1 /* 2 * Copyright (c) 2003,2004,2009 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.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 35 /* 36 * lwkt_token - Implement soft token locks. 37 * 38 * Tokens are locks which serialize a thread only while the thread is 39 * running. If the thread blocks all tokens are released, then reacquired 40 * when the thread resumes. 41 * 42 * This implementation requires no critical sections or spin locks, but 43 * does use atomic_cmpset_ptr(). 44 * 45 * Tokens may be recursively acquired by the same thread. However the 46 * caller must be sure to release such tokens in reverse order. 47 */ 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/kernel.h> 51 #include <sys/proc.h> 52 #include <sys/rtprio.h> 53 #include <sys/queue.h> 54 #include <sys/sysctl.h> 55 #include <sys/ktr.h> 56 #include <sys/kthread.h> 57 #include <machine/cpu.h> 58 #include <sys/lock.h> 59 #include <sys/caps.h> 60 #include <sys/spinlock.h> 61 62 #include <sys/thread2.h> 63 #include <sys/spinlock2.h> 64 #include <sys/mplock2.h> 65 66 #include <vm/vm.h> 67 #include <vm/vm_param.h> 68 #include <vm/vm_kern.h> 69 #include <vm/vm_object.h> 70 #include <vm/vm_page.h> 71 #include <vm/vm_map.h> 72 #include <vm/vm_pager.h> 73 #include <vm/vm_extern.h> 74 #include <vm/vm_zone.h> 75 76 #include <machine/stdarg.h> 77 #include <machine/smp.h> 78 79 #ifndef LWKT_NUM_POOL_TOKENS 80 #define LWKT_NUM_POOL_TOKENS 1024 /* power of 2 */ 81 #endif 82 #define LWKT_MASK_POOL_TOKENS (LWKT_NUM_POOL_TOKENS - 1) 83 84 static lwkt_token pool_tokens[LWKT_NUM_POOL_TOKENS]; 85 86 #define TOKEN_STRING "REF=%p TOK=%p TD=%p" 87 #define CONTENDED_STRING "REF=%p TOK=%p TD=%p (contention started)" 88 #define UNCONTENDED_STRING "REF=%p TOK=%p TD=%p (contention stopped)" 89 #if !defined(KTR_TOKENS) 90 #define KTR_TOKENS KTR_ALL 91 #endif 92 93 KTR_INFO_MASTER(tokens); 94 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, sizeof(void *) * 3); 95 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, sizeof(void *) * 3); 96 #if 0 97 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, sizeof(void *) * 3); 98 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, sizeof(void *) * 3); 99 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, sizeof(void *) * 3); 100 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, sizeof(void *) * 3); 101 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, sizeof(void *) * 3); 102 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, sizeof(void *) * 3); 103 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, sizeof(void *) * 3); 104 #endif 105 106 #define logtoken(name, ref) \ 107 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread) 108 109 /* 110 * Global tokens. These replace the MP lock for major subsystem locking. 111 * These tokens are initially used to lockup both global and individual 112 * operations. 113 * 114 * Once individual structures get their own locks these tokens are used 115 * only to protect global lists & other variables and to interlock 116 * allocations and teardowns and such. 117 * 118 * The UP initializer causes token acquisition to also acquire the MP lock 119 * for maximum compatibility. The feature may be enabled and disabled at 120 * any time, the MP state is copied to the tokref when the token is acquired 121 * and will not race against sysctl changes. 