| 1 | /* |
| 2 | * Implementation of select and poll |
| 3 | * |
| 4 | * Copyright 2011-2012 Intel Corporation. |
| 5 | * |
| 6 | * This file is a derivative of fs/select.c from within the Linux kernel |
| 7 | * source distribution, version 2.6.34; it has been modified (starting |
| 8 | * in May 2011) to work within the context of the SCIF driver. |
| 9 | * |
| 10 | * This program is free software; you can redistribute it and/or modify |
| 11 | * it under the terms of the GNU General Public License, version 2, as |
| 12 | * published by the Free Software Foundation. |
| 13 | * |
| 14 | * This program is distributed in the hope that it will be useful, but |
| 15 | * WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 17 | * General Public License for more details. |
| 18 | * |
| 19 | * You should have received a copy of the GNU General Public License |
| 20 | * along with this program; if not, write to the Free Software |
| 21 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 |
| 22 | * USA. |
| 23 | * |
| 24 | * Initial comment from fs/select.c: |
| 25 | * |
| 26 | * This file contains the procedures for the handling of select and poll |
| 27 | * |
| 28 | * Created for Linux based loosely upon Mathius Lattner's minix |
| 29 | * patches by Peter MacDonald. Heavily edited by Linus. |
| 30 | * |
| 31 | * 4 February 1994 |
| 32 | * COFF/ELF binary emulation. If the process has the STICKY_TIMEOUTS |
| 33 | * flag set in its personality we do *not* modify the given timeout |
| 34 | * parameter to reflect time remaining. |
| 35 | * |
| 36 | * 24 January 2000 |
| 37 | * Changed sys_poll()/do_poll() to use PAGE_SIZE chunk-based allocation |
| 38 | * of fds to overcome nfds < 16390 descriptors limit (Tigran Aivazian). |
| 39 | */ |
| 40 | |
| 41 | #include <linux/kernel.h> |
| 42 | #include <linux/sched.h> |
| 43 | #include <linux/file.h> |
| 44 | #include <linux/hrtimer.h> |
| 45 | #include <linux/module.h> |
| 46 | |
| 47 | #include "mic/micscif.h" |
| 48 | |
| 49 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 50 | #include <linux/sched/rt.h> |
| 51 | #endif |
| 52 | |
| 53 | struct poll_table_page { |
| 54 | struct poll_table_page *next; |
| 55 | struct poll_table_entry *entry; |
| 56 | struct poll_table_entry entries[0]; |
| 57 | }; |
| 58 | |
| 59 | /* |
| 60 | * Estimate expected accuracy in ns from a timeval. |
| 61 | * |
| 62 | * After quite a bit of churning around, we've settled on |
| 63 | * a simple thing of taking 0.1% of the timeout as the |
| 64 | * slack, with a cap of 100 msec. |
| 65 | * "nice" tasks get a 0.5% slack instead. |
| 66 | * |
| 67 | * Consider this comment an open invitation to come up with even |
| 68 | * better solutions.. |
| 69 | */ |
| 70 | |
| 71 | #define MAX_SLACK (100 * NSEC_PER_MSEC) |
| 72 | |
| 73 | static long __estimate_accuracy(struct timespec *tv) |
| 74 | { |
| 75 | long slack; |
| 76 | int divfactor = 1000; |
| 77 | |
| 78 | if (tv->tv_sec < 0) |
| 79 | return 0; |
| 80 | |
| 81 | if (task_nice(current) > 0) |
| 82 | divfactor = divfactor / 5; |
| 83 | |
| 84 | if (tv->tv_sec > MAX_SLACK / (NSEC_PER_SEC/divfactor)) |
| 85 | return MAX_SLACK; |
| 86 | |
