fs/select.c

   1 /*
   2  * This file contains the procedures for the handling of select and poll
   3  *
   4  * Created for Linux based loosely upon Mathius Lattner's minix
   5  * patches by Peter MacDonald. Heavily edited by Linus.
   6  *
   7  *  4 February 1994
   8  *     COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS
   9  *     flag set in its personality we do *not* modify the given timeout
  10  *     parameter to reflect time remaining.
  11  *
  12  *  24 January 2000
  13  *     Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation
  14  *     of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian).
  15  */
  16
  17 #include <linux/syscalls.h>
  18 #include <linux/module.h>
  19 #include <linux/slab.h>
  20 #include <linux/smp_lock.h>
  21 #include <linux/poll.h>
  22 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
  23 #include <linux/file.h>
  24 #include <linux/fs.h>
  25
  26 #include <asm/uaccess.h>
  27
  28 #define ROUND_UP(x,y) (((x)+(y)-1)/(y))
  29 #define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
  30
  31 struct poll_table_entry {
  32         struct file * filp;
  33         wait_queue_t wait;
  34         wait_queue_head_t * wait_address;
  35 };
  36
  37 struct poll_table_page {
  38         struct poll_table_page * next;
  39         struct poll_table_entry * entry;
  40         struct poll_table_entry entries[0];
  41 };
  42
  43 #define POLL_TABLE_FULL(table) \
  44         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  45
  46 /*
  47  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
  48  * I have rewritten this, taking some shortcuts: This code may not be easy to
  49  * follow, but it should be free of race-conditions, and it's practical. If you
  50  * understand what I'm doing here, then you understand how the linux
  51  * sleep/wakeup mechanism works.
  52  *
  53  * Two very simple procedures, poll_wait() and poll_freewait() make all the
  54  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
  55  * as all select/poll functions have to call it to add an entry to the
  56  * poll table.
  57  */
  58 void __pollwait(struct file *filp, wait_queue_head_t *wait_address, poll_table *p);
  59
  60 void poll_initwait(struct poll_wqueues *pwq)
  61 {
  62         init_poll_funcptr(&pwq->pt, __pollwait);
  63         pwq->error = 0;
  64         pwq->table = NULL;
  65 }
  66
  67 EXPORT_SYMBOL(poll_initwait);
  68
  69 void poll_freewait(struct poll_wqueues *pwq)
  70 {
  71         struct poll_table_page * p = pwq->table;
  72         while (p) {
  73                 struct poll_table_entry * entry;
  74                 struct poll_table_page *old;
  75
  76                 entry = p->entry;
  77                 do {
  78                         entry--;
  79                         remove_wait_queue(entry->wait_address,&entry->wait);
  80                         fput(entry->filp);
  81                 } while (entry > p->entries);
  82                 old = p;
  83                 p = p->next;
  84                 free_page((unsigned long) old);
  85         }
  86 }
  87
  88 EXPORT_SYMBOL(poll_freewait);
  89
  90 void __pollwait(struct file *filp, wait_queue_head_t *wait_address, poll_table *_p)
  91 {
  92         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
  93         struct poll_table_page *table = p->table;
  94
  95         if (!table || POLL_TABLE_FULL(table)) {
  96                 struct poll_table_page *new_table;
  97
  98                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
  99                 if (!new_table) {
 100                         p->error = -ENOMEM;
 101                         __set_current_state(TASK_RUNNING);
 102                         return;
 103                 }
 104                 new_table->entry = new_table->entries;
 105                 new_table->next = table;
 106                 p->table = new_table;
 107                 table = new_table;
 108         }
 109
 110         /* Add a new entry */
 111         {
 112                 struct poll_table_entry * entry = table->entry;
 113                 table->entry = entry+1;
 114                 get_file(filp);
 115                 entry->filp = filp;
 116                 entry->wait_address = wait_address;
 117                 init_waitqueue_entry(&entry->wait, current);
 118                 add_wait_queue(wait_address,&entry->wait);
 119         }
 120 }
 121
 122 #define FDS_IN(fds, n)          (fds->in + n)
 123 #define FDS_OUT(fds, n)         (fds->out + n)
 124 #define FDS_EX(fds, n)          (fds->ex + n)
 125
 126 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 127
 128 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 129 {
 130         unsigned long *open_fds;
 131         unsigned long set;
 132         int max;
 133
 134         /* handle last in-complete long-word first */
 135         set = ~(~0UL << (n & (__NFDBITS-1)));
 136         n /= __NFDBITS;
 137         open_fds = current->files->open_fds->fds_bits+n;
 138         max = 0;
 139         if (set) {
 140                 set &= BITS(fds, n);
 141                 if (set) {
 142                         if (!(set & ~*open_fds))
 143                                 goto get_max;
 144                         return -EBADF;
 145                 }
 146         }
 147         while (n) {
