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 without 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 1.55 2008/08/17 17:21:36 nth 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>
54 #include <sys/eventhandler.h>
56 #include <cpu/lwbuf.h>
59 #include <vm/vm_kern.h>
60 #include <vm/vm_param.h>
63 #include <vm/vm_map.h>
64 #include <vm/vm_object.h>
65 #include <vm/vm_extern.h>
67 #include <machine/elf.h>
68 #include <machine/md_var.h>
69 #include <sys/mount.h>
71 #define OLD_EI_BRAND 8
76 static int elf_check_header (const Elf_Ehdr *hdr);
77 static int elf_freebsd_fixup (register_t **stack_base,
78 struct image_params *imgp);
79 static int elf_load_file (struct proc *p, const char *file, u_long *addr,
81 static int elf_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,
85 static int exec_elf_imgact (struct image_params *imgp);
87 static int elf_trace = 0;
88 SYSCTL_INT(_debug, OID_AUTO, elf_trace, CTLFLAG_RW, &elf_trace, 0, "");
89 static int elf_legacy_coredump = 0;
90 SYSCTL_INT(_debug, OID_AUTO, elf_legacy_coredump, CTLFLAG_RW,
91 &elf_legacy_coredump, 0,
92 "Do not dump inaccessible mappings, legacy coredump mode");
94 static int dragonfly_match_abi_note(const Elf_Note *);
95 static int freebsd_match_abi_note(const Elf_Note *);
97 static struct sysentvec elf_freebsd_sysvec = {
111 #if defined(__x86_64__)
121 static Elf_Brandinfo freebsd_brand_info = {
124 freebsd_match_abi_note,
126 "/usr/libexec/ld-elf.so.1",
130 static Elf_Brandinfo dragonfly_brand_info = {
133 dragonfly_match_abi_note,
135 "/usr/libexec/ld-elf.so.2",
139 static Elf_Brandinfo *elf_brand_list[MAX_BRANDS] = {
140 &dragonfly_brand_info,
147 freebsd_match_abi_note(const Elf_Note *abi_note)
149 const char *abi_name = (const char *)
150 ((const uint8_t *)abi_note + sizeof(*abi_note));
152 if (abi_note->n_namesz != sizeof("FreeBSD"))
154 if (memcmp(abi_name, "FreeBSD", sizeof("FreeBSD")))
160 dragonfly_match_abi_note(const Elf_Note *abi_note)
162 const char *abi_name = (const char *)
163 ((const uint8_t *)abi_note + sizeof(*abi_note));
165 if (abi_note->n_namesz != sizeof("DragonFly"))
167 if (memcmp(abi_name, "DragonFly", sizeof("DragonFly")))
173 elf_insert_brand_entry(Elf_Brandinfo *entry)
177 for (i=1; i<MAX_BRANDS; i++) {
178 if (elf_brand_list[i] == NULL) {
179 elf_brand_list[i] = entry;
189 elf_remove_brand_entry(Elf_Brandinfo *entry)
193 for (i=1; i<MAX_BRANDS; i++) {
194 if (elf_brand_list[i] == entry) {
195 elf_brand_list[i] = NULL;
205 * Check if an elf brand is being used anywhere in the system.
207 * Used by the linux emulation module unloader. This isn't safe from
210 struct elf_brand_inuse_info {
212 Elf_Brandinfo *entry;
215 static int elf_brand_inuse_callback(struct proc *p, void *data);
218 elf_brand_inuse(Elf_Brandinfo *entry)
220 struct elf_brand_inuse_info info;
224 allproc_scan(elf_brand_inuse_callback, entry);
230 elf_brand_inuse_callback(struct proc *p, void *data)
232 struct elf_brand_inuse_info *info = data;
234 if (p->p_sysent == info->entry->sysvec) {
242 elf_check_header(const Elf_Ehdr *hdr)
245 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
246 hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
247 hdr->e_ident[EI_VERSION] != EV_CURRENT ||
248 hdr->e_phentsize != sizeof(Elf_Phdr) ||
249 hdr->e_ehsize != sizeof(Elf_Ehdr) ||
250 hdr->e_version != ELF_TARG_VER)
253 if (!ELF_MACHINE_OK(hdr->e_machine))
259 static Elf_Brandinfo *
260 elf_check_abi_note(struct image_params *imgp, const Elf_Phdr *ph)
262 Elf_Brandinfo *match = NULL;
263 const Elf_Note *tmp_note;
275 firstoff = off & PAGE_MASK;
276 firstlen = PAGE_SIZE - firstoff;
278 if (len < sizeof(Elf_Note) || len > PAGE_SIZE)
279 return NULL; /* ENOEXEC? */
281 if (exec_map_page(imgp, off >> PAGE_SHIFT, &lwb, &page))
285 * Crosses page boundary? Is that allowed?
