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 loff_t lseek_kcore(struct file * file, loff_t offset, int origin);
  31
  32 static ssize_t read_kcore(struct file *, char *, size_t, loff_t *);
  33
  34 struct file_operations proc_kcore_operations = {
  35         read:           read_kcore,
  36         open:           open_kcore,
  37         llseek:         lseek_kcore,
  38 };
  39
  40 #ifdef CONFIG_KCORE_AOUT
  41 static ssize_t read_kcore(struct file *file, char *buf, size_t count, loff_t *ppos)
  42 {
  43         unsigned long long p = *ppos, memsize;
  44         ssize_t read;
  45         ssize_t count1;
  46         char * pnt;
  47         struct user dump;
  48 #if defined (__i386__) || defined (__mc68000__) || defined(__x86_64__)
  49 #       define FIRST_MAPPED     PAGE_SIZE       /* we don't have page 0 mapped on x86.. */
  50 #else
  51 #       define FIRST_MAPPED     0
  52 #endif
  53
  54         memset(&dump, 0, sizeof(struct user));
  55         dump.magic = CMAGIC;
  56         dump.u_dsize = (virt_to_phys(high_memory) >> PAGE_SHIFT);
  57 #if defined (__i386__) || defined(__x86_64__)
  58         dump.start_code = PAGE_OFFSET;
  59 #endif
  60 #ifdef __alpha__
  61         dump.start_data = PAGE_OFFSET;
  62 #endif
  63
  64         memsize = virt_to_phys(high_memory);
  65         if (p >= memsize)
  66                 return 0;
  67         if (count > memsize - p)
  68                 count = memsize - p;
  69         read = 0;
  70
  71         if (p < sizeof(struct user) && count > 0) {
  72                 count1 = count;
  73                 if (p + count1 > sizeof(struct user))
  74                         count1 = sizeof(struct user)-p;
  75                 pnt = (char *) &dump + p;
  76                 if (copy_to_user(buf,(void *) pnt, count1))
  77                         return -EFAULT;
  78                 buf += count1;
  79                 p += count1;
  80                 count -= count1;
  81                 read += count1;
  82         }
  83
  84         if (count > 0 && p < PAGE_SIZE + FIRST_MAPPED) {
  85                 count1 = PAGE_SIZE + FIRST_MAPPED - p;
  86                 if (count1 > count)
  87                         count1 = count;
  88                 if (clear_user(buf, count1))
  89                         return -EFAULT;
  90                 buf += count1;
  91                 p += count1;
  92                 count -= count1;
  93                 read += count1;
  94         }
  95         if (count > 0) {
  96                 if (copy_to_user(buf, (void *) (PAGE_OFFSET+p-PAGE_SIZE), count))
  97                         return -EFAULT;
  98                 read += count;
  99                 p += count;
 100         }
 101         *ppos = p;
 102         return read;
 103 }
 104 #else /* CONFIG_KCORE_AOUT */
 105
 106 #define roundup(x, y)  ((((x)+((y)-1))/(y))*(y))
 107
 108 /* An ELF note in memory */
 109 struct memelfnote
 110 {
 111         const char *name;
 112         int type;
 113         unsigned int datasz;
 114         void *data;
 115 };
 116
 117 extern char saved_command_line[];
 118
 119 static unsigned long get_kcore_size(int *num_vma, size_t *elf_buflen)
 120 {
 121         unsigned long try, size;
 122         struct vm_struct *m;
 123
 124         *num_vma = 0;
 125         size = ((size_t)high_memory - PAGE_OFFSET + PAGE_SIZE);
 126         if (!vmlist) {
 127                 *elf_buflen = PAGE_SIZE;
 128                 return (size);
 129         }
 130
 131         for (m=vmlist; m; m=m->next) {
 132                 try = (unsigned long)m->addr + m->size;
 133                 if (try > size)
 134                         size = try;
 135                 *num_vma = *num_vma + 1;
 136         }
 137         *elf_buflen =   sizeof(struct elfhdr) +
 138                         (*num_vma + 2)*sizeof(struct elf_phdr) +
 139                         3 * (sizeof(struct elf_note) + 4) +
 140                         sizeof(struct elf_prstatus) +
 141                         sizeof(struct elf_prpsinfo) +
 142                         sizeof(struct task_struct);
 143         *elf_buflen = PAGE_ALIGN(*elf_buflen);
 144         return (size - PAGE_OFFSET + *elf_buflen);
 145 }
 146
 147
 148 /*****************************************************************************/
 149 /*
 150  * determine size of ELF note
 151  */
 152 static int notesize(struct memelfnote *en)
 153 {
 154         int sz;
 155
 156         sz = sizeof(struct elf_note);
 157         sz += roundup(strlen(en->name), 4);
