fs/proc/kcore.c

   1 /*
   2  *      fs/proc/kcore.c kernel ELF core dumper
   3  *
   4  *      Modelled on fs/exec.c:aout_core_dump()
   5  *      Jeremy Fitzhardinge <jeremy@sw.oz.au>
   6  *      ELF version written by David Howells <David.Howells@nexor.co.uk>
   7  *      Modified and incorporated into 2.3.x by Tigran Aivazian <tigran@veritas.com>
   8  *      Support to dump vmalloc'd areas (ELF only), Tigran Aivazian <tigran@veritas.com>
   9  *      Safe accesses to vmalloc/direct-mapped discontiguous areas, Kanoj Sarcar <kanoj@sgi.com>
  10  */
  11
  12 #include <linux/config.h>
  13 #include <linux/mm.h>
  14 #include <linux/proc_fs.h>
  15 #include <linux/user.h>
  16 #include <linux/a.out.h>
  17 #include <linux/elf.h>
  18 #include <linux/elfcore.h>
  19 #include <linux/vmalloc.h>
  20 #include <linux/highmem.h>
  21 #include <linux/init.h>
  22 #include <asm/uaccess.h>
  23 #include <asm/io.h>
  24
  25
  26 static int open_kcore(struct inode * inode, struct file * filp)
  27 {
  28         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
  29 }
  30
  31 static ssize_t read_kcore(struct file *, char __user *, size_t, loff_t *);
  32
  33 struct file_operations proc_kcore_operations = {
  34         .read           = read_kcore,
  35         .open           = open_kcore,
  36 };
  37
  38 #ifndef kc_vaddr_to_offset
  39 #define kc_vaddr_to_offset(v) ((v) - PAGE_OFFSET)
  40 #endif
  41 #ifndef kc_offset_to_vaddr
  42 #define kc_offset_to_vaddr(o) ((o) + PAGE_OFFSET)
  43 #endif
  44
  45 #define roundup(x, y)  ((((x)+((y)-1))/(y))*(y))
  46
  47 /* An ELF note in memory */
  48 struct memelfnote
  49 {
  50         const char *name;
  51         int type;
  52         unsigned int datasz;
  53         void *data;
  54 };
  55
  56 static struct kcore_list *kclist;
  57 static rwlock_t kclist_lock = RW_LOCK_UNLOCKED;
  58
  59 void
  60 kclist_add(struct kcore_list *new, void *addr, size_t size)
  61 {
  62         new->addr = (unsigned long)addr;
  63         new->size = size;
  64
  65         write_lock(&kclist_lock);
  66         new->next = kclist;
  67         kclist = new;
  68         write_unlock(&kclist_lock);
  69 }
  70
  71 struct kcore_list *
  72 kclist_del(void *addr)
  73 {
  74         struct kcore_list *m, **p = &kclist;
  75
  76         write_lock(&kclist_lock);
  77         for (m = *p; m; p = &m->next) {
  78                 if (m->addr == (unsigned long)addr) {
  79                         *p = m->next;
  80                         write_unlock(&kclist_lock);
  81                         return m;
  82                 }
  83         }
  84         write_unlock(&kclist_lock);
  85         return NULL;
  86 }
  87
  88 static size_t get_kcore_size(int *nphdr, size_t *elf_buflen)
  89 {
  90         size_t try, size;
  91         struct kcore_list *m;
  92
  93         *nphdr = 1; /* PT_NOTE */
  94         size = 0;
  95
  96         for (m=kclist; m; m=m->next) {
  97                 try = kc_vaddr_to_offset((size_t)m->addr + m->size);
  98                 if (try > size)
  99                         size = try;
 100                 *nphdr = *nphdr + 1;
 101         }
 102         *elf_buflen =   sizeof(struct elfhdr) +
 103                         (*nphdr + 2)*sizeof(struct elf_phdr) +
 104                         3 * sizeof(struct memelfnote) +
 105                         sizeof(struct elf_prstatus) +
 106                         sizeof(struct elf_prpsinfo) +
 107                         sizeof(struct task_struct);
 108         *elf_buflen = PAGE_ALIGN(*elf_buflen);
 109         return size + *elf_buflen;
 110 }
 111
 112
 113 /*****************************************************************************/
 114 /*
 115  * determine size of ELF note
 116  */
 117 static int notesize(struct memelfnote *en)
 118 {
 119         int sz;
 120
 121         sz = sizeof(struct elf_note);
 122         sz += roundup(strlen(en->name), 4);
