| 1 | /* |
| 2 | * Copyright (c) 2009, 2010 The DragonFly Project. All rights reserved. |
| 3 | * |
| 4 | * This code is derived from software contributed to The DragonFly Project |
| 5 | * by Alex Hornung <ahornung@gmail.com> |
| 6 | * |
| 7 | * Redistribution and use in source and binary forms, with or without |
| 8 | * modification, are permitted provided that the following conditions |
| 9 | * are met: |
| 10 | * |
| 11 | * 1. Redistributions of source code must retain the above copyright |
| 12 | * notice, this list of conditions and the following disclaimer. |
| 13 | * 2. Redistributions in binary form must reproduce the above copyright |
| 14 | * notice, this list of conditions and the following disclaimer in |
| 15 | * the documentation and/or other materials provided with the |
| 16 | * distribution. |
| 17 | * 3. Neither the name of The DragonFly Project nor the names of its |
| 18 | * contributors may be used to endorse or promote products derived |
| 19 | * from this software without specific, prior written permission. |
| 20 | * |
| 21 | * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 22 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 23 | * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS |
| 24 | * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE |
| 25 | * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, |
| 26 | * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, |
| 27 | * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 28 | * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED |
| 29 | * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 30 | * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT |
| 31 | * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 32 | * SUCH DAMAGE. |
| 33 | */ |
| 34 | #include <sys/param.h> |
| 35 | #include <sys/systm.h> |
| 36 | #include <sys/kernel.h> |
| 37 | #include <sys/proc.h> |
| 38 | #include <sys/sysctl.h> |
| 39 | #include <sys/buf.h> |
| 40 | #include <sys/conf.h> |
| 41 | #include <sys/diskslice.h> |
| 42 | #include <sys/disk.h> |
| 43 | #include <sys/malloc.h> |
| 44 | #include <machine/md_var.h> |
| 45 | #include <sys/ctype.h> |
| 46 | #include <sys/syslog.h> |
| 47 | #include <sys/device.h> |
| 48 | #include <sys/msgport.h> |
| 49 | #include <sys/msgport2.h> |
| 50 | #include <sys/buf2.h> |
| 51 | #include <sys/dsched.h> |
| 52 | #include <sys/fcntl.h> |
| 53 | #include <machine/varargs.h> |
| 54 | |
| 55 | TAILQ_HEAD(tdio_list_head, dsched_thread_io); |
| 56 | |
| 57 | MALLOC_DEFINE(M_DSCHED, "dsched", "dsched allocs"); |
| 58 | |
| 59 | static dsched_prepare_t noop_prepare; |
| 60 | static dsched_teardown_t noop_teardown; |
| 61 | static dsched_cancel_t noop_cancel; |
| 62 | static dsched_queue_t noop_queue; |
| 63 | |
| 64 | static void dsched_sysctl_add_disk(struct dsched_disk_ctx *diskctx, char *name); |
| 65 | static void dsched_disk_ctx_destroy(struct dsched_disk_ctx *diskctx); |
| 66 | static void dsched_thread_io_destroy(struct dsched_thread_io *tdio); |
| 67 | static void dsched_thread_ctx_destroy(struct dsched_thread_ctx *tdctx); |
| 68 | |
| 69 | static int dsched_inited = 0; |
| 70 | static int default_set = 0; |
| 71 | |
| 72 | struct lock dsched_lock; |
| 73 | static int dsched_debug_enable = 0; |
| 74 | |
| 75 | struct dsched_stats dsched_stats; |
| 76 | |
| 77 | struct objcache_malloc_args dsched_disk_ctx_malloc_args = { |
| 78 | DSCHED_DISK_CTX_MAX_SZ, M_DSCHED }; |
| 79 | struct objcache_malloc_args dsched_thread_io_malloc_args = { |
| 80 | DSCHED_THREAD_IO_MAX_SZ, M_DSCHED }; |
| 81 | struct objcache_malloc_args dsched_thread_ctx_malloc_args = { |
| 82 | DSCHED_THREAD_CTX_MAX_SZ, M_DSCHED }; |
| 83 | |
| 84 | static struct objcache *dsched_diskctx_cache; |
| 85 | static struct objcache *dsched_tdctx_cache; |
| 86 | static struct objcache *dsched_tdio_cache; |
| 87 | |
| 88 | TAILQ_HEAD(, dsched_thread_ctx) dsched_tdctx_list = |
| 89 | TAILQ_HEAD_INITIALIZER(dsched_tdctx_list); |
| 90 | |
| 91 | struct lock dsched_tdctx_lock; |
| 92 | |
| 93 | static struct dsched_policy_head dsched_policy_list = |
| 94 | TAILQ_HEAD_INITIALIZER(dsched_policy_list); |
| 95 | |
| 96 | static struct dsched_policy dsched_noop_policy = { |
| 97 | .name = "noop", |
| 98 | |
| 99 | .prepare = noop_prepare, |
| 100 | .teardown = noop_teardown, |
| 101 | .cancel_all = noop_cancel, |
| 102 | .bio_queue = noop_queue |
| 103 | }; |
| 104 | |
| 105 | static struct dsched_policy *default_policy = &dsched_noop_policy; |
| 106 | |
| 107 | /* |
| 108 | * dsched_debug() is a SYSCTL and TUNABLE controlled debug output function |
| 109 | * using kvprintf |
| 110 | */ |
| 111 | int |
| 112 | dsched_debug(int level, char *fmt, ...) |
| 113 | { |
| 114 | __va_list ap; |
| 115 | |
| 116 | __va_start(ap, fmt); |
| 117 | if (level <= dsched_debug_enable) |
| 118 | kvprintf(fmt, ap); |
| 119 | __va_end(ap); |
| 120 | |
| 121 | return 0; |
| 122 | } |
| 123 | |
| 124 | /* |
| 125 | * Called on disk_create() |
| 126 | * tries to read which policy to use from loader.conf, if there's |
| 127 | * none specified, the default policy is used. |
| 128 | */ |
| 129 | void |
| 130 | dsched_disk_create_callback(struct disk *dp, const char *head_name, int unit) |
| 131 | { |
| 132 | char tunable_key[SPECNAMELEN + 48]; |
| 133 | char sched_policy[DSCHED_POLICY_NAME_LENGTH]; |
| 134 | char *ptr; |
| 135 | struct dsched_policy *policy = NULL; |
| 136 | |
| 137 | /* Also look for serno stuff? */ |
| 138 | /* kprintf("dsched_disk_create_callback() for disk %s%d\n", head_name, unit); */ |
| 139 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 140 | |
| 141 | ksnprintf(tunable_key, sizeof(tunable_key), "dsched.policy.%s%d", |
| 142 | head_name, unit); |