122 */ 123 struct lwkt_token mp_token = LWKT_TOKEN_MP_INITIALIZER(mp_token); 124 struct lwkt_token pmap_token = LWKT_TOKEN_UP_INITIALIZER(pmap_token); 125 struct lwkt_token dev_token = LWKT_TOKEN_UP_INITIALIZER(dev_token); 126 struct lwkt_token vm_token = LWKT_TOKEN_UP_INITIALIZER(vm_token); 127 struct lwkt_token vmspace_token = LWKT_TOKEN_UP_INITIALIZER(vmspace_token); 128 struct lwkt_token kvm_token = LWKT_TOKEN_UP_INITIALIZER(kvm_token); 129 struct lwkt_token proc_token = LWKT_TOKEN_UP_INITIALIZER(proc_token); 130 struct lwkt_token tty_token = LWKT_TOKEN_UP_INITIALIZER(tty_token); 131 struct lwkt_token vnode_token = LWKT_TOKEN_UP_INITIALIZER(vnode_token); 132 struct lwkt_token vmobj_token = LWKT_TOKEN_UP_INITIALIZER(vmobj_token); 133 134 SYSCTL_INT(_lwkt, OID_AUTO, pmap_mpsafe, CTLFLAG_RW, 135 &pmap_token.t_flags, 0, "Require MP lock for pmap_token"); 136 SYSCTL_INT(_lwkt, OID_AUTO, dev_mpsafe, CTLFLAG_RW, 137 &dev_token.t_flags, 0, "Require MP lock for dev_token"); 138 SYSCTL_INT(_lwkt, OID_AUTO, vm_mpsafe, CTLFLAG_RW, 139 &vm_token.t_flags, 0, "Require MP lock for vm_token"); 140 SYSCTL_INT(_lwkt, OID_AUTO, vmspace_mpsafe, CTLFLAG_RW, 141 &vmspace_token.t_flags, 0, "Require MP lock for vmspace_token"); 142 SYSCTL_INT(_lwkt, OID_AUTO, kvm_mpsafe, CTLFLAG_RW, 143 &kvm_token.t_flags, 0, "Require MP lock for kvm_token"); 144 SYSCTL_INT(_lwkt, OID_AUTO, proc_mpsafe, CTLFLAG_RW, 145 &proc_token.t_flags, 0, "Require MP lock for proc_token"); 146 SYSCTL_INT(_lwkt, OID_AUTO, tty_mpsafe, CTLFLAG_RW, 147 &tty_token.t_flags, 0, "Require MP lock for tty_token"); 148 SYSCTL_INT(_lwkt, OID_AUTO, vnode_mpsafe, CTLFLAG_RW, 149 &vnode_token.t_flags, 0, "Require MP lock for vnode_token"); 150 SYSCTL_INT(_lwkt, OID_AUTO, vmobj_mpsafe, CTLFLAG_RW, 151 &vmobj_token.t_flags, 0, "Require MP lock for vmobj_token"); 152 153 static int lwkt_token_ipi_dispatch = 4; 154 SYSCTL_INT(_lwkt, OID_AUTO, token_ipi_dispatch, CTLFLAG_RW, 155 &lwkt_token_ipi_dispatch, 0, "Number of IPIs to dispatch on token release"); 156 157 /* 158 * The collision count is bumped every time the LWKT scheduler fails 159 * to acquire needed tokens in addition to a normal lwkt_gettoken() 160 * stall. 161 */ 162 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW, 163 &mp_token.t_collisions, 0, "Collision counter of mp_token"); 164 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW, 165 &pmap_token.t_collisions, 0, "Collision counter of pmap_token"); 166 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW, 167 &dev_token.t_collisions, 0, "Collision counter of dev_token"); 168 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW, 169 &vm_token.t_collisions, 0, "Collision counter of vm_token"); 170 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW, 171 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token"); 172 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW, 173 &kvm_token.t_collisions, 0, "Collision counter of kvm_token"); 174 SYSCTL_LONG(_lwkt, OID_AUTO, proc_collisions, CTLFLAG_RW, 175 &proc_token.t_collisions, 0, "Collision counter of proc_token"); 176 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW, 