xref: /freebsd/stand/common/load_elf.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
1 /*-
2  * Copyright (c) 1998 Michael Smith <msmith@freebsd.org>
3  * Copyright (c) 1998 Peter Wemm <peter@freebsd.org>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
30 
31 #include <sys/param.h>
32 #include <sys/endian.h>
33 #include <sys/exec.h>
34 #include <sys/linker.h>
35 #include <sys/module.h>
36 #include <sys/stdint.h>
37 #include <string.h>
38 #include <machine/elf.h>
39 #include <stand.h>
40 #define FREEBSD_ELF
41 #include <sys/link_elf.h>
42 
43 #include "bootstrap.h"
44 
45 #define COPYOUT(s,d,l)	archsw.arch_copyout((vm_offset_t)(s), d, l)
46 
47 #if defined(__i386__) && __ELF_WORD_SIZE == 64
48 #undef ELF_TARG_CLASS
49 #undef ELF_TARG_MACH
50 #define ELF_TARG_CLASS  ELFCLASS64
51 #define ELF_TARG_MACH   EM_X86_64
52 #endif
53 
54 typedef struct elf_file {
55     Elf_Phdr 	*ph;
56     Elf_Ehdr	*ehdr;
57     Elf_Sym	*symtab;
58     Elf_Hashelt	*hashtab;
59     Elf_Hashelt	nbuckets;
60     Elf_Hashelt	nchains;
61     Elf_Hashelt	*buckets;
62     Elf_Hashelt	*chains;
63     Elf_Rel	*rel;
64     size_t	relsz;
65     Elf_Rela	*rela;
66     size_t	relasz;
67     char	*strtab;
68     size_t	strsz;
69     int		fd;
70     caddr_t	firstpage;
71     size_t	firstlen;
72     int		kernel;
73     u_int64_t	off;
74 } *elf_file_t;
75 
76 static int __elfN(loadimage)(struct preloaded_file *mp, elf_file_t ef, u_int64_t loadaddr);
77 static int __elfN(lookup_symbol)(struct preloaded_file *mp, elf_file_t ef, const char* name, Elf_Sym* sym);
78 static int __elfN(reloc_ptr)(struct preloaded_file *mp, elf_file_t ef,
79     Elf_Addr p, void *val, size_t len);
80 static int __elfN(parse_modmetadata)(struct preloaded_file *mp, elf_file_t ef,
81     Elf_Addr p_start, Elf_Addr p_end);
82 static symaddr_fn __elfN(symaddr);
83 static char	*fake_modname(const char *name);
84 
85 const char	*__elfN(kerneltype) = "elf kernel";
86 const char	*__elfN(moduletype) = "elf module";
87 
88 u_int64_t	__elfN(relocation_offset) = 0;
89 
90 extern void elf_wrong_field_size(void);
91 #define CONVERT_FIELD(b, f, e)			\
92 	switch (sizeof((b)->f)) {		\
93 	case 2:					\
94 		(b)->f = e ## 16toh((b)->f);	\
95 		break;				\
96 	case 4:					\
97 		(b)->f = e ## 32toh((b)->f);	\
98 		break;				\
99 	case 8:					\
100 		(b)->f = e ## 64toh((b)->f);	\
101 		break;				\
102 	default:				\
103 		/* Force a link time error. */	\
104 		elf_wrong_field_size();		\
105 		break;				\
106 	}
107 
108 #define CONVERT_SWITCH(h, d, f)			\
109 	switch ((h)->e_ident[EI_DATA]) {	\
110 	case ELFDATA2MSB:			\
111 		f(d, be);			\
112 		break;				\
113 	case ELFDATA2LSB:			\
114 		f(d, le);			\
115 		break;				\
116 	default:				\
117 		return (EINVAL);		\
118 	}
119 
120 
121 static int elf_header_convert(Elf_Ehdr *ehdr)
122 {
123 	/*
124 	 * Fixup ELF header endianness.
125 	 *
126 	 * The Xhdr structure was loaded using block read call to optimize file
127 	 * accesses. It might happen, that the endianness of the system memory
128 	 * is different that endianness of the ELF header.  Swap fields here to
129 	 * guarantee that Xhdr always contain valid data regardless of
130 	 * architecture.
