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 1.55 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/sfbuf.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_param.h>
61 #include <vm/vm_map.h>
62 #include <vm/vm_object.h>
63 #include <vm/vm_extern.h>
65 #include <machine/elf.h>
66 #include <machine/md_var.h>
67 #include <sys/mount.h>
69 #define OLD_EI_BRAND 8
74 static int elf_check_header (const Elf_Ehdr *hdr);
75 static int elf_freebsd_fixup (register_t **stack_base,
76 struct image_params *imgp);
77 static int elf_load_file (struct proc *p, const char *file, u_long *addr,
79 static int elf_load_section (struct proc *p,
80 struct vmspace *vmspace, struct vnode *vp,
81 vm_offset_t offset, caddr_t vmaddr, size_t memsz, size_t filsz,
83 static int exec_elf_imgact (struct image_params *imgp);
85 static int elf_trace = 0;
86 SYSCTL_INT(_debug, OID_AUTO, elf_trace, CTLFLAG_RW, &elf_trace, 0, "");
87 static int elf_legacy_coredump = 0;
88 SYSCTL_INT(_debug, OID_AUTO, elf_legacy_coredump, CTLFLAG_RW,
89 &elf_legacy_coredump, 0, "");
91 static int dragonfly_match_abi_note(const Elf_Note *);
92 static int freebsd_match_abi_note(const Elf_Note *);
94 static struct sysentvec elf_freebsd_sysvec = {
114 static Elf_Brandinfo freebsd_brand_info = {
117 freebsd_match_abi_note,
119 "/usr/libexec/ld-elf.so.1",
123 static Elf_Brandinfo dragonfly_brand_info = {
126 dragonfly_match_abi_note,
128 "/usr/libexec/ld-elf.so.2",
132 static Elf_Brandinfo *elf_brand_list[MAX_BRANDS] = {
133 &dragonfly_brand_info,
140 freebsd_match_abi_note(const Elf_Note *abi_note)
142 const char *abi_name = (const char *)
143 ((const uint8_t *)abi_note + sizeof(*abi_note));
145 if (abi_note->n_namesz != sizeof("FreeBSD"))
147 if (memcmp(abi_name, "FreeBSD", sizeof("FreeBSD")))
153 dragonfly_match_abi_note(const Elf_Note *abi_note)
155 const char *abi_name = (const char *)
156 ((const uint8_t *)abi_note + sizeof(*abi_note));
158 if (abi_note->n_namesz != sizeof("DragonFly"))
160 if (memcmp(abi_name, "DragonFly", sizeof("DragonFly")))
166 elf_insert_brand_entry(Elf_Brandinfo *entry)
170 for (i=1; i<MAX_BRANDS; i++) {
171 if (elf_brand_list[i] == NULL) {
172 elf_brand_list[i] = entry;
182 elf_remove_brand_entry(Elf_Brandinfo *entry)
186 for (i=1; i<MAX_BRANDS; i++) {
187 if (elf_brand_list[i] == entry) {
188 elf_brand_list[i] = NULL;
198 * Check if an elf brand is being used anywhere in the system.
200 * Used by the linux emulation module unloader. This isn't safe from
203 struct elf_brand_inuse_info {
205 Elf_Brandinfo *entry;
208 static int elf_brand_inuse_callback(struct proc *p, void *data);
211 elf_brand_inuse(Elf_Brandinfo *entry)
213 struct elf_brand_inuse_info info;
217 allproc_scan(elf_brand_inuse_callback, entry);
223 elf_brand_inuse_callback(struct proc *p, void *data)
225 struct elf_brand_inuse_info *info = data;
227 if (p->p_sysent == info->entry->sysvec) {
235 elf_check_header(const Elf_Ehdr *hdr)
238 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
239 hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
240 hdr->e_ident[EI_VERSION] != EV_CURRENT ||
241 hdr->e_phentsize != sizeof(Elf_Phdr) ||
242 hdr->e_ehsize != sizeof(Elf_Ehdr) ||
243 hdr->e_version != ELF_TARG_VER)
246 if (!ELF_MACHINE_OK(hdr->e_machine))
252 static Elf_Brandinfo *
253 elf_check_abi_note(struct image_params *imgp, const Elf_Phdr *ph)
255 Elf_Brandinfo *match = NULL;
256 const Elf_Note *tmp_note;
268 firstoff = off & PAGE_MASK;
269 firstlen = PAGE_SIZE - firstoff;
271 if (len < sizeof(Elf_Note) || len > PAGE_SIZE)
272 return NULL; /* ENOEXEC? */
274 if (exec_map_page(imgp, off >> PAGE_SHIFT, &sfb, &page))
278 * Crosses page boundary? Is that allowed?
