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 1.41 2006/09/05 00:55:45 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 * Check if an elf brand is being used anywhere in the system.
199 * Used by the linux emulatino module unloader. This isn't safe from
202 struct elf_brand_inuse_info {
204 Elf_Brandinfo *entry;
207 static int elf_brand_inuse_callback(struct proc *p, void *data);
210 elf_brand_inuse(Elf_Brandinfo *entry)
212 struct elf_brand_inuse_info info;
216 allproc_scan(elf_brand_inuse_callback, entry);
222 elf_brand_inuse_callback(struct proc *p, void *data)
224 struct elf_brand_inuse_info *info = data;
226 if (p->p_sysent == info->entry->sysvec) {
234 elf_check_header(const Elf_Ehdr *hdr)
237 hdr->e_ident[EI_CLASS] != ELF_TARG_CLASS ||
238 hdr->e_ident[EI_DATA] != ELF_TARG_DATA ||
239 hdr->e_ident[EI_VERSION] != EV_CURRENT ||
240 hdr->e_phentsize != sizeof(Elf_Phdr) ||
241 hdr->e_ehsize != sizeof(Elf_Ehdr) ||
242 hdr->e_version != ELF_TARG_VER)
245 if (!ELF_MACHINE_OK(hdr->e_machine))
252 elf_load_section(struct proc *p, struct vmspace *vmspace, struct vnode *vp,
253 vm_offset_t offset, caddr_t vmaddr, size_t memsz,
254 size_t filsz, vm_prot_t prot)
257 vm_offset_t map_addr;
262 vm_offset_t file_addr;
263 vm_offset_t data_buf = 0;
265 object = vp->v_object;
269 * It's necessary to fail if the filsz + offset taken from the
270 * header is greater than the actual file pager object's size.
271 * If we were to allow this, then the vm_map_find() below would
272 * walk right off the end of the file object and into the ether.
274 * While I'm here, might as well check for something else that
275 * is invalid: filsz cannot be greater than memsz.
277 if ((off_t)filsz + offset > vp->v_filesize || filsz > memsz) {
278 uprintf("elf_load_section: truncated ELF file\n");
282 map_addr = trunc_page((vm_offset_t)vmaddr);
283 file_addr = trunc_page(offset);
286 * We have two choices. We can either clear the data in the last page
287 * of an oversized mapping, or we can start the anon mapping a page
288 * early and copy the initialized data into that first page. We
289 * choose the second..
292 map_len = trunc_page(offset+filsz) - file_addr;
294 map_len = round_page(offset+filsz) - file_addr;
297 vm_object_reference(object);
299 /* cow flags: don't dump readonly sections in core */
300 cow = MAP_COPY_ON_WRITE | MAP_PREFAULT |
301 (prot & VM_PROT_WRITE ? 0 : MAP_DISABLE_COREDUMP);
303 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
304 vm_map_lock(&vmspace->vm_map);
305 rv = vm_map_insert(&vmspace->vm_map, &count,
307 file_addr, /* file offset */
308 map_addr, /* virtual start */
309 map_addr + map_len,/* virtual end */
313 vm_map_unlock(&vmspace->vm_map);
314 vm_map_entry_release(count);
315 if (rv != KERN_SUCCESS) {
316 vm_object_deallocate(object);
320 /* we can stop now if we've covered it all */
321 if (memsz == filsz) {
328 * We have to get the remaining bit of the file into the first part
329 * of the oversized map segment. This is normally because the .data
330 * segment in the file is extended to provide bss. It's a neat idea
331 * to try and save a page, but it's a pain in the behind to implement.
333 copy_len = (offset + filsz) - trunc_page(offset + filsz);
334 map_addr = trunc_page((vm_offset_t)vmaddr + filsz);
335 map_len = round_page((vm_offset_t)vmaddr + memsz) - map_addr;
337 /* This had damn well better be true! */
339 count = vm_map_entry_reserve(MAP_RESERVE_COUNT);
340 vm_map_lock(&vmspace->vm_map);
341 rv = vm_map_insert(&vmspace->vm_map, &count,
343 map_addr, map_addr + map_len,
344 VM_PROT_ALL, VM_PROT_ALL, 0);
345 vm_map_unlock(&vmspace->vm_map);
346 vm_map_entry_release(count);
347 if (rv != KERN_SUCCESS) {
353 vm_object_reference(object);
354 rv = vm_map_find(exec_map,
356 trunc_page(offset + filsz),
362 MAP_COPY_ON_WRITE | MAP_PREFAULT_PARTIAL);
363 if (rv != KERN_SUCCESS) {
364 vm_object_deallocate(object);
368 /* send the page fragment to user space */
369 error = copyout((caddr_t)data_buf, (caddr_t)map_addr, copy_len);
370 vm_map_remove(exec_map, data_buf, data_buf + PAGE_SIZE);
377 * set it to the specified protection
379 vm_map_protect(&vmspace->vm_map, map_addr, map_addr + map_len, prot,
386 * Load the file "file" into memory. It may be either a shared object
389 * The "addr" reference parameter is in/out. On entry, it specifies
390 * the address where a shared object should be loaded. If the file is
391 * an executable, this value is ignored. On exit, "addr" specifies
392 * where the file was actually loaded.
