2 * Copyright (c) 1995-1996 Søren Schmidt
3 * Copyright (c) 1996 Peter Wemm
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer
11 * in this position and unchanged.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. The name of the author may not be used to endorse or promote products
16 * derived from this software withough specific prior written permission
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
19 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
20 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
22 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
23 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
24 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
25 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
27 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 * $FreeBSD: src/sys/kern/imgact_elf.c,v 1.73.2.13 2002/12/28 19:49:41 dillon Exp $
30 * $DragonFly: src/sys/kern/imgact_elf.c,v 1.35 2006/03/29 18:44:50 dillon Exp $
33 #include <sys/param.h>
35 #include <sys/fcntl.h>
37 #include <sys/imgact.h>
38 #include <sys/imgact_elf.h>
39 #include <sys/kernel.h>
40 #include <sys/malloc.h>
42 #include <sys/systm.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>
50 #include <sys/syscall.h>
51 #include <sys/sysctl.h>
52 #include <sys/sysent.h>
53 #include <sys/vnode.h>
56 #include <vm/vm_kern.h>
57 #include <vm/vm_param.h>
60 #include <vm/vm_map.h>
61 #include <vm/vm_object.h>
62 #include <vm/vm_extern.h>
64 #include <machine/elf.h>
65 #include <machine/md_var.h>
66 #include <sys/mount.h>
68 #define OLD_EI_BRAND 8
73 static int elf_check_header (const Elf_Ehdr *hdr);
74 static int elf_freebsd_fixup (register_t **stack_base,
75 struct image_params *imgp);
76 static int elf_load_file (struct proc *p, const char *file, u_long *addr,
78 static int elf_load_section (struct proc *p,
79 struct vmspace *vmspace, struct vnode *vp,
80 vm_offset_t offset, caddr_t vmaddr, size_t memsz, size_t filsz,
82 static int exec_elf_imgact (struct image_params *imgp);
84 static int elf_trace = 0;
85 SYSCTL_INT(_debug, OID_AUTO, elf_trace, CTLFLAG_RW, &elf_trace, 0, "");
86 static int elf_legacy_coredump = 0;
87 SYSCTL_INT(_debug, OID_AUTO, elf_legacy_coredump, CTLFLAG_RW,
88 &elf_legacy_coredump, 0, "");
90 static int dragonfly_match_abi_note(const Elf_Note *);
91 static int freebsd_match_abi_note(const Elf_Note *);
93 static struct sysentvec elf_freebsd_sysvec = {
113 static Elf_Brandinfo freebsd_brand_info = {
116 freebsd_match_abi_note,
118 "/usr/libexec/ld-elf.so.1",
122 static Elf_Brandinfo dragonfly_brand_info = {
125 dragonfly_match_abi_note,
127 "/usr/libexec/ld-elf.so.2",
131 static Elf_Brandinfo *elf_brand_list[MAX_BRANDS] = {
132 &dragonfly_brand_info,
139 freebsd_match_abi_note(const Elf_Note *abi_note)
141 const char *abi_name = (const char *)
142 ((const uint8_t *)abi_note + sizeof(*abi_note));
144 if (abi_note->n_namesz != sizeof("FreeBSD"))
146 if (memcmp(abi_name, "FreeBSD", sizeof("FreeBSD")))
152 dragonfly_match_abi_note(const Elf_Note *abi_note)
154 const char *abi_name = (const char *)
155 ((const uint8_t *)abi_note + sizeof(*abi_note));
157 if (abi_note->n_namesz != sizeof("DragonFly"))
159 if (memcmp(abi_name, "DragonFly", sizeof("DragonFly")))
165 elf_insert_brand_entry(Elf_Brandinfo *entry)
169 for (i=1; i<MAX_BRANDS; i++) {
170 if (elf_brand_list[i] == NULL) {
171 elf_brand_list[i] = entry;
181 elf_remove_brand_entry(Elf_Brandinfo *entry)
185 for (i=1; i<MAX_BRANDS; i++) {
186 if (elf_brand_list[i] == entry) {
187 elf_brand_list[i] = NULL;
197 elf_brand_inuse(Elf_Brandinfo *entry)
202 FOREACH_PROC_IN_SYSTEM(p) {
203 if (p->p_sysent == entry->sysvec) {
213 elf_check_header(const Elf_Ehdr *hdr)
216 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
217 hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
218 hdr->e_ident[EI_VERSION] != EV_CURRENT ||
219 hdr->e_phentsize != sizeof(Elf_Phdr) ||
220 hdr->e_ehsize != sizeof(Elf_Ehdr) ||
221 hdr->e_version != ELF_TARG_VER)
224 if (!ELF_MACHINE_OK(hdr->e_machine))
231 elf_load_section(struct proc *p, struct vmspace *vmspace, struct vnode *vp,
232 vm_offset_t offset, caddr_t vmaddr, size_t memsz,
233 size_t filsz, vm_prot_t prot)
236 vm_offset_t map_addr;
241 vm_offset_t file_addr;
242 vm_offset_t data_buf = 0;
244 object = vp->v_object;
248 * It's necessary to fail if the filsz + offset taken from the
249 * header is greater than the actual file pager object's size.