 148                 open_fds--;
 149                 n--;
 150                 set = BITS(fds, n);
 151                 if (!set)
 152                         continue;
 153                 if (set & ~*open_fds)
 154                         return -EBADF;
 155                 if (max)
 156                         continue;
 157 get_max:
 158                 do {
 159                         max++;
 160                         set >>= 1;
 161                 } while (set);
 162                 max += n * __NFDBITS;
 163         }
 164
 165         return max;
 166 }
 167
 168 #define BIT(i)          (1UL << ((i)&(__NFDBITS-1)))
 169 #define MEM(i,m)        ((m)+(unsigned)(i)/__NFDBITS)
 170 #define ISSET(i,m)      (((i)&*(m)) != 0)
 171 #define SET(i,m)        (*(m) |= (i))
 172
 173 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 174 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 175 #define POLLEX_SET (POLLPRI)
 176
 177 int do_select(int n, fd_set_bits *fds, long *timeout)
 178 {
 179         struct poll_wqueues table;
 180         poll_table *wait;
 181         int retval, i;
 182         long __timeout = *timeout;
 183
 184         spin_lock(&current->files->file_lock);
 185         retval = max_select_fd(n, fds);
 186         spin_unlock(&current->files->file_lock);
 187
 188         if (retval < 0)
 189                 return retval;
 190         n = retval;
 191
 192         poll_initwait(&table);
 193         wait = &table.pt;
 194         if (!__timeout)
 195                 wait = NULL;
 196         retval = 0;
 197         for (;;) {
 198                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 199
 200                 set_current_state(TASK_INTERRUPTIBLE);
 201
 202                 inp = fds->in; outp = fds->out; exp = fds->ex;
 203                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 204
 205                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 206                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 207                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 208                         struct file_operations *f_op = NULL;
 209                         struct file *file = NULL;
 210
 211                         in = *inp++; out = *outp++; ex = *exp++;
 212                         all_bits = in | out | ex;
 213                         if (all_bits == 0) {
 214                                 i += __NFDBITS;
 215                                 continue;
 216                         }
 217
 218                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 219                                 if (i >= n)
 220                                         break;
 221                                 if (!(bit & all_bits))
 222                                         continue;
 223                                 file = fget(i);
 224                                 if (file) {
 225                                         f_op = file->f_op;
 226                                         mask = DEFAULT_POLLMASK;
 227                                         if (f_op && f_op->poll)
 228                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 229                                         fput(file);
 230                                         if ((mask & POLLIN_SET) && (in & bit)) {
 231                                                 res_in |= bit;
 232                                                 retval++;
 233                                         }
 234                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 235                                                 res_out |= bit;
 236                                                 retval++;
 237                                         }
 238                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 239                                                 res_ex |= bit;
 240                                                 retval++;
 241                                         }
 242                                 }
 243                         }
 244                         if (res_in)
 245                                 *rinp = res_in;
 246                         if (res_out)
 247                                 *routp = res_out;
 248                         if (res_ex)
 249                                 *rexp = res_ex;
 250                 }
 251                 wait = NULL;
 252                 if (retval || !__timeout || signal_pending(current))
 253                         break;
 254                 if(table.error) {
 255                         retval = table.error;
 256                         break;
 257                 }
 258                 __timeout = schedule_timeout(__timeout);
 259         }
 260         __set_current_state(TASK_RUNNING);
 261
 262         poll_freewait(&table);
 263
 264         /*
 265          * Up-to-date the caller timeout.
 266          */
 267         *timeout = __timeout;
 268         return retval;
 269 }
 270
 271 static void *select_bits_alloc(int size)
 272 {
 273         return kmalloc(6 * size, GFP_KERNEL);
 274 }
 275
 276 static void select_bits_free(void *bits, int size)
 277 {
 278         kfree(bits);
 279 }
 280
 281 /*
 282  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 283  * like to be certain this leads to no problems. So I return
 284  * EINTR just for safety.