287 if (firstlen < len) {
288 data = kmalloc(len, M_TEMP, M_WAITOK);
290 bcopy(page + firstoff, data, firstlen);
292 exec_unmap_page(lwb);
293 if (exec_map_page(imgp, (off >> PAGE_SHIFT) + 1, &lwb, &page)) {
297 bcopy(page, data + firstlen, len - firstlen);
298 tmp_note = (void *)data;
300 tmp_note = (const void *)(page + firstoff);
303 while (len >= sizeof(Elf_Note)) {
305 size_t nlen = roundup(tmp_note->n_namesz, sizeof(Elf_Word)) +
306 roundup(tmp_note->n_descsz, sizeof(Elf_Word)) +
312 if (tmp_note->n_type != 1)
315 for (i = 0; i < MAX_BRANDS; i++) {
316 Elf_Brandinfo *bi = elf_brand_list[i];
318 if (bi != NULL && bi->match_abi_note != NULL &&
319 bi->match_abi_note(tmp_note)) {
335 exec_unmap_page(lwb);
341 elf_load_section(struct proc *p, struct vmspace *vmspace, struct vnode *vp,
342 vm_offset_t offset, caddr_t vmaddr, size_t memsz,
343 size_t filsz, vm_prot_t prot)
346 vm_offset_t map_addr;
351 vm_offset_t file_addr;
353 object = vp->v_object;
357 * It's necessary to fail if the filsz + offset taken from the
358 * header is greater than the actual file pager object's size.
359 * If we were to allow this, then the vm_map_find() below would
360 * walk right off the end of the file object and into the ether.
362 * While I'm here, might as well check for something else that
363 * is invalid: filsz cannot be greater than memsz.
365 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) {
366 uprintf("elf_load_section: truncated ELF file\n");
370 map_addr = trunc_page((vm_offset_t)vmaddr);
371 file_addr = trunc_page(offset);
374 * We have two choices. We can either clear the data in the last page
375 * of an oversized mapping, or we can start the anon mapping a page
376 * early and copy the initialized data into that first page. We
377 * choose the second..
380 map_len = trunc_page(offset+filsz) - file_addr;
382 map_len = round_page(offset+filsz) - file_addr;
385 vm_object_reference(object);
387 /* cow flags: don't dump readonly sections in core */
388 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT |
389 (prot & VM_PROT_WRITE ? 0 : MAP_DISABLE_COREDUMP);
391 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
392 vm_map_lock(&vmspace->vm_map);
393 rv = vm_map_insert(&vmspace->vm_map, &count,
395 file_addr, /* file offset */
396 map_addr, /* virtual start */
397 map_addr + map_len,/* virtual end */
401 vm_map_unlock(&vmspace->vm_map);
402 vm_map_entry_release(count);
403 if (rv != KERN_SUCCESS) {
404 vm_object_deallocate(object);
408 /* we can stop now if we've covered it all */
409 if (memsz == filsz) {
416 * We have to get the remaining bit of the file into the first part
417 * of the oversized map segment. This is normally because the .data
418 * segment in the file is extended to provide bss. It's a neat idea
419 * to try and save a page, but it's a pain in the behind to implement.
421 copy_len = (offset + filsz) - trunc_page(offset + filsz);
422 map_addr = trunc_page((vm_offset_t)vmaddr + filsz);
423 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr;
425 /* This had damn well better be true! */
427 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
428 vm_map_lock(&vmspace->vm_map);
429 rv = vm_map_insert(&vmspace->vm_map, &count,
431 map_addr, map_addr + map_len,
433 VM_PROT_ALL, VM_PROT_ALL,
435 vm_map_unlock(&vmspace->vm_map);
436 vm_map_entry_release(count);
437 if (rv != KERN_SUCCESS) {
446 m = vm_fault_object_page(object, trunc_page(offset + filsz),
447 VM_PROT_READ, 0, &error);
449 lwb = lwbuf_alloc(m);
450 error = copyout((caddr_t)lwbuf_kva(lwb),
451 (caddr_t)map_addr, copy_len);
461 * set it to the specified protection
463 vm_map_protect(&vmspace->vm_map, map_addr, map_addr + map_len, prot,
470 * Load the file "file" into memory. It may be either a shared object
473 * The "addr" reference parameter is in/out. On entry, it specifies
474 * the address where a shared object should be loaded. If the file is
475 * an executable, this value is ignored. On exit, "addr" specifies
476 * where the file was actually loaded.
478 * The "entry" reference parameter is out only. On exit, it specifies
479 * the entry point for the loaded file.
482 elf_load_file(struct proc *p, const char *file, u_long *addr, u_long *entry)
485 struct nlookupdata nd;
487 struct image_params image_params;
489 const Elf_Ehdr *hdr = NULL;
490 const Elf_Phdr *phdr = NULL;
491 struct nlookupdata *nd;
492 struct vmspace *vmspace = p->p_vmspace;
494 struct image_params *imgp;
495 struct mount *topmnt;
498 u_long base_addr = 0;
499 int error, i, numsegs;
501 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_WAITOK);
503 attr = &tempdata->attr;
504 imgp = &tempdata->image_params;
507 * Initialize part of the common data
511 imgp->firstpage = NULL;
512 imgp->image_header = NULL;
515 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW);
519 error = cache_vget(&nd->nl_nch, nd->nl_cred, LK_EXCLUSIVE, &imgp->vp);
520 topmnt = nd->nl_nch.mount;
526 * Check permissions, modes, uid, etc on the file, and "open" it.