280 if (firstlen < len) {
281 data = kmalloc(len, M_TEMP, M_WAITOK);
283 bcopy(page + firstoff, data, firstlen);
285 exec_unmap_page(sfb);
286 if (exec_map_page(imgp, (off >> PAGE_SHIFT) + 1, &sfb, &page)) {
290 bcopy(page, data + firstlen, len - firstlen);
291 tmp_note = (void *)data;
293 tmp_note = (const void *)(page + firstoff);
296 while (len >= sizeof(Elf_Note)) {
298 size_t nlen = roundup(tmp_note->n_namesz, sizeof(Elf_Word)) +
299 roundup(tmp_note->n_descsz, sizeof(Elf_Word)) +
305 if (tmp_note->n_type != 1)
308 for (i = 0; i < MAX_BRANDS; i++) {
309 Elf_Brandinfo *bi = elf_brand_list[i];
311 if (bi != NULL && bi->match_abi_note != NULL &&
312 bi->match_abi_note(tmp_note)) {
328 exec_unmap_page(sfb);
334 elf_load_section(struct proc *p, struct vmspace *vmspace, struct vnode *vp,
335 vm_offset_t offset, caddr_t vmaddr, size_t memsz,
336 size_t filsz, vm_prot_t prot)
339 vm_offset_t map_addr;
344 vm_offset_t file_addr;
346 object = vp->v_object;
350 * It's necessary to fail if the filsz + offset taken from the
351 * header is greater than the actual file pager object's size.
352 * If we were to allow this, then the vm_map_find() below would
353 * walk right off the end of the file object and into the ether.
355 * While I'm here, might as well check for something else that
356 * is invalid: filsz cannot be greater than memsz.
358 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) {
359 uprintf("elf_load_section: truncated ELF file\n");
363 map_addr = trunc_page((vm_offset_t)vmaddr);
364 file_addr = trunc_page(offset);
367 * We have two choices. We can either clear the data in the last page
368 * of an oversized mapping, or we can start the anon mapping a page
369 * early and copy the initialized data into that first page. We
370 * choose the second..
373 map_len = trunc_page(offset+filsz) - file_addr;
375 map_len = round_page(offset+filsz) - file_addr;
378 vm_object_reference(object);
380 /* cow flags: don't dump readonly sections in core */
381 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT |
382 (prot & VM_PROT_WRITE ? 0 : MAP_DISABLE_COREDUMP);
384 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
385 vm_map_lock(&vmspace->vm_map);
386 rv = vm_map_insert(&vmspace->vm_map, &count,
388 file_addr, /* file offset */
389 map_addr, /* virtual start */
390 map_addr + map_len,/* virtual end */
394 vm_map_unlock(&vmspace->vm_map);
395 vm_map_entry_release(count);
396 if (rv != KERN_SUCCESS) {
397 vm_object_deallocate(object);
401 /* we can stop now if we've covered it all */
402 if (memsz == filsz) {
409 * We have to get the remaining bit of the file into the first part
410 * of the oversized map segment. This is normally because the .data
411 * segment in the file is extended to provide bss. It's a neat idea
412 * to try and save a page, but it's a pain in the behind to implement.
414 copy_len = (offset + filsz) - trunc_page(offset + filsz);
415 map_addr = trunc_page((vm_offset_t)vmaddr + filsz);
416 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr;
418 /* This had damn well better be true! */
420 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
421 vm_map_lock(&vmspace->vm_map);
422 rv = vm_map_insert(&vmspace->vm_map, &count,
424 map_addr, map_addr + map_len,
426 VM_PROT_ALL, VM_PROT_ALL,
428 vm_map_unlock(&vmspace->vm_map);
429 vm_map_entry_release(count);
430 if (rv != KERN_SUCCESS) {
439 m = vm_fault_object_page(object, trunc_page(offset + filsz),
440 VM_PROT_READ, 0, &error);
442 sf = sf_buf_alloc(m, SFB_CPUPRIVATE);
443 error = copyout((caddr_t)sf_buf_kva(sf),
444 (caddr_t)map_addr, copy_len);
454 * set it to the specified protection
456 vm_map_protect(&vmspace->vm_map, map_addr, map_addr + map_len, prot,
463 * Load the file "file" into memory. It may be either a shared object
466 * The "addr" reference parameter is in/out. On entry, it specifies
467 * the address where a shared object should be loaded. If the file is
468 * an executable, this value is ignored. On exit, "addr" specifies
469 * where the file was actually loaded.
471 * The "entry" reference parameter is out only. On exit, it specifies
472 * the entry point for the loaded file.
475 elf_load_file(struct proc *p, const char *file, u_long *addr, u_long *entry)
478 struct nlookupdata nd;
480 struct image_params image_params;
482 const Elf_Ehdr *hdr = NULL;
483 const Elf_Phdr *phdr = NULL;
484 struct nlookupdata *nd;
485 struct vmspace *vmspace = p->p_vmspace;
487 struct image_params *imgp;
490 u_long base_addr = 0;
491 int error, i, numsegs;
493 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_WAITOK);
495 attr = &tempdata->attr;
496 imgp = &tempdata->image_params;
499 * Initialize part of the common data
503 imgp->firstpage = NULL;
504 imgp->image_header = NULL;
507 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW);
511 error = cache_vget(&nd->nl_nch, nd->nl_cred, LK_EXCLUSIVE, &imgp->vp);
517 * Check permissions, modes, uid, etc on the file, and "open" it.
