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
 122 #define __IN(fds, n)            (fds->in + n)
 123 #define __OUT(fds, n)           (fds->out + n)
 124 #define __EX(fds, n)            (fds->ex + n)
 125 #define __RES_IN(fds, n)        (fds->res_in + n)
 126 #define __RES_OUT(fds, n)       (fds->res_out + n)
 127 #define __RES_EX(fds, n)        (fds->res_ex + n)
 128
 129 #define BITS(fds, n)            (*__IN(fds, n)|*__OUT(fds, n)|*__EX(fds, n))
 130
 131 static int max_select_fd(unsigned long n, fd_set_bits *fds)
 132 {
 133         unsigned long *open_fds;
 134         unsigned long set;
 135         int max;
 136
 137         /* handle last in-complete long-word first */
 138         set = ~(~0UL << (n & (__NFDBITS-1)));
 139         n /= __NFDBITS;
 140         open_fds = current->files->open_fds->fds_bits+n;
 141         max = 0;
 142         if (set) {
 143                 set &= BITS(fds, n);
 144                 if (set) {
 145                         if (!(set & ~*open_fds))
 146                                 goto get_max;
 147                         return -EBADF;
 148                 }
 149         }
 150         while (n) {
 151                 open_fds--;
 152                 n--;
 153                 set = BITS(fds, n);
 154                 if (!set)
 155                         continue;
 156                 if (set & ~*open_fds)
 157                         return -EBADF;
 158                 if (max)
 159                         continue;
 160 get_max:
 161                 do {
 162                         max++;
 163                         set >>= 1;
 164                 } while (set);
 165                 max += n * __NFDBITS;
 166         }
 167
 168         return max;
 169 }
 170
 171 #define BIT(i)          (1UL << ((i)&(__NFDBITS-1)))
 172 #define MEM(i,m)        ((m)+(unsigned)(i)/__NFDBITS)
 173 #define ISSET(i,m)      (((i)&*(m)) != 0)
 174 #define SET(i,m)        (*(m) |= (i))
 175
 176 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
 177 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
 178 #define POLLEX_SET (POLLPRI)
 179
 180 int do_select(int n, fd_set_bits *fds, long *timeout)
 181 {
 182         struct poll_wqueues table;
 183         poll_table *wait;
 184         int retval, i;
 185         long __timeout = *timeout;
 186
 187         spin_lock(&current->files->file_lock);
 188         retval = max_select_fd(n, fds);
 189         spin_unlock(&current->files->file_lock);
 190
 191         if (retval < 0)
 192                 return retval;
 193         n = retval;
 194
 195         poll_initwait(&table);
 196         wait = &table.pt;
 197         if (!__timeout)
 198                 wait = NULL;
 199         retval = 0;
 200         for (;;) {
 201                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
 202
 203                 set_current_state(TASK_INTERRUPTIBLE);
 204
 205                 inp = fds->in; outp = fds->out; exp = fds->ex;
 206                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
 207
 208                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
 209                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
 210                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
 211                         struct file_operations *f_op = NULL;
 212                         struct file *file = NULL;
 213
 214                         in = *inp++; out = *outp++; ex = *exp++;
 215                         all_bits = in | out | ex;
 216                         if (all_bits == 0) {
 217                                 i += __NFDBITS;
 218                                 continue;
 219                         }
 220
 221                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
 222                                 if (i >= n)
 223                                         break;
 224                                 if (!(bit & all_bits))
 225                                         continue;
 226                                 file = fget(i);
 227                                 if (file) {
 228                                         f_op = file->f_op;
 229                                         mask = DEFAULT_POLLMASK;
 230                                         if (f_op && f_op->poll)
 231                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
 232                                         fput(file);
 233                                         if ((mask & POLLIN_SET) && (in & bit)) {
 234                                                 res_in |= bit;
 235                                                 retval++;
 236                                         }
 237                                         if ((mask & POLLOUT_SET) && (out & bit)) {
 238                                                 res_out |= bit;
 239                                                 retval++;
 240                                         }
 241                                         if ((mask & POLLEX_SET) && (ex & bit)) {
 242                                                 res_ex |= bit;
 243                                                 retval++;
 244                                         }
 245                                 }
 246                                 cond_resched();
 247                         }
 248                         if (res_in)
 249                                 *rinp = res_in;
 250                         if (res_out)
 251                                 *routp = res_out;
 252                         if (res_ex)
 253                                 *rexp = res_ex;
 254                 }
 255                 wait = NULL;
 256                 if (retval || !__timeout || signal_pending(current))
 257                         break;
 258                 if(table.error) {
 259                         retval = table.error;
 260                         break;
 261                 }
 262                 __timeout = schedule_timeout(__timeout);
 263         }
 264         __set_current_state(TASK_RUNNING);
 265
 266         poll_freewait(&table);
 267
 268         /*
 269          * Up-to-date the caller timeout.