131 	 */
132 #define HEADER_FIELDS(b, e)			\
133 	CONVERT_FIELD(b, e_type, e);		\
134 	CONVERT_FIELD(b, e_machine, e);		\
135 	CONVERT_FIELD(b, e_version, e);		\
136 	CONVERT_FIELD(b, e_entry, e);		\
137 	CONVERT_FIELD(b, e_phoff, e);		\
138 	CONVERT_FIELD(b, e_shoff, e);		\
139 	CONVERT_FIELD(b, e_flags, e);		\
140 	CONVERT_FIELD(b, e_ehsize, e);		\
141 	CONVERT_FIELD(b, e_phentsize, e);	\
142 	CONVERT_FIELD(b, e_phnum, e);		\
143 	CONVERT_FIELD(b, e_shentsize, e);	\
144 	CONVERT_FIELD(b, e_shnum, e);		\
145 	CONVERT_FIELD(b, e_shstrndx, e)
146 
147 	CONVERT_SWITCH(ehdr, ehdr, HEADER_FIELDS);
148 
149 #undef HEADER_FIELDS
150 
151 	return (0);
152 }
153 
154 static int elf_program_header_convert(const Elf_Ehdr *ehdr, Elf_Phdr *phdr)
155 {
156 #define PROGRAM_HEADER_FIELDS(b, e)		\
157 	CONVERT_FIELD(b, p_type, e);		\
158 	CONVERT_FIELD(b, p_flags, e);		\
159 	CONVERT_FIELD(b, p_offset, e);		\
160 	CONVERT_FIELD(b, p_vaddr, e);		\
161 	CONVERT_FIELD(b, p_paddr, e);		\
162 	CONVERT_FIELD(b, p_filesz, e);		\
163 	CONVERT_FIELD(b, p_memsz, e);		\
164 	CONVERT_FIELD(b, p_align, e)
165 
166 	CONVERT_SWITCH(ehdr, phdr, PROGRAM_HEADER_FIELDS);
167 
168 #undef PROGRAM_HEADER_FIELDS
169 
170 	return (0);
171 }
172 
173 static int elf_section_header_convert(const Elf_Ehdr *ehdr, Elf_Shdr *shdr)
174 {
175 #define SECTION_HEADER_FIELDS(b, e)		\
176 	CONVERT_FIELD(b, sh_name, e);		\
177 	CONVERT_FIELD(b, sh_type, e);		\
178 	CONVERT_FIELD(b, sh_link, e);		\
179 	CONVERT_FIELD(b, sh_info, e);		\
180 	CONVERT_FIELD(b, sh_flags, e);		\
181 	CONVERT_FIELD(b, sh_addr, e);		\
182 	CONVERT_FIELD(b, sh_offset, e);		\
183 	CONVERT_FIELD(b, sh_size, e);		\
184 	CONVERT_FIELD(b, sh_addralign, e);	\
185 	CONVERT_FIELD(b, sh_entsize, e)
186 
187 	CONVERT_SWITCH(ehdr, shdr, SECTION_HEADER_FIELDS);
188 
189 #undef SECTION_HEADER_FIELDS
190 
191 	return (0);
192 }
193 #undef CONVERT_SWITCH
194 #undef CONVERT_FIELD
195 
196 static int
197 __elfN(load_elf_header)(char *filename, elf_file_t ef)
198 {
199 	ssize_t			 bytes_read;
200 	Elf_Ehdr		*ehdr;
201 	int 			 err;
202 
203 	/*
204 	* Open the image, read and validate the ELF header
205 	*/
206 	if (filename == NULL)	/* can't handle nameless */
207 		return (EFTYPE);
208 	if ((ef->fd = open(filename, O_RDONLY)) == -1)
209 		return (errno);
210 	ef->firstpage = malloc(PAGE_SIZE);
211 	if (ef->firstpage == NULL) {
212 		close(ef->fd);
213 		return (ENOMEM);
214 	}
215 	bytes_read = read(ef->fd, ef->firstpage, PAGE_SIZE);
216 	ef->firstlen = (size_t)bytes_read;
217 	if (bytes_read < 0 || ef->firstlen <= sizeof(Elf_Ehdr)) {
218 		err = EFTYPE; /* could be EIO, but may be small file */
219 		goto error;
220 	}
221 	ehdr = ef->ehdr = (Elf_Ehdr *)ef->firstpage;
222 
223 	/* Is it ELF? */
224 	if (!IS_ELF(*ehdr)) {
225 		err = EFTYPE;
226 		goto error;
227 	}
228 
229 	if (ehdr->e_ident[EI_CLASS] != ELF_TARG_CLASS || /* Layout ? */
230 	    ehdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
231 	    ehdr->e_ident[EI_VERSION] != EV_CURRENT) /* Version ? */ {
232 		err = EFTYPE;
233 		goto error;
234 	}
235 
236 	err = elf_header_convert(ehdr);
237 	if (err)
238 		goto error;
239 
240 	if (ehdr->e_version != EV_CURRENT || ehdr->e_machine != ELF_TARG_MACH) { /* Machine ? */
241 		err = EFTYPE;
242 		goto error;
243 	}
244 
245 	return (0);
246 
247 error:
248 	if (ef->firstpage != NULL) {
249 		free(ef->firstpage);
250 		ef->firstpage = NULL;
251 	}
252 	if (ef->fd != -1) {
253 		close(ef->fd);
254 		ef->fd = -1;
255 	}
256 	return (err);
257 }
258 
259 /*
260  * Attempt to load the file (file) as an ELF module.  It will be stored at
261  * (dest), and a pointer to a module structure describing the loaded object
262  * will be saved in (result).
263  */
264 int
265 __elfN(loadfile)(char *filename, u_int64_t dest, struct preloaded_file **result)
266 {
267 	return (__elfN(loadfile_raw)(filename, dest, result, 0));
268 }
269 
270 int
271 __elfN(loadfile_raw)(char *filename, u_int64_t dest,
272     struct preloaded_file **result, int multiboot)
273 {
274     struct preloaded_file	*fp, *kfp;
275     struct elf_file		ef;
276     Elf_Ehdr 			*ehdr;
277     int				err;
278 
279     fp = NULL;
280     bzero(&ef, sizeof(struct elf_file));
281     ef.fd = -1;
282 
283     err = __elfN(load_elf_header)(filename, &ef);
284     if (err != 0)
285     	return (err);
286 
287     ehdr = ef.ehdr;
288 
289     /*
290      * Check to see what sort of module we are.
291      */
292     kfp = file_findfile(NULL, __elfN(kerneltype));
293 #ifdef __powerpc__
294     /*
295      * Kernels can be ET_DYN, so just assume the first loaded object is the
296      * kernel. This assumption will be checked later.
297      */
298     if (kfp == NULL)
299         ef.kernel = 1;
300 #endif
301     if (ef.kernel || ehdr->e_type == ET_EXEC) {
302 	/* Looks like a kernel */
303 	if (kfp != NULL) {
304 	    printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: kernel already loaded\n");
305 	    err = EPERM;
306 	    goto oerr;
307 	}
308 	/*
309 	 * Calculate destination address based on kernel entrypoint.
310 	 *
311 	 * For ARM, the destination address is independent of any values in the
312 	 * elf header (an ARM kernel can be loaded at any 2MB boundary), so we
313 	 * leave dest set to the value calculated by archsw.arch_loadaddr() and
314 	 * passed in to this function.