| 143 | if (TUNABLE_STR_FETCH(tunable_key, sched_policy, |
| 144 | sizeof(sched_policy)) != 0) { |
| 145 | policy = dsched_find_policy(sched_policy); |
| 146 | } |
| 147 | |
| 148 | ksnprintf(tunable_key, sizeof(tunable_key), "dsched.policy.%s", |
| 149 | head_name); |
| 150 | for (ptr = tunable_key; *ptr; ptr++) { |
| 151 | if (*ptr == '/') |
| 152 | *ptr = '-'; |
| 153 | } |
| 154 | if (!policy && (TUNABLE_STR_FETCH(tunable_key, sched_policy, |
| 155 | sizeof(sched_policy)) != 0)) { |
| 156 | policy = dsched_find_policy(sched_policy); |
| 157 | } |
| 158 | |
| 159 | ksnprintf(tunable_key, sizeof(tunable_key), "dsched.policy.default"); |
| 160 | if (!policy && !default_set && (TUNABLE_STR_FETCH(tunable_key, sched_policy, |
| 161 | sizeof(sched_policy)) != 0)) { |
| 162 | policy = dsched_find_policy(sched_policy); |
| 163 | } |
| 164 | |
| 165 | if (!policy) { |
| 166 | if (!default_set && bootverbose) { |
| 167 | dsched_debug(0, |
| 168 | "No policy for %s%d specified, " |
| 169 | "or policy not found\n", |
| 170 | head_name, unit); |
| 171 | } |
| 172 | dsched_set_policy(dp, default_policy); |
| 173 | } else { |
| 174 | dsched_set_policy(dp, policy); |
| 175 | } |
| 176 | |
| 177 | if (strncmp(head_name, "mapper/", strlen("mapper/")) == 0) |
| 178 | ksnprintf(tunable_key, sizeof(tunable_key), "%s", head_name); |
| 179 | else |
| 180 | ksnprintf(tunable_key, sizeof(tunable_key), "%s%d", head_name, unit); |
| 181 | for (ptr = tunable_key; *ptr; ptr++) { |
| 182 | if (*ptr == '/') |
| 183 | *ptr = '-'; |
| 184 | } |
| 185 | dsched_sysctl_add_disk( |
| 186 | (struct dsched_disk_ctx *)dsched_get_disk_priv(dp), |
| 187 | tunable_key); |
| 188 | |
| 189 | lockmgr(&dsched_lock, LK_RELEASE); |
| 190 | } |
| 191 | |
| 192 | /* |
| 193 | * Called from disk_setdiskinfo (or rather _setdiskinfo). This will check if |
| 194 | * there's any policy associated with the serial number of the device. |
| 195 | */ |
| 196 | void |
| 197 | dsched_disk_update_callback(struct disk *dp, struct disk_info *info) |
| 198 | { |
| 199 | char tunable_key[SPECNAMELEN + 48]; |
| 200 | char sched_policy[DSCHED_POLICY_NAME_LENGTH]; |
| 201 | struct dsched_policy *policy = NULL; |
| 202 | |
| 203 | if (info->d_serialno == NULL) |
| 204 | return; |
| 205 | |
| 206 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 207 | |
| 208 | ksnprintf(tunable_key, sizeof(tunable_key), "dsched.policy.%s", |
| 209 | info->d_serialno); |
| 210 | |
| 211 | if((TUNABLE_STR_FETCH(tunable_key, sched_policy, |
| 212 | sizeof(sched_policy)) != 0)) { |
| 213 | policy = dsched_find_policy(sched_policy); |
| 214 | } |
| 215 | |
| 216 | if (policy) { |
| 217 | dsched_switch(dp, policy); |
| 218 | } |
| 219 | |
| 220 | dsched_sysctl_add_disk( |
| 221 | (struct dsched_disk_ctx *)dsched_get_disk_priv(dp), |
| 222 | info->d_serialno); |
| 223 | |
| 224 | lockmgr(&dsched_lock, LK_RELEASE); |
| 225 | } |
| 226 | |
| 227 | /* |
| 228 | * Called on disk_destroy() |
| 229 | * shuts down the scheduler core and cancels all remaining bios |
| 230 | */ |
| 231 | void |
| 232 | dsched_disk_destroy_callback(struct disk *dp) |
| 233 | { |
| 234 | struct dsched_policy *old_policy; |
| 235 | struct dsched_disk_ctx *diskctx; |
| 236 | |
| 237 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 238 | |
| 239 | diskctx = dsched_get_disk_priv(dp); |
| 240 | |
| 241 | old_policy = dp->d_sched_policy; |
| 242 | dp->d_sched_policy = &dsched_noop_policy; |
| 243 | old_policy->cancel_all(dsched_get_disk_priv(dp)); |
| 244 | old_policy->teardown(dsched_get_disk_priv(dp)); |
| 245 | |
| 246 | if (diskctx->flags & DSCHED_SYSCTL_CTX_INITED) |
| 247 | sysctl_ctx_free(&diskctx->sysctl_ctx); |
| 248 | |
| 249 | policy_destroy(dp); |
| 250 | atomic_subtract_int(&old_policy->ref_count, 1); |
| 251 | KKASSERT(old_policy->ref_count >= 0); |
| 252 | |
| 253 | lockmgr(&dsched_lock, LK_RELEASE); |
| 254 | } |
| 255 | |
| 256 | |
| 257 | void |
| 258 | dsched_queue(struct disk *dp, struct bio *bio) |
| 259 | { |
| 260 | struct dsched_thread_ctx *tdctx; |
| 261 | struct dsched_thread_io *tdio; |
| 262 | struct dsched_disk_ctx *diskctx; |
| 263 | |
| 264 | int found = 0, error = 0; |
| 265 | |
| 266 | tdctx = dsched_get_buf_priv(bio->bio_buf); |
| 267 | if (tdctx == NULL) { |
| 268 | /* We don't handle this case, let dsched dispatch */ |
| 269 | atomic_add_int(&dsched_stats.no_tdctx, 1); |
| 270 | dsched_strategy_raw(dp, bio); |
| 271 | return; |
| 272 | } |
| 273 | |
| 274 | DSCHED_THREAD_CTX_LOCK(tdctx); |
| 275 | |
| 276 | KKASSERT(!TAILQ_EMPTY(&tdctx->tdio_list)); |
| 277 | /* |
| 278 | * XXX: |
| 279 | * iterate in reverse to make sure we find the most up-to-date |
| 280 | * tdio for a given disk. After a switch it may take some time |
| 281 | * for everything to clean up. |
| 282 | */ |
| 283 | TAILQ_FOREACH_REVERSE(tdio, &tdctx->tdio_list, tdio_list_head, link) { |
| 284 | if (tdio->dp == dp) { |
| 285 | dsched_thread_io_ref(tdio); |
| 286 | found = 1; |
| 287 | break; |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | DSCHED_THREAD_CTX_UNLOCK(tdctx); |
| 292 | dsched_clr_buf_priv(bio->bio_buf); |
| 293 | dsched_thread_ctx_unref(tdctx); /* acquired on new_buf */ |
| 294 | |
| 295 | KKASSERT(found == 1); |
| 296 | diskctx = dsched_get_disk_priv(dp); |
| 297 | dsched_disk_ctx_ref(diskctx); |
| 298 | |
| 299 | if (dp->d_sched_policy != &dsched_noop_policy) |
| 300 | KKASSERT(tdio->debug_policy == dp->d_sched_policy); |
| 301 | |
| 302 | KKASSERT(tdio->debug_inited == 0xF00F1234); |
| 303 | |
| 304 | error = dp->d_sched_policy->bio_queue(diskctx, tdio, bio); |
| 305 | |
| 306 | if (error) { |
| 307 | dsched_strategy_raw(dp, bio); |
| 308 | } |
| 309 | dsched_disk_ctx_unref(diskctx); |
| 310 | dsched_thread_io_unref(tdio); |