 285  *
 286  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 287  * I'm trying ERESTARTNOHAND which restart only when you want to.
 288  */
 289 #define MAX_SELECT_SECONDS \
 290         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 291
 292 asmlinkage long
 293 sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, struct timeval __user *tvp)
 294 {
 295         fd_set_bits fds;
 296         char *bits;
 297         long timeout;
 298         int ret, size, max_fdset;
 299
 300         timeout = MAX_SCHEDULE_TIMEOUT;
 301         if (tvp) {
 302                 time_t sec, usec;
 303
 304                 if ((ret = verify_area(VERIFY_READ, tvp, sizeof(*tvp)))
 305                     || (ret = __get_user(sec, &tvp->tv_sec))
 306                     || (ret = __get_user(usec, &tvp->tv_usec)))
 307                         goto out_nofds;
 308
 309                 ret = -EINVAL;
 310                 if (sec < 0 || usec < 0)
 311                         goto out_nofds;
 312
 313                 if ((unsigned long) sec < MAX_SELECT_SECONDS) {
 314                         timeout = ROUND_UP(usec, 1000000/HZ);
 315                         timeout += sec * (unsigned long) HZ;
 316                 }
 317         }
 318
 319         ret = -EINVAL;
 320         if (n < 0)
 321                 goto out_nofds;
 322
 323         /* max_fdset can increase, so grab it once to avoid race */
 324         max_fdset = current->files->max_fdset;
 325         if (n > max_fdset)
 326                 n = max_fdset;
 327
 328         /*
 329          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 330          * since we used fdset we need to allocate memory in units of
 331          * long-words.
 332          */
 333         ret = -ENOMEM;
 334         size = FDS_BYTES(n);
 335         bits = select_bits_alloc(size);
 336         if (!bits)
 337                 goto out_nofds;
 338         fds.in      = (unsigned long *)  bits;
 339         fds.out     = (unsigned long *) (bits +   size);
 340         fds.ex      = (unsigned long *) (bits + 2*size);
 341         fds.res_in  = (unsigned long *) (bits + 3*size);
 342         fds.res_out = (unsigned long *) (bits + 4*size);
 343         fds.res_ex  = (unsigned long *) (bits + 5*size);
 344
 345         if ((ret = get_fd_set(n, inp, fds.in)) ||
 346             (ret = get_fd_set(n, outp, fds.out)) ||
 347             (ret = get_fd_set(n, exp, fds.ex)))
 348                 goto out;
 349         zero_fd_set(n, fds.res_in);
 350         zero_fd_set(n, fds.res_out);
 351         zero_fd_set(n, fds.res_ex);
 352
 353         ret = do_select(n, &fds, &timeout);
 354
 355         if (tvp && !(current->personality & STICKY_TIMEOUTS)) {
 356                 time_t sec = 0, usec = 0;
 357                 if (timeout) {
 358                         sec = timeout / HZ;
 359                         usec = timeout % HZ;
 360                         usec *= (1000000/HZ);
 361                 }
 362                 put_user(sec, &tvp->tv_sec);
 363                 put_user(usec, &tvp->tv_usec);
 364         }
 365
 366         if (ret < 0)
 367                 goto out;
 368         if (!ret) {
 369                 ret = -ERESTARTNOHAND;
 370                 if (signal_pending(current))
 371                         goto out;
 372                 ret = 0;
 373         }
 374
 375         if (set_fd_set(n, inp, fds.res_in) ||
 376             set_fd_set(n, outp, fds.res_out) ||
 377             set_fd_set(n, exp, fds.res_ex))
 378                 ret = -EFAULT;
 379
 380 out:
 381         select_bits_free(bits, size);
 382 out_nofds:
 383         return ret;
 384 }
 385
 386 struct poll_list {
 387         struct poll_list *next;
 388         int len;
 389         struct pollfd entries[0];
 390 };
 391
 392 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 393
 394 static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 395         poll_table ** pwait, int *count)
 396 {
 397         int i;
 398
 399         for (i = 0; i < num; i++) {
 400                 int fd;
 401                 unsigned int mask;
 402                 struct pollfd *fdp;
 403
 404                 mask = 0;
 405                 fdp = fdpage+i;
 406                 fd = fdp->fd;
 407                 if (fd >= 0) {
 408                         struct file * file = fget(fd);
 409                         mask = POLLNVAL;
 410                         if (file != NULL) {
 411                                 mask = DEFAULT_POLLMASK;