315 	 */
316 #ifndef __arm__
317         if (ehdr->e_type == ET_EXEC)
318 	    dest = (ehdr->e_entry & ~PAGE_MASK);
319 #endif
320 	if ((ehdr->e_entry & ~PAGE_MASK) == 0) {
321 	    printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: not a kernel (maybe static binary?)\n");
322 	    err = EPERM;
323 	    goto oerr;
324 	}
325 	ef.kernel = 1;
326 
327     } else if (ehdr->e_type == ET_DYN) {
328 	/* Looks like a kld module */
329 	if (multiboot != 0) {
330 		printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: can't load module as multiboot\n");
331 		err = EPERM;
332 		goto oerr;
333 	}
334 	if (kfp == NULL) {
335 	    printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: can't load module before kernel\n");
336 	    err = EPERM;
337 	    goto oerr;
338 	}
339 	if (strcmp(__elfN(kerneltype), kfp->f_type)) {
340 	    printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: can't load module with kernel type '%s'\n", kfp->f_type);
341 	    err = EPERM;
342 	    goto oerr;
343 	}
344 	/* Looks OK, got ahead */
345 	ef.kernel = 0;
346 
347     } else {
348 	err = EFTYPE;
349 	goto oerr;
350     }
351 
352     if (archsw.arch_loadaddr != NULL)
353 	dest = archsw.arch_loadaddr(LOAD_ELF, ehdr, dest);
354     else
355 	dest = roundup(dest, PAGE_SIZE);
356 
357     /*
358      * Ok, we think we should handle this.
359      */
360     fp = file_alloc();
361     if (fp == NULL) {
362 	    printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadfile: cannot allocate module info\n");
363 	    err = EPERM;
364 	    goto out;
365     }
366     if (ef.kernel == 1 && multiboot == 0)
367 	setenv("kernelname", filename, 1);
368     fp->f_name = strdup(filename);
369     if (multiboot == 0)
370     	fp->f_type = strdup(ef.kernel ?
371     	    __elfN(kerneltype) : __elfN(moduletype));
372     else
373     	fp->f_type = strdup("elf multiboot kernel");
374 
375 #ifdef ELF_VERBOSE
376     if (ef.kernel)
377 	printf("%s entry at 0x%jx\n", filename, (uintmax_t)ehdr->e_entry);
378 #else
379     printf("%s ", filename);
380 #endif
381 
382     fp->f_size = __elfN(loadimage)(fp, &ef, dest);
383     if (fp->f_size == 0 || fp->f_addr == 0)
384 	goto ioerr;
385 
386     /* save exec header as metadata */
387     file_addmetadata(fp, MODINFOMD_ELFHDR, sizeof(*ehdr), ehdr);
388 
389     /* Load OK, return module pointer */
390     *result = (struct preloaded_file *)fp;
391     err = 0;
392     goto out;
393 
394  ioerr:
395     err = EIO;
396  oerr:
397     file_discard(fp);
398  out:
399     if (ef.firstpage)
400 	free(ef.firstpage);
401     if (ef.fd != -1)
402     	close(ef.fd);
403     return(err);
404 }
405 
406 /*
407  * With the file (fd) open on the image, and (ehdr) containing
408  * the Elf header, load the image at (off)
409  */
410 static int
411 __elfN(loadimage)(struct preloaded_file *fp, elf_file_t ef, u_int64_t off)
412 {
413     int 	i;
414     u_int	j;
415     Elf_Ehdr	*ehdr;
416     Elf_Phdr	*phdr, *php;
417     Elf_Shdr	*shdr;
418     char	*shstr;
419     int		ret;
420     vm_offset_t firstaddr;
421     vm_offset_t lastaddr;
422     size_t	chunk;
423     ssize_t	result;
424     Elf_Addr	ssym, esym;
425     Elf_Dyn	*dp;
426     Elf_Addr	adp;
427     Elf_Addr	ctors;
428     int		ndp;
429     int		symstrindex;
430     int		symtabindex;
431     Elf_Size	size;
432     u_int	fpcopy;
433     Elf_Sym	sym;
434     Elf_Addr	p_start, p_end;
435 
436     dp = NULL;
437     shdr = NULL;
438     ret = 0;
439     firstaddr = lastaddr = 0;
440     ehdr = ef->ehdr;
441     if (ehdr->e_type == ET_EXEC) {
442 #if defined(__i386__) || defined(__amd64__)
443 #if __ELF_WORD_SIZE == 64
444 	off = - (off & 0xffffffffff000000ull);/* x86_64 relocates after locore */
445 #else
446 	off = - (off & 0xff000000u);	/* i386 relocates after locore */
447 #endif
448 #elif defined(__powerpc__)
449 	/*
450 	 * On the purely virtual memory machines like e500, the kernel is
451 	 * linked against its final VA range, which is most often not
452 	 * available at the loader stage, but only after kernel initializes
453 	 * and completes its VM settings. In such cases we cannot use p_vaddr
454 	 * field directly to load ELF segments, but put them at some
455 	 * 'load-time' locations.
456 	 */
457 	if (off & 0xf0000000u) {
458 	    off = -(off & 0xf0000000u);
459 	    /*
460 	     * XXX the physical load address should not be hardcoded. Note
461 	     * that the Book-E kernel assumes that it's loaded at a 16MB
462 	     * boundary for now...
463 	     */
464 	    off += 0x01000000;
465 	    ehdr->e_entry += off;
466 #ifdef ELF_VERBOSE
467 	    printf("Converted entry 0x%08x\n", ehdr->e_entry);
468 #endif
469 	} else
470 	    off = 0;
471 #elif defined(__arm__) && !defined(EFI)
472 	/*
473 	 * The elf headers in arm kernels specify virtual addresses in all
474 	 * header fields, even the ones that should be physical addresses.
475 	 * We assume the entry point is in the first page, and masking the page
476 	 * offset will leave us with the virtual address the kernel was linked
477 	 * at.  We subtract that from the load offset, making 'off' into the
478 	 * value which, when added to a virtual address in an elf header,
479 	 * translates it to a physical address.  We do the va->pa conversion on
480 	 * the entry point address in the header now, so that later we can
481 	 * launch the kernel by just jumping to that address.
