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/module.h>
  18 #include <linux/slab.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/poll.h>
  21 #include <linux/personality.h> /* for STICKY_TIMEOUTS */
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24
  25 #include <asm/uaccess.h>
  26
  27 #define ROUND_UP(x,y) (((x)+(y)-1)/(y))
  28 #define DEFAULT_POLLMASK (POLLIN | POLLOUT | POLLRDNORM | POLLWRNORM)
  29
  30 struct poll_table_entry {
  31         struct file * filp;
  32         wait_queue_t wait;
  33         wait_queue_head_t * wait_address;
  34 };
  35
  36 struct poll_table_page {
  37         struct poll_table_page * next;
  38         struct poll_table_entry * entry;
  39         struct poll_table_entry entries[0];
  40 };
  41
  42 #define POLL_TABLE_FULL(table) \
  43         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
  44
  45 /*
  46  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
  47  * I have rewritten this, taking some shortcuts: This code may not be easy to
  48  * follow, but it should be free of race-conditions, and it's practical. If you
  49  * understand what I'm doing here, then you understand how the linux
  50  * sleep/wakeup mechanism works.
  51  *
  52  * Two very simple procedures, poll_wait() and poll_freewait() make all the
  53  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
  54  * as all select/poll functions have to call it to add an entry to the
  55  * poll table.
  56  */
  57 void __pollwait(struct file *filp, wait_queue_head_t *wait_address, poll_table *p);
  58
  59 void poll_initwait(struct poll_wqueues *pwq)
  60 {
  61         init_poll_funcptr(&pwq->pt, __pollwait);
  62         pwq->error = 0;
  63         pwq->table = NULL;
  64 }
  65
  66 EXPORT_SYMBOL(poll_initwait);
  67
  68 void poll_freewait(struct poll_wqueues *pwq)
  69 {
  70         struct poll_table_page * p = pwq->table;
  71         while (p) {
  72                 struct poll_table_entry * entry;
  73                 struct poll_table_page *old;
  74
  75                 entry = p->entry;
  76                 do {
  77                         entry--;
  78                         remove_wait_queue(entry->wait_address,&entry->wait);
  79                         fput(entry->filp);
  80                 } while (entry > p->entries);
  81                 old = p;
  82                 p = p->next;
  83                 free_page((unsigned long) old);
  84         }
  85 }
  86
  87 EXPORT_SYMBOL(poll_freewait);
  88
  89 void __pollwait(struct file *filp, wait_queue_head_t *wait_address, poll_table *_p)
  90 {
  91         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
  92         struct poll_table_page *table = p->table;
  93
  94         if (!table || POLL_TABLE_FULL(table)) {
  95                 struct poll_table_page *new_table;
  96
  97                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
  98                 if (!new_table) {
  99                         p->error = -ENOMEM;
 100                         __set_current_state(TASK_RUNNING);
 101                         return;
 102                 }
 103                 new_table->entry = new_table->entries;
 104                 new_table->next = table;
 105                 p->table = new_table;
 106                 table = new_table;
 107         }
 108
 109         /* Add a new entry */
 110         {
 111                 struct poll_table_entry * entry = table->entry;
 112                 table->entry = entry+1;
 113                 get_file(filp);
 114                 entry->filp = filp;
 115                 entry->wait_address = wait_address;
 116                 init_waitqueue_entry(&entry->wait, current);
 117                 add_wait_queue(wait_address,&entry->wait);
 118         }
 119 }
 120
 121 #define FDS_IN(fds, n)          (fds->in + n)
 122 #define FDS_OUT(fds, n)         (fds->out + n)
 123 #define FDS_EX(fds, n)          (fds->ex + n)
 124
 125 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
 126
 127 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 128 {
 129         unsigned long *open_fds;
 130         unsigned long set;
 131         int max;
 132
 133         /* handle last in-complete long-word first */
 134         set = ~(~0UL << (n & (__NFDBITS-1)));
 135         n /= __NFDBITS;
 136         open_fds = current->files->open_fds->fds_bits+n;
 137         max = 0;
 138         if (set) {
 139                 set &= BITS(fds, n);
 140                 if (set) {
 141                         if (!(set & ~*open_fds))
 142                                 goto get_max;
 143                         return -EBADF;
 144                 }
 145         }
 146         while (n) {
 147                 open_fds--;
 148                 n--;
 149                 set = BITS(fds, n);
 150                 if (!set)
 151                         continue;
 152                 if (set & ~*open_fds)
 153                         return -EBADF;
 154                 if (max)
 155                         continue;
 156 get_max:
 157                 do {
 158                         max++;
 159                         set >>= 1;
