Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
Milind Arun Choudhary | 022a169 | 2007-05-08 00:29:02 -0700 | [diff] [blame] | 17 | #include <linux/kernel.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 18 | #include <linux/syscalls.h> |
| 19 | #include <linux/module.h> |
| 20 | #include <linux/slab.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 21 | #include <linux/poll.h> |
| 22 | #include <linux/personality.h> /* for STICKY_TIMEOUTS */ |
| 23 | #include <linux/file.h> |
Al Viro | 9f3acc3 | 2008-04-24 07:44:08 -0400 | [diff] [blame] | 24 | #include <linux/fdtable.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 25 | #include <linux/fs.h> |
Dipankar Sarma | b835996 | 2005-09-09 13:04:14 -0700 | [diff] [blame] | 26 | #include <linux/rcupdate.h> |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 27 | #include <linux/hrtimer.h> |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 28 | |
| 29 | #include <asm/uaccess.h> |
| 30 | |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 31 | |
| 32 | /* |
| 33 | * Estimate expected accuracy in ns from a timeval. |
| 34 | * |
| 35 | * After quite a bit of churning around, we've settled on |
| 36 | * a simple thing of taking 0.1% of the timeout as the |
| 37 | * slack, with a cap of 100 msec. |
| 38 | * "nice" tasks get a 0.5% slack instead. |
| 39 | * |
| 40 | * Consider this comment an open invitation to come up with even |
| 41 | * better solutions.. |
| 42 | */ |
| 43 | |
| 44 | static unsigned long __estimate_accuracy(struct timespec *tv) |
| 45 | { |
| 46 | unsigned long slack; |
| 47 | int divfactor = 1000; |
| 48 | |
| 49 | if (task_nice(current)) |
| 50 | divfactor = divfactor / 5; |
| 51 | |
| 52 | slack = tv->tv_nsec / divfactor; |
| 53 | slack += tv->tv_sec * (NSEC_PER_SEC/divfactor); |
| 54 | |
| 55 | if (slack > 100 * NSEC_PER_MSEC) |
| 56 | slack = 100 * NSEC_PER_MSEC; |
| 57 | return slack; |
| 58 | } |
| 59 | |
| 60 | static unsigned long estimate_accuracy(struct timespec *tv) |
| 61 | { |
| 62 | unsigned long ret; |
| 63 | struct timespec now; |
| 64 | |
| 65 | /* |
| 66 | * Realtime tasks get a slack of 0 for obvious reasons. |
| 67 | */ |
| 68 | |
| 69 | if (current->policy == SCHED_FIFO || |
| 70 | current->policy == SCHED_RR) |
| 71 | return 0; |
| 72 | |
| 73 | ktime_get_ts(&now); |
| 74 | now = timespec_sub(*tv, now); |
| 75 | ret = __estimate_accuracy(&now); |
| 76 | if (ret < current->timer_slack_ns) |
| 77 | return current->timer_slack_ns; |
| 78 | return ret; |
| 79 | } |
| 80 | |
| 81 | |
| 82 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 83 | struct poll_table_page { |
| 84 | struct poll_table_page * next; |
| 85 | struct poll_table_entry * entry; |
| 86 | struct poll_table_entry entries[0]; |
| 87 | }; |
| 88 | |
| 89 | #define POLL_TABLE_FULL(table) \ |
| 90 | ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table)) |
| 91 | |
| 92 | /* |
| 93 | * Ok, Peter made a complicated, but straightforward multiple_wait() function. |
| 94 | * I have rewritten this, taking some shortcuts: This code may not be easy to |
| 95 | * follow, but it should be free of race-conditions, and it's practical. If you |
| 96 | * understand what I'm doing here, then you understand how the linux |
| 97 | * sleep/wakeup mechanism works. |
| 98 | * |
| 99 | * Two very simple procedures, poll_wait() and poll_freewait() make all the |
| 100 | * work. poll_wait() is an inline-function defined in <linux/poll.h>, |
| 101 | * as all select/poll functions have to call it to add an entry to the |
| 102 | * poll table. |
| 103 | */ |
Adrian Bunk | 75c96f8 | 2005-05-05 16:16:09 -0700 | [diff] [blame] | 104 | static void __pollwait(struct file *filp, wait_queue_head_t *wait_address, |
| 105 | poll_table *p); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 106 | |
| 107 | void poll_initwait(struct poll_wqueues *pwq) |
| 108 | { |
| 109 | init_poll_funcptr(&pwq->pt, __pollwait); |
| 110 | pwq->error = 0; |
| 111 | pwq->table = NULL; |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 112 | pwq->inline_index = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 113 | } |
| 114 | |
| 115 | EXPORT_SYMBOL(poll_initwait); |
| 116 | |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 117 | static void free_poll_entry(struct poll_table_entry *entry) |
| 118 | { |
WANG Cong | ccf6780 | 2007-05-09 07:10:02 +0200 | [diff] [blame] | 119 | remove_wait_queue(entry->wait_address, &entry->wait); |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 120 | fput(entry->filp); |
| 121 | } |
| 122 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 123 | void poll_freewait(struct poll_wqueues *pwq) |
| 124 | { |
| 125 | struct poll_table_page * p = pwq->table; |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 126 | int i; |
| 127 | for (i = 0; i < pwq->inline_index; i++) |
| 128 | free_poll_entry(pwq->inline_entries + i); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 129 | while (p) { |
| 130 | struct poll_table_entry * entry; |
| 131 | struct poll_table_page *old; |
| 132 | |
| 133 | entry = p->entry; |
| 134 | do { |