394 * The "entry" reference parameter is out only. On exit, it specifies
395 * the entry point for the loaded file.
398 elf_load_file(struct proc *p, const char *file, u_long *addr, u_long *entry)
401 struct nlookupdata nd;
403 struct image_params image_params;
405 const Elf_Ehdr *hdr = NULL;
406 const Elf_Phdr *phdr = NULL;
407 struct nlookupdata *nd;
408 struct vmspace *vmspace = p->p_vmspace;
410 struct image_params *imgp;
413 u_long base_addr = 0;
414 int error, i, numsegs;
416 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_WAITOK);
418 attr = &tempdata->attr;
419 imgp = &tempdata->image_params;
422 * Initialize part of the common data
426 imgp->firstpage = NULL;
427 imgp->image_header = NULL;
430 error = nlookup_init(nd, file, UIO_SYSSPACE, NLC_FOLLOW);
434 error = cache_vget(nd->nl_ncp, nd->nl_cred, LK_EXCLUSIVE, &imgp->vp);
440 * Check permissions, modes, uid, etc on the file, and "open" it.
442 error = exec_check_permissions(imgp);
448 error = exec_map_first_page(imgp);
450 * Also make certain that the interpreter stays the same, so set
451 * its VTEXT flag, too.
454 imgp->vp->v_flag |= VTEXT;
459 hdr = (const Elf_Ehdr *)imgp->image_header;
460 if ((error = elf_check_header(hdr)) != 0)
462 if (hdr->e_type == ET_DYN)
464 else if (hdr->e_type == ET_EXEC)
471 /* Only support headers that fit within first page for now
472 * (multiplication of two Elf_Half fields will not overflow) */
473 if ((hdr->e_phoff > PAGE_SIZE) ||
474 (hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE - hdr->e_phoff) {
479 phdr = (const Elf_Phdr *)(imgp->image_header + hdr->e_phoff);
481 for (i = 0, numsegs = 0; i < hdr->e_phnum; i++) {
482 if (phdr[i].p_type == PT_LOAD) { /* Loadable segment */
484 if (phdr[i].p_flags & PF_X)
485 prot |= VM_PROT_EXECUTE;
486 if (phdr[i].p_flags & PF_W)
487 prot |= VM_PROT_WRITE;
488 if (phdr[i].p_flags & PF_R)
489 prot |= VM_PROT_READ;
491 error = elf_load_section(
492 p, vmspace, imgp->vp,
494 (caddr_t)phdr[i].p_vaddr +
497 phdr[i].p_filesz, prot);
501 * Establish the base address if this is the
505 base_addr = trunc_page(phdr[i].p_vaddr + rbase);
510 *entry=(unsigned long)hdr->e_entry + rbase;
514 exec_unmap_first_page(imgp);
519 kfree(tempdata, M_TEMP);
525 * non static, as it can be overridden by start_init()
527 int fallback_elf_brand = -1;
528 SYSCTL_INT(_kern, OID_AUTO, fallback_elf_brand, CTLFLAG_RW,
529 &fallback_elf_brand, -1,
530 "ELF brand of last resort");
533 exec_elf_imgact(struct image_params *imgp)
535 const Elf_Ehdr *hdr = (const Elf_Ehdr *) imgp->image_header;
536 const Elf_Phdr *phdr;
537 Elf_Auxargs *elf_auxargs = NULL;
538 struct vmspace *vmspace;
540 u_long text_size = 0, data_size = 0, total_size = 0;
541 u_long text_addr = 0, data_addr = 0;
542 u_long seg_size, seg_addr;
543 u_long addr, entry = 0, proghdr = 0;
545 const char *interp = NULL;
546 const Elf_Note *abi_note = NULL;
547 Elf_Brandinfo *brand_info;
553 * Do we have a valid ELF header ?
555 if (elf_check_header(hdr) != 0 || hdr->e_type != ET_EXEC)
559 * From here on down, we return an errno, not -1, as we've
560 * detected an ELF file.
563 if ((hdr->e_phoff > PAGE_SIZE) ||
564 (hdr->e_phoff + hdr->e_phentsize * hdr->e_phnum) > PAGE_SIZE) {
565 /* Only support headers in first page for now */
568 phdr = (const Elf_Phdr*)(imgp->image_header + hdr->e_phoff);
571 * From this point on, we may have resources that need to be freed.
