Merge branch 'for-linus' of git://linux-nfs.org/~bfields/linux
[linux-2.6] / 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/kernel.h>
18 #include <linux/syscalls.h>
19 #include <linux/module.h>
20 #include <linux/slab.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 #include <linux/rcupdate.h>
26
27 #include <asm/uaccess.h>
28
29 struct poll_table_page {
30         struct poll_table_page * next;
31         struct poll_table_entry * entry;
32         struct poll_table_entry entries[0];
33 };
34
35 #define POLL_TABLE_FULL(table) \
36         ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table))
37
38 /*
39  * Ok, Peter made a complicated, but straightforward multiple_wait() function.
40  * I have rewritten this, taking some shortcuts: This code may not be easy to
41  * follow, but it should be free of race-conditions, and it's practical. If you
42  * understand what I'm doing here, then you understand how the linux
43  * sleep/wakeup mechanism works.
44  *
45  * Two very simple procedures, poll_wait() and poll_freewait() make all the
46  * work.  poll_wait() is an inline-function defined in <linux/poll.h>,
47  * as all select/poll functions have to call it to add an entry to the
48  * poll table.
49  */
50 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
51                        poll_table *p);
52
53 void poll_initwait(struct poll_wqueues *pwq)
54 {
55         init_poll_funcptr(&pwq->pt, __pollwait);
56         pwq->error = 0;
57         pwq->table = NULL;
58         pwq->inline_index = 0;
59 }
60
61 EXPORT_SYMBOL(poll_initwait);
62
63 static void free_poll_entry(struct poll_table_entry *entry)
64 {
65         remove_wait_queue(entry->wait_address, &entry->wait);
66         fput(entry->filp);
67 }
68
69 void poll_freewait(struct poll_wqueues *pwq)
70 {
71         struct poll_table_page * p = pwq->table;
72         int i;
73         for (i = 0; i < pwq->inline_index; i++)
74                 free_poll_entry(pwq->inline_entries + i);
75         while (p) {
76                 struct poll_table_entry * entry;
77                 struct poll_table_page *old;
78
79                 entry = p->entry;
80                 do {
81                         entry--;
82                         free_poll_entry(entry);
83                 } while (entry > p->entries);
84                 old = p;
85                 p = p->next;
86                 free_page((unsigned long) old);
87         }
88 }
89
90 EXPORT_SYMBOL(poll_freewait);
91
92 static struct poll_table_entry *poll_get_entry(poll_table *_p)
93 {
94         struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt);
95         struct poll_table_page *table = p->table;
96
97         if (p->inline_index < N_INLINE_POLL_ENTRIES)
98                 return p->inline_entries + p->inline_index++;
99
100         if (!table || POLL_TABLE_FULL(table)) {
101                 struct poll_table_page *new_table;
102
103                 new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL);
104                 if (!new_table) {
105                         p->error = -ENOMEM;
106                         __set_current_state(TASK_RUNNING);
107                         return NULL;
108                 }
109                 new_table->entry = new_table->entries;
110                 new_table->next = table;
111                 p->table = new_table;
112                 table = new_table;
113         }
114
115         return table->entry++;
116 }
117
118 /* Add a new entry */
119 static void __pollwait(struct file *filp, wait_queue_head_t *wait_address,
120                                 poll_table *p)
121 {
122         struct poll_table_entry *entry = poll_get_entry(p);
123         if (!entry)
124                 return;
125         get_file(filp);
126         entry->filp = filp;
127         entry->wait_address = wait_address;
128         init_waitqueue_entry(&entry->wait, current);
129         add_wait_queue(wait_address, &entry->wait);
130 }
131
132 #define FDS_IN(fds, n)          (fds->in + n)
133 #define FDS_OUT(fds, n)         (fds->out + n)
134 #define FDS_EX(fds, n)          (fds->ex + n)
135
136 #define BITS(fds, n)    (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n))
137
138 static int max_select_fd(unsigned long n, fd_set_bits *fds)
139 {
140         unsigned long *open_fds;
141         unsigned long set;
142         int max;
143         struct fdtable *fdt;
144
145         /* handle last in-complete long-word first */
146         set = ~(~0UL << (n & (__NFDBITS-1)));
147         n /= __NFDBITS;
148         fdt = files_fdtable(current->files);
149         open_fds = fdt->open_fds->fds_bits+n;
150         max = 0;
151         if (set) {
152                 set &= BITS(fds, n);
153                 if (set) {
154                         if (!(set & ~*open_fds))
155                                 goto get_max;
156                         return -EBADF;
157                 }
158         }
159         while (n) {
160                 open_fds--;
161                 n--;
162                 set = BITS(fds, n);
163                 if (!set)
164                         continue;
165                 if (set & ~*open_fds)
166                         return -EBADF;
167                 if (max)
168                         continue;
169 get_max:
170                 do {
171                         max++;
172                         set >>= 1;