250 * If we were to allow this, then the vm_map_find() below would
251 * walk right off the end of the file object and into the ether.
253 * While I'm here, might as well check for something else that
254 * is invalid: filsz cannot be greater than memsz.
256 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) {
257 uprintf("elf_load_section: truncated ELF file\n");
261 map_addr = trunc_page((vm_offset_t)vmaddr);
262 file_addr = trunc_page(offset);
265 * We have two choices. We can either clear the data in the last page
266 * of an oversized mapping, or we can start the anon mapping a page
267 * early and copy the initialized data into that first page. We
268 * choose the second..
271 map_len = trunc_page(offset+filsz) - file_addr;
273 map_len = round_page(offset+filsz) - file_addr;
276 vm_object_reference(object);
278 /* cow flags: don't dump readonly sections in core */
279 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT |
280 (prot & VM_PROT_WRITE ? 0 : MAP_DISABLE_COREDUMP);
282 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
283 vm_map_lock(&vmspace->vm_map);
284 rv = vm_map_insert(&vmspace->vm_map, &count,
286 file_addr, /* file offset */
287 map_addr, /* virtual start */
288 map_addr + map_len,/* virtual end */
292 vm_map_unlock(&vmspace->vm_map);
293 vm_map_entry_release(count);
294 if (rv != KERN_SUCCESS) {
295 vm_object_deallocate(object);
299 /* we can stop now if we've covered it all */
300 if (memsz == filsz) {
307 * We have to get the remaining bit of the file into the first part
308 * of the oversized map segment. This is normally because the .data
309 * segment in the file is extended to provide bss. It's a neat idea
310 * to try and save a page, but it's a pain in the behind to implement.
312 copy_len = (offset + filsz) - trunc_page(offset + filsz);
313 map_addr = trunc_page((vm_offset_t)vmaddr + filsz);
314 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr;
316 /* This had damn well better be true! */
318 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
319 vm_map_lock(&vmspace->vm_map);
320 rv = vm_map_insert(&vmspace->vm_map, &count,
322 map_addr, map_addr + map_len,
323 VM_PROT_ALL, VM_PROT_ALL, 0);
324 vm_map_unlock(&vmspace->vm_map);
325 vm_map_entry_release(count);
326 if (rv != KERN_SUCCESS) {
332 vm_object_reference(object);
333 rv = vm_map_find(exec_map,
335 trunc_page(offset + filsz),
341 MAP_COPY_ON_WRITE | MAP_PREFAULT_PARTIAL);
342 if (rv != KERN_SUCCESS) {
343 vm_object_deallocate(object);
347 /* send the page fragment to user space */
348 error = copyout((caddr_t)data_buf, (caddr_t)map_addr, copy_len);
349 vm_map_remove(exec_map, data_buf, data_buf + PAGE_SIZE);
356 * set it to the specified protection
358 vm_map_protect(&vmspace->vm_map, map_addr, map_addr + map_len, prot,
365 * Load the file "file" into memory. It may be either a shared object
368 * The "addr" reference parameter is in/out. On entry, it specifies
369 * the address where a shared object should be loaded. If the file is
370 * an executable, this value is ignored. On exit, "addr" specifies
371 * where the file was actually loaded.
373 * The "entry" reference parameter is out only. On exit, it specifies
374 * the entry point for the loaded file.
377 elf_load_file(struct proc *p, const char *file, u_long *addr, u_long *entry)
380 struct nlookupdata nd;
382 struct image_params image_params;
384 const Elf_Ehdr *hdr = NULL;
385 const Elf_Phdr *phdr = NULL;
386 struct nlookupdata *nd;
387 struct vmspace *vmspace = p->p_vmspace;
389 struct image_params *imgp;
392 u_long base_addr = 0;
393 int error, i, numsegs;
394 struct thread *td = p->p_thread;
396 tempdata = malloc(sizeof(*tempdata), M_TEMP, M_WAITOK);
398 attr = &tempdata->attr;
399 imgp = &tempdata->image_params;
402 * Initialize part of the common data
406 imgp->firstpage = NULL;
407 imgp->image_header = NULL;
410 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW);
414 error = cache_vget(nd->nl_ncp, nd->nl_cred, LK_EXCLUSIVE, &imgp->vp);
420 * Check permissions, modes, uid, etc on the file, and "open" it.
