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 /* XXX EXCLREQ should work */ 335 /* 336 * It is possible to get the token shared. 337 */ 338 if (atomic_cmpset_long(&tok->t_count, count, 339 count + TOK_INCR)) { 340 return TRUE; 341 } 342 /* retry */ 343 } else if ((count & TOK_EXCLUSIVE) && 344 oref >= &td->td_toks_base && 345 oref < td->td_toks_stop) { 346 /* 347 * We own the exclusive bit on the token so 348 * we can in fact also get it shared. 349 */ 350 atomic_add_long(&tok->t_count, TOK_INCR); 351 return TRUE; 352 } else { 353 /* 354 * We failed to get the token shared 355 */ 356 return FALSE; 357 } 358 /* retry */ 359 } 360 } 361 } 362 363 static __inline 364 int 365 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode) 366 { 367 int spin; 368 369 if (_lwkt_trytokref(ref, td, mode)) { 370 #ifdef DEBUG_LOCKS_LATENCY 371 long j; 372 for (j = tokens_add_latency; j > 0; --j) 373 cpu_ccfence(); 374 #endif 375 return TRUE; 376 } 377 for (spin = lwkt_token_spin; spin > 0; --spin) { 378 if (lwkt_token_delay) 379 tsc_delay(lwkt_token_delay); 380 else 381 cpu_pause(); 382 if (_lwkt_trytokref(ref, td, mode)) { 383 #ifdef DEBUG_LOCKS_LATENCY 384 long j; 385 for (j = tokens_add_latency; j > 0; --j) 386 cpu_ccfence(); 387 #endif 388 return TRUE; 389 } 390 } 391 return FALSE; 392 } 393 394 /* 395 * Release a token that we hold. 396 */ 397 static __inline 398 void 399 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td) 400 { 401 lwkt_token_t tok; 402 long count; 403 404 tok = ref->tr_tok; 405 for (;;) { 406 count = tok->t_count; 407 cpu_ccfence(); 408 if (tok->t_ref == ref) { 409 /* 410 * We are an exclusive holder. We must clear tr_ref 411 * before we clear the TOK_EXCLUSIVE bit. If we are 412 * unable to clear the bit we must restore 413 * tok->t_ref. 414 */ 415 KKASSERT(count & TOK_EXCLUSIVE); 416 tok->t_ref = NULL; 417 if (atomic_cmpset_long(&tok->t_count, count, 418 count & ~TOK_EXCLUSIVE)) { 419 return; 420 } 421 tok->t_ref = ref; 422 /* retry */ 423 } else { 424 /* 425 * We are a shared holder 426 */ 427 KKASSERT(count & TOK_COUNTMASK); 428 if (atomic_cmpset_long(&tok->t_count, count, 429 count - TOK_INCR)) { 430 return; 431 } 432 /* retry */ 433 } 434 /* retry */ 435 } 436 } 437 438 /* 439 * Obtain all the tokens required by the specified thread on the current 440 * cpu, return 0 on failure and non-zero on success. If a failure occurs 441 * any partially acquired tokens will be released prior to return. 442 * 443 * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all 444 * tokens that the thread had to release when it switched away. 445 * 446 * If spinning is non-zero this function acquires the tokens in a particular 447 * order to deal with potential deadlocks. We simply use address order for 448 * the case. 449 * 450 * Called from a critical section. 451 */ 452 int 453 lwkt_getalltokens(thread_t td, int spinning) 454 { 455 lwkt_tokref_t scan; 456 lwkt_token_t tok; 457 458 if (spinning) 459 return(_lwkt_getalltokens_sorted(td)); 460 461 /* 462 * Acquire tokens in forward order, assign or validate tok->t_ref. 463 */ 464 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 465 tok = scan->tr_tok; 466 for (;;) { 467 /* 468 * Only try really hard on the last token 469 */ 470 if (scan == td->td_toks_stop - 1) { 471 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 472 break; 473 } else { 474 if (_lwkt_trytokref(scan, td, scan->tr_count)) 475 break; 476 } 477 478 /* 479 * Otherwise we failed to acquire all the tokens. 