177 &tty_token.t_collisions, 0, "Collision counter of tty_token"); 178 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW, 179 &vnode_token.t_collisions, 0, "Collision counter of vnode_token"); 180 181 #ifdef SMP 182 /* 183 * Acquire the initial mplock 184 * 185 * (low level boot only) 186 */ 187 void 188 cpu_get_initial_mplock(void) 189 { 190 KKASSERT(mp_token.t_ref == NULL); 191 if (lwkt_trytoken(&mp_token) == FALSE) 192 panic("cpu_get_initial_mplock"); 193 } 194 #endif 195 196 /* 197 * Return a pool token given an address 198 */ 199 static __inline 200 lwkt_token_t 201 _lwkt_token_pool_lookup(void *ptr) 202 { 203 int i; 204 205 i = ((int)(intptr_t)ptr >> 2) ^ ((int)(intptr_t)ptr >> 12); 206 return(&pool_tokens[i & LWKT_MASK_POOL_TOKENS]); 207 } 208 209 /* 210 * Initialize a tokref_t prior to making it visible in the thread's 211 * token array. 212 * 213 * As an optimization we set the MPSAFE flag if the thread is already 214 * holding the mp_token. This bypasses unncessary calls to get_mplock() and 215 * rel_mplock() on tokens which are not normally MPSAFE when the thread 216 * is already holding the MP lock. 217 */ 218 static __inline 219 intptr_t 220 _lwkt_tok_flags(lwkt_token_t tok, thread_t td) 221 { 222 intptr_t flags; 223 224 /* 225 * tok->t_flags can change out from under us, make sure we have 226 * a local copy. 227 */ 228 flags = tok->t_flags; 229 cpu_ccfence(); 230 #ifdef SMP 231 if ((flags & LWKT_TOKEN_MPSAFE) == 0 && 232 _lwkt_token_held(&mp_token, td)) { 233 return (flags | LWKT_TOKEN_MPSAFE); 234 } else { 235 return (flags); 236 } 237 #else 238 return (flags | LWKT_TOKEN_MPSAFE); 239 #endif 240 } 241 242 static __inline 243 void 244 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, 245 intptr_t flags) 246 { 247 ref->tr_tok = tok; 248 ref->tr_owner = td; 249 ref->tr_flags = flags; 250 } 251 252 #ifdef SMP 253 /* 254 * Force a LWKT reschedule on the target cpu when a requested token 255 * becomes available. 256 */ 257 static 258 void 259 lwkt_reltoken_mask_remote(void *arg, int arg2, struct intrframe *frame) 260 { 261 need_lwkt_resched(); 262 } 263 #endif 264 265 /* 266 * This bit of code sends a LWKT reschedule request to whatever other cpus 267 * had contended on the token being released. We could wake up all the cpus 268 * but generally speaking if there is a lot of contention we really only want 269 * to wake up a subset of cpus to avoid aggregating O(N^2) IPIs. The current 270 * cpuid is used as a basis to select which other cpus to wake up. 271 * 272 * For the selected cpus we can avoid issuing the actual IPI if the target 273 * cpu's RQF_WAKEUP is already set. In this case simply setting the 274 * reschedule flag RQF_AST_LWKT_RESCHED will be sufficient. 275 * 276 * lwkt.token_ipi_dispatch specifies the maximum number of IPIs to dispatch 277 * on a token release. 