519 error = exec_check_permissions(imgp);
525 error = exec_map_first_page(imgp);
527 * Also make certain that the interpreter stays the same, so set
528 * its VTEXT flag, too.
531 vsetflags(imgp->vp, VTEXT);
536 hdr = (const Elf_Ehdr *)imgp->image_header;
537 if ((error = elf_check_header(hdr)) != 0)
539 if (hdr->e_type == ET_DYN)
541 else if (hdr->e_type == ET_EXEC)
548 /* Only support headers that fit within first page for now
549 * (multiplication of two Elf_Half fields will not overflow) */
550 if ((hdr->e_phoff > PAGE_SIZE) ||
551 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) {
556 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
558 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) {
559 if (phdr[i].p_type == PT_LOAD) { /* Loadable segment */
561 if (phdr[i].p_flags & PF_X)
562 prot |= VM_PROT_EXECUTE;
563 if (phdr[i].p_flags & PF_W)
564 prot |= VM_PROT_WRITE;
565 if (phdr[i].p_flags & PF_R)
566 prot |= VM_PROT_READ;
568 error = elf_load_section(
569 p, vmspace, imgp->vp,
571 (caddr_t)phdr[i].p_vaddr +
574 phdr[i].p_filesz, prot);
578 * Establish the base address if this is the
582 base_addr = trunc_page(phdr[i].p_vaddr + rbase);
587 *entry=(unsigned long)hdr->e_entry + rbase;
591 exec_unmap_first_page(imgp);
596 kfree(tempdata, M_TEMP);
602 * non static, as it can be overridden by start_init()
604 int fallback_elf_brand = -1;
605 SYSCTL_INT(_kern, OID_AUTO, fallback_elf_brand, CTLFLAG_RW,
606 &fallback_elf_brand, -1,
607 "ELF brand of last resort");
609 static int can_exec_dyn = 1;
610 SYSCTL_INT(_kern, OID_AUTO, elf_exec_dyn, CTLFLAG_RW,
612 "ELF: can exec shared libraries");
615 exec_elf_imgact(struct image_params *imgp)
617 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header;
618 const Elf_Phdr *phdr;
619 Elf_Auxargs *elf_auxargs = NULL;
620 struct vmspace *vmspace;
622 u_long text_size = 0, data_size = 0, total_size = 0;
623 u_long text_addr = 0, data_addr = 0;
624 u_long seg_size, seg_addr;
625 u_long addr, entry = 0, proghdr = 0;
627 const char *interp = NULL;
628 const Elf_Note *abi_note = NULL;
629 Elf_Brandinfo *brand_info = NULL;
635 * Do we have a valid ELF header ?
636 * We allow execution of ET_EXEC and, if kern.elf_exec_dyn is 1, ET_DYN.
638 if (elf_check_header(hdr) != 0 ||
639 (hdr->e_type != ET_EXEC && (!can_exec_dyn || hdr->e_type != ET_DYN)))
643 * From here on down, we return an errno, not -1, as we've
644 * detected an ELF file.
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 */
652 phdr = (const Elf_Phdr*)(imgp->image_header + hdr->e_phoff);
655 * From this point on, we may have resources that need to be freed.
658 exec_new_vmspace(imgp, NULL);
661 * Yeah, I'm paranoid. There is every reason in the world to get
662 * VTEXT now since from here on out, there are places we can have
663 * a context switch. Better safe than sorry; I really don't want
664 * the file to change while it's being loaded.
666 vsetflags(imgp->vp, VTEXT);
668 vmspace = imgp->proc->p_vmspace;
670 for (i = 0; i < hdr->e_phnum; i++) {
671 switch(phdr[i].p_type) {
673 case PT_LOAD: /* Loadable segment */
675 if (phdr[i].p_flags & PF_X)
676 prot |= VM_PROT_EXECUTE;
677 if (phdr[i].p_flags & PF_W)
678 prot |= VM_PROT_WRITE;
679 if (phdr[i].p_flags & PF_R)
680 prot |= VM_PROT_READ;
682 if ((error = elf_load_section(imgp->proc,
685 (caddr_t)phdr[i].p_vaddr,
687 phdr[i].p_filesz, prot)) != 0)
691 * If this segment contains the program headers,
692 * remember their virtual address for the AT_PHDR
693 * aux entry. Static binaries don't usually include
696 if (phdr[i].p_offset == 0 &&
697 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize
699 proghdr = phdr[i].p_vaddr + hdr->e_phoff;
701 seg_addr = trunc_page(phdr[i].p_vaddr);
702 seg_size = round_page(phdr[i].p_memsz +
703 phdr[i].p_vaddr - seg_addr);
706 * Is this .text or .data? We can't use
707 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the
708 * alpha terribly and possibly does other bad
709 * things so we stick to the old way of figuring
710 * it out: If the segment contains the program
711 * entry point, it's a text segment, otherwise it
714 * Note that obreak() assumes that data_addr +
715 * data_size == end of data load area, and the ELF
716 * file format expects segments to be sorted by
717 * address. If multiple data segments exist, the
718 * last one will be used.