| 135 | entry--; |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 136 | free_poll_entry(entry); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 137 | } while (entry > p->entries); |
| 138 | old = p; |
| 139 | p = p->next; |
| 140 | free_page((unsigned long) old); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | EXPORT_SYMBOL(poll_freewait); |
| 145 | |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 146 | static struct poll_table_entry *poll_get_entry(poll_table *_p) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 147 | { |
| 148 | struct poll_wqueues *p = container_of(_p, struct poll_wqueues, pt); |
| 149 | struct poll_table_page *table = p->table; |
| 150 | |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 151 | if (p->inline_index < N_INLINE_POLL_ENTRIES) |
| 152 | return p->inline_entries + p->inline_index++; |
| 153 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 154 | if (!table || POLL_TABLE_FULL(table)) { |
| 155 | struct poll_table_page *new_table; |
| 156 | |
| 157 | new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL); |
| 158 | if (!new_table) { |
| 159 | p->error = -ENOMEM; |
| 160 | __set_current_state(TASK_RUNNING); |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 161 | return NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 162 | } |
| 163 | new_table->entry = new_table->entries; |
| 164 | new_table->next = table; |
| 165 | p->table = new_table; |
| 166 | table = new_table; |
| 167 | } |
| 168 | |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 169 | return table->entry++; |
| 170 | } |
| 171 | |
| 172 | /* Add a new entry */ |
| 173 | static void __pollwait(struct file *filp, wait_queue_head_t *wait_address, |
| 174 | poll_table *p) |
| 175 | { |
| 176 | struct poll_table_entry *entry = poll_get_entry(p); |
| 177 | if (!entry) |
| 178 | return; |
| 179 | get_file(filp); |
| 180 | entry->filp = filp; |
| 181 | entry->wait_address = wait_address; |
| 182 | init_waitqueue_entry(&entry->wait, current); |
WANG Cong | ccf6780 | 2007-05-09 07:10:02 +0200 | [diff] [blame] | 183 | add_wait_queue(wait_address, &entry->wait); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 184 | } |
| 185 | |
Thomas Gleixner | b773ad4 | 2008-08-31 08:16:57 -0700 | [diff] [blame] | 186 | /** |
| 187 | * poll_select_set_timeout - helper function to setup the timeout value |
| 188 | * @to: pointer to timespec variable for the final timeout |
| 189 | * @sec: seconds (from user space) |
| 190 | * @nsec: nanoseconds (from user space) |
| 191 | * |
| 192 | * Note, we do not use a timespec for the user space value here, That |
| 193 | * way we can use the function for timeval and compat interfaces as well. |
| 194 | * |
| 195 | * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0. |
| 196 | */ |
| 197 | int poll_select_set_timeout(struct timespec *to, long sec, long nsec) |
| 198 | { |
| 199 | struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec}; |
| 200 | |
| 201 | if (!timespec_valid(&ts)) |
| 202 | return -EINVAL; |
| 203 | |
| 204 | /* Optimize for the zero timeout value here */ |
| 205 | if (!sec && !nsec) { |
| 206 | to->tv_sec = to->tv_nsec = 0; |
| 207 | } else { |
| 208 | ktime_get_ts(to); |
| 209 | *to = timespec_add_safe(*to, ts); |
| 210 | } |
| 211 | return 0; |
| 212 | } |
| 213 | |
| 214 | static int poll_select_copy_remaining(struct timespec *end_time, void __user *p, |
| 215 | int timeval, int ret) |
| 216 | { |
| 217 | struct timespec rts; |
| 218 | struct timeval rtv; |
| 219 | |
| 220 | if (!p) |
| 221 | return ret; |
| 222 | |
| 223 | if (current->personality & STICKY_TIMEOUTS) |
| 224 | goto sticky; |
| 225 | |
| 226 | /* No update for zero timeout */ |
| 227 | if (!end_time->tv_sec && !end_time->tv_nsec) |
| 228 | return ret; |
| 229 | |
| 230 | ktime_get_ts(&rts); |
| 231 | rts = timespec_sub(*end_time, rts); |
| 232 | if (rts.tv_sec < 0) |
| 233 | rts.tv_sec = rts.tv_nsec = 0; |
| 234 | |
| 235 | if (timeval) { |
| 236 | rtv.tv_sec = rts.tv_sec; |
| 237 | rtv.tv_usec = rts.tv_nsec / NSEC_PER_USEC; |
| 238 | |
| 239 | if (!copy_to_user(p, &rtv, sizeof(rtv))) |
| 240 | return ret; |
| 241 | |
| 242 | } else if (!copy_to_user(p, &rts, sizeof(rts))) |
| 243 | return ret; |
| 244 | |
| 245 | /* |
| 246 | * If an application puts its timeval in read-only memory, we |
| 247 | * don't want the Linux-specific update to the timeval to |
| 248 | * cause a fault after the select has completed |
| 249 | * successfully. However, because we're not updating the |
| 250 | * timeval, we can't restart the system call. |
| 251 | */ |
| 252 | |
| 253 | sticky: |
| 254 | if (ret == -ERESTARTNOHAND) |
| 255 | ret = -EINTR; |
| 256 | return ret; |
| 257 | } |
| 258 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 259 | #define FDS_IN(fds, n) (fds->in + n) |
| 260 | #define FDS_OUT(fds, n) (fds->out + n) |
| 261 | #define FDS_EX(fds, n) (fds->ex + n) |
| 262 | |
| 263 | #define BITS(fds, n) (*FDS_IN(fds, n)|*FDS_OUT(fds, n)|*FDS_EX(fds, n)) |
| 264 | |
| 265 | static int max_select_fd(unsigned long n, fd_set_bits *fds) |
| 266 | { |
| 267 | unsigned long *open_fds; |
| 268 | unsigned long set; |
| 269 | int max; |
