fs/proc/kcore.c

   1 /*
   2  *      fs/proc/kcore.c kernel ELF/AOUT 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 <asm/uaccess.h>
  22 #include <asm/io.h>
  23
  24
  25 static int open_kcore(struct inode * inode, struct file * filp)
  26 {
  27         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
  28 }
  29
  30 static ssize_t read_kcore(struct file *, char *, size_t, loff_t *);
  31
  32 struct file_operations proc_kcore_operations = {
  33         read:           read_kcore,
  34         open:           open_kcore,
  35 };
  36
  37 #ifdef CONFIG_KCORE_AOUT
  38 static ssize_t read_kcore(struct file *file, char *buf, size_t count, loff_t *ppos)
  39 {
  40         unsigned long long p = *ppos, memsize;
  41         ssize_t read;
  42         ssize_t count1;
  43         char * pnt;
  44         struct user dump;
  45 #if defined (__i386__) || defined (__mc68000__) || defined(__x86_64__)
  46 #       define FIRST_MAPPED     PAGE_SIZE       /* we don't have page 0 mapped on x86.. */
  47 #else
  48 #       define FIRST_MAPPED     0
  49 #endif
  50
  51         memset(&dump, 0, sizeof(struct user));
  52         dump.magic = CMAGIC;
  53         dump.u_dsize = (virt_to_phys(high_memory) >> PAGE_SHIFT);
  54 #if defined (__i386__) || defined(__x86_64__)
  55         dump.start_code = PAGE_OFFSET;
  56 #endif
  57 #ifdef __alpha__
  58         dump.start_data = PAGE_OFFSET;
  59 #endif
  60
  61         memsize = virt_to_phys(high_memory);
  62         if (p >= memsize)
  63                 return 0;
  64         if (count > memsize - p)
  65                 count = memsize - p;
  66         read = 0;
  67
  68         if (p < sizeof(struct user) && count > 0) {
  69                 count1 = count;
  70                 if (p + count1 > sizeof(struct user))
  71                         count1 = sizeof(struct user)-p;
  72                 pnt = (char *) &dump + p;
  73                 if (copy_to_user(buf,(void *) pnt, count1))
  74                         return -EFAULT;
  75                 buf += count1;
  76                 p += count1;
  77                 count -= count1;
  78                 read += count1;
  79         }
  80
  81         if (count > 0 && p < PAGE_SIZE + FIRST_MAPPED) {
  82                 count1 = PAGE_SIZE + FIRST_MAPPED - p;
  83                 if (count1 > count)
  84                         count1 = count;
  85                 if (clear_user(buf, count1))
  86                         return -EFAULT;
  87                 buf += count1;
  88                 p += count1;
  89                 count -= count1;
  90                 read += count1;
  91         }
  92         if (count > 0) {
  93                 if (copy_to_user(buf, (void *) (PAGE_OFFSET+p-PAGE_SIZE), count))
  94                         return -EFAULT;
  95                 read += count;
  96         }
  97         *ppos += read;
  98         return read;
  99 }
 100 #else /* CONFIG_KCORE_AOUT */
 101
 102 #define roundup(x, y)  ((((x)+((y)-1))/(y))*(y))
 103
 104 /* An ELF note in memory */
 105 struct memelfnote
 106 {
 107         const char *name;
 108         int type;
 109         unsigned int datasz;
 110         void *data;
 111 };
 112
 113 extern char saved_command_line[];
 114
 115 static size_t get_kcore_size(int *num_vma, size_t *elf_buflen)
 116 {
 117         size_t try, size;
 118         struct vm_struct *m;
 119
 120         *num_vma = 0;
 121         size = ((size_t)high_memory - PAGE_OFFSET + PAGE_SIZE);
 122         if (!vmlist) {
 123                 *elf_buflen = PAGE_SIZE;
 124                 return (size);
 125         }
 126
 127         for (m=vmlist; m; m=m->next) {
 128                 try = (size_t)m->addr + m->size;
 129                 if (try > size)
 130                         size = try;
 131                 *num_vma = *num_vma + 1;
 132         }
 133         *elf_buflen =   sizeof(struct elfhdr) +
 134                         (*num_vma + 2)*sizeof(struct elf_phdr) +
 135                         3 * sizeof(struct memelfnote);
 136         *elf_buflen = PAGE_ALIGN(*elf_buflen);
 137         return (size - PAGE_OFFSET + *elf_buflen);
 138 }
 139
 140
 141 /*****************************************************************************/
 142 /*
 143  * determine size of ELF note
 144  */
 145 static int notesize(struct memelfnote *en)
 146 {
 147         int sz;
 148
 149         sz = sizeof(struct elf_note);
 150         sz += roundup(strlen(en->name), 4);
 151         sz += roundup(en->datasz, 4);
 152