| 311 | } |
| 312 | |
| 313 | |
| 314 | /* |
| 315 | * Called from each module_init or module_attach of each policy |
| 316 | * registers the policy in the local policy list. |
| 317 | */ |
| 318 | int |
| 319 | dsched_register(struct dsched_policy *d_policy) |
| 320 | { |
| 321 | struct dsched_policy *policy; |
| 322 | int error = 0; |
| 323 | |
| 324 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 325 | |
| 326 | policy = dsched_find_policy(d_policy->name); |
| 327 | |
| 328 | if (!policy) { |
| 329 | TAILQ_INSERT_TAIL(&dsched_policy_list, d_policy, link); |
| 330 | atomic_add_int(&d_policy->ref_count, 1); |
| 331 | } else { |
| 332 | dsched_debug(LOG_ERR, "Policy with name %s already registered!\n", |
| 333 | d_policy->name); |
| 334 | error = EEXIST; |
| 335 | } |
| 336 | |
| 337 | lockmgr(&dsched_lock, LK_RELEASE); |
| 338 | return error; |
| 339 | } |
| 340 | |
| 341 | /* |
| 342 | * Called from each module_detach of each policy |
| 343 | * unregisters the policy |
| 344 | */ |
| 345 | int |
| 346 | dsched_unregister(struct dsched_policy *d_policy) |
| 347 | { |
| 348 | struct dsched_policy *policy; |
| 349 | |
| 350 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 351 | policy = dsched_find_policy(d_policy->name); |
| 352 | |
| 353 | if (policy) { |
| 354 | if (policy->ref_count > 1) { |
| 355 | lockmgr(&dsched_lock, LK_RELEASE); |
| 356 | return EBUSY; |
| 357 | } |
| 358 | TAILQ_REMOVE(&dsched_policy_list, policy, link); |
| 359 | atomic_subtract_int(&policy->ref_count, 1); |
| 360 | KKASSERT(policy->ref_count == 0); |
| 361 | } |
| 362 | lockmgr(&dsched_lock, LK_RELEASE); |
| 363 | |
| 364 | return 0; |
| 365 | } |
| 366 | |
| 367 | |
| 368 | /* |
| 369 | * switches the policy by first removing the old one and then |
| 370 | * enabling the new one. |
| 371 | */ |
| 372 | int |
| 373 | dsched_switch(struct disk *dp, struct dsched_policy *new_policy) |
| 374 | { |
| 375 | struct dsched_policy *old_policy; |
| 376 | |
| 377 | /* If we are asked to set the same policy, do nothing */ |
| 378 | if (dp->d_sched_policy == new_policy) |
| 379 | return 0; |
| 380 | |
| 381 | /* lock everything down, diskwise */ |
| 382 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 383 | old_policy = dp->d_sched_policy; |
| 384 | |
| 385 | atomic_subtract_int(&old_policy->ref_count, 1); |
| 386 | KKASSERT(old_policy->ref_count >= 0); |
| 387 | |
| 388 | dp->d_sched_policy = &dsched_noop_policy; |
| 389 | old_policy->teardown(dsched_get_disk_priv(dp)); |
| 390 | policy_destroy(dp); |
| 391 | |
| 392 | /* Bring everything back to life */ |
| 393 | dsched_set_policy(dp, new_policy); |
| 394 | lockmgr(&dsched_lock, LK_RELEASE); |
| 395 | |
| 396 | return 0; |
| 397 | } |
| 398 | |
| 399 | |
| 400 | /* |
| 401 | * Loads a given policy and attaches it to the specified disk. |
| 402 | * Also initializes the core for the policy |
| 403 | */ |
| 404 | void |
| 405 | dsched_set_policy(struct disk *dp, struct dsched_policy *new_policy) |
| 406 | { |
| 407 | int locked = 0; |
| 408 | |
| 409 | /* Check if it is locked already. if not, we acquire the devfs lock */ |
| 410 | if (!(lockstatus(&dsched_lock, curthread)) == LK_EXCLUSIVE) { |
| 411 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 412 | locked = 1; |
| 413 | } |
| 414 | |
| 415 | DSCHED_GLOBAL_THREAD_CTX_LOCK(); |
| 416 | |
| 417 | policy_new(dp, new_policy); |
| 418 | new_policy->prepare(dsched_get_disk_priv(dp)); |
| 419 | dp->d_sched_policy = new_policy; |
| 420 | |
| 421 | DSCHED_GLOBAL_THREAD_CTX_UNLOCK(); |
| 422 | |
| 423 | atomic_add_int(&new_policy->ref_count, 1); |
| 424 | kprintf("disk scheduler: set policy of %s to %s\n", dp->d_cdev->si_name, |
| 425 | new_policy->name); |
| 426 | |
| 427 | /* If we acquired the lock, we also get rid of it */ |
| 428 | if (locked) |
| 429 | lockmgr(&dsched_lock, LK_RELEASE); |
| 430 | } |
| 431 | |
| 432 | struct dsched_policy* |
| 433 | dsched_find_policy(char *search) |
| 434 | { |
| 435 | struct dsched_policy *policy; |
| 436 | struct dsched_policy *policy_found = NULL; |
| 437 | int locked = 0; |
| 438 | |
| 439 | /* Check if it is locked already. if not, we acquire the devfs lock */ |
| 440 | if (!(lockstatus(&dsched_lock, curthread)) == LK_EXCLUSIVE) { |
| 441 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 442 | locked = 1; |
| 443 | } |
| 444 | |
| 445 | TAILQ_FOREACH(policy, &dsched_policy_list, link) { |
| 446 | if (!strcmp(policy->name, search)) { |
| 447 | policy_found = policy; |
| 448 | break; |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | /* If we acquired the lock, we also get rid of it */ |
| 453 | if (locked) |
| 454 | lockmgr(&dsched_lock, LK_RELEASE); |
| 455 | |
| 456 | return policy_found; |
| 457 | } |
| 458 | |
| 459 | struct disk* |
| 460 | dsched_find_disk(char *search) |
| 461 | { |
| 462 | struct disk *dp_found = NULL; |
| 463 | struct disk *dp = NULL; |
| 464 | |
| 465 | while((dp = disk_enumerate(dp))) { |
| 466 | if (!strcmp(dp->d_cdev->si_name, search)) { |
| 467 | dp_found = dp; |
| 468 | break; |
| 469 | } |
| 470 | } |
| 471 | |
| 472 | return dp_found; |
| 473 | } |
| 474 | |
| 475 | struct disk* |
| 476 | dsched_disk_enumerate(struct disk *dp, struct dsched_policy *policy) |
| 477 | { |
| 478 | while ((dp = disk_enumerate(dp))) { |
| 479 | if (dp->d_sched_policy == policy) |
| 480 | return dp; |
| 481 | } |
| 482 | |
| 483 | return NULL; |
| 484 | } |
| 485 | |
| 486 | struct dsched_policy * |
| 487 | dsched_policy_enumerate(struct dsched_policy *pol) |
| 488 | { |
| 489 | if (!pol) |
| 490 | return (TAILQ_FIRST(&dsched_policy_list)); |
| 491 | else |
| 492 | return (TAILQ_NEXT(pol, link)); |
| 493 | } |
| 494 | |
| 495 | void |
| 496 | dsched_cancel_bio(struct bio *bp) |