| 87 | slack = tv->tv_nsec / divfactor; |
| 88 | slack += tv->tv_sec * (NSEC_PER_SEC/divfactor); |
| 89 | |
| 90 | if (slack > MAX_SLACK) |
| 91 | return MAX_SLACK; |
| 92 | |
| 93 | return slack; |
| 94 | } |
| 95 | |
| 96 | static long estimate_accuracy(struct timespec *tv) |
| 97 | { |
| 98 | unsigned long ret; |
| 99 | struct timespec now; |
| 100 | |
| 101 | /* |
| 102 | * Realtime tasks get a slack of 0 for obvious reasons. |
| 103 | */ |
| 104 | |
| 105 | if (rt_task(current)) |
| 106 | return 0; |
| 107 | |
| 108 | ktime_get_ts(&now); |
| 109 | now = timespec_sub(*tv, now); |
| 110 | ret = __estimate_accuracy(&now); |
| 111 | if (ret < current->timer_slack_ns) |
| 112 | return current->timer_slack_ns; |
| 113 | return ret; |
| 114 | } |
| 115 | |
| 116 | #define POLL_TABLE_FULL(table) \ |
| 117 | ((unsigned long)((table)->entry+1) > PAGE_SIZE + (unsigned long)(table)) |
| 118 | |
| 119 | /* |
| 120 | * Ok, Peter made a complicated, but straightforward multiple_wait() function. |
| 121 | * I have rewritten this, taking some shortcuts: This code may not be easy to |
| 122 | * follow, but it should be free of race-conditions, and it's practical. If you |
| 123 | * understand what I'm doing here, then you understand how the linux |
| 124 | * sleep/wakeup mechanism works. |
| 125 | * |
| 126 | * Two very simple procedures, poll_wait() and poll_freewait() make all the |
| 127 | * work. poll_wait() is an inline-function defined in <linux/poll.h>, |
| 128 | * as all select/poll functions have to call it to add an entry to the |
| 129 | * poll table. |
| 130 | */ |
| 131 | static void __pollwait(struct file *filp __attribute__((unused)), wait_queue_head_t *wait_address, |
| 132 | poll_table *p); |
| 133 | |
| 134 | static void scif_poll_initwait(struct poll_wqueues *pwq) |
| 135 | { |
| 136 | init_poll_funcptr(&pwq->pt, __pollwait); |
| 137 | pwq->polling_task = current; |
| 138 | pwq->triggered = 0; |
| 139 | pwq->error = 0; |
| 140 | pwq->table = NULL; |
| 141 | pwq->inline_index = 0; |
| 142 | } |
| 143 | |
| 144 | static void free_poll_entry(struct poll_table_entry *entry) |
| 145 | { |
| 146 | remove_wait_queue(entry->wait_address, &entry->wait); |
| 147 | } |
| 148 | |
| 149 | static void scif_poll_freewait(struct poll_wqueues *pwq) |
| 150 | { |
| 151 | struct poll_table_page * p = pwq->table; |
| 152 | int i; |
| 153 | for (i = 0; i < pwq->inline_index; i++) |
| 154 | free_poll_entry(pwq->inline_entries + i); |
| 155 | while (p) { |
| 156 | struct poll_table_entry *entry; |
| 157 | struct poll_table_page *old; |
| 158 | |
| 159 | entry = p->entry; |
| 160 | do { |
| 161 | entry--; |
| 162 | free_poll_entry(entry); |
| 163 | } while (entry > p->entries); |
| 164 | old = p; |
| 165 | p = p->next; |
| 166 | free_page((unsigned long) old); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | static struct poll_table_entry *poll_get_entry(struct poll_wqueues *p) |
| 171 | { |
| 172 | struct poll_table_page *table = p->table; |
| 173 | |
| 174 | if (p->inline_index < N_INLINE_POLL_ENTRIES) |
| 175 | return p->inline_entries + p->inline_index++; |
| 176 | |
| 177 | if (!table || POLL_TABLE_FULL(table)) { |