173                 } while (set);
174                 max += n * __NFDBITS;
175         }
176
177         return max;
178 }
179
180 #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR)
181 #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR)
182 #define POLLEX_SET (POLLPRI)
183
184 int do_select(int n, fd_set_bits *fds, s64 *timeout)
185 {
186         struct poll_wqueues table;
187         poll_table *wait;
188         int retval, i;
189
190         rcu_read_lock();
191         retval = max_select_fd(n, fds);
192         rcu_read_unlock();
193
194         if (retval < 0)
195                 return retval;
196         n = retval;
197
198         poll_initwait(&table);
199         wait = &table.pt;
200         if (!*timeout)
201                 wait = NULL;
202         retval = 0;
203         for (;;) {
204                 unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp;
205                 long __timeout;
206
207                 set_current_state(TASK_INTERRUPTIBLE);
208
209                 inp = fds->in; outp = fds->out; exp = fds->ex;
210                 rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex;
211
212                 for (i = 0; i < n; ++rinp, ++routp, ++rexp) {
213                         unsigned long in, out, ex, all_bits, bit = 1, mask, j;
214                         unsigned long res_in = 0, res_out = 0, res_ex = 0;
215                         const struct file_operations *f_op = NULL;
216                         struct file *file = NULL;
217
218                         in = *inp++; out = *outp++; ex = *exp++;
219                         all_bits = in | out | ex;
220                         if (all_bits == 0) {
221                                 i += __NFDBITS;
222                                 continue;
223                         }
224
225                         for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) {
226                                 int fput_needed;
227                                 if (i >= n)
228                                         break;
229                                 if (!(bit & all_bits))
230                                         continue;
231                                 file = fget_light(i, &fput_needed);
232                                 if (file) {
233                                         f_op = file->f_op;
234                                         mask = DEFAULT_POLLMASK;
235                                         if (f_op && f_op->poll)
236                                                 mask = (*f_op->poll)(file, retval ? NULL : wait);
237                                         fput_light(file, fput_needed);
238                                         if ((mask & POLLIN_SET) && (in & bit)) {
239                                                 res_in |= bit;
240                                                 retval++;
241                                         }
242                                         if ((mask & POLLOUT_SET) && (out & bit)) {
243                                                 res_out |= bit;
244                                                 retval++;
245                                         }
246                                         if ((mask & POLLEX_SET) && (ex & bit)) {
247                                                 res_ex |= bit;
248                                                 retval++;
249                                         }
250                                 }
251                                 cond_resched();
252                         }
253                         if (res_in)
254                                 *rinp = res_in;
255                         if (res_out)
256                                 *routp = res_out;
257                         if (res_ex)
258                                 *rexp = res_ex;
259                 }
260                 wait = NULL;
261                 if (retval || !*timeout || signal_pending(current))
262                         break;
263                 if(table.error) {
264                         retval = table.error;
265                         break;
266                 }
267
268                 if (*timeout < 0) {
269                         /* Wait indefinitely */
270                         __timeout = MAX_SCHEDULE_TIMEOUT;
271                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT - 1)) {
272                         /* Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in a loop */
273                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
274                         *timeout -= __timeout;
275                 } else {
276                         __timeout = *timeout;
277                         *timeout = 0;
278                 }
279                 __timeout = schedule_timeout(__timeout);
280                 if (*timeout >= 0)
281                         *timeout += __timeout;
282         }
283         __set_current_state(TASK_RUNNING);
284
285         poll_freewait(&table);
286
287         return retval;
288 }
289
290 /*
291  * We can actually return ERESTARTSYS instead of EINTR, but I'd
292  * like to be certain this leads to no problems. So I return
293  * EINTR just for safety.
294  *
295  * Update: ERESTARTSYS breaks at least the xview clock binary, so
296  * I'm trying ERESTARTNOHAND which restart only when you want to.
