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/spinlock.h> 60 61 #include <sys/thread2.h> 62 #include <sys/spinlock2.h> 63 #include <sys/mplock2.h> 64 65 #include <vm/vm.h> 66 #include <vm/vm_param.h> 67 #include <vm/vm_kern.h> 68 #include <vm/vm_object.h> 69 #include <vm/vm_page.h> 70 #include <vm/vm_map.h> 71 #include <vm/vm_pager.h> 72 #include <vm/vm_extern.h> 73 #include <vm/vm_zone.h> 74 75 #include <machine/stdarg.h> 76 #include <machine/smp.h> 77 78 #include "opt_ddb.h" 79 #ifdef DDB 80 #include <ddb/ddb.h> 81 #endif 82 83 extern int lwkt_sched_debug; 84 85 #ifndef LWKT_NUM_POOL_TOKENS 86 #define LWKT_NUM_POOL_TOKENS 4001 /* prime number */ 87 #endif 88 89 struct lwkt_pool_token { 90 struct lwkt_token token; 91 } __cachealign; 92 93 static struct lwkt_pool_token pool_tokens[LWKT_NUM_POOL_TOKENS]; 94 struct spinlock tok_debug_spin = SPINLOCK_INITIALIZER(&tok_debug_spin, "tok_debug_spin"); 95 96 #define TOKEN_STRING "REF=%p TOK=%p TD=%p" 97 #define TOKEN_ARGS lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td 98 #define CONTENDED_STRING TOKEN_STRING " (contention started)" 99 #define UNCONTENDED_STRING TOKEN_STRING " (contention stopped)" 100 #if !defined(KTR_TOKENS) 101 #define KTR_TOKENS KTR_ALL 102 #endif 103 104 KTR_INFO_MASTER(tokens); 105 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS); 106 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS); 107 #if 0 108 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS); 109 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS); 110 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS); 111 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS); 112 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS); 113 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS); 114 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS); 115 #endif 116 117 #define logtoken(name, ref) \ 118 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread) 119 120 /* 121 * Global tokens. These replace the MP lock for major subsystem locking. 122 * These tokens are initially used to lockup both global and individual 123 * operations. 124 * 125 * Once individual structures get their own locks these tokens are used 126 * only to protect global lists & other variables and to interlock 127 * allocations and teardowns and such. 128 * 129 * The UP initializer causes token acquisition to also acquire the MP lock 130 * for maximum compatibility. The feature may be enabled and disabled at 131 * any time, the MP state is copied to the tokref when the token is acquired 132 * and will not race against sysctl changes. 133 */ 134 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token); 135 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token); 136 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token); 137 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token); 138 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token); 139 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token); 140 struct lwkt_token sigio_token = LWKT_TOKEN_INITIALIZER(sigio_token); 141 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token); 142 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token); 143 144 static int lwkt_token_spin = 5; 145 SYSCTL_INT(_lwkt, OID_AUTO, token_spin, CTLFLAG_RW, 146 &lwkt_token_spin, 0, "Decontention spin loops"); 147 static int lwkt_token_delay = 0; 148 SYSCTL_INT(_lwkt, OID_AUTO, token_delay, CTLFLAG_RW, 149 &lwkt_token_delay, 0, "Decontention spin delay in ns"); 150 151 /* 152 * The collision count is bumped every time the LWKT scheduler fails 153 * to acquire needed tokens in addition to a normal lwkt_gettoken() 154 * stall. 