?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
kernel - Improve physio performance
[dragonfly.git] / sys / kern / imgact_elf.c
blob78f855b12fdf3bd943d4a3ce50d0eaef88b35279
1 /*-
2 * Copyright (c) 2000 David O'Brien
3 * Copyright (c) 1995-1996 Søren Schmidt
4 * Copyright (c) 1996 Peter Wemm
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 * in this position and unchanged.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. The name of the author may not be used to endorse or promote products
17 * derived from this software without specific prior written permission
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
20 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
21 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
22 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
24 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
28 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 *
30 * $FreeBSD: src/sys/kern/imgact_elf.c,v 1.73.2.13 2002/12/28 19:49:41 dillon Exp $
31 */
32
33 #include <sys/param.h>
34 #include <sys/exec.h>
35 #include <sys/fcntl.h>
36 #include <sys/file.h>
37 #include <sys/imgact.h>
38 #include <sys/imgact_elf.h>
39 #include <sys/kernel.h>
40 #include <sys/malloc.h>
41 #include <sys/mman.h>
42 #include <sys/systm.h>
43 #include <sys/proc.h>
44 #include <sys/nlookup.h>
45 #include <sys/pioctl.h>
46 #include <sys/procfs.h>
47 #include <sys/resourcevar.h>
48 #include <sys/signalvar.h>
49 #include <sys/stat.h>
50 #include <sys/syscall.h>
51 #include <sys/sysctl.h>
52 #include <sys/sysent.h>
53 #include <sys/vnode.h>
54 #include <sys/eventhandler.h>
55
56 #include <cpu/lwbuf.h>
57
58 #include <vm/vm.h>
59 #include <vm/vm_kern.h>
60 #include <vm/vm_param.h>
61 #include <vm/pmap.h>
62 #include <sys/lock.h>
63 #include <vm/vm_map.h>
64 #include <vm/vm_object.h>
65 #include <vm/vm_extern.h>
66
67 #include <machine/elf.h>
68 #include <machine/md_var.h>
69 #include <sys/mount.h>
70 #include <sys/ckpt.h>
71
72 #define OLD_EI_BRAND 8
73 #define truncps(va,ps) ((va) & ~(ps - 1))
74 #define aligned(a,t) (truncps((u_long)(a), sizeof(t)) == (u_long)(a))
75
76 static int __elfN(check_header)(const Elf_Ehdr *hdr);
77 static Elf_Brandinfo *__elfN(get_brandinfo)(struct image_params *imgp,
78 const char *interp, int32_t *osrel);
79 static int __elfN(load_file)(struct proc *p, const char *file, u_long *addr,
80 u_long *entry);
81 static int __elfN(load_section)(struct proc *p,
82 struct vmspace *vmspace, struct vnode *vp,
83 vm_offset_t offset, caddr_t vmaddr, size_t memsz, size_t filsz,
84 vm_prot_t prot);
85 static int __CONCAT(exec_, __elfN(imgact))(struct image_params *imgp);
86 static boolean_t __elfN(bsd_trans_osrel)(const Elf_Note *note,
87 int32_t *osrel);
88 static boolean_t __elfN(check_note)(struct image_params *imgp,
89 Elf_Brandnote *checknote, int32_t *osrel);
90 static vm_prot_t __elfN(trans_prot)(Elf_Word);
91 static Elf_Word __elfN(untrans_prot)(vm_prot_t);
92 static boolean_t check_PT_NOTE(struct image_params *imgp,
93 Elf_Brandnote *checknote, int32_t *osrel, const Elf_Phdr * pnote);
94 static boolean_t extract_interpreter(struct image_params *imgp,
95 const Elf_Phdr *pinterpreter, char *data);
96 static u_long pie_base_hint(struct proc *p);
97
98 static int elf_legacy_coredump = 0;
99 static int __elfN(fallback_brand) = -1;
100 static int elf_pie_base_mmap = 0;
101 #if defined(__x86_64__)
102 SYSCTL_NODE(_kern, OID_AUTO, elf64, CTLFLAG_RW, 0, "");
103 SYSCTL_INT(_debug, OID_AUTO, elf64_legacy_coredump, CTLFLAG_RW,
104 &elf_legacy_coredump, 0, "legacy coredump mode");
105 SYSCTL_INT(_kern_elf64, OID_AUTO, fallback_brand, CTLFLAG_RW,
106 &elf64_fallback_brand, 0, "ELF64 brand of last resort");
107 TUNABLE_INT("kern.elf64.fallback_brand", &elf64_fallback_brand);
108 SYSCTL_INT(_kern_elf64, OID_AUTO, pie_base_mmap, CTLFLAG_RW,
109 &elf_pie_base_mmap, 0,
110 "choose a base address for PIE as if it is mapped with mmap()");
111 TUNABLE_INT("kern.elf64.pie_base_mmap", &elf_pie_base_mmap);
112 #else /* i386 assumed */
113 SYSCTL_NODE(_kern, OID_AUTO, elf32, CTLFLAG_RW, 0, "");
114 SYSCTL_INT(_debug, OID_AUTO, elf32_legacy_coredump, CTLFLAG_RW,
115 &elf_legacy_coredump, 0, "legacy coredump mode");
116 SYSCTL_INT(_kern_elf32, OID_AUTO, fallback_brand, CTLFLAG_RW,
117 &elf32_fallback_brand, 0, "ELF32 brand of last resort");
118 TUNABLE_INT("kern.elf32.fallback_brand", &elf32_fallback_brand);
119 SYSCTL_INT(_kern_elf32, OID_AUTO, pie_base_mmap, CTLFLAG_RW,
120 &elf_pie_base_mmap, 0,
121 "choose a base address for PIE as if it is mapped with mmap()");
122 TUNABLE_INT("kern.elf32.pie_base_mmap", &elf_pie_base_mmap);
123 #endif
124
125 static Elf_Brandinfo *elf_brand_list[MAX_BRANDS];
126
127 static const char DRAGONFLY_ABI_VENDOR[] = "DragonFly";
128 static const char FREEBSD_ABI_VENDOR[] = "FreeBSD";
129
130 Elf_Brandnote __elfN(dragonfly_brandnote) = {
131 .hdr.n_namesz = sizeof(DRAGONFLY_ABI_VENDOR),
132 .hdr.n_descsz = sizeof(int32_t),
133 .hdr.n_type = 1,
134 .vendor = DRAGONFLY_ABI_VENDOR,
135 .flags = BN_TRANSLATE_OSREL,
136 .trans_osrel = __elfN(bsd_trans_osrel),
137 };
138
139 Elf_Brandnote __elfN(freebsd_brandnote) = {
140 .hdr.n_namesz = sizeof(FREEBSD_ABI_VENDOR),
141 .hdr.n_descsz = sizeof(int32_t),
142 .hdr.n_type = 1,
143 .vendor = FREEBSD_ABI_VENDOR,
144 .flags = BN_TRANSLATE_OSREL,
145 .trans_osrel = __elfN(bsd_trans_osrel),
146 };
147
148 int
149 __elfN(insert_brand_entry)(Elf_Brandinfo *entry)
150 {
151 int i;
152
153 for (i = 0; i < MAX_BRANDS; i++) {
154 if (elf_brand_list[i] == NULL) {
155 elf_brand_list[i] = entry;
156 break;
157 }
158 }
159 if (i == MAX_BRANDS) {
160 uprintf("WARNING: %s: could not insert brandinfo entry: %p\n",
161 __func__, entry);
162 return (-1);
163 }
164 return (0);
165 }
166
167 int
168 __elfN(remove_brand_entry)(Elf_Brandinfo *entry)
169 {
170 int i;
171
172 for (i = 0; i < MAX_BRANDS; i++) {
173 if (elf_brand_list[i] == entry) {
174 elf_brand_list[i] = NULL;
175 break;
176 }
177 }
178 if (i == MAX_BRANDS)
179 return (-1);
180 return (0);
181 }
182
183 /*
184 * Check if an elf brand is being used anywhere in the system.
185 *
186 * Used by the linux emulation module unloader. This isn't safe from
187 * races.