482 	 *
483 	 * When booting from UEFI the copyin and copyout functions handle
484 	 * adjusting the location relative to the first virtual address.
485 	 * Because of this there is no need to adjust the offset or entry
486 	 * point address as these will both be handled by the efi code.
487 	 */
488 	off -= ehdr->e_entry & ~PAGE_MASK;
489 	ehdr->e_entry += off;
490 #ifdef ELF_VERBOSE
491 	printf("ehdr->e_entry 0x%08x, va<->pa off %llx\n", ehdr->e_entry, off);
492 #endif
493 #else
494 	off = 0;		/* other archs use direct mapped kernels */
495 #endif
496     }
497     ef->off = off;
498 
499     if (ef->kernel)
500 	__elfN(relocation_offset) = off;
501 
502     if ((ehdr->e_phoff + ehdr->e_phnum * sizeof(*phdr)) > ef->firstlen) {
503 	printf("elf" __XSTRING(__ELF_WORD_SIZE) "_loadimage: program header not within first page\n");
504 	goto out;
505     }
506     phdr = (Elf_Phdr *)(ef->firstpage + ehdr->e_phoff);
507 
508     for (i = 0; i < ehdr->e_phnum; i++) {
509 	if (elf_program_header_convert(ehdr, phdr))
510 	    continue;
511 
512 	/* We want to load PT_LOAD segments only.. */
513 	if (phdr[i].p_type != PT_LOAD)
514 	    continue;
515 
516 #ifdef ELF_VERBOSE
517 	printf("Segment: 0x%lx@0x%lx -> 0x%lx-0x%lx",
518 	    (long)phdr[i].p_filesz, (long)phdr[i].p_offset,
519 	    (long)(phdr[i].p_vaddr + off),
520 	    (long)(phdr[i].p_vaddr + off + phdr[i].p_memsz - 1));
521 #else
522 	if ((phdr[i].p_flags & PF_W) == 0) {
523 	    printf("text=0x%lx ", (long)phdr[i].p_filesz);
524 	} else {
525 	    printf("data=0x%lx", (long)phdr[i].p_filesz);
526 	    if (phdr[i].p_filesz < phdr[i].p_memsz)
527 		printf("+0x%lx", (long)(phdr[i].p_memsz -phdr[i].p_filesz));
528 	    printf(" ");
529 	}
530 #endif
531 	fpcopy = 0;
532 	if (ef->firstlen > phdr[i].p_offset) {
533 	    fpcopy = ef->firstlen - phdr[i].p_offset;
534 	    archsw.arch_copyin(ef->firstpage + phdr[i].p_offset,
535 			       phdr[i].p_vaddr + off, fpcopy);
536 	}
537 	if (phdr[i].p_filesz > fpcopy) {
538 	    if (kern_pread(ef->fd, phdr[i].p_vaddr + off + fpcopy,
539 		phdr[i].p_filesz - fpcopy, phdr[i].p_offset + fpcopy) != 0) {
540 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
541 		    "_loadimage: read failed\n");
542 		goto out;
543 	    }
544 	}
545 	/* clear space from oversized segments; eg: bss */
546 	if (phdr[i].p_filesz < phdr[i].p_memsz) {
547 #ifdef ELF_VERBOSE
548 	    printf(" (bss: 0x%lx-0x%lx)",
549 		(long)(phdr[i].p_vaddr + off + phdr[i].p_filesz),
550 		(long)(phdr[i].p_vaddr + off + phdr[i].p_memsz - 1));
551 #endif
552 
553 	    kern_bzero(phdr[i].p_vaddr + off + phdr[i].p_filesz,
554 		phdr[i].p_memsz - phdr[i].p_filesz);
555 	}
556 #ifdef ELF_VERBOSE
557 	printf("\n");
558 #endif
559 
560 	if (archsw.arch_loadseg != NULL)
561 	    archsw.arch_loadseg(ehdr, phdr + i, off);
562 
563 	if (firstaddr == 0 || firstaddr > (phdr[i].p_vaddr + off))
564 	    firstaddr = phdr[i].p_vaddr + off;
565 	if (lastaddr == 0 || lastaddr < (phdr[i].p_vaddr + off + phdr[i].p_memsz))
566 	    lastaddr = phdr[i].p_vaddr + off + phdr[i].p_memsz;
567     }
568     lastaddr = roundup(lastaddr, sizeof(long));
569 
570     /*
571      * Get the section headers.  We need this for finding the .ctors
572      * section as well as for loading any symbols.  Both may be hard
573      * to do if reading from a .gz file as it involves seeking.  I
574      * think the rule is going to have to be that you must strip a
575      * file to remove symbols before gzipping it.
576      */
577     chunk = (size_t)ehdr->e_shnum * (size_t)ehdr->e_shentsize;
578     if (chunk == 0 || ehdr->e_shoff == 0)
579 	goto nosyms;
580     shdr = alloc_pread(ef->fd, ehdr->e_shoff, chunk);
581     if (shdr == NULL) {
582 	printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
583 	    "_loadimage: failed to read section headers");
584 	goto nosyms;
585     }
586 
587     for (i = 0; i < ehdr->e_shnum; i++)
588 	elf_section_header_convert(ehdr, &shdr[i]);
589 
590     file_addmetadata(fp, MODINFOMD_SHDR, chunk, shdr);
591 
592     /*
593      * Read the section string table and look for the .ctors section.
594      * We need to tell the kernel where it is so that it can call the
595      * ctors.