574 exec_new_vmspace(imgp, NULL);
577 * Yeah, I'm paranoid. There is every reason in the world to get
578 * VTEXT now since from here on out, there are places we can have
579 * a context switch. Better safe than sorry; I really don't want
580 * the file to change while it's being loaded.
582 vsetflags(imgp->vp, VTEXT);
584 vmspace = imgp->proc->p_vmspace;
586 for (i = 0; i < hdr->e_phnum; i++) {
587 switch(phdr[i].p_type) {
589 case PT_LOAD: /* Loadable segment */
591 if (phdr[i].p_flags & PF_X)
592 prot |= VM_PROT_EXECUTE;
593 if (phdr[i].p_flags & PF_W)
594 prot |= VM_PROT_WRITE;
595 if (phdr[i].p_flags & PF_R)
596 prot |= VM_PROT_READ;
598 if ((error = elf_load_section(imgp->proc,
601 (caddr_t)phdr[i].p_vaddr,
603 phdr[i].p_filesz, prot)) != 0)
607 * If this segment contains the program headers,
608 * remember their virtual address for the AT_PHDR
609 * aux entry. Static binaries don't usually include
612 if (phdr[i].p_offset == 0 &&
613 hdr->e_phoff + hdr->e_phnum * hdr->e_phentsize
615 proghdr = phdr[i].p_vaddr + hdr->e_phoff;
617 seg_addr = trunc_page(phdr[i].p_vaddr);
618 seg_size = round_page(phdr[i].p_memsz +
619 phdr[i].p_vaddr - seg_addr);
622 * Is this .text or .data? We can't use
623 * VM_PROT_WRITE or VM_PROT_EXEC, it breaks the
624 * alpha terribly and possibly does other bad
625 * things so we stick to the old way of figuring
626 * it out: If the segment contains the program
627 * entry point, it's a text segment, otherwise it
630 * Note that obreak() assumes that data_addr +
631 * data_size == end of data load area, and the ELF
632 * file format expects segments to be sorted by
633 * address. If multiple data segments exist, the
634 * last one will be used.
636 if (hdr->e_entry >= phdr[i].p_vaddr &&
637 hdr->e_entry < (phdr[i].p_vaddr +
639 text_size = seg_size;
640 text_addr = seg_addr;
641 entry = (u_long)hdr->e_entry;
643 data_size = seg_size;
644 data_addr = seg_addr;
646 total_size += seg_size;
649 * Check limits. It should be safe to check the
650 * limits after loading the segment since we do
651 * not actually fault in all the segment's pages.
654 imgp->proc->p_rlimit[RLIMIT_DATA].rlim_cur ||
655 text_size > maxtsiz ||
657 imgp->proc->p_rlimit[RLIMIT_VMEM].rlim_cur) {
662 case PT_INTERP: /* Path to interpreter */
663 if (phdr[i].p_filesz > MAXPATHLEN ||
664 phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE) {
668 interp = imgp->image_header + phdr[i].p_offset;
670 case PT_NOTE: /* Check for .note.ABI-tag */
672 const Elf_Note *tmp_note;
673 /* XXX handle anything outside the first page */
674 if (phdr[i].p_offset + phdr[i].p_filesz > PAGE_SIZE)
676 if (phdr[i].p_filesz < sizeof(Elf_Note))
677 continue; /* ENOEXEC? */
678 tmp_note = (const Elf_Note *)(imgp->image_header + phdr[i].p_offset);
679 if (tmp_note->n_type != 1)
681 if (tmp_note->n_namesz + sizeof(Elf_Note) +
682 tmp_note->n_descsz > phdr[i].p_filesz)
683 continue; /* ENOEXEC? */
687 case PT_PHDR: /* Program header table info */
688 proghdr = phdr[i].p_vaddr;
695 vmspace->vm_tsize = text_size >> PAGE_SHIFT;
696 vmspace->vm_taddr = (caddr_t)(uintptr_t)text_addr;
697 vmspace->vm_dsize = data_size >> PAGE_SHIFT;
698 vmspace->vm_daddr = (caddr_t)(uintptr_t)data_addr;
700 addr = ELF_RTLD_ADDR(vmspace);
702 imgp->entry_addr = entry;
706 /* We support three types of branding -- (1) the ELF EI_OSABI field
707 * that SCO added to the ELF spec, (2) FreeBSD 3.x's traditional string
708 * branding w/in the ELF header, and (3) path of the `interp_path'
709 * field. We should also look for an ".note.ABI-tag" ELF section now
710 * in all Linux ELF binaries, FreeBSD 4.1+, and some NetBSD ones.
