1/* 2 * GPL HEADER START 3 * 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 only, 8 * as published by the Free Software Foundation. 9 * 10 * This program is distributed in the hope that it will be useful, but 11 * WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 13 * General Public License version 2 for more details (a copy is included 14 * in the LICENSE file that accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License 17 * version 2 along with this program; If not, see 18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf 19 * 20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 21 * CA 95054 USA or visit www.sun.com if you need additional information or 22 * have any questions. 23 * 24 * GPL HEADER END 25 */ 26/* 27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved. 28 * Use is subject to license terms. 29 * 30 * Copyright (c) 2012, Intel Corporation. 31 */ 32/* 33 * This file is part of Lustre, http://www.lustre.org/ 34 * Lustre is a trademark of Sun Microsystems, Inc. 35 * 36 * lnet/lnet/lib-eq.c 37 * 38 * Library level Event queue management routines 39 */ 40 41#define DEBUG_SUBSYSTEM S_LNET 42#include "../../include/linux/lnet/lib-lnet.h" 43 44/** 45 * Create an event queue that has room for \a count number of events. 46 * 47 * The event queue is circular and older events will be overwritten by new 48 * ones if they are not removed in time by the user using the functions 49 * LNetEQGet(), LNetEQWait(), or LNetEQPoll(). It is up to the user to 50 * determine the appropriate size of the event queue to prevent this loss 51 * of events. Note that when EQ handler is specified in \a callback, no 52 * event loss can happen, since the handler is run for each event deposited 53 * into the EQ. 54 * 55 * \param count The number of events to be stored in the event queue. It 56 * will be rounded up to the next power of two. 57 * \param callback A handler function that runs when an event is deposited 58 * into the EQ. The constant value LNET_EQ_HANDLER_NONE can be used to 59 * indicate that no event handler is desired. 60 * \param handle On successful return, this location will hold a handle for 61 * the newly created EQ. 62 * 63 * \retval 0 On success. 64 * \retval -EINVAL If an parameter is not valid. 65 * \retval -ENOMEM If memory for the EQ can't be allocated. 66 * 67 * \see lnet_eq_handler_t for the discussion on EQ handler semantics. 68 */ 69int 70LNetEQAlloc(unsigned int count, lnet_eq_handler_t callback, 71 lnet_handle_eq_t *handle) 72{ 73 lnet_eq_t *eq; 74 75 LASSERT(the_lnet.ln_refcount > 0); 76 77 /* 78 * We need count to be a power of 2 so that when eq_{enq,deq}_seq 79 * overflow, they don't skip entries, so the queue has the same 80 * apparent capacity at all times 81 */ 82 if (count) 83 count = roundup_pow_of_two(count); 84 85 if (callback != LNET_EQ_HANDLER_NONE && count) 86 CWARN("EQ callback is guaranteed to get every event, do you still want to set eqcount %d for polling event which will have locking overhead? Please contact with developer to confirm\n", count); 87 88 /* 89 * count can be 0 if only need callback, we can eliminate 90 * overhead of enqueue event 91 */ 92 if (!count && callback == LNET_EQ_HANDLER_NONE) 93 return -EINVAL; 94 95 eq = lnet_eq_alloc(); 96 if (!eq) 97 return -ENOMEM; 98 99 if (count) { 100 LIBCFS_ALLOC(eq->eq_events, count * sizeof(lnet_event_t)); 101 if (!eq->eq_events) 102 goto failed; 103 /* 104 * NB allocator has set all event sequence numbers to 0, 105 * so all them should be earlier than eq_deq_seq 106 */ 107 } 108 109 eq->eq_deq_seq = 1; 110 eq->eq_enq_seq = 1; 111 eq->eq_size = count; 112 eq->eq_callback = callback; 113 114 eq->eq_refs = cfs_percpt_alloc(lnet_cpt_table(), 115 sizeof(*eq->eq_refs[0])); 116 if (!eq->eq_refs) 117 goto failed; 118 119 /* MUST hold both exclusive lnet_res_lock */ 120 lnet_res_lock(LNET_LOCK_EX); 121 /* 122 * NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do 123 * both EQ lookup and poll event with only lnet_eq_wait_lock 124 */ 125 lnet_eq_wait_lock(); 126 127 lnet_res_lh_initialize(&the_lnet.ln_eq_container, &eq->eq_lh); 128 list_add(&eq->eq_list, &the_lnet.ln_eq_container.rec_active); 129 130 lnet_eq_wait_unlock(); 131 lnet_res_unlock(LNET_LOCK_EX); 132 133 lnet_eq2handle(handle, eq); 134 return 0; 135 136failed: 137 if (eq->eq_events) 138 LIBCFS_FREE(eq->eq_events, count * sizeof(lnet_event_t)); 139 140 if (eq->eq_refs) 141 cfs_percpt_free(eq->eq_refs); 142 143 lnet_eq_free(eq); 144 return -ENOMEM; 145} 146EXPORT_SYMBOL(LNetEQAlloc); 147 148/** 149 * Release the resources associated with an event queue if it's idle; 150 * otherwise do nothing and it's up to the user to try again. 