720 if (hdr->e_entry >= phdr[i].p_vaddr &&
721 hdr->e_entry < (phdr[i].p_vaddr +
723 text_size = seg_size;
724 text_addr = seg_addr;
725 entry = (u_long)hdr->e_entry;
727 data_size = seg_size;
728 data_addr = seg_addr;
730 total_size += seg_size;
733 * Check limits. It should be safe to check the
734 * limits after loading the segment since we do
735 * not actually fault in all the segment's pages.
738 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur ||
739 text_size > maxtsiz ||
741 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
746 case PT_INTERP: /* Path to interpreter */
747 if (phdr[i].p_filesz > MAXPATHLEN ||
748 phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE) {
752 interp = imgp->image_header + phdr[i].p_offset;
754 case PT_NOTE: /* Check for .note.ABI-tag */
755 if (brand_info == NULL)
756 brand_info = elf_check_abi_note(imgp, &phdr[i]);
758 case PT_PHDR: /* Program header table info */
759 proghdr = phdr[i].p_vaddr;
766 vmspace->vm_tsize = text_size >> PAGE_SHIFT;
767 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr;
768 vmspace->vm_dsize = data_size >> PAGE_SHIFT;
769 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr;
771 addr = ELF_RTLD_ADDR(vmspace);
773 imgp->entry_addr = entry;
775 /* We support three types of branding -- (1) the ELF EI_OSABI field
776 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string
777 * branding w/in the ELF header, and (3) path of the `interp_path'
778 * field. We should also look for an ".note.ABI-tag" ELF section now
779 * in all Linux ELF binaries, FreeBSD 4.1+, and some NetBSD ones.
782 /* If the executable has a brand, search for it in the brand list. */
783 if (brand_info == NULL && hdr->e_ident[EI_OSABI] != ELFOSABI_NONE) {
784 for (i = 0; i < MAX_BRANDS; i++) {
785 Elf_Brandinfo *bi = elf_brand_list[i];
788 (hdr->e_ident[EI_OSABI] == bi->brand
790 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND],
791 bi->compat_3_brand, strlen(bi->compat_3_brand)))) {
798 /* Search for a recognized ABI. */
799 if (brand_info == NULL && abi_note != NULL) {
803 * ELFOSABI_NONE == ELFOSABI_SYSV, so a SYSV binary misses all
804 * checks so far, since it is neither branded nor does it have
805 * an ABI note. If the EI_OSABI field is ELFOSABI_NONE, assume
806 * it is svr4 and look for an entry in the elf_brand_list with
807 * match_abi_note == NULL.
809 if (brand_info == NULL && hdr->e_ident[EI_OSABI] == ELFOSABI_NONE) {
810 for (i = 0; i < MAX_BRANDS; i++) {
811 Elf_Brandinfo *bi = elf_brand_list[i];
813 if (bi != NULL && bi->match_abi_note == NULL &&
814 ELFOSABI_SYSV == bi->brand) {
821 /* Lacking a recognized ABI, search for a recognized interpreter. */
822 if (brand_info == NULL && interp != NULL) {
823 for (i = 0; i < MAX_BRANDS; i++) {
824 Elf_Brandinfo *bi = elf_brand_list[i];
827 strcmp(interp, bi->interp_path) == 0) {
834 /* Lacking a recognized interpreter, try the default brand */
835 if (brand_info == NULL) {
836 for (i = 0; i < MAX_BRANDS; i++) {
837 Elf_Brandinfo *bi = elf_brand_list[i];
839 if (bi != NULL && fallback_elf_brand == bi->brand) {
846 if (brand_info == NULL) {
847 uprintf("ELF binary type \"%u\" not known.\n",
848 hdr->e_ident[EI_OSABI]);
853 imgp->proc->p_sysent = brand_info->sysvec;
854 if (interp != NULL) {
855 path = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
856 ksnprintf(path, MAXPATHLEN, "%s%s",
857 brand_info->emul_path, interp);
858 if ((error = elf_load_file(imgp->proc, path, &addr,
859 &imgp->entry_addr)) != 0) {
860 if ((error = elf_load_file(imgp->proc, interp, &addr,
861 &imgp->entry_addr)) != 0) {
862 uprintf("ELF interpreter %s not found\n", path);
873 * Construct auxargs table (used by the fixup routine)
875 elf_auxargs = kmalloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK);
876 elf_auxargs->execfd = -1;
877 elf_auxargs->phdr = proghdr;
878 elf_auxargs->phent = hdr->e_phentsize;
879 elf_auxargs->phnum = hdr->e_phnum;
880 elf_auxargs->pagesz = PAGE_SIZE;
881 elf_auxargs->base = addr;
882 elf_auxargs->flags = 0;
883 elf_auxargs->entry = entry;
884 elf_auxargs->trace = elf_trace;
886 imgp->auxargs = elf_auxargs;
887 imgp->interpreted = 0;
894 elf_freebsd_fixup(register_t **stack_base, struct image_params *imgp)
896 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs;
899 pos = *stack_base + (imgp->args->argc + imgp->args->envc + 2);
902 AUXARGS_ENTRY(pos, AT_DEBUG, 1);
904 if (args->execfd != -1) {
905 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
907 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
908 AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
909 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
910 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
911 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
912 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
913 AUXARGS_ENTRY(pos, AT_BASE, args->base);
914 AUXARGS_ENTRY(pos, AT_NULL, 0);
916 kfree(imgp->auxargs, M_TEMP);
917 imgp->auxargs = NULL;
920 suword(*stack_base, (long) imgp->args->argc);
925 * Code for generating ELF core dumps.