 153         return sz;
 154 } /* end notesize() */
 155
 156 /*****************************************************************************/
 157 /*
 158  * store a note in the header buffer
 159  */
 160 static char *storenote(struct memelfnote *men, char *bufp)
 161 {
 162         struct elf_note en;
 163
 164 #define DUMP_WRITE(addr,nr) do { memcpy(bufp,addr,nr); bufp += nr; } while(0)
 165
 166         en.n_namesz = strlen(men->name);
 167         en.n_descsz = men->datasz;
 168         en.n_type = men->type;
 169
 170         DUMP_WRITE(&en, sizeof(en));
 171         DUMP_WRITE(men->name, en.n_namesz);
 172
 173         /* XXX - cast from long long to long to avoid need for libgcc.a */
 174         bufp = (char*) roundup((unsigned long)bufp,4);
 175         DUMP_WRITE(men->data, men->datasz);
 176         bufp = (char*) roundup((unsigned long)bufp,4);
 177
 178 #undef DUMP_WRITE
 179
 180         return bufp;
 181 } /* end storenote() */
 182
 183 /*
 184  * store an ELF coredump header in the supplied buffer
 185  * num_vma is the number of elements in vmlist
 186  */
 187 static void elf_kcore_store_hdr(char *bufp, int num_vma, int dataoff)
 188 {
 189         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
 190         struct elf_prpsinfo prpsinfo;   /* NT_PRPSINFO */
 191         struct elf_phdr *nhdr, *phdr;
 192         struct elfhdr *elf;
 193         struct memelfnote notes[3];
 194         off_t offset = 0;
 195         struct vm_struct *m;
 196
 197         /* setup ELF header */
 198         elf = (struct elfhdr *) bufp;
 199         bufp += sizeof(struct elfhdr);
 200         offset += sizeof(struct elfhdr);
 201         memcpy(elf->e_ident, ELFMAG, SELFMAG);
 202         elf->e_ident[EI_CLASS]  = ELF_CLASS;
 203         elf->e_ident[EI_DATA]   = ELF_DATA;
 204         elf->e_ident[EI_VERSION]= EV_CURRENT;
 205         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
 206         elf->e_type     = ET_CORE;
 207         elf->e_machine  = ELF_ARCH;
 208         elf->e_version  = EV_CURRENT;
 209         elf->e_entry    = 0;
 210         elf->e_phoff    = sizeof(struct elfhdr);
 211         elf->e_shoff    = 0;
 212         elf->e_flags    = 0;
 213         elf->e_ehsize   = sizeof(struct elfhdr);
 214         elf->e_phentsize= sizeof(struct elf_phdr);
 215         elf->e_phnum    = 2 + num_vma;
 216         elf->e_shentsize= 0;
 217         elf->e_shnum    = 0;
 218         elf->e_shstrndx = 0;
 219
 220         /* setup ELF PT_NOTE program header */
 221         nhdr = (struct elf_phdr *) bufp;
 222         bufp += sizeof(struct elf_phdr);
 223         offset += sizeof(struct elf_phdr);
 224         nhdr->p_type    = PT_NOTE;
 225         nhdr->p_offset  = 0;
 226         nhdr->p_vaddr   = 0;
 227         nhdr->p_paddr   = 0;
 228         nhdr->p_filesz  = 0;
 229         nhdr->p_memsz   = 0;
 230         nhdr->p_flags   = 0;
 231         nhdr->p_align   = 0;
 232
 233         /* setup ELF PT_LOAD program header for the
 234          * virtual range 0xc0000000 -> high_memory */
 235         phdr = (struct elf_phdr *) bufp;
 236         bufp += sizeof(struct elf_phdr);
 237         offset += sizeof(struct elf_phdr);
 238         phdr->p_type    = PT_LOAD;
 239         phdr->p_flags   = PF_R|PF_W|PF_X;
 240         phdr->p_offset  = dataoff;
 241         phdr->p_vaddr   = PAGE_OFFSET;
 242         phdr->p_paddr   = __pa(PAGE_OFFSET);
 243         phdr->p_filesz  = phdr->p_memsz = ((unsigned long)high_memory - PAGE_OFFSET);
 244         phdr->p_align   = PAGE_SIZE;
 245
 246         /* setup ELF PT_LOAD program header for every vmalloc'd area */
 247         for (m=vmlist; m; m=m->next) {
 248                 if (m->flags & VM_IOREMAP) /* don't dump ioremap'd stuff! (TA) */
 249                         continue;
 250
 251                 phdr = (struct elf_phdr *) bufp;
 252                 bufp += sizeof(struct elf_phdr);
 253                 offset += sizeof(struct elf_phdr);
 254
 255                 phdr->p_type    = PT_LOAD;
 256                 phdr->p_flags   = PF_R|PF_W|PF_X;
 257                 phdr->p_offset  = (size_t)m->addr - PAGE_OFFSET + dataoff;
 258                 phdr->p_vaddr   = (size_t)m->addr;
 259                 phdr->p_paddr   = __pa(m->addr);
 260                 phdr->p_filesz  = phdr->p_memsz = m->size;
 261                 phdr->p_align   = PAGE_SIZE;
 262         }
 263
 264         /*
 265          * Set up the notes in similar form to SVR4 core dumps made
 266          * with info from their /proc.