 412                                 if (file->f_op && file->f_op->poll)
 413                                         mask = file->f_op->poll(file, *pwait);
 414                                 mask &= fdp->events | POLLERR | POLLHUP;
 415                                 fput(file);
 416                         }
 417                         if (mask) {
 418                                 *pwait = NULL;
 419                                 (*count)++;
 420                         }
 421                 }
 422                 fdp->revents = mask;
 423         }
 424 }
 425
 426 static int do_poll(unsigned int nfds,  struct poll_list *list,
 427                         struct poll_wqueues *wait, long timeout)
 428 {
 429         int count = 0;
 430         poll_table* pt = &wait->pt;
 431
 432         if (!timeout)
 433                 pt = NULL;
 434
 435         for (;;) {
 436                 struct poll_list *walk;
 437                 set_current_state(TASK_INTERRUPTIBLE);
 438                 walk = list;
 439                 while(walk != NULL) {
 440                         do_pollfd( walk->len, walk->entries, &pt, &count);
 441                         walk = walk->next;
 442                 }
 443                 pt = NULL;
 444                 if (count || !timeout || signal_pending(current))
 445                         break;
 446                 count = wait->error;
 447                 if (count)
 448                         break;
 449                 timeout = schedule_timeout(timeout);
 450         }
 451         __set_current_state(TASK_RUNNING);
 452         return count;
 453 }
 454
 455 asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long timeout)
 456 {
 457         struct poll_wqueues table;
 458         int fdcount, err;
 459         unsigned int i;
 460         struct poll_list *head;
 461         struct poll_list *walk;
 462
 463         /* Do a sanity check on nfds ... */
 464         if (nfds > current->files->max_fdset && nfds > OPEN_MAX)
 465                 return -EINVAL;
 466
 467         if (timeout) {
 468                 /* Careful about overflow in the intermediate values */
 469                 if ((unsigned long) timeout < MAX_SCHEDULE_TIMEOUT / HZ)
 470                         timeout = (unsigned long)(timeout*HZ+999)/1000+1;
 471                 else /* Negative or overflow */
 472                         timeout = MAX_SCHEDULE_TIMEOUT;
 473         }
 474
 475         poll_initwait(&table);
 476
 477         head = NULL;
 478         walk = NULL;
 479         i = nfds;
 480         err = -ENOMEM;
 481         while(i!=0) {
 482                 struct poll_list *pp;
 483                 pp = kmalloc(sizeof(struct poll_list)+
 484                                 sizeof(struct pollfd)*
 485                                 (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i),
 486                                         GFP_KERNEL);
 487                 if(pp==NULL)
 488                         goto out_fds;
 489                 pp->next=NULL;
 490                 pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i);
 491                 if (head == NULL)
 492                         head = pp;
 493                 else
 494                         walk->next = pp;
 495
 496                 walk = pp;
 497                 if (copy_from_user(pp->entries, ufds + nfds-i,
 498                                 sizeof(struct pollfd)*pp->len)) {
 499                         err = -EFAULT;
 500                         goto out_fds;
 501                 }
 502                 i -= pp->len;
 503         }
 504         fdcount = do_poll(nfds, head, &table, timeout);
 505
 506         /* OK, now copy the revents fields back to user space. */
 507         walk = head;
 508         err = -EFAULT;
 509         while(walk != NULL) {
 510                 struct pollfd *fds = walk->entries;
 511                 int j;
 512
 513                 for (j=0; j < walk->len; j++, ufds++) {
 514                         if(__put_user(fds[j].revents, &ufds->revents))
 515                                 goto out_fds;
 516                 }
 517                 walk = walk->next;
 518         }
 519         err = fdcount;
 520         if (!fdcount && signal_pending(current))
 521                 err = -EINTR;
 522 out_fds:
 523         walk = head;
 524         while(walk!=NULL) {
 525                 struct poll_list *pp = walk->next;
 526                 kfree(walk);
 527                 walk = pp;
 528         }
 529         poll_freewait(&table);
 530         return err;
 531 }