188 */
189 struct elf_brand_inuse_info {
190 int rval;
191 Elf_Brandinfo *entry;
192 };
193
194 static int elf_brand_inuse_callback(struct proc *p, void *data);
195
196 int
197 __elfN(brand_inuse)(Elf_Brandinfo *entry)
198 {
199 struct elf_brand_inuse_info info;
200
201 info.rval = FALSE;
202 info.entry = entry;
203 allproc_scan(elf_brand_inuse_callback, &info);
204 return (info.rval);
205 }
206
207 static
208 int
209 elf_brand_inuse_callback(struct proc *p, void *data)
210 {
211 struct elf_brand_inuse_info *info = data;
212
213 if (p->p_sysent == info->entry->sysvec) {
214 info->rval = TRUE;
215 return (-1);
216 }
217 return (0);
218 }
219
220 static int
221 __elfN(check_header)(const Elf_Ehdr *hdr)
222 {
223 Elf_Brandinfo *bi;
224 int i;
225
226 if (!IS_ELF(*hdr) ||
227 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
228 hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
229 hdr->e_ident[EI_VERSION] != EV_CURRENT ||
230 hdr->e_phentsize != sizeof(Elf_Phdr) ||
231 hdr->e_ehsize != sizeof(Elf_Ehdr) ||
232 hdr->e_version != ELF_TARG_VER)
233 return (ENOEXEC);
234
235 /*
236 * Make sure we have at least one brand for this machine.
237 */
238
239 for (i = 0; i < MAX_BRANDS; i++) {
240 bi = elf_brand_list[i];
241 if (bi != NULL && bi->machine == hdr->e_machine)
242 break;
243 }
244 if (i == MAX_BRANDS)
245 return (ENOEXEC);
246
247 return (0);
248 }
249
250 static int
251 __elfN(load_section)(struct proc *p, struct vmspace *vmspace, struct vnode *vp,
252 vm_offset_t offset, caddr_t vmaddr, size_t memsz,
253 size_t filsz, vm_prot_t prot)
254 {
255 size_t map_len;
256 vm_offset_t map_addr;
257 int error, rv, cow;
258 int count;
259 int shared;
260 size_t copy_len;
261 vm_object_t object;
262 vm_offset_t file_addr;
263
264 object = vp->v_object;
265 error = 0;
266
267 /*
268 * In most cases we will be able to use a shared lock on the
269 * object we are inserting into the map. The lock will be
270 * upgraded in situations where new VM pages must be allocated.
271 */
272 vm_object_hold_shared(object);
273 shared = 1;
274
275 /*
276 * It's necessary to fail if the filsz + offset taken from the
277 * header is greater than the actual file pager object's size.
278 * If we were to allow this, then the vm_map_find() below would
279 * walk right off the end of the file object and into the ether.
280 *
281 * While I'm here, might as well check for something else that
282 * is invalid: filsz cannot be greater than memsz.
283 */
284 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) {
285 uprintf("elf_load_section: truncated ELF file\n");
286 vm_object_drop(object);
287 return (ENOEXEC);
288 }
289
290 map_addr = trunc_page((vm_offset_t)vmaddr);
291 file_addr = trunc_page(offset);
292
293 /*
294 * We have two choices. We can either clear the data in the last page
295 * of an oversized mapping, or we can start the anon mapping a page
296 * early and copy the initialized data into that first page. We
297 * choose the second..
298 */
299 if (memsz > filsz)
300 map_len = trunc_page(offset+filsz) - file_addr;
301 else
302 map_len = round_page(offset+filsz) - file_addr;
303
304 if (map_len != 0) {
305 vm_object_reference_locked(object);
306
307 /* cow flags: don't dump readonly sections in core */
308 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT;
309 if ((prot & VM_PROT_WRITE) == 0)
310 cow |= MAP_DISABLE_COREDUMP;
311 if (shared == 0)
312 cow |= MAP_PREFAULT_RELOCK;
313
314 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
315 vm_map_lock(&vmspace->vm_map);
316 rv = vm_map_insert(&vmspace->vm_map, &count,
317 object, NULL,
318 file_addr, /* file offset */
319 map_addr, /* virtual start */
320 map_addr + map_len,/* virtual end */
321 VM_MAPTYPE_NORMAL,
322 prot, VM_PROT_ALL,
323 cow);
324 vm_map_unlock(&vmspace->vm_map);
325 vm_map_entry_release(count);
326
327 /*
328 * NOTE: Object must have a hold ref when calling
329 * vm_object_deallocate().
330 */
331 if (rv != KERN_SUCCESS) {
332 vm_object_drop(object);
333 vm_object_deallocate(object);
334 return (EINVAL);
335 }
336
337 /* we can stop now if we've covered it all */
338 if (memsz == filsz) {
339 vm_object_drop(object);
340 return (0);
341 }
342 }
343
344 /*
345 * We have to get the remaining bit of the file into the first part
346 * of the oversized map segment. This is normally because the .data
347 * segment in the file is extended to provide bss. It's a neat idea
348 * to try and save a page, but it's a pain in the behind to implement.
349 */
350 copy_len = (offset + filsz) - trunc_page(offset + filsz);
351 map_addr = trunc_page((vm_offset_t)vmaddr + filsz);
352 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr;
353
354 /* This had damn well better be true! */
355 if (map_len != 0) {
356 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
357 vm_map_lock(&vmspace->vm_map);
358 rv = vm_map_insert(&vmspace->vm_map, &count,
359 NULL, NULL,
360 0,
361 map_addr,
362 map_addr + map_len,
363 VM_MAPTYPE_NORMAL,
364 VM_PROT_ALL, VM_PROT_ALL,
365 0);
366 vm_map_unlock(&vmspace->vm_map);
367 vm_map_entry_release(count);
368 if (rv != KERN_SUCCESS) {
369 vm_object_drop(object);
370 return (EINVAL);
371 }
372 }
373
374 if (copy_len != 0) {
375 struct lwbuf *lwb;
376 struct lwbuf lwb_cache;
377 vm_page_t m;
378
379 m = vm_fault_object_page(object, trunc_page(offset + filsz),
380 VM_PROT_READ, 0, &shared, &error);
381 vm_object_drop(object);
382 if (m) {
383 lwb = lwbuf_alloc(m, &lwb_cache);
384 error = copyout((caddr_t)lwbuf_kva(lwb),
385 (caddr_t)map_addr, copy_len);
386 lwbuf_free(lwb);
387 vm_page_unhold(m);
388 }
389 } else {
390 vm_object_drop(object);
391 }
392
393 /*
394 * set it to the specified protection
395 */
396 if (error == 0) {
397 vm_map_protect(&vmspace->vm_map,
398 map_addr, map_addr + map_len,
399 prot, FALSE);
400 }
401 return (error);
402 }
403
404 /*
405 * Load the file "file" into memory. It may be either a shared object
406 * or an executable.
407 *
408 * The "addr" reference parameter is in/out. On entry, it specifies
409 * the address where a shared object should be loaded. If the file is
410 * an executable, this value is ignored. On exit, "addr" specifies
411 * where the file was actually loaded.
412 *
413 * The "entry" reference parameter is out only. On exit, it specifies
414 * the entry point for the loaded file.
415 */
416 static int
417 __elfN(load_file)(struct proc *p, const char *file, u_long *addr, u_long *entry)
418 {
419 struct {
420 struct nlookupdata nd;
421 struct vattr attr;
422 struct image_params image_params;
423 } *tempdata;
424 const Elf_Ehdr *hdr = NULL;
425 const Elf_Phdr *phdr = NULL;
426 struct nlookupdata *nd;
427 struct vmspace *vmspace = p->p_vmspace;
428 struct vattr *attr;
429 struct image_params *imgp;
430 struct mount *topmnt;
431 vm_prot_t prot;
432 u_long rbase;
433 u_long base_addr = 0;
434 int error, i, numsegs;
435
436 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_WAITOK);
437 nd = &tempdata->nd;
438 attr = &tempdata->attr;
439 imgp = &tempdata->image_params;
440
441 /*
442 * Initialize part of the common data
443 */
444 imgp->proc = p;
445 imgp->attr = attr;
446 imgp->firstpage = NULL;
447 imgp->image_header = NULL;
448 imgp->vp = NULL;
449
450 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW);
451 if (error == 0)
452 error = nlookup(nd);
453 if (error == 0)
454 error = cache_vget(&nd->nl_nch, nd->nl_cred,
455 LK_SHARED, &imgp->vp);
456 topmnt = nd->nl_nch.mount;
457 nlookup_done(nd);
458 if (error)
459 goto fail;
460
461 /*
462 * Check permissions, modes, uid, etc on the file, and "open" it.
463 */
464 error = exec_check_permissions(imgp, topmnt);
465 if (error) {
466 vn_unlock(imgp->vp);
467 goto fail;
468 }
469
470 error = exec_map_first_page(imgp);
471 /*
472 * Also make certain that the interpreter stays the same, so set
473 * its VTEXT flag, too.