 158         sz += roundup(en->datasz, 4);
 159
 160         return sz;
 161 } /* end notesize() */
 162
 163 /*****************************************************************************/
 164 /*
 165  * store a note in the header buffer
 166  */
 167 static char *storenote(struct memelfnote *men, char *bufp)
 168 {
 169         struct elf_note en;
 170
 171 #define DUMP_WRITE(addr,nr) do { memcpy(bufp,addr,nr); bufp += nr; } while(0)
 172
 173         en.n_namesz = strlen(men->name);
 174         en.n_descsz = men->datasz;
 175         en.n_type = men->type;
 176
 177         DUMP_WRITE(&en, sizeof(en));
 178         DUMP_WRITE(men->name, en.n_namesz);
 179
 180         /* XXX - cast from long long to long to avoid need for libgcc.a */
 181         bufp = (char*) roundup((unsigned long)bufp,4);
 182         DUMP_WRITE(men->data, men->datasz);
 183         bufp = (char*) roundup((unsigned long)bufp,4);
 184
 185 #undef DUMP_WRITE
 186
 187         return bufp;
 188 } /* end storenote() */
 189
 190 /*
 191  * store an ELF coredump header in the supplied buffer
 192  * num_vma is the number of elements in vmlist
 193  */
 194 static void elf_kcore_store_hdr(char *bufp, int num_vma, int dataoff)
 195 {
 196         struct elf_prstatus prstatus;   /* NT_PRSTATUS */
 197         struct elf_prpsinfo prpsinfo;   /* NT_PRPSINFO */
 198         struct elf_phdr *nhdr, *phdr;
 199         struct elfhdr *elf;
 200         struct memelfnote notes[3];
 201         off_t offset = 0;
 202         struct vm_struct *m;
 203
 204         /* setup ELF header */
 205         elf = (struct elfhdr *) bufp;
 206         bufp += sizeof(struct elfhdr);
 207         offset += sizeof(struct elfhdr);
 208         memcpy(elf->e_ident, ELFMAG, SELFMAG);
 209         elf->e_ident[EI_CLASS]  = ELF_CLASS;
 210         elf->e_ident[EI_DATA]   = ELF_DATA;
 211         elf->e_ident[EI_VERSION]= EV_CURRENT;
 212         memset(elf->e_ident+EI_PAD, 0, EI_NIDENT-EI_PAD);
 213         elf->e_type     = ET_CORE;
 214         elf->e_machine  = ELF_ARCH;
 215         elf->e_version  = EV_CURRENT;
 216         elf->e_entry    = 0;
 217         elf->e_phoff    = sizeof(struct elfhdr);
 218         elf->e_shoff    = 0;
 219         elf->e_flags    = 0;
 220         elf->e_ehsize   = sizeof(struct elfhdr);
 221         elf->e_phentsize= sizeof(struct elf_phdr);
 222         elf->e_phnum    = 2 + num_vma;
 223         elf->e_shentsize= 0;
 224         elf->e_shnum    = 0;
 225         elf->e_shstrndx = 0;
 226
 227         /* setup ELF PT_NOTE program header */
 228         nhdr = (struct elf_phdr *) bufp;
 229         bufp += sizeof(struct elf_phdr);
 230         offset += sizeof(struct elf_phdr);
 231         nhdr->p_type    = PT_NOTE;
 232         nhdr->p_offset  = 0;
 233         nhdr->p_vaddr   = 0;
 234         nhdr->p_paddr   = 0;
 235         nhdr->p_filesz  = 0;
 236         nhdr->p_memsz   = 0;
 237         nhdr->p_flags   = 0;
 238         nhdr->p_align   = 0;
 239
 240         /* setup ELF PT_LOAD program header for the
 241          * virtual range 0xc0000000 -> high_memory */
 242         phdr = (struct elf_phdr *) bufp;
 243         bufp += sizeof(struct elf_phdr);
 244         offset += sizeof(struct elf_phdr);
 245         phdr->p_type    = PT_LOAD;
 246         phdr->p_flags   = PF_R|PF_W|PF_X;
 247         phdr->p_offset  = dataoff;
 248         phdr->p_vaddr   = PAGE_OFFSET;
 249         phdr->p_paddr   = __pa(PAGE_OFFSET);
 250         phdr->p_filesz  = phdr->p_memsz = ((unsigned long)high_memory - PAGE_OFFSET);
 251         phdr->p_align   = PAGE_SIZE;
 252
 253         /* setup ELF PT_LOAD program header for every vmalloc'd area */
 254         for (m=vmlist; m; m=m->next) {
 255                 if (m->flags & VM_IOREMAP) /* don't dump ioremap'd stuff! (TA) */
 256                         continue;
 257
 258                 phdr = (struct elf_phdr *) bufp;
 259                 bufp += sizeof(struct elf_phdr);
 260                 offset += sizeof(struct elf_phdr);
 261
 262                 phdr->p_type    = PT_LOAD;
 263                 phdr->p_flags   = PF_R|PF_W|PF_X;
 264                 phdr->p_offset  = (size_t)m->addr - PAGE_OFFSET + dataoff;
 265                 phdr->p_vaddr   = (size_t)m->addr;
 266                 phdr->p_paddr   = __pa(m->addr);
 267                 phdr->p_filesz  = phdr->p_memsz = m->size;
 268                 phdr->p_align   = PAGE_SIZE;
 269         }
 270
 271         /*
 272          * Set up the notes in similar form to SVR4 core dumps made
 273          * with info from their /proc.