713 /* If the executable has a brand, search for it in the brand list. */
714 if (brand_info == NULL && hdr->e_ident[EI_OSABI] != ELFOSABI_NONE) {
715 for (i = 0; i < MAX_BRANDS; i++) {
716 Elf_Brandinfo *bi = elf_brand_list[i];
719 (hdr->e_ident[EI_OSABI] == bi->brand
721 strncmp((const char *)&hdr->e_ident[OLD_EI_BRAND],
722 bi->compat_3_brand, strlen(bi->compat_3_brand)))) {
729 /* Search for a recognized ABI. */
730 if (brand_info == NULL && abi_note != NULL) {
731 for (i = 0; i < MAX_BRANDS; i++) {
732 Elf_Brandinfo *bi = elf_brand_list[i];
734 if (bi != NULL && bi->match_abi_note != NULL &&
735 (*bi->match_abi_note)(abi_note)) {
743 * ELFOSABI_NONE == ELFOSABI_SYSV, so a SYSV binary misses all
744 * checks so far, since it is neither branded nor does it have
745 * an ABI note. If the EI_OSABI field is ELFOSABI_NONE, assume
746 * it is svr4 and look for an entry in the elf_brand_list with
747 * match_abi_note == NULL.
749 if (brand_info == NULL && hdr->e_ident[EI_OSABI] == ELFOSABI_NONE) {
750 for (i = 0; i < MAX_BRANDS; i++) {
751 Elf_Brandinfo *bi = elf_brand_list[i];
753 if (bi != NULL && bi->match_abi_note == NULL &&
754 ELFOSABI_SYSV == bi->brand) {
761 /* Lacking a recognized ABI, search for a recognized interpreter. */
762 if (brand_info == NULL && interp != NULL) {
763 for (i = 0; i < MAX_BRANDS; i++) {
764 Elf_Brandinfo *bi = elf_brand_list[i];
767 strcmp(interp, bi->interp_path) == 0) {
774 /* Lacking a recognized interpreter, try the default brand */
775 if (brand_info == NULL) {
776 for (i = 0; i < MAX_BRANDS; i++) {
777 Elf_Brandinfo *bi = elf_brand_list[i];
779 if (bi != NULL && fallback_elf_brand == bi->brand) {
786 if (brand_info == NULL) {
787 uprintf("ELF binary type \"%u\" not known.\n",
788 hdr->e_ident[EI_OSABI]);
793 imgp->proc->p_sysent = brand_info->sysvec;
794 if (interp != NULL) {
795 path = kmalloc(MAXPATHLEN, M_TEMP, M_WAITOK);
796 snprintf(path, MAXPATHLEN, "%s%s",
797 brand_info->emul_path, interp);
798 if ((error = elf_load_file(imgp->proc, path, &addr,
799 &imgp->entry_addr)) != 0) {
800 if ((error = elf_load_file(imgp->proc, interp, &addr,
801 &imgp->entry_addr)) != 0) {
802 uprintf("ELF interpreter %s not found\n", path);
811 * Construct auxargs table (used by the fixup routine)
813 elf_auxargs = kmalloc(sizeof(Elf_Auxargs), M_TEMP, M_WAITOK);
814 elf_auxargs->execfd = -1;
815 elf_auxargs->phdr = proghdr;
816 elf_auxargs->phent = hdr->e_phentsize;
817 elf_auxargs->phnum = hdr->e_phnum;
818 elf_auxargs->pagesz = PAGE_SIZE;
819 elf_auxargs->base = addr;
820 elf_auxargs->flags = 0;
821 elf_auxargs->entry = entry;
822 elf_auxargs->trace = elf_trace;
824 imgp->auxargs = elf_auxargs;
825 imgp->interpreted = 0;
832 elf_freebsd_fixup(register_t **stack_base, struct image_params *imgp)
834 Elf_Auxargs *args = (Elf_Auxargs *)imgp->auxargs;
837 pos = *stack_base + (imgp->args->argc + imgp->args->envc + 2);
840 AUXARGS_ENTRY(pos, AT_DEBUG, 1);
842 if (args->execfd != -1) {
843 AUXARGS_ENTRY(pos, AT_EXECFD, args->execfd);
845 AUXARGS_ENTRY(pos, AT_PHDR, args->phdr);
846 AUXARGS_ENTRY(pos, AT_PHENT, args->phent);
847 AUXARGS_ENTRY(pos, AT_PHNUM, args->phnum);
848 AUXARGS_ENTRY(pos, AT_PAGESZ, args->pagesz);
849 AUXARGS_ENTRY(pos, AT_FLAGS, args->flags);
850 AUXARGS_ENTRY(pos, AT_ENTRY, args->entry);
851 AUXARGS_ENTRY(pos, AT_BASE, args->base);
852 AUXARGS_ENTRY(pos, AT_NULL, 0);
854 kfree(imgp->auxargs, M_TEMP);
855 imgp->auxargs = NULL;
858 suword(*stack_base, (long) imgp->args->argc);
863 * Code for generating ELF core dumps.