Dipankar Sarma | badf166 | 2005-09-09 13:04:10 -0700 | [diff] [blame] | 270 | struct fdtable *fdt; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 271 | |
| 272 | /* handle last in-complete long-word first */ |
| 273 | set = ~(~0UL << (n & (__NFDBITS-1))); |
| 274 | n /= __NFDBITS; |
Dipankar Sarma | badf166 | 2005-09-09 13:04:10 -0700 | [diff] [blame] | 275 | fdt = files_fdtable(current->files); |
| 276 | open_fds = fdt->open_fds->fds_bits+n; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 277 | max = 0; |
| 278 | if (set) { |
| 279 | set &= BITS(fds, n); |
| 280 | if (set) { |
| 281 | if (!(set & ~*open_fds)) |
| 282 | goto get_max; |
| 283 | return -EBADF; |
| 284 | } |
| 285 | } |
| 286 | while (n) { |
| 287 | open_fds--; |
| 288 | n--; |
| 289 | set = BITS(fds, n); |
| 290 | if (!set) |
| 291 | continue; |
| 292 | if (set & ~*open_fds) |
| 293 | return -EBADF; |
| 294 | if (max) |
| 295 | continue; |
| 296 | get_max: |
| 297 | do { |
| 298 | max++; |
| 299 | set >>= 1; |
| 300 | } while (set); |
| 301 | max += n * __NFDBITS; |
| 302 | } |
| 303 | |
| 304 | return max; |
| 305 | } |
| 306 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 307 | #define POLLIN_SET (POLLRDNORM | POLLRDBAND | POLLIN | POLLHUP | POLLERR) |
| 308 | #define POLLOUT_SET (POLLWRBAND | POLLWRNORM | POLLOUT | POLLERR) |
| 309 | #define POLLEX_SET (POLLPRI) |
| 310 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 311 | int do_select(int n, fd_set_bits *fds, struct timespec *end_time) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 312 | { |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 313 | ktime_t expire, *to = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 314 | struct poll_wqueues table; |
| 315 | poll_table *wait; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 316 | int retval, i, timed_out = 0; |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 317 | unsigned long slack = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 318 | |
Dipankar Sarma | b835996 | 2005-09-09 13:04:14 -0700 | [diff] [blame] | 319 | rcu_read_lock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 320 | retval = max_select_fd(n, fds); |
Dipankar Sarma | b835996 | 2005-09-09 13:04:14 -0700 | [diff] [blame] | 321 | rcu_read_unlock(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 322 | |
| 323 | if (retval < 0) |
| 324 | return retval; |
| 325 | n = retval; |
| 326 | |
| 327 | poll_initwait(&table); |
| 328 | wait = &table.pt; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 329 | if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 330 | wait = NULL; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 331 | timed_out = 1; |
| 332 | } |
| 333 | |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 334 | if (end_time) |
| 335 | slack = estimate_accuracy(end_time); |
| 336 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 337 | retval = 0; |
| 338 | for (;;) { |
| 339 | unsigned long *rinp, *routp, *rexp, *inp, *outp, *exp; |
| 340 | |
| 341 | set_current_state(TASK_INTERRUPTIBLE); |
| 342 | |
| 343 | inp = fds->in; outp = fds->out; exp = fds->ex; |
| 344 | rinp = fds->res_in; routp = fds->res_out; rexp = fds->res_ex; |
| 345 | |
| 346 | for (i = 0; i < n; ++rinp, ++routp, ++rexp) { |
| 347 | unsigned long in, out, ex, all_bits, bit = 1, mask, j; |
| 348 | unsigned long res_in = 0, res_out = 0, res_ex = 0; |
Arjan van de Ven | 99ac48f | 2006-03-28 01:56:41 -0800 | [diff] [blame] | 349 | const struct file_operations *f_op = NULL; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 350 | struct file *file = NULL; |
| 351 | |
| 352 | in = *inp++; out = *outp++; ex = *exp++; |
| 353 | all_bits = in | out | ex; |
| 354 | if (all_bits == 0) { |
| 355 | i += __NFDBITS; |
| 356 | continue; |
| 357 | } |
| 358 | |
| 359 | for (j = 0; j < __NFDBITS; ++j, ++i, bit <<= 1) { |
Eric Dumazet | e4a1f12 | 2006-03-28 01:56:34 -0800 | [diff] [blame] | 360 | int fput_needed; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 361 | if (i >= n) |
| 362 | break; |
| 363 | if (!(bit & all_bits)) |
| 364 | continue; |
Eric Dumazet | e4a1f12 | 2006-03-28 01:56:34 -0800 | [diff] [blame] | 365 | file = fget_light(i, &fput_needed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 366 | if (file) { |
| 367 | f_op = file->f_op; |
| 368 | mask = DEFAULT_POLLMASK; |
| 369 | if (f_op && f_op->poll) |
| 370 | mask = (*f_op->poll)(file, retval ? NULL : wait); |
Eric Dumazet | e4a1f12 | 2006-03-28 01:56:34 -0800 | [diff] [blame] | 371 | fput_light(file, fput_needed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 372 | if ((mask & POLLIN_SET) && (in & bit)) { |
| 373 | res_in |= bit; |
| 374 | retval++; |
| 375 | } |
| 376 | if ((mask & POLLOUT_SET) && (out & bit)) { |
| 377 | res_out |= bit; |
| 378 | retval++; |
| 379 | } |
| 380 | if ((mask & POLLEX_SET) && (ex & bit)) { |
| 381 | res_ex |= bit; |
| 382 | retval++; |
| 383 | } |
| 384 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 385 | } |
| 386 | if (res_in) |
| 387 | *rinp = res_in; |
| 388 | if (res_out) |
| 389 | *routp = res_out; |
| 390 | if (res_ex) |
| 391 | *rexp = res_ex; |