278 */ 279 static __inline 280 void 281 _lwkt_reltoken_mask(lwkt_token_t tok) 282 { 283 #ifdef SMP 284 globaldata_t ngd; 285 cpumask_t mask; 286 cpumask_t tmpmask; 287 cpumask_t wumask; /* wakeup mask */ 288 cpumask_t remask; /* clear mask */ 289 int wucount; /* wakeup count */ 290 int cpuid; 291 int reqflags; 292 293 /* 294 * Mask of contending cpus we want to wake up. 295 */ 296 mask = tok->t_collmask; 297 cpu_ccfence(); 298 if (mask == 0) 299 return; 300 301 /* 302 * Degenerate case - IPI to all contending cpus 303 */ 304 wucount = lwkt_token_ipi_dispatch; 305 if (wucount <= 0 || wucount >= ncpus) { 306 wucount = 0; 307 wumask = mask; 308 remask = mask; 309 } else { 310 wumask = 0; 311 remask = 0; 312 } 313 314 /* 315 * Calculate which cpus to IPI. These cpus are potentially in a 316 * HLT state waiting for token contention to go away. 317 * 318 * Ask the cpu LWKT scheduler to reschedule by setting 319 * RQF_AST_LWKT_RESCHEDULE. Signal the cpu if RQF_WAKEUP is not 320 * set (otherwise it has already been signalled or will check the 321 * flag very soon anyway). Both bits must be adjusted atomically 322 * all in one go to avoid races. 323 * 324 * The collision mask is cleared for all cpus we set the resched 325 * flag for, but we only IPI the ones that need signalling. 326 */ 327 while (wucount && mask) { 328 tmpmask = mask & ~(CPUMASK(mycpu->gd_cpuid) - 1); 329 if (tmpmask) 330 cpuid = BSFCPUMASK(tmpmask); 331 else 332 cpuid = BSFCPUMASK(mask); 333 ngd = globaldata_find(cpuid); 334 for (;;) { 335 reqflags = ngd->gd_reqflags; 336 if (atomic_cmpset_int(&ngd->gd_reqflags, reqflags, 337 reqflags | 338 (RQF_WAKEUP | 339 RQF_AST_LWKT_RESCHED))) { 340 break; 341 } 342 } 343 if ((reqflags & RQF_WAKEUP) == 0) { 344 wumask |= CPUMASK(cpuid); 345 --wucount; 346 } 347 remask |= CPUMASK(cpuid); 348 mask &= ~CPUMASK(cpuid); 349 } 350 if (remask) { 351 atomic_clear_cpumask(&tok->t_collmask, remask); 352 lwkt_send_ipiq3_mask(wumask, lwkt_reltoken_mask_remote, 353 NULL, 0); 354 } 355 #endif 356 } 357 358 /* 359 * Obtain all the tokens required by the specified thread on the current 360 * cpu, return 0 on failure and non-zero on success. If a failure occurs 361 * any partially acquired tokens will be released prior to return. 362 * 363 * lwkt_getalltokens is called by the LWKT scheduler to acquire all 364 * tokens that the thread had acquired prior to going to sleep. 365 * 366 * We always clear the collision mask on token aquision. 367 * 368 * Called from a critical section. 369 */ 370 int 371 lwkt_getalltokens(thread_t td) 372 { 373 lwkt_tokref_t scan; 374 lwkt_tokref_t ref; 375 lwkt_token_t tok; 376 377 /* 378 * Acquire tokens in forward order, assign or validate tok->t_ref. 379 */ 380 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 381 tok = scan->tr_tok; 382 for (;;) { 383 /* 384 * Try to acquire the token if we do not already have 385 * it. 386 * 387 * NOTE: If atomic_cmpset_ptr() fails we have to 388 * loop and try again. It just means we 389 * lost a cpu race. 390 */ 391 ref = tok->t_ref; 392 if (ref == NULL) { 393 if (atomic_cmpset_ptr(&tok->t_ref, NULL, scan)) 394 { 395 if (tok->t_collmask & td->td_gd->gd_cpumask) { 396 atomic_clear_cpumask(&tok->t_collmask, 397 td->td_gd->gd_cpumask); 398 } 399 break; 400 } 401 continue; 402 } 403 404 /* 405 * Someone holds the token. 406 * 407 * Test if ref is already recursively held by this 408 * thread. We cannot safely dereference tok->t_ref 409 * (it might belong to another thread and is thus 410 * unstable), but we don't have to. We can simply 411 * range-check it. 412 */ 413 if (ref >= &td->td_toks_base && ref < td->td_toks_stop) 414 break; 415 416 #ifdef SMP 417 /* 418 * Otherwise we failed to acquire all the tokens. 