474 */
475 if (error == 0)
476 vsetflags(imgp->vp, VTEXT);
477 vn_unlock(imgp->vp);
478 if (error)
479 goto fail;
480
481 hdr = (const Elf_Ehdr *)imgp->image_header;
482 if ((error = __elfN(check_header)(hdr)) != 0)
483 goto fail;
484 if (hdr->e_type == ET_DYN)
485 rbase = *addr;
486 else if (hdr->e_type == ET_EXEC)
487 rbase = 0;
488 else {
489 error = ENOEXEC;
490 goto fail;
491 }
492
493 /* Only support headers that fit within first page for now */
494 /* (multiplication of two Elf_Half fields will not overflow) */
495 if ((hdr->e_phoff > PAGE_SIZE) ||
496 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) {
497 error = ENOEXEC;
498 goto fail;
499 }
500
501 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
502 if (!aligned(phdr, Elf_Addr)) {
503 error = ENOEXEC;
504 goto fail;
505 }
506
507 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) {
508 if (phdr[i].p_type == PT_LOAD && phdr[i].p_memsz != 0) {
509 /* Loadable segment */
510 prot = __elfN(trans_prot)(phdr[i].p_flags);
511 error = __elfN(load_section)(
512 p, vmspace, imgp->vp,
513 phdr[i].p_offset,
514 (caddr_t)phdr[i].p_vaddr +
515 rbase,
516 phdr[i].p_memsz,
517 phdr[i].p_filesz, prot);
518 if (error != 0)
519 goto fail;
520 /*
521 * Establish the base address if this is the
522 * first segment.
523 */
524 if (numsegs == 0)
525 base_addr = trunc_page(phdr[i].p_vaddr + rbase);
526 numsegs++;
527 }
528 }
529 *addr = base_addr;
530 *entry = (unsigned long)hdr->e_entry + rbase;
531
532 fail:
533 if (imgp->firstpage)
534 exec_unmap_first_page(imgp);
535 if (imgp->vp) {
536 vrele(imgp->vp);
537 imgp->vp = NULL;
538 }
539 kfree(tempdata, M_TEMP);
540
541 return (error);
542 }
543
544 static Elf_Brandinfo *
545 __elfN(get_brandinfo)(struct image_params *imgp, const char *interp,
546 int32_t *osrel)
547 {
548 const Elf_Ehdr *hdr = (const Elf_Ehdr *)imgp->image_header;
549 Elf_Brandinfo *bi;
550 boolean_t ret;
551 int i;
552
553 /* We support four types of branding -- (1) the ELF EI_OSABI field
554 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string
555 * branding within the ELF header, (3) path of the `interp_path' field,
556 * and (4) the ".note.ABI-tag" ELF section.
557 */
558
559 /* Look for an ".note.ABI-tag" ELF section */
560 for (i = 0; i < MAX_BRANDS; i++) {
561 bi = elf_brand_list[i];
562
563 if (bi == NULL)
564 continue;
565 if (hdr->e_machine == bi->machine && (bi->flags &
566 (BI_BRAND_NOTE|BI_BRAND_NOTE_MANDATORY)) != 0) {
567 ret = __elfN(check_note)(imgp, bi->brand_note, osrel);
568 if (ret)
569 return (bi);
570 }
571 }
572
573 /* If the executable has a brand, search for it in the brand list. */
574 for (i = 0; i < MAX_BRANDS; i++) {
575 bi = elf_brand_list[i];
576
577 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY)
578 continue;
579 if (hdr->e_machine == bi->machine &&
580 (hdr->e_ident[EI_OSABI] == bi->brand ||
581 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND],
582 bi->compat_3_brand, strlen(bi->compat_3_brand)) == 0))
583 return (bi);
584 }
585
586 /* Lacking a known brand, search for a recognized interpreter. */
587 if (interp != NULL) {
588 for (i = 0; i < MAX_BRANDS; i++) {
589 bi = elf_brand_list[i];
590
591 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY)
592 continue;
593 if (hdr->e_machine == bi->machine &&
594 strcmp(interp, bi->interp_path) == 0)
595 return (bi);
596 }
597 }
598
599 /* Lacking a recognized interpreter, try the default brand */
600 for (i = 0; i < MAX_BRANDS; i++) {
601 bi = elf_brand_list[i];
602
603 if (bi == NULL || bi->flags & BI_BRAND_NOTE_MANDATORY)
604 continue;
605 if (hdr->e_machine == bi->machine &&
606 __elfN(fallback_brand) == bi->brand)
607 return (bi);
608 }
609 return (NULL);
610 }
611
612 static int
613 __CONCAT(exec_,__elfN(imgact))(struct image_params *imgp)
614 {
615 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header;
616 const Elf_Phdr *phdr;
617 Elf_Auxargs *elf_auxargs;
618 struct vmspace *vmspace;
619 vm_prot_t prot;
620 u_long text_size = 0, data_size = 0, total_size = 0;
621 u_long text_addr = 0, data_addr = 0;
622 u_long seg_size, seg_addr;
623 u_long addr, baddr, et_dyn_addr = 0, entry = 0, proghdr = 0;
624 int32_t osrel = 0;
625 int error = 0, i, n;
626 boolean_t failure;
627 char *interp = NULL;
628 const char *newinterp = NULL;
629 Elf_Brandinfo *brand_info;
630 char *path;
631
632 /*
633 * Do we have a valid ELF header ?
634 *
635 * Only allow ET_EXEC & ET_DYN here, reject ET_DYN later if a particular
636 * brand doesn't support it. Both DragonFly platforms do by default.
637 */
638 if (__elfN(check_header)(hdr) != 0 ||
639 (hdr->e_type != ET_EXEC && hdr->e_type != ET_DYN))
640 return (-1);
641
642 /*
643 * From here on down, we return an errno, not -1, as we've
644 * detected an ELF file.
645 */
646
647 if ((hdr->e_phoff > PAGE_SIZE) ||
648 (hdr->e_phoff + hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE) {
649 /* Only support headers in first page for now */
650 return (ENOEXEC);
651 }
652 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
653 if (!aligned(phdr, Elf_Addr))
654 return (ENOEXEC);
655 n = 0;
656 baddr = 0;
657 for (i = 0; i < hdr->e_phnum; i++) {
658 if (phdr[i].p_type == PT_LOAD) {
659 if (n == 0)
660 baddr = phdr[i].p_vaddr;
661 n++;
662 continue;
663 }
664 if (phdr[i].p_type == PT_INTERP) {
665 /*
666 * If interp is already defined there are more than
667 * one PT_INTERP program headers present. Take only
668 * the first one and ignore the rest.
669 */
670 if (interp != NULL)
671 continue;
672
673 if (phdr[i].p_filesz == 0 ||
674 phdr[i].p_filesz > PAGE_SIZE ||
675 phdr[i].p_filesz > MAXPATHLEN)
676 return (ENOEXEC);
677
678 interp = kmalloc(phdr[i].p_filesz, M_TEMP, M_WAITOK);
679 failure = extract_interpreter(imgp, &phdr[i], interp);
680 if (failure) {
681 kfree(interp, M_TEMP);
682 return (ENOEXEC);
683 }
684 continue;
685 }
686 }
687
688 brand_info = __elfN(get_brandinfo)(imgp, interp, &osrel);
689 if (brand_info == NULL) {
690 uprintf("ELF binary type \"%u\" not known.\n",
691 hdr->e_ident[EI_OSABI]);
692 if (interp != NULL)
693 kfree(interp, M_TEMP);
694 return (ENOEXEC);
695 }
696 if (hdr->e_type == ET_DYN) {
697 if ((brand_info->flags & BI_CAN_EXEC_DYN) == 0) {
698 if (interp != NULL)
699 kfree(interp, M_TEMP);
700 return (ENOEXEC);
701 }
702 /*
703 * If p_vaddr field of PT_LOAD program header is zero and type of an
704 * executale is ET_DYN, then it must be a position independent
705 * executable (PIE). In this case the system needs to pick a base
706 * address for us. Set et_dyn_addr to non-zero and choose the actual
707 * address when we are ready.
708 */
709 if (baddr == 0)
710 et_dyn_addr = 1;
711 }
712
713 if (interp != NULL && brand_info->interp_newpath != NULL)
714 newinterp = brand_info->interp_newpath;
715
716 exec_new_vmspace(imgp, NULL);
717
718 /*
719 * Yeah, I'm paranoid. There is every reason in the world to get
720 * VTEXT now since from here on out, there are places we can have
721 * a context switch. Better safe than sorry; I really don't want
722 * the file to change while it's being loaded.