Linus Torvalds | 55d8538 | 2008-06-22 12:23:15 -0700 | [diff] [blame] | 392 | cond_resched(); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 393 | } |
| 394 | wait = NULL; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 395 | if (retval || timed_out || signal_pending(current)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 396 | break; |
Pavel Machek | f526448 | 2008-04-21 22:15:06 +0000 | [diff] [blame] | 397 | if (table.error) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 398 | retval = table.error; |
| 399 | break; |
| 400 | } |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 401 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 402 | /* |
| 403 | * If this is the first loop and we have a timeout |
| 404 | * given, then we convert to ktime_t and set the to |
| 405 | * pointer to the expiry value. |
| 406 | */ |
| 407 | if (end_time && !to) { |
| 408 | expire = timespec_to_ktime(*end_time); |
| 409 | to = &expire; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 410 | } |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 411 | |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 412 | if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 413 | timed_out = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 414 | } |
| 415 | __set_current_state(TASK_RUNNING); |
| 416 | |
| 417 | poll_freewait(&table); |
| 418 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 419 | return retval; |
| 420 | } |
| 421 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 422 | /* |
| 423 | * We can actually return ERESTARTSYS instead of EINTR, but I'd |
| 424 | * like to be certain this leads to no problems. So I return |
| 425 | * EINTR just for safety. |
| 426 | * |
| 427 | * Update: ERESTARTSYS breaks at least the xview clock binary, so |
| 428 | * I'm trying ERESTARTNOHAND which restart only when you want to. |
| 429 | */ |
| 430 | #define MAX_SELECT_SECONDS \ |
| 431 | ((unsigned long) (MAX_SCHEDULE_TIMEOUT / HZ)-1) |
| 432 | |
Al Viro | a2dcb44 | 2008-04-23 14:05:15 -0400 | [diff] [blame] | 433 | int core_sys_select(int n, fd_set __user *inp, fd_set __user *outp, |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 434 | fd_set __user *exp, struct timespec *end_time) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 435 | { |
| 436 | fd_set_bits fds; |
Andrew Morton | 29ff2db | 2006-04-10 22:52:46 -0700 | [diff] [blame] | 437 | void *bits; |
Vadim Lobanov | bbea9f6 | 2006-12-10 02:21:12 -0800 | [diff] [blame] | 438 | int ret, max_fds; |
Mitchell Blank Jr | b04eb6a | 2006-04-10 22:54:08 -0700 | [diff] [blame] | 439 | unsigned int size; |
Dipankar Sarma | badf166 | 2005-09-09 13:04:10 -0700 | [diff] [blame] | 440 | struct fdtable *fdt; |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 441 | /* Allocate small arguments on the stack to save memory and be faster */ |
Jes Sorensen | 30c14e4 | 2006-03-31 11:18:57 -0500 | [diff] [blame] | 442 | long stack_fds[SELECT_STACK_ALLOC/sizeof(long)]; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 443 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 444 | ret = -EINVAL; |
| 445 | if (n < 0) |
| 446 | goto out_nofds; |
| 447 | |
Vadim Lobanov | bbea9f6 | 2006-12-10 02:21:12 -0800 | [diff] [blame] | 448 | /* max_fds can increase, so grab it once to avoid race */ |
Dipankar Sarma | b835996 | 2005-09-09 13:04:14 -0700 | [diff] [blame] | 449 | rcu_read_lock(); |
Dipankar Sarma | badf166 | 2005-09-09 13:04:10 -0700 | [diff] [blame] | 450 | fdt = files_fdtable(current->files); |
Vadim Lobanov | bbea9f6 | 2006-12-10 02:21:12 -0800 | [diff] [blame] | 451 | max_fds = fdt->max_fds; |
Dipankar Sarma | b835996 | 2005-09-09 13:04:14 -0700 | [diff] [blame] | 452 | rcu_read_unlock(); |
Vadim Lobanov | bbea9f6 | 2006-12-10 02:21:12 -0800 | [diff] [blame] | 453 | if (n > max_fds) |
| 454 | n = max_fds; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 455 | |
| 456 | /* |
| 457 | * We need 6 bitmaps (in/out/ex for both incoming and outgoing), |
| 458 | * since we used fdset we need to allocate memory in units of |
| 459 | * long-words. |
| 460 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 461 | size = FDS_BYTES(n); |
Mitchell Blank Jr | b04eb6a | 2006-04-10 22:54:08 -0700 | [diff] [blame] | 462 | bits = stack_fds; |
| 463 | if (size > sizeof(stack_fds) / 6) { |
| 464 | /* Not enough space in on-stack array; must use kmalloc */ |
| 465 | ret = -ENOMEM; |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 466 | bits = kmalloc(6 * size, GFP_KERNEL); |
Mitchell Blank Jr | b04eb6a | 2006-04-10 22:54:08 -0700 | [diff] [blame] | 467 | if (!bits) |
| 468 | goto out_nofds; |
| 469 | } |
Andrew Morton | 29ff2db | 2006-04-10 22:52:46 -0700 | [diff] [blame] | 470 | fds.in = bits; |
| 471 | fds.out = bits + size; |
| 472 | fds.ex = bits + 2*size; |
| 473 | fds.res_in = bits + 3*size; |
| 474 | fds.res_out = bits + 4*size; |
| 475 | fds.res_ex = bits + 5*size; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 476 | |
| 477 | if ((ret = get_fd_set(n, inp, fds.in)) || |
| 478 | (ret = get_fd_set(n, outp, fds.out)) || |
| 479 | (ret = get_fd_set(n, exp, fds.ex))) |
| 480 | goto out; |