528 error = exec_check_permissions(imgp, topmnt);
534 error = exec_map_first_page(imgp);
536 * Also make certain that the interpreter stays the same, so set
537 * its VTEXT flag, too.
540 vsetflags(imgp->vp, VTEXT);
545 hdr = (const Elf_Ehdr *)imgp->image_header;
546 if ((error = elf_check_header(hdr)) != 0)
548 if (hdr->e_type == ET_DYN)
550 else if (hdr->e_type == ET_EXEC)
557 /* Only support headers that fit within first page for now
558 * (multiplication of two Elf_Half fields will not overflow) */
559 if ((hdr->e_phoff > PAGE_SIZE) ||
560 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) {
565 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
567 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) {
568 if (phdr[i].p_type == PT_LOAD) { /* Loadable segment */
570 if (phdr[i].p_flags & PF_X)
571 prot |= VM_PROT_EXECUTE;
572 if (phdr[i].p_flags & PF_W)
573 prot |= VM_PROT_WRITE;
574 if (phdr[i].p_flags & PF_R)
575 prot |= VM_PROT_READ;
577 error = elf_load_section(
578 p, vmspace, imgp->vp,
580 (caddr_t)phdr[i].p_vaddr +
583 phdr[i].p_filesz, prot);
587 * Establish the base address if this is the
591 base_addr = trunc_page(phdr[i].p_vaddr + rbase);
596 *entry=(unsigned long)hdr->e_entry + rbase;
600 exec_unmap_first_page(imgp);
605 kfree(tempdata, M_TEMP);
611 * non static, as it can be overridden by start_init()
613 int fallback_elf_brand = -1;
614 SYSCTL_INT(_kern, OID_AUTO, fallback_elf_brand, CTLFLAG_RW,
615 &fallback_elf_brand, -1,
616 "ELF brand of last resort");
618 static int can_exec_dyn = 1;
619 SYSCTL_INT(_kern, OID_AUTO, elf_exec_dyn, CTLFLAG_RW,
621 "ELF: can exec shared libraries");
624 exec_elf_imgact(struct image_params *imgp)
626 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header;
627 const Elf_Phdr *phdr;
628 Elf_Auxargs *elf_auxargs = NULL;
629 struct vmspace *vmspace;
631 u_long text_size = 0, data_size = 0, total_size = 0;
632 u_long text_addr = 0, data_addr = 0;
633 u_long seg_size, seg_addr;
634 u_long addr, entry = 0, proghdr = 0;
636 const char *interp = NULL;
637 const Elf_Note *abi_note = NULL;
638 Elf_Brandinfo *brand_info = NULL;
644 * Do we have a valid ELF header ?
645 * We allow execution of ET_EXEC and, if kern.elf_exec_dyn is 1, ET_DYN.
647 if (elf_check_header(hdr) != 0 ||
648 (hdr->e_type != ET_EXEC && (!can_exec_dyn || hdr->e_type != ET_DYN)))
652 * From here on down, we return an errno, not -1, as we've
653 * detected an ELF file.
656 if ((hdr->e_phoff > PAGE_SIZE) ||
657 (hdr->e_phoff + hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE) {
658 /* Only support headers in first page for now */
661 phdr = (const Elf_Phdr*)(imgp->image_header + hdr->e_phoff);
664 * From this point on, we may have resources that need to be freed.
667 exec_new_vmspace(imgp, NULL);
670 * Yeah, I'm paranoid. There is every reason in the world to get
671 * VTEXT now since from here on out, there are places we can have
672 * a context switch. Better safe than sorry; I really don't want
673 * the file to change while it's being loaded.
675 vsetflags(imgp->vp, VTEXT);
677 vmspace = imgp->proc->p_vmspace;
679 for (i = 0; i < hdr->e_phnum; i++) {
680 switch(phdr[i].p_type) {
682 case PT_LOAD: /* Loadable segment */
684 if (phdr[i].p_flags & PF_X)
685 prot |= VM_PROT_EXECUTE;
686 if (phdr[i].p_flags & PF_W)
687 prot |= VM_PROT_WRITE;
688 if (phdr[i].p_flags & PF_R)
689 prot |= VM_PROT_READ;
691 if ((error = elf_load_section(imgp->proc,
694 (caddr_t)phdr[i].p_vaddr,
696 phdr[i].p_filesz, prot)) != 0)
700 * If this segment contains the program headers,
701 * remember their virtual address for the AT_PHDR
702 * aux entry. Static binaries don't usually include
705 if (phdr[i].p_offset == 0 &&
706 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize
708 proghdr = phdr[i].p_vaddr + hdr->e_phoff;
710 seg_addr = trunc_page(phdr[i].p_vaddr);
711 seg_size = round_page(phdr[i].p_memsz +
712 phdr[i].p_vaddr - seg_addr);
715 * Is this .text or .data? We can't use
716 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the
717 * alpha terribly and possibly does other bad
718 * things so we stick to the old way of figuring
719 * it out: If the segment contains the program
720 * entry point, it's a text segment, otherwise it
723 * Note that obreak() assumes that data_addr +
724 * data_size == end of data load area, and the ELF
725 * file format expects segments to be sorted by
726 * address. If multiple data segments exist, the
727 * last one will be used.