866 typedef int (*segment_callback) (vm_map_entry_t, void *);
868 /* Closure for cb_put_phdr(). */
869 struct phdr_closure {
870 Elf_Phdr *phdr; /* Program header to fill in (incremented) */
871 Elf_Phdr *phdr_max; /* Pointer bound for error check */
872 Elf_Off offset; /* Offset of segment in core file */
875 /* Closure for cb_size_segment(). */
876 struct sseg_closure {
877 int count; /* Count of writable segments. */
878 size_t vsize; /* Total size of all writable segments. */
881 /* Closure for cb_put_fp(). */
884 struct vn_hdr *vnh_max;
889 typedef struct elf_buf {
895 static void *target_reserve(elf_buf_t target, size_t bytes, int *error);
897 static int cb_put_phdr (vm_map_entry_t, void *);
898 static int cb_size_segment (vm_map_entry_t, void *);
899 static int cb_fpcount_segment(vm_map_entry_t, void *);
900 static int cb_put_fp(vm_map_entry_t, void *);
903 static int each_segment (struct proc *, segment_callback, void *, int);
904 static int elf_corehdr (struct proc *, struct file *, struct ucred *,
906 static int elf_puthdr (struct proc *, elf_buf_t, const prstatus_t *,
907 const prfpregset_t *, const prpsinfo_t *, int);
908 static int elf_putnote (elf_buf_t, const char *, int, const void *, size_t);
910 static int elf_putsigs(struct proc *, elf_buf_t);
911 static int elf_puttextvp(struct proc *, elf_buf_t);
912 static int elf_putfiles(struct proc *, elf_buf_t);
914 extern int osreldate;
917 elf_coredump(struct proc *p, struct vnode *vp, off_t limit)
922 if ((error = falloc(NULL, &fp, NULL)) != 0)
924 fsetcred(fp, p->p_ucred);
929 fp->f_type = DTYPE_VNODE;
930 fp->f_flag = O_CREAT|O_WRONLY|O_NOFOLLOW;
931 fp->f_ops = &vnode_fileops;
935 error = generic_elf_coredump(p, fp, limit);
939 fp->f_ops = &badfileops;
946 generic_elf_coredump(struct proc *p, struct file *fp, off_t limit)
948 struct ucred *cred = p->p_ucred;
950 struct sseg_closure seginfo;
951 struct elf_buf target;
954 printf("can't dump core - null fp\n");
957 * Size the program segments
961 each_segment(p, cb_size_segment, &seginfo, 1);
964 * Calculate the size of the core file header area by making
965 * a dry run of generating it. Nothing is written, but the
966 * size is calculated.
968 bzero(&target, sizeof(target));
969 elf_puthdr(p, &target, NULL, NULL, NULL, seginfo.count);
971 if (target.off + seginfo.vsize >= limit)
975 * Allocate memory for building the header, fill it up,
978 target.off_max = target.off;
980 target.buf = kmalloc(target.off_max, M_TEMP, M_WAITOK|M_ZERO);
982 if (target.buf == NULL)
984 error = elf_corehdr(p, fp, cred, seginfo.count, &target);
986 /* Write the contents of all of the writable segments. */
992 php = (Elf_Phdr *)(target.buf + sizeof(Elf_Ehdr)) + 1;
993 for (i = 0; i < seginfo.count; i++) {
994 error = fp_write(fp, (caddr_t)php->p_vaddr,
995 php->p_filesz, &nbytes);
1001 kfree(target.buf, M_TEMP);
1007 * A callback for each_segment() to write out the segment's
1008 * program header entry.
1011 cb_put_phdr(vm_map_entry_t entry, void *closure)
1013 struct phdr_closure *phc = closure;
1014 Elf_Phdr *phdr = phc->phdr;
1016 if (phc->phdr == phc->phdr_max)
1019 phc->offset = round_page(phc->offset);
1021 phdr->p_type = PT_LOAD;
1022 phdr->p_offset = phc->offset;
1023 phdr->p_vaddr = entry->start;
1025 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1026 phdr->p_align = PAGE_SIZE;
1028 if (entry->protection & VM_PROT_READ)
1029 phdr->p_flags |= PF_R;
1030 if (entry->protection & VM_PROT_WRITE)
1031 phdr->p_flags |= PF_W;
1032 if (entry->protection & VM_PROT_EXECUTE)
1033 phdr->p_flags |= PF_X;
1035 phc->offset += phdr->p_filesz;
1041 * A callback for each_writable_segment() to gather information about
1042 * the number of segments and their total size.
1045 cb_size_segment(vm_map_entry_t entry, void *closure)
1047 struct sseg_closure *ssc = closure;
1050 ssc->vsize += entry->end - entry->start;
1055 * A callback for each_segment() to gather information about
1056 * the number of text segments.