 123         sz += roundup(en->datasz, 4);
 124
 125         return sz;
 126 } /* end notesize() */
 127
 128 /*****************************************************************************/
 129 /*
 130  * store a note in the header buffer
 131  */
 132 static char *storenote(struct memelfnote *men, char *bufp)
 133 {
 134         struct elf_note en;
 135
 136 #define DUMP_WRITE(addr,nr) do { memcpy(bufp,addr,nr); bufp += nr; } while(0)
 137
 138         en.n_namesz = strlen(men->name);
 139         en.n_descsz = men->datasz;
 140         en.n_type = men->type;
 141
 142         DUMP_WRITE(&en, sizeof(en));
 143         DUMP_WRITE(men->name, en.n_namesz);
 144
 145         /* XXX - cast from long long to long to avoid need for libgcc.a */
 146         bufp = (char*) roundup((unsigned long)bufp,4);
 147         DUMP_WRITE(men->data, men->datasz);
 148         bufp = (char*) roundup((unsigned long)bufp,4);
 149
 150 #undef DUMP_WRITE
 151
 152         return bufp;
 153 } /* end storenote() */
 154
 155 /*
 156  * store an ELF coredump header in the supplied buffer
 157  * nphdr is the number of elf_phdr to insert
 158  */
 159 static void elf_kcore_store_hdr(char *bufp, int nphdr, int dataoff)
 160 {
 161         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
 162         struct elf_prpsinfo prpsinfo;   /* NT_PRPSINFO */
 163         struct elf_phdr *nhdr, *phdr;
 164         struct elfhdr *elf;
 165         struct memelfnote notes[3];
 166         off_t offset = 0;
 167         struct kcore_list *m;
 168
 169         /* setup ELF header */
 170         elf = (struct elfhdr *) bufp;
 171         bufp += sizeof(struct elfhdr);
 172         offset += sizeof(struct elfhdr);
 173         memcpy(elf->e_ident, ELFMAG, SELFMAG);
 174         elf->e_ident[EI_CLASS]  = ELF_CLASS;
 175         elf->e_ident[EI_DATA]   = ELF_DATA;
 176         elf->e_ident[EI_VERSION]= EV_CURRENT;
 177         elf->e_ident[EI_OSABI] = ELF_OSABI;
 178         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
 179         elf->e_type     = ET_CORE;
 180         elf->e_machine  = ELF_ARCH;
 181         elf->e_version  = EV_CURRENT;
 182         elf->e_entry    = 0;
 183         elf->e_phoff    = sizeof(struct elfhdr);
 184         elf->e_shoff    = 0;
 185 #if defined(CONFIG_H8300)
 186         elf->e_flags    = ELF_FLAGS;
 187 #else
 188         elf->e_flags    = 0;
 189 #endif
 190         elf->e_ehsize   = sizeof(struct elfhdr);
 191         elf->e_phentsize= sizeof(struct elf_phdr);
 192         elf->e_phnum    = nphdr;
 193         elf->e_shentsize= 0;
 194         elf->e_shnum    = 0;
 195         elf->e_shstrndx = 0;
 196
 197         /* setup ELF PT_NOTE program header */
 198         nhdr = (struct elf_phdr *) bufp;
 199         bufp += sizeof(struct elf_phdr);
 200         offset += sizeof(struct elf_phdr);
 201         nhdr->p_type    = PT_NOTE;
 202         nhdr->p_offset  = 0;
 203         nhdr->p_vaddr   = 0;
 204         nhdr->p_paddr   = 0;
 205         nhdr->p_filesz  = 0;
 206         nhdr->p_memsz   = 0;
 207         nhdr->p_flags   = 0;
 208         nhdr->p_align   = 0;
 209
 210         /* setup ELF PT_LOAD program header for every area */
 211         for (m=kclist; m; m=m->next) {
 212                 phdr = (struct elf_phdr *) bufp;
 213                 bufp += sizeof(struct elf_phdr);
 214                 offset += sizeof(struct elf_phdr);
 215
 216                 phdr->p_type    = PT_LOAD;
 217                 phdr->p_flags   = PF_R|PF_W|PF_X;
 218                 phdr->p_offset  = kc_vaddr_to_offset(m->addr) + dataoff;
 219                 phdr->p_vaddr   = (size_t)m->addr;
 220                 phdr->p_paddr   = 0;
 221                 phdr->p_filesz  = phdr->p_memsz = m->size;
 222                 phdr->p_align   = PAGE_SIZE;
 223         }
 224
 225         /*
 226          * Set up the notes in similar form to SVR4 core dumps made
 227          * with info from their /proc.
