *	  (currently just ignored)
37  *	- consider the amount of memory in the system.
38  *	- consider the speed of the underlying device.
39  *	- consider filesystem block size / block layout.
40  */
41 
42 #include <sys/cdefs.h>
43 __KERNEL_RCSID(0, "$NetBSD: uvm_readahead.c,v 1.4 2007/05/11 12:11:09 tsutsui Exp $");
44 
45 #include <sys/param.h>
46 #include <sys/pool.h>
47 
48 #include <uvm/uvm.h>
49 #include <uvm/uvm_readahead.h>
50 
51 #if defined(READAHEAD_DEBUG)
52 #define	DPRINTF(a)	printf a
53 #else /* defined(READAHEAD_DEBUG) */
54 #define	DPRINTF(a)	/* nothing */
55 #endif /* defined(READAHEAD_DEBUG) */
56 
57 /*
58  * uvm_ractx: read-ahead context.
59  */
60 
61 struct uvm_ractx {
62 	int ra_flags;
63 #define	RA_VALID	1
64 	off_t ra_winstart;	/* window start offset */
65 	size_t ra_winsize;	/* window size */
66 	off_t ra_next;		/* next offset to read-ahead */
67 };
68 
69 #if defined(sun2) || (defined(sun3) && defined(_SUN3_))
70 /* XXX: on sun2 and sun3 (but not sun3x) MAXPHYS is 0xe000 */
71 #undef MAXPHYS
72 #define MAXPHYS		0x8000	/* XXX */
73 #endif
74 
75 #define	RA_WINSIZE_INIT	MAXPHYS			/* initial window size */
76 #define	RA_WINSIZE_MAX	(MAXPHYS * 8)		/* max window size */
77 #define	RA_WINSIZE_SEQENTIAL	RA_WINSIZE_MAX	/* fixed window size used for
78 						   SEQUENTIAL hint */
79 #define	RA_MINSIZE	(MAXPHYS * 2)		/* min size to start i/o */
80 #define	RA_IOCHUNK	MAXPHYS			/* read-ahead i/o chunk size */
81 
82 static off_t ra_startio(struct uvm_object *, off_t, size_t);
83 static struct uvm_ractx *ra_allocctx(void);
84 static void ra_freectx(struct uvm_ractx *);
85 
86 static POOL_INIT(ractx_pool, sizeof(struct uvm_ractx), 0, 0, 0, "ractx",
87     &pool_allocator_nointr, IPL_NONE);
88 
89 static struct uvm_ractx *
90 ra_allocctx(void)
91 {
92 
93 	return pool_get(&ractx_pool, PR_NOWAIT);
94 }
95 
96 static void
97 ra_freectx(struct uvm_ractx *ra)
98 {
99 
100 	pool_put(&ractx_pool, ra);
101 }
102 
103 /*
104  * ra_startio: start i/o for read-ahead.
105  *
106  * => start i/o for each RA_IOCHUNK sized chunk.
107  * => return offset to which we started i/o.
108  */
109 
110 static off_t
111 ra_startio(struct uvm_object *uobj, off_t off, size_t sz)
112 {
113 	const off_t endoff = off + sz;
114 
115 	DPRINTF(("%s: uobj=%p, off=%" PRIu64 ", endoff=%" PRIu64 "\n",
116 	    __func__, uobj, off, endoff));
117 	off = trunc_page(off);
118 	while (off < endoff) {
119 		const size_t chunksize = RA_IOCHUNK;
120 		int error;
121 		size_t donebytes;
122 		int npages;
123 		int orignpages;
124 		size_t bytelen;
125 
126 		KASSERT((chunksize & (chunksize - 1)) == 0);
127 		KASSERT((off & PAGE_MASK) == 0);
128 		bytelen = ((off + chunksize) & -(off_t)chunksize) - off;
129 		KASSERT((bytelen & PAGE_MASK) == 0);
130 		npages = orignpages = bytelen >> PAGE_SHIFT;
131 		KASSERT(npages != 0);
132 
133 		/*
134 		 * use UVM_ADV_RANDOM to avoid recursion.
135 		 */
136 
137 		simple_lock(&uobj->vmobjlock);
138 		error = (*uobj->pgops->pgo_get)(uobj, off, NULL,
139 		    &npages, 0, VM_PROT_READ, UVM_ADV_RANDOM, 0);
140 		DPRINTF(("%s:  off=%" PRIu64 ", bytelen=%zu -> %d\n",
141 		    __func__, off, bytelen, error));
142 		if (error != 0 && error != EBUSY) {
143 			if (error != EINVAL) { /* maybe past EOF */
144 				DPRINTF(("%s: error=%d\n", __func__, error));
145 			}
146 			break;
147 		}
148 		KASSERT(orignpages == npages);
149 		donebytes = orignpages << PAGE_SHIFT;
150 		off += donebytes;
151 	}
152 
153 	return off;
154 }
155 
156 /* ------------------------------------------------------------ */
157 
158 /*
159  * uvm_ra_allocctx: allocate a context.