1059 cb_fpcount_segment(vm_map_entry_t entry, void *closure)
1061 int *count = closure;
1064 if (entry->object.vm_object->type == OBJT_VNODE) {
1065 vp = (struct vnode *)entry->object.vm_object->handle;
1066 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1074 cb_put_fp(vm_map_entry_t entry, void *closure)
1076 struct fp_closure *fpc = closure;
1077 struct vn_hdr *vnh = fpc->vnh;
1078 Elf_Phdr *phdr = &vnh->vnh_phdr;
1083 * If an entry represents a vnode then write out a file handle.
1085 * If we are checkpointing a checkpoint-restored program we do
1086 * NOT record the filehandle for the old checkpoint vnode (which
1087 * is mapped all over the place). Instead we rely on the fact
1088 * that a checkpoint-restored program does not mmap() the checkpt
1089 * vnode NOCORE, so its contents will be written out to the
1090 * new checkpoint file. This is necessary because the 'old'
1091 * checkpoint file is typically destroyed when a new one is created
1092 * and thus cannot be used to restore the new checkpoint.
1094 * Theoretically we could create a chain of checkpoint files and
1095 * operate the checkpointing operation kinda like an incremental
1096 * checkpoint, but a checkpoint restore would then likely wind up
1097 * referencing many prior checkpoint files and that is a bit over
1098 * the top for the purpose of the checkpoint API.
1100 if (entry->object.vm_object->type == OBJT_VNODE) {
1101 vp = (struct vnode *)entry->object.vm_object->handle;
1102 if ((vp->v_flag & VCKPT) && curproc->p_textvp == vp)
1104 if (vnh == fpc->vnh_max)
1108 vnh->vnh_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1109 error = VFS_VPTOFH(vp, &vnh->vnh_fh.fh_fid);
1111 char *freepath, *fullpath;
1113 if (vn_fullpath(curproc, vp, &fullpath, &freepath)) {
1114 printf("Warning: coredump, error %d: cannot store file handle for vnode %p\n", error, vp);
1116 printf("Warning: coredump, error %d: cannot store file handle for %s\n", error, fullpath);
1117 kfree(freepath, M_TEMP);
1122 phdr->p_type = PT_LOAD;
1123 phdr->p_offset = 0; /* not written to core */
1124 phdr->p_vaddr = entry->start;
1126 phdr->p_filesz = phdr->p_memsz = entry->end - entry->start;
1127 phdr->p_align = PAGE_SIZE;
1129 if (entry->protection & VM_PROT_READ)
1130 phdr->p_flags |= PF_R;
1131 if (entry->protection & VM_PROT_WRITE)
1132 phdr->p_flags |= PF_W;
1133 if (entry->protection & VM_PROT_EXECUTE)
1134 phdr->p_flags |= PF_X;
1142 * For each writable segment in the process's memory map, call the given
1143 * function with a pointer to the map entry and some arbitrary
1144 * caller-supplied data.
1147 each_segment(struct proc *p, segment_callback func, void *closure, int writable)
1150 vm_map_t map = &p->p_vmspace->vm_map;
1151 vm_map_entry_t entry;
1153 for (entry = map->header.next; error == 0 && entry != &map->header;
1154 entry = entry->next) {
1158 * Don't dump inaccessible mappings, deal with legacy
1161 * Note that read-only segments related to the elf binary
1162 * are marked MAP_ENTRY_NOCOREDUMP now so we no longer
1163 * need to arbitrarily ignore such segments.
1165 if (elf_legacy_coredump) {
1166 if (writable && (entry->protection & VM_PROT_RW) != VM_PROT_RW)
1169 if (writable && (entry->protection & VM_PROT_ALL) == 0)
1174 * Dont include memory segment in the coredump if
1175 * MAP_NOCORE is set in mmap(2) or MADV_NOCORE in
1176 * madvise(2). Do not dump submaps (i.e. parts of the
1179 if (writable && entry->eflags & (MAP_ENTRY_NOCOREDUMP|MAP_ENTRY_IS_SUB_MAP))
1182 if ((obj = entry->object.vm_object) == NULL)
1185 /* Find the deepest backing object. */
1186 while (obj->backing_object != NULL)
1187 obj = obj->backing_object;
1189 /* Ignore memory-mapped devices and such things. */
1190 if (obj->type != OBJT_DEFAULT &&
1191 obj->type != OBJT_SWAP &&
1192 obj->type != OBJT_VNODE)
1195 error = (*func)(entry, closure);
1202 target_reserve(elf_buf_t target, size_t bytes, int *error)
1207 if (target->off + bytes > target->off_max)
1210 res = target->buf + target->off;
1212 target->off += bytes;
1217 * Write the core file header to the file, including padding up to
1218 * the page boundary.
