xref: /illumos-gate/usr/src/cmd/sgs/libelf/common/input.c (revision 2c5ec7a875dcd76853e6618614e990f1e8cdd56d)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*	Copyright (c) 1988 AT&T	*/
28 /*	  All Rights Reserved  	*/
29 
30 #include <unistd.h>
31 #include <stdlib.h>
32 #include <memory.h>
33 #include <errno.h>
34 #include <sys/mman.h>
35 #include <sys/param.h>
36 #include <libelf.h>
37 #include "decl.h"
38 #include "msg.h"
39 
40 
41 /*
42  * File input
43  *	These functions read input files.
44  *	On SVR4 and newer systems use mmap(2).  On older systems (or on
45  *	file systems that don't support mmap, this code simulates mmap.
46  *	When reading a file, enough memory is allocated to hold the file's
47  *	image, and reads are delayed.  When another part of the library
48  *	wants to use a part of the file, it "fetches" the needed regions.
49  *
50  *	An elf descriptor has a bit array to manage this.  Each bit
51  *	represents one "page" of the file.  Pages are grouped into regions.
52  *	The page size is tunable.  Its value should be at least one disk
53  *	block and small enough to avoid superfluous traffic.
54  *
55  *	NBITS	The number of bits in an unsigned.  Each unsigned object
56  *		holds a "REGION."  A byte must have at least 8 bits;
57  *		it may have more, though the extra bits at the top of
58  *		the unsigned will be unused.  Thus, for 9-bit bytes and
59  *		36-bit words, 4 bits at the top will stay empty.
60  *
61  *	This mechanism gives significant performance gains for library
62  *	handling (among other things), because programs typically don't
63  *	need to look at entire libraries.  The fastest I/O is no I/O.
64  */
65 
66 /*
67  * This global is used to hold the value of the PAGESIZE macro.
68  *
69  * This is because the PAGESIZE macro actually calls the
70  * sysconfig(_CONFIG_PAGESIZE) system call and we don't want
71  * to repeatedly call this through out libelf.
72  */
73 static unsigned long	_elf_pagesize = 0;
74 
75 #define	NBITS		(8 * sizeof (unsigned))
76 #define	REGSZ		(NBITS * _elf_pagesize)
77 #define	PGNUM(off)	((off % REGSZ) / _elf_pagesize)
78 #define	REGNUM(off)	(off / REGSZ)
79 
80 
81 
82 Okay
83 _elf_vm(Elf * elf, size_t base, size_t sz)
84 {
85 	register unsigned	*hdreg, hdbit;
86 	unsigned		*tlreg, tlbit;
87 	size_t			tail;
88 	off_t			off;
89 	Elf_Void		*iop;
90 
91 
92 	/*
93 	 * always validate region
94 	 */
95 
96 	if ((base + sz) > elf->ed_fsz) {
97 		/*
98 		 * range outside of file bounds.
99 		 */
100 		_elf_seterr(EFMT_VM, 0);
101 		return (OK_NO);
102 	}
103 
104 	/*
105 	 * If file is mmap()'d and/or the read size is 0
106 	 * their is nothing else for us to do.
107 	 */
108 	if (elf->ed_vm == 0 || sz == 0)
109 		return (OK_YES);
110 	/*
111 	 * This uses arithmetic instead of masking because
112 	 * sizeof (unsigned) might not be a power of 2.
113 	 *
114 	 * Tail gives one beyond the last offset that must be retrieved,
115 	 * NOT the last in the region.
116 	 */
117 
118 	if (elf->ed_parent && elf->ed_parent->ed_fd == -1)
119 		elf->ed_fd = -1;
120 
121 	base += elf->ed_baseoff;
122 	tail = base + sz + _elf_pagesize - 1;
123 	off = base - base % _elf_pagesize;
124 	hdbit = 1 << PGNUM(base);
125 	tlbit = 1 << PGNUM(tail);
126 	hdreg = &elf->ed_vm[REGNUM(base)];
127 	tlreg = &elf->ed_vm[REGNUM(tail)];
128 	sz = 0;
129 
130 	/*
131 	 * Scan through the files 'page table' and make sure
132 	 * that all of the pages in the specified range have been
133 	 * loaded into memory.  As the pages are loaded the appropriate
134 	 * bit in the 'page table' is set.
135 	 *
136 	 * Note: This loop will only read in those pages which havn't
137 	 *	 been previously loaded into memory, if the page is
138 	 *	 already present it will not be re-loaded.
139 	 */
140 	while ((hdreg != tlreg) || (hdbit != tlbit)) {
141 		if (*hdreg & hdbit) {
142 			if (sz != 0) {
143 				/*
144 				 * Read in a 'chunk' of the elf image.