596      */
597     chunk = shdr[ehdr->e_shstrndx].sh_size;
598     if (chunk) {
599 	shstr = alloc_pread(ef->fd, shdr[ehdr->e_shstrndx].sh_offset, chunk);
600 	if (shstr) {
601 	    for (i = 0; i < ehdr->e_shnum; i++) {
602 		if (strcmp(shstr + shdr[i].sh_name, ".ctors") != 0)
603 		    continue;
604 		ctors = shdr[i].sh_addr;
605 		file_addmetadata(fp, MODINFOMD_CTORS_ADDR, sizeof(ctors),
606 		    &ctors);
607 		size = shdr[i].sh_size;
608 		file_addmetadata(fp, MODINFOMD_CTORS_SIZE, sizeof(size),
609 		    &size);
610 		break;
611 	    }
612 	    free(shstr);
613 	}
614     }
615 
616     /*
617      * Now load any symbols.
618      */
619     symtabindex = -1;
620     symstrindex = -1;
621     for (i = 0; i < ehdr->e_shnum; i++) {
622 	if (shdr[i].sh_type != SHT_SYMTAB)
623 	    continue;
624 	for (j = 0; j < ehdr->e_phnum; j++) {
625 	    if (phdr[j].p_type != PT_LOAD)
626 		continue;
627 	    if (shdr[i].sh_offset >= phdr[j].p_offset &&
628 		(shdr[i].sh_offset + shdr[i].sh_size <=
629 		 phdr[j].p_offset + phdr[j].p_filesz)) {
630 		shdr[i].sh_offset = 0;
631 		shdr[i].sh_size = 0;
632 		break;
633 	    }
634 	}
635 	if (shdr[i].sh_offset == 0 || shdr[i].sh_size == 0)
636 	    continue;		/* alread loaded in a PT_LOAD above */
637 	/* Save it for loading below */
638 	symtabindex = i;
639 	symstrindex = shdr[i].sh_link;
640     }
641     if (symtabindex < 0 || symstrindex < 0)
642 	goto nosyms;
643 
644     /* Ok, committed to a load. */
645 #ifndef ELF_VERBOSE
646     printf("syms=[");
647 #endif
648     ssym = lastaddr;
649     for (i = symtabindex; i >= 0; i = symstrindex) {
650 #ifdef ELF_VERBOSE
651 	char	*secname;
652 
653 	switch(shdr[i].sh_type) {
654 	    case SHT_SYMTAB:		/* Symbol table */
655 		secname = "symtab";
656 		break;
657 	    case SHT_STRTAB:		/* String table */
658 		secname = "strtab";
659 		break;
660 	    default:
661 		secname = "WHOA!!";
662 		break;
663 	}
664 #endif
665 	size = shdr[i].sh_size;
666 #if defined(__powerpc__)
667   #if __ELF_WORD_SIZE == 64
668 	size = htobe64(size);
669   #else
670 	size = htobe32(size);
671   #endif
672 #endif
673 
674 	archsw.arch_copyin(&size, lastaddr, sizeof(size));
675 	lastaddr += sizeof(size);
676 
677 #ifdef ELF_VERBOSE
678 	printf("\n%s: 0x%jx@0x%jx -> 0x%jx-0x%jx", secname,
679 	    (uintmax_t)shdr[i].sh_size, (uintmax_t)shdr[i].sh_offset,
680 	    (uintmax_t)lastaddr, (uintmax_t)(lastaddr + shdr[i].sh_size));
681 #else
682 	if (i == symstrindex)
683 	    printf("+");
684 	printf("0x%lx+0x%lx", (long)sizeof(size), (long)size);
685 #endif
686 
687 	if (lseek(ef->fd, (off_t)shdr[i].sh_offset, SEEK_SET) == -1) {
688 	    printf("\nelf" __XSTRING(__ELF_WORD_SIZE) "_loadimage: could not seek for symbols - skipped!");
689 	    lastaddr = ssym;
690 	    ssym = 0;
691 	    goto nosyms;
692 	}
693 	result = archsw.arch_readin(ef->fd, lastaddr, shdr[i].sh_size);
694 	if (result < 0 || (size_t)result != shdr[i].sh_size) {
695 	    printf("\nelf" __XSTRING(__ELF_WORD_SIZE) "_loadimage: could not read symbols - skipped! (%ju != %ju)", (uintmax_t)result,
696 		(uintmax_t)shdr[i].sh_size);
697 	    lastaddr = ssym;
698 	    ssym = 0;
699 	    goto nosyms;
700 	}
701 	/* Reset offsets relative to ssym */
702 	lastaddr += shdr[i].sh_size;
703 	lastaddr = roundup(lastaddr, sizeof(size));
704 	if (i == symtabindex)
705 	    symtabindex = -1;
706 	else if (i == symstrindex)
707 	    symstrindex = -1;
708     }
709     esym = lastaddr;
710 #ifndef ELF_VERBOSE
711     printf("]");
712 #endif
713 
714 #if defined(__powerpc__)
715   /* On PowerPC we always need to provide BE data to the kernel */
716   #if __ELF_WORD_SIZE == 64
717     ssym = htobe64((uint64_t)ssym);
718     esym = htobe64((uint64_t)esym);
719   #else
720     ssym = htobe32((uint32_t)ssym);
721     esym = htobe32((uint32_t)esym);
722   #endif
723 #endif
724 
725     file_addmetadata(fp, MODINFOMD_SSYM, sizeof(ssym), &ssym);
726     file_addmetadata(fp, MODINFOMD_ESYM, sizeof(esym), &esym);
727 
728 nosyms:
729     printf("\n");
730 
731     ret = lastaddr - firstaddr;
732     fp->f_addr = firstaddr;
733 
734     php = NULL;
735     for (i = 0; i < ehdr->e_phnum; i++) {
736 	if (phdr[i].p_type == PT_DYNAMIC) {
737 	    php = phdr + i;
738 	    adp = php->p_vaddr;
739 	    file_addmetadata(fp, MODINFOMD_DYNAMIC, sizeof(adp), &adp);
740 	    break;
741 	}
742     }
743 