1221 elf_corehdr(struct proc *p, struct file *fp, struct ucred *cred, int numsegs,
1226 prfpregset_t fpregset;
1231 prfpregset_t *fpregset;
1234 tempdata = kmalloc(sizeof(*tempdata), M_TEMP, M_ZERO | M_WAITOK);
1235 status = &tempdata->status;
1236 fpregset = &tempdata->fpregset;
1237 psinfo = &tempdata->psinfo;
1239 /* Gather the information for the header. */
1240 status->pr_version = PRSTATUS_VERSION;
1241 status->pr_statussz = sizeof(prstatus_t);
1242 status->pr_gregsetsz = sizeof(gregset_t);
1243 status->pr_fpregsetsz = sizeof(fpregset_t);
1244 status->pr_osreldate = osreldate;
1245 status->pr_cursig = p->p_sig;
1246 status->pr_pid = p->p_pid;
1247 fill_regs(&p->p_lwp, &status->pr_reg);
1249 fill_fpregs(&p->p_lwp, fpregset);
1251 psinfo->pr_version = PRPSINFO_VERSION;
1252 psinfo->pr_psinfosz = sizeof(prpsinfo_t);
1253 strncpy(psinfo->pr_fname, p->p_comm, sizeof(psinfo->pr_fname) - 1);
1255 /* XXX - We don't fill in the command line arguments properly yet. */
1256 strncpy(psinfo->pr_psargs, p->p_comm, PRARGSZ);
1258 /* Fill in the header. */
1259 error = elf_puthdr(p, target, status, fpregset, psinfo, numsegs);
1261 kfree(tempdata, M_TEMP);
1263 /* Write it to the core file. */
1265 error = fp_write(fp, target->buf, target->off, &nbytes);
1270 elf_puthdr(struct proc *p, elf_buf_t target, const prstatus_t *status,
1271 const prfpregset_t *fpregset, const prpsinfo_t *psinfo, int numsegs)
1280 ehdr = target_reserve(target, sizeof(Elf_Ehdr), &error);
1282 phoff = target->off;
1283 phdr = target_reserve(target, (numsegs + 1) * sizeof(Elf_Phdr), &error);
1285 noteoff = target->off;
1287 error = elf_putnote(target, "FreeBSD", NT_PRSTATUS,
1288 status, sizeof *status);
1291 error = elf_putnote(target, "FreeBSD", NT_FPREGSET,
1292 fpregset, sizeof *fpregset);
1295 error = elf_putnote(target, "FreeBSD", NT_PRPSINFO,
1296 psinfo, sizeof *psinfo);
1298 notesz = target->off - noteoff;
1301 * put extra cruft for dumping process state here
1302 * - we really want it be before all the program
1304 * - we just need to update the offset accordingly
1305 * and GDB will be none the wiser.
1308 error = elf_puttextvp(p, target);
1310 error = elf_putsigs(p, target);
1312 error = elf_putfiles(p, target);
1315 * Align up to a page boundary for the program segments. The
1316 * actual data will be written to the outptu file, not to elf_buf_t,
1317 * so we do not have to do any further bounds checking.
1319 target->off = round_page(target->off);
1320 if (error == 0 && ehdr != NULL) {
1322 * Fill in the ELF header.
1324 ehdr->e_ident[EI_MAG0] = ELFMAG0;
1325 ehdr->e_ident[EI_MAG1] = ELFMAG1;
1326 ehdr->e_ident[EI_MAG2] = ELFMAG2;
1327 ehdr->e_ident[EI_MAG3] = ELFMAG3;
1328 ehdr->e_ident[EI_CLASS] = ELF_CLASS;
1329 ehdr->e_ident[EI_DATA] = ELF_DATA;
1330 ehdr->e_ident[EI_VERSION] = EV_CURRENT;
1331 ehdr->e_ident[EI_OSABI] = ELFOSABI_FREEBSD;
1332 ehdr->e_ident[EI_ABIVERSION] = 0;
1333 ehdr->e_ident[EI_PAD] = 0;
1334 ehdr->e_type = ET_CORE;
1335 ehdr->e_machine = ELF_ARCH;
1336 ehdr->e_version = EV_CURRENT;
1338 ehdr->e_phoff = phoff;
1340 ehdr->e_ehsize = sizeof(Elf_Ehdr);
1341 ehdr->e_phentsize = sizeof(Elf_Phdr);
1342 ehdr->e_phnum = numsegs + 1;
1343 ehdr->e_shentsize = sizeof(Elf_Shdr);
1345 ehdr->e_shstrndx = SHN_UNDEF;
1347 if (error == 0 && phdr != NULL) {
1349 * Fill in the program header entries.