297  */
298 #define MAX_SELECT_SECONDS \
299         ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1)
300
301 static int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp,
302                            fd_set __user *exp, s64 *timeout)
303 {
304         fd_set_bits fds;
305         void *bits;
306         int ret, max_fds;
307         unsigned int size;
308         struct fdtable *fdt;
309         /* Allocate small arguments on the stack to save memory and be faster */
310         long stack_fds[SELECT_STACK_ALLOC/sizeof(long)];
311
312         ret = -EINVAL;
313         if (n < 0)
314                 goto out_nofds;
315
316         /* max_fds can increase, so grab it once to avoid race */
317         rcu_read_lock();
318         fdt = files_fdtable(current->files);
319         max_fds = fdt->max_fds;
320         rcu_read_unlock();
321         if (n > max_fds)
322                 n = max_fds;
323
324         /*
325          * We need 6 bitmaps (in/out/ex for both incoming and outgoing),
326          * since we used fdset we need to allocate memory in units of
327          * long-words. 
328          */
329         size = FDS_BYTES(n);
330         bits = stack_fds;
331         if (size > sizeof(stack_fds) / 6) {
332                 /* Not enough space in on-stack array; must use kmalloc */
333                 ret = -ENOMEM;
334                 bits = kmalloc(6 * size, GFP_KERNEL);
335                 if (!bits)
336                         goto out_nofds;
337         }
338         fds.in      = bits;
339         fds.out     = bits +   size;
340         fds.ex      = bits + 2*size;
341         fds.res_in  = bits + 3*size;
342         fds.res_out = bits + 4*size;
343         fds.res_ex  = 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 (ret < 0)
356                 goto out;
357         if (!ret) {
358                 ret = -ERESTARTNOHAND;
359                 if (signal_pending(current))
360                         goto out;
361                 ret = 0;
362         }
363
364         if (set_fd_set(n, inp, fds.res_in) ||
365             set_fd_set(n, outp, fds.res_out) ||
366             set_fd_set(n, exp, fds.res_ex))
367                 ret = -EFAULT;
368
369 out:
370         if (bits != stack_fds)
371                 kfree(bits);
372 out_nofds:
373         return ret;
374 }
375
376 asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp,
377                         fd_set __user *exp, struct timeval __user *tvp)
378 {
379         s64 timeout = -1;
380         struct timeval tv;
381         int ret;
382
383         if (tvp) {
384                 if (copy_from_user(&tv, tvp, sizeof(tv)))
385                         return -EFAULT;
386
387                 if (tv.tv_sec < 0 || tv.tv_usec < 0)
388                         return -EINVAL;
389
390                 /* Cast to u64 to make GCC stop complaining */
391                 if ((u64)tv.tv_sec >= (u64)MAX_INT64_SECONDS)
392                         timeout = -1;   /* infinite */
393                 else {
394                         timeout = DIV_ROUND_UP(tv.tv_usec, USEC_PER_SEC/HZ);
395                         timeout += tv.tv_sec * HZ;
396                 }
397         }
398
399         ret = core_sys_select(n, inp, outp, exp, &timeout);
400
401         if (tvp) {
402                 struct timeval rtv;
403
404                 if (current->personality & STICKY_TIMEOUTS)
405                         goto sticky;
406                 rtv.tv_usec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ));
407                 rtv.tv_sec = timeout;
408                 if (timeval_compare(&rtv, &tv) >= 0)
409                         rtv = tv;
410                 if (copy_to_user(tvp, &rtv, sizeof(rtv))) {
411 sticky:
412                         /*
413                          * If an application puts its timeval in read-only
414                          * memory, we don't want the Linux-specific update to
415                          * the timeval to cause a fault after the select has
416                          * completed successfully. However, because we're not
417                          * updating the timeval, we can't restart the system
418                          * call.
