xref: /netbsd-src/sys/uvm/uvm_readahead.c (revision 59e0001f2c76f540e0d2715b1f987bb546a238ec)
1 /*	$NetBSD: uvm_readahead.c,v 1.16 2023/09/23 18:21:12 ad Exp $	*/
2 
3 /*-
4  * Copyright (c)2003, 2005, 2009 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 1.16 2023/09/23 18:21:12 ad Exp $");
44 
45 #include <sys/param.h>
46 #include <sys/kmem.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)
70 /* XXX: on sun2 and sun3 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 * 16)		/* 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 /*
87  * uvm_ra_init: initialize readahead module.
88  */
89 
90 void
uvm_ra_init(void)91 uvm_ra_init(void)
92 {
93 
94 }
95 
96 static struct uvm_ractx *
ra_allocctx(void)97 ra_allocctx(void)
98 {
99 
100 	return kmem_alloc(sizeof(struct uvm_ractx), KM_NOSLEEP);
101 }
102 
103 static void
ra_freectx(struct uvm_ractx * ra)104 ra_freectx(struct uvm_ractx *ra)
105 {
106 
107 	kmem_free(ra, sizeof(struct uvm_ractx));
108 }
109 
110 /*
111  * ra_startio: start i/o for read-ahead.
112  *
113  * => start i/o for each RA_IOCHUNK sized chunk.
114  * => return offset to which we started i/o.
115  */
116 
117 static off_t
ra_startio(struct uvm_object * uobj,off_t off,size_t sz)118 ra_startio(struct uvm_object *uobj, off_t off, size_t sz)
119 {
120 	const off_t endoff = off + sz;
121 
122 	DPRINTF(("%s: uobj=%p, off=%" PRIu64 ", endoff=%" PRIu64 "\n",
123 	    __func__, uobj, off, endoff));
124 
125 	KASSERT(rw_write_held(uobj->vmobjlock));
126 
127 	/*
128 	 * Don't issue read-ahead if the last page of the range is already cached.
129 	 * The assumption is that since the access is sequential, the intermediate
130 	 * pages would have similar LRU stats, and hence likely to be still in cache
131 	 * too. This speeds up I/O using cache, since it avoids lookups and temporary
132 	 * allocations done by full pgo_get.
133 	 */
134 	struct vm_page *pg = uvm_pagelookup(uobj, trunc_page(endoff - 1));
135 	if (pg != NULL) {
136 		DPRINTF(("%s:  off=%" PRIu64 ", sz=%zu already cached\n",
137 		    __func__, off, sz));
138 		return endoff;
139 	}
140 
141 	off = trunc_page(off);
142 	while (off < endoff) {
143 		const size_t chunksize = RA_IOCHUNK;
144 		int error;
145 		size_t donebytes;
146 		int npages;
147 		int orignpages;
148 		size_t bytelen;
149 
150 		KASSERT((chunksize & (chunksize - 1)) == 0);
151 		KASSERT((off & PAGE_MASK) == 0);
152 		bytelen = ((off + chunksize) & -(off_t)chunksize) - off;
153 		KASSERT((bytelen & PAGE_MASK) == 0);
154 		npages = orignpages = bytelen >> PAGE_SHIFT;
155 		KASSERT(npages != 0);
156 
157 		/*
158 		 * use UVM_ADV_RANDOM to avoid recursion.
159 		 */
160 
161 		error = (*uobj->pgops->pgo_get)(uobj, off, NULL,
162 		    &npages, 0, VM_PROT_READ, UVM_ADV_RANDOM, PGO_NOTIMESTAMP);
163 		rw_enter(uobj->vmobjlock, RW_WRITER);
164 		DPRINTF(("%s:  off=%" PRIu64 ", bytelen=%zu -> %d\n",
165 		    __func__, off, bytelen, error));
166 		if (error != 0 && error != EBUSY) {
167 			if (error != EINVAL) { /* maybe past EOF */
168 				DPRINTF(("%s: error=%d\n", __func__, error));
169 			}
170 			break;
171 		}
172 		KASSERT(orignpages == npages);
173 		donebytes = orignpages << PAGE_SHIFT;
174 		off += donebytes;
175 	}
176 
177 	return off;
178 }
179 
180 /* ------------------------------------------------------------ */
181 
182 /*
183  * uvm_ra_allocctx: allocate a context.
184  */
185 
186 struct uvm_ractx *
uvm_ra_allocctx(void)187 uvm_ra_allocctx(void)
188 {
189 	struct uvm_ractx *ra;
190 
191 	ra = ra_allocctx();
192 	if (ra != NULL) {
193 		ra->ra_flags = 0;
194 	}
195 
196 	return ra;
197 }
198 
199 /*
200  * uvm_ra_freectx: free a context.
201  */
202 
203 void
uvm_ra_freectx(struct uvm_ractx * ra)204 uvm_ra_freectx(struct uvm_ractx *ra)
205 {
206 
207 	KASSERT(ra != NULL);
208 	ra_freectx(ra);
209 }
210 
211 /*
212  * uvm_ra_request: update a read-ahead context and start i/o if appropriate.
213  *
214  * => called when [reqoff, reqoff+reqsize) is requested.