| 497 | { |
| 498 | bp->bio_buf->b_error = ENXIO; |
| 499 | bp->bio_buf->b_flags |= B_ERROR; |
| 500 | bp->bio_buf->b_resid = bp->bio_buf->b_bcount; |
| 501 | |
| 502 | biodone(bp); |
| 503 | } |
| 504 | |
| 505 | void |
| 506 | dsched_strategy_raw(struct disk *dp, struct bio *bp) |
| 507 | { |
| 508 | /* |
| 509 | * Ideally, this stuff shouldn't be needed... but just in case, we leave it in |
| 510 | * to avoid panics |
| 511 | */ |
| 512 | KASSERT(dp->d_rawdev != NULL, ("dsched_strategy_raw sees NULL d_rawdev!!")); |
| 513 | if(bp->bio_track != NULL) { |
| 514 | dsched_debug(LOG_INFO, |
| 515 | "dsched_strategy_raw sees non-NULL bio_track!! " |
| 516 | "bio: %p\n", bp); |
| 517 | bp->bio_track = NULL; |
| 518 | } |
| 519 | dev_dstrategy(dp->d_rawdev, bp); |
| 520 | } |
| 521 | |
| 522 | void |
| 523 | dsched_strategy_sync(struct disk *dp, struct bio *bio) |
| 524 | { |
| 525 | struct buf *bp, *nbp; |
| 526 | struct bio *nbio; |
| 527 | |
| 528 | bp = bio->bio_buf; |
| 529 | |
| 530 | nbp = getpbuf(NULL); |
| 531 | nbio = &nbp->b_bio1; |
| 532 | |
| 533 | nbp->b_cmd = bp->b_cmd; |
| 534 | nbp->b_bufsize = bp->b_bufsize; |
| 535 | nbp->b_runningbufspace = bp->b_runningbufspace; |
| 536 | nbp->b_bcount = bp->b_bcount; |
| 537 | nbp->b_resid = bp->b_resid; |
| 538 | nbp->b_data = bp->b_data; |
| 539 | #if 0 |
| 540 | /* |
| 541 | * Buffers undergoing device I/O do not need a kvabase/size. |
| 542 | */ |
| 543 | nbp->b_kvabase = bp->b_kvabase; |
| 544 | nbp->b_kvasize = bp->b_kvasize; |
| 545 | #endif |
| 546 | nbp->b_dirtyend = bp->b_dirtyend; |
| 547 | |
| 548 | nbio->bio_done = biodone_sync; |
| 549 | nbio->bio_flags |= BIO_SYNC; |
| 550 | nbio->bio_track = NULL; |
| 551 | |
| 552 | nbio->bio_caller_info1.ptr = dp; |
| 553 | nbio->bio_offset = bio->bio_offset; |
| 554 | |
| 555 | dev_dstrategy(dp->d_rawdev, nbio); |
| 556 | biowait(nbio, "dschedsync"); |
| 557 | bp->b_resid = nbp->b_resid; |
| 558 | bp->b_error = nbp->b_error; |
| 559 | biodone(bio); |
| 560 | #if 0 |
| 561 | nbp->b_kvabase = NULL; |
| 562 | nbp->b_kvasize = 0; |
| 563 | #endif |
| 564 | relpbuf(nbp, NULL); |
| 565 | } |
| 566 | |
| 567 | void |
| 568 | dsched_strategy_async(struct disk *dp, struct bio *bio, biodone_t *done, void *priv) |
| 569 | { |
| 570 | struct bio *nbio; |
| 571 | |
| 572 | nbio = push_bio(bio); |
| 573 | nbio->bio_done = done; |
| 574 | nbio->bio_offset = bio->bio_offset; |
| 575 | |
| 576 | dsched_set_bio_dp(nbio, dp); |
| 577 | dsched_set_bio_priv(nbio, priv); |
| 578 | |
| 579 | getmicrotime(&nbio->bio_caller_info3.tv); |
| 580 | dev_dstrategy(dp->d_rawdev, nbio); |
| 581 | } |
| 582 | |
| 583 | /* |
| 584 | * A special bio done call back function |
| 585 | * used by policy having request polling implemented. |
| 586 | */ |
| 587 | static void |
| 588 | request_polling_biodone(struct bio *bp) |
| 589 | { |
| 590 | struct dsched_disk_ctx *diskctx = NULL; |
| 591 | struct disk *dp = NULL; |
| 592 | struct bio *obio; |
| 593 | struct dsched_policy *policy; |
| 594 | |
| 595 | dp = dsched_get_bio_dp(bp); |
| 596 | policy = dp->d_sched_policy; |
| 597 | diskctx = dsched_get_disk_priv(dp); |
| 598 | KKASSERT(diskctx && policy); |
| 599 | dsched_disk_ctx_ref(diskctx); |
| 600 | |
| 601 | /* |
| 602 | * XXX: |
| 603 | * the bio_done function should not be blocked ! |
| 604 | */ |
| 605 | if (diskctx->dp->d_sched_policy->bio_done) |
| 606 | diskctx->dp->d_sched_policy->bio_done(bp); |
| 607 | |
| 608 | obio = pop_bio(bp); |
| 609 | biodone(obio); |
| 610 | |
| 611 | atomic_subtract_int(&diskctx->current_tag_queue_depth, 1); |
| 612 | |
| 613 | /* call the polling function, |
| 614 | * XXX: |
| 615 | * the polling function should not be blocked! |
| 616 | */ |
| 617 | if (policy->polling_func) |
| 618 | policy->polling_func(diskctx); |
| 619 | else |
| 620 | dsched_debug(0, "dsched: the policy uses request polling without a polling function!\n"); |
| 621 | dsched_disk_ctx_unref(diskctx); |
| 622 | } |
| 623 | |
| 624 | /* |
| 625 | * A special dsched strategy used by policy having request polling |
| 626 | * (polling function) implemented. |
| 627 | * |
| 628 | * The strategy is the just like dsched_strategy_async(), but |
| 629 | * the biodone call back is set to a preset one. |
| 630 | * |
| 631 | * If the policy needs its own biodone callback, it should |
| 632 | * register it in the policy structure. (bio_done field) |
| 633 | * |
| 634 | * The current_tag_queue_depth is maintained by this function |
| 635 | * and the request_polling_biodone() function |
| 636 | */ |
| 637 | |
| 638 | void |
| 639 | dsched_strategy_request_polling(struct disk *dp, struct bio *bio, struct dsched_disk_ctx *diskctx) |
| 640 | { |
| 641 | atomic_add_int(&diskctx->current_tag_queue_depth, 1); |
| 642 | dsched_strategy_async(dp, bio, request_polling_biodone, dsched_get_bio_priv(bio)); |
| 643 | } |
| 644 | |
| 645 | /* |
| 646 | * Ref and deref various structures. The 1->0 transition of the reference |
| 647 | * count actually transitions 1->0x80000000 and causes the object to be |
| 648 | * destroyed. It is possible for transitory references to occur on the |
| 649 | * object while it is being destroyed. We use bit 31 to indicate that |
| 650 | * destruction is in progress and to prevent nested destructions. |
| 651 | */ |
| 652 | void |
| 653 | dsched_disk_ctx_ref(struct dsched_disk_ctx *diskctx) |
| 654 | { |
| 655 | int refcount; |
| 656 | |
| 657 | refcount = atomic_fetchadd_int(&diskctx->refcount, 1); |
| 658 | } |
| 659 | |
| 660 | void |
| 661 | dsched_thread_io_ref(struct dsched_thread_io *tdio) |
| 662 | { |
| 663 | int refcount; |
| 664 | |
| 665 | refcount = atomic_fetchadd_int(&tdio->refcount, 1); |
| 666 | } |
| 667 | |
| 668 | void |