155 */ 156 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW, 157 &mp_token.t_collisions, 0, "Collision counter of mp_token"); 158 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW, 159 &pmap_token.t_collisions, 0, "Collision counter of pmap_token"); 160 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW, 161 &dev_token.t_collisions, 0, "Collision counter of dev_token"); 162 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW, 163 &vm_token.t_collisions, 0, "Collision counter of vm_token"); 164 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW, 165 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token"); 166 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW, 167 &kvm_token.t_collisions, 0, "Collision counter of kvm_token"); 168 SYSCTL_LONG(_lwkt, OID_AUTO, sigio_collisions, CTLFLAG_RW, 169 &sigio_token.t_collisions, 0, "Collision counter of sigio_token"); 170 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW, 171 &tty_token.t_collisions, 0, "Collision counter of tty_token"); 172 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW, 173 &vnode_token.t_collisions, 0, "Collision counter of vnode_token"); 174 175 int tokens_debug_output; 176 SYSCTL_INT(_lwkt, OID_AUTO, tokens_debug_output, CTLFLAG_RW, 177 &tokens_debug_output, 0, "Generate stack trace N times"); 178 179 180 #ifdef DEBUG_LOCKS_LATENCY 181 182 static long tokens_add_latency; 183 SYSCTL_LONG(_debug, OID_AUTO, tokens_add_latency, CTLFLAG_RW, 184 &tokens_add_latency, 0, 185 "Add spinlock latency"); 186 187 #endif 188 189 190 static int _lwkt_getalltokens_sorted(thread_t td); 191 192 /* 193 * Acquire the initial mplock 194 * 195 * (low level boot only) 196 */ 197 void 198 cpu_get_initial_mplock(void) 199 { 200 KKASSERT(mp_token.t_ref == NULL); 201 if (lwkt_trytoken(&mp_token) == FALSE) 202 panic("cpu_get_initial_mplock"); 203 } 204 205 /* 206 * Return a pool token given an address. Use a prime number to reduce 207 * overlaps. 208 */ 209 static __inline 210 lwkt_token_t 211 _lwkt_token_pool_lookup(void *ptr) 212 { 213 u_int i; 214 215 i = (u_int)(uintptr_t)ptr % LWKT_NUM_POOL_TOKENS; 216 return (&pool_tokens[i].token); 217 } 218 219 /* 220 * Initialize a tokref_t prior to making it visible in the thread's 221 * token array. 222 */ 223 static __inline 224 void 225 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl) 226 { 227 ref->tr_tok = tok; 228 ref->tr_count = excl; 229 ref->tr_owner = td; 230 } 231 232 /* 233 * Attempt to acquire a shared or exclusive token. Returns TRUE on success, 234 * FALSE on failure. 235 * 236 * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive 237 * token, otherwise are attempting to get a shared token. 238 * 239 * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise 240 * it is a non-blocking operation (for both exclusive or shared acquisions). 241 */ 242 static __inline 243 int 244 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode) 245 { 246 lwkt_token_t tok; 247 lwkt_tokref_t oref; 248 long count; 249 250 tok = ref->tr_tok; 251 KASSERT(((mode & TOK_EXCLREQ) == 0 || /* non blocking */ 252 td->td_gd->gd_intr_nesting_level == 0 || 253 panic_cpu_gd == mycpu), 254 ("Attempt to acquire token %p not already " 255 "held in hard code section", tok)); 256 257 if (mode & TOK_EXCLUSIVE) { 258 /* 259 * Attempt to get an exclusive token 260 */ 261 for (;;) { 262 count = tok->t_count; 263 oref = tok->t_ref; /* can be NULL */ 264 cpu_ccfence(); 265 if ((count & ~TOK_EXCLREQ) == 0) { 266 /* 267 * It is possible to get the exclusive bit. 268 * We must clear TOK_EXCLREQ on successful 269 * acquisition. 