729 if (hdr->e_entry >= phdr[i].p_vaddr &&
730 hdr->e_entry < (phdr[i].p_vaddr +
732 text_size = seg_size;
733 text_addr = seg_addr;
734 entry = (u_long)hdr->e_entry;
736 data_size = seg_size;
737 data_addr = seg_addr;
739 total_size += seg_size;
742 * Check limits. It should be safe to check the
743 * limits after loading the segment since we do
744 * not actually fault in all the segment's pages.
747 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur ||
748 text_size > maxtsiz ||
750 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
755 case PT_INTERP: /* Path to interpreter */
756 if (phdr[i].p_filesz > MAXPATHLEN ||
757 phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE) {
761 interp = imgp->image_header + phdr[i].p_offset;
763 case PT_NOTE: /* Check for .note.ABI-tag */
764 if (brand_info == NULL)
765 brand_info = elf_check_abi_note(imgp, &phdr[i]);
767 case PT_PHDR: /* Program header table info */
768 proghdr = phdr[i].p_vaddr;
775 vmspace->vm_tsize = text_size >> PAGE_SHIFT;
776 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr;
777 vmspace->vm_dsize = data_size >> PAGE_SHIFT;
778 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr;
780 addr = ELF_RTLD_ADDR(vmspace);
782 imgp->entry_addr = entry;
784 /* We support three types of branding -- (1) the ELF EI_OSABI field
785 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string
786 * branding w/in the ELF header, and (3) path of the `interp_path'
787 * field. We should also look for an ".note.ABI-tag" ELF section now
788 * in all Linux ELF binaries, FreeBSD 4.1+, and some NetBSD ones.
791 /* If the executable has a brand, search for it in the brand list. */
792 if (brand_info == NULL && hdr->e_ident[EI_OSABI] != ELFOSABI_NONE) {
793 for (i = 0; i < MAX_BRANDS; i++) {
794 Elf_Brandinfo *bi = elf_brand_list[i];
797 (hdr->e_ident[EI_OSABI] == bi->brand
799 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND],
800 bi->compat_3_brand, strlen(bi->compat_3_brand)))) {
807 /* Search for a recognized ABI. */
808 if (brand_info == NULL && abi_note != NULL) {
812 * ELFOSABI_NONE == ELFOSABI_SYSV, so a SYSV binary misses all
813 * checks so far, since it is neither branded nor does it have
814 * an ABI note. If the EI_OSABI field is ELFOSABI_NONE, assume
815 * it is svr4 and look for an entry in the elf_brand_list with
816 * match_abi_note == NULL.
818 if (brand_info == NULL && hdr->e_ident[EI_OSABI] == ELFOSABI_NONE) {
819 for (i = 0; i < MAX_BRANDS; i++) {
820 Elf_Brandinfo *bi = elf_brand_list[i];
822 if (bi != NULL && bi->match_abi_note == NULL &&
823 ELFOSABI_SYSV == bi->brand) {
830 /* Lacking a recognized ABI, search for a recognized interpreter. */
831 if (brand_info == NULL && interp != NULL) {
832 for (i = 0; i < MAX_BRANDS; i++) {
833 Elf_Brandinfo *bi = elf_brand_list[i];
836 strcmp(interp, bi->interp_path) == 0) {
843 /* Lacking a recognized interpreter, try the default brand */
844 if (brand_info == NULL) {
845 for (i = 0; i < MAX_BRANDS; i++) {
846 Elf_Brandinfo *bi = elf_brand_list[i];
848 if (bi != NULL && fallback_elf_brand == bi->brand) {
855 if (brand_info == NULL) {
856 uprintf("ELF binary type \"%u\" not known.\n",
857 hdr->e_ident[EI_OSABI]);
862 imgp->proc->p_sysent = brand_info->sysvec;
863 EVENTHANDLER_INVOKE(process_exec, imgp);
865 if (interp != NULL) {
866 path = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
867 ksnprintf(path, MAXPATHLEN, "%s%s",
868 brand_info->emul_path, interp);
869 if ((error = elf_load_file(imgp->proc, path, &addr,
870 &imgp->entry_addr)) != 0) {
871 if ((error = elf_load_file(imgp->proc, interp, &addr,
872 &imgp->entry_addr)) != 0) {
873 uprintf("ELF interpreter %s not found\n", path);
884 * Construct auxargs table (used by the fixup routine)
886 elf_auxargs = kmalloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK);
887 elf_auxargs->execfd = -1;
888 elf_auxargs->phdr = proghdr;
889 elf_auxargs->phent = hdr->e_phentsize;
890 elf_auxargs->phnum = hdr->e_phnum;
891 elf_auxargs->pagesz = PAGE_SIZE;
892 elf_auxargs->base = addr;
893 elf_auxargs->flags = 0;
894 elf_auxargs->entry = entry;
895 elf_auxargs->trace = elf_trace;
897 imgp->auxargs = elf_auxargs;
898 imgp->interpreted = 0;
905 elf_freebsd_fixup(register_t **stack_base, struct image_params *imgp)
907 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs;
910 pos = *stack_base + (imgp->args->argc + imgp->args->envc + 2);
913 AUXARGS_ENTRY(pos, AT_DEBUG, 1);
915 if (args->execfd != -1) {
916 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
918 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
919 AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
920 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
921 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
922 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
923 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
924 AUXARGS_ENTRY(pos, AT_BASE, args->base);
925 AUXARGS_ENTRY(pos, AT_NULL, 0);
927 kfree(imgp->auxargs, M_TEMP);
928 imgp->auxargs = NULL;
931 suword(*stack_base, (long) imgp->args->argc);
936 * Code for generating ELF core dumps.