419 * Undo and return. We have to try once more after 420 * setting cpumask to cover possible races against 421 * the checking of t_collmask. 422 */ 423 atomic_set_cpumask(&tok->t_collmask, 424 td->td_gd->gd_cpumask); 425 if (atomic_cmpset_ptr(&tok->t_ref, NULL, scan)) { 426 if (tok->t_collmask & td->td_gd->gd_cpumask) { 427 atomic_clear_cpumask(&tok->t_collmask, 428 td->td_gd->gd_cpumask); 429 } 430 break; 431 } 432 #endif 433 td->td_wmesg = tok->t_desc; 434 atomic_add_long(&tok->t_collisions, 1); 435 lwkt_relalltokens(td); 436 return(FALSE); 437 } 438 } 439 return (TRUE); 440 } 441 442 /* 443 * Release all tokens owned by the specified thread on the current cpu. 444 * 445 * This code is really simple. Even in cases where we own all the tokens 446 * note that t_ref may not match the scan for recursively held tokens, 447 * or for the case where a lwkt_getalltokens() failed. 448 * 449 * The scheduler is responsible for maintaining the MP lock count, so 450 * we don't need to deal with tr_flags here. 451 * 452 * Called from a critical section. 453 */ 454 void 455 lwkt_relalltokens(thread_t td) 456 { 457 lwkt_tokref_t scan; 458 lwkt_token_t tok; 459 460 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 461 tok = scan->tr_tok; 462 if (tok->t_ref == scan) { 463 tok->t_ref = NULL; 464 _lwkt_reltoken_mask(tok); 465 } 466 } 467 } 468 469 /* 470 * Token acquisition helper function. The caller must have already 471 * made nref visible by adjusting td_toks_stop and will be responsible 472 * for the disposition of nref on either success or failure. 473 * 474 * When acquiring tokens recursively we want tok->t_ref to point to 475 * the outer (first) acquisition so it gets cleared only on the last 476 * release. 477 */ 478 static __inline 479 int 480 _lwkt_trytokref2(lwkt_tokref_t nref, thread_t td, int blocking) 481 { 482 lwkt_token_t tok; 483 lwkt_tokref_t ref; 484 485 /* 486 * Make sure the compiler does not reorder prior instructions 487 * beyond this demark. 488 */ 489 cpu_ccfence(); 490 491 /* 492 * Attempt to gain ownership 493 */ 494 tok = nref->tr_tok; 495 for (;;) { 496 /* 497 * Try to acquire the token if we do not already have 498 * it. This is not allowed if we are in a hard code 499 * section (because it 'might' have blocked). 500 */ 501 ref = tok->t_ref; 502 if (ref == NULL) { 503 KASSERT((blocking == 0 || 504 td->td_gd->gd_intr_nesting_level == 0 || 505 panic_cpu_gd == mycpu), 506 ("Attempt to acquire token %p not already " 507 "held in hard code section", tok)); 508 509 /* 510 * NOTE: If atomic_cmpset_ptr() fails we have to 511 * loop and try again. It just means we 512 * lost a cpu race. 513 */ 514 if (atomic_cmpset_ptr(&tok->t_ref, NULL, nref)) 515 return (TRUE); 516 continue; 517 } 518 519 /* 520 * Test if ref is already recursively held by this 521 * thread. We cannot safely dereference tok->t_ref 522 * (it might belong to another thread and is thus 523 * unstable), but we don't have to. We can simply 524 * range-check it. 525 * 526 * It is ok to acquire a token that is already held 527 * by the current thread when in a hard code section. 