723 */
724 vsetflags(imgp->vp, VTEXT);
725
726 vmspace = imgp->proc->p_vmspace;
727 /* Choose the base address for dynamic executables if we need to. */
728 if (et_dyn_addr)
729 et_dyn_addr = pie_base_hint(imgp->proc);
730
731 for (i = 0; i < hdr->e_phnum; i++) {
732 switch (phdr[i].p_type) {
733 case PT_LOAD: /* Loadable segment */
734 if (phdr[i].p_memsz == 0)
735 break;
736 prot = __elfN(trans_prot)(phdr[i].p_flags);
737
738 if ((error = __elfN(load_section)(
739 imgp->proc,
740 vmspace,
741 imgp->vp,
742 phdr[i].p_offset,
743 (caddr_t)phdr[i].p_vaddr + et_dyn_addr,
744 phdr[i].p_memsz,
745 phdr[i].p_filesz,
746 prot)) != 0) {
747 if (interp != NULL)
748 kfree (interp, M_TEMP);
749 return (error);
750 }
751
752 /*
753 * If this segment contains the program headers,
754 * remember their virtual address for the AT_PHDR
755 * aux entry. Static binaries don't usually include
756 * a PT_PHDR entry.
757 */
758 if (phdr[i].p_offset == 0 &&
759 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize
760 <= phdr[i].p_filesz)
761 proghdr = phdr[i].p_vaddr + hdr->e_phoff +
762 et_dyn_addr;
763
764 seg_addr = trunc_page(phdr[i].p_vaddr + et_dyn_addr);
765 seg_size = round_page(phdr[i].p_memsz +
766 phdr[i].p_vaddr + et_dyn_addr - seg_addr);
767
768 /*
769 * Is this .text or .data? We can't use
770 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the
771 * alpha terribly and possibly does other bad
772 * things so we stick to the old way of figuring
773 * it out: If the segment contains the program
774 * entry point, it's a text segment, otherwise it
775 * is a data segment.
776 *
777 * Note that obreak() assumes that data_addr +
778 * data_size == end of data load area, and the ELF
779 * file format expects segments to be sorted by
780 * address. If multiple data segments exist, the
781 * last one will be used.
782 */
783 if (hdr->e_entry >= phdr[i].p_vaddr &&
784 hdr->e_entry < (phdr[i].p_vaddr +
785 phdr[i].p_memsz)) {
786 text_size = seg_size;
787 text_addr = seg_addr;
788 entry = (u_long)hdr->e_entry + et_dyn_addr;
789 } else {
790 data_size = seg_size;
791 data_addr = seg_addr;
792 }
793 total_size += seg_size;
794
795 /*
796 * Check limits. It should be safe to check the
797 * limits after loading the segment since we do
798 * not actually fault in all the segment's pages.
799 */
800 if (data_size >
801 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur ||
802 text_size > maxtsiz ||
803 total_size >
804 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
805 if (interp != NULL)
806 kfree(interp, M_TEMP);
807 error = ENOMEM;
808 return (error);
809 }
810 break;
811 case PT_PHDR: /* Program header table info */
812 proghdr = phdr[i].p_vaddr + et_dyn_addr;
813 break;
814 default:
815 break;
816 }
817 }
818
819 vmspace->vm_tsize = text_size >> PAGE_SHIFT;
820 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr;
821 vmspace->vm_dsize = data_size >> PAGE_SHIFT;
822 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr;
823
824 addr = ELF_RTLD_ADDR(vmspace);
825
826 imgp->entry_addr = entry;
827
828 imgp->proc->p_sysent = brand_info->sysvec;
829 EVENTHANDLER_INVOKE(process_exec, imgp);
830
831 if (interp != NULL) {
832 int have_interp = FALSE;
833 if (brand_info->emul_path != NULL &&
834 brand_info->emul_path[0] != '\0') {
835 path = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
836 ksnprintf(path, MAXPATHLEN, "%s%s",
837 brand_info->emul_path, interp);
838 error = __elfN(load_file)(imgp->proc, path, &addr,
839 &imgp->entry_addr);
840 kfree(path, M_TEMP);
841 if (error == 0)
842 have_interp = TRUE;
843 }
844 if (!have_interp && newinterp != NULL) {
845 error = __elfN(load_file)(imgp->proc, newinterp,
846 &addr, &imgp->entry_addr);
847 if (error == 0)
848 have_interp = TRUE;
849 }
850 if (!have_interp) {
851 error = __elfN(load_file)(imgp->proc, interp, &addr,
852 &imgp->entry_addr);
853 }
854 if (error != 0) {
855 uprintf("ELF interpreter %s not found\n", interp);
856 kfree(interp, M_TEMP);
857 return (error);
858 }
859 kfree(interp, M_TEMP);
860 } else
861 addr = et_dyn_addr;
862
863 /*
864 * Construct auxargs table (used by the fixup routine)
865 */
866 elf_auxargs = kmalloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK);
867 elf_auxargs->execfd = -1;
868 elf_auxargs->phdr = proghdr;
869 elf_auxargs->phent = hdr->e_phentsize;
870 elf_auxargs->phnum = hdr->e_phnum;
871 elf_auxargs->pagesz = PAGE_SIZE;
872 elf_auxargs->base = addr;
873 elf_auxargs->flags = 0;
874 elf_auxargs->entry = entry;
875
876 imgp->auxargs = elf_auxargs;
877 imgp->interpreted = 0;
878 imgp->proc->p_osrel = osrel;
879
880 return (error);
881 }
882
883 int
884 __elfN(dragonfly_fixup)(register_t **stack_base, struct image_params *imgp)
885 {
886 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs;
887 Elf_Addr *base;
888 Elf_Addr *pos;
889
890 base = (Elf_Addr *)*stack_base;
891 pos = base + (imgp->args->argc + imgp->args->envc + 2);
892
893 if (args->execfd != -1)
894 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
895 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
896 AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
897 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
898 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
899 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
900 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
901 AUXARGS_ENTRY(pos, AT_BASE, args->base);
902 if (imgp->execpathp != 0)
903 AUXARGS_ENTRY(pos, AT_EXECPATH, imgp->execpathp);
904 AUXARGS_ENTRY(pos, AT_OSRELDATE, osreldate);
905 AUXARGS_ENTRY(pos, AT_NULL, 0);
906
907 kfree(imgp->auxargs, M_TEMP);
908 imgp->auxargs = NULL;
909
910 base--;
911 suword(base, (long)imgp->args->argc);
912 *stack_base = (register_t *)base;
913 return (0);
914 }
915
916 /*
917 * Code for generating ELF core dumps.
918 */
919
920 typedef int (*segment_callback)(vm_map_entry_t, void *);
921
922 /* Closure for cb_put_phdr(). */
923 struct phdr_closure {
924 Elf_Phdr *phdr; /* Program header to fill in (incremented) */
925 Elf_Phdr *phdr_max; /* Pointer bound for error check */
926 Elf_Off offset; /* Offset of segment in core file */
927 };
928
929 /* Closure for cb_size_segment(). */
930 struct sseg_closure {
931 int count; /* Count of writable segments. */
932 size_t vsize; /* Total size of all writable segments. */
933 };
934
935 /* Closure for cb_put_fp(). */
936 struct fp_closure {
937 struct vn_hdr *vnh;
938 struct vn_hdr *vnh_max;
939 int count;
940 struct stat *sb;
941 };
942
943 typedef struct elf_buf {
944 char *buf;
945 size_t off;
946 size_t off_max;
947 } *elf_buf_t;
948
949 static void *target_reserve(elf_buf_t target, size_t bytes, int *error);
950
951 static int cb_put_phdr (vm_map_entry_t, void *);
952 static int cb_size_segment (vm_map_entry_t, void *);
953 static int cb_fpcount_segment(vm_map_entry_t, void *);
954 static int cb_put_fp(vm_map_entry_t, void *);
955
956
957 static int each_segment (struct proc *, segment_callback, void *, int);
958 static int __elfN(corehdr)(struct lwp *, int, struct file *, struct ucred *,
959 int, elf_buf_t);
960 enum putmode { WRITE, DRYRUN };
961 static int __elfN(puthdr)(struct lwp *, elf_buf_t, int sig, enum putmode,
962 int, struct file *);
963 static int elf_putallnotes(struct lwp *, elf_buf_t, int, enum putmode);
964 static int __elfN(putnote)(elf_buf_t, const char *, int, const void *, size_t);
965
966 static int elf_putsigs(struct lwp *, elf_buf_t);
967 static int elf_puttextvp(struct proc *, elf_buf_t);
968 static int elf_putfiles(struct proc *, elf_buf_t, struct file *);
969
970 int
971 __elfN(coredump)(struct lwp *lp, int sig, struct vnode *vp, off_t limit)
972 {
973 struct file *fp;
974 int error;
975
976 if ((error = falloc(NULL, &fp, NULL)) != 0)
977 return (error);
978 fsetcred(fp, lp->lwp_proc->p_ucred);
979
980 /*
981 * XXX fixme.