 267          */
 268         nhdr->p_offset  = offset;
 269
 270         /* set up the process status */
 271         notes[0].name = "CORE";
 272         notes[0].type = NT_PRSTATUS;
 273         notes[0].datasz = sizeof(struct elf_prstatus);
 274         notes[0].data = &prstatus;
 275
 276         memset(&prstatus, 0, sizeof(struct elf_prstatus));
 277
 278         nhdr->p_filesz  = notesize(&notes[0]);
 279         bufp = storenote(&notes[0], bufp);
 280
 281         /* set up the process info */
 282         notes[1].name   = "CORE";
 283         notes[1].type   = NT_PRPSINFO;
 284         notes[1].datasz = sizeof(struct elf_prpsinfo);
 285         notes[1].data   = &prpsinfo;
 286
 287         memset(&prpsinfo, 0, sizeof(struct elf_prpsinfo));
 288         prpsinfo.pr_state       = 0;
 289         prpsinfo.pr_sname       = 'R';
 290         prpsinfo.pr_zomb        = 0;
 291
 292         strcpy(prpsinfo.pr_fname, "vmlinux");
 293         strncpy(prpsinfo.pr_psargs, saved_command_line, ELF_PRARGSZ);
 294
 295         nhdr->p_filesz  = notesize(&notes[1]);
 296         bufp = storenote(&notes[1], bufp);
 297
 298         /* set up the task structure */
 299         notes[2].name   = "CORE";
 300         notes[2].type   = NT_TASKSTRUCT;
 301         notes[2].datasz = sizeof(struct task_struct);
 302         notes[2].data   = current;
 303
 304         nhdr->p_filesz  = notesize(&notes[2]);
 305         bufp = storenote(&notes[2], bufp);
 306
 307 } /* end elf_kcore_store_hdr() */
 308
 309 /*****************************************************************************/
 310 /*
 311  * read from the ELF header and then kernel memory
 312  */
 313 static ssize_t read_kcore(struct file *file, char *buffer, size_t buflen, loff_t *fpos)
 314 {
 315         ssize_t acc = 0;
 316         size_t size, tsz;
 317         size_t elf_buflen;
 318         int num_vma;
 319         unsigned long start;
 320
 321         read_lock(&vmlist_lock);
 322         proc_root_kcore->size = size = get_kcore_size(&num_vma, &elf_buflen);
 323         if (buflen == 0 || *fpos >= size) {
 324                 read_unlock(&vmlist_lock);
 325                 return 0;
 326         }
 327
 328         /* trim buflen to not go beyond EOF */
 329         if (buflen > size - *fpos)
 330                 buflen = size - *fpos;
 331
 332         /* construct an ELF core header if we'll need some of it */
 333         if (*fpos < elf_buflen) {
 334                 char * elf_buf;
 335
 336                 tsz = elf_buflen - *fpos;
 337                 if (buflen < tsz)
 338                         tsz = buflen;
 339                 elf_buf = kmalloc(elf_buflen, GFP_ATOMIC);
 340                 if (!elf_buf) {
 341                         read_unlock(&vmlist_lock);
 342                         return -ENOMEM;
 343                 }
 344                 memset(elf_buf, 0, elf_buflen);
 345                 elf_kcore_store_hdr(elf_buf, num_vma, elf_buflen);
 346                 read_unlock(&vmlist_lock);
 347                 if (copy_to_user(buffer, elf_buf + *fpos, tsz)) {
 348                         kfree(elf_buf);
 349                         return -EFAULT;
 350                 }
 351                 kfree(elf_buf);
 352                 buflen -= tsz;
 353                 *fpos += tsz;
 354                 buffer += tsz;
 355                 acc += tsz;
 356
 357                 /* leave now if filled buffer already */
 358                 if (buflen == 0)
 359                         return acc;
 360         } else
 361                 read_unlock(&vmlist_lock);
 362
 363         /* where page 0 not mapped, write zeros into buffer */
 364 #if defined (__i386__) || defined (__mc68000__) || defined(__x86_64__)
 365         if (*fpos < PAGE_SIZE + elf_buflen) {
 366                 /* work out how much to clear */
 367                 tsz = PAGE_SIZE + elf_buflen - *fpos;
 368                 if (buflen < tsz)
 369                         tsz = buflen;
 370
 371                 /* write zeros to buffer */
 372                 if (clear_user(buffer, tsz))
 373                         return -EFAULT;
 374                 buflen -= tsz;
 375                 *fpos += tsz;
 376                 buffer += tsz;
 377                 acc += tsz;
 378
 379                 /* leave now if filled buffer already */
 380                 if (buflen == 0)
 381                         return tsz;
 382         }
 383 #endif
 384
 385         /*
 386          * Fill the remainder of the buffer from kernel VM space.