| 178 | struct poll_table_page *new_table; |
| 179 | |
| 180 | new_table = (struct poll_table_page *) __get_free_page(GFP_KERNEL); |
| 181 | if (!new_table) { |
| 182 | p->error = -ENOMEM; |
| 183 | return NULL; |
| 184 | } |
| 185 | new_table->entry = new_table->entries; |
| 186 | new_table->next = table; |
| 187 | p->table = new_table; |
| 188 | table = new_table; |
| 189 | } |
| 190 | |
| 191 | return table->entry++; |
| 192 | } |
| 193 | |
| 194 | static int __pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key) |
| 195 | { |
| 196 | struct poll_wqueues *pwq = wait->private; |
| 197 | DECLARE_WAITQUEUE(dummy_wait, pwq->polling_task); |
| 198 | |
| 199 | /* |
| 200 | * Although this function is called under waitqueue lock, LOCK |
| 201 | * doesn't imply write barrier and the users expect write |
| 202 | * barrier semantics on wakeup functions. The following |
| 203 | * smp_wmb() is equivalent to smp_wmb() in try_to_wake_up() |
| 204 | * and is paired with set_mb() in poll_schedule_timeout. |
| 205 | */ |
| 206 | smp_wmb(); |
| 207 | pwq->triggered = 1; |
| 208 | |
| 209 | /* |
| 210 | * Perform the default wake up operation using a dummy |
| 211 | * waitqueue. |
| 212 | * |
| 213 | * TODO: This is hacky but there currently is no interface to |
| 214 | * pass in @sync. @sync is scheduled to be removed and once |
| 215 | * that happens, wake_up_process() can be used directly. |
| 216 | */ |
| 217 | return default_wake_function(&dummy_wait, mode, sync, key); |
| 218 | } |
| 219 | |
| 220 | static int pollwake(wait_queue_t *wait, unsigned mode, int sync, void *key) |
| 221 | { |
| 222 | struct poll_table_entry *entry; |
| 223 | |
| 224 | entry = container_of(wait, struct poll_table_entry, wait); |
| 225 | if (key && !((unsigned long)key & entry->key)) |
| 226 | return 0; |
| 227 | return __pollwake(wait, mode, sync, key); |
| 228 | } |
| 229 | |
| 230 | /* Add a new entry */ |
| 231 | static void __pollwait(struct file *filp __attribute__((unused)), wait_queue_head_t *wait_address, |
| 232 | poll_table *p) |
| 233 | { |
| 234 | struct poll_wqueues *pwq = container_of(p, struct poll_wqueues, pt); |
| 235 | struct poll_table_entry *entry = poll_get_entry(pwq); |
| 236 | if (!entry) |
| 237 | return; |
| 238 | entry->filp = NULL; |
| 239 | entry->wait_address = wait_address; |
| 240 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 241 | entry->key = p->_key; |
| 242 | #else |
| 243 | entry->key = p->key; |
| 244 | #endif |
| 245 | init_waitqueue_func_entry(&entry->wait, pollwake); |
| 246 | entry->wait.private = pwq; |
| 247 | add_wait_queue(wait_address, &entry->wait); |
| 248 | } |
| 249 | |
| 250 | int poll_schedule_timeout(struct poll_wqueues *pwq, int state, |
| 251 | ktime_t *expires, unsigned long slack) |
| 252 | { |
| 253 | int rc = -EINTR; |
| 254 | |
| 255 | set_current_state(state); |
| 256 | if (!pwq->triggered) |
| 257 | rc = schedule_hrtimeout_range(expires, slack, HRTIMER_MODE_ABS); |
| 258 | __set_current_state(TASK_RUNNING); |
| 259 | |
| 260 | /* |
| 261 | * Prepare for the next iteration. |
| 262 | * |
| 263 | * The following set_mb() serves two purposes. First, it's |
| 264 | * the counterpart rmb of the wmb in pollwake() such that data |