| 669 | dsched_thread_ctx_ref(struct dsched_thread_ctx *tdctx) |
| 670 | { |
| 671 | int refcount; |
| 672 | |
| 673 | refcount = atomic_fetchadd_int(&tdctx->refcount, 1); |
| 674 | } |
| 675 | |
| 676 | void |
| 677 | dsched_disk_ctx_unref(struct dsched_disk_ctx *diskctx) |
| 678 | { |
| 679 | int refs; |
| 680 | int nrefs; |
| 681 | |
| 682 | /* |
| 683 | * Handle 1->0 transitions for diskctx and nested destruction |
| 684 | * recursions. If the refs are already in destruction mode (bit 31 |
| 685 | * set) on the 1->0 transition we don't try to destruct it again. |
| 686 | * |
| 687 | * 0x80000001->0x80000000 transitions are handled normally and |
| 688 | * thus avoid nested dstruction. |
| 689 | */ |
| 690 | for (;;) { |
| 691 | refs = diskctx->refcount; |
| 692 | cpu_ccfence(); |
| 693 | nrefs = refs - 1; |
| 694 | |
| 695 | KKASSERT(((refs ^ nrefs) & 0x80000000) == 0); |
| 696 | if (nrefs) { |
| 697 | if (atomic_cmpset_int(&diskctx->refcount, refs, nrefs)) |
| 698 | break; |
| 699 | continue; |
| 700 | } |
| 701 | nrefs = 0x80000000; |
| 702 | if (atomic_cmpset_int(&diskctx->refcount, refs, nrefs)) { |
| 703 | dsched_disk_ctx_destroy(diskctx); |
| 704 | break; |
| 705 | } |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | static |
| 710 | void |
| 711 | dsched_disk_ctx_destroy(struct dsched_disk_ctx *diskctx) |
| 712 | { |
| 713 | struct dsched_thread_io *tdio; |
| 714 | |
| 715 | #if 0 |
| 716 | kprintf("diskctx (%p) destruction started, trace:\n", diskctx); |
| 717 | print_backtrace(4); |
| 718 | #endif |
| 719 | lockmgr(&diskctx->lock, LK_EXCLUSIVE); |
| 720 | while ((tdio = TAILQ_FIRST(&diskctx->tdio_list)) != NULL) { |
| 721 | KKASSERT(tdio->flags & DSCHED_LINKED_DISK_CTX); |
| 722 | TAILQ_REMOVE(&diskctx->tdio_list, tdio, dlink); |
| 723 | atomic_clear_int(&tdio->flags, DSCHED_LINKED_DISK_CTX); |
| 724 | tdio->diskctx = NULL; |
| 725 | /* XXX tdio->diskctx->dp->d_sched_policy->destroy_tdio(tdio);*/ |
| 726 | dsched_thread_io_unref(tdio); |
| 727 | } |
| 728 | lockmgr(&diskctx->lock, LK_RELEASE); |
| 729 | if (diskctx->dp->d_sched_policy->destroy_diskctx) |
| 730 | diskctx->dp->d_sched_policy->destroy_diskctx(diskctx); |
| 731 | KKASSERT(diskctx->refcount == 0x80000000); |
| 732 | objcache_put(dsched_diskctx_cache, diskctx); |
| 733 | atomic_subtract_int(&dsched_stats.diskctx_allocations, 1); |
| 734 | } |
| 735 | |
| 736 | void |
| 737 | dsched_thread_io_unref(struct dsched_thread_io *tdio) |
| 738 | { |
| 739 | int refs; |
| 740 | int nrefs; |
| 741 | |
| 742 | /* |
| 743 | * Handle 1->0 transitions for tdio and nested destruction |
| 744 | * recursions. If the refs are already in destruction mode (bit 31 |
| 745 | * set) on the 1->0 transition we don't try to destruct it again. |
| 746 | * |
| 747 | * 0x80000001->0x80000000 transitions are handled normally and |
| 748 | * thus avoid nested dstruction. |
| 749 | */ |
| 750 | for (;;) { |
| 751 | refs = tdio->refcount; |
| 752 | cpu_ccfence(); |
| 753 | nrefs = refs - 1; |
| 754 | |
| 755 | KKASSERT(((refs ^ nrefs) & 0x80000000) == 0); |
| 756 | if (nrefs) { |
| 757 | if (atomic_cmpset_int(&tdio->refcount, refs, nrefs)) |
| 758 | break; |
| 759 | continue; |
| 760 | } |
| 761 | nrefs = 0x80000000; |
| 762 | if (atomic_cmpset_int(&tdio->refcount, refs, nrefs)) { |
| 763 | dsched_thread_io_destroy(tdio); |
| 764 | break; |
| 765 | } |
| 766 | } |
| 767 | } |
| 768 | |
| 769 | static void |
| 770 | dsched_thread_io_destroy(struct dsched_thread_io *tdio) |
| 771 | { |
| 772 | struct dsched_thread_ctx *tdctx; |
| 773 | struct dsched_disk_ctx *diskctx; |
| 774 | |
| 775 | #if 0 |
| 776 | kprintf("tdio (%p) destruction started, trace:\n", tdio); |
| 777 | print_backtrace(8); |
| 778 | #endif |
| 779 | KKASSERT(tdio->qlength == 0); |
| 780 | |
| 781 | while ((diskctx = tdio->diskctx) != NULL) { |
| 782 | dsched_disk_ctx_ref(diskctx); |
| 783 | lockmgr(&diskctx->lock, LK_EXCLUSIVE); |
| 784 | if (diskctx != tdio->diskctx) { |
| 785 | lockmgr(&diskctx->lock, LK_RELEASE); |
| 786 | dsched_disk_ctx_unref(diskctx); |
| 787 | continue; |
| 788 | } |
| 789 | KKASSERT(tdio->flags & DSCHED_LINKED_DISK_CTX); |
| 790 | if (diskctx->dp->d_sched_policy->destroy_tdio) |
| 791 | diskctx->dp->d_sched_policy->destroy_tdio(tdio); |
| 792 | TAILQ_REMOVE(&diskctx->tdio_list, tdio, dlink); |
| 793 | atomic_clear_int(&tdio->flags, DSCHED_LINKED_DISK_CTX); |
| 794 | tdio->diskctx = NULL; |
| 795 | lockmgr(&diskctx->lock, LK_RELEASE); |
| 796 | dsched_disk_ctx_unref(diskctx); |
| 797 | } |
| 798 | while ((tdctx = tdio->tdctx) != NULL) { |
| 799 | dsched_thread_ctx_ref(tdctx); |
| 800 | lockmgr(&tdctx->lock, LK_EXCLUSIVE); |
| 801 | if (tdctx != tdio->tdctx) { |
| 802 | lockmgr(&tdctx->lock, LK_RELEASE); |
| 803 | dsched_thread_ctx_unref(tdctx); |
| 804 | continue; |
| 805 | } |
| 806 | KKASSERT(tdio->flags & DSCHED_LINKED_THREAD_CTX); |
| 807 | TAILQ_REMOVE(&tdctx->tdio_list, tdio, link); |
| 808 | atomic_clear_int(&tdio->flags, DSCHED_LINKED_THREAD_CTX); |
| 809 | tdio->tdctx = NULL; |
| 810 | lockmgr(&tdctx->lock, LK_RELEASE); |
| 811 | dsched_thread_ctx_unref(tdctx); |
| 812 | } |
| 813 | KKASSERT(tdio->refcount == 0x80000000); |
| 814 | objcache_put(dsched_tdio_cache, tdio); |
| 815 | atomic_subtract_int(&dsched_stats.tdio_allocations, 1); |
| 816 | #if 0 |
| 817 | dsched_disk_ctx_unref(diskctx); |
| 818 | #endif |
| 819 | } |
| 820 | |
| 821 | void |
| 822 | dsched_thread_ctx_unref(struct dsched_thread_ctx *tdctx) |
| 823 | { |
| 824 | int refs; |
| 825 | int nrefs; |
| 826 | |
| 827 | /* |
| 828 | * Handle 1->0 transitions for tdctx and nested destruction |
| 829 | * recursions. If the refs are already in destruction mode (bit 31 |
| 830 | * set) on the 1->0 transition we don't try to destruct it again. |