939 typedef int (*segment_callback) (vm_map_entry_t, void *);
941 /* Closure for cb_put_phdr(). */
942 struct phdr_closure {
943 Elf_Phdr *phdr; /* Program header to fill in (incremented) */
944 Elf_Phdr *phdr_max; /* Pointer bound for error check */
945 Elf_Off offset; /* Offset of segment in core file */
948 /* Closure for cb_size_segment(). */
949 struct sseg_closure {
950 int count; /* Count of writable segments. */
951 size_t vsize; /* Total size of all writable segments. */
954 /* Closure for cb_put_fp(). */
957 struct vn_hdr *vnh_max;
962 typedef struct elf_buf {
968 static void *target_reserve(elf_buf_t target, size_t bytes, int *error);
970 static int cb_put_phdr (vm_map_entry_t, void *);
971 static int cb_size_segment (vm_map_entry_t, void *);
972 static int cb_fpcount_segment(vm_map_entry_t, void *);
973 static int cb_put_fp(vm_map_entry_t, void *);
976 static int each_segment (struct proc *, segment_callback, void *, int);
977 static int elf_corehdr (struct lwp *, int, struct file *, struct ucred *,
979 enum putmode { WRITE, DRYRUN };
980 static int elf_puthdr (struct lwp *, elf_buf_t, int sig, enum putmode,
982 static int elf_putallnotes(struct lwp *, elf_buf_t, int, enum putmode);
983 static int elf_putnote (elf_buf_t, const char *, int, const void *, size_t);
985 static int elf_putsigs(struct lwp *, elf_buf_t);
986 static int elf_puttextvp(struct proc *, elf_buf_t);
987 static int elf_putfiles(struct proc *, elf_buf_t, struct file *);
989 extern int osreldate;
992 elf_coredump(struct lwp *lp, int sig, struct vnode *vp, off_t limit)
997 if ((error = falloc(NULL, &fp, NULL)) != 0)
999 fsetcred(fp, lp->lwp_proc->p_ucred);
1004 fp->f_type = DTYPE_VNODE;
1005 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW;
1006 fp->f_ops = &vnode_fileops;
1010 error = generic_elf_coredump(lp, sig, fp, limit);
1014 fp->f_ops = &badfileops;
1021 generic_elf_coredump(struct lwp *lp, int sig, struct file *fp, off_t limit)
1023 struct proc *p = lp->lwp_proc;
1024 struct ucred *cred = p->p_ucred;
1026 struct sseg_closure seginfo;
1027 struct elf_buf target;
1030 kprintf("can't dump core - null fp\n");
1033 * Size the program segments
1037 each_segment(p, cb_size_segment, &seginfo, 1);
1040 * Calculate the size of the core file header area by making
1041 * a dry run of generating it. Nothing is written, but the
1042 * size is calculated.
1044 bzero(&target, sizeof(target));
1045 elf_puthdr(lp, &target, sig, DRYRUN, seginfo.count, fp);
1047 if (target.off + seginfo.vsize >= limit)
1051 * Allocate memory for building the header, fill it up,
1054 target.off_max = target.off;
1056 target.buf = kmalloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO);
1058 error = elf_corehdr(lp, sig, fp, cred, seginfo.count, &target);
1060 /* Write the contents of all of the writable segments. */
1066 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1;
1067 for (i = 0; i < seginfo.count; i++) {
1068 error = fp_write(fp, (caddr_t)php->p_vaddr,
1069 php->p_filesz, &nbytes, UIO_USERSPACE);
1075 kfree(target.buf, M_TEMP);
1081 * A callback for each_segment() to write out the segment's
1082 * program header entry.
1085 cb_put_phdr(vm_map_entry_t entry, void *closure)
1087 struct phdr_closure *phc = closure;
1088 Elf_Phdr *phdr = phc->phdr;
1090 if (phc->phdr == phc->phdr_max)
1093 phc->offset = round_page(phc->offset);
1095 phdr->p_type = PT_LOAD;
1096 phdr->p_offset = phc->offset;
1097 phdr->p_vaddr = entry->start;
1099 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1100 phdr->p_align = PAGE_SIZE;
1102 if (entry->protection & VM_PROT_READ)
1103 phdr->p_flags |= PF_R;
1104 if (entry->protection & VM_PROT_WRITE)
1105 phdr->p_flags |= PF_W;
1106 if (entry->protection & VM_PROT_EXECUTE)
1107 phdr->p_flags |= PF_X;
1109 phc->offset += phdr->p_filesz;
1115 * A callback for each_writable_segment() to gather information about
1116 * the number of segments and their total size.