| 481 | zero_fd_set(n, fds.res_in); |
| 482 | zero_fd_set(n, fds.res_out); |
| 483 | zero_fd_set(n, fds.res_ex); |
| 484 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 485 | ret = do_select(n, &fds, end_time); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 486 | |
| 487 | if (ret < 0) |
| 488 | goto out; |
| 489 | if (!ret) { |
| 490 | ret = -ERESTARTNOHAND; |
| 491 | if (signal_pending(current)) |
| 492 | goto out; |
| 493 | ret = 0; |
| 494 | } |
| 495 | |
| 496 | if (set_fd_set(n, inp, fds.res_in) || |
| 497 | set_fd_set(n, outp, fds.res_out) || |
| 498 | set_fd_set(n, exp, fds.res_ex)) |
| 499 | ret = -EFAULT; |
| 500 | |
| 501 | out: |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 502 | if (bits != stack_fds) |
| 503 | kfree(bits); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 504 | out_nofds: |
| 505 | return ret; |
| 506 | } |
| 507 | |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 508 | asmlinkage long sys_select(int n, fd_set __user *inp, fd_set __user *outp, |
| 509 | fd_set __user *exp, struct timeval __user *tvp) |
| 510 | { |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 511 | struct timespec end_time, *to = NULL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 512 | struct timeval tv; |
| 513 | int ret; |
| 514 | |
| 515 | if (tvp) { |
| 516 | if (copy_from_user(&tv, tvp, sizeof(tv))) |
| 517 | return -EFAULT; |
| 518 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 519 | to = &end_time; |
| 520 | if (poll_select_set_timeout(to, tv.tv_sec, |
| 521 | tv.tv_usec * NSEC_PER_USEC)) |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 522 | return -EINVAL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 523 | } |
| 524 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 525 | ret = core_sys_select(n, inp, outp, exp, to); |
| 526 | ret = poll_select_copy_remaining(&end_time, tvp, 1, ret); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 527 | |
| 528 | return ret; |
| 529 | } |
| 530 | |
Roland McGrath | f3de272 | 2008-04-30 00:53:09 -0700 | [diff] [blame] | 531 | #ifdef HAVE_SET_RESTORE_SIGMASK |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 532 | asmlinkage long sys_pselect7(int n, fd_set __user *inp, fd_set __user *outp, |
| 533 | fd_set __user *exp, struct timespec __user *tsp, |
| 534 | const sigset_t __user *sigmask, size_t sigsetsize) |
| 535 | { |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 536 | sigset_t ksigmask, sigsaved; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 537 | struct timespec ts, end_time, *to = NULL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 538 | int ret; |
| 539 | |
| 540 | if (tsp) { |
| 541 | if (copy_from_user(&ts, tsp, sizeof(ts))) |
| 542 | return -EFAULT; |
| 543 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 544 | to = &end_time; |
| 545 | if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 546 | return -EINVAL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 547 | } |
| 548 | |
| 549 | if (sigmask) { |
| 550 | /* XXX: Don't preclude handling different sized sigset_t's. */ |
| 551 | if (sigsetsize != sizeof(sigset_t)) |
| 552 | return -EINVAL; |
| 553 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) |
| 554 | return -EFAULT; |
| 555 | |
| 556 | sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); |
| 557 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); |
| 558 | } |
| 559 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 560 | ret = core_sys_select(n, inp, outp, exp, &end_time); |
| 561 | ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 562 | |
| 563 | if (ret == -ERESTARTNOHAND) { |
| 564 | /* |
| 565 | * Don't restore the signal mask yet. Let do_signal() deliver |
| 566 | * the signal on the way back to userspace, before the signal |
| 567 | * mask is restored. |
| 568 | */ |
| 569 | if (sigmask) { |
| 570 | memcpy(¤t->saved_sigmask, &sigsaved, |
| 571 | sizeof(sigsaved)); |
Roland McGrath | 4e4c22c | 2008-04-30 00:53:06 -0700 | [diff] [blame] | 572 | set_restore_sigmask(); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 573 | } |
| 574 | } else if (sigmask) |
| 575 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); |
| 576 | |
| 577 | return ret; |
| 578 | } |
| 579 | |
| 580 | /* |
| 581 | * Most architectures can't handle 7-argument syscalls. So we provide a |
| 582 | * 6-argument version where the sixth argument is a pointer to a structure |
| 583 | * which has a pointer to the sigset_t itself followed by a size_t containing |
| 584 | * the sigset size. |
| 585 | */ |
| 586 | asmlinkage long sys_pselect6(int n, fd_set __user *inp, fd_set __user *outp, |
| 587 | fd_set __user *exp, struct timespec __user *tsp, void __user *sig) |
| 588 | { |
| 589 | size_t sigsetsize = 0; |
| 590 | sigset_t __user *up = NULL; |
| 591 | |
| 592 | if (sig) { |
| 593 | if (!access_ok(VERIFY_READ, sig, sizeof(void *)+sizeof(size_t)) |
Al Viro | e110ab9 | 2006-02-01 05:26:09 -0500 | [diff] [blame] | 594 | || __get_user(up, (sigset_t __user * __user *)sig) |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 595 | || __get_user(sigsetsize, |