270 */ 271 if (atomic_cmpset_long(&tok->t_count, count, 272 (count & ~TOK_EXCLREQ) | 273 TOK_EXCLUSIVE)) { 274 KKASSERT(tok->t_ref == NULL); 275 tok->t_ref = ref; 276 return TRUE; 277 } 278 /* retry */ 279 } else if ((count & TOK_EXCLUSIVE) && 280 oref >= &td->td_toks_base && 281 oref < td->td_toks_stop) { 282 /* 283 * Our thread already holds the exclusive 284 * bit, we treat this tokref as a shared 285 * token (sorta) to make the token release 286 * code easier. 287 * 288 * NOTE: oref cannot race above if it 289 * happens to be ours, so we're good. 290 * But we must still have a stable 291 * variable for both parts of the 292 * comparison. 293 * 294 * NOTE: Since we already have an exclusive 295 * lock and don't need to check EXCLREQ 296 * we can just use an atomic_add here 297 */ 298 atomic_add_long(&tok->t_count, TOK_INCR); 299 ref->tr_count &= ~TOK_EXCLUSIVE; 300 return TRUE; 301 } else if ((mode & TOK_EXCLREQ) && 302 (count & TOK_EXCLREQ) == 0) { 303 /* 304 * Unable to get the exclusive bit but being 305 * asked to set the exclusive-request bit. 306 * Since we are going to retry anyway just 307 * set the bit unconditionally. 308 */ 309 atomic_set_long(&tok->t_count, TOK_EXCLREQ); 310 return FALSE; 311 } else { 312 /* 313 * Unable to get the exclusive bit and not 314 * being asked to set the exclusive-request 315 * (aka lwkt_trytoken()), or EXCLREQ was 316 * already set. 317 */ 318 cpu_pause(); 319 return FALSE; 320 } 321 /* retry */ 322 } 323 } else { 324 /* 325 * Attempt to get a shared token. Note that TOK_EXCLREQ 326 * for shared tokens simply means the caller intends to 327 * block. We never actually set the bit in tok->t_count. 328 */ 329 for (;;) { 330 count = tok->t_count; 331 oref = tok->t_ref; /* can be NULL */ 332 cpu_ccfence(); 333 if ((count & (TOK_EXCLUSIVE/*|TOK_EXCLREQ*/)) == 0) { 334 /* 335 * It may be possible to get the token shared. 336 */ 337 if ((atomic_fetchadd_long(&tok->t_count, TOK_INCR) & TOK_EXCLUSIVE) == 0) { 338 return TRUE; 339 } 340 atomic_fetchadd_long(&tok->t_count, -TOK_INCR); 341 /* retry */ 342 } else if ((count & TOK_EXCLUSIVE) && 343 oref >= &td->td_toks_base && 344 oref < td->td_toks_stop) { 345 /* 346 * We own the exclusive bit on the token so 347 * we can in fact also get it shared. 348 */ 349 atomic_add_long(&tok->t_count, TOK_INCR); 350 return TRUE; 351 } else { 352 /* 353 * We failed to get the token shared 354 */ 355 return FALSE; 356 } 357 /* retry */ 358 } 359 } 360 } 361 362 static __inline 363 int 364 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode) 365 { 366 int spin; 367 368 if (_lwkt_trytokref(ref, td, mode)) { 369 #ifdef DEBUG_LOCKS_LATENCY 370 long j; 371 for (j = tokens_add_latency; j > 0; --j) 372 cpu_ccfence(); 373 #endif 374 return TRUE; 375 } 376 for (spin = lwkt_token_spin; spin > 0; --spin) { 377 if (lwkt_token_delay) 378 tsc_delay(lwkt_token_delay); 379 else 380 cpu_pause(); 381 if (_lwkt_trytokref(ref, td, mode)) { 382 #ifdef DEBUG_LOCKS_LATENCY 383 long j; 384 for (j = tokens_add_latency; j > 0; --j) 385 cpu_ccfence(); 386 #endif 387 return TRUE; 388 } 389 } 390 return FALSE; 391 } 392 393 /* 394 * Release a token that we hold. 395 */ 396 static __inline 397 void 398 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td) 399 { 400 lwkt_token_t tok; 401 long count; 