422 error = exec_check_permissions(imgp);
424 VOP_UNLOCK(imgp->vp, 0, td);
428 error = exec_map_first_page(imgp);
430 * Also make certain that the interpreter stays the same, so set
431 * its VTEXT flag, too.
434 imgp->vp->v_flag |= VTEXT;
435 VOP_UNLOCK(imgp->vp, 0, td);
439 hdr = (const Elf_Ehdr *)imgp->image_header;
440 if ((error = elf_check_header(hdr)) != 0)
442 if (hdr->e_type == ET_DYN)
444 else if (hdr->e_type == ET_EXEC)
451 /* Only support headers that fit within first page for now
452 * (multiplication of two Elf_Half fields will not overflow) */
453 if ((hdr->e_phoff > PAGE_SIZE) ||
454 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) {
459 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
461 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) {
462 if (phdr[i].p_type == PT_LOAD) { /* Loadable segment */
464 if (phdr[i].p_flags & PF_X)
465 prot |= VM_PROT_EXECUTE;
466 if (phdr[i].p_flags & PF_W)
467 prot |= VM_PROT_WRITE;
468 if (phdr[i].p_flags & PF_R)
469 prot |= VM_PROT_READ;
471 error = elf_load_section(
472 p, vmspace, imgp->vp,
474 (caddr_t)phdr[i].p_vaddr +
477 phdr[i].p_filesz, prot);
481 * Establish the base address if this is the
485 base_addr = trunc_page(phdr[i].p_vaddr + rbase);
490 *entry=(unsigned long)hdr->e_entry + rbase;
494 exec_unmap_first_page(imgp);
499 free(tempdata, M_TEMP);
505 * non static, as it can be overridden by start_init()
507 int fallback_elf_brand = -1;
508 SYSCTL_INT(_kern, OID_AUTO, fallback_elf_brand, CTLFLAG_RW,
509 &fallback_elf_brand, -1,
510 "ELF brand of last resort");
513 exec_elf_imgact(struct image_params *imgp)
515 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header;
516 const Elf_Phdr *phdr;
517 Elf_Auxargs *elf_auxargs = NULL;
518 struct vmspace *vmspace;
520 u_long text_size = 0, data_size = 0, total_size = 0;
521 u_long text_addr = 0, data_addr = 0;
522 u_long seg_size, seg_addr;
523 u_long addr, entry = 0, proghdr = 0;
525 const char *interp = NULL;
526 const Elf_Note *abi_note = NULL;
527 Elf_Brandinfo *brand_info;
533 * Do we have a valid ELF header ?
535 if (elf_check_header(hdr) != 0 || hdr->e_type != ET_EXEC)
539 * From here on down, we return an errno, not -1, as we've
540 * detected an ELF file.
543 if ((hdr->e_phoff > PAGE_SIZE) ||
544 (hdr->e_phoff + hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE) {
545 /* Only support headers in first page for now */
548 phdr = (const Elf_Phdr*)(imgp->image_header + hdr->e_phoff);
551 * From this point on, we may have resources that need to be freed.
554 exec_new_vmspace(imgp, NULL);
557 * Yeah, I'm paranoid. There is every reason in the world to get
558 * VTEXT now since from here on out, there are places we can have
559 * a context switch. Better safe than sorry; I really don't want
560 * the file to change while it's being loaded.
562 vsetflags(imgp->vp, VTEXT);
564 vmspace = imgp->proc->p_vmspace;
566 for (i = 0; i < hdr->e_phnum; i++) {
567 switch(phdr[i].p_type) {
569 case PT_LOAD: /* Loadable segment */
571 if (phdr[i].p_flags & PF_X)
572 prot |= VM_PROT_EXECUTE;
573 if (phdr[i].p_flags & PF_W)
574 prot |= VM_PROT_WRITE;
575 if (phdr[i].p_flags & PF_R)
576 prot |= VM_PROT_READ;
578 if ((error = elf_load_section(imgp->proc,
581 (caddr_t)phdr[i].p_vaddr,
583 phdr[i].p_filesz, prot)) != 0)
587 * If this segment contains the program headers,
588 * remember their virtual address for the AT_PHDR
589 * aux entry. Static binaries don't usually include
592 if (phdr[i].p_offset == 0 &&
593 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize
595 proghdr = phdr[i].p_vaddr + hdr->e_phoff;
597 seg_addr = trunc_page(phdr[i].p_vaddr);
598 seg_size = round_page(phdr[i].p_memsz +
599 phdr[i].p_vaddr - seg_addr);
602 * Is this .text or .data? We can't use
603 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the
604 * alpha terribly and possibly does other bad
605 * things so we stick to the old way of figuring
606 * it out: If the segment contains the program
607 * entry point, it's a text segment, otherwise it
610 * Note that obreak() assumes that data_addr +
611 * data_size == end of data load area, and the ELF
612 * file format expects segments to be sorted by
613 * address. If multiple data segments exist, the
614 * last one will be used.