 160                 } while (set);
 161                 max += n * __NFDBITS;
 162         }
 163
 164         return max;
 165 }
 166
 167 #define BIT(i)          (1UL << ((i)&(__NFDBITS-1)))
 168 #define MEM(i,m)        ((m)+(unsigned)(i)/__NFDBITS)
 169 #define ISSET(i,m)      (((i)&*(m)) != 0)
 170 #define SET(i,m)        (*(m) |= (i))
 171
 172 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 173 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 174 #define POLLEX_SET (POLLPRI)
 175
 176 int do_select(int n, fd_set_bits *fds, long *timeout)
 177 {
 178         struct poll_wqueues table;
 179         poll_table *wait;
 180         int retval, i;
 181         long __timeout = *timeout;
 182
 183         spin_lock(&current->files->file_lock);
 184         retval = max_select_fd(n, fds);
 185         spin_unlock(&current->files->file_lock);
 186
 187         if (retval < 0)
 188                 return retval;
 189         n = retval;
 190
 191         poll_initwait(&table);
 192         wait = &table.pt;
 193         if (!__timeout)
 194                 wait = NULL;
 195         retval = 0;
 196         for (;;) {
 197                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 198
 199                 set_current_state(TASK_INTERRUPTIBLE);
 200
 201                 inp = fds->in; outp = fds->out; exp = fds->ex;
 202                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 203
 204                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 205                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 206                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 207                         struct file_operations *f_op = NULL;
 208                         struct file *file = NULL;
 209
 210                         in = *inp++; out = *outp++; ex = *exp++;
 211                         all_bits = in | out | ex;
 212                         if (all_bits == 0) {
 213                                 i += __NFDBITS;
 214                                 continue;
 215                         }
 216
 217                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 218                                 if (i >= n)
 219                                         break;
 220                                 if (!(bit & all_bits))
 221                                         continue;
 222                                 file = fget(i);
 223                                 if (file) {
 224                                         f_op = file->f_op;
 225                                         mask = DEFAULT_POLLMASK;
 226                                         if (f_op && f_op->poll)
 227                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 228                                         fput(file);
 229                                         if ((mask & POLLIN_SET) && (in & bit)) {
 230                                                 res_in |= bit;
 231                                                 retval++;
 232                                         }
 233                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 234                                                 res_out |= bit;
 235                                                 retval++;
 236                                         }
 237                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 238                                                 res_ex |= bit;
 239                                                 retval++;
 240                                         }
 241                                 }
 242                         }
 243                         if (res_in)
 244                                 *rinp = res_in;
 245                         if (res_out)
 246                                 *routp = res_out;
 247                         if (res_ex)
 248                                 *rexp = res_ex;
 249                 }
 250                 wait = NULL;
 251                 if (retval || !__timeout || signal_pending(current))
 252                         break;
 253                 if(table.error) {
 254                         retval = table.error;
 255                         break;
 256                 }
 257                 __timeout = schedule_timeout(__timeout);
 258         }
 259         __set_current_state(TASK_RUNNING);
 260
 261         poll_freewait(&table);
 262
 263         /*
 264          * Up-to-date the caller timeout.
 265          */
 266         *timeout = __timeout;
 267         return retval;
 268 }
 269
 270 EXPORT_SYMBOL(do_select);
 271
 272 static void *select_bits_alloc(int size)
 273 {
 274         return kmalloc(6 * size, GFP_KERNEL);
 275 }
 276
 277 static void select_bits_free(void *bits, int size)
 278 {
 279         kfree(bits);
 280 }
 281
 282 /*
 283  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 284  * like to be certain this leads to no problems. So I return
 285  * EINTR just for safety.
 286  *
 287  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 288  * I'm trying ERESTARTNOHAND which restart only when you want to.