 274          */
 275         nhdr->p_offset  = offset;
 276
 277         /* set up the process status */
 278         notes[0].name = "CORE";
 279         notes[0].type = NT_PRSTATUS;
 280         notes[0].datasz = sizeof(struct elf_prstatus);
 281         notes[0].data = &prstatus;
 282
 283         memset(&prstatus, 0, sizeof(struct elf_prstatus));
 284
 285         nhdr->p_filesz += notesize(&notes[0]);
 286         bufp = storenote(&notes[0], bufp);
 287
 288         /* set up the process info */
 289         notes[1].name   = "CORE";
 290         notes[1].type   = NT_PRPSINFO;
 291         notes[1].datasz = sizeof(struct elf_prpsinfo);
 292         notes[1].data   = &prpsinfo;
 293
 294         memset(&prpsinfo, 0, sizeof(struct elf_prpsinfo));
 295         prpsinfo.pr_state       = 0;
 296         prpsinfo.pr_sname       = 'R';
 297         prpsinfo.pr_zomb        = 0;
 298
 299         strcpy(prpsinfo.pr_fname, "vmlinux");
 300         strncpy(prpsinfo.pr_psargs, saved_command_line, ELF_PRARGSZ);
 301
 302         nhdr->p_filesz += notesize(&notes[1]);
 303         bufp = storenote(&notes[1], bufp);
 304
 305         /* set up the task structure */
 306         notes[2].name   = "CORE";
 307         notes[2].type   = NT_TASKSTRUCT;
 308         notes[2].datasz = sizeof(struct task_struct);
 309         notes[2].data   = current;
 310
 311         nhdr->p_filesz += notesize(&notes[2]);
 312         bufp = storenote(&notes[2], bufp);
 313
 314 } /* end elf_kcore_store_hdr() */
 315
 316 /*****************************************************************************/
 317 /*
 318  * read from the ELF header and then kernel memory
 319  */
 320 static ssize_t read_kcore(struct file *file, char *buffer, size_t buflen, loff_t *fpos)
 321 {
 322         ssize_t acc = 0;
 323         unsigned long size, tsz;
 324         size_t elf_buflen;
 325         int num_vma;
 326         unsigned long start;
 327
 328         read_lock(&vmlist_lock);
 329         proc_root_kcore->size = size = get_kcore_size(&num_vma, &elf_buflen);
 330         if (buflen == 0 || (unsigned long long)*fpos >= size) {
 331                 read_unlock(&vmlist_lock);
 332                 return 0;
 333         }
 334
 335         /* trim buflen to not go beyond EOF */
 336         if (buflen > size - *fpos)
 337                 buflen = size - *fpos;
 338
 339         /* construct an ELF core header if we'll need some of it */
 340         if (*fpos < elf_buflen) {
 341                 char * elf_buf;
 342
 343                 tsz = elf_buflen - *fpos;
 344                 if (buflen < tsz)
 345                         tsz = buflen;
 346                 elf_buf = kmalloc(elf_buflen, GFP_ATOMIC);
 347                 if (!elf_buf) {
 348                         read_unlock(&vmlist_lock);
 349                         return -ENOMEM;
 350                 }
 351                 memset(elf_buf, 0, elf_buflen);
 352                 elf_kcore_store_hdr(elf_buf, num_vma, elf_buflen);
 353                 read_unlock(&vmlist_lock);
 354                 if (copy_to_user(buffer, elf_buf + *fpos, tsz)) {
 355                         kfree(elf_buf);
 356                         return -EFAULT;
 357                 }
 358                 kfree(elf_buf);
 359                 buflen -= tsz;
 360                 *fpos += tsz;
 361                 buffer += tsz;
 362                 acc += tsz;
 363
 364                 /* leave now if filled buffer already */
 365                 if (buflen == 0)
 366                         return acc;
 367         } else
 368                 read_unlock(&vmlist_lock);
 369
 370         /* where page 0 not mapped, write zeros into buffer */
 371 #if defined (__i386__) || defined (__mc68000__) || defined(__x86_64__)
 372         if (*fpos < PAGE_SIZE + elf_buflen) {
 373                 /* work out how much to clear */
 374                 tsz = PAGE_SIZE + elf_buflen - *fpos;
 375                 if (buflen < tsz)
 376                         tsz = buflen;
 377
 378                 /* write zeros to buffer */
 379                 if (clear_user(buffer, tsz))
 380                         return -EFAULT;
 381                 buflen -= tsz;
 382                 *fpos += tsz;
 383                 buffer += tsz;
 384                 acc += tsz;
 385
 386                 /* leave now if filled buffer already */
 387                 if (buflen == 0)
 388                         return tsz;
 389         }
 390 #endif
 391
 392         /*
 393          * Fill the remainder of the buffer from kernel VM space.