480 * Release whatever we did get. 481 */ 482 KASSERT(tok->t_desc, 483 ("token %p is not initialized", tok)); 484 strncpy(td->td_gd->gd_cnt.v_lock_name, 485 tok->t_desc, 486 sizeof(td->td_gd->gd_cnt.v_lock_name) - 1); 487 488 if (lwkt_sched_debug > 0) { 489 --lwkt_sched_debug; 490 kprintf("toka %p %s %s\n", 491 tok, tok->t_desc, td->td_comm); 492 } 493 td->td_wmesg = tok->t_desc; 494 ++tok->t_collisions; 495 while (--scan >= &td->td_toks_base) 496 _lwkt_reltokref(scan, td); 497 return(FALSE); 498 } 499 } 500 return (TRUE); 501 } 502 503 /* 504 * Release all tokens owned by the specified thread on the current cpu. 505 * 506 * This code is really simple. Even in cases where we own all the tokens 507 * note that t_ref may not match the scan for recursively held tokens which 508 * are held deeper in the stack, or for the case where a lwkt_getalltokens() 509 * failed. 510 * 511 * Tokens are released in reverse order to reduce chasing race failures. 512 * 513 * Called from a critical section. 514 */ 515 void 516 lwkt_relalltokens(thread_t td) 517 { 518 lwkt_tokref_t scan; 519 520 /* 521 * Weird order is to try to avoid a panic loop 522 */ 523 if (td->td_toks_have) { 524 scan = td->td_toks_have; 525 td->td_toks_have = NULL; 526 } else { 527 scan = td->td_toks_stop; 528 } 529 while (--scan >= &td->td_toks_base) 530 _lwkt_reltokref(scan, td); 531 } 532 533 /* 534 * This is the decontention version of lwkt_getalltokens(). The tokens are 535 * acquired in address-sorted order to deal with any deadlocks. Ultimately 536 * token failures will spin into the scheduler and get here. 537 * 538 * Called from critical section 539 */ 540 static 541 int 542 _lwkt_getalltokens_sorted(thread_t td) 543 { 544 lwkt_tokref_t sort_array[LWKT_MAXTOKENS]; 545 lwkt_tokref_t scan; 546 lwkt_token_t tok; 547 int i; 548 int j; 549 int n; 550 551 /* 552 * Sort the token array. Yah yah, I know this isn't fun. 553 * 554 * NOTE: Recursively acquired tokens are ordered the same as in the 555 * td_toks_array so we can always get the earliest one first. 556 */ 557 i = 0; 558 scan = &td->td_toks_base; 559 while (scan < td->td_toks_stop) { 560 for (j = 0; j < i; ++j) { 561 if (scan->tr_tok < sort_array[j]->tr_tok) 562 break; 563 } 564 if (j != i) { 565 bcopy(sort_array + j, sort_array + j + 1, 566 (i - j) * sizeof(lwkt_tokref_t)); 567 } 568 sort_array[j] = scan; 569 ++scan; 570 ++i; 571 } 572 n = i; 573 574 /* 575 * Acquire tokens in forward order, assign or validate tok->t_ref. 576 */ 577 for (i = 0; i < n; ++i) { 578 scan = sort_array[i]; 579 tok = scan->tr_tok; 580 for (;;) { 581 /* 582 * Only try really hard on the last token 583 */ 584 if (scan == td->td_toks_stop - 1) { 585 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 586 break; 587 } else { 588 if (_lwkt_trytokref(scan, td, scan->tr_count)) 589 break; 590 } 591 592 /* 593 * Otherwise we failed to acquire all the tokens. 594 * Release whatever we did get. 595 */ 596 if (lwkt_sched_debug > 0) { 597 --lwkt_sched_debug; 598 kprintf("tokb %p %s %s\n", 599 tok, tok->t_desc, td->td_comm); 600 } 601 td->td_wmesg = tok->t_desc; 602 ++tok->t_collisions; 603 while (--i >= 0) { 604 scan = sort_array[i]; 605 _lwkt_reltokref(scan, td); 606 } 607 return(FALSE); 608 } 609 } 610 611 /* 612 * We were successful, there is no need for another core to signal 613 * us. 614 */ 615 return (TRUE); 616 } 617 618 /* 619 * Get a serializing token. This routine can block. 620 */ 621 void 622 lwkt_gettoken(lwkt_token_t tok) 623 { 624 thread_t td = curthread; 625 lwkt_tokref_t ref; 626 627 ref = td->td_toks_stop; 628 KKASSERT(ref < &td->td_toks_end); 629 ++td->td_toks_stop; 630 cpu_ccfence(); 631 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 632 633 #ifdef DEBUG_LOCKS 634 /* 635 * Taking an exclusive token after holding it shared will 636 * livelock. Scan for that case and assert. 637 */ 638 lwkt_tokref_t tk; 639 int found = 0; 640 for (tk = &td->td_toks_base; tk < ref; tk++) { 641 if (tk->tr_tok != tok) 642 continue; 643 644 found++; 645 if (tk->tr_count & TOK_EXCLUSIVE) 646 goto good; 647 } 648 /* We found only shared instances of this token if found >0 here */ 649 KASSERT((found == 0), ("Token %p s/x livelock", tok)); 650 good: 651 #endif 652 653 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ)) 654 return; 655 656 /* 657 * Give up running if we can't acquire the token right now. 658 * 659 * Since the tokref is already active the scheduler now 660 * takes care of acquisition, so we need only call 661 * lwkt_switch(). 