Al Viro | e110ab9 | 2006-02-01 05:26:09 -0500 | [diff] [blame] | 596 | (size_t __user *)(sig+sizeof(void *)))) |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 597 | return -EFAULT; |
| 598 | } |
| 599 | |
| 600 | return sys_pselect7(n, inp, outp, exp, tsp, up, sigsetsize); |
| 601 | } |
Roland McGrath | f3de272 | 2008-04-30 00:53:09 -0700 | [diff] [blame] | 602 | #endif /* HAVE_SET_RESTORE_SIGMASK */ |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 603 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 604 | struct poll_list { |
| 605 | struct poll_list *next; |
| 606 | int len; |
| 607 | struct pollfd entries[0]; |
| 608 | }; |
| 609 | |
| 610 | #define POLLFD_PER_PAGE ((PAGE_SIZE-sizeof(struct poll_list)) / sizeof(struct pollfd)) |
| 611 | |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 612 | /* |
| 613 | * Fish for pollable events on the pollfd->fd file descriptor. We're only |
| 614 | * interested in events matching the pollfd->events mask, and the result |
| 615 | * matching that mask is both recorded in pollfd->revents and returned. The |
| 616 | * pwait poll_table will be used by the fd-provided poll handler for waiting, |
| 617 | * if non-NULL. |
| 618 | */ |
| 619 | static inline unsigned int do_pollfd(struct pollfd *pollfd, poll_table *pwait) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 620 | { |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 621 | unsigned int mask; |
| 622 | int fd; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 623 | |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 624 | mask = 0; |
| 625 | fd = pollfd->fd; |
| 626 | if (fd >= 0) { |
| 627 | int fput_needed; |
| 628 | struct file * file; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 629 | |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 630 | file = fget_light(fd, &fput_needed); |
| 631 | mask = POLLNVAL; |
| 632 | if (file != NULL) { |
| 633 | mask = DEFAULT_POLLMASK; |
| 634 | if (file->f_op && file->f_op->poll) |
| 635 | mask = file->f_op->poll(file, pwait); |
| 636 | /* Mask out unneeded events. */ |
| 637 | mask &= pollfd->events | POLLERR | POLLHUP; |
| 638 | fput_light(file, fput_needed); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 639 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 640 | } |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 641 | pollfd->revents = mask; |
| 642 | |
| 643 | return mask; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 644 | } |
| 645 | |
| 646 | static int do_poll(unsigned int nfds, struct poll_list *list, |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 647 | struct poll_wqueues *wait, struct timespec *end_time) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 648 | { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 649 | poll_table* pt = &wait->pt; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 650 | ktime_t expire, *to = NULL; |
| 651 | int timed_out = 0, count = 0; |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 652 | unsigned long slack = 0; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 653 | |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 654 | /* Optimise the no-wait case */ |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 655 | if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 656 | pt = NULL; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 657 | timed_out = 1; |
| 658 | } |
Oleg Nesterov | 9bf084f | 2007-10-16 23:26:18 -0700 | [diff] [blame] | 659 | |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 660 | if (end_time) |
| 661 | slack = estimate_accuracy(end_time); |
| 662 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 663 | for (;;) { |
| 664 | struct poll_list *walk; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 665 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 666 | set_current_state(TASK_INTERRUPTIBLE); |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 667 | for (walk = list; walk != NULL; walk = walk->next) { |
| 668 | struct pollfd * pfd, * pfd_end; |
| 669 | |
| 670 | pfd = walk->entries; |
| 671 | pfd_end = pfd + walk->len; |
| 672 | for (; pfd != pfd_end; pfd++) { |
| 673 | /* |
| 674 | * Fish for events. If we found one, record it |
| 675 | * and kill the poll_table, so we don't |
| 676 | * needlessly register any other waiters after |
| 677 | * this. They'll get immediately deregistered |
| 678 | * when we break out and return. |
| 679 | */ |
| 680 | if (do_pollfd(pfd, pt)) { |
| 681 | count++; |
| 682 | pt = NULL; |
| 683 | } |
| 684 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 685 | } |
Vadim Lobanov | 4a4b69f | 2006-06-23 02:05:16 -0700 | [diff] [blame] | 686 | /* |
| 687 | * All waiters have already been registered, so don't provide |
| 688 | * a poll_table to them on the next loop iteration. |
| 689 | */ |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 690 | pt = NULL; |
Oleg Nesterov | 9bf084f | 2007-10-16 23:26:18 -0700 | [diff] [blame] | 691 | if (!count) { |
| 692 | count = wait->error; |
| 693 | if (signal_pending(current)) |
| 694 | count = -EINTR; |