982 */
983 fp->f_type = DTYPE_VNODE;
984 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW;
985 fp->f_ops = &vnode_fileops;
986 fp->f_data = vp;
987
988 error = generic_elf_coredump(lp, sig, fp, limit);
989
990 fp->f_type = 0;
991 fp->f_flag = 0;
992 fp->f_ops = &badfileops;
993 fp->f_data = NULL;
994 fdrop(fp);
995 return (error);
996 }
997
998 int
999 generic_elf_coredump(struct lwp *lp, int sig, struct file *fp, off_t limit)
1000 {
1001 struct proc *p = lp->lwp_proc;
1002 struct ucred *cred = p->p_ucred;
1003 int error = 0;
1004 struct sseg_closure seginfo;
1005 struct elf_buf target;
1006
1007 if (!fp)
1008 kprintf("can't dump core - null fp\n");
1009
1010 /*
1011 * Size the program segments
1012 */
1013 seginfo.count = 0;
1014 seginfo.vsize = 0;
1015 each_segment(p, cb_size_segment, &seginfo, 1);
1016
1017 /*
1018 * Calculate the size of the core file header area by making
1019 * a dry run of generating it. Nothing is written, but the
1020 * size is calculated.
1021 */
1022 bzero(&target, sizeof(target));
1023 __elfN(puthdr)(lp, &target, sig, DRYRUN, seginfo.count, fp);
1024
1025 if (target.off + seginfo.vsize >= limit)
1026 return (EFAULT);
1027
1028 /*
1029 * Allocate memory for building the header, fill it up,
1030 * and write it out.
1031 */
1032 target.off_max = target.off;
1033 target.off = 0;
1034 target.buf = kmalloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO);
1035
1036 error = __elfN(corehdr)(lp, sig, fp, cred, seginfo.count, &target);
1037
1038 /* Write the contents of all of the writable segments. */
1039 if (error == 0) {
1040 Elf_Phdr *php;
1041 int i;
1042 ssize_t nbytes;
1043
1044 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1;
1045 for (i = 0; i < seginfo.count; i++) {
1046 error = fp_write(fp, (caddr_t)php->p_vaddr,
1047 php->p_filesz, &nbytes, UIO_USERSPACE);
1048 if (error != 0)
1049 break;
1050 php++;
1051 }
1052 }
1053 kfree(target.buf, M_TEMP);
1054
1055 return (error);
1056 }
1057
1058 /*
1059 * A callback for each_segment() to write out the segment's
1060 * program header entry.
1061 */
1062 static int
1063 cb_put_phdr(vm_map_entry_t entry, void *closure)
1064 {
1065 struct phdr_closure *phc = closure;
1066 Elf_Phdr *phdr = phc->phdr;
1067
1068 if (phc->phdr == phc->phdr_max)
1069 return (EINVAL);
1070
1071 phc->offset = round_page(phc->offset);
1072
1073 phdr->p_type = PT_LOAD;
1074 phdr->p_offset = phc->offset;
1075 phdr->p_vaddr = entry->start;
1076 phdr->p_paddr = 0;
1077 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1078 phdr->p_align = PAGE_SIZE;
1079 phdr->p_flags = __elfN(untrans_prot)(entry->protection);
1080
1081 phc->offset += phdr->p_filesz;
1082 ++phc->phdr;
1083 return (0);
1084 }
1085
1086 /*
1087 * A callback for each_writable_segment() to gather information about
1088 * the number of segments and their total size.
1089 */
1090 static int
1091 cb_size_segment(vm_map_entry_t entry, void *closure)
1092 {
1093 struct sseg_closure *ssc = closure;
1094
1095 ++ssc->count;
1096 ssc->vsize += entry->end - entry->start;
1097 return (0);
1098 }
1099
1100 /*
1101 * A callback for each_segment() to gather information about
1102 * the number of text segments.
1103 */
1104 static int
1105 cb_fpcount_segment(vm_map_entry_t entry, void *closure)
1106 {
1107 int *count = closure;
1108 struct vnode *vp;
1109
1110 if (entry->object.vm_object->type == OBJT_VNODE) {
1111 vp = (struct vnode *)entry->object.vm_object->handle;
1112 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1113 return (0);
1114 ++*count;
1115 }
1116 return (0);
1117 }
1118
1119 static int
1120 cb_put_fp(vm_map_entry_t entry, void *closure)
1121 {
1122 struct fp_closure *fpc = closure;
1123 struct vn_hdr *vnh = fpc->vnh;
1124 Elf_Phdr *phdr = &vnh->vnh_phdr;
1125 struct vnode *vp;
1126 int error;
1127
1128 /*
1129 * If an entry represents a vnode then write out a file handle.
1130 *
1131 * If we are checkpointing a checkpoint-restored program we do
1132 * NOT record the filehandle for the old checkpoint vnode (which
1133 * is mapped all over the place). Instead we rely on the fact
1134 * that a checkpoint-restored program does not mmap() the checkpt
1135 * vnode NOCORE, so its contents will be written out to the
1136 * new checkpoint file. This is necessary because the 'old'
1137 * checkpoint file is typically destroyed when a new one is created
1138 * and thus cannot be used to restore the new checkpoint.
1139 *
1140 * Theoretically we could create a chain of checkpoint files and
1141 * operate the checkpointing operation kinda like an incremental
1142 * checkpoint, but a checkpoint restore would then likely wind up
1143 * referencing many prior checkpoint files and that is a bit over
1144 * the top for the purpose of the checkpoint API.
1145 */
1146 if (entry->object.vm_object->type == OBJT_VNODE) {
1147 vp = (struct vnode *)entry->object.vm_object->handle;
1148 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1149 return (0);
1150 if (vnh == fpc->vnh_max)
1151 return (EINVAL);
1152
1153 if (vp->v_mount)
1154 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1155 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid);
1156 if (error) {
1157 char *freepath, *fullpath;
1158
1159 if (vn_fullpath(curproc, vp, &fullpath, &freepath, 0)) {
1160 kprintf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp);
1161 } else {
1162 kprintf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath);
1163 kfree(freepath, M_TEMP);
1164 }
1165 error = 0;
1166 }
1167
1168 phdr->p_type = PT_LOAD;
1169 phdr->p_offset = 0; /* not written to core */
1170 phdr->p_vaddr = entry->start;
1171 phdr->p_paddr = 0;
1172 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1173 phdr->p_align = PAGE_SIZE;
1174 phdr->p_flags = 0;
1175 if (entry->protection & VM_PROT_READ)
1176 phdr->p_flags |= PF_R;
1177 if (entry->protection & VM_PROT_WRITE)
1178 phdr->p_flags |= PF_W;
1179 if (entry->protection & VM_PROT_EXECUTE)
1180 phdr->p_flags |= PF_X;
1181 ++fpc->vnh;
1182 ++fpc->count;
1183 }
1184 return (0);
1185 }
1186
1187 /*
1188 * For each writable segment in the process's memory map, call the given
1189 * function with a pointer to the map entry and some arbitrary
1190 * caller-supplied data.
1191 */
1192 static int
1193 each_segment(struct proc *p, segment_callback func, void *closure, int writable)
1194 {
1195 int error = 0;
1196 vm_map_t map = &p->p_vmspace->vm_map;
1197 vm_map_entry_t entry;
1198
1199 for (entry = map->header.next; error == 0 && entry != &map->header;
1200 entry = entry->next) {
1201 vm_object_t obj;
1202 vm_object_t lobj;
1203 vm_object_t tobj;
1204
1205 /*
1206 * Don't dump inaccessible mappings, deal with legacy
1207 * coredump mode.
1208 *
1209 * Note that read-only segments related to the elf binary
1210 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer
1211 * need to arbitrarily ignore such segments.
1212 */
1213 if (elf_legacy_coredump) {
1214 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW)
1215 continue;
1216 } else {
1217 if (writable && (entry->protection & VM_PROT_ALL) == 0)
1218 continue;
1219 }
1220
1221 /*
1222 * Dont include memory segment in the coredump if
1223 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in
1224 * madvise(2).
1225 *
1226 * Currently we only dump normal VM object maps. We do
1227 * not dump submaps or virtual page tables.
1228 */
1229 if (writable && (entry->eflags & MAP_ENTRY_NOCOREDUMP))
1230 continue;
1231 if (entry->maptype != VM_MAPTYPE_NORMAL)
1232 continue;
1233 if ((obj = entry->object.vm_object) == NULL)
1234 continue;
1235
1236 /*
1237 * Find the bottom-most object, leaving the base object
1238 * and the bottom-most object held (but only one hold
1239 * if they happen to be the same).