1119 cb_size_segment(vm_map_entry_t entry, void *closure)
1121 struct sseg_closure *ssc = closure;
1124 ssc->vsize += entry->end - entry->start;
1129 * A callback for each_segment() to gather information about
1130 * the number of text segments.
1133 cb_fpcount_segment(vm_map_entry_t entry, void *closure)
1135 int *count = closure;
1138 if (entry->object.vm_object->type == OBJT_VNODE) {
1139 vp = (struct vnode *)entry->object.vm_object->handle;
1140 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1148 cb_put_fp(vm_map_entry_t entry, void *closure)
1150 struct fp_closure *fpc = closure;
1151 struct vn_hdr *vnh = fpc->vnh;
1152 Elf_Phdr *phdr = &vnh->vnh_phdr;
1157 * If an entry represents a vnode then write out a file handle.
1159 * If we are checkpointing a checkpoint-restored program we do
1160 * NOT record the filehandle for the old checkpoint vnode (which
1161 * is mapped all over the place). Instead we rely on the fact
1162 * that a checkpoint-restored program does not mmap() the checkpt
1163 * vnode NOCORE, so its contents will be written out to the
1164 * new checkpoint file. This is necessary because the 'old'
1165 * checkpoint file is typically destroyed when a new one is created
1166 * and thus cannot be used to restore the new checkpoint.
1168 * Theoretically we could create a chain of checkpoint files and
1169 * operate the checkpointing operation kinda like an incremental
1170 * checkpoint, but a checkpoint restore would then likely wind up
1171 * referencing many prior checkpoint files and that is a bit over
1172 * the top for the purpose of the checkpoint API.
1174 if (entry->object.vm_object->type == OBJT_VNODE) {
1175 vp = (struct vnode *)entry->object.vm_object->handle;
1176 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1178 if (vnh == fpc->vnh_max)
1182 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1183 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid);
1185 char *freepath, *fullpath;
1187 if (vn_fullpath(curproc, vp, &fullpath, &freepath, 0)) {
1188 kprintf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp);
1190 kprintf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath);
1191 kfree(freepath, M_TEMP);
1196 phdr->p_type = PT_LOAD;
1197 phdr->p_offset = 0; /* not written to core */
1198 phdr->p_vaddr = entry->start;
1200 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1201 phdr->p_align = PAGE_SIZE;
1203 if (entry->protection & VM_PROT_READ)
1204 phdr->p_flags |= PF_R;
1205 if (entry->protection & VM_PROT_WRITE)
1206 phdr->p_flags |= PF_W;
1207 if (entry->protection & VM_PROT_EXECUTE)
1208 phdr->p_flags |= PF_X;
1216 * For each writable segment in the process's memory map, call the given
1217 * function with a pointer to the map entry and some arbitrary
1218 * caller-supplied data.
1221 each_segment(struct proc *p, segment_callback func, void *closure, int writable)
1224 vm_map_t map = &p->p_vmspace->vm_map;
1225 vm_map_entry_t entry;
1227 for (entry = map->header.next; error == 0 && entry != &map->header;
1228 entry = entry->next) {
1232 * Don't dump inaccessible mappings, deal with legacy
1235 * Note that read-only segments related to the elf binary
1236 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer
1237 * need to arbitrarily ignore such segments.
1239 if (elf_legacy_coredump) {
1240 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW)
1243 if (writable && (entry->protection & VM_PROT_ALL) == 0)
1248 * Dont include memory segment in the coredump if
1249 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in
1252 * Currently we only dump normal VM object maps. We do
1253 * not dump submaps or virtual page tables.
1255 if (writable && (entry->eflags & MAP_ENTRY_NOCOREDUMP))
1257 if (entry->maptype != VM_MAPTYPE_NORMAL)
1259 if ((obj = entry->object.vm_object) == NULL)
1262 /* Find the deepest backing object. */
1263 while (obj->backing_object != NULL)
1264 obj = obj->backing_object;
1266 /* Ignore memory-mapped devices and such things. */
1267 if (obj->type != OBJT_DEFAULT &&
1268 obj->type != OBJT_SWAP &&
1269 obj->type != OBJT_VNODE)
1272 error = (*func)(entry, closure);
1279 target_reserve(elf_buf_t target, size_t bytes, int *error)
1284 if (target->off + bytes > target->off_max)
1287 res = target->buf + target->off;
1289 target->off += bytes;
1294 * Write the core file header to the file, including padding up to
1295 * the page boundary.
1298 elf_corehdr(struct lwp *lp, int sig, struct file *fp, struct ucred *cred,
1299 int numsegs, elf_buf_t target)
1305 * Fill in the header. The fp is passed so we can detect and flag
1306 * a checkpoint file pointer within the core file itself, because
1307 * it may not be restored from the same file handle.