616 if (hdr->e_entry >= phdr[i].p_vaddr &&
617 hdr->e_entry < (phdr[i].p_vaddr +
619 text_size = seg_size;
620 text_addr = seg_addr;
621 entry = (u_long)hdr->e_entry;
623 data_size = seg_size;
624 data_addr = seg_addr;
626 total_size += seg_size;
629 * Check limits. It should be safe to check the
630 * limits after loading the segment since we do
631 * not actually fault in all the segment's pages.
634 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur ||
635 text_size > maxtsiz ||
637 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
642 case PT_INTERP: /* Path to interpreter */
643 if (phdr[i].p_filesz > MAXPATHLEN ||
644 phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE) {
648 interp = imgp->image_header + phdr[i].p_offset;
650 case PT_NOTE: /* Check for .note.ABI-tag */
652 const Elf_Note *tmp_note;
653 /* XXX handle anything outside the first page */
654 if (phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE)
656 if (phdr[i].p_filesz < sizeof(Elf_Note))
657 continue; /* ENOEXEC? */
658 tmp_note = (const Elf_Note *)(imgp->image_header + phdr[i].p_offset);
659 if (tmp_note->n_type != 1)
661 if (tmp_note->n_namesz + sizeof(Elf_Note) +
662 tmp_note->n_descsz > phdr[i].p_filesz)
663 continue; /* ENOEXEC? */
667 case PT_PHDR: /* Program header table info */
668 proghdr = phdr[i].p_vaddr;
675 vmspace->vm_tsize = text_size >> PAGE_SHIFT;
676 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr;
677 vmspace->vm_dsize = data_size >> PAGE_SHIFT;
678 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr;
680 addr = ELF_RTLD_ADDR(vmspace);
682 imgp->entry_addr = entry;
686 /* We support three types of branding -- (1) the ELF EI_OSABI field
687 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string
688 * branding w/in the ELF header, and (3) path of the `interp_path'
689 * field. We should also look for an ".note.ABI-tag" ELF section now
690 * in all Linux ELF binaries, FreeBSD 4.1+, and some NetBSD ones.
693 /* If the executable has a brand, search for it in the brand list. */
694 if (brand_info == NULL && hdr->e_ident[EI_OSABI] != ELFOSABI_NONE) {
695 for (i = 0; i < MAX_BRANDS; i++) {
696 Elf_Brandinfo *bi = elf_brand_list[i];
699 (hdr->e_ident[EI_OSABI] == bi->brand
701 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND],
702 bi->compat_3_brand, strlen(bi->compat_3_brand)))) {
709 /* Search for a recognized ABI. */
710 if (brand_info == NULL && abi_note != NULL) {
711 for (i = 0; i < MAX_BRANDS; i++) {
712 Elf_Brandinfo *bi = elf_brand_list[i];
714 if (bi != NULL && bi->match_abi_note != NULL &&
715 (*bi->match_abi_note)(abi_note)) {
723 * ELFOSABI_NONE == ELFOSABI_SYSV, so a SYSV binary misses all
724 * checks so far, since it is neither branded nor does it have
725 * an ABI note. If the EI_OSABI field is ELFOSABI_NONE, assume
726 * it is svr4 and look for an entry in the elf_brand_list with
727 * match_abi_note == NULL.
729 if (brand_info == NULL && hdr->e_ident[EI_OSABI] == ELFOSABI_NONE) {
730 for (i = 0; i < MAX_BRANDS; i++) {
731 Elf_Brandinfo *bi = elf_brand_list[i];
733 if (bi != NULL && bi->match_abi_note == NULL &&
734 ELFOSABI_SYSV == bi->brand) {
741 /* Lacking a recognized ABI, search for a recognized interpreter. */
742 if (brand_info == NULL && interp != NULL) {
743 for (i = 0; i < MAX_BRANDS; i++) {
744 Elf_Brandinfo *bi = elf_brand_list[i];
747 strcmp(interp, bi->interp_path) == 0) {
754 /* Lacking a recognized interpreter, try the default brand */
755 if (brand_info == NULL) {
756 for (i = 0; i < MAX_BRANDS; i++) {
757 Elf_Brandinfo *bi = elf_brand_list[i];
759 if (bi != NULL && fallback_elf_brand == bi->brand) {
766 if (brand_info == NULL) {
767 uprintf("ELF binary type \"%u\" not known.\n",
768 hdr->e_ident[EI_OSABI]);
773 imgp->proc->p_sysent = brand_info->sysvec;
774 if (interp != NULL) {
775 path = malloc(MAXPATHLEN, M_TEMP, M_WAITOK);
776 snprintf(path, MAXPATHLEN, "%s%s",
777 brand_info->emul_path, interp);
778 if ((error = elf_load_file(imgp->proc, path, &addr,
779 &imgp->entry_addr)) != 0) {
780 if ((error = elf_load_file(imgp->proc, interp, &addr,
781 &imgp->entry_addr)) != 0) {
782 uprintf("ELF interpreter %s not found\n", path);
791 * Construct auxargs table (used by the fixup routine)
793 elf_auxargs = malloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK);