1240 */
1241 vm_object_hold_shared(obj);
1242
1243 lobj = obj;
1244 while (lobj && (tobj = lobj->backing_object) != NULL) {
1245 KKASSERT(tobj != obj);
1246 vm_object_hold_shared(tobj);
1247 if (tobj == lobj->backing_object) {
1248 if (lobj != obj) {
1249 vm_object_lock_swap();
1250 vm_object_drop(lobj);
1251 }
1252 lobj = tobj;
1253 } else {
1254 vm_object_drop(tobj);
1255 }
1256 }
1257
1258 /*
1259 * The callback only applies to default, swap, or vnode
1260 * objects. Other types of objects such as memory-mapped
1261 * devices are ignored.
1262 */
1263 if (lobj->type == OBJT_DEFAULT || lobj->type == OBJT_SWAP ||
1264 lobj->type == OBJT_VNODE) {
1265 error = (*func)(entry, closure);
1266 }
1267 if (lobj != obj)
1268 vm_object_drop(lobj);
1269 vm_object_drop(obj);
1270 }
1271 return (error);
1272 }
1273
1274 static
1275 void *
1276 target_reserve(elf_buf_t target, size_t bytes, int *error)
1277 {
1278 void *res = NULL;
1279
1280 if (target->buf) {
1281 if (target->off + bytes > target->off_max)
1282 *error = EINVAL;
1283 else
1284 res = target->buf + target->off;
1285 }
1286 target->off += bytes;
1287 return (res);
1288 }
1289
1290 /*
1291 * Write the core file header to the file, including padding up to
1292 * the page boundary.
1293 */
1294 static int
1295 __elfN(corehdr)(struct lwp *lp, int sig, struct file *fp, struct ucred *cred,
1296 int numsegs, elf_buf_t target)
1297 {
1298 int error;
1299 ssize_t nbytes;
1300
1301 /*
1302 * Fill in the header. The fp is passed so we can detect and flag
1303 * a checkpoint file pointer within the core file itself, because
1304 * it may not be restored from the same file handle.
1305 */
1306 error = __elfN(puthdr)(lp, target, sig, WRITE, numsegs, fp);
1307
1308 /* Write it to the core file. */
1309 if (error == 0) {
1310 error = fp_write(fp, target->buf, target->off, &nbytes,
1311 UIO_SYSSPACE);
1312 }
1313 return (error);
1314 }
1315
1316 static int
1317 __elfN(puthdr)(struct lwp *lp, elf_buf_t target, int sig, enum putmode mode,
1318 int numsegs, struct file *fp)
1319 {
1320 struct proc *p = lp->lwp_proc;
1321 int error = 0;
1322 size_t phoff;
1323 size_t noteoff;
1324 size_t notesz;
1325 Elf_Ehdr *ehdr;
1326 Elf_Phdr *phdr;
1327
1328 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error);
1329
1330 phoff = target->off;
1331 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error);
1332
1333 noteoff = target->off;
1334 if (error == 0)
1335 elf_putallnotes(lp, target, sig, mode);
1336 notesz = target->off - noteoff;
1337
1338 /*
1339 * put extra cruft for dumping process state here
1340 * - we really want it be before all the program
1341 * mappings
1342 * - we just need to update the offset accordingly
1343 * and GDB will be none the wiser.
1344 */
1345 if (error == 0)
1346 error = elf_puttextvp(p, target);
1347 if (error == 0)
1348 error = elf_putsigs(lp, target);
1349 if (error == 0)
1350 error = elf_putfiles(p, target, fp);
1351
1352 /*
1353 * Align up to a page boundary for the program segments. The
1354 * actual data will be written to the outptu file, not to elf_buf_t,
1355 * so we do not have to do any further bounds checking.
1356 */
1357 target->off = round_page(target->off);
1358 if (error == 0 && ehdr != NULL) {
1359 /*
1360 * Fill in the ELF header.
1361 */
1362 ehdr->e_ident[EI_MAG0] = ELFMAG0;
1363 ehdr->e_ident[EI_MAG1] = ELFMAG1;
1364 ehdr->e_ident[EI_MAG2] = ELFMAG2;
1365 ehdr->e_ident[EI_MAG3] = ELFMAG3;
1366 ehdr->e_ident[EI_CLASS] = ELF_CLASS;
1367 ehdr->e_ident[EI_DATA] = ELF_DATA;
1368 ehdr->e_ident[EI_VERSION] = EV_CURRENT;
1369 ehdr->e_ident[EI_OSABI] = ELFOSABI_NONE;
1370 ehdr->e_ident[EI_ABIVERSION] = 0;
1371 ehdr->e_ident[EI_PAD] = 0;
1372 ehdr->e_type = ET_CORE;
1373 ehdr->e_machine = ELF_ARCH;
1374 ehdr->e_version = EV_CURRENT;
1375 ehdr->e_entry = 0;
1376 ehdr->e_phoff = phoff;
1377 ehdr->e_flags = 0;
1378 ehdr->e_ehsize = sizeof(Elf_Ehdr);
1379 ehdr->e_phentsize = sizeof(Elf_Phdr);
1380 ehdr->e_phnum = numsegs + 1;
1381 ehdr->e_shentsize = sizeof(Elf_Shdr);
1382 ehdr->e_shnum = 0;
1383 ehdr->e_shstrndx = SHN_UNDEF;
1384 }
1385 if (error == 0 && phdr != NULL) {
1386 /*
1387 * Fill in the program header entries.
1388 */
1389 struct phdr_closure phc;
1390
1391 /* The note segement. */
1392 phdr->p_type = PT_NOTE;
1393 phdr->p_offset = noteoff;
1394 phdr->p_vaddr = 0;
1395 phdr->p_paddr = 0;
1396 phdr->p_filesz = notesz;
1397 phdr->p_memsz = 0;
1398 phdr->p_flags = 0;
1399 phdr->p_align = 0;
1400 ++phdr;
1401
1402 /* All the writable segments from the program. */
1403 phc.phdr = phdr;
1404 phc.phdr_max = phdr + numsegs;
1405 phc.offset = target->off;
1406 each_segment(p, cb_put_phdr, &phc, 1);
1407 }
1408 return (error);
1409 }
1410
1411 /*
1412 * Append core dump notes to target ELF buffer or simply update target size
1413 * if dryrun selected.
1414 */
1415 static int
1416 elf_putallnotes(struct lwp *corelp, elf_buf_t target, int sig,
1417 enum putmode mode)
1418 {
1419 struct proc *p = corelp->lwp_proc;
1420 int error;
1421 struct {
1422 prstatus_t status;
1423 prfpregset_t fpregs;
1424 prpsinfo_t psinfo;
1425 } *tmpdata;
1426 prstatus_t *status;
1427 prfpregset_t *fpregs;
1428 prpsinfo_t *psinfo;
1429 struct lwp *lp;
1430
1431 /*
1432 * Allocate temporary storage for notes on heap to avoid stack overflow.
1433 */
1434 if (mode != DRYRUN) {
1435 tmpdata = kmalloc(sizeof(*tmpdata), M_TEMP, M_ZERO | M_WAITOK);
1436 status = &tmpdata->status;
1437 fpregs = &tmpdata->fpregs;
1438 psinfo = &tmpdata->psinfo;
1439 } else {
1440 tmpdata = NULL;
1441 status = NULL;
1442 fpregs = NULL;
1443 psinfo = NULL;
1444 }
1445
1446 /*
1447 * Append LWP-agnostic note.
1448 */
1449 if (mode != DRYRUN) {
1450 psinfo->pr_version = PRPSINFO_VERSION;
1451 psinfo->pr_psinfosz = sizeof(prpsinfo_t);
1452 strlcpy(psinfo->pr_fname, p->p_comm,
1453 sizeof(psinfo->pr_fname));
1454 /*
1455 * XXX - We don't fill in the command line arguments
1456 * properly yet.
1457 */
1458 strlcpy(psinfo->pr_psargs, p->p_comm,
1459 sizeof(psinfo->pr_psargs));
1460 }
1461 error =
1462 __elfN(putnote)(target, "CORE", NT_PRPSINFO, psinfo, sizeof *psinfo);
1463 if (error)
1464 goto exit;
1465
1466 /*
1467 * Append first note for LWP that triggered core so that it is
1468 * the selected one when the debugger starts.
1469 */
1470 if (mode != DRYRUN) {
1471 status->pr_version = PRSTATUS_VERSION;
1472 status->pr_statussz = sizeof(prstatus_t);
1473 status->pr_gregsetsz = sizeof(gregset_t);
1474 status->pr_fpregsetsz = sizeof(fpregset_t);
1475 status->pr_osreldate = osreldate;
1476 status->pr_cursig = sig;
1477 /*
1478 * XXX GDB needs unique pr_pid for each LWP and does not
1479 * not support pr_pid==0 but lwp_tid can be 0, so hack unique
1480 * value.