215  * => object must be locked by caller, will return locked.
216  */
217 
218 void
uvm_ra_request(struct uvm_ractx * ra,int advice,struct uvm_object * uobj,off_t reqoff,size_t reqsize)219 uvm_ra_request(struct uvm_ractx *ra, int advice, struct uvm_object *uobj,
220     off_t reqoff, size_t reqsize)
221 {
222 
223 	KASSERT(rw_write_held(uobj->vmobjlock));
224 
225 	if (ra == NULL || advice == UVM_ADV_RANDOM) {
226 		return;
227 	}
228 
229 	if (advice == UVM_ADV_SEQUENTIAL) {
230 
231 		/*
232 		 * always do read-ahead with a large window.
233 		 */
234 
235 		if ((ra->ra_flags & RA_VALID) == 0) {
236 			ra->ra_winstart = ra->ra_next = 0;
237 			ra->ra_flags |= RA_VALID;
238 		}
239 		if (reqoff < ra->ra_winstart) {
240 			ra->ra_next = reqoff;
241 		}
242 		ra->ra_winsize = RA_WINSIZE_SEQENTIAL;
243 		goto do_readahead;
244 	}
245 
246 	/*
247 	 * a request with UVM_ADV_NORMAL hint.  (ie. no hint)
248 	 *
249 	 * we keep a sliding window in order to determine:
250 	 *	- if the previous read-ahead was successful or not.
251 	 *	- how many bytes to read-ahead.
252 	 */
253 
254 	/*
255 	 * if it's the first request for this context,
256 	 * initialize context and return.
257 	 */
258 
259 	if ((ra->ra_flags & RA_VALID) == 0) {
260 initialize:
261 		ra->ra_winstart = ra->ra_next = reqoff + reqsize;
262 		ra->ra_winsize = RA_WINSIZE_INIT;
263 		ra->ra_flags |= RA_VALID;
264 		goto done;
265 	}
266 
267 	/*
268 	 * if it isn't in our window,
269 	 * initialize context and return.
270 	 * (read-ahead miss)
271 	 */
272 
273 	if (reqoff < ra->ra_winstart ||
274 	    ra->ra_winstart + ra->ra_winsize < reqoff) {
275 
276 		/*
277 		 * ... unless we seem to be reading the same chunk repeatedly.
278 		 *
279 		 * XXX should have some margin?
280 		 */
281 
282 		if (reqoff + reqsize == ra->ra_winstart) {
283 			DPRINTF(("%s: %p: same block: off=%" PRIu64
284 			    ", size=%zd, winstart=%" PRIu64 "\n",
285 			    __func__, ra, reqoff, reqsize, ra->ra_winstart));
286 			goto done;
287 		}
288 		goto initialize;
289 	}
290 
291 	/*
292 	 * it's in our window. (read-ahead hit)
293 	 *	- start read-ahead i/o if appropriate.
294 	 *	- advance and enlarge window.
295 	 */
296 
297 do_readahead:
298 
299 	/*
300 	 * don't bother to read-ahead behind current request.
301 	 */
302 
303 	if (reqoff > ra->ra_next) {
304 		ra->ra_next = reqoff;
305 	}
306 
307 	/*
308 	 * try to make [reqoff, reqoff+ra_winsize) in-core.
309 	 * note that [reqoff, ra_next) is considered already done.
310 	 */
311 
312 	if (reqoff + ra->ra_winsize > ra->ra_next) {
313 		off_t raoff = MAX(reqoff, ra->ra_next);
314 		size_t rasize = reqoff + ra->ra_winsize - ra->ra_next;
315 
316 #if defined(DIAGNOSTIC)
317 		if (rasize > RA_WINSIZE_MAX) {
318 			printf("%s: corrupted context", __func__);
319 			rasize = RA_WINSIZE_MAX;
320 		}
321 #endif /* defined(DIAGNOSTIC) */
322 
323 		/*
324 		 * issue read-ahead only if we can start big enough i/o.
325 		 * otherwise we end up with a stream of small i/o.
326 		 */
327 
328 		if (rasize >= RA_MINSIZE) {
329 			off_t next;
330 
331 			next = ra_startio(uobj, raoff, rasize);
332 			ra->ra_next = next;
333 		}
334 	}
335 
336 	/*
337 	 * update window.
338 	 *
339 	 * enlarge window by reqsize, so that it grows in a predictable manner
340 	 * regardless of the size of each read(2).
341 	 */
342 
343 	ra->ra_winstart = reqoff + reqsize;
344 	ra->ra_winsize = MIN(RA_WINSIZE_MAX, ra->ra_winsize + reqsize);
345 
346 done:;
347 }
348 
349 int
uvm_readahead(struct uvm_object * uobj,off_t off,off_t size)350 uvm_readahead(struct uvm_object *uobj, off_t off, off_t size)
351 {
352 
353 	/*
354 	 * don't allow too much read-ahead.
355 	 */
356 	if (size > RA_WINSIZE_MAX) {
357 		size = RA_WINSIZE_MAX;
358 	}
359 	rw_enter(uobj->vmobjlock, RW_WRITER);
360 	ra_startio(uobj, off, size);
361 	rw_exit(uobj->vmobjlock);
362 	return 0;
363 }
364