662 * 663 * Since we failed this was not a recursive token so upon 664 * return tr_tok->t_ref should be assigned to this specific 665 * ref. 666 */ 667 td->td_wmesg = tok->t_desc; 668 ++tok->t_collisions; 669 logtoken(fail, ref); 670 td->td_toks_have = td->td_toks_stop - 1; 671 672 if (tokens_debug_output > 0) { 673 --tokens_debug_output; 674 spin_lock(&tok_debug_spin); 675 kprintf("Excl Token thread %p %s %s\n", 676 td, tok->t_desc, td->td_comm); 677 print_backtrace(6); 678 kprintf("\n"); 679 spin_unlock(&tok_debug_spin); 680 } 681 682 lwkt_switch(); 683 logtoken(succ, ref); 684 KKASSERT(tok->t_ref == ref); 685 } 686 687 /* 688 * Similar to gettoken but we acquire a shared token instead of an exclusive 689 * token. 690 */ 691 void 692 lwkt_gettoken_shared(lwkt_token_t tok) 693 { 694 thread_t td = curthread; 695 lwkt_tokref_t ref; 696 697 ref = td->td_toks_stop; 698 KKASSERT(ref < &td->td_toks_end); 699 ++td->td_toks_stop; 700 cpu_ccfence(); 701 _lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ); 702 703 #ifdef DEBUG_LOCKS 704 /* 705 * Taking a pool token in shared mode is a bad idea; other 706 * addresses deeper in the call stack may hash to the same pool 707 * token and you may end up with an exclusive-shared livelock. 708 * Warn in this condition. 709 */ 710 if ((tok >= &pool_tokens[0].token) && 711 (tok < &pool_tokens[LWKT_NUM_POOL_TOKENS].token)) 712 kprintf("Warning! Taking pool token %p in shared mode\n", tok); 713 #endif 714 715 716 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ)) 717 return; 718 719 /* 720 * Give up running if we can't acquire the token right now. 721 * 722 * Since the tokref is already active the scheduler now 723 * takes care of acquisition, so we need only call 724 * lwkt_switch(). 725 * 726 * Since we failed this was not a recursive token so upon 727 * return tr_tok->t_ref should be assigned to this specific 728 * ref. 729 */ 730 td->td_wmesg = tok->t_desc; 731 ++tok->t_collisions; 732 logtoken(fail, ref); 733 td->td_toks_have = td->td_toks_stop - 1; 734 735 if (tokens_debug_output > 0) { 736 --tokens_debug_output; 737 spin_lock(&tok_debug_spin); 738 kprintf("Shar Token thread %p %s %s\n", 739 td, tok->t_desc, td->td_comm); 740 print_backtrace(6); 741 kprintf("\n"); 742 spin_unlock(&tok_debug_spin); 743 } 744 745 lwkt_switch(); 746 logtoken(succ, ref); 747 } 748 749 /* 750 * Attempt to acquire a token, return TRUE on success, FALSE on failure. 751 * 752 * We setup the tokref in case we actually get the token (if we switch later 753 * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without 754 * TOK_EXCLREQ in case we fail. 755 */ 756 int 757 lwkt_trytoken(lwkt_token_t tok) 758 { 759 thread_t td = curthread; 760 lwkt_tokref_t ref; 761 762 ref = td->td_toks_stop; 763 KKASSERT(ref < &td->td_toks_end); 764 ++td->td_toks_stop; 765 cpu_ccfence(); 766 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 767 768 if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE)) 769 return TRUE; 770 771 /* 772 * Failed, unpend the request 773 */ 774 cpu_ccfence(); 775 --td->td_toks_stop; 776 ++tok->t_collisions; 777 return FALSE; 778 } 779 780 781 void 782 lwkt_gettoken_hard(lwkt_token_t tok) 783 { 784 lwkt_gettoken(tok); 785 crit_enter_hard(); 786 } 787 788 lwkt_token_t 789 lwkt_getpooltoken(void *ptr) 790 { 791 lwkt_token_t tok; 792 793 tok = _lwkt_token_pool_lookup(ptr); 794 lwkt_gettoken(tok); 795 return (tok); 796 } 797 798 /* 799 * Release a serializing token. 800 * 801 * WARNING! All tokens must be released in reverse order. This will be 802 * asserted. 