1351 struct phdr_closure phc;
1353 /* The note segement. */
1354 phdr->p_type = PT_NOTE;
1355 phdr->p_offset = noteoff;
1358 phdr->p_filesz = notesz;
1364 /* All the writable segments from the program. */
1366 phc.phdr_max = phdr + numsegs;
1367 phc.offset = target->off;
1368 each_segment(p, cb_put_phdr, &phc, 1);
1374 elf_putnote(elf_buf_t target, const char *name, int type,
1375 const void *desc, size_t descsz)
1381 note.n_namesz = strlen(name) + 1;
1382 note.n_descsz = descsz;
1384 dst = target_reserve(target, sizeof(note), &error);
1386 bcopy(¬e, dst, sizeof note);
1387 dst = target_reserve(target, note.n_namesz, &error);
1389 bcopy(name, dst, note.n_namesz);
1390 target->off = roundup2(target->off, sizeof(Elf_Size));
1391 dst = target_reserve(target, note.n_descsz, &error);
1393 bcopy(desc, dst, note.n_descsz);
1394 target->off = roundup2(target->off, sizeof(Elf_Size));
1400 elf_putsigs(struct proc *p, elf_buf_t target)
1403 struct ckpt_siginfo *csi;
1405 csi = target_reserve(target, sizeof(struct ckpt_siginfo), &error);
1407 csi->csi_ckptpisz = sizeof(struct ckpt_siginfo);
1408 bcopy(p->p_procsig, &csi->csi_procsig, sizeof(struct procsig));
1409 bcopy(p->p_procsig->ps_sigacts, &csi->csi_sigacts, sizeof(struct sigacts));
1410 bcopy(&p->p_realtimer, &csi->csi_itimerval, sizeof(struct itimerval));
1411 bcopy(&p->p_sigmask, &csi->csi_sigmask, sizeof(sigset_t));
1412 csi->csi_sigparent = p->p_sigparent;
1418 elf_putfiles(struct proc *p, elf_buf_t target)
1422 struct ckpt_filehdr *cfh = NULL;
1423 struct ckpt_fileinfo *cfi;
1427 * the duplicated loop is gross, but it was the only way
1428 * to eliminate uninitialized variable warnings
1430 cfh = target_reserve(target, sizeof(struct ckpt_filehdr), &error);
1432 cfh->cfh_nfiles = 0;
1436 * ignore STDIN/STDERR/STDOUT.
1438 for (i = 3; error == 0 && i < p->p_fd->fd_nfiles; i++) {
1439 fp = holdfp(p->p_fd, i, -1);
1443 * XXX Only checkpoint vnodes for now.
1445 if (fp->f_type != DTYPE_VNODE) {
1449 cfi = target_reserve(target, sizeof(struct ckpt_fileinfo),
1455 cfi->cfi_index = -1;
1456 cfi->cfi_type = fp->f_type;
1457 cfi->cfi_flags = fp->f_flag;
1458 cfi->cfi_offset = fp->f_offset;
1459 /* f_count and f_msgcount should not be saved/restored */
1460 /* XXX save cred info */
1462 switch(fp->f_type) {
1464 vp = (struct vnode *)fp->f_data;
1466 * it looks like a bug in ptrace is marking
1467 * a non-vnode as a vnode - until we find the
1468 * root cause this will at least prevent
1469 * further panics from truss
1471 if (vp == NULL || vp->v_mount == NULL)
1475 cfi->cfi_fh.fh_fsid = vp->v_mount->mnt_stat.f_fsid;
1476 error = VFS_VPTOFH(vp, &cfi->cfi_fh.fh_fid);
1487 elf_puttextvp(struct proc *p, elf_buf_t target)
1491 struct fp_closure fpc;
1492 struct ckpt_vminfo *vminfo;
1494 vminfo = target_reserve(target, sizeof(struct ckpt_vminfo), &error);
1495 if (vminfo != NULL) {
1496 vminfo->cvm_dsize = p->p_vmspace->vm_dsize;
1497 vminfo->cvm_tsize = p->p_vmspace->vm_tsize;
1498 vminfo->cvm_daddr = p->p_vmspace->vm_daddr;
1499 vminfo->cvm_taddr = p->p_vmspace->vm_taddr;
1503 vn_count = target_reserve(target, sizeof(int), &error);
1504 if (target->buf != NULL) {
1505 fpc.vnh = (struct vn_hdr *)(target->buf + target->off);
1506 fpc.vnh_max = fpc.vnh +
1507 (target->off_max - target->off) / sizeof(struct vn_hdr);
1508 error = each_segment(p, cb_put_fp, &fpc, 0);
1510 *vn_count = fpc.count;
1512 error = each_segment(p, cb_fpcount_segment, &fpc.count, 0);
1514 target->off += fpc.count * sizeof(struct vn_hdr);
1520 * Tell kern_execve.c about it, with a little help from the linker.
1522 static struct execsw elf_execsw = {exec_elf_imgact, "ELF"};
1523 EXEC_SET_ORDERED(elf, elf_execsw, SI_ORDER_FIRST);