402 403 tok = ref->tr_tok; 404 for (;;) { 405 count = tok->t_count; 406 cpu_ccfence(); 407 if (tok->t_ref == ref) { 408 /* 409 * We are an exclusive holder. We must clear tr_ref 410 * before we clear the TOK_EXCLUSIVE bit. If we are 411 * unable to clear the bit we must restore 412 * tok->t_ref. 413 */ 414 KKASSERT(count & TOK_EXCLUSIVE); 415 tok->t_ref = NULL; 416 if (atomic_cmpset_long(&tok->t_count, count, 417 count & ~TOK_EXCLUSIVE)) { 418 return; 419 } 420 tok->t_ref = ref; 421 /* retry */ 422 } else { 423 /* 424 * We are a shared holder 425 */ 426 KKASSERT(count & TOK_COUNTMASK); 427 if (atomic_cmpset_long(&tok->t_count, count, 428 count - TOK_INCR)) { 429 return; 430 } 431 /* retry */ 432 } 433 /* retry */ 434 } 435 } 436 437 /* 438 * Obtain all the tokens required by the specified thread on the current 439 * cpu, return 0 on failure and non-zero on success. If a failure occurs 440 * any partially acquired tokens will be released prior to return. 441 * 442 * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all 443 * tokens that the thread had to release when it switched away. 444 * 445 * If spinning is non-zero this function acquires the tokens in a particular 446 * order to deal with potential deadlocks. We simply use address order for 447 * the case. 448 * 449 * Called from a critical section. 450 */ 451 int 452 lwkt_getalltokens(thread_t td, int spinning) 453 { 454 lwkt_tokref_t scan; 455 lwkt_token_t tok; 456 457 if (spinning) 458 return(_lwkt_getalltokens_sorted(td)); 459 460 /* 461 * Acquire tokens in forward order, assign or validate tok->t_ref. 462 */ 463 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 464 tok = scan->tr_tok; 465 for (;;) { 466 /* 467 * Only try really hard on the last token 468 */ 469 if (scan == td->td_toks_stop - 1) { 470 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 471 break; 472 } else { 473 if (_lwkt_trytokref(scan, td, scan->tr_count)) 474 break; 475 } 476 477 /* 478 * Otherwise we failed to acquire all the tokens. 479 * Release whatever we did get. 480 */ 481 KASSERT(tok->t_desc, 482 ("token %p is not initialized", tok)); 483 strncpy(td->td_gd->gd_cnt.v_lock_name, 484 tok->t_desc, 485 sizeof(td->td_gd->gd_cnt.v_lock_name) - 1); 486 487 if (lwkt_sched_debug > 0) { 488 --lwkt_sched_debug; 489 kprintf("toka %p %s %s\n", 490 tok, tok->t_desc, td->td_comm); 491 } 492 td->td_wmesg = tok->t_desc; 493 ++tok->t_collisions; 494 while (--scan >= &td->td_toks_base) 495 _lwkt_reltokref(scan, td); 496 return(FALSE); 497 } 498 } 499 return (TRUE); 500 } 501 502 /* 503 * Release all tokens owned by the specified thread on the current cpu. 504 * 505 * This code is really simple. Even in cases where we own all the tokens 506 * note that t_ref may not match the scan for recursively held tokens which 507 * are held deeper in the stack, or for the case where a lwkt_getalltokens() 508 * failed. 509 * 510 * Tokens are released in reverse order to reduce chasing race failures. 511 * 512 * Called from a critical section. 