928 typedef int (*segment_callback) (vm_map_entry_t, void *);
930 /* Closure for cb_put_phdr(). */
931 struct phdr_closure {
932 Elf_Phdr *phdr; /* Program header to fill in (incremented) */
933 Elf_Phdr *phdr_max; /* Pointer bound for error check */
934 Elf_Off offset; /* Offset of segment in core file */
937 /* Closure for cb_size_segment(). */
938 struct sseg_closure {
939 int count; /* Count of writable segments. */
940 size_t vsize; /* Total size of all writable segments. */
943 /* Closure for cb_put_fp(). */
946 struct vn_hdr *vnh_max;
951 typedef struct elf_buf {
957 static void *target_reserve(elf_buf_t target, size_t bytes, int *error);
959 static int cb_put_phdr (vm_map_entry_t, void *);
960 static int cb_size_segment (vm_map_entry_t, void *);
961 static int cb_fpcount_segment(vm_map_entry_t, void *);
962 static int cb_put_fp(vm_map_entry_t, void *);
965 static int each_segment (struct proc *, segment_callback, void *, int);
966 static int elf_corehdr (struct lwp *, int, struct file *, struct ucred *,
968 enum putmode { WRITE, DRYRUN };
969 static int elf_puthdr (struct lwp *, elf_buf_t, int sig, enum putmode,
971 static int elf_putallnotes(struct lwp *, elf_buf_t, int, enum putmode);
972 static int elf_putnote (elf_buf_t, const char *, int, const void *, size_t);
974 static int elf_putsigs(struct lwp *, elf_buf_t);
975 static int elf_puttextvp(struct proc *, elf_buf_t);
976 static int elf_putfiles(struct proc *, elf_buf_t, struct file *);
978 extern int osreldate;
981 elf_coredump(struct lwp *lp, int sig, struct vnode *vp, off_t limit)
986 if ((error = falloc(NULL, &fp, NULL)) != 0)
988 fsetcred(fp, lp->lwp_proc->p_ucred);
993 fp->f_type = DTYPE_VNODE;
994 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW;
995 fp->f_ops = &vnode_fileops;
999 error = generic_elf_coredump(lp, sig, fp, limit);
1003 fp->f_ops = &badfileops;
1010 generic_elf_coredump(struct lwp *lp, int sig, struct file *fp, off_t limit)
1012 struct proc *p = lp->lwp_proc;
1013 struct ucred *cred = p->p_ucred;
1015 struct sseg_closure seginfo;
1016 struct elf_buf target;
1019 kprintf("can't dump core - null fp\n");
1022 * Size the program segments
1026 each_segment(p, cb_size_segment, &seginfo, 1);
1029 * Calculate the size of the core file header area by making
1030 * a dry run of generating it. Nothing is written, but the
1031 * size is calculated.
1033 bzero(&target, sizeof(target));
1034 elf_puthdr(lp, &target, sig, DRYRUN, seginfo.count, fp);
1036 if (target.off + seginfo.vsize >= limit)
1040 * Allocate memory for building the header, fill it up,
1043 target.off_max = target.off;
1045 target.buf = kmalloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO);
1047 error = elf_corehdr(lp, sig, fp, cred, seginfo.count, &target);
1049 /* Write the contents of all of the writable segments. */
1055 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1;
1056 for (i = 0; i < seginfo.count; i++) {
1057 error = fp_write(fp, (caddr_t)php->p_vaddr,
1058 php->p_filesz, &nbytes, UIO_USERSPACE);
1064 kfree(target.buf, M_TEMP);
1070 * A callback for each_segment() to write out the segment's
1071 * program header entry.
1074 cb_put_phdr(vm_map_entry_t entry, void *closure)
1076 struct phdr_closure *phc = closure;
1077 Elf_Phdr *phdr = phc->phdr;
1079 if (phc->phdr == phc->phdr_max)
1082 phc->offset = round_page(phc->offset);
1084 phdr->p_type = PT_LOAD;
1085 phdr->p_offset = phc->offset;
1086 phdr->p_vaddr = entry->start;
1088 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1089 phdr->p_align = PAGE_SIZE;
1091 if (entry->protection & VM_PROT_READ)
1092 phdr->p_flags |= PF_R;
1093 if (entry->protection & VM_PROT_WRITE)
1094 phdr->p_flags |= PF_W;
1095 if (entry->protection & VM_PROT_EXECUTE)
1096 phdr->p_flags |= PF_X;
1098 phc->offset += phdr->p_filesz;
1104 * A callback for each_writable_segment() to gather information about
1105 * the number of segments and their total size.