794 elf_auxargs->execfd = -1;
795 elf_auxargs->phdr = proghdr;
796 elf_auxargs->phent = hdr->e_phentsize;
797 elf_auxargs->phnum = hdr->e_phnum;
798 elf_auxargs->pagesz = PAGE_SIZE;
799 elf_auxargs->base = addr;
800 elf_auxargs->flags = 0;
801 elf_auxargs->entry = entry;
802 elf_auxargs->trace = elf_trace;
804 imgp->auxargs = elf_auxargs;
805 imgp->interpreted = 0;
812 elf_freebsd_fixup(register_t **stack_base, struct image_params *imgp)
814 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs;
817 pos = *stack_base + (imgp->args->argc + imgp->args->envc + 2);
820 AUXARGS_ENTRY(pos, AT_DEBUG, 1);
822 if (args->execfd != -1) {
823 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
825 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
826 AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
827 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
828 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
829 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
830 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
831 AUXARGS_ENTRY(pos, AT_BASE, args->base);
832 AUXARGS_ENTRY(pos, AT_NULL, 0);
834 free(imgp->auxargs, M_TEMP);
835 imgp->auxargs = NULL;
838 suword(*stack_base, (long) imgp->args->argc);
843 * Code for generating ELF core dumps.
846 typedef int (*segment_callback) (vm_map_entry_t, void *);
848 /* Closure for cb_put_phdr(). */
849 struct phdr_closure {
850 Elf_Phdr *phdr; /* Program header to fill in (incremented) */
851 Elf_Phdr *phdr_max; /* Pointer bound for error check */
852 Elf_Off offset; /* Offset of segment in core file */
855 /* Closure for cb_size_segment(). */
856 struct sseg_closure {
857 int count; /* Count of writable segments. */
858 size_t vsize; /* Total size of all writable segments. */
861 /* Closure for cb_put_fp(). */
864 struct vn_hdr *vnh_max;
869 typedef struct elf_buf {
875 static void *target_reserve(elf_buf_t target, size_t bytes, int *error);
877 static int cb_put_phdr (vm_map_entry_t, void *);
878 static int cb_size_segment (vm_map_entry_t, void *);
879 static int cb_fpcount_segment(vm_map_entry_t, void *);
880 static int cb_put_fp(vm_map_entry_t, void *);
883 static int each_segment (struct proc *, segment_callback, void *, int);
884 static int elf_corehdr (struct proc *, struct file *, struct ucred *,
886 static int elf_puthdr (struct proc *, elf_buf_t, const prstatus_t *,
887 const prfpregset_t *, const prpsinfo_t *, int);
888 static int elf_putnote (elf_buf_t, const char *, int, const void *, size_t);
890 static int elf_putsigs(struct proc *, elf_buf_t);
891 static int elf_puttextvp(struct proc *, elf_buf_t);
892 static int elf_putfiles(struct proc *, elf_buf_t);
894 extern int osreldate;
897 elf_coredump(struct proc *p, struct vnode *vp, off_t limit)
902 if ((error = falloc(NULL, &fp, NULL)) != 0)
904 fsetcred(fp, p->p_ucred);
909 fp->f_type = DTYPE_VNODE;
910 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW;
911 fp->f_ops = &vnode_fileops;
913 VOP_UNLOCK(vp, 0, p->p_thread);
915 error = generic_elf_coredump(p, fp, limit);
919 fp->f_ops = &badfileops;
921 fdrop(fp, p->p_thread);
926 generic_elf_coredump(struct proc *p, struct file *fp, off_t limit)
928 struct ucred *cred = p->p_ucred;
930 struct sseg_closure seginfo;
931 struct elf_buf target;
934 printf("can't dump core - null fp\n");
937 * Size the program segments
941 each_segment(p, cb_size_segment, &seginfo, 1);
944 * Calculate the size of the core file header area by making
945 * a dry run of generating it. Nothing is written, but the
946 * size is calculated.
948 bzero(&target, sizeof(target));
949 elf_puthdr(p, &target, NULL, NULL, NULL, seginfo.count);
951 if (target.off + seginfo.vsize >= limit)
955 * Allocate memory for building the header, fill it up,
958 target.off_max = target.off;
960 target.buf = malloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO);
962 if (target.buf == NULL)
964 error = elf_corehdr(p, fp, cred, seginfo.count, &target);
966 /* Write the contents of all of the writable segments. */
972 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1;
973 for (i = 0; i < seginfo.count; i++) {
974 error = fp_write(fp, (caddr_t)php->p_vaddr,
975 php->p_filesz, &nbytes);
981 free(target.buf, M_TEMP);
987 * A callback for each_segment() to write out the segment's
988 * program header entry.