 387          * We said in the ELF header that the data which starts
 388          * at 'elf_buflen' is virtual address PAGE_OFFSET. --rmk
 389          */
 390         start = PAGE_OFFSET + (*fpos - elf_buflen);
 391         if ((tsz = (PAGE_SIZE - (start & ~PAGE_MASK))) > buflen)
 392                 tsz = buflen;
 393
 394         while (buflen) {
 395                 if ((start >= VMALLOC_START) && (start < VMALLOC_END)) {
 396                         char * elf_buf;
 397                         struct vm_struct *m;
 398                         unsigned long curstart = start;
 399                         unsigned long cursize = tsz;
 400
 401                         elf_buf = kmalloc(tsz, GFP_KERNEL);
 402                         if (!elf_buf)
 403                                 return -ENOMEM;
 404                         memset(elf_buf, 0, tsz);
 405
 406                         read_lock(&vmlist_lock);
 407                         for (m=vmlist; m && cursize; m=m->next) {
 408                                 unsigned long vmstart;
 409                                 unsigned long vmsize;
 410                                 unsigned long msize = m->size - PAGE_SIZE;
 411
 412                                 if (((unsigned long)m->addr + msize) <
 413                                                                 curstart)
 414                                         continue;
 415                                 if ((unsigned long)m->addr > (curstart +
 416                                                                 cursize))
 417                                         break;
 418                                 vmstart = (curstart < (unsigned long)m->addr ?
 419                                         (unsigned long)m->addr : curstart);
 420                                 if (((unsigned long)m->addr + msize) >
 421                                                         (curstart + cursize))
 422                                         vmsize = curstart + cursize - vmstart;
 423                                 else
 424                                         vmsize = (unsigned long)m->addr +
 425                                                         msize - vmstart;
 426                                 curstart = vmstart + vmsize;
 427                                 cursize -= vmsize;
 428                                 /* don't dump ioremap'd stuff! (TA) */
 429                                 if (m->flags & VM_IOREMAP)
 430                                         continue;
 431                                 memcpy(elf_buf + (vmstart - start),
 432                                         (char *)vmstart, vmsize);
 433                         }
 434                         read_unlock(&vmlist_lock);
 435                         if (copy_to_user(buffer, elf_buf, tsz)) {
 436                                 kfree(elf_buf);
 437                                 return -EFAULT;
 438                         }
 439                         kfree(elf_buf);
 440                 } else if ((start > PAGE_OFFSET) && (start <
 441                                                 (unsigned long)high_memory)) {
 442                         if (kern_addr_valid(start)) {
 443                                 if (copy_to_user(buffer, (char *)start, tsz))
 444                                         return -EFAULT;
 445                         } else {
 446                                 if (clear_user(buffer, tsz))
 447                                         return -EFAULT;
 448                         }
 449                 } else {
 450                         if (clear_user(buffer, tsz))
 451                                 return -EFAULT;
 452                 }
 453                 buflen -= tsz;
 454                 *fpos += tsz;
 455                 buffer += tsz;
 456                 acc += tsz;
 457                 start += tsz;
 458                 tsz = (buflen > PAGE_SIZE ? PAGE_SIZE : buflen);
 459         }
 460
 461         return acc;
 462 }
 463 #endif /* CONFIG_KCORE_AOUT */