1108 cb_size_segment(vm_map_entry_t entry, void *closure)
1110 struct sseg_closure *ssc = closure;
1113 ssc->vsize += entry->end - entry->start;
1118 * A callback for each_segment() to gather information about
1119 * the number of text segments.
1122 cb_fpcount_segment(vm_map_entry_t entry, void *closure)
1124 int *count = closure;
1127 if (entry->object.vm_object->type == OBJT_VNODE) {
1128 vp = (struct vnode *)entry->object.vm_object->handle;
1129 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1137 cb_put_fp(vm_map_entry_t entry, void *closure)
1139 struct fp_closure *fpc = closure;
1140 struct vn_hdr *vnh = fpc->vnh;
1141 Elf_Phdr *phdr = &vnh->vnh_phdr;
1146 * If an entry represents a vnode then write out a file handle.
1148 * If we are checkpointing a checkpoint-restored program we do
1149 * NOT record the filehandle for the old checkpoint vnode (which
1150 * is mapped all over the place). Instead we rely on the fact
1151 * that a checkpoint-restored program does not mmap() the checkpt
1152 * vnode NOCORE, so its contents will be written out to the
1153 * new checkpoint file. This is necessary because the 'old'
1154 * checkpoint file is typically destroyed when a new one is created
1155 * and thus cannot be used to restore the new checkpoint.
1157 * Theoretically we could create a chain of checkpoint files and
1158 * operate the checkpointing operation kinda like an incremental
1159 * checkpoint, but a checkpoint restore would then likely wind up
1160 * referencing many prior checkpoint files and that is a bit over
1161 * the top for the purpose of the checkpoint API.
1163 if (entry->object.vm_object->type == OBJT_VNODE) {
1164 vp = (struct vnode *)entry->object.vm_object->handle;
1165 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1167 if (vnh == fpc->vnh_max)
1171 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1172 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid);
1174 char *freepath, *fullpath;
1176 if (vn_fullpath(curproc, vp, &fullpath, &freepath)) {
1177 kprintf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp);
1179 kprintf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath);
1180 kfree(freepath, M_TEMP);
1185 phdr->p_type = PT_LOAD;
1186 phdr->p_offset = 0; /* not written to core */
1187 phdr->p_vaddr = entry->start;
1189 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1190 phdr->p_align = PAGE_SIZE;
1192 if (entry->protection & VM_PROT_READ)
1193 phdr->p_flags |= PF_R;
1194 if (entry->protection & VM_PROT_WRITE)
1195 phdr->p_flags |= PF_W;
1196 if (entry->protection & VM_PROT_EXECUTE)
1197 phdr->p_flags |= PF_X;
1205 * For each writable segment in the process's memory map, call the given
1206 * function with a pointer to the map entry and some arbitrary
1207 * caller-supplied data.
1210 each_segment(struct proc *p, segment_callback func, void *closure, int writable)
1213 vm_map_t map = &p->p_vmspace->vm_map;
1214 vm_map_entry_t entry;
1216 for (entry = map->header.next; error == 0 && entry != &map->header;
1217 entry = entry->next) {
1221 * Don't dump inaccessible mappings, deal with legacy
1224 * Note that read-only segments related to the elf binary
1225 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer
1226 * need to arbitrarily ignore such segments.
1228 if (elf_legacy_coredump) {
1229 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW)
1232 if (writable && (entry->protection & VM_PROT_ALL) == 0)
1237 * Dont include memory segment in the coredump if
1238 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in
1241 * Currently we only dump normal VM object maps. We do
1242 * not dump submaps or virtual page tables.
1244 if (writable && (entry->eflags & MAP_ENTRY_NOCOREDUMP))
1246 if (entry->maptype != VM_MAPTYPE_NORMAL)
1248 if ((obj = entry->object.vm_object) == NULL)
1251 /* Find the deepest backing object. */
1252 while (obj->backing_object != NULL)
1253 obj = obj->backing_object;
1255 /* Ignore memory-mapped devices and such things. */
1256 if (obj->type != OBJT_DEFAULT &&
1257 obj->type != OBJT_SWAP &&
1258 obj->type != OBJT_VNODE)
1261 error = (*func)(entry, closure);
1268 target_reserve(elf_buf_t target, size_t bytes, int *error)
1273 if (target->off + bytes > target->off_max)
1276 res = target->buf + target->off;
1278 target->off += bytes;
1283 * Write the core file header to the file, including padding up to
1284 * the page boundary.
1287 elf_corehdr(struct lwp *lp, int sig, struct file *fp, struct ucred *cred,
1288 int numsegs, elf_buf_t target)
1294 * Fill in the header. The fp is passed so we can detect and flag
1295 * a checkpoint file pointer within the core file itself, because
1296 * it may not be restored from the same file handle.