145 				 */
146 				iop = (Elf_Void *)(elf->ed_image + off);
147 				/*
148 				 * do not read past the end of the file
149 				 */
150 				if (elf->ed_imagesz - off < sz)
151 					sz = elf->ed_imagesz - off;
152 				if ((lseek(elf->ed_fd, off,
153 				    SEEK_SET) != off) ||
154 				    (read(elf->ed_fd, iop, sz) != sz)) {
155 					_elf_seterr(EIO_VM, errno);
156 					return (OK_NO);
157 				}
158 				off += sz;
159 				sz = 0;
160 			}
161 			off += _elf_pagesize;
162 		} else {
163 			if (elf->ed_fd < 0) {
164 				_elf_seterr(EREQ_NOFD, 0);
165 				return (OK_NO);
166 			}
167 			sz += _elf_pagesize;
168 			*hdreg |= hdbit;
169 		}
170 		if (hdbit == ((unsigned)1 << (NBITS - 1))) {
171 			hdbit = 1;
172 			++hdreg;
173 		} else
174 			hdbit <<= 1;
175 	}
176 
177 	if (sz != 0) {
178 		iop = (Elf_Void *)(elf->ed_image + off);
179 		/*
180 		 * do not read past the end of the file
181 		 */
182 		if ((elf->ed_imagesz - off) < sz)
183 			sz = elf->ed_imagesz - off;
184 		if ((lseek(elf->ed_fd, off, SEEK_SET) != off) ||
185 		    (read(elf->ed_fd, iop, sz) != sz)) {
186 			_elf_seterr(EIO_VM, errno);
187 			return (OK_NO);
188 		}
189 	}
190 	return (OK_YES);
191 }
192 
193 
194 Okay
195 _elf_inmap(Elf * elf)
196 {
197 	int		fd = elf->ed_fd;
198 	register size_t	sz;
199 
200 	{
201 		register off_t	off = lseek(fd, (off_t)0, SEEK_END);
202 
203 		if (off == 0)
204 			return (OK_YES);
205 
206 		if (off == -1) {
207 			_elf_seterr(EIO_FSZ, errno);
208 			return (OK_NO);
209 		}
210 
211 		if ((sz = (size_t)off) != off) {
212 			_elf_seterr(EIO_FBIG, 0);
213 			return (OK_NO);
214 		}
215 	}
216 	/*
217 	 *	If the file is mapped, elf->ed_vm will stay null
218 	 *	and elf->ed_image will need to be unmapped someday.
219 	 *	If the file is read, elf->ed_vm and the file image
220 	 *	are allocated together; free() elf->ed_vm.
221 	 *
222 	 *	If the file can be written, disallow mmap.
223 	 *	Otherwise, the input mapping and the output mapping
224 	 *	can collide.  Moreover, elf_update will truncate
225 	 *	the file, possibly invalidating the input mapping.
226 	 *	Disallowing input mmap forces the library to malloc
227 	 *	and read the space, which will make output mmap safe.
228 	 *	Using mmap for output reduces the swap space needed
229 	 *	for the process, so that is given preference.
230 	 */
231 
232 	{
233 		register char	*p;
234 
235 		if ((elf->ed_myflags & EDF_WRITE) == 0 &&
236 		    (p = mmap((char *)0, sz, PROT_READ,
237 		    MAP_PRIVATE, fd, (off_t)0)) != (char *)-1) {
238 			elf->ed_image = elf->ed_ident = p;
239 			elf->ed_imagesz = elf->ed_fsz = elf->ed_identsz = sz;
240 			return (OK_YES);
241 		}
242 	}
243 
244 	if (_elf_pagesize == 0)
245 		_elf_pagesize = PAGESIZE;
246 
247 	/*
248 	 * If mmap fails, try read.  Some file systems don't mmap
249 	 */
250 	{
251 		register size_t	vmsz = sizeof (unsigned) * (REGNUM(sz) + 1);
252 
253 		if (vmsz % sizeof (Elf64) != 0)
254 			vmsz += sizeof (Elf64) - vmsz % sizeof (Elf64);
255 		if ((elf->ed_vm = (unsigned *)malloc(vmsz + sz)) == 0) {
256 			_elf_seterr(EMEM_VM, errno);
257 			return (OK_NO);
258 		}
259 		(void) memset(elf->ed_vm, 0, vmsz);
260 		elf->ed_vmsz = vmsz / sizeof (unsigned);
261 		elf->ed_image = elf->ed_ident = (char *)elf->ed_vm + vmsz;
262 		elf->ed_imagesz = elf->ed_fsz = elf->ed_identsz = sz;
263 	}
264 	return (_elf_vm(elf, (size_t)0, (size_t)1));
265 }
266 
267 
268 void
269 _elf_unmap(char *p, size_t sz)
270 {
271 	if (p == 0 || sz == 0)
272 		return;
273 	(void) munmap(p, sz);
274 }
275