1481 */
1482 status->pr_pid = corelp->lwp_tid;
1483 fill_regs(corelp, &status->pr_reg);
1484 fill_fpregs(corelp, fpregs);
1485 }
1486 error =
1487 __elfN(putnote)(target, "CORE", NT_PRSTATUS, status, sizeof *status);
1488 if (error)
1489 goto exit;
1490 error =
1491 __elfN(putnote)(target, "CORE", NT_FPREGSET, fpregs, sizeof *fpregs);
1492 if (error)
1493 goto exit;
1494
1495 /*
1496 * Then append notes for other LWPs.
1497 */
1498 FOREACH_LWP_IN_PROC(lp, p) {
1499 if (lp == corelp)
1500 continue;
1501 /* skip lwps being created */
1502 if (lp->lwp_thread == NULL)
1503 continue;
1504 if (mode != DRYRUN) {
1505 status->pr_pid = lp->lwp_tid;
1506 fill_regs(lp, &status->pr_reg);
1507 fill_fpregs(lp, fpregs);
1508 }
1509 error = __elfN(putnote)(target, "CORE", NT_PRSTATUS,
1510 status, sizeof *status);
1511 if (error)
1512 goto exit;
1513 error = __elfN(putnote)(target, "CORE", NT_FPREGSET,
1514 fpregs, sizeof *fpregs);
1515 if (error)
1516 goto exit;
1517 }
1518
1519 exit:
1520 if (tmpdata != NULL)
1521 kfree(tmpdata, M_TEMP);
1522 return (error);
1523 }
1524
1525 /*
1526 * Generate a note sub-structure.
1527 *
1528 * NOTE: 4-byte alignment.
1529 */
1530 static int
1531 __elfN(putnote)(elf_buf_t target, const char *name, int type,
1532 const void *desc, size_t descsz)
1533 {
1534 int error = 0;
1535 char *dst;
1536 Elf_Note note;
1537
1538 note.n_namesz = strlen(name) + 1;
1539 note.n_descsz = descsz;
1540 note.n_type = type;
1541 dst = target_reserve(target, sizeof(note), &error);
1542 if (dst != NULL)
1543 bcopy(&note, dst, sizeof note);
1544 dst = target_reserve(target, note.n_namesz, &error);
1545 if (dst != NULL)
1546 bcopy(name, dst, note.n_namesz);
1547 target->off = roundup2(target->off, sizeof(Elf_Word));
1548 dst = target_reserve(target, note.n_descsz, &error);
1549 if (dst != NULL)
1550 bcopy(desc, dst, note.n_descsz);
1551 target->off = roundup2(target->off, sizeof(Elf_Word));
1552 return (error);
1553 }
1554
1555
1556 static int
1557 elf_putsigs(struct lwp *lp, elf_buf_t target)
1558 {
1559 /* XXX lwp handle more than one lwp */
1560 struct proc *p = lp->lwp_proc;
1561 int error = 0;
1562 struct ckpt_siginfo *csi;
1563
1564 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error);
1565 if (csi) {
1566 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo);
1567 bcopy(p->p_sigacts, &csi->csi_sigacts, sizeof(*p->p_sigacts));
1568 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval));
1569 bcopy(&lp->lwp_sigmask, &csi->csi_sigmask,
1570 sizeof(sigset_t));
1571 csi->csi_sigparent = p->p_sigparent;
1572 }
1573 return (error);
1574 }
1575
1576 static int
1577 elf_putfiles(struct proc *p, elf_buf_t target, struct file *ckfp)
1578 {
1579 int error = 0;
1580 int i;
1581 struct ckpt_filehdr *cfh = NULL;
1582 struct ckpt_fileinfo *cfi;
1583 struct file *fp;
1584 struct vnode *vp;
1585 /*
1586 * the duplicated loop is gross, but it was the only way
1587 * to eliminate uninitialized variable warnings
1588 */
1589 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error);
1590 if (cfh) {
1591 cfh->cfh_nfiles = 0;
1592 }
1593
1594 /*
1595 * ignore STDIN/STDERR/STDOUT.
1596 */
1597 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) {
1598 fp = holdfp(p->p_fd, i, -1);
1599 if (fp == NULL)
1600 continue;
1601 /*
1602 * XXX Only checkpoint vnodes for now.
1603 */
1604 if (fp->f_type != DTYPE_VNODE) {
1605 fdrop(fp);
1606 continue;
1607 }
1608 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo),
1609 &error);
1610 if (cfi == NULL) {
1611 fdrop(fp);
1612 continue;
1613 }
1614 cfi->cfi_index = -1;
1615 cfi->cfi_type = fp->f_type;
1616 cfi->cfi_flags = fp->f_flag;
1617 cfi->cfi_offset = fp->f_offset;
1618 cfi->cfi_ckflags = 0;
1619
1620 if (fp == ckfp)
1621 cfi->cfi_ckflags |= CKFIF_ISCKPTFD;
1622 /* f_count and f_msgcount should not be saved/restored */
1623 /* XXX save cred info */
1624
1625 switch(fp->f_type) {
1626 case DTYPE_VNODE:
1627 vp = (struct vnode *)fp->f_data;
1628 /*
1629 * it looks like a bug in ptrace is marking
1630 * a non-vnode as a vnode - until we find the
1631 * root cause this will at least prevent
1632 * further panics from truss
1633 */
1634 if (vp == NULL || vp->v_mount == NULL)
1635 break;
1636 cfh->cfh_nfiles++;
1637 cfi->cfi_index = i;
1638 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1639 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid);
1640 break;
1641 default:
1642 break;
1643 }
1644 fdrop(fp);
1645 }
1646 return (error);
1647 }
1648
1649 static int
1650 elf_puttextvp(struct proc *p, elf_buf_t target)
1651 {
1652 int error = 0;
1653 int *vn_count;
1654 struct fp_closure fpc;
1655 struct ckpt_vminfo *vminfo;
1656
1657 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error);
1658 if (vminfo != NULL) {
1659 vminfo->cvm_dsize = p->p_vmspace->vm_dsize;
1660 vminfo->cvm_tsize = p->p_vmspace->vm_tsize;
1661 vminfo->cvm_daddr = p->p_vmspace->vm_daddr;
1662 vminfo->cvm_taddr = p->p_vmspace->vm_taddr;
1663 }
1664
1665 fpc.count = 0;
1666 vn_count = target_reserve(target, sizeof(int), &error);
1667 if (target->buf != NULL) {
1668 fpc.vnh = (struct vn_hdr *)(target->buf + target->off);
1669 fpc.vnh_max = fpc.vnh +
1670 (target->off_max - target->off) / sizeof(struct vn_hdr);
1671 error = each_segment(p, cb_put_fp, &fpc, 0);
1672 if (vn_count)
1673 *vn_count = fpc.count;
1674 } else {
1675 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0);
1676 }
1677 target->off += fpc.count * sizeof(struct vn_hdr);
1678 return (error);
1679 }
1680
1681 /*
1682 * Try to find the appropriate ABI-note section for checknote,
1683 * The entire image is searched if necessary, not only the first page.
1684 */
1685 static boolean_t
1686 __elfN(check_note)(struct image_params *imgp, Elf_Brandnote *checknote,
1687 int32_t *osrel)
1688 {
1689 boolean_t valid_note_found;
1690 const Elf_Phdr *phdr, *pnote;
1691 const Elf_Ehdr *hdr;
1692 int i;
1693
1694 valid_note_found = FALSE;
1695 hdr = (const Elf_Ehdr *)imgp->image_header;
1696 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
1697
1698 for (i = 0; i < hdr->e_phnum; i++) {
1699 if (phdr[i].p_type == PT_NOTE) {
1700 pnote = &phdr[i];
1701 valid_note_found = check_PT_NOTE (imgp, checknote,
1702 osrel, pnote);
1703 if (valid_note_found)
1704 break;
1705 }
1706 }
1707 return valid_note_found;
1708 }
1709
1710 /*
1711 * Be careful not to create new overflow conditions when checking
1712 * for overflow.