 289  */
 290 #define MAX_SELECT_SECONDS \
 291         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 292
 293 asmlinkage long
 294 sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, struct timeval __user *tvp)
 295 {
 296         fd_set_bits fds;
 297         char *bits;
 298         long timeout;
 299         int ret, size, max_fdset;
 300
 301         timeout = MAX_SCHEDULE_TIMEOUT;
 302         if (tvp) {
 303                 time_t sec, usec;
 304
 305                 if ((ret = verify_area(VERIFY_READ, tvp, sizeof(*tvp)))
 306                     || (ret = __get_user(sec, &tvp->tv_sec))
 307                     || (ret = __get_user(usec, &tvp->tv_usec)))
 308                         goto out_nofds;
 309
 310                 ret = -EINVAL;
 311                 if (sec < 0 || usec < 0)
 312                         goto out_nofds;
 313
 314                 if ((unsigned long) sec < MAX_SELECT_SECONDS) {
 315                         timeout = ROUND_UP(usec, 1000000/HZ);
 316                         timeout += sec * (unsigned long) HZ;
 317                 }
 318         }
 319
 320         ret = -EINVAL;
 321         if (n < 0)
 322                 goto out_nofds;
 323
 324         /* max_fdset can increase, so grab it once to avoid race */
 325         max_fdset = current->files->max_fdset;
 326         if (n > max_fdset)
 327                 n = max_fdset;
 328
 329         /*
 330          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 331          * since we used fdset we need to allocate memory in units of
 332          * long-words.
 333          */
 334         ret = -ENOMEM;
 335         size = FDS_BYTES(n);
 336         bits = select_bits_alloc(size);
 337         if (!bits)
 338                 goto out_nofds;
 339         fds.in      = (unsigned long *)  bits;
 340         fds.out     = (unsigned long *) (bits +   size);
 341         fds.ex      = (unsigned long *) (bits + 2*size);
 342         fds.res_in  = (unsigned long *) (bits + 3*size);
 343         fds.res_out = (unsigned long *) (bits + 4*size);
 344         fds.res_ex  = (unsigned long *) (bits + 5*size);
 345
 346         if ((ret = get_fd_set(n, inp, fds.in)) ||
 347             (ret = get_fd_set(n, outp, fds.out)) ||
 348             (ret = get_fd_set(n, exp, fds.ex)))
 349                 goto out;
 350         zero_fd_set(n, fds.res_in);
 351         zero_fd_set(n, fds.res_out);
 352         zero_fd_set(n, fds.res_ex);
 353
 354         ret = do_select(n, &fds, &timeout);
 355
 356         if (tvp && !(current->personality & STICKY_TIMEOUTS)) {
 357                 time_t sec = 0, usec = 0;
 358                 if (timeout) {
 359                         sec = timeout / HZ;
 360                         usec = timeout % HZ;
 361                         usec *= (1000000/HZ);
 362                 }
 363                 put_user(sec, &tvp->tv_sec);
 364                 put_user(usec, &tvp->tv_usec);
 365         }
 366
 367         if (ret < 0)
 368                 goto out;
 369         if (!ret) {
 370                 ret = -ERESTARTNOHAND;
 371                 if (signal_pending(current))
 372                         goto out;
 373                 ret = 0;
 374         }
 375
 376         if (set_fd_set(n, inp, fds.res_in) ||
 377             set_fd_set(n, outp, fds.res_out) ||
 378             set_fd_set(n, exp, fds.res_ex))
 379                 ret = -EFAULT;
 380
 381 out:
 382         select_bits_free(bits, size);
 383 out_nofds:
 384         return ret;
 385 }
 386
 387 struct poll_list {
 388         struct poll_list *next;
 389         int len;
 390         struct pollfd entries[0];
 391 };
 392
 393 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 394
 395 static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 396         poll_table ** pwait, int *count)
 397 {
 398         int i;
 399
 400         for (i = 0; i < num; i++) {
 401                 int fd;
 402                 unsigned int mask;
 403                 struct pollfd *fdp;
 404
 405                 mask = 0;
 406                 fdp = fdpage+i;
 407                 fd = fdp->fd;
 408                 if (fd >= 0) {
 409                         struct file * file = fget(fd);
 410                         mask = POLLNVAL;
 411                         if (file != NULL) {
 412                                 mask = DEFAULT_POLLMASK;
 413                                 if (file->f_op && file->f_op->poll)