528 */ 529 if (ref >= &td->td_toks_base && ref < td->td_toks_stop) 530 return(TRUE); 531 532 /* 533 * Otherwise we failed, and it is not ok to attempt to 534 * acquire a token in a hard code section. 535 */ 536 KASSERT((blocking == 0 || 537 td->td_gd->gd_intr_nesting_level == 0), 538 ("Attempt to acquire token %p not already " 539 "held in hard code section", tok)); 540 541 return(FALSE); 542 } 543 } 544 545 /* 546 * Get a serializing token. This routine can block. 547 */ 548 void 549 lwkt_gettoken(lwkt_token_t tok) 550 { 551 thread_t td = curthread; 552 lwkt_tokref_t ref; 553 intptr_t flags; 554 555 flags = _lwkt_tok_flags(tok, td); 556 if ((flags & LWKT_TOKEN_MPSAFE) == 0) 557 get_mplock(); 558 559 ref = td->td_toks_stop; 560 KKASSERT(ref < &td->td_toks_end); 561 ++td->td_toks_stop; 562 cpu_ccfence(); 563 _lwkt_tokref_init(ref, tok, td, flags); 564 565 if (_lwkt_trytokref2(ref, td, 1) == FALSE) { 566 /* 567 * Give up running if we can't acquire the token right now. 568 * 569 * Since the tokref is already active the scheduler now 570 * takes care of acquisition, so we need only call 571 * lwkt_switch(). 572 * 573 * Since we failed this was not a recursive token so upon 574 * return tr_tok->t_ref should be assigned to this specific 575 * ref. 576 */ 577 #ifdef SMP 578 #if 0 579 /* 580 * (DISABLED ATM) - Do not set t_collmask on a token 581 * acquisition failure, the scheduler will spin at least 582 * once and deal with hlt/spin semantics. 583 */ 584 atomic_set_cpumask(&tok->t_collmask, td->td_gd->gd_cpumask); 585 if (atomic_cmpset_ptr(&tok->t_ref, NULL, ref)) { 586 atomic_clear_cpumask(&tok->t_collmask, 587 td->td_gd->gd_cpumask); 588 return; 589 } 590 #endif 591 #endif 592 td->td_wmesg = tok->t_desc; 593 atomic_add_long(&tok->t_collisions, 1); 594 logtoken(fail, ref); 595 lwkt_switch(); 596 logtoken(succ, ref); 597 KKASSERT(tok->t_ref == ref); 598 } 599 } 600 601 void 602 lwkt_gettoken_hard(lwkt_token_t tok) 603 { 604 thread_t td = curthread; 605 lwkt_tokref_t ref; 606 intptr_t flags; 607 608 flags = _lwkt_tok_flags(tok, td); 609 if ((flags & LWKT_TOKEN_MPSAFE) == 0) 610 get_mplock(); 611 612 ref = td->td_toks_stop; 613 KKASSERT(ref < &td->td_toks_end); 614 ++td->td_toks_stop; 615 cpu_ccfence(); 616 _lwkt_tokref_init(ref, tok, td, flags); 617 618 if (_lwkt_trytokref2(ref, td, 1) == FALSE) { 619 /* 620 * Give up running if we can't acquire the token right now. 621 * 622 * Since the tokref is already active the scheduler now 623 * takes care of acquisition, so we need only call 624 * lwkt_switch(). 625 * 626 * Since we failed this was not a recursive token so upon 627 * return tr_tok->t_ref should be assigned to this specific 628 * ref. 629 */ 630 #ifdef SMP 631 #if 0 632 /* 633 * (DISABLED ATM) - Do not set t_collmask on a token 634 * acquisition failure, the scheduler will spin at least 635 * once and deal with hlt/spin semantics. 