 394          * We said in the ELF header that the data which starts
 395          * at 'elf_buflen' is virtual address PAGE_OFFSET. --rmk
 396          */
 397         start = PAGE_OFFSET + (*fpos - elf_buflen);
 398         if ((tsz = (PAGE_SIZE - (start & ~PAGE_MASK))) > buflen)
 399                 tsz = buflen;
 400         while (buflen) {
 401                 int err;
 402
 403                 if ((start > PAGE_OFFSET) && (start < (unsigned long)high_memory)) {
 404                         if (kern_addr_valid(start)) {
 405                                 err = copy_to_user(buffer, (char *)start, tsz);
 406                         } else {
 407                                 err = clear_user(buffer, tsz);
 408                         }
 409                 } else {
 410                         char * elf_buf;
 411                         struct vm_struct *m;
 412                         unsigned long curstart = start;
 413                         unsigned long cursize = tsz;
 414
 415                         elf_buf = kmalloc(tsz, GFP_KERNEL);
 416                         if (!elf_buf)
 417                                 return -ENOMEM;
 418                         memset(elf_buf, 0, tsz);
 419
 420                         read_lock(&vmlist_lock);
 421                         for (m=vmlist; m && cursize; m=m->next) {
 422                                 unsigned long vmstart;
 423                                 unsigned long vmsize;
 424                                 unsigned long msize = m->size - PAGE_SIZE;
 425
 426                                 if (((unsigned long)m->addr + msize) <
 427                                                                 curstart)
 428                                         continue;
 429                                 if ((unsigned long)m->addr > (curstart +
 430                                                                 cursize))
 431                                         break;
 432                                 vmstart = (curstart < (unsigned long)m->addr ?
 433                                         (unsigned long)m->addr : curstart);
 434                                 if (((unsigned long)m->addr + msize) >
 435                                                         (curstart + cursize))
 436                                         vmsize = curstart + cursize - vmstart;
 437                                 else
 438                                         vmsize = (unsigned long)m->addr +
 439                                                         msize - vmstart;
 440                                 curstart = vmstart + vmsize;
 441                                 cursize -= vmsize;
 442                                 /* don't dump ioremap'd stuff! (TA) */
 443                                 if (m->flags & VM_IOREMAP)
 444                                         continue;
 445                                 memcpy(elf_buf + (vmstart - start),
 446                                         (char *)vmstart, vmsize);
 447                         }
 448                         read_unlock(&vmlist_lock);
 449                         err = copy_to_user(buffer, elf_buf, tsz);
 450                                 kfree(elf_buf);
 451                         }
 452                 if (err)
 453                                         return -EFAULT;
 454                 buflen -= tsz;
 455                 *fpos += tsz;
 456                 buffer += tsz;
 457                 acc += tsz;
 458                 start += tsz;
 459                 tsz = (buflen > PAGE_SIZE ? PAGE_SIZE : buflen);
 460         }
 461
 462         return acc;
 463 }
 464 #endif /* CONFIG_KCORE_AOUT */
 465
 466 static loff_t lseek_kcore(struct file * file, loff_t offset, int origin)
 467 {
 468         switch (origin) {
 469                 case 2:
 470                         offset += file->f_dentry->d_inode->i_size;
 471                         break;
 472                 case 1:
 473                         offset += file->f_pos;
 474         }
 475         /* RED-PEN user can fake an error here by setting offset to >=-4095 && <0  */
 476         file->f_pos = offset;
 477         return offset;
 478 }