513 */ 514 void 515 lwkt_relalltokens(thread_t td) 516 { 517 lwkt_tokref_t scan; 518 519 /* 520 * Weird order is to try to avoid a panic loop 521 */ 522 if (td->td_toks_have) { 523 scan = td->td_toks_have; 524 td->td_toks_have = NULL; 525 } else { 526 scan = td->td_toks_stop; 527 } 528 while (--scan >= &td->td_toks_base) 529 _lwkt_reltokref(scan, td); 530 } 531 532 /* 533 * This is the decontention version of lwkt_getalltokens(). The tokens are 534 * acquired in address-sorted order to deal with any deadlocks. Ultimately 535 * token failures will spin into the scheduler and get here. 536 * 537 * Called from critical section 538 */ 539 static 540 int 541 _lwkt_getalltokens_sorted(thread_t td) 542 { 543 lwkt_tokref_t sort_array[LWKT_MAXTOKENS]; 544 lwkt_tokref_t scan; 545 lwkt_token_t tok; 546 int i; 547 int j; 548 int n; 549 550 /* 551 * Sort the token array. Yah yah, I know this isn't fun. 552 * 553 * NOTE: Recursively acquired tokens are ordered the same as in the 554 * td_toks_array so we can always get the earliest one first. 555 */ 556 i = 0; 557 scan = &td->td_toks_base; 558 while (scan < td->td_toks_stop) { 559 for (j = 0; j < i; ++j) { 560 if (scan->tr_tok < sort_array[j]->tr_tok) 561 break; 562 } 563 if (j != i) { 564 bcopy(sort_array + j, sort_array + j + 1, 565 (i - j) * sizeof(lwkt_tokref_t)); 566 } 567 sort_array[j] = scan; 568 ++scan; 569 ++i; 570 } 571 n = i; 572 573 /* 574 * Acquire tokens in forward order, assign or validate tok->t_ref. 575 */ 576 for (i = 0; i < n; ++i) { 577 scan = sort_array[i]; 578 tok = scan->tr_tok; 579 for (;;) { 580 /* 581 * Only try really hard on the last token 582 */ 583 if (scan == td->td_toks_stop - 1) { 584 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 585 break; 586 } else { 587 if (_lwkt_trytokref(scan, td, scan->tr_count)) 588 break; 589 } 590 591 /* 592 * Otherwise we failed to acquire all the tokens. 593 * Release whatever we did get. 594 */ 595 if (lwkt_sched_debug > 0) { 596 --lwkt_sched_debug; 597 kprintf("tokb %p %s %s\n", 598 tok, tok->t_desc, td->td_comm); 599 } 600 td->td_wmesg = tok->t_desc; 601 ++tok->t_collisions; 602 while (--i >= 0) { 603 scan = sort_array[i]; 604 _lwkt_reltokref(scan, td); 605 } 606 return(FALSE); 607 } 608 } 609 610 /* 611 * We were successful, there is no need for another core to signal 612 * us. 613 */ 614 return (TRUE); 615 } 616 617 /* 618 * Get a serializing token. This routine can block. 619 */ 620 void 621 lwkt_gettoken(lwkt_token_t tok) 622 { 623 thread_t td = curthread; 624 lwkt_tokref_t ref; 625 626 ref = td->td_toks_stop; 627 KKASSERT(ref < &td->td_toks_end); 628 ++td->td_toks_stop; 629 cpu_ccfence(); 630 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 631 632 #ifdef DEBUG_LOCKS 633 /* 634 * Taking an exclusive token after holding it shared will 635 * livelock. Scan for that case and assert. 636 */ 637 lwkt_tokref_t tk; 638 int found = 0; 639 for (tk = &td->td_toks_base; tk < ref; tk++) { 640 if (tk->tr_tok != tok) 641 continue; 642 643 found++; 644 if (tk->tr_count & TOK_EXCLUSIVE) 645 goto good; 646 } 647 /* We found only shared instances of this token if found >0 here */ 648 KASSERT((found == 0), ("Token %p s/x livelock", tok)); 649 good: 650 #endif 651 652 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ)) 653 return; 654 655 /* 656 * Give up running if we can't acquire the token right now. 657 * 658 * Since the tokref is already active the scheduler now 659 * takes care of acquisition, so we need only call 660 * lwkt_switch(). 661 * 662 * Since we failed this was not a recursive token so upon 663 * return tr_tok->t_ref should be assigned to this specific 664 * ref. 665 */ 666 td->td_wmesg = tok->t_desc; 667 ++tok->t_collisions; 668 logtoken(fail, ref); 669 td->td_toks_have = td->td_toks_stop - 1; 670 671 if (tokens_debug_output > 0) { 672 --tokens_debug_output; 673 spin_lock(&tok_debug_spin); 674 kprintf("Excl Token thread %p %s %s\n", 675 td, tok->t_desc, td->td_comm); 676 print_backtrace(6); 677 kprintf("\n"); 678 spin_unlock(&tok_debug_spin); 679 } 680 681 lwkt_switch(); 682 logtoken(succ, ref); 683 KKASSERT(tok->t_ref == ref); 684 } 685 686 /* 687 * Similar to gettoken but we acquire a shared token instead of an exclusive 688 * token. 