| 695 | } |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 696 | if (count || timed_out) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 697 | break; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 698 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 699 | /* |
| 700 | * If this is the first loop and we have a timeout |
| 701 | * given, then we convert to ktime_t and set the to |
| 702 | * pointer to the expiry value. |
| 703 | */ |
| 704 | if (end_time && !to) { |
| 705 | expire = timespec_to_ktime(*end_time); |
| 706 | to = &expire; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 707 | } |
| 708 | |
Arjan van de Ven | 90d6e24 | 2008-09-01 15:55:35 -0700 | [diff] [blame^] | 709 | if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS)) |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 710 | timed_out = 1; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 711 | } |
| 712 | __set_current_state(TASK_RUNNING); |
| 713 | return count; |
| 714 | } |
| 715 | |
Andi Kleen | 70674f9 | 2006-03-28 01:56:33 -0800 | [diff] [blame] | 716 | #define N_STACK_PPS ((sizeof(stack_pps) - sizeof(struct poll_list)) / \ |
| 717 | sizeof(struct pollfd)) |
| 718 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 719 | int do_sys_poll(struct pollfd __user *ufds, unsigned int nfds, |
| 720 | struct timespec *end_time) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 721 | { |
| 722 | struct poll_wqueues table; |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 723 | int err = -EFAULT, fdcount, len, size; |
Jes Sorensen | 30c14e4 | 2006-03-31 11:18:57 -0500 | [diff] [blame] | 724 | /* Allocate small arguments on the stack to save memory and be |
| 725 | faster - use long to make sure the buffer is aligned properly |
| 726 | on 64 bit archs to avoid unaligned access */ |
| 727 | long stack_pps[POLL_STACK_ALLOC/sizeof(long)]; |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 728 | struct poll_list *const head = (struct poll_list *)stack_pps; |
| 729 | struct poll_list *walk = head; |
| 730 | unsigned long todo = nfds; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 731 | |
Chris Snook | 4e6fd33 | 2006-09-29 02:01:33 -0700 | [diff] [blame] | 732 | if (nfds > current->signal->rlim[RLIMIT_NOFILE].rlim_cur) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 733 | return -EINVAL; |
| 734 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 735 | len = min_t(unsigned int, nfds, N_STACK_PPS); |
| 736 | for (;;) { |
| 737 | walk->next = NULL; |
| 738 | walk->len = len; |
| 739 | if (!len) |
| 740 | break; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 741 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 742 | if (copy_from_user(walk->entries, ufds + nfds-todo, |
| 743 | sizeof(struct pollfd) * walk->len)) |
| 744 | goto out_fds; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 745 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 746 | todo -= walk->len; |
| 747 | if (!todo) |
| 748 | break; |
| 749 | |
| 750 | len = min(todo, POLLFD_PER_PAGE); |
| 751 | size = sizeof(struct poll_list) + sizeof(struct pollfd) * len; |
| 752 | walk = walk->next = kmalloc(size, GFP_KERNEL); |
| 753 | if (!walk) { |
| 754 | err = -ENOMEM; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 755 | goto out_fds; |
| 756 | } |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 757 | } |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 758 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 759 | poll_initwait(&table); |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 760 | fdcount = do_poll(nfds, head, &table, end_time); |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 761 | poll_freewait(&table); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 762 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 763 | for (walk = head; walk; walk = walk->next) { |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 764 | struct pollfd *fds = walk->entries; |
| 765 | int j; |
| 766 | |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 767 | for (j = 0; j < walk->len; j++, ufds++) |
| 768 | if (__put_user(fds[j].revents, &ufds->revents)) |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 769 | goto out_fds; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 770 | } |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 771 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 772 | err = fdcount; |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 773 | out_fds: |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 774 | walk = head->next; |
| 775 | while (walk) { |
| 776 | struct poll_list *pos = walk; |
| 777 | walk = walk->next; |
| 778 | kfree(pos); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 779 | } |
Oleg Nesterov | 252e572 | 2007-10-16 23:26:17 -0700 | [diff] [blame] | 780 | |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 781 | return err; |
| 782 | } |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 783 | |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 784 | static long do_restart_poll(struct restart_block *restart_block) |
| 785 | { |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 786 | struct pollfd __user *ufds = restart_block->poll.ufds; |