744     if (php == NULL)	/* this is bad, we cannot get to symbols or _DYNAMIC */
745 	goto out;
746 
747     ndp = php->p_filesz / sizeof(Elf_Dyn);
748     if (ndp == 0)
749 	goto out;
750     dp = malloc(php->p_filesz);
751     if (dp == NULL)
752 	goto out;
753     archsw.arch_copyout(php->p_vaddr + off, dp, php->p_filesz);
754 
755     ef->strsz = 0;
756     for (i = 0; i < ndp; i++) {
757 	if (dp[i].d_tag == 0)
758 	    break;
759 	switch (dp[i].d_tag) {
760 	case DT_HASH:
761 	    ef->hashtab = (Elf_Hashelt*)(uintptr_t)(dp[i].d_un.d_ptr + off);
762 	    break;
763 	case DT_STRTAB:
764 	    ef->strtab = (char *)(uintptr_t)(dp[i].d_un.d_ptr + off);
765 	    break;
766 	case DT_STRSZ:
767 	    ef->strsz = dp[i].d_un.d_val;
768 	    break;
769 	case DT_SYMTAB:
770 	    ef->symtab = (Elf_Sym*)(uintptr_t)(dp[i].d_un.d_ptr + off);
771 	    break;
772 	case DT_REL:
773 	    ef->rel = (Elf_Rel *)(uintptr_t)(dp[i].d_un.d_ptr + off);
774 	    break;
775 	case DT_RELSZ:
776 	    ef->relsz = dp[i].d_un.d_val;
777 	    break;
778 	case DT_RELA:
779 	    ef->rela = (Elf_Rela *)(uintptr_t)(dp[i].d_un.d_ptr + off);
780 	    break;
781 	case DT_RELASZ:
782 	    ef->relasz = dp[i].d_un.d_val;
783 	    break;
784 	default:
785 	    break;
786 	}
787     }
788     if (ef->hashtab == NULL || ef->symtab == NULL ||
789 	ef->strtab == NULL || ef->strsz == 0)
790 	goto out;
791     COPYOUT(ef->hashtab, &ef->nbuckets, sizeof(ef->nbuckets));
792     COPYOUT(ef->hashtab + 1, &ef->nchains, sizeof(ef->nchains));
793     ef->buckets = ef->hashtab + 2;
794     ef->chains = ef->buckets + ef->nbuckets;
795 
796     if (__elfN(lookup_symbol)(fp, ef, "__start_set_modmetadata_set", &sym) != 0)
797 	return 0;
798     p_start = sym.st_value + ef->off;
799     if (__elfN(lookup_symbol)(fp, ef, "__stop_set_modmetadata_set", &sym) != 0)
800 	return ENOENT;
801     p_end = sym.st_value + ef->off;
802 
803     if (__elfN(parse_modmetadata)(fp, ef, p_start, p_end) == 0)
804 	goto out;
805 
806     if (ef->kernel)			/* kernel must not depend on anything */
807 	goto out;
808 
809 out:
810     if (dp)
811 	free(dp);
812     if (shdr)
813 	free(shdr);
814     return ret;
815 }
816 
817 static char invalid_name[] = "bad";
818 
819 char *
820 fake_modname(const char *name)
821 {
822     const char *sp, *ep;
823     char *fp;
824     size_t len;
825 
826     sp = strrchr(name, '/');
827     if (sp)
828 	sp++;
829     else
830 	sp = name;
831     ep = strrchr(name, '.');
832     if (ep) {
833 	    if (ep == name) {
834 		sp = invalid_name;
835 		ep = invalid_name + sizeof(invalid_name) - 1;
836 	    }
837     } else
838 	ep = name + strlen(name);
839     len = ep - sp;
840     fp = malloc(len + 1);
841     if (fp == NULL)
842 	return NULL;
843     memcpy(fp, sp, len);
844     fp[len] = '\0';
845     return fp;
846 }
847 
848 #if (defined(__i386__) || defined(__powerpc__)) && __ELF_WORD_SIZE == 64
849 struct mod_metadata64 {
850 	int		md_version;	/* structure version MDTV_* */
851 	int		md_type;	/* type of entry MDT_* */
852 	u_int64_t	md_data;	/* specific data */
853 	u_int64_t	md_cval;	/* common string label */
854 };
855 #endif
856 #if defined(__amd64__) && __ELF_WORD_SIZE == 32
857 struct mod_metadata32 {
858 	int		md_version;	/* structure version MDTV_* */
859 	int		md_type;	/* type of entry MDT_* */
860 	u_int32_t	md_data;	/* specific data */
861 	u_int32_t	md_cval;	/* common string label */
862 };
863 #endif
864 
865 int
866 __elfN(load_modmetadata)(struct preloaded_file *fp, u_int64_t dest)
867 {
868 	struct elf_file		 ef;
869 	int			 err, i, j;
870 	Elf_Shdr		*sh_meta, *shdr = NULL;
871 	Elf_Shdr		*sh_data[2];
872 	char			*shstrtab = NULL;
873 	size_t			 size;
874 	Elf_Addr		 p_start, p_end;
875 
876 	bzero(&ef, sizeof(struct elf_file));
877 	ef.fd = -1;
878 
879 	err = __elfN(load_elf_header)(fp->f_name, &ef);
880 	if (err != 0)
881 		goto out;
882 
883 	if (ef.kernel == 1 || ef.ehdr->e_type == ET_EXEC) {
884 		ef.kernel = 1;
885 	} else if (ef.ehdr->e_type != ET_DYN) {
886 		err = EFTYPE;
887 		goto out;
888 	}
889 
890 	size = (size_t)ef.ehdr->e_shnum * (size_t)ef.ehdr->e_shentsize;
891 	shdr = alloc_pread(ef.fd, ef.ehdr->e_shoff, size);
892 	if (shdr == NULL) {
893 		err = ENOMEM;
894 		goto out;
895 	}
896 
897 	/* Load shstrtab. */
898 	shstrtab = alloc_pread(ef.fd, shdr[ef.ehdr->e_shstrndx].sh_offset,
899 	    shdr[ef.ehdr->e_shstrndx].sh_size);