636 */ 637 atomic_set_cpumask(&tok->t_collmask, td->td_gd->gd_cpumask); 638 if (atomic_cmpset_ptr(&tok->t_ref, NULL, ref)) { 639 atomic_clear_cpumask(&tok->t_collmask, 640 td->td_gd->gd_cpumask); 641 goto success; 642 } 643 #endif 644 #endif 645 td->td_wmesg = tok->t_desc; 646 atomic_add_long(&tok->t_collisions, 1); 647 logtoken(fail, ref); 648 lwkt_switch(); 649 logtoken(succ, ref); 650 KKASSERT(tok->t_ref == ref); 651 } 652 #ifdef SMP 653 #if 0 654 success: 655 #endif 656 #endif 657 crit_enter_hard_gd(td->td_gd); 658 } 659 660 lwkt_token_t 661 lwkt_getpooltoken(void *ptr) 662 { 663 thread_t td = curthread; 664 lwkt_token_t tok; 665 lwkt_tokref_t ref; 666 intptr_t flags; 667 668 tok = _lwkt_token_pool_lookup(ptr); 669 flags = _lwkt_tok_flags(tok, td); 670 if ((flags & LWKT_TOKEN_MPSAFE) == 0) 671 get_mplock(); 672 673 ref = td->td_toks_stop; 674 KKASSERT(ref < &td->td_toks_end); 675 ++td->td_toks_stop; 676 cpu_ccfence(); 677 _lwkt_tokref_init(ref, tok, td, flags); 678 679 if (_lwkt_trytokref2(ref, td, 1) == FALSE) { 680 /* 681 * Give up running if we can't acquire the token right now. 682 * 683 * Since the tokref is already active the scheduler now 684 * takes care of acquisition, so we need only call 685 * lwkt_switch(). 686 * 687 * Since we failed this was not a recursive token so upon 688 * return tr_tok->t_ref should be assigned to this specific 689 * ref. 690 */ 691 #ifdef SMP 692 #if 0 693 /* 694 * (DISABLED ATM) - Do not set t_collmask on a token 695 * acquisition failure, the scheduler will spin at least 696 * once and deal with hlt/spin semantics. 697 */ 698 atomic_set_cpumask(&tok->t_collmask, td->td_gd->gd_cpumask); 699 if (atomic_cmpset_ptr(&tok->t_ref, NULL, ref)) { 700 atomic_clear_cpumask(&tok->t_collmask, 701 td->td_gd->gd_cpumask); 702 goto success; 703 } 704 #endif 705 #endif 706 td->td_wmesg = tok->t_desc; 707 atomic_add_long(&tok->t_collisions, 1); 708 logtoken(fail, ref); 709 lwkt_switch(); 710 logtoken(succ, ref); 711 KKASSERT(tok->t_ref == ref); 712 } 713 #ifdef SMP 714 #if 0 715 success: 716 #endif 717 #endif 718 return(tok); 719 } 720 721 /* 722 * Attempt to acquire a token, return TRUE on success, FALSE on failure. 723 */ 724 int 725 lwkt_trytoken(lwkt_token_t tok) 726 { 727 thread_t td = curthread; 728 lwkt_tokref_t ref; 729 intptr_t flags; 730 731 flags = _lwkt_tok_flags(tok, td); 732 if ((flags & LWKT_TOKEN_MPSAFE) == 0) { 733 if (try_mplock() == 0) 734 return (FALSE); 735 } 736 737 ref = td->td_toks_stop; 738 KKASSERT(ref < &td->td_toks_end); 739 ++td->td_toks_stop; 740 cpu_ccfence(); 741 _lwkt_tokref_init(ref, tok, td, flags); 742 743 if (_lwkt_trytokref2(ref, td, 0) == FALSE) { 744 /* 745 * Cleanup, deactivate the failed token. 746 */ 747 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) { 748 cpu_ccfence(); 749 --td->td_toks_stop; 750 cpu_ccfence(); 751 rel_mplock(); 752 } else { 753 cpu_ccfence(); 754 --td->td_toks_stop; 755 } 756 return (FALSE); 757 } 758 return (TRUE); 759 } 760 761 /* 762 * Release a serializing token. 763 * 764 * WARNING! All tokens must be released in reverse order. This will be 765 * asserted. 766 */ 767 void 768 lwkt_reltoken(lwkt_token_t tok) 769 { 770 thread_t td = curthread; 771 lwkt_tokref_t ref; 772 773 /* 774 * Remove ref from thread token list and assert that it matches 775 * the token passed in. Tokens must be released in reverse order. 776 */ 777 ref = td->td_toks_stop - 1; 778 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok); 779 780 /* 781 * Only clear the token if it matches ref. If ref was a recursively 782 * acquired token it may not match. Then adjust td_toks_stop. 783 * 784 * Some comparisons must be run prior to adjusting td_toks_stop 785 * to avoid racing against a fast interrupt/ ipi which tries to 786 * acquire a token. 