1298 error = elf_puthdr(lp, target, sig, WRITE, numsegs, fp);
1300 /* Write it to the core file. */
1302 error = fp_write(fp, target->buf, target->off, &nbytes,
1309 elf_puthdr(struct lwp *lp, elf_buf_t target, int sig, enum putmode mode,
1310 int numsegs, struct file *fp)
1312 struct proc *p = lp->lwp_proc;
1320 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error);
1322 phoff = target->off;
1323 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error);
1325 noteoff = target->off;
1327 elf_putallnotes(lp, target, sig, mode);
1328 notesz = target->off - noteoff;
1331 * put extra cruft for dumping process state here
1332 * - we really want it be before all the program
1334 * - we just need to update the offset accordingly
1335 * and GDB will be none the wiser.
1338 error = elf_puttextvp(p, target);
1340 error = elf_putsigs(lp, target);
1342 error = elf_putfiles(p, target, fp);
1345 * Align up to a page boundary for the program segments. The
1346 * actual data will be written to the outptu file, not to elf_buf_t,
1347 * so we do not have to do any further bounds checking.
1349 target->off = round_page(target->off);
1350 if (error == 0 && ehdr != NULL) {
1352 * Fill in the ELF header.
1354 ehdr->e_ident[EI_MAG0] = ELFMAG0;
1355 ehdr->e_ident[EI_MAG1] = ELFMAG1;
1356 ehdr->e_ident[EI_MAG2] = ELFMAG2;
1357 ehdr->e_ident[EI_MAG3] = ELFMAG3;
1358 ehdr->e_ident[EI_CLASS] = ELF_CLASS;
1359 ehdr->e_ident[EI_DATA] = ELF_DATA;
1360 ehdr->e_ident[EI_VERSION] = EV_CURRENT;
1361 ehdr->e_ident[EI_OSABI] = ELFOSABI_NONE;
1362 ehdr->e_ident[EI_ABIVERSION] = 0;
1363 ehdr->e_ident[EI_PAD] = 0;
1364 ehdr->e_type = ET_CORE;
1365 ehdr->e_machine = ELF_ARCH;
1366 ehdr->e_version = EV_CURRENT;
1368 ehdr->e_phoff = phoff;
1370 ehdr->e_ehsize = sizeof(Elf_Ehdr);
1371 ehdr->e_phentsize = sizeof(Elf_Phdr);
1372 ehdr->e_phnum = numsegs + 1;
1373 ehdr->e_shentsize = sizeof(Elf_Shdr);
1375 ehdr->e_shstrndx = SHN_UNDEF;
1377 if (error == 0 && phdr != NULL) {
1379 * Fill in the program header entries.
1381 struct phdr_closure phc;
1383 /* The note segement. */
1384 phdr->p_type = PT_NOTE;
1385 phdr->p_offset = noteoff;
1388 phdr->p_filesz = notesz;
1394 /* All the writable segments from the program. */
1396 phc.phdr_max = phdr + numsegs;
1397 phc.offset = target->off;
1398 each_segment(p, cb_put_phdr, &phc, 1);
1404 * Append core dump notes to target ELF buffer or simply update target size
1405 * if dryrun selected.
1408 elf_putallnotes(struct lwp *corelp, elf_buf_t target, int sig,
1411 struct proc *p = corelp->lwp_proc;
1415 prfpregset_t fpregs;
1419 prfpregset_t *fpregs;
1424 * Allocate temporary storage for notes on heap to avoid stack overflow.
1426 if (mode != DRYRUN) {
1427 tmpdata = kmalloc(sizeof(*tmpdata), M_TEMP, M_ZERO | M_WAITOK);
1428 status = &tmpdata->status;
1429 fpregs = &tmpdata->fpregs;
1430 psinfo = &tmpdata->psinfo;
1439 * Append LWP-agnostic note.
1441 if (mode != DRYRUN) {
1442 psinfo->pr_version = PRPSINFO_VERSION;
1443 psinfo->pr_psinfosz = sizeof(prpsinfo_t);
1444 strncpy(psinfo->pr_fname, p->p_comm,
1445 sizeof(psinfo->pr_fname) - 1);
1447 * XXX - We don't fill in the command line arguments
1450 strncpy(psinfo->pr_psargs, p->p_comm, PRARGSZ);
1453 elf_putnote(target, "CORE", NT_PRPSINFO, psinfo, sizeof *psinfo);
1458 * Append first note for LWP that triggered core so that it is
1459 * the selected one when the debugger starts.
1461 if (mode != DRYRUN) {
1462 status->pr_version = PRSTATUS_VERSION;
1463 status->pr_statussz = sizeof(prstatus_t);
1464 status->pr_gregsetsz = sizeof(gregset_t);
1465 status->pr_fpregsetsz = sizeof(fpregset_t);
1466 status->pr_osreldate = osreldate;
1467 status->pr_cursig = sig;
1469 * XXX GDB needs unique pr_pid for each LWP and does not
1470 * not support pr_pid==0 but lwp_tid can be 0, so hack unique
1473 status->pr_pid = corelp->lwp_tid;
1474 fill_regs(corelp, &status->pr_reg);
1475 fill_fpregs(corelp, fpregs);
1478 elf_putnote(target, "CORE", NT_PRSTATUS, status, sizeof *status);
1482 elf_putnote(target, "CORE", NT_FPREGSET, fpregs, sizeof *fpregs);
1487 * Then append notes for other LWPs.