| 265 | * written before wake up is always visible after wake up. |
| 266 | * Second, the full barrier guarantees that triggered clearing |
| 267 | * doesn't pass event check of the next iteration. Note that |
| 268 | * this problem doesn't exist for the first iteration as |
| 269 | * add_wait_queue() has full barrier semantics. |
| 270 | */ |
| 271 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(4,2,0)) |
| 272 | smp_store_mb(pwq->triggered, 0); |
| 273 | #else |
| 274 | set_mb(pwq->triggered, 0); |
| 275 | #endif |
| 276 | |
| 277 | return rc; |
| 278 | } |
| 279 | |
| 280 | static unsigned int scif_poll_kernel(poll_table *pwait, struct endpt *ep) |
| 281 | { |
| 282 | return __scif_pollfd(NULL, pwait, ep); |
| 283 | } |
| 284 | |
| 285 | /* |
| 286 | * Fish for pollable events on the pollfd->fd file descriptor. We're only |
| 287 | * interested in events matching the pollfd->events mask, and the result |
| 288 | * matching that mask is both recorded in pollfd->revents and returned. The |
| 289 | * pwait poll_table will be used by the fd-provided poll handler for waiting, |
| 290 | * if non-NULL. |
| 291 | */ |
| 292 | static inline unsigned int do_pollfd(struct scif_pollepd *pollfd, poll_table *pwait) |
| 293 | { |
| 294 | unsigned int mask; |
| 295 | scif_epd_t epd; |
| 296 | |
| 297 | mask = 0; |
| 298 | epd = pollfd->epd; |
| 299 | if (epd) { |
| 300 | mask = POLLNVAL; |
| 301 | mask = DEFAULT_POLLMASK; |
| 302 | if (pwait) |
| 303 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,10,0)) |
| 304 | pwait->_key = pollfd->events | POLLERR | POLLHUP; |
| 305 | #else |
| 306 | pwait->key = pollfd->events | POLLERR | POLLHUP; |
| 307 | #endif |
| 308 | mask = scif_poll_kernel(pwait, epd); |
| 309 | /* Mask out unneeded events. */ |
| 310 | mask &= pollfd->events | POLLERR | POLLHUP; |
| 311 | } |
| 312 | pollfd->revents = mask; |
| 313 | |
| 314 | return mask; |
| 315 | } |
| 316 | |
| 317 | static int do_poll(unsigned int nfds, struct scif_pollepd *ufds, |
| 318 | struct poll_wqueues *wait, struct timespec *end_time) |
| 319 | { |
| 320 | poll_table* pt = &wait->pt; |
| 321 | ktime_t expire, *to = NULL; |
| 322 | int timed_out = 0, count = 0, i = 0; |
| 323 | unsigned long slack = 0; |
| 324 | |
| 325 | /* Optimise the no-wait case */ |
| 326 | if (end_time && !end_time->tv_sec && !end_time->tv_nsec) { |
| 327 | pt = NULL; |
| 328 | timed_out = 1; |
| 329 | } |
| 330 | |
| 331 | if (end_time && !timed_out) |
| 332 | slack = estimate_accuracy(end_time); |
| 333 | |
| 334 | for (;;) { |
| 335 | for (i = 0; i < nfds; i++) { |
| 336 | /* |
| 337 | * Fish for events. If we found one, record it |
| 338 | * and kill the poll_table, so we don't |
| 339 | * needlessly register any other waiters after |
| 340 | * this. They'll get immediately deregistered |
| 341 | * when we break out and return. |
| 342 | */ |
| 343 | if (do_pollfd(ufds + i, pt)) { |
| 344 | count++; |
| 345 | pt = NULL; |
| 346 | } |
| 347 | } |
| 348 | /* |
| 349 | * All waiters have already been registered, so don't provide |
| 350 | * a poll_table to them on the next loop iteration. |
| 351 | */ |
| 352 | pt = NULL; |