 414                                         mask = file->f_op->poll(file, *pwait);
 415                                 mask &= fdp->events | POLLERR | POLLHUP;
 416                                 fput(file);
 417                         }
 418                         if (mask) {
 419                                 *pwait = NULL;
 420                                 (*count)++;
 421                         }
 422                 }
 423                 fdp->revents = mask;
 424         }
 425 }
 426
 427 static int do_poll(unsigned int nfds,  struct poll_list *list,
 428                         struct poll_wqueues *wait, long timeout)
 429 {
 430         int count = 0;
 431         poll_table* pt = &wait->pt;
 432
 433         if (!timeout)
 434                 pt = NULL;
 435
 436         for (;;) {
 437                 struct poll_list *walk;
 438                 set_current_state(TASK_INTERRUPTIBLE);
 439                 walk = list;
 440                 while(walk != NULL) {
 441                         do_pollfd( walk->len, walk->entries, &pt, &count);
 442                         walk = walk->next;
 443                 }
 444                 pt = NULL;
 445                 if (count || !timeout || signal_pending(current))
 446                         break;
 447                 count = wait->error;
 448                 if (count)
 449                         break;
 450                 timeout = schedule_timeout(timeout);
 451         }
 452         __set_current_state(TASK_RUNNING);
 453         return count;
 454 }
 455
 456 asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long timeout)
 457 {
 458         struct poll_wqueues table;
 459         int fdcount, err;
 460         unsigned int i;
 461         struct poll_list *head;
 462         struct poll_list *walk;
 463
 464         /* Do a sanity check on nfds ... */
 465         if (nfds > current->files->max_fdset && nfds > OPEN_MAX)
 466                 return -EINVAL;
 467
 468         if (timeout) {
 469                 /* Careful about overflow in the intermediate values */
 470                 if ((unsigned long) timeout < MAX_SCHEDULE_TIMEOUT / HZ)
 471                         timeout = (unsigned long)(timeout*HZ+999)/1000+1;
 472                 else /* Negative or overflow */
 473                         timeout = MAX_SCHEDULE_TIMEOUT;
 474         }
 475
 476         poll_initwait(&table);
 477
 478         head = NULL;
 479         walk = NULL;
 480         i = nfds;
 481         err = -ENOMEM;
 482         while(i!=0) {
 483                 struct poll_list *pp;
 484                 pp = kmalloc(sizeof(struct poll_list)+
 485                                 sizeof(struct pollfd)*
 486                                 (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i),
 487                                         GFP_KERNEL);
 488                 if(pp==NULL)
 489                         goto out_fds;
 490                 pp->next=NULL;
 491                 pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i);
 492                 if (head == NULL)
 493                         head = pp;
 494                 else
 495                         walk->next = pp;
 496
 497                 walk = pp;
 498                 if (copy_from_user(pp->entries, ufds + nfds-i,
 499                                 sizeof(struct pollfd)*pp->len)) {
 500                         err = -EFAULT;
 501                         goto out_fds;
 502                 }
 503                 i -= pp->len;
 504         }
 505         fdcount = do_poll(nfds, head, &table, timeout);
 506
 507         /* OK, now copy the revents fields back to user space. */
 508         walk = head;
 509         err = -EFAULT;
 510         while(walk != NULL) {
 511                 struct pollfd *fds = walk->entries;
 512                 int j;
 513
 514                 for (j=0; j < walk->len; j++, ufds++) {
 515                         if(__put_user(fds[j].revents, &ufds->revents))
 516                                 goto out_fds;
 517                 }
 518                 walk = walk->next;
 519         }
 520         err = fdcount;
 521         if (!fdcount && signal_pending(current))
 522                 err = -EINTR;
 523 out_fds:
 524         walk = head;
 525         while(walk!=NULL) {
 526                 struct poll_list *pp = walk->next;
 527                 kfree(walk);
 528                 walk = pp;
 529         }
 530         poll_freewait(&table);
 531         return err;
 532 }