900 	if (shstrtab == NULL) {
901 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
902 		    "load_modmetadata: unable to load shstrtab\n");
903 		err = EFTYPE;
904 		goto out;
905 	}
906 
907 	/* Find set_modmetadata_set and data sections. */
908 	sh_data[0] = sh_data[1] = sh_meta = NULL;
909 	for (i = 0, j = 0; i < ef.ehdr->e_shnum; i++) {
910 		if (strcmp(&shstrtab[shdr[i].sh_name],
911 		    "set_modmetadata_set") == 0) {
912 			sh_meta = &shdr[i];
913 		}
914 		if ((strcmp(&shstrtab[shdr[i].sh_name], ".data") == 0) ||
915 		    (strcmp(&shstrtab[shdr[i].sh_name], ".rodata") == 0)) {
916 			sh_data[j++] = &shdr[i];
917 		}
918 	}
919 	if (sh_meta == NULL || sh_data[0] == NULL || sh_data[1] == NULL) {
920 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
921     "load_modmetadata: unable to find set_modmetadata_set or data sections\n");
922 		err = EFTYPE;
923 		goto out;
924 	}
925 
926 	/* Load set_modmetadata_set into memory */
927 	err = kern_pread(ef.fd, dest, sh_meta->sh_size, sh_meta->sh_offset);
928 	if (err != 0) {
929 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
930     "load_modmetadata: unable to load set_modmetadata_set: %d\n", err);
931 		goto out;
932 	}
933 	p_start = dest;
934 	p_end = dest + sh_meta->sh_size;
935 	dest += sh_meta->sh_size;
936 
937 	/* Load data sections into memory. */
938 	err = kern_pread(ef.fd, dest, sh_data[0]->sh_size,
939 	    sh_data[0]->sh_offset);
940 	if (err != 0) {
941 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
942 		    "load_modmetadata: unable to load data: %d\n", err);
943 		goto out;
944 	}
945 
946 	/*
947 	 * We have to increment the dest, so that the offset is the same into
948 	 * both the .rodata and .data sections.
949 	 */
950 	ef.off = -(sh_data[0]->sh_addr - dest);
951 	dest +=	(sh_data[1]->sh_addr - sh_data[0]->sh_addr);
952 
953 	err = kern_pread(ef.fd, dest, sh_data[1]->sh_size,
954 	    sh_data[1]->sh_offset);
955 	if (err != 0) {
956 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
957 		    "load_modmetadata: unable to load data: %d\n", err);
958 		goto out;
959 	}
960 
961 	err = __elfN(parse_modmetadata)(fp, &ef, p_start, p_end);
962 	if (err != 0) {
963 		printf("\nelf" __XSTRING(__ELF_WORD_SIZE)
964 		    "load_modmetadata: unable to parse metadata: %d\n", err);
965 		goto out;
966 	}
967 
968 out:
969 	if (shstrtab != NULL)
970 		free(shstrtab);
971 	if (shdr != NULL)
972 		free(shdr);
973 	if (ef.firstpage != NULL)
974 		free(ef.firstpage);
975 	if (ef.fd != -1)
976 		close(ef.fd);
977 	return (err);
978 }
979 
980 int
981 __elfN(parse_modmetadata)(struct preloaded_file *fp, elf_file_t ef,
982     Elf_Addr p_start, Elf_Addr p_end)
983 {
984     struct mod_metadata md;
985 #if (defined(__i386__) || defined(__powerpc__)) && __ELF_WORD_SIZE == 64
986     struct mod_metadata64 md64;
987 #elif defined(__amd64__) && __ELF_WORD_SIZE == 32
988     struct mod_metadata32 md32;
989 #endif
990     struct mod_depend *mdepend;
991     struct mod_version mver;
992     char *s;
993     int error, modcnt, minfolen;
994     Elf_Addr v, p;
995 
996     modcnt = 0;
997     p = p_start;
998     while (p < p_end) {
999 	COPYOUT(p, &v, sizeof(v));
1000 	error = __elfN(reloc_ptr)(fp, ef, p, &v, sizeof(v));
1001 	if (error == EOPNOTSUPP)
1002 	    v += ef->off;
1003 	else if (error != 0)
1004 	    return (error);
1005 #if (defined(__i386__) || defined(__powerpc__)) && __ELF_WORD_SIZE == 64
1006 	COPYOUT(v, &md64, sizeof(md64));
1007 	error = __elfN(reloc_ptr)(fp, ef, v, &md64, sizeof(md64));
1008 	if (error == EOPNOTSUPP) {
1009 	    md64.md_cval += ef->off;
1010 	    md64.md_data += ef->off;
1011 	} else if (error != 0)
1012 	    return (error);
1013 	md.md_version = md64.md_version;
1014 	md.md_type = md64.md_type;
1015 	md.md_cval = (const char *)(uintptr_t)md64.md_cval;
1016 	md.md_data = (void *)(uintptr_t)md64.md_data;
1017 #elif defined(__amd64__) && __ELF_WORD_SIZE == 32
1018 	COPYOUT(v, &md32, sizeof(md32));
1019 	error = __elfN(reloc_ptr)(fp, ef, v, &md32, sizeof(md32));
1020 	if (error == EOPNOTSUPP) {
1021 	    md32.md_cval += ef->off;
1022 	    md32.md_data += ef->off;
1023 	} else if (error != 0)
1024 	    return (error);
1025 	md.md_version = md32.md_version;
1026 	md.md_type = md32.md_type;
1027 	md.md_cval = (const char *)(uintptr_t)md32.md_cval;
1028 	md.md_data = (void *)(uintptr_t)md32.md_data;
1029 #else