1713 */
1714 static boolean_t
1715 note_overflow(const Elf_Note *note, size_t maxsize)
1716 {
1717 if (sizeof(*note) > maxsize)
1718 return TRUE;
1719 if (note->n_namesz > maxsize - sizeof(*note))
1720 return TRUE;
1721 return FALSE;
1722 }
1723
1724 static boolean_t
1725 hdr_overflow(__ElfN(Off) off_beg, __ElfN(Size) size)
1726 {
1727 __ElfN(Off) off_end;
1728
1729 off_end = off_beg + size;
1730 if (off_end < off_beg)
1731 return TRUE;
1732 return FALSE;
1733 }
1734
1735 static boolean_t
1736 check_PT_NOTE(struct image_params *imgp, Elf_Brandnote *checknote,
1737 int32_t *osrel, const Elf_Phdr * pnote)
1738 {
1739 boolean_t limited_to_first_page;
1740 boolean_t found = FALSE;
1741 const Elf_Note *note, *note0, *note_end;
1742 const char *note_name;
1743 __ElfN(Off) noteloc, firstloc;
1744 __ElfN(Size) notesz, firstlen, endbyte;
1745 struct lwbuf *lwb;
1746 struct lwbuf lwb_cache;
1747 const char *page;
1748 char *data = NULL;
1749 int n;
1750
1751 if (hdr_overflow(pnote->p_offset, pnote->p_filesz))
1752 return (FALSE);
1753 notesz = pnote->p_filesz;
1754 noteloc = pnote->p_offset;
1755 endbyte = noteloc + notesz;
1756 limited_to_first_page = noteloc < PAGE_SIZE && endbyte < PAGE_SIZE;
1757
1758 if (limited_to_first_page) {
1759 note = (const Elf_Note *)(imgp->image_header + noteloc);
1760 note_end = (const Elf_Note *)(imgp->image_header + endbyte);
1761 note0 = note;
1762 } else {
1763 firstloc = noteloc & PAGE_MASK;
1764 firstlen = PAGE_SIZE - firstloc;
1765 if (notesz < sizeof(Elf_Note) || notesz > PAGE_SIZE)
1766 return (FALSE);
1767
1768 lwb = &lwb_cache;
1769 if (exec_map_page(imgp, noteloc >> PAGE_SHIFT, &lwb, &page))
1770 return (FALSE);
1771 if (firstlen < notesz) { /* crosses page boundary */
1772 data = kmalloc(notesz, M_TEMP, M_WAITOK);
1773 bcopy(page + firstloc, data, firstlen);
1774
1775 exec_unmap_page(lwb);
1776 lwb = &lwb_cache;
1777 if (exec_map_page(imgp, (noteloc >> PAGE_SHIFT) + 1,
1778 &lwb, &page)) {
1779 kfree(data, M_TEMP);
1780 return (FALSE);
1781 }
1782 bcopy(page, data + firstlen, notesz - firstlen);
1783 note = note0 = (const Elf_Note *)(data);
1784 note_end = (const Elf_Note *)(data + notesz);
1785 } else {
1786 note = note0 = (const Elf_Note *)(page + firstloc);
1787 note_end = (const Elf_Note *)(page + firstloc +
1788 firstlen);
1789 }
1790 }
1791
1792 for (n = 0; n < 100 && note >= note0 && note < note_end; n++) {
1793 if (!aligned(note, Elf32_Addr))
1794 break;
1795 if (note_overflow(note, (const char *)note_end -
1796 (const char *)note)) {
1797 break;
1798 }
1799 note_name = (const char *)(note + 1);
1800
1801 if (note->n_namesz == checknote->hdr.n_namesz
1802 && note->n_descsz == checknote->hdr.n_descsz
1803 && note->n_type == checknote->hdr.n_type
1804 && (strncmp(checknote->vendor, note_name,
1805 checknote->hdr.n_namesz) == 0)) {
1806 /* Fetch osreldata from ABI.note-tag */
1807 if ((checknote->flags & BN_TRANSLATE_OSREL) != 0 &&
1808 checknote->trans_osrel != NULL)
1809 checknote->trans_osrel(note, osrel);
1810 found = TRUE;
1811 break;
1812 }
1813 note = (const Elf_Note *)((const char *)(note + 1) +
1814 roundup2(note->n_namesz, sizeof(Elf32_Addr)) +
1815 roundup2(note->n_descsz, sizeof(Elf32_Addr)));
1816 }
1817
1818 if (!limited_to_first_page) {
1819 if (data != NULL)
1820 kfree(data, M_TEMP);
1821 exec_unmap_page(lwb);
1822 }
1823 return (found);
1824 }
1825
1826 /*
1827 * The interpreter program header may be located beyond the first page, so
1828 * regardless of its location, a copy of the interpreter path is created so
1829 * that it may be safely referenced by the calling function in all case. The
1830 * memory is allocated by calling function, and the copying is done here.
1831 */
1832 static boolean_t
1833 extract_interpreter(struct image_params *imgp, const Elf_Phdr *pinterpreter,
1834 char *data)
1835 {
1836 boolean_t limited_to_first_page;
1837 const boolean_t result_success = FALSE;
1838 const boolean_t result_failure = TRUE;
1839 __ElfN(Off) pathloc, firstloc;
1840 __ElfN(Size) pathsz, firstlen, endbyte;
1841 struct lwbuf *lwb;
1842 struct lwbuf lwb_cache;
1843 const char *page;
1844
1845 if (hdr_overflow(pinterpreter->p_offset, pinterpreter->p_filesz))
1846 return (result_failure);
1847 pathsz = pinterpreter->p_filesz;
1848 pathloc = pinterpreter->p_offset;
1849 endbyte = pathloc + pathsz;
1850
1851 limited_to_first_page = pathloc < PAGE_SIZE && endbyte < PAGE_SIZE;
1852 if (limited_to_first_page) {
1853 bcopy(imgp->image_header + pathloc, data, pathsz);
1854 return (result_success);
1855 }
1856
1857 firstloc = pathloc & PAGE_MASK;
1858 firstlen = PAGE_SIZE - firstloc;
1859
1860 lwb = &lwb_cache;
1861 if (exec_map_page(imgp, pathloc >> PAGE_SHIFT, &lwb, &page))
1862 return (result_failure);
1863
1864 if (firstlen < pathsz) { /* crosses page boundary */
1865 bcopy(page + firstloc, data, firstlen);
1866
1867 exec_unmap_page(lwb);
1868 lwb = &lwb_cache;
1869 if (exec_map_page(imgp, (pathloc >> PAGE_SHIFT) + 1, &lwb,
1870 &page))
1871 return (result_failure);
1872 bcopy(page, data + firstlen, pathsz - firstlen);
1873 } else
1874 bcopy(page + firstloc, data, pathsz);
1875
1876 exec_unmap_page(lwb);
1877 return (result_success);
1878 }
1879
1880 static boolean_t
1881 __elfN(bsd_trans_osrel)(const Elf_Note *note, int32_t *osrel)
1882 {
1883 uintptr_t p;
1884
1885 p = (uintptr_t)(note + 1);
1886 p += roundup2(note->n_namesz, sizeof(Elf32_Addr));
1887 *osrel = *(const int32_t *)(p);
1888
1889 return (TRUE);
1890 }
1891
1892 /*
1893 * Tell kern_execve.c about it, with a little help from the linker.
1894 */
1895 #if defined(__x86_64__)
1896 static struct execsw elf_execsw = {exec_elf64_imgact, "ELF64"};
1897 EXEC_SET_ORDERED(elf64, elf_execsw, SI_ORDER_FIRST);
1898 #else /* i386 assumed */
1899 static struct execsw elf_execsw = {exec_elf32_imgact, "ELF32"};
1900 EXEC_SET_ORDERED(elf32, elf_execsw, SI_ORDER_FIRST);
1901 #endif
1902
1903 static vm_prot_t
1904 __elfN(trans_prot)(Elf_Word flags)
1905 {
1906 vm_prot_t prot;
1907
1908 prot = 0;
1909 if (flags & PF_X)
1910 prot |= VM_PROT_EXECUTE;
1911 if (flags & PF_W)
1912 prot |= VM_PROT_WRITE;
1913 if (flags & PF_R)
1914 prot |= VM_PROT_READ;
1915 return (prot);
1916 }
1917
1918 static Elf_Word
1919 __elfN(untrans_prot)(vm_prot_t prot)
1920 {
1921 Elf_Word flags;
1922
1923 flags = 0;
1924 if (prot & VM_PROT_EXECUTE)
1925 flags |= PF_X;
1926 if (prot & VM_PROT_READ)
1927 flags |= PF_R;
1928 if (prot & VM_PROT_WRITE)
1929 flags |= PF_W;
1930 return (flags);
1931 }
1932
1933 static u_long
1934 pie_base_hint(struct proc *p)
1935 {
1936 u_long base;
1937
1938 if (elf_pie_base_mmap)
1939 base = vm_map_hint(p, 0, VM_PROT_READ | VM_PROT_EXECUTE);
1940 else
1941 base = ET_DYN_LOAD_ADDR;
1942 return base;
1943 }
DragonFly BSD