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