 228          */
 229         nhdr->p_offset  = offset;
 230
 231         /* set up the process status */
 232         notes[0].name = "CORE";
 233         notes[0].type = NT_PRSTATUS;
 234         notes[0].datasz = sizeof(struct elf_prstatus);
 235         notes[0].data = &prstatus;
 236
 237         memset(&prstatus, 0, sizeof(struct elf_prstatus));
 238
 239         nhdr->p_filesz  = notesize(&notes[0]);
 240         bufp = storenote(&notes[0], bufp);
 241
 242         /* set up the process info */
 243         notes[1].name   = "CORE";
 244         notes[1].type   = NT_PRPSINFO;
 245         notes[1].datasz = sizeof(struct elf_prpsinfo);
 246         notes[1].data   = &prpsinfo;
 247
 248         memset(&prpsinfo, 0, sizeof(struct elf_prpsinfo));
 249         prpsinfo.pr_state       = 0;
 250         prpsinfo.pr_sname       = 'R';
 251         prpsinfo.pr_zomb        = 0;
 252
 253         strcpy(prpsinfo.pr_fname, "vmlinux");
 254         strncpy(prpsinfo.pr_psargs, saved_command_line, ELF_PRARGSZ);
 255
 256         nhdr->p_filesz  += notesize(&notes[1]);
 257         bufp = storenote(&notes[1], bufp);
 258
 259         /* set up the task structure */
 260         notes[2].name   = "CORE";
 261         notes[2].type   = NT_TASKSTRUCT;
 262         notes[2].datasz = sizeof(struct task_struct);
 263         notes[2].data   = current;
 264
 265         nhdr->p_filesz  += notesize(&notes[2]);
 266         bufp = storenote(&notes[2], bufp);
 267
 268 } /* end elf_kcore_store_hdr() */
 269
 270 /*****************************************************************************/
 271 /*
 272  * read from the ELF header and then kernel memory
 273  */
 274 static ssize_t
 275 read_kcore(struct file *file, char __user *buffer, size_t buflen, loff_t *fpos)
 276 {
 277         ssize_t acc = 0;
 278         size_t size, tsz;
 279         size_t elf_buflen;
 280         int nphdr;
 281         unsigned long start;
 282
 283         read_lock(&kclist_lock);
 284         tsz =  get_kcore_size(&nphdr, &elf_buflen);
 285         proc_root_kcore->size = size = tsz + elf_buflen;
 286         if (buflen == 0 || *fpos >= size) {
 287                 read_unlock(&kclist_lock);
 288                 return 0;
 289         }
 290
 291         /* trim buflen to not go beyond EOF */
 292         if (buflen > size - *fpos)
 293                 buflen = size - *fpos;
 294
 295         /* construct an ELF core header if we'll need some of it */
 296         if (*fpos < elf_buflen) {
 297                 char * elf_buf;
 298
 299                 tsz = elf_buflen - *fpos;
 300                 if (buflen < tsz)
 301                         tsz = buflen;
 302                 elf_buf = kmalloc(elf_buflen, GFP_ATOMIC);
 303                 if (!elf_buf) {
 304                         read_unlock(&kclist_lock);
 305                         return -ENOMEM;
 306                 }
 307                 memset(elf_buf, 0, elf_buflen);
 308                 elf_kcore_store_hdr(elf_buf, nphdr, elf_buflen);
 309                 read_unlock(&kclist_lock);
 310                 if (copy_to_user(buffer, elf_buf + *fpos, tsz)) {
 311                         kfree(elf_buf);
 312                         return -EFAULT;
 313                 }
 314                 kfree(elf_buf);
 315                 buflen -= tsz;
 316                 *fpos += tsz;
 317                 buffer += tsz;
 318                 acc += tsz;
 319
 320                 /* leave now if filled buffer already */
 321                 if (buflen == 0)
 322                         return acc;
 323         } else
 324                 read_unlock(&kclist_lock);
 325
 326         /*
 327          * Check to see if our file offset matches with any of
 328          * the addresses in the elf_phdr on our list.