991 cb_put_phdr(vm_map_entry_t entry, void *closure)
993 struct phdr_closure *phc = closure;
994 Elf_Phdr *phdr = phc->phdr;
996 if (phc->phdr == phc->phdr_max)
999 phc->offset = round_page(phc->offset);
1001 phdr->p_type = PT_LOAD;
1002 phdr->p_offset = phc->offset;
1003 phdr->p_vaddr = entry->start;
1005 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1006 phdr->p_align = PAGE_SIZE;
1008 if (entry->protection & VM_PROT_READ)
1009 phdr->p_flags |= PF_R;
1010 if (entry->protection & VM_PROT_WRITE)
1011 phdr->p_flags |= PF_W;
1012 if (entry->protection & VM_PROT_EXECUTE)
1013 phdr->p_flags |= PF_X;
1015 phc->offset += phdr->p_filesz;
1021 * A callback for each_writable_segment() to gather information about
1022 * the number of segments and their total size.
1025 cb_size_segment(vm_map_entry_t entry, void *closure)
1027 struct sseg_closure *ssc = closure;
1030 ssc->vsize += entry->end - entry->start;
1035 * A callback for each_segment() to gather information about
1036 * the number of text segments.
1039 cb_fpcount_segment(vm_map_entry_t entry, void *closure)
1041 int *count = closure;
1044 if (entry->object.vm_object->type == OBJT_VNODE) {
1045 vp = (struct vnode *)entry->object.vm_object->handle;
1046 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1054 cb_put_fp(vm_map_entry_t entry, void *closure)
1056 struct fp_closure *fpc = closure;
1057 struct vn_hdr *vnh = fpc->vnh;
1058 Elf_Phdr *phdr = &vnh->vnh_phdr;
1063 * If an entry represents a vnode then write out a file handle.
1065 * If we are checkpointing a checkpoint-restored program we do
1066 * NOT record the filehandle for the old checkpoint vnode (which
1067 * is mapped all over the place). Instead we rely on the fact
1068 * that a checkpoint-restored program does not mmap() the checkpt
1069 * vnode NOCORE, so its contents will be written out to the
1070 * new checkpoint file. This is necessary because the 'old'
1071 * checkpoint file is typically destroyed when a new one is created
1072 * and thus cannot be used to restore the new checkpoint.
1074 * Theoretically we could create a chain of checkpoint files and
1075 * operate the checkpointing operation kinda like an incremental
1076 * checkpoint, but a checkpoint restore would then likely wind up
1077 * referencing many prior checkpoint files and that is a bit over
1078 * the top for the purpose of the checkpoint API.
1080 if (entry->object.vm_object->type == OBJT_VNODE) {
1081 vp = (struct vnode *)entry->object.vm_object->handle;
1082 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1084 if (vnh == fpc->vnh_max)
1088 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1089 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid);
1091 char *freepath, *fullpath;
1093 if (vn_fullpath(curproc, vp, &fullpath, &freepath)) {
1094 printf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp);
1096 printf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath);
1097 free(freepath, M_TEMP);
1102 phdr->p_type = PT_LOAD;
1103 phdr->p_offset = 0; /* not written to core */
1104 phdr->p_vaddr = entry->start;
1106 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1107 phdr->p_align = PAGE_SIZE;
1109 if (entry->protection & VM_PROT_READ)
1110 phdr->p_flags |= PF_R;
1111 if (entry->protection & VM_PROT_WRITE)
1112 phdr->p_flags |= PF_W;
1113 if (entry->protection & VM_PROT_EXECUTE)
1114 phdr->p_flags |= PF_X;
1122 * For each writable segment in the process's memory map, call the given
1123 * function with a pointer to the map entry and some arbitrary
1124 * caller-supplied data.
1127 each_segment(struct proc *p, segment_callback func, void *closure, int writable)
1130 vm_map_t map = &p->p_vmspace->vm_map;
1131 vm_map_entry_t entry;
1133 for (entry = map->header.next; error == 0 && entry != &map->header;
1134 entry = entry->next) {
1138 * Don't dump inaccessible mappings, deal with legacy
1141 * Note that read-only segments related to the elf binary
1142 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer
1143 * need to arbitrarily ignore such segments.