| 831 | * |
| 832 | * 0x80000001->0x80000000 transitions are handled normally and |
| 833 | * thus avoid nested dstruction. |
| 834 | */ |
| 835 | for (;;) { |
| 836 | refs = tdctx->refcount; |
| 837 | cpu_ccfence(); |
| 838 | nrefs = refs - 1; |
| 839 | |
| 840 | KKASSERT(((refs ^ nrefs) & 0x80000000) == 0); |
| 841 | if (nrefs) { |
| 842 | if (atomic_cmpset_int(&tdctx->refcount, refs, nrefs)) |
| 843 | break; |
| 844 | continue; |
| 845 | } |
| 846 | nrefs = 0x80000000; |
| 847 | if (atomic_cmpset_int(&tdctx->refcount, refs, nrefs)) { |
| 848 | dsched_thread_ctx_destroy(tdctx); |
| 849 | break; |
| 850 | } |
| 851 | } |
| 852 | } |
| 853 | |
| 854 | static void |
| 855 | dsched_thread_ctx_destroy(struct dsched_thread_ctx *tdctx) |
| 856 | { |
| 857 | struct dsched_thread_io *tdio; |
| 858 | |
| 859 | #if 0 |
| 860 | kprintf("tdctx (%p) destruction started, trace:\n", tdctx); |
| 861 | print_backtrace(8); |
| 862 | #endif |
| 863 | DSCHED_GLOBAL_THREAD_CTX_LOCK(); |
| 864 | |
| 865 | lockmgr(&tdctx->lock, LK_EXCLUSIVE); |
| 866 | |
| 867 | while ((tdio = TAILQ_FIRST(&tdctx->tdio_list)) != NULL) { |
| 868 | KKASSERT(tdio->flags & DSCHED_LINKED_THREAD_CTX); |
| 869 | TAILQ_REMOVE(&tdctx->tdio_list, tdio, link); |
| 870 | atomic_clear_int(&tdio->flags, DSCHED_LINKED_THREAD_CTX); |
| 871 | tdio->tdctx = NULL; |
| 872 | dsched_thread_io_unref(tdio); |
| 873 | } |
| 874 | KKASSERT(tdctx->refcount == 0x80000000); |
| 875 | TAILQ_REMOVE(&dsched_tdctx_list, tdctx, link); |
| 876 | |
| 877 | lockmgr(&tdctx->lock, LK_RELEASE); |
| 878 | |
| 879 | DSCHED_GLOBAL_THREAD_CTX_UNLOCK(); |
| 880 | |
| 881 | objcache_put(dsched_tdctx_cache, tdctx); |
| 882 | atomic_subtract_int(&dsched_stats.tdctx_allocations, 1); |
| 883 | } |
| 884 | |
| 885 | struct dsched_thread_io * |
| 886 | dsched_thread_io_alloc(struct disk *dp, struct dsched_thread_ctx *tdctx, |
| 887 | struct dsched_policy *pol) |
| 888 | { |
| 889 | struct dsched_thread_io *tdio; |
| 890 | #if 0 |
| 891 | dsched_disk_ctx_ref(dsched_get_disk_priv(dp)); |
| 892 | #endif |
| 893 | tdio = objcache_get(dsched_tdio_cache, M_WAITOK); |
| 894 | bzero(tdio, DSCHED_THREAD_IO_MAX_SZ); |
| 895 | |
| 896 | /* XXX: maybe we do need another ref for the disk list for tdio */ |
| 897 | dsched_thread_io_ref(tdio); |
| 898 | |
| 899 | DSCHED_THREAD_IO_LOCKINIT(tdio); |
| 900 | tdio->dp = dp; |
| 901 | |
| 902 | tdio->diskctx = dsched_get_disk_priv(dp); |
| 903 | TAILQ_INIT(&tdio->queue); |
| 904 | |
| 905 | if (pol->new_tdio) |
| 906 | pol->new_tdio(tdio); |
| 907 | |
| 908 | lockmgr(&tdio->diskctx->lock, LK_EXCLUSIVE); |
| 909 | TAILQ_INSERT_TAIL(&tdio->diskctx->tdio_list, tdio, dlink); |
| 910 | atomic_set_int(&tdio->flags, DSCHED_LINKED_DISK_CTX); |
| 911 | lockmgr(&tdio->diskctx->lock, LK_RELEASE); |
| 912 | |
| 913 | if (tdctx) { |
| 914 | tdio->tdctx = tdctx; |
| 915 | tdio->p = tdctx->p; |
| 916 | |
| 917 | /* Put the tdio in the tdctx list */ |
| 918 | DSCHED_THREAD_CTX_LOCK(tdctx); |
| 919 | TAILQ_INSERT_TAIL(&tdctx->tdio_list, tdio, link); |
| 920 | DSCHED_THREAD_CTX_UNLOCK(tdctx); |
| 921 | atomic_set_int(&tdio->flags, DSCHED_LINKED_THREAD_CTX); |
| 922 | } |
| 923 | |
| 924 | tdio->debug_policy = pol; |
| 925 | tdio->debug_inited = 0xF00F1234; |
| 926 | |
| 927 | atomic_add_int(&dsched_stats.tdio_allocations, 1); |
| 928 | return tdio; |
| 929 | } |
| 930 | |
| 931 | |
| 932 | struct dsched_disk_ctx * |
| 933 | dsched_disk_ctx_alloc(struct disk *dp, struct dsched_policy *pol) |
| 934 | { |
| 935 | struct dsched_disk_ctx *diskctx; |
| 936 | |
| 937 | diskctx = objcache_get(dsched_diskctx_cache, M_WAITOK); |
| 938 | bzero(diskctx, DSCHED_DISK_CTX_MAX_SZ); |
| 939 | dsched_disk_ctx_ref(diskctx); |
| 940 | diskctx->dp = dp; |
| 941 | DSCHED_DISK_CTX_LOCKINIT(diskctx); |
| 942 | TAILQ_INIT(&diskctx->tdio_list); |
| 943 | /* |
| 944 | * XXX: magic number 32: most device has a tag queue |
| 945 | * of depth 32. |
| 946 | * Better to retrive more precise value from the driver |
| 947 | */ |
| 948 | diskctx->max_tag_queue_depth = 32; |
| 949 | diskctx->current_tag_queue_depth = 0; |
| 950 | |
| 951 | atomic_add_int(&dsched_stats.diskctx_allocations, 1); |
| 952 | if (pol->new_diskctx) |
| 953 | pol->new_diskctx(diskctx); |
| 954 | return diskctx; |
| 955 | } |
| 956 | |
| 957 | |
| 958 | struct dsched_thread_ctx * |
| 959 | dsched_thread_ctx_alloc(struct proc *p) |
| 960 | { |
| 961 | struct dsched_thread_ctx *tdctx; |
| 962 | struct dsched_thread_io *tdio; |
| 963 | struct disk *dp = NULL; |
| 964 | |
| 965 | tdctx = objcache_get(dsched_tdctx_cache, M_WAITOK); |
| 966 | bzero(tdctx, DSCHED_THREAD_CTX_MAX_SZ); |
| 967 | dsched_thread_ctx_ref(tdctx); |
| 968 | #if 0 |
| 969 | kprintf("dsched_thread_ctx_alloc, new tdctx = %p\n", tdctx); |
| 970 | #endif |
| 971 | DSCHED_THREAD_CTX_LOCKINIT(tdctx); |
| 972 | TAILQ_INIT(&tdctx->tdio_list); |
| 973 | tdctx->p = p; |
| 974 | |
| 975 | DSCHED_GLOBAL_THREAD_CTX_LOCK(); |
| 976 | while ((dp = disk_enumerate(dp))) { |
| 977 | tdio = dsched_thread_io_alloc(dp, tdctx, dp->d_sched_policy); |
| 978 | } |
| 979 | |
| 980 | TAILQ_INSERT_TAIL(&dsched_tdctx_list, tdctx, link); |
| 981 | DSCHED_GLOBAL_THREAD_CTX_UNLOCK(); |
| 982 | |
| 983 | atomic_add_int(&dsched_stats.tdctx_allocations, 1); |
| 984 | /* XXX: no callback here */ |
| 985 | return tdctx; |
| 986 | } |
| 987 | |
| 988 | void |
| 989 | policy_new(struct disk *dp, struct dsched_policy *pol) { |
| 990 | struct dsched_thread_ctx *tdctx; |
| 991 | struct dsched_disk_ctx *diskctx; |
| 992 | struct dsched_thread_io *tdio; |
| 993 | |
| 994 | diskctx = dsched_disk_ctx_alloc(dp, pol); |
| 995 | dsched_disk_ctx_ref(diskctx); |
| 996 | dsched_set_disk_priv(dp, diskctx); |