1309 error = elf_puthdr(lp, target, sig, WRITE, numsegs, fp);
1311 /* Write it to the core file. */
1313 error = fp_write(fp, target->buf, target->off, &nbytes,
1320 elf_puthdr(struct lwp *lp, elf_buf_t target, int sig, enum putmode mode,
1321 int numsegs, struct file *fp)
1323 struct proc *p = lp->lwp_proc;
1331 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error);
1333 phoff = target->off;
1334 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error);
1336 noteoff = target->off;
1338 elf_putallnotes(lp, target, sig, mode);
1339 notesz = target->off - noteoff;
1342 * put extra cruft for dumping process state here
1343 * - we really want it be before all the program
1345 * - we just need to update the offset accordingly
1346 * and GDB will be none the wiser.
1349 error = elf_puttextvp(p, target);
1351 error = elf_putsigs(lp, target);
1353 error = elf_putfiles(p, target, fp);
1356 * Align up to a page boundary for the program segments. The
1357 * actual data will be written to the outptu file, not to elf_buf_t,
1358 * so we do not have to do any further bounds checking.
1360 target->off = round_page(target->off);
1361 if (error == 0 && ehdr != NULL) {
1363 * Fill in the ELF header.
1365 ehdr->e_ident[EI_MAG0] = ELFMAG0;
1366 ehdr->e_ident[EI_MAG1] = ELFMAG1;
1367 ehdr->e_ident[EI_MAG2] = ELFMAG2;
1368 ehdr->e_ident[EI_MAG3] = ELFMAG3;
1369 ehdr->e_ident[EI_CLASS] = ELF_CLASS;
1370 ehdr->e_ident[EI_DATA] = ELF_DATA;
1371 ehdr->e_ident[EI_VERSION] = EV_CURRENT;
1372 ehdr->e_ident[EI_OSABI] = ELFOSABI_NONE;
1373 ehdr->e_ident[EI_ABIVERSION] = 0;
1374 ehdr->e_ident[EI_PAD] = 0;
1375 ehdr->e_type = ET_CORE;
1376 ehdr->e_machine = ELF_ARCH;
1377 ehdr->e_version = EV_CURRENT;
1379 ehdr->e_phoff = phoff;
1381 ehdr->e_ehsize = sizeof(Elf_Ehdr);
1382 ehdr->e_phentsize = sizeof(Elf_Phdr);
1383 ehdr->e_phnum = numsegs + 1;
1384 ehdr->e_shentsize = sizeof(Elf_Shdr);
1386 ehdr->e_shstrndx = SHN_UNDEF;
1388 if (error == 0 && phdr != NULL) {
1390 * Fill in the program header entries.
1392 struct phdr_closure phc;
1394 /* The note segement. */
1395 phdr->p_type = PT_NOTE;
1396 phdr->p_offset = noteoff;
1399 phdr->p_filesz = notesz;
1405 /* All the writable segments from the program. */
1407 phc.phdr_max = phdr + numsegs;
1408 phc.offset = target->off;
1409 each_segment(p, cb_put_phdr, &phc, 1);
1415 * Append core dump notes to target ELF buffer or simply update target size
1416 * if dryrun selected.
1419 elf_putallnotes(struct lwp *corelp, elf_buf_t target, int sig,
1422 struct proc *p = corelp->lwp_proc;
1426 prfpregset_t fpregs;
1430 prfpregset_t *fpregs;
1435 * Allocate temporary storage for notes on heap to avoid stack overflow.
1437 if (mode != DRYRUN) {
1438 tmpdata = kmalloc(sizeof(*tmpdata), M_TEMP, M_ZERO | M_WAITOK);
1439 status = &tmpdata->status;
1440 fpregs = &tmpdata->fpregs;
1441 psinfo = &tmpdata->psinfo;
1450 * Append LWP-agnostic note.
1452 if (mode != DRYRUN) {
1453 psinfo->pr_version = PRPSINFO_VERSION;
1454 psinfo->pr_psinfosz = sizeof(prpsinfo_t);
1455 strncpy(psinfo->pr_fname, p->p_comm,
1456 sizeof(psinfo->pr_fname) - 1);
1458 * XXX - We don't fill in the command line arguments
1461 strncpy(psinfo->pr_psargs, p->p_comm, PRARGSZ);
1464 elf_putnote(target, "CORE", NT_PRPSINFO, psinfo, sizeof *psinfo);
1469 * Append first note for LWP that triggered core so that it is
1470 * the selected one when the debugger starts.
1472 if (mode != DRYRUN) {
1473 status->pr_version = PRSTATUS_VERSION;
1474 status->pr_statussz = sizeof(prstatus_t);
1475 status->pr_gregsetsz = sizeof(gregset_t);
1476 status->pr_fpregsetsz = sizeof(fpregset_t);
1477 status->pr_osreldate = osreldate;
1478 status->pr_cursig = sig;
1480 * XXX GDB needs unique pr_pid for each LWP and does not
1481 * not support pr_pid==0 but lwp_tid can be 0, so hack unique
1484 status->pr_pid = corelp->lwp_tid;
1485 fill_regs(corelp, &status->pr_reg);
1486 fill_fpregs(corelp, fpregs);
1489 elf_putnote(target, "CORE", NT_PRSTATUS, status, sizeof *status);
1493 elf_putnote(target, "CORE", NT_FPREGSET, fpregs, sizeof *fpregs);
1498 * Then append notes for other LWPs.