160  */
161 
162 struct uvm_ractx *
163 uvm_ra_allocctx(void)
164 {
165 	struct uvm_ractx *ra;
166 
167 	ra = ra_allocctx();
168 	if (ra != NULL) {
169 		ra->ra_flags = 0;
170 	}
171 
172 	return ra;
173 }
174 
175 /*
176  * uvm_ra_freectx: free a context.
177  */
178 
179 void
180 uvm_ra_freectx(struct uvm_ractx *ra)
181 {
182 
183 	KASSERT(ra != NULL);
184 	ra_freectx(ra);
185 }
186 
187 /*
188  * uvm_ra_request: update a read-ahead context and start i/o if appropriate.
189  *
190  * => called when [reqoff, reqoff+reqsize) is requested.
191  */
192 
193 void
194 uvm_ra_request(struct uvm_ractx *ra, int advice, struct uvm_object *uobj,
195     off_t reqoff, size_t reqsize)
196 {
197 
198 	if (ra == NULL || advice == UVM_ADV_RANDOM) {
199 		return;
200 	}
201 
202 	/*
203 	 * XXX needs locking?  maybe.
204 	 * but the worst effect is merely a bad read-ahead.
205 	 */
206 
207 	if (advice == UVM_ADV_SEQUENTIAL) {
208 
209 		/*
210 		 * always do read-ahead with a large window.
211 		 */
212 
213 		if ((ra->ra_flags & RA_VALID) == 0) {
214 			ra->ra_winstart = ra->ra_next = 0;
215 			ra->ra_flags |= RA_VALID;
216 		}
217 		if (reqoff < ra->ra_winstart) {
218 			ra->ra_next = reqoff;
219 		}
220 		ra->ra_winsize = RA_WINSIZE_SEQENTIAL;
221 		goto do_readahead;
222 	}
223 
224 	/*
225 	 * a request with UVM_ADV_NORMAL hint.  (ie. no hint)
226 	 *
227 	 * we keep a sliding window in order to determine:
228 	 *	- if the previous read-ahead was successful or not.
229 	 *	- how many bytes to read-ahead.
230 	 */
231 
232 	/*
233 	 * if it's the first request for this context,
234 	 * initialize context and return.
235 	 */
236 
237 	if ((ra->ra_flags & RA_VALID) == 0) {
238 initialize:
239 		ra->ra_winstart = ra->ra_next = reqoff + reqsize;
240 		ra->ra_winsize = RA_WINSIZE_INIT;
241 		ra->ra_flags |= RA_VALID;
242 		goto done;
243 	}
244 
245 	/*
246 	 * if it isn't in our window,
247 	 * initialize context and return.
248 	 * (read-ahead miss)
249 	 */
250 
251 	if (reqoff < ra->ra_winstart ||
252 	    ra->ra_winstart + ra->ra_winsize < reqoff) {
253 
254 		/*
255 		 * ... unless we seem to be reading the same chunk repeatedly.
256 		 *
257 		 * XXX should have some margin?
258 		 */
259 
260 		if (reqoff + reqsize == ra->ra_winstart) {
261 			DPRINTF(("%s: %p: same block: off=%" PRIu64
262 			    ", size=%zd, winstart=%" PRIu64 "\n",
263 			    __func__, ra, reqoff, reqsize, ra->ra_winstart));
264 			goto done;
265 		}
266 		goto initialize;
267 	}
268 
269 	/*
270 	 * it's in our window. (read-ahead hit)
271 	 *	- start read-ahead i/o if appropriate.
272 	 *	- advance and enlarge window.
273 	 */
274 
275 do_readahead:
276 
277 	/*
278 	 * don't bother to read-ahead behind current request.
279 	 */
280 
281 	if (reqoff > ra->ra_next) {
282 		ra->ra_next = reqoff;
283 	}
284 
285 	/*
286 	 * try to make [reqoff, reqoff+ra_winsize) in-core.
287 	 * note that [reqoff, ra_next) is considered already done.
288 	 */
289 
290 	if (reqoff + ra->ra_winsize > ra->ra_next) {
291 		off_t raoff = MAX(reqoff, ra->ra_next);
292 		size_t rasize = reqoff + ra->ra_winsize - ra->ra_next;
293 
294 #if defined(DIAGNOSTIC)
295 		if (rasize > RA_WINSIZE_MAX) {
296 
297 			/*
298 			 * shouldn't happen as far as we're protected by
299 			 * kernel_lock.
300 			 */
301 
302 			printf("%s: corrupted context", __func__);
303 			rasize = RA_WINSIZE_MAX;
304 		}
305 #endif /* defined(DIAGNOSTIC) */
306 
307 		/*
308 		 * issue read-ahead only if we can start big enough i/o.
309 		 * otherwise we end up with a stream of small i/o.
310 		 */
311 
312 		if (rasize >= RA_MINSIZE) {
313 			ra->ra_next = ra_startio(uobj, raoff, rasize);
314 		}
315 	}
316 
317 	/*
318 	 * update window.
319 	 *
320 	 * enlarge window by reqsize, so that it grows in a predictable manner
321 	 * regardless of the size of each read(2).
322 	 */
323 
324 	ra->ra_winstart = reqoff + reqsize;
325 	ra->ra_winsize = MIN(RA_WINSIZE_MAX, ra->ra_winsize + reqsize);
326 
327 done:;
328 }
329