| 997 | |
| 998 | TAILQ_FOREACH(tdctx, &dsched_tdctx_list, link) { |
| 999 | tdio = dsched_thread_io_alloc(dp, tdctx, pol); |
| 1000 | } |
| 1001 | } |
| 1002 | |
| 1003 | void |
| 1004 | policy_destroy(struct disk *dp) { |
| 1005 | struct dsched_disk_ctx *diskctx; |
| 1006 | |
| 1007 | diskctx = dsched_get_disk_priv(dp); |
| 1008 | KKASSERT(diskctx != NULL); |
| 1009 | |
| 1010 | dsched_disk_ctx_unref(diskctx); /* from prepare */ |
| 1011 | dsched_disk_ctx_unref(diskctx); /* from alloc */ |
| 1012 | |
| 1013 | dsched_set_disk_priv(dp, NULL); |
| 1014 | } |
| 1015 | |
| 1016 | void |
| 1017 | dsched_new_buf(struct buf *bp) |
| 1018 | { |
| 1019 | struct dsched_thread_ctx *tdctx = NULL; |
| 1020 | |
| 1021 | if (dsched_inited == 0) |
| 1022 | return; |
| 1023 | |
| 1024 | if (curproc != NULL) { |
| 1025 | tdctx = dsched_get_proc_priv(curproc); |
| 1026 | } else { |
| 1027 | /* This is a kernel thread, so no proc info is available */ |
| 1028 | tdctx = dsched_get_thread_priv(curthread); |
| 1029 | } |
| 1030 | |
| 1031 | #if 0 |
| 1032 | /* |
| 1033 | * XXX: hack. we don't want this assert because we aren't catching all |
| 1034 | * threads. mi_startup() is still getting away without an tdctx. |
| 1035 | */ |
| 1036 | |
| 1037 | /* by now we should have an tdctx. if not, something bad is going on */ |
| 1038 | KKASSERT(tdctx != NULL); |
| 1039 | #endif |
| 1040 | |
| 1041 | if (tdctx) { |
| 1042 | dsched_thread_ctx_ref(tdctx); |
| 1043 | } |
| 1044 | dsched_set_buf_priv(bp, tdctx); |
| 1045 | } |
| 1046 | |
| 1047 | void |
| 1048 | dsched_exit_buf(struct buf *bp) |
| 1049 | { |
| 1050 | struct dsched_thread_ctx *tdctx; |
| 1051 | |
| 1052 | tdctx = dsched_get_buf_priv(bp); |
| 1053 | if (tdctx != NULL) { |
| 1054 | dsched_clr_buf_priv(bp); |
| 1055 | dsched_thread_ctx_unref(tdctx); |
| 1056 | } |
| 1057 | } |
| 1058 | |
| 1059 | void |
| 1060 | dsched_new_proc(struct proc *p) |
| 1061 | { |
| 1062 | struct dsched_thread_ctx *tdctx; |
| 1063 | |
| 1064 | if (dsched_inited == 0) |
| 1065 | return; |
| 1066 | |
| 1067 | KKASSERT(p != NULL); |
| 1068 | |
| 1069 | tdctx = dsched_thread_ctx_alloc(p); |
| 1070 | tdctx->p = p; |
| 1071 | dsched_thread_ctx_ref(tdctx); |
| 1072 | |
| 1073 | dsched_set_proc_priv(p, tdctx); |
| 1074 | atomic_add_int(&dsched_stats.nprocs, 1); |
| 1075 | } |
| 1076 | |
| 1077 | |
| 1078 | void |
| 1079 | dsched_new_thread(struct thread *td) |
| 1080 | { |
| 1081 | struct dsched_thread_ctx *tdctx; |
| 1082 | |
| 1083 | if (dsched_inited == 0) |
| 1084 | return; |
| 1085 | |
| 1086 | KKASSERT(td != NULL); |
| 1087 | |
| 1088 | tdctx = dsched_thread_ctx_alloc(NULL); |
| 1089 | tdctx->td = td; |
| 1090 | dsched_thread_ctx_ref(tdctx); |
| 1091 | |
| 1092 | dsched_set_thread_priv(td, tdctx); |
| 1093 | atomic_add_int(&dsched_stats.nthreads, 1); |
| 1094 | } |
| 1095 | |
| 1096 | void |
| 1097 | dsched_exit_proc(struct proc *p) |
| 1098 | { |
| 1099 | struct dsched_thread_ctx *tdctx; |
| 1100 | |
| 1101 | if (dsched_inited == 0) |
| 1102 | return; |
| 1103 | |
| 1104 | KKASSERT(p != NULL); |
| 1105 | |
| 1106 | tdctx = dsched_get_proc_priv(p); |
| 1107 | KKASSERT(tdctx != NULL); |
| 1108 | |
| 1109 | tdctx->dead = 0xDEAD; |
| 1110 | dsched_set_proc_priv(p, NULL); |
| 1111 | |
| 1112 | dsched_thread_ctx_unref(tdctx); /* one for alloc, */ |
| 1113 | dsched_thread_ctx_unref(tdctx); /* one for ref */ |
| 1114 | atomic_subtract_int(&dsched_stats.nprocs, 1); |
| 1115 | } |
| 1116 | |
| 1117 | |
| 1118 | void |
| 1119 | dsched_exit_thread(struct thread *td) |
| 1120 | { |
| 1121 | struct dsched_thread_ctx *tdctx; |
| 1122 | |
| 1123 | if (dsched_inited == 0) |
| 1124 | return; |
| 1125 | |
| 1126 | KKASSERT(td != NULL); |
| 1127 | |
| 1128 | tdctx = dsched_get_thread_priv(td); |
| 1129 | KKASSERT(tdctx != NULL); |
| 1130 | |
| 1131 | tdctx->dead = 0xDEAD; |
| 1132 | dsched_set_thread_priv(td, 0); |
| 1133 | |
| 1134 | dsched_thread_ctx_unref(tdctx); /* one for alloc, */ |
| 1135 | dsched_thread_ctx_unref(tdctx); /* one for ref */ |
| 1136 | atomic_subtract_int(&dsched_stats.nthreads, 1); |
| 1137 | } |
| 1138 | |
| 1139 | struct dsched_thread_io * |
| 1140 | dsched_new_policy_thread_tdio(struct dsched_disk_ctx *diskctx, |
| 1141 | struct dsched_policy *pol) { |
| 1142 | struct dsched_thread_ctx *tdctx; |
| 1143 | struct dsched_thread_io *tdio; |
| 1144 | |
| 1145 | DSCHED_GLOBAL_THREAD_CTX_LOCK(); |
| 1146 | |
| 1147 | tdctx = dsched_get_thread_priv(curthread); |
| 1148 | KKASSERT(tdctx != NULL); |
| 1149 | tdio = dsched_thread_io_alloc(diskctx->dp, tdctx, pol); |
| 1150 | |
| 1151 | DSCHED_GLOBAL_THREAD_CTX_UNLOCK(); |
| 1152 | |
| 1153 | return tdio; |
| 1154 | } |
| 1155 | |
| 1156 | /* DEFAULT NOOP POLICY */ |
| 1157 | |
| 1158 | static int |
| 1159 | noop_prepare(struct dsched_disk_ctx *diskctx) |
| 1160 | { |
| 1161 | return 0; |
| 1162 | } |
| 1163 | |
| 1164 | static void |
| 1165 | noop_teardown(struct dsched_disk_ctx *diskctx) |
| 1166 | { |
| 1167 | |
| 1168 | } |
| 1169 | |
| 1170 | static void |
| 1171 | noop_cancel(struct dsched_disk_ctx *diskctx) |
| 1172 | { |
| 1173 | |
| 1174 | } |
| 1175 | |
| 1176 | static int |
| 1177 | noop_queue(struct dsched_disk_ctx *diskctx, struct dsched_thread_io *tdio, |
| 1178 | struct bio *bio) |
| 1179 | { |
| 1180 | dsched_strategy_raw(diskctx->dp, bio); |
| 1181 | #if 0 |
| 1182 | dsched_strategy_async(diskctx->dp, bio, noop_completed, NULL); |
| 1183 | #endif |
| 1184 | return 0; |
| 1185 | } |
| 1186 | |
| 1187 | /* |
| 1188 | * SYSINIT stuff |
| 1189 | */ |
| 1190 | static void |
| 1191 | dsched_init(void) |
| 1192 | { |
| 1193 | dsched_tdio_cache = objcache_create("dsched-tdio-cache", 0, 0, |
| 1194 | NULL, NULL, NULL, |
| 1195 | objcache_malloc_alloc, |