1030 	COPYOUT(v, &md, sizeof(md));
1031 	error = __elfN(reloc_ptr)(fp, ef, v, &md, sizeof(md));
1032 	if (error == EOPNOTSUPP) {
1033 	    md.md_cval += ef->off;
1034 	    md.md_data = (void *)((uintptr_t)md.md_data + (uintptr_t)ef->off);
1035 	} else if (error != 0)
1036 	    return (error);
1037 #endif
1038 	p += sizeof(Elf_Addr);
1039 	switch(md.md_type) {
1040 	  case MDT_DEPEND:
1041 	    if (ef->kernel)		/* kernel must not depend on anything */
1042 	      break;
1043 	    s = strdupout((vm_offset_t)md.md_cval);
1044 	    minfolen = sizeof(*mdepend) + strlen(s) + 1;
1045 	    mdepend = malloc(minfolen);
1046 	    if (mdepend == NULL)
1047 		return ENOMEM;
1048 	    COPYOUT((vm_offset_t)md.md_data, mdepend, sizeof(*mdepend));
1049 	    strcpy((char*)(mdepend + 1), s);
1050 	    free(s);
1051 	    file_addmetadata(fp, MODINFOMD_DEPLIST, minfolen, mdepend);
1052 	    free(mdepend);
1053 	    break;
1054 	  case MDT_VERSION:
1055 	    s = strdupout((vm_offset_t)md.md_cval);
1056 	    COPYOUT((vm_offset_t)md.md_data, &mver, sizeof(mver));
1057 	    file_addmodule(fp, s, mver.mv_version, NULL);
1058 	    free(s);
1059 	    modcnt++;
1060 	    break;
1061 	}
1062     }
1063     if (modcnt == 0) {
1064 	s = fake_modname(fp->f_name);
1065 	file_addmodule(fp, s, 1, NULL);
1066 	free(s);
1067     }
1068     return 0;
1069 }
1070 
1071 static unsigned long
1072 elf_hash(const char *name)
1073 {
1074     const unsigned char *p = (const unsigned char *) name;
1075     unsigned long h = 0;
1076     unsigned long g;
1077 
1078     while (*p != '\0') {
1079 	h = (h << 4) + *p++;
1080 	if ((g = h & 0xf0000000) != 0)
1081 	    h ^= g >> 24;
1082 	h &= ~g;
1083     }
1084     return h;
1085 }
1086 
1087 static const char __elfN(bad_symtable)[] = "elf" __XSTRING(__ELF_WORD_SIZE) "_lookup_symbol: corrupt symbol table\n";
1088 int
1089 __elfN(lookup_symbol)(struct preloaded_file *fp, elf_file_t ef, const char* name,
1090 		  Elf_Sym *symp)
1091 {
1092     Elf_Hashelt symnum;
1093     Elf_Sym sym;
1094     char *strp;
1095     unsigned long hash;
1096 
1097     hash = elf_hash(name);
1098     COPYOUT(&ef->buckets[hash % ef->nbuckets], &symnum, sizeof(symnum));
1099 
1100     while (symnum != STN_UNDEF) {
1101 	if (symnum >= ef->nchains) {
1102 	    printf(__elfN(bad_symtable));
1103 	    return ENOENT;
1104 	}
1105 
1106 	COPYOUT(ef->symtab + symnum, &sym, sizeof(sym));
1107 	if (sym.st_name == 0) {
1108 	    printf(__elfN(bad_symtable));
1109 	    return ENOENT;
1110 	}
1111 
1112 	strp = strdupout((vm_offset_t)(ef->strtab + sym.st_name));
1113 	if (strcmp(name, strp) == 0) {
1114 	    free(strp);
1115 	    if (sym.st_shndx != SHN_UNDEF ||
1116 		(sym.st_value != 0 &&
1117 		 ELF_ST_TYPE(sym.st_info) == STT_FUNC)) {
1118 		*symp = sym;
1119 		return 0;
1120 	    }
1121 	    return ENOENT;
1122 	}
1123 	free(strp);
1124 	COPYOUT(&ef->chains[symnum], &symnum, sizeof(symnum));
1125     }
1126     return ENOENT;
1127 }
1128 
1129 /*
1130  * Apply any intra-module relocations to the value. p is the load address
1131  * of the value and val/len is the value to be modified. This does NOT modify
1132  * the image in-place, because this is done by kern_linker later on.
1133  *
1134  * Returns EOPNOTSUPP if no relocation method is supplied.
1135  */
1136 static int
1137 __elfN(reloc_ptr)(struct preloaded_file *mp, elf_file_t ef,
1138     Elf_Addr p, void *val, size_t len)
1139 {
1140 	size_t n;
1141 	Elf_Rela a;
1142 	Elf_Rel r;
1143 	int error;
1144 
1145 	/*
1146 	 * The kernel is already relocated, but we still want to apply
1147 	 * offset adjustments.
1148 	 */
1149 	if (ef->kernel)
1150 		return (EOPNOTSUPP);
1151 
1152 	for (n = 0; n < ef->relsz / sizeof(r); n++) {
1153 		COPYOUT(ef->rel + n, &r, sizeof(r));
1154 
1155 		error = __elfN(reloc)(ef, __elfN(symaddr), &r, ELF_RELOC_REL,
1156 		    ef->off, p, val, len);
1157 		if (error != 0)
1158 			return (error);
1159 	}
1160 	for (n = 0; n < ef->relasz / sizeof(a); n++) {
1161 		COPYOUT(ef->rela + n, &a, sizeof(a));
1162 
1163 		error = __elfN(reloc)(ef, __elfN(symaddr), &a, ELF_RELOC_RELA,
1164 		    ef->off, p, val, len);
1165 		if (error != 0)
1166 			return (error);
1167 	}
1168 
1169 	return (0);
1170 }
1171 
1172 static Elf_Addr
1173 __elfN(symaddr)(struct elf_file *ef, Elf_Size symidx)
1174 {
1175 
1176 	/* Symbol lookup by index not required here. */
1177 	return (0);
1178 }
1179