| 787 | int nfds = restart_block->poll.nfds; |
| 788 | struct timespec *to = NULL, end_time; |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 789 | int ret; |
| 790 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 791 | if (restart_block->poll.has_timeout) { |
| 792 | end_time.tv_sec = restart_block->poll.tv_sec; |
| 793 | end_time.tv_nsec = restart_block->poll.tv_nsec; |
| 794 | to = &end_time; |
| 795 | } |
| 796 | |
| 797 | ret = do_sys_poll(ufds, nfds, to); |
| 798 | |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 799 | if (ret == -EINTR) { |
| 800 | restart_block->fn = do_restart_poll; |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 801 | ret = -ERESTART_RESTARTBLOCK; |
| 802 | } |
| 803 | return ret; |
| 804 | } |
| 805 | |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 806 | asmlinkage long sys_poll(struct pollfd __user *ufds, unsigned int nfds, |
| 807 | long timeout_msecs) |
| 808 | { |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 809 | struct timespec end_time, *to = NULL; |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 810 | int ret; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 811 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 812 | if (timeout_msecs >= 0) { |
| 813 | to = &end_time; |
| 814 | poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC, |
| 815 | NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC)); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 816 | } |
| 817 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 818 | ret = do_sys_poll(ufds, nfds, to); |
| 819 | |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 820 | if (ret == -EINTR) { |
| 821 | struct restart_block *restart_block; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 822 | |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 823 | restart_block = ¤t_thread_info()->restart_block; |
| 824 | restart_block->fn = do_restart_poll; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 825 | restart_block->poll.ufds = ufds; |
| 826 | restart_block->poll.nfds = nfds; |
| 827 | |
| 828 | if (timeout_msecs >= 0) { |
| 829 | restart_block->poll.tv_sec = end_time.tv_sec; |
| 830 | restart_block->poll.tv_nsec = end_time.tv_nsec; |
| 831 | restart_block->poll.has_timeout = 1; |
| 832 | } else |
| 833 | restart_block->poll.has_timeout = 0; |
| 834 | |
Chris Wright | 3075d9d | 2007-10-16 23:27:18 -0700 | [diff] [blame] | 835 | ret = -ERESTART_RESTARTBLOCK; |
| 836 | } |
| 837 | return ret; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 838 | } |
| 839 | |
Roland McGrath | f3de272 | 2008-04-30 00:53:09 -0700 | [diff] [blame] | 840 | #ifdef HAVE_SET_RESTORE_SIGMASK |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 841 | asmlinkage long sys_ppoll(struct pollfd __user *ufds, unsigned int nfds, |
| 842 | struct timespec __user *tsp, const sigset_t __user *sigmask, |
| 843 | size_t sigsetsize) |
| 844 | { |
| 845 | sigset_t ksigmask, sigsaved; |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 846 | struct timespec ts, end_time, *to = NULL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 847 | int ret; |
| 848 | |
| 849 | if (tsp) { |
| 850 | if (copy_from_user(&ts, tsp, sizeof(ts))) |
| 851 | return -EFAULT; |
| 852 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 853 | to = &end_time; |
| 854 | if (poll_select_set_timeout(to, ts.tv_sec, ts.tv_nsec)) |
| 855 | return -EINVAL; |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 856 | } |
| 857 | |
| 858 | if (sigmask) { |
| 859 | /* XXX: Don't preclude handling different sized sigset_t's. */ |
| 860 | if (sigsetsize != sizeof(sigset_t)) |
| 861 | return -EINVAL; |
| 862 | if (copy_from_user(&ksigmask, sigmask, sizeof(ksigmask))) |
| 863 | return -EFAULT; |
| 864 | |
| 865 | sigdelsetmask(&ksigmask, sigmask(SIGKILL)|sigmask(SIGSTOP)); |
| 866 | sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved); |
| 867 | } |
| 868 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 869 | ret = do_sys_poll(ufds, nfds, to); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 870 | |
| 871 | /* We can restart this syscall, usually */ |
| 872 | if (ret == -EINTR) { |
| 873 | /* |
| 874 | * Don't restore the signal mask yet. Let do_signal() deliver |
| 875 | * the signal on the way back to userspace, before the signal |
| 876 | * mask is restored. |
| 877 | */ |
| 878 | if (sigmask) { |
| 879 | memcpy(¤t->saved_sigmask, &sigsaved, |
| 880 | sizeof(sigsaved)); |
Roland McGrath | 4e4c22c | 2008-04-30 00:53:06 -0700 | [diff] [blame] | 881 | set_restore_sigmask(); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 882 | } |
| 883 | ret = -ERESTARTNOHAND; |
| 884 | } else if (sigmask) |
| 885 | sigprocmask(SIG_SETMASK, &sigsaved, NULL); |
| 886 | |
Arjan van de Ven | 8ff3e8e | 2008-08-31 08:26:40 -0700 | [diff] [blame] | 887 | ret = poll_select_copy_remaining(&end_time, tsp, 0, ret); |
David Woodhouse | 9f72949 | 2006-01-18 17:44:05 -0800 | [diff] [blame] | 888 | |
| 889 | return ret; |
| 890 | } |
Roland McGrath | f3de272 | 2008-04-30 00:53:09 -0700 | [diff] [blame] | 891 | #endif /* HAVE_SET_RESTORE_SIGMASK */ |