419                          */
420                         if (ret == -ERESTARTNOHAND)
421                                 ret = -EINTR;
422                 }
423         }
424
425         return ret;
426 }
427
428 #ifdef TIF_RESTORE_SIGMASK
429 asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp,
430                 fd_set __user *exp, struct timespec __user *tsp,
431                 const sigset_t __user *sigmask, size_t sigsetsize)
432 {
433         s64 timeout = MAX_SCHEDULE_TIMEOUT;
434         sigset_t ksigmask, sigsaved;
435         struct timespec ts;
436         int ret;
437
438         if (tsp) {
439                 if (copy_from_user(&ts, tsp, sizeof(ts)))
440                         return -EFAULT;
441
442                 if (ts.tv_sec < 0 || ts.tv_nsec < 0)
443                         return -EINVAL;
444
445                 /* Cast to u64 to make GCC stop complaining */
446                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
447                         timeout = -1;   /* infinite */
448                 else {
449                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
450                         timeout += ts.tv_sec * HZ;
451                 }
452         }
453
454         if (sigmask) {
455                 /* XXX: Don't preclude handling different sized sigset_t's.  */
456                 if (sigsetsize != sizeof(sigset_t))
457                         return -EINVAL;
458                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
459                         return -EFAULT;
460
461                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
462                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
463         }
464
465         ret = core_sys_select(n, inp, outp, exp, &timeout);
466
467         if (tsp) {
468                 struct timespec rts;
469
470                 if (current->personality & STICKY_TIMEOUTS)
471                         goto sticky;
472                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
473                                                 1000;
474                 rts.tv_sec = timeout;
475                 if (timespec_compare(&rts, &ts) >= 0)
476                         rts = ts;
477                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
478 sticky:
479                         /*
480                          * If an application puts its timeval in read-only
481                          * memory, we don't want the Linux-specific update to
482                          * the timeval to cause a fault after the select has
483                          * completed successfully. However, because we're not
484                          * updating the timeval, we can't restart the system
485                          * call.
486                          */
487                         if (ret == -ERESTARTNOHAND)
488                                 ret = -EINTR;
489                 }
490         }
491
492         if (ret == -ERESTARTNOHAND) {
493                 /*
494                  * Don't restore the signal mask yet. Let do_signal() deliver
495                  * the signal on the way back to userspace, before the signal
496                  * mask is restored.
497                  */
498                 if (sigmask) {
499                         memcpy(&current->saved_sigmask, &sigsaved,
500                                         sizeof(sigsaved));
501                         set_thread_flag(TIF_RESTORE_SIGMASK);
502                 }
503         } else if (sigmask)
504                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
505
506         return ret;
507 }
508
509 /*
510  * Most architectures can't handle 7-argument syscalls. So we provide a
511  * 6-argument version where the sixth argument is a pointer to a structure
512  * which has a pointer to the sigset_t itself followed by a size_t containing
513  * the sigset size.
514  */
515 asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp,
516         fd_set __user *exp, struct timespec __user *tsp, void __user *sig)
517 {
518         size_t sigsetsize = 0;
519         sigset_t __user *up = NULL;
520
521         if (sig) {
522                 if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t))
523                     || __get_user(up, (sigset_t __user * __user *)sig)
524                     || __get_user(sigsetsize,
525                                 (size_t __user *)(sig+sizeof(void *))))
526                         return -EFAULT;
527         }
528
529         return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize);
530 }
531 #endif /* TIF_RESTORE_SIGMASK */
532
533 struct poll_list {
534         struct poll_list *next;
535         int len;
536         struct pollfd entries[0];
537 };
538
539 #define POLLFD_PER_PAGE  ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd))
540
541 /*
542  * Fish for pollable events on the pollfd->fd file descriptor. We're only
543  * interested in events matching the pollfd->events mask, and the result
544  * matching that mask is both recorded in pollfd->revents and returned. The
545  * pwait poll_table will be used by the fd-provided poll handler for waiting,
546  * if non-NULL.
547  */
548 static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait)
549 {
550         unsigned int mask;
551         int fd;
552
553         mask = 0;
554         fd = pollfd->fd;
555         if (fd >= 0) {
556                 int fput_needed;
557                 struct file * file;
558
559                 file = fget_light(fd, &fput_needed);
560                 mask = POLLNVAL;
561                 if (file != NULL) {
562                         mask = DEFAULT_POLLMASK;
563                         if (file->f_op && file->f_op->poll)
564                                 mask = file->f_op->poll(file, pwait);
565                         /* Mask out unneeded events. */
566                         mask &= pollfd->events | POLLERR | POLLHUP;
567                         fput_light(file, fput_needed);
568                 }
569         }
570         pollfd->revents = mask;
571
572         return mask;
573 }
574
575 static int do_poll(unsigned int nfds,  struct poll_list *list,
576                    struct poll_wqueues *wait, s64 *timeout)
577 {
578         int count = 0;
579         poll_table* pt = &wait->pt;
580
581         /* Optimise the no-wait case */
582         if (!(*timeout))
583                 pt = NULL;
584
585         for (;;) {
586                 struct poll_list *walk;
587                 long __timeout;
588
589                 set_current_state(TASK_INTERRUPTIBLE);
590                 for (walk = list; walk != NULL; walk = walk->next) {
591                         struct pollfd * pfd, * pfd_end;
592
593                         pfd = walk->entries;
594                         pfd_end = pfd + walk->len;
595                         for (; pfd != pfd_end; pfd++) {
596                                 /*
597                                  * Fish for events. If we found one, record it
598                                  * and kill the poll_table, so we don't
599                                  * needlessly register any other waiters after
600                                  * this. They'll get immediately deregistered
601                                  * when we break out and return.