803 */ 804 void 805 lwkt_reltoken(lwkt_token_t tok) 806 { 807 thread_t td = curthread; 808 lwkt_tokref_t ref; 809 810 /* 811 * Remove ref from thread token list and assert that it matches 812 * the token passed in. Tokens must be released in reverse order. 813 */ 814 ref = td->td_toks_stop - 1; 815 KKASSERT(ref >= &td->td_toks_base && ref->tr_tok == tok); 816 _lwkt_reltokref(ref, td); 817 cpu_sfence(); 818 td->td_toks_stop = ref; 819 } 820 821 void 822 lwkt_reltoken_hard(lwkt_token_t tok) 823 { 824 lwkt_reltoken(tok); 825 crit_exit_hard(); 826 } 827 828 /* 829 * It is faster for users of lwkt_getpooltoken() to use the returned 830 * token and just call lwkt_reltoken(), but for convenience we provide 831 * this function which looks the token up based on the ident. 832 */ 833 void 834 lwkt_relpooltoken(void *ptr) 835 { 836 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr); 837 lwkt_reltoken(tok); 838 } 839 840 /* 841 * Return a count of the number of token refs the thread has to the 842 * specified token, whether it currently owns the token or not. 843 */ 844 int 845 lwkt_cnttoken(lwkt_token_t tok, thread_t td) 846 { 847 lwkt_tokref_t scan; 848 int count = 0; 849 850 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 851 if (scan->tr_tok == tok) 852 ++count; 853 } 854 return(count); 855 } 856 857 /* 858 * Pool tokens are used to provide a type-stable serializing token 859 * pointer that does not race against disappearing data structures. 860 * 861 * This routine is called in early boot just after we setup the BSP's 862 * globaldata structure. 863 */ 864 void 865 lwkt_token_pool_init(void) 866 { 867 int i; 868 869 for (i = 0; i < LWKT_NUM_POOL_TOKENS; ++i) 870 lwkt_token_init(&pool_tokens[i].token, "pool"); 871 } 872 873 lwkt_token_t 874 lwkt_token_pool_lookup(void *ptr) 875 { 876 return (_lwkt_token_pool_lookup(ptr)); 877 } 878 879 /* 880 * Initialize a token. 881 */ 882 void 883 lwkt_token_init(lwkt_token_t tok, const char *desc) 884 { 885 tok->t_count = 0; 886 tok->t_ref = NULL; 887 tok->t_collisions = 0; 888 tok->t_desc = desc; 889 } 890 891 void 892 lwkt_token_uninit(lwkt_token_t tok) 893 { 894 /* empty */ 895 } 896 897 /* 898 * Exchange the two most recent tokens on the tokref stack. This allows 899 * you to release a token out of order. 900 * 901 * We have to be careful about the case where the top two tokens are 902 * the same token. In this case tok->t_ref will point to the deeper 903 * ref and must remain pointing to the deeper ref. If we were to swap 904 * it the first release would clear the token even though a second 905 * ref is still present. 906 * 907 * Only exclusively held tokens contain a reference to the tokref which 908 * has to be flipped along with the swap. 909 */ 910 void 911 lwkt_token_swap(void) 912 { 913 lwkt_tokref_t ref1, ref2; 914 lwkt_token_t tok1, tok2; 915 long count1, count2; 916 thread_t td = curthread; 917 918 crit_enter(); 919 920 ref1 = td->td_toks_stop - 1; 921 ref2 = td->td_toks_stop - 2; 922 KKASSERT(ref1 >= &td->td_toks_base); 923 KKASSERT(ref2 >= &td->td_toks_base); 924 925 tok1 = ref1->tr_tok; 926 tok2 = ref2->tr_tok; 927 count1 = ref1->tr_count; 928 count2 = ref2->tr_count; 929 930 if (tok1 != tok2) { 931 ref1->tr_tok = tok2; 932 ref1->tr_count = count2; 933 ref2->tr_tok = tok1; 934 ref2->tr_count = count1; 935 if (tok1->t_ref == ref1) 936 tok1->t_ref = ref2; 937 if (tok2->t_ref == ref2) 938 tok2->t_ref = ref1; 939 } 940 941 crit_exit(); 942 } 943 944 #ifdef DDB 945 DB_SHOW_COMMAND(tokens, db_tok_all) 946 { 947 struct lwkt_token *tok, **ptr; 948 struct lwkt_token *toklist[16] = { 949 &mp_token, 950 &pmap_token, 951 &dev_token, 952 &vm_token, 953 &vmspace_token, 954 &kvm_token, 955 &sigio_token, 956 &tty_token, 957 &vnode_token, 958 NULL 959 }; 960 961 ptr = toklist; 962 for (tok = *ptr; tok; tok = *(++ptr)) { 963 db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok, 964 (tok->t_ref ? tok->t_ref->tr_owner : NULL), 965 tok->t_collisions, tok->t_desc); 966 } 967 } 968 #endif /* DDB */ 969