689 */ 690 void 691 lwkt_gettoken_shared(lwkt_token_t tok) 692 { 693 thread_t td = curthread; 694 lwkt_tokref_t ref; 695 696 ref = td->td_toks_stop; 697 KKASSERT(ref < &td->td_toks_end); 698 ++td->td_toks_stop; 699 cpu_ccfence(); 700 _lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ); 701 702 #ifdef DEBUG_LOCKS 703 /* 704 * Taking a pool token in shared mode is a bad idea; other 705 * addresses deeper in the call stack may hash to the same pool 706 * token and you may end up with an exclusive-shared livelock. 707 * Warn in this condition. 708 */ 709 if ((tok >= &pool_tokens[0].token) && 710 (tok < &pool_tokens[LWKT_NUM_POOL_TOKENS].token)) 711 kprintf("Warning! Taking pool token %p in shared mode\n", tok); 712 #endif 713 714 715 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ)) 716 return; 717 718 /* 719 * Give up running if we can't acquire the token right now. 720 * 721 * Since the tokref is already active the scheduler now 722 * takes care of acquisition, so we need only call 723 * lwkt_switch(). 724 * 725 * Since we failed this was not a recursive token so upon 726 * return tr_tok->t_ref should be assigned to this specific 727 * ref. 728 */ 729 td->td_wmesg = tok->t_desc; 730 ++tok->t_collisions; 731 logtoken(fail, ref); 732 td->td_toks_have = td->td_toks_stop - 1; 733 734 if (tokens_debug_output > 0) { 735 --tokens_debug_output; 736 spin_lock(&tok_debug_spin); 737 kprintf("Shar Token thread %p %s %s\n", 738 td, tok->t_desc, td->td_comm); 739 print_backtrace(6); 740 kprintf("\n"); 741 spin_unlock(&tok_debug_spin); 742 } 743 744 lwkt_switch(); 745 logtoken(succ, ref); 746 } 747 748 /* 749 * Attempt to acquire a token, return TRUE on success, FALSE on failure. 750 * 751 * We setup the tokref in case we actually get the token (if we switch later 752 * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without 753 * TOK_EXCLREQ in case we fail. 754 */ 755 int 756 lwkt_trytoken(lwkt_token_t tok) 757 { 758 thread_t td = curthread; 759 lwkt_tokref_t ref; 760 761 ref = td->td_toks_stop; 762 KKASSERT(ref < &td->td_toks_end); 763 ++td->td_toks_stop; 764 cpu_ccfence(); 765 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 766 767 if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE)) 768 return TRUE; 769 770 /* 771 * Failed, unpend the request 772 */ 773 cpu_ccfence(); 774 --td->td_toks_stop; 775 ++tok->t_collisions; 776 return FALSE; 777 } 778 779 780 void 781 lwkt_gettoken_hard(lwkt_token_t tok) 782 { 783 lwkt_gettoken(tok); 784 crit_enter_hard(); 785 } 786 787 lwkt_token_t 788 lwkt_getpooltoken(void *ptr) 789 { 790 lwkt_token_t tok; 791 792 tok = _lwkt_token_pool_lookup(ptr); 793 lwkt_gettoken(tok); 794 return (tok); 795 } 796 797 /* 798 * Release a serializing token. 799 * 800 * WARNING! All tokens must be released in reverse order. This will be 801 * asserted. 802 */ 803 void 804 lwkt_reltoken(lwkt_token_t tok) 805 { 806 thread_t td = curthread; 807 lwkt_tokref_t ref; 808 809 /* 810 * Remove ref from thread token list and assert that it matches 811 * the token passed in. Tokens must be released in reverse order. 