 270          */
 271         *timeout = __timeout;
 272         return retval;
 273 }
 274
 275 EXPORT_SYMBOL(do_select);
 276
 277 static void *select_bits_alloc(int size)
 278 {
 279         return kmalloc(6 * size, GFP_KERNEL);
 280 }
 281
 282 static void select_bits_free(void *bits, int size)
 283 {
 284         kfree(bits);
 285 }
 286
 287 /*
 288  * We can actually return ERESTARTSYS instead of EINTR, but I'd
 289  * like to be certain this leads to no problems. So I return
 290  * EINTR just for safety.
 291  *
 292  * Update: ERESTARTSYS breaks at least the xview clock binary, so
 293  * I'm trying ERESTARTNOHAND which restart only when you want to.
 294  */
 295 #define MAX_SELECT_SECONDS \
 296         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
 297
 298 asmlinkage long
 299 sys_select(int n, fd_set __user *inp, fd_set __user *outp, fd_set __user *exp, struct timeval __user *tvp)
 300 {
 301         fd_set_bits fds;
 302         char *bits;
 303         long timeout;
 304         int ret, size, max_fdset;
 305
 306         timeout = MAX_SCHEDULE_TIMEOUT;
 307         if (tvp) {
 308                 time_t sec, usec;
 309
 310                 if ((ret = verify_area(VERIFY_READ, tvp, sizeof(*tvp)))
 311                     || (ret = __get_user(sec, &tvp->tv_sec))
 312                     || (ret = __get_user(usec, &tvp->tv_usec)))
 313                         goto out_nofds;
 314
 315                 ret = -EINVAL;
 316                 if (sec < 0 || usec < 0)
 317                         goto out_nofds;
 318
 319                 if ((unsigned long) sec < MAX_SELECT_SECONDS) {
 320                         timeout = ROUND_UP(usec, 1000000/HZ);
 321                         timeout += sec * (unsigned long) HZ;
 322                 }
 323         }
 324
 325         ret = -EINVAL;
 326         if (n < 0)
 327                 goto out_nofds;
 328
 329         /* max_fdset can increase, so grab it once to avoid race */
 330         max_fdset = current->files->max_fdset;
 331         if (n > max_fdset)
 332                 n = max_fdset;
 333
 334         /*
 335          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
 336          * since we used fdset we need to allocate memory in units of
 337          * long-words.
 338          */
 339         ret = -ENOMEM;
 340         size = FDS_BYTES(n);
 341         bits = select_bits_alloc(size);
 342         if (!bits)
 343                 goto out_nofds;
 344         fds.in      = (unsigned long *)  bits;
 345         fds.out     = (unsigned long *) (bits +   size);
 346         fds.ex      = (unsigned long *) (bits + 2*size);
 347         fds.res_in  = (unsigned long *) (bits + 3*size);
 348         fds.res_out = (unsigned long *) (bits + 4*size);
 349         fds.res_ex  = (unsigned long *) (bits + 5*size);
 350
 351         if ((ret = get_fd_set(n, inp, fds.in)) ||
 352             (ret = get_fd_set(n, outp, fds.out)) ||
 353             (ret = get_fd_set(n, exp, fds.ex)))
 354                 goto out;
 355         zero_fd_set(n, fds.res_in);
 356         zero_fd_set(n, fds.res_out);
 357         zero_fd_set(n, fds.res_ex);
 358
 359         ret = do_select(n, &fds, &timeout);
 360
 361         if (tvp && !(current->personality & STICKY_TIMEOUTS)) {
 362                 time_t sec = 0, usec = 0;
 363                 if (timeout) {
 364                         sec = timeout / HZ;
 365                         usec = timeout % HZ;
 366                         usec *= (1000000/HZ);
 367                 }
 368                 put_user(sec, &tvp->tv_sec);
 369                 put_user(usec, &tvp->tv_usec);
 370         }
 371
 372         if (ret < 0)
 373                 goto out;
 374         if (!ret) {
 375                 ret = -ERESTARTNOHAND;
 376                 if (signal_pending(current))
 377                         goto out;
 378                 ret = 0;
 379         }
 380
 381         set_fd_set(n, inp, fds.res_in);
 382         set_fd_set(n, outp, fds.res_out);
 383         set_fd_set(n, exp, fds.res_ex);
 384
 385 out:
 386         select_bits_free(bits, size);
 387 out_nofds:
 388         return ret;
 389 }
 390
 391 struct poll_list {
 392         struct poll_list *next;
 393         int len;
 394         struct pollfd entries[0];
 395 };
 396
 397 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
 398
 399 static void do_pollfd(unsigned int num, struct pollfd * fdpage,
 400         poll_table ** pwait, int *count)
 401 {
 402         int i;
 403
 404         for (i = 0; i < num; i++) {
 405                 int fd;
 406                 unsigned int mask;
 407                 struct pollfd *fdp;
 408
 409                 mask = 0;
 410                 fdp = fdpage+i;
 411                 fd = fdp->fd;
 412                 if (fd >= 0) {