| 353 | if (!count) { |
| 354 | count = wait->error; |
| 355 | if (signal_pending(current)) |
| 356 | count = -EINTR; |
| 357 | } |
| 358 | if (count || timed_out) |
| 359 | break; |
| 360 | |
| 361 | /* |
| 362 | * If this is the first loop and we have a timeout |
| 363 | * given, then we convert to ktime_t and set the to |
| 364 | * pointer to the expiry value. |
| 365 | */ |
| 366 | if (end_time && !to) { |
| 367 | expire = timespec_to_ktime(*end_time); |
| 368 | to = &expire; |
| 369 | } |
| 370 | |
| 371 | if (!poll_schedule_timeout(wait, TASK_INTERRUPTIBLE, to, slack)) |
| 372 | timed_out = 1; |
| 373 | } |
| 374 | return count; |
| 375 | } |
| 376 | |
| 377 | static int do_scif_poll(struct scif_pollepd *ufds, unsigned int nfds, |
| 378 | struct timespec *end_time) |
| 379 | { |
| 380 | struct poll_wqueues table; |
| 381 | int epdcount; |
| 382 | |
| 383 | scif_poll_initwait(&table); |
| 384 | epdcount = do_poll(nfds, ufds, &table, end_time); |
| 385 | scif_poll_freewait(&table); |
| 386 | |
| 387 | return epdcount; |
| 388 | } |
| 389 | |
| 390 | /* |
| 391 | * Add two timespec values and do a safety check for overflow. |
| 392 | * It's assumed that both values are valid (>= 0) |
| 393 | */ |
| 394 | static struct timespec scif_timespec_add_safe(const struct timespec lhs, |
| 395 | const struct timespec rhs) |
| 396 | { |
| 397 | struct timespec res; |
| 398 | |
| 399 | set_normalized_timespec(&res, lhs.tv_sec + rhs.tv_sec, |
| 400 | lhs.tv_nsec + rhs.tv_nsec); |
| 401 | |
| 402 | if (res.tv_sec < lhs.tv_sec || res.tv_sec < rhs.tv_sec) |
| 403 | res.tv_sec = TIME_T_MAX; |
| 404 | |
| 405 | return res; |
| 406 | } |
| 407 | /** |
| 408 | * poll_select_set_timeout - helper function to setup the timeout value |
| 409 | * @to: pointer to timespec variable for the final timeout |
| 410 | * @sec: seconds (from user space) |
| 411 | * @nsec: nanoseconds (from user space) |
| 412 | * |
| 413 | * Note, we do not use a timespec for the user space value here, That |
| 414 | * way we can use the function for timeval and compat interfaces as well. |
| 415 | * |
| 416 | * Returns -EINVAL if sec/nsec are not normalized. Otherwise 0. |
| 417 | */ |
| 418 | static int scif_poll_select_set_timeout(struct timespec *to, long sec, long nsec) |
| 419 | { |
| 420 | struct timespec ts = {.tv_sec = sec, .tv_nsec = nsec}; |
| 421 | |
| 422 | if (!timespec_valid(&ts)) |
| 423 | return -EINVAL; |
| 424 | |
| 425 | /* Optimize for the zero timeout value here */ |
| 426 | if (!sec && !nsec) { |
| 427 | to->tv_sec = to->tv_nsec = 0; |
| 428 | } else { |
| 429 | ktime_get_ts(to); |
| 430 | *to = scif_timespec_add_safe(*to, ts); |
| 431 | } |
| 432 | return 0; |
| 433 | } |
| 434 | |
| 435 | int scif_poll(struct scif_pollepd *ufds, unsigned int nfds, long timeout_msecs) |
| 436 | { |
| 437 | struct timespec end_time, *to = NULL; |
| 438 | if (timeout_msecs >= 0) { |
| 439 | to = &end_time; |
| 440 | scif_poll_select_set_timeout(to, timeout_msecs / MSEC_PER_SEC, |
| 441 | NSEC_PER_MSEC * (timeout_msecs % MSEC_PER_SEC)); |
| 442 | } |
| 443 | |
| 444 | return do_scif_poll(ufds, nfds, to); |
| 445 | } |
| 446 | EXPORT_SYMBOL(scif_poll); |