602                                  */
603                                 if (do_pollfd(pfd, pt)) {
604                                         count++;
605                                         pt = NULL;
606                                 }
607                         }
608                 }
609                 /*
610                  * All waiters have already been registered, so don't provide
611                  * a poll_table to them on the next loop iteration.
612                  */
613                 pt = NULL;
614                 if (!count) {
615                         count = wait->error;
616                         if (signal_pending(current))
617                                 count = -EINTR;
618                 }
619                 if (count || !*timeout)
620                         break;
621
622                 if (*timeout < 0) {
623                         /* Wait indefinitely */
624                         __timeout = MAX_SCHEDULE_TIMEOUT;
625                 } else if (unlikely(*timeout >= (s64)MAX_SCHEDULE_TIMEOUT-1)) {
626                         /*
627                          * Wait for longer than MAX_SCHEDULE_TIMEOUT. Do it in
628                          * a loop
629                          */
630                         __timeout = MAX_SCHEDULE_TIMEOUT - 1;
631                         *timeout -= __timeout;
632                 } else {
633                         __timeout = *timeout;
634                         *timeout = 0;
635                 }
636
637                 __timeout = schedule_timeout(__timeout);
638                 if (*timeout >= 0)
639                         *timeout += __timeout;
640         }
641         __set_current_state(TASK_RUNNING);
642         return count;
643 }
644
645 #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list))  / \
646                         sizeof(struct pollfd))
647
648 int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, s64 *timeout)
649 {
650         struct poll_wqueues table;
651         int err = -EFAULT, fdcount, len, size;
652         /* Allocate small arguments on the stack to save memory and be
653            faster - use long to make sure the buffer is aligned properly
654            on 64 bit archs to avoid unaligned access */
655         long stack_pps[POLL_STACK_ALLOC/sizeof(long)];
656         struct poll_list *const head = (struct poll_list *)stack_pps;
657         struct poll_list *walk = head;
658         unsigned long todo = nfds;
659
660         if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur)
661                 return -EINVAL;
662
663         len = min_t(unsigned int, nfds, N_STACK_PPS);
664         for (;;) {
665                 walk->next = NULL;
666                 walk->len = len;
667                 if (!len)
668                         break;
669
670                 if (copy_from_user(walk->entries, ufds + nfds-todo,
671                                         sizeof(struct pollfd) * walk->len))
672                         goto out_fds;
673
674                 todo -= walk->len;
675                 if (!todo)
676                         break;
677
678                 len = min(todo, POLLFD_PER_PAGE);
679                 size = sizeof(struct poll_list) + sizeof(struct pollfd) * len;
680                 walk = walk->next = kmalloc(size, GFP_KERNEL);
681                 if (!walk) {
682                         err = -ENOMEM;
683                         goto out_fds;
684                 }
685         }
686
687         poll_initwait(&table);
688         fdcount = do_poll(nfds, head, &table, timeout);
689         poll_freewait(&table);
690
691         for (walk = head; walk; walk = walk->next) {
692                 struct pollfd *fds = walk->entries;
693                 int j;
694
695                 for (j = 0; j < walk->len; j++, ufds++)
696                         if (__put_user(fds[j].revents, &ufds->revents))
697                                 goto out_fds;
698         }
699
700         err = fdcount;
701 out_fds:
702         walk = head->next;
703         while (walk) {
704                 struct poll_list *pos = walk;
705                 walk = walk->next;
706                 kfree(pos);
707         }
708
709         return err;
710 }
711
712 static long do_restart_poll(struct restart_block *restart_block)
713 {
714         struct pollfd __user *ufds = (struct pollfd __user*)restart_block->arg0;
715         int nfds = restart_block->arg1;
716         s64 timeout = ((s64)restart_block->arg3<<32) | (s64)restart_block->arg2;
717         int ret;
718
719         ret = do_sys_poll(ufds, nfds, &timeout);
720         if (ret == -EINTR) {
721                 restart_block->fn = do_restart_poll;
722                 restart_block->arg2 = timeout & 0xFFFFFFFF;
723                 restart_block->arg3 = (u64)timeout >> 32;
724                 ret = -ERESTART_RESTARTBLOCK;
725         }