787 * 788 * We must also be absolutely sure that the compiler does not 789 * reorder the clearing of t_ref and the adjustment of td_toks_stop, 790 * or reorder the adjustment of td_toks_stop against the conditional. 791 * 792 * NOTE: The mplock is a token also so sequencing is a bit complex. 793 */ 794 if (tok->t_ref == ref) { 795 tok->t_ref = NULL; 796 _lwkt_reltoken_mask(tok); 797 } 798 cpu_sfence(); 799 if ((ref->tr_flags & LWKT_TOKEN_MPSAFE) == 0) { 800 cpu_ccfence(); 801 td->td_toks_stop = ref; 802 cpu_ccfence(); 803 rel_mplock(); 804 } else { 805 cpu_ccfence(); 806 td->td_toks_stop = ref; 807 cpu_ccfence(); 808 } 809 KKASSERT(tok->t_ref != ref); 810 } 811 812 void 813 lwkt_reltoken_hard(lwkt_token_t tok) 814 { 815 lwkt_reltoken(tok); 816 crit_exit_hard(); 817 } 818 819 /* 820 * It is faster for users of lwkt_getpooltoken() to use the returned 821 * token and just call lwkt_reltoken(), but for convenience we provide 822 * this function which looks the token up based on the ident. 823 */ 824 void 825 lwkt_relpooltoken(void *ptr) 826 { 827 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr); 828 lwkt_reltoken(tok); 829 } 830 831 /* 832 * Return a count of the number of token refs the thread has to the 833 * specified token, whether it currently owns the token or not. 834 */ 835 int 836 lwkt_cnttoken(lwkt_token_t tok, thread_t td) 837 { 838 lwkt_tokref_t scan; 839 int count = 0; 840 841 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 842 if (scan->tr_tok == tok) 843 ++count; 844 } 845 return(count); 846 } 847 848 849 /* 850 * Pool tokens are used to provide a type-stable serializing token 851 * pointer that does not race against disappearing data structures. 852 * 853 * This routine is called in early boot just after we setup the BSP's 854 * globaldata structure. 855 */ 856 void 857 lwkt_token_pool_init(void) 858 { 859 int i; 860 861 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i) 862 lwkt_token_init(&pool_tokens[i], 1, "pool"); 863 } 864 865 lwkt_token_t 866 lwkt_token_pool_lookup(void *ptr) 867 { 868 return (_lwkt_token_pool_lookup(ptr)); 869 } 870 871 /* 872 * Initialize a token. If mpsafe is 0, the MP lock is acquired before 873 * acquiring the token and released after releasing the token. 874 */ 875 void 876 lwkt_token_init(lwkt_token_t tok, int mpsafe, const char *desc) 877 { 878 tok->t_ref = NULL; 879 tok->t_flags = mpsafe ? LWKT_TOKEN_MPSAFE : 0; 880 tok->t_collisions = 0; 881 tok->t_collmask = 0; 882 tok->t_desc = desc; 883 } 884 885 void 886 lwkt_token_uninit(lwkt_token_t tok) 887 { 888 /* empty */ 889 } 890 891 #if 0 892 int 893 lwkt_token_is_stale(lwkt_tokref_t ref) 894 { 895 lwkt_token_t tok = ref->tr_tok; 896 897 KKASSERT(tok->t_owner == curthread && ref->tr_state == 1 && 898 tok->t_count > 0); 899 900 /* Token is not stale */ 901 if (tok->t_lastowner == tok->t_owner) 902 return (FALSE); 903 904 /* 905 * The token is stale. Reset to not stale so that the next call to 906 * lwkt_token_is_stale will return "not stale" unless the token 907 * was acquired in-between by another thread. 908 */ 909 tok->t_lastowner = tok->t_owner; 910 return (TRUE); 911 } 912 #endif 913