812 */ 813 ref = td->td_toks_stop - 1; 814 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok); 815 _lwkt_reltokref(ref, td); 816 cpu_sfence(); 817 td->td_toks_stop = ref; 818 } 819 820 void 821 lwkt_reltoken_hard(lwkt_token_t tok) 822 { 823 lwkt_reltoken(tok); 824 crit_exit_hard(); 825 } 826 827 /* 828 * It is faster for users of lwkt_getpooltoken() to use the returned 829 * token and just call lwkt_reltoken(), but for convenience we provide 830 * this function which looks the token up based on the ident. 831 */ 832 void 833 lwkt_relpooltoken(void *ptr) 834 { 835 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr); 836 lwkt_reltoken(tok); 837 } 838 839 /* 840 * Return a count of the number of token refs the thread has to the 841 * specified token, whether it currently owns the token or not. 842 */ 843 int 844 lwkt_cnttoken(lwkt_token_t tok, thread_t td) 845 { 846 lwkt_tokref_t scan; 847 int count = 0; 848 849 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 850 if (scan->tr_tok == tok) 851 ++count; 852 } 853 return(count); 854 } 855 856 /* 857 * Pool tokens are used to provide a type-stable serializing token 858 * pointer that does not race against disappearing data structures. 859 * 860 * This routine is called in early boot just after we setup the BSP's 861 * globaldata structure. 862 */ 863 void 864 lwkt_token_pool_init(void) 865 { 866 int i; 867 868 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i) 869 lwkt_token_init(&pool_tokens[i].token, "pool"); 870 } 871 872 lwkt_token_t 873 lwkt_token_pool_lookup(void *ptr) 874 { 875 return (_lwkt_token_pool_lookup(ptr)); 876 } 877 878 /* 879 * Initialize a token. 880 */ 881 void 882 lwkt_token_init(lwkt_token_t tok, const char *desc) 883 { 884 tok->t_count = 0; 885 tok->t_ref = NULL; 886 tok->t_collisions = 0; 887 tok->t_desc = desc; 888 } 889 890 void 891 lwkt_token_uninit(lwkt_token_t tok) 892 { 893 /* empty */ 894 } 895 896 /* 897 * Exchange the two most recent tokens on the tokref stack. This allows 898 * you to release a token out of order. 899 * 900 * We have to be careful about the case where the top two tokens are 901 * the same token. In this case tok->t_ref will point to the deeper 902 * ref and must remain pointing to the deeper ref. If we were to swap 903 * it the first release would clear the token even though a second 904 * ref is still present. 905 * 906 * Only exclusively held tokens contain a reference to the tokref which 907 * has to be flipped along with the swap. 908 */ 909 void 910 lwkt_token_swap(void) 911 { 912 lwkt_tokref_t ref1, ref2; 913 lwkt_token_t tok1, tok2; 914 long count1, count2; 915 thread_t td = curthread; 916 917 crit_enter(); 918 919 ref1 = td->td_toks_stop - 1; 920 ref2 = td->td_toks_stop - 2; 921 KKASSERT(ref1 >= &td->td_toks_base); 922 KKASSERT(ref2 >= &td->td_toks_base); 923 924 tok1 = ref1->tr_tok; 925 tok2 = ref2->tr_tok; 926 count1 = ref1->tr_count; 927 count2 = ref2->tr_count; 928 929 if (tok1 != tok2) { 930 ref1->tr_tok = tok2; 931 ref1->tr_count = count2; 932 ref2->tr_tok = tok1; 933 ref2->tr_count = count1; 934 if (tok1->t_ref == ref1) 935 tok1->t_ref = ref2; 936 if (tok2->t_ref == ref2) 937 tok2->t_ref = ref1; 938 } 939 940 crit_exit(); 941 } 942 943 #ifdef DDB 944 DB_SHOW_COMMAND(tokens, db_tok_all) 945 { 946 struct lwkt_token *tok, **ptr; 947 struct lwkt_token *toklist[16] = { 948 &mp_token, 949 &pmap_token, 950 &dev_token, 951 &vm_token, 952 &vmspace_token, 953 &kvm_token, 954 &sigio_token, 955 &tty_token, 956 &vnode_token, 957 NULL 958 }; 959 960 ptr = toklist; 961 for (tok = *ptr; tok; tok = *(++ptr)) { 962 db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok, 963 (tok->t_ref ? tok->t_ref->tr_owner : NULL), 964 tok->t_collisions, tok->t_desc); 965 } 966 } 967 #endif /* DDB */ 968