1500 FOREACH_LWP_IN_PROC(lp, p) {
1503 /* skip lwps being created */
1504 if (lp->lwp_thread == NULL)
1506 if (mode != DRYRUN) {
1507 status->pr_pid = lp->lwp_tid;
1508 fill_regs(lp, &status->pr_reg);
1509 fill_fpregs(lp, fpregs);
1511 error = elf_putnote(target, "CORE", NT_PRSTATUS,
1512 status, sizeof *status);
1515 error = elf_putnote(target, "CORE", NT_FPREGSET,
1516 fpregs, sizeof *fpregs);
1522 if (tmpdata != NULL)
1523 kfree(tmpdata, M_TEMP);
1528 * Generate a note sub-structure.
1530 * NOTE: 4-byte alignment.
1533 elf_putnote(elf_buf_t target, const char *name, int type,
1534 const void *desc, size_t descsz)
1540 note.n_namesz = strlen(name) + 1;
1541 note.n_descsz = descsz;
1543 dst = target_reserve(target, sizeof(note), &error);
1545 bcopy(¬e, dst, sizeof note);
1546 dst = target_reserve(target, note.n_namesz, &error);
1548 bcopy(name, dst, note.n_namesz);
1549 target->off = roundup2(target->off, sizeof(Elf_Word));
1550 dst = target_reserve(target, note.n_descsz, &error);
1552 bcopy(desc, dst, note.n_descsz);
1553 target->off = roundup2(target->off, sizeof(Elf_Word));
1559 elf_putsigs(struct lwp *lp, elf_buf_t target)
1561 /* XXX lwp handle more than one lwp */
1562 struct proc *p = lp->lwp_proc;
1564 struct ckpt_siginfo *csi;
1566 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error);
1568 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo);
1569 bcopy(p->p_sigacts, &csi->csi_sigacts, sizeof(*p->p_sigacts));
1570 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval));
1571 bcopy(&lp->lwp_sigmask, &csi->csi_sigmask,
1573 csi->csi_sigparent = p->p_sigparent;
1579 elf_putfiles(struct proc *p, elf_buf_t target, struct file *ckfp)
1583 struct ckpt_filehdr *cfh = NULL;
1584 struct ckpt_fileinfo *cfi;
1588 * the duplicated loop is gross, but it was the only way
1589 * to eliminate uninitialized variable warnings
1591 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error);
1593 cfh->cfh_nfiles = 0;
1597 * ignore STDIN/STDERR/STDOUT.
1599 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) {
1600 fp = holdfp(p->p_fd, i, -1);
1604 * XXX Only checkpoint vnodes for now.
1606 if (fp->f_type != DTYPE_VNODE) {
1610 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo),
1616 cfi->cfi_index = -1;
1617 cfi->cfi_type = fp->f_type;
1618 cfi->cfi_flags = fp->f_flag;
1619 cfi->cfi_offset = fp->f_offset;
1620 cfi->cfi_ckflags = 0;
1623 cfi->cfi_ckflags |= CKFIF_ISCKPTFD;
1624 /* f_count and f_msgcount should not be saved/restored */
1625 /* XXX save cred info */
1627 switch(fp->f_type) {
1629 vp = (struct vnode *)fp->f_data;
1631 * it looks like a bug in ptrace is marking
1632 * a non-vnode as a vnode - until we find the
1633 * root cause this will at least prevent
1634 * further panics from truss
1636 if (vp == NULL || vp->v_mount == NULL)
1640 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1641 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid);
1652 elf_puttextvp(struct proc *p, elf_buf_t target)
1656 struct fp_closure fpc;
1657 struct ckpt_vminfo *vminfo;
1659 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error);
1660 if (vminfo != NULL) {
1661 vminfo->cvm_dsize = p->p_vmspace->vm_dsize;
1662 vminfo->cvm_tsize = p->p_vmspace->vm_tsize;
1663 vminfo->cvm_daddr = p->p_vmspace->vm_daddr;
1664 vminfo->cvm_taddr = p->p_vmspace->vm_taddr;
1668 vn_count = target_reserve(target, sizeof(int), &error);
1669 if (target->buf != NULL) {
1670 fpc.vnh = (struct vn_hdr *)(target->buf + target->off);
1671 fpc.vnh_max = fpc.vnh +
1672 (target->off_max - target->off) / sizeof(struct vn_hdr);
1673 error = each_segment(p, cb_put_fp, &fpc, 0);
1675 *vn_count = fpc.count;
1677 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0);
1679 target->off += fpc.count * sizeof(struct vn_hdr);
1685 * Tell kern_execve.c about it, with a little help from the linker.
1687 static struct execsw elf_execsw = {exec_elf_imgact, "ELF"};
1688 EXEC_SET_ORDERED(elf, elf_execsw, SI_ORDER_FIRST);