151 * 152 * \param eqh A handle for the event queue to be released. 153 * 154 * \retval 0 If the EQ is not in use and freed. 155 * \retval -ENOENT If \a eqh does not point to a valid EQ. 156 * \retval -EBUSY If the EQ is still in use by some MDs. 157 */ 158int 159LNetEQFree(lnet_handle_eq_t eqh) 160{ 161 struct lnet_eq *eq; 162 lnet_event_t *events = NULL; 163 int **refs = NULL; 164 int *ref; 165 int rc = 0; 166 int size = 0; 167 int i; 168 169 LASSERT(the_lnet.ln_refcount > 0); 170 171 lnet_res_lock(LNET_LOCK_EX); 172 /* 173 * NB: hold lnet_eq_wait_lock for EQ link/unlink, so we can do 174 * both EQ lookup and poll event with only lnet_eq_wait_lock 175 */ 176 lnet_eq_wait_lock(); 177 178 eq = lnet_handle2eq(&eqh); 179 if (!eq) { 180 rc = -ENOENT; 181 goto out; 182 } 183 184 cfs_percpt_for_each(ref, i, eq->eq_refs) { 185 LASSERT(*ref >= 0); 186 if (!*ref) 187 continue; 188 189 CDEBUG(D_NET, "Event equeue (%d: %d) busy on destroy.\n", 190 i, *ref); 191 rc = -EBUSY; 192 goto out; 193 } 194 195 /* stash for free after lock dropped */ 196 events = eq->eq_events; 197 size = eq->eq_size; 198 refs = eq->eq_refs; 199 200 lnet_res_lh_invalidate(&eq->eq_lh); 201 list_del(&eq->eq_list); 202 lnet_eq_free(eq); 203 out: 204 lnet_eq_wait_unlock(); 205 lnet_res_unlock(LNET_LOCK_EX); 206 207 if (events) 208 LIBCFS_FREE(events, size * sizeof(lnet_event_t)); 209 if (refs) 210 cfs_percpt_free(refs); 211 212 return rc; 213} 214EXPORT_SYMBOL(LNetEQFree); 215 216void 217lnet_eq_enqueue_event(lnet_eq_t *eq, lnet_event_t *ev) 218{ 219 /* MUST called with resource lock hold but w/o lnet_eq_wait_lock */ 220 int index; 221 222 if (!eq->eq_size) { 223 LASSERT(eq->eq_callback != LNET_EQ_HANDLER_NONE); 224 eq->eq_callback(ev); 225 return; 226 } 227 228 lnet_eq_wait_lock(); 229 ev->sequence = eq->eq_enq_seq++; 230 231 LASSERT(eq->eq_size == LOWEST_BIT_SET(eq->eq_size)); 232 index = ev->sequence & (eq->eq_size - 1); 233 234 eq->eq_events[index] = *ev; 235 236 if (eq->eq_callback != LNET_EQ_HANDLER_NONE) 237 eq->eq_callback(ev); 238 239 /* Wake anyone waiting in LNetEQPoll() */ 240 if (waitqueue_active(&the_lnet.ln_eq_waitq)) 241 wake_up_all(&the_lnet.ln_eq_waitq); 242 lnet_eq_wait_unlock(); 243} 244 245static int 246lnet_eq_dequeue_event(lnet_eq_t *eq, lnet_event_t *ev) 247{ 248 int new_index = eq->eq_deq_seq & (eq->eq_size - 1); 249 lnet_event_t *new_event = &eq->eq_events[new_index]; 250 int rc; 251 252 /* must called with lnet_eq_wait_lock hold */ 253 if (LNET_SEQ_GT(eq->eq_deq_seq, new_event->sequence)) 254 return 0; 255 256 /* We've got a new event... */ 257 *ev = *new_event; 258 259 CDEBUG(D_INFO, "event: %p, sequence: %lu, eq->size: %u\n", 260 new_event, eq->eq_deq_seq, eq->eq_size); 261 262 /* ...but did it overwrite an event we've not seen yet? */ 263 if (eq->eq_deq_seq == new_event->sequence) { 264 rc = 1; 265 } else { 266 /* 267 * don't complain with CERROR: some EQs are sized small 268 * anyway; if it's important, the caller should complain 269 */ 270 CDEBUG(D_NET, "Event Queue Overflow: eq seq %lu ev seq %lu\n", 271 eq->eq_deq_seq, new_event->sequence); 272 rc = -EOVERFLOW; 273 } 274 275 eq->eq_deq_seq = new_event->sequence + 1; 276 return rc; 277} 278 279/** 280 * A nonblocking function that can be used to get the next event in an EQ. 281 * If an event handler is associated with the EQ, the handler will run before 282 * this function returns successfully. The event is removed from the queue. 283 * 284 * \param eventq A handle for the event queue. 285 * \param event On successful return (1 or -EOVERFLOW), this location will 286 * hold the next event in the EQ. 287 * 288 * \retval 0 No pending event in the EQ. 289 * \retval 1 Indicates success. 