1489 FOREACH_LWP_IN_PROC(lp, p) {
1492 /* skip lwps being created */
1493 if (lp->lwp_thread == NULL)
1495 if (mode != DRYRUN) {
1496 status->pr_pid = lp->lwp_tid;
1497 fill_regs(lp, &status->pr_reg);
1498 fill_fpregs(lp, fpregs);
1500 error = elf_putnote(target, "CORE", NT_PRSTATUS,
1501 status, sizeof *status);
1504 error = elf_putnote(target, "CORE", NT_FPREGSET,
1505 fpregs, sizeof *fpregs);
1511 if (tmpdata != NULL)
1512 kfree(tmpdata, M_TEMP);
1517 * Generate a note sub-structure.
1519 * NOTE: 4-byte alignment.
1522 elf_putnote(elf_buf_t target, const char *name, int type,
1523 const void *desc, size_t descsz)
1529 note.n_namesz = strlen(name) + 1;
1530 note.n_descsz = descsz;
1532 dst = target_reserve(target, sizeof(note), &error);
1534 bcopy(¬e, dst, sizeof note);
1535 dst = target_reserve(target, note.n_namesz, &error);
1537 bcopy(name, dst, note.n_namesz);
1538 target->off = roundup2(target->off, sizeof(Elf_Word));
1539 dst = target_reserve(target, note.n_descsz, &error);
1541 bcopy(desc, dst, note.n_descsz);
1542 target->off = roundup2(target->off, sizeof(Elf_Word));
1548 elf_putsigs(struct lwp *lp, elf_buf_t target)
1550 /* XXX lwp handle more than one lwp */
1551 struct proc *p = lp->lwp_proc;
1553 struct ckpt_siginfo *csi;
1555 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error);
1557 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo);
1558 bcopy(p->p_sigacts, &csi->csi_sigacts, sizeof(*p->p_sigacts));
1559 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval));
1560 bcopy(&lp->lwp_sigmask, &csi->csi_sigmask,
1562 csi->csi_sigparent = p->p_sigparent;
1568 elf_putfiles(struct proc *p, elf_buf_t target, struct file *ckfp)
1572 struct ckpt_filehdr *cfh = NULL;
1573 struct ckpt_fileinfo *cfi;
1577 * the duplicated loop is gross, but it was the only way
1578 * to eliminate uninitialized variable warnings
1580 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error);
1582 cfh->cfh_nfiles = 0;
1586 * ignore STDIN/STDERR/STDOUT.
1588 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) {
1589 fp = holdfp(p->p_fd, i, -1);
1593 * XXX Only checkpoint vnodes for now.
1595 if (fp->f_type != DTYPE_VNODE) {
1599 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo),
1605 cfi->cfi_index = -1;
1606 cfi->cfi_type = fp->f_type;
1607 cfi->cfi_flags = fp->f_flag;
1608 cfi->cfi_offset = fp->f_offset;
1609 cfi->cfi_ckflags = 0;
1612 cfi->cfi_ckflags |= CKFIF_ISCKPTFD;
1613 /* f_count and f_msgcount should not be saved/restored */
1614 /* XXX save cred info */
1616 switch(fp->f_type) {
1618 vp = (struct vnode *)fp->f_data;
1620 * it looks like a bug in ptrace is marking
1621 * a non-vnode as a vnode - until we find the
1622 * root cause this will at least prevent
1623 * further panics from truss
1625 if (vp == NULL || vp->v_mount == NULL)
1629 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1630 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid);
1641 elf_puttextvp(struct proc *p, elf_buf_t target)
1645 struct fp_closure fpc;
1646 struct ckpt_vminfo *vminfo;
1648 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error);
1649 if (vminfo != NULL) {
1650 vminfo->cvm_dsize = p->p_vmspace->vm_dsize;
1651 vminfo->cvm_tsize = p->p_vmspace->vm_tsize;
1652 vminfo->cvm_daddr = p->p_vmspace->vm_daddr;
1653 vminfo->cvm_taddr = p->p_vmspace->vm_taddr;
1657 vn_count = target_reserve(target, sizeof(int), &error);
1658 if (target->buf != NULL) {
1659 fpc.vnh = (struct vn_hdr *)(target->buf + target->off);
1660 fpc.vnh_max = fpc.vnh +
1661 (target->off_max - target->off) / sizeof(struct vn_hdr);
1662 error = each_segment(p, cb_put_fp, &fpc, 0);
1664 *vn_count = fpc.count;
1666 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0);
1668 target->off += fpc.count * sizeof(struct vn_hdr);
1674 * Tell kern_execve.c about it, with a little help from the linker.
1676 static struct execsw elf_execsw = {exec_elf_imgact, "ELF"};
1677 EXEC_SET_ORDERED(elf, elf_execsw, SI_ORDER_FIRST);