1145 if (elf_legacy_coredump) {
1146 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW)
1149 if (writable && (entry->protection & VM_PROT_ALL) == 0)
1154 * Dont include memory segment in the coredump if
1155 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in
1156 * madvise(2). Do not dump submaps (i.e. parts of the
1159 if (writable && entry->eflags & (MAP_ENTRY_NOCOREDUMP|MAP_ENTRY_IS_SUB_MAP))
1162 if ((obj = entry->object.vm_object) == NULL)
1165 /* Find the deepest backing object. */
1166 while (obj->backing_object != NULL)
1167 obj = obj->backing_object;
1169 /* Ignore memory-mapped devices and such things. */
1170 if (obj->type != OBJT_DEFAULT &&
1171 obj->type != OBJT_SWAP &&
1172 obj->type != OBJT_VNODE)
1175 error = (*func)(entry, closure);
1182 target_reserve(elf_buf_t target, size_t bytes, int *error)
1187 if (target->off + bytes > target->off_max)
1190 res = target->buf + target->off;
1192 target->off += bytes;
1197 * Write the core file header to the file, including padding up to
1198 * the page boundary.
1201 elf_corehdr(struct proc *p, struct file *fp, struct ucred *cred, int numsegs,
1206 prfpregset_t fpregset;
1211 prfpregset_t *fpregset;
1214 tempdata = malloc(sizeof(*tempdata), M_TEMP, M_ZERO | M_WAITOK);
1215 status = &tempdata->status;
1216 fpregset = &tempdata->fpregset;
1217 psinfo = &tempdata->psinfo;
1219 /* Gather the information for the header. */
1220 status->pr_version = PRSTATUS_VERSION;
1221 status->pr_statussz = sizeof(prstatus_t);
1222 status->pr_gregsetsz = sizeof(gregset_t);
1223 status->pr_fpregsetsz = sizeof(fpregset_t);
1224 status->pr_osreldate = osreldate;
1225 status->pr_cursig = p->p_sig;
1226 status->pr_pid = p->p_pid;
1227 fill_regs(&p->p_lwp, &status->pr_reg);
1229 fill_fpregs(&p->p_lwp, fpregset);
1231 psinfo->pr_version = PRPSINFO_VERSION;
1232 psinfo->pr_psinfosz = sizeof(prpsinfo_t);
1233 strncpy(psinfo->pr_fname, p->p_comm, sizeof(psinfo->pr_fname) - 1);
1235 /* XXX - We don't fill in the command line arguments properly yet. */
1236 strncpy(psinfo->pr_psargs, p->p_comm, PRARGSZ);
1238 /* Fill in the header. */
1239 error = elf_puthdr(p, target, status, fpregset, psinfo, numsegs);
1241 free(tempdata, M_TEMP);
1243 /* Write it to the core file. */
1245 error = fp_write(fp, target->buf, target->off, &nbytes);
1250 elf_puthdr(struct proc *p, elf_buf_t target, const prstatus_t *status,
1251 const prfpregset_t *fpregset, const prpsinfo_t *psinfo, int numsegs)
1260 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error);
1262 phoff = target->off;
1263 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error);
1265 noteoff = target->off;
1267 error = elf_putnote(target, "FreeBSD", NT_PRSTATUS,
1268 status, sizeof *status);
1271 error = elf_putnote(target, "FreeBSD", NT_FPREGSET,
1272 fpregset, sizeof *fpregset);
1275 error = elf_putnote(target, "FreeBSD", NT_PRPSINFO,
1276 psinfo, sizeof *psinfo);
1278 notesz = target->off - noteoff;
1281 * put extra cruft for dumping process state here
1282 * - we really want it be before all the program
1284 * - we just need to update the offset accordingly
1285 * and GDB will be none the wiser.
1288 error = elf_puttextvp(p, target);
1290 error = elf_putsigs(p, target);
1292 error = elf_putfiles(p, target);
1295 * Align up to a page boundary for the program segments. The
1296 * actual data will be written to the outptu file, not to elf_buf_t,
1297 * so we do not have to do any further bounds checking.
1299 target->off = round_page(target->off);
1300 if (error == 0 && ehdr != NULL) {
1302 * Fill in the ELF header.
1304 ehdr->e_ident[EI_MAG0] = ELFMAG0;
1305 ehdr->e_ident[EI_MAG1] = ELFMAG1;
1306 ehdr->e_ident[EI_MAG2] = ELFMAG2;
1307 ehdr->e_ident[EI_MAG3] = ELFMAG3;
1308 ehdr->e_ident[EI_CLASS] = ELF_CLASS;
1309 ehdr->e_ident[EI_DATA] = ELF_DATA;
1310 ehdr->e_ident[EI_VERSION] = EV_CURRENT;
1311 ehdr->e_ident[EI_OSABI] = ELFOSABI_FREEBSD;
1312 ehdr->e_ident[EI_ABIVERSION] = 0;
1313 ehdr->e_ident[EI_PAD] = 0;
1314 ehdr->e_type = ET_CORE;
1315 ehdr->e_machine = ELF_ARCH;
1316 ehdr->e_version = EV_CURRENT;
1318 ehdr->e_phoff = phoff;
1320 ehdr->e_ehsize = sizeof(Elf_Ehdr);
1321 ehdr->e_phentsize = sizeof(Elf_Phdr);
1322 ehdr->e_phnum = numsegs + 1;
1323 ehdr->e_shentsize = sizeof(Elf_Shdr);
1325 ehdr->e_shstrndx = SHN_UNDEF;
1327 if (error == 0 && phdr != NULL) {
1329 * Fill in the program header entries.