290 * \retval -ENOENT If \a eventq does not point to a valid EQ. 291 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that 292 * at least one event between this event and the last event obtained from the 293 * EQ has been dropped due to limited space in the EQ. 294 */ 295 296/** 297 * Block the calling process until there is an event in the EQ. 298 * If an event handler is associated with the EQ, the handler will run before 299 * this function returns successfully. This function returns the next event 300 * in the EQ and removes it from the EQ. 301 * 302 * \param eventq A handle for the event queue. 303 * \param event On successful return (1 or -EOVERFLOW), this location will 304 * hold the next event in the EQ. 305 * 306 * \retval 1 Indicates success. 307 * \retval -ENOENT If \a eventq does not point to a valid EQ. 308 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that 309 * at least one event between this event and the last event obtained from the 310 * EQ has been dropped due to limited space in the EQ. 311 */ 312 313static int 314lnet_eq_wait_locked(int *timeout_ms) 315__must_hold(&the_lnet.ln_eq_wait_lock) 316{ 317 int tms = *timeout_ms; 318 int wait; 319 wait_queue_t wl; 320 unsigned long now; 321 322 if (!tms) 323 return -ENXIO; /* don't want to wait and no new event */ 324 325 init_waitqueue_entry(&wl, current); 326 set_current_state(TASK_INTERRUPTIBLE); 327 add_wait_queue(&the_lnet.ln_eq_waitq, &wl); 328 329 lnet_eq_wait_unlock(); 330 331 if (tms < 0) { 332 schedule(); 333 } else { 334 now = jiffies; 335 schedule_timeout(msecs_to_jiffies(tms)); 336 tms -= jiffies_to_msecs(jiffies - now); 337 if (tms < 0) /* no more wait but may have new event */ 338 tms = 0; 339 } 340 341 wait = tms; /* might need to call here again */ 342 *timeout_ms = tms; 343 344 lnet_eq_wait_lock(); 345 remove_wait_queue(&the_lnet.ln_eq_waitq, &wl); 346 347 return wait; 348} 349 350/** 351 * Block the calling process until there's an event from a set of EQs or 352 * timeout happens. 353 * 354 * If an event handler is associated with the EQ, the handler will run before 355 * this function returns successfully, in which case the corresponding event 356 * is consumed. 357 * 358 * LNetEQPoll() provides a timeout to allow applications to poll, block for a 359 * fixed period, or block indefinitely. 360 * 361 * \param eventqs,neq An array of EQ handles, and size of the array. 362 * \param timeout_ms Time in milliseconds to wait for an event to occur on 363 * one of the EQs. The constant LNET_TIME_FOREVER can be used to indicate an 364 * infinite timeout. 365 * \param event,which On successful return (1 or -EOVERFLOW), \a event will 366 * hold the next event in the EQs, and \a which will contain the index of the 367 * EQ from which the event was taken. 368 * 369 * \retval 0 No pending event in the EQs after timeout. 370 * \retval 1 Indicates success. 371 * \retval -EOVERFLOW Indicates success (i.e., an event is returned) and that 372 * at least one event between this event and the last event obtained from the 373 * EQ indicated by \a which has been dropped due to limited space in the EQ. 374 * \retval -ENOENT If there's an invalid handle in \a eventqs. 375 */ 376int 377LNetEQPoll(lnet_handle_eq_t *eventqs, int neq, int timeout_ms, 378 lnet_event_t *event, int *which) 379{ 380 int wait = 1; 381 int rc; 382 int i; 383 384 LASSERT(the_lnet.ln_refcount > 0); 385 386 if (neq < 1) 387 return -ENOENT; 388 389 lnet_eq_wait_lock(); 390 391 for (;;) { 392 for (i = 0; i < neq; i++) { 393 lnet_eq_t *eq = lnet_handle2eq(&eventqs[i]); 394 395 if (!eq) { 396 lnet_eq_wait_unlock(); 397 return -ENOENT; 398 } 399 400 rc = lnet_eq_dequeue_event(eq, event); 401 if (rc) { 402 lnet_eq_wait_unlock(); 403 *which = i; 404 return rc; 405 } 406 } 407 408 if (!wait) 409 break; 410 411 /* 412 * return value of lnet_eq_wait_locked: 413 * -1 : did nothing and it's sure no new event 414 * 1 : sleep inside and wait until new event 415 * 0 : don't want to wait anymore, but might have new event 416 * so need to call dequeue again 417 */ 418 wait = lnet_eq_wait_locked(&timeout_ms); 419 if (wait < 0) /* no new event */ 420 break; 421 } 422 423 lnet_eq_wait_unlock(); 424 return 0; 425} 426