 329          */
 330         start = kc_offset_to_vaddr(*fpos - elf_buflen);
 331         if ((tsz = (PAGE_SIZE - (start & ~PAGE_MASK))) > buflen)
 332                 tsz = buflen;
 333
 334         while (buflen) {
 335                 struct kcore_list *m;
 336
 337                 read_lock(&kclist_lock);
 338                 for (m=kclist; m; m=m->next) {
 339                         if (start >= m->addr && start < (m->addr+m->size))
 340                                 break;
 341                 }
 342                 read_unlock(&kclist_lock);
 343
 344                 if (m == NULL) {
 345                         if (clear_user(buffer, tsz))
 346                                 return -EFAULT;
 347                 } else if ((start >= VMALLOC_START) && (start < VMALLOC_END)) {
 348                         char * elf_buf;
 349                         struct vm_struct *m;
 350                         unsigned long curstart = start;
 351                         unsigned long cursize = tsz;
 352
 353                         elf_buf = kmalloc(tsz, GFP_KERNEL);
 354                         if (!elf_buf)
 355                                 return -ENOMEM;
 356                         memset(elf_buf, 0, tsz);
 357
 358                         read_lock(&vmlist_lock);
 359                         for (m=vmlist; m && cursize; m=m->next) {
 360                                 unsigned long vmstart;
 361                                 unsigned long vmsize;
 362                                 unsigned long msize = m->size - PAGE_SIZE;
 363
 364                                 if (((unsigned long)m->addr + msize) <
 365                                                                 curstart)
 366                                         continue;
 367                                 if ((unsigned long)m->addr > (curstart +
 368                                                                 cursize))
 369                                         break;
 370                                 vmstart = (curstart < (unsigned long)m->addr ?
 371                                         (unsigned long)m->addr : curstart);
 372                                 if (((unsigned long)m->addr + msize) >
 373                                                         (curstart + cursize))
 374                                         vmsize = curstart + cursize - vmstart;
 375                                 else
 376                                         vmsize = (unsigned long)m->addr +
 377                                                         msize - vmstart;
 378                                 curstart = vmstart + vmsize;
 379                                 cursize -= vmsize;
 380                                 /* don't dump ioremap'd stuff! (TA) */
 381                                 if (m->flags & VM_IOREMAP)
 382                                         continue;
 383                                 memcpy(elf_buf + (vmstart - start),
 384                                         (char *)vmstart, vmsize);
 385                         }
 386                         read_unlock(&vmlist_lock);
 387                         if (copy_to_user(buffer, elf_buf, tsz)) {
 388                                 kfree(elf_buf);
 389                                 return -EFAULT;
 390                         }
 391                         kfree(elf_buf);
 392                 } else {
 393                         if (kern_addr_valid(start)) {
 394                                 unsigned long n;
 395
 396                                 n = copy_to_user(buffer, (char *)start, tsz);
 397                                 /*
 398                                  * We cannot distingush between fault on source
 399                                  * and fault on destination. When this happens
 400                                  * we clear too and hope it will trigger the
 401                                  * EFAULT again.
 402                                  */
 403                                 if (n) {
 404                                         if (clear_user(buffer + tsz - n,
 405                                                                 tsz - n))
 406                                                 return -EFAULT;
 407                                 }
 408                         } else {
 409                                 if (clear_user(buffer, tsz))
 410                                         return -EFAULT;
 411                         }
 412                 }
 413                 buflen -= tsz;
 414                 *fpos += tsz;
 415                 buffer += tsz;
 416                 acc += tsz;
 417                 start += tsz;
 418                 tsz = (buflen > PAGE_SIZE ? PAGE_SIZE : buflen);
 419         }
 420
 421         return acc;
 422 }