 413                         struct file * file = fget(fd);
 414                         mask = POLLNVAL;
 415                         if (file != NULL) {
 416                                 mask = DEFAULT_POLLMASK;
 417                                 if (file->f_op && file->f_op->poll)
 418                                         mask = file->f_op->poll(file, *pwait);
 419                                 mask &= fdp->events | POLLERR | POLLHUP;
 420                                 fput(file);
 421                         }
 422                         if (mask) {
 423                                 *pwait = NULL;
 424                                 (*count)++;
 425                         }
 426                 }
 427                 fdp->revents = mask;
 428         }
 429 }
 430
 431 static int do_poll(unsigned int nfds,  struct poll_list *list,
 432                         struct poll_wqueues *wait, long timeout)
 433 {
 434         int count = 0;
 435         poll_table* pt = &wait->pt;
 436
 437         if (!timeout)
 438                 pt = NULL;
 439
 440         for (;;) {
 441                 struct poll_list *walk;
 442                 set_current_state(TASK_INTERRUPTIBLE);
 443                 walk = list;
 444                 while(walk != NULL) {
 445                         do_pollfd( walk->len, walk->entries, &pt, &count);
 446                         walk = walk->next;
 447                 }
 448                 pt = NULL;
 449                 if (count || !timeout || signal_pending(current))
 450                         break;
 451                 count = wait->error;
 452                 if (count)
 453                         break;
 454                 timeout = schedule_timeout(timeout);
 455         }
 456         __set_current_state(TASK_RUNNING);
 457         return count;
 458 }
 459
 460 asmlinkage long sys_poll(struct pollfd __user * ufds, unsigned int nfds, long timeout)
 461 {
 462         struct poll_wqueues table;
 463         int fdcount, err;
 464         unsigned int i;
 465         struct poll_list *head;
 466         struct poll_list *walk;
 467
 468         /* Do a sanity check on nfds ... */
 469         if (nfds > current->files->max_fdset && nfds > OPEN_MAX)
 470                 return -EINVAL;
 471
 472         if (timeout) {
 473                 /* Careful about overflow in the intermediate values */
 474                 if ((unsigned long) timeout < MAX_SCHEDULE_TIMEOUT / HZ)
 475                         timeout = (unsigned long)(timeout*HZ+999)/1000+1;
 476                 else /* Negative or overflow */
 477                         timeout = MAX_SCHEDULE_TIMEOUT;
 478         }
 479
 480         poll_initwait(&table);
 481
 482         head = NULL;
 483         walk = NULL;
 484         i = nfds;
 485         err = -ENOMEM;
 486         while(i!=0) {
 487                 struct poll_list *pp;
 488                 pp = kmalloc(sizeof(struct poll_list)+
 489                                 sizeof(struct pollfd)*
 490                                 (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i),
 491                                         GFP_KERNEL);
 492                 if(pp==NULL)
 493                         goto out_fds;
 494                 pp->next=NULL;
 495                 pp->len = (i>POLLFD_PER_PAGE?POLLFD_PER_PAGE:i);
 496                 if (head == NULL)
 497                         head = pp;
 498                 else
 499                         walk->next = pp;
 500
 501                 walk = pp;
 502                 if (copy_from_user(pp->entries, ufds + nfds-i,
 503                                 sizeof(struct pollfd)*pp->len)) {
 504                         err = -EFAULT;
 505                         goto out_fds;
 506                 }
 507                 i -= pp->len;
 508         }
 509         fdcount = do_poll(nfds, head, &table, timeout);
 510
 511         /* OK, now copy the revents fields back to user space. */
 512         walk = head;
 513         err = -EFAULT;
 514         while(walk != NULL) {
 515                 struct pollfd *fds = walk->entries;
 516                 int j;
 517
 518                 for (j=0; j < walk->len; j++, ufds++) {
 519                         if(__put_user(fds[j].revents, &ufds->revents))
 520                                 goto out_fds;
 521                 }
 522                 walk = walk->next;
 523         }
 524         err = fdcount;
 525         if (!fdcount && signal_pending(current))
 526                 err = -EINTR;
 527 out_fds:
 528         walk = head;
 529         while(walk!=NULL) {
 530                 struct poll_list *pp = walk->next;
 531                 kfree(walk);
 532                 walk = pp;
 533         }
 534         poll_freewait(&table);
 535         return err;
 536 }