| 1196 | objcache_malloc_free, |
| 1197 | &dsched_thread_io_malloc_args ); |
| 1198 | |
| 1199 | dsched_tdctx_cache = objcache_create("dsched-tdctx-cache", 0, 0, |
| 1200 | NULL, NULL, NULL, |
| 1201 | objcache_malloc_alloc, |
| 1202 | objcache_malloc_free, |
| 1203 | &dsched_thread_ctx_malloc_args ); |
| 1204 | |
| 1205 | dsched_diskctx_cache = objcache_create("dsched-diskctx-cache", 0, 0, |
| 1206 | NULL, NULL, NULL, |
| 1207 | objcache_malloc_alloc, |
| 1208 | objcache_malloc_free, |
| 1209 | &dsched_disk_ctx_malloc_args ); |
| 1210 | |
| 1211 | bzero(&dsched_stats, sizeof(struct dsched_stats)); |
| 1212 | |
| 1213 | lockinit(&dsched_lock, "dsched lock", 0, LK_CANRECURSE); |
| 1214 | DSCHED_GLOBAL_THREAD_CTX_LOCKINIT(); |
| 1215 | |
| 1216 | dsched_register(&dsched_noop_policy); |
| 1217 | |
| 1218 | dsched_inited = 1; |
| 1219 | } |
| 1220 | |
| 1221 | static void |
| 1222 | dsched_uninit(void) |
| 1223 | { |
| 1224 | } |
| 1225 | |
| 1226 | SYSINIT(subr_dsched_register, SI_SUB_CREATE_INIT-1, SI_ORDER_FIRST, dsched_init, NULL); |
| 1227 | SYSUNINIT(subr_dsched_register, SI_SUB_CREATE_INIT-1, SI_ORDER_ANY, dsched_uninit, NULL); |
| 1228 | |
| 1229 | /* |
| 1230 | * SYSCTL stuff |
| 1231 | */ |
| 1232 | static int |
| 1233 | sysctl_dsched_stats(SYSCTL_HANDLER_ARGS) |
| 1234 | { |
| 1235 | return (sysctl_handle_opaque(oidp, &dsched_stats, sizeof(struct dsched_stats), req)); |
| 1236 | } |
| 1237 | |
| 1238 | static int |
| 1239 | sysctl_dsched_list_policies(SYSCTL_HANDLER_ARGS) |
| 1240 | { |
| 1241 | struct dsched_policy *pol = NULL; |
| 1242 | int error, first = 1; |
| 1243 | |
| 1244 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 1245 | |
| 1246 | while ((pol = dsched_policy_enumerate(pol))) { |
| 1247 | if (!first) { |
| 1248 | error = SYSCTL_OUT(req, " ", 1); |
| 1249 | if (error) |
| 1250 | break; |
| 1251 | } else { |
| 1252 | first = 0; |
| 1253 | } |
| 1254 | error = SYSCTL_OUT(req, pol->name, strlen(pol->name)); |
| 1255 | if (error) |
| 1256 | break; |
| 1257 | |
| 1258 | } |
| 1259 | |
| 1260 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1261 | |
| 1262 | error = SYSCTL_OUT(req, "", 1); |
| 1263 | |
| 1264 | return error; |
| 1265 | } |
| 1266 | |
| 1267 | static int |
| 1268 | sysctl_dsched_policy(SYSCTL_HANDLER_ARGS) |
| 1269 | { |
| 1270 | char buf[DSCHED_POLICY_NAME_LENGTH]; |
| 1271 | struct dsched_disk_ctx *diskctx = arg1; |
| 1272 | struct dsched_policy *pol = NULL; |
| 1273 | int error; |
| 1274 | |
| 1275 | if (diskctx == NULL) { |
| 1276 | return 0; |
| 1277 | } |
| 1278 | |
| 1279 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 1280 | |
| 1281 | pol = diskctx->dp->d_sched_policy; |
| 1282 | memcpy(buf, pol->name, DSCHED_POLICY_NAME_LENGTH); |
| 1283 | |
| 1284 | error = sysctl_handle_string(oidp, buf, DSCHED_POLICY_NAME_LENGTH, req); |
| 1285 | if (error || req->newptr == NULL) { |
| 1286 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1287 | return (error); |
| 1288 | } |
| 1289 | |
| 1290 | pol = dsched_find_policy(buf); |
| 1291 | if (pol == NULL) { |
| 1292 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1293 | return 0; |
| 1294 | } |
| 1295 | |
| 1296 | dsched_switch(diskctx->dp, pol); |
| 1297 | |
| 1298 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1299 | |
| 1300 | return error; |
| 1301 | } |
| 1302 | |
| 1303 | static int |
| 1304 | sysctl_dsched_default_policy(SYSCTL_HANDLER_ARGS) |
| 1305 | { |
| 1306 | char buf[DSCHED_POLICY_NAME_LENGTH]; |
| 1307 | struct dsched_policy *pol = NULL; |
| 1308 | int error; |
| 1309 | |
| 1310 | lockmgr(&dsched_lock, LK_EXCLUSIVE); |
| 1311 | |
| 1312 | pol = default_policy; |
| 1313 | memcpy(buf, pol->name, DSCHED_POLICY_NAME_LENGTH); |
| 1314 | |
| 1315 | error = sysctl_handle_string(oidp, buf, DSCHED_POLICY_NAME_LENGTH, req); |
| 1316 | if (error || req->newptr == NULL) { |
| 1317 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1318 | return (error); |
| 1319 | } |
| 1320 | |
| 1321 | pol = dsched_find_policy(buf); |
| 1322 | if (pol == NULL) { |
| 1323 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1324 | return 0; |
| 1325 | } |
| 1326 | |
| 1327 | default_set = 1; |
| 1328 | default_policy = pol; |
| 1329 | |
| 1330 | lockmgr(&dsched_lock, LK_RELEASE); |
| 1331 | |
| 1332 | return error; |
| 1333 | } |
| 1334 | |
| 1335 | SYSCTL_NODE(, OID_AUTO, dsched, CTLFLAG_RD, NULL, |
| 1336 | "Disk Scheduler Framework (dsched) magic"); |
| 1337 | SYSCTL_NODE(_dsched, OID_AUTO, policy, CTLFLAG_RW, NULL, |
| 1338 | "List of disks and their policies"); |
| 1339 | SYSCTL_INT(_dsched, OID_AUTO, debug, CTLFLAG_RW, &dsched_debug_enable, |
| 1340 | 0, "Enable dsched debugging"); |
| 1341 | SYSCTL_PROC(_dsched, OID_AUTO, stats, CTLTYPE_OPAQUE|CTLFLAG_RD, |
| 1342 | 0, sizeof(struct dsched_stats), sysctl_dsched_stats, "dsched_stats", |
| 1343 | "dsched statistics"); |
| 1344 | SYSCTL_PROC(_dsched, OID_AUTO, policies, CTLTYPE_STRING|CTLFLAG_RD, |
| 1345 | NULL, 0, sysctl_dsched_list_policies, "A", "names of available policies"); |
| 1346 | SYSCTL_PROC(_dsched_policy, OID_AUTO, default, CTLTYPE_STRING|CTLFLAG_RW, |
| 1347 | NULL, 0, sysctl_dsched_default_policy, "A", "default dsched policy"); |
| 1348 | |
| 1349 | static void |
| 1350 | dsched_sysctl_add_disk(struct dsched_disk_ctx *diskctx, char *name) |
| 1351 | { |
| 1352 | if (!(diskctx->flags & DSCHED_SYSCTL_CTX_INITED)) { |
| 1353 | diskctx->flags |= DSCHED_SYSCTL_CTX_INITED; |
| 1354 | sysctl_ctx_init(&diskctx->sysctl_ctx); |
| 1355 | } |
| 1356 | |
| 1357 | SYSCTL_ADD_PROC(&diskctx->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_dsched_policy), |
| 1358 | OID_AUTO, name, CTLTYPE_STRING|CTLFLAG_RW, |
| 1359 | diskctx, 0, sysctl_dsched_policy, "A", "policy"); |
| 1360 | } |