1331 struct phdr_closure phc;
1333 /* The note segement. */
1334 phdr->p_type = PT_NOTE;
1335 phdr->p_offset = noteoff;
1338 phdr->p_filesz = notesz;
1344 /* All the writable segments from the program. */
1346 phc.phdr_max = phdr + numsegs;
1347 phc.offset = target->off;
1348 each_segment(p, cb_put_phdr, &phc, 1);
1354 elf_putnote(elf_buf_t target, const char *name, int type,
1355 const void *desc, size_t descsz)
1361 note.n_namesz = strlen(name) + 1;
1362 note.n_descsz = descsz;
1364 dst = target_reserve(target, sizeof(note), &error);
1366 bcopy(¬e, dst, sizeof note);
1367 dst = target_reserve(target, note.n_namesz, &error);
1369 bcopy(name, dst, note.n_namesz);
1370 target->off = roundup2(target->off, sizeof(Elf_Size));
1371 dst = target_reserve(target, note.n_descsz, &error);
1373 bcopy(desc, dst, note.n_descsz);
1374 target->off = roundup2(target->off, sizeof(Elf_Size));
1380 elf_putsigs(struct proc *p, elf_buf_t target)
1383 struct ckpt_siginfo *csi;
1385 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error);
1387 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo);
1388 bcopy(p->p_procsig, &csi->csi_procsig, sizeof(struct procsig));
1389 bcopy(p->p_procsig->ps_sigacts, &csi->csi_sigacts, sizeof(struct sigacts));
1390 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval));
1391 bcopy(&p->p_sigmask, &csi->csi_sigmask, sizeof(sigset_t));
1392 csi->csi_sigparent = p->p_sigparent;
1398 elf_putfiles(struct proc *p, elf_buf_t target)
1402 struct ckpt_filehdr *cfh = NULL;
1403 struct ckpt_fileinfo *cfi;
1407 * the duplicated loop is gross, but it was the only way
1408 * to eliminate uninitialized variable warnings
1410 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error);
1412 cfh->cfh_nfiles = 0;
1416 * ignore STDIN/STDERR/STDOUT.
1418 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) {
1419 if ((fp = p->p_fd->fd_files[i].fp) == NULL)
1422 * XXX Only checkpoint vnodes for now.
1424 if (fp->f_type != DTYPE_VNODE)
1426 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo),
1430 cfi->cfi_index = -1;
1431 cfi->cfi_type = fp->f_type;
1432 cfi->cfi_flags = fp->f_flag;
1433 cfi->cfi_offset = fp->f_offset;
1434 /* f_count and f_msgcount should not be saved/restored */
1435 /* XXX save cred info */
1437 switch(fp->f_type) {
1439 vp = (struct vnode *)fp->f_data;
1441 * it looks like a bug in ptrace is marking
1442 * a non-vnode as a vnode - until we find the
1443 * root cause this will at least prevent
1444 * further panics from truss
1446 if (vp == NULL || vp->v_mount == NULL)
1450 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1451 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid);
1461 elf_puttextvp(struct proc *p, elf_buf_t target)
1465 struct fp_closure fpc;
1466 struct ckpt_vminfo *vminfo;
1468 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error);
1469 if (vminfo != NULL) {
1470 vminfo->cvm_dsize = p->p_vmspace->vm_dsize;
1471 vminfo->cvm_tsize = p->p_vmspace->vm_tsize;
1472 vminfo->cvm_daddr = p->p_vmspace->vm_daddr;
1473 vminfo->cvm_taddr = p->p_vmspace->vm_taddr;
1477 vn_count = target_reserve(target, sizeof(int), &error);
1478 if (target->buf != NULL) {
1479 fpc.vnh = (struct vn_hdr *)(target->buf + target->off);
1480 fpc.vnh_max = fpc.vnh +
1481 (target->off_max - target->off) / sizeof(struct vn_hdr);
1482 error = each_segment(p, cb_put_fp, &fpc, 0);
1484 *vn_count = fpc.count;
1486 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0);
1488 target->off += fpc.count * sizeof(struct vn_hdr);
1494 * Tell kern_execve.c about it, with a little help from the linker.
1496 static struct execsw elf_execsw = {exec_elf_imgact, "ELF"};
1497 EXEC_SET_ORDERED(elf, elf_execsw, SI_ORDER_FIRST);