726         return ret;
727 }
728
729 asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds,
730                         long timeout_msecs)
731 {
732         s64 timeout_jiffies;
733         int ret;
734
735         if (timeout_msecs > 0) {
736 #if HZ > 1000
737                 /* We can only overflow if HZ > 1000 */
738                 if (timeout_msecs / 1000 > (s64)0x7fffffffffffffffULL / (s64)HZ)
739                         timeout_jiffies = -1;
740                 else
741 #endif
742                         timeout_jiffies = msecs_to_jiffies(timeout_msecs) + 1;
743         } else {
744                 /* Infinite (< 0) or no (0) timeout */
745                 timeout_jiffies = timeout_msecs;
746         }
747
748         ret = do_sys_poll(ufds, nfds, &timeout_jiffies);
749         if (ret == -EINTR) {
750                 struct restart_block *restart_block;
751                 restart_block = &current_thread_info()->restart_block;
752                 restart_block->fn = do_restart_poll;
753                 restart_block->arg0 = (unsigned long)ufds;
754                 restart_block->arg1 = nfds;
755                 restart_block->arg2 = timeout_jiffies & 0xFFFFFFFF;
756                 restart_block->arg3 = (u64)timeout_jiffies >> 32;
757                 ret = -ERESTART_RESTARTBLOCK;
758         }
759         return ret;
760 }
761
762 #ifdef TIF_RESTORE_SIGMASK
763 asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds,
764         struct timespec __user *tsp, const sigset_t __user *sigmask,
765         size_t sigsetsize)
766 {
767         sigset_t ksigmask, sigsaved;
768         struct timespec ts;
769         s64 timeout = -1;
770         int ret;
771
772         if (tsp) {
773                 if (copy_from_user(&ts, tsp, sizeof(ts)))
774                         return -EFAULT;
775
776                 /* Cast to u64 to make GCC stop complaining */
777                 if ((u64)ts.tv_sec >= (u64)MAX_INT64_SECONDS)
778                         timeout = -1;   /* infinite */
779                 else {
780                         timeout = DIV_ROUND_UP(ts.tv_nsec, NSEC_PER_SEC/HZ);
781                         timeout += ts.tv_sec * HZ;
782                 }
783         }
784
785         if (sigmask) {
786                 /* XXX: Don't preclude handling different sized sigset_t's.  */
787                 if (sigsetsize != sizeof(sigset_t))
788                         return -EINVAL;
789                 if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask)))
790                         return -EFAULT;
791
792                 sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP));
793                 sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
794         }
795
796         ret = do_sys_poll(ufds, nfds, &timeout);
797
798         /* We can restart this syscall, usually */
799         if (ret == -EINTR) {
800                 /*
801                  * Don't restore the signal mask yet. Let do_signal() deliver
802                  * the signal on the way back to userspace, before the signal
803                  * mask is restored.
804                  */
805                 if (sigmask) {
806                         memcpy(&current->saved_sigmask, &sigsaved,
807                                         sizeof(sigsaved));
808                         set_thread_flag(TIF_RESTORE_SIGMASK);
809                 }
810                 ret = -ERESTARTNOHAND;
811         } else if (sigmask)
812                 sigprocmask(SIG_SETMASK, &sigsaved, NULL);
813
814         if (tsp && timeout >= 0) {
815                 struct timespec rts;
816
817                 if (current->personality & STICKY_TIMEOUTS)
818                         goto sticky;
819                 /* Yes, we know it's actually an s64, but it's also positive. */
820                 rts.tv_nsec = jiffies_to_usecs(do_div((*(u64*)&timeout), HZ)) *
821                                                 1000;
822                 rts.tv_sec = timeout;
823                 if (timespec_compare(&rts, &ts) >= 0)
824                         rts = ts;
825                 if (copy_to_user(tsp, &rts, sizeof(rts))) {
826                 sticky:
827                         /*
828                          * If an application puts its timeval in read-only
829                          * memory, we don't want the Linux-specific update to
830                          * the timeval to cause a fault after the select has
831                          * completed successfully. However, because we're not
832                          * updating the timeval, we can't restart the system
833                          * call.
834                          */
835                         if (ret == -ERESTARTNOHAND && timeout >= 0)
836                                 ret = -EINTR;
837                 }
838         }
839
840         return ret;
841 }
842 #endif /* TIF_RESTORE_SIGMASK */