1 /* 2 * Copyright (c) 2003-2006,2009-2019 The DragonFly Project. 3 * All rights reserved. 4 * 5 * This code is derived from software contributed to The DragonFly Project 6 * by Matthew Dillon <dillon@backplane.com> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in 16 * the documentation and/or other materials provided with the 17 * distribution. 18 * 3. Neither the name of The DragonFly Project nor the names of its 19 * contributors may be used to endorse or promote products derived 20 * from this software without specific, prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 33 * SUCH DAMAGE. 34 */ 35 36 /* 37 * lwkt_token - Implement soft token locks. 38 * 39 * Tokens are locks which serialize a thread only while the thread is 40 * running. If the thread blocks all tokens are released, then reacquired 41 * when the thread resumes. 42 * 43 * This implementation requires no critical sections or spin locks, but 44 * does use atomic_cmpset_ptr(). 45 * 46 * Tokens may be recursively acquired by the same thread. However the 47 * caller must be sure to release such tokens in reverse order. 48 */ 49 #include <sys/param.h> 50 #include <sys/systm.h> 51 #include <sys/kernel.h> 52 #include <sys/proc.h> 53 #include <sys/rtprio.h> 54 #include <sys/queue.h> 55 #include <sys/sysctl.h> 56 #include <sys/ktr.h> 57 #include <sys/kthread.h> 58 #include <machine/cpu.h> 59 #include <sys/lock.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 #include "opt_ddb.h" 80 #ifdef DDB 81 #include <ddb/ddb.h> 82 #endif 83 84 extern int lwkt_sched_debug; 85 86 #define LWKT_POOL_TOKENS 16384 /* must be power of 2 */ 87 #define LWKT_POOL_MASK (LWKT_POOL_TOKENS - 1) 88 89 struct lwkt_pool_token { 90 struct lwkt_token token; 91 } __cachealign; 92 93 static struct lwkt_pool_token pool_tokens[LWKT_POOL_TOKENS]; 94 static struct spinlock tok_debug_spin = 95 SPINLOCK_INITIALIZER(&tok_debug_spin, "tok_debug_spin"); 96 97 #define TOKEN_STRING "REF=%p TOK=%p TD=%p" 98 #define TOKEN_ARGS lwkt_tokref_t ref, lwkt_token_t tok, struct thread *td 99 #define CONTENDED_STRING TOKEN_STRING " (contention started)" 100 #define UNCONTENDED_STRING TOKEN_STRING " (contention stopped)" 101 #if !defined(KTR_TOKENS) 102 #define KTR_TOKENS KTR_ALL 103 #endif 104 105 KTR_INFO_MASTER(tokens); 106 KTR_INFO(KTR_TOKENS, tokens, fail, 0, TOKEN_STRING, TOKEN_ARGS); 107 KTR_INFO(KTR_TOKENS, tokens, succ, 1, TOKEN_STRING, TOKEN_ARGS); 108 #if 0 109 KTR_INFO(KTR_TOKENS, tokens, release, 2, TOKEN_STRING, TOKEN_ARGS); 110 KTR_INFO(KTR_TOKENS, tokens, remote, 3, TOKEN_STRING, TOKEN_ARGS); 111 KTR_INFO(KTR_TOKENS, tokens, reqremote, 4, TOKEN_STRING, TOKEN_ARGS); 112 KTR_INFO(KTR_TOKENS, tokens, reqfail, 5, TOKEN_STRING, TOKEN_ARGS); 113 KTR_INFO(KTR_TOKENS, tokens, drain, 6, TOKEN_STRING, TOKEN_ARGS); 114 KTR_INFO(KTR_TOKENS, tokens, contention_start, 7, CONTENDED_STRING, TOKEN_ARGS); 115 KTR_INFO(KTR_TOKENS, tokens, contention_stop, 7, UNCONTENDED_STRING, TOKEN_ARGS); 116 #endif 117 118 #define logtoken(name, ref) \ 119 KTR_LOG(tokens_ ## name, ref, ref->tr_tok, curthread) 120 121 /* 122 * Global tokens. These replace the MP lock for major subsystem locking. 123 * These tokens are initially used to lockup both global and individual 124 * operations. 125 * 126 * Once individual structures get their own locks these tokens are used 127 * only to protect global lists & other variables and to interlock 128 * allocations and teardowns and such. 129 * 130 * The UP initializer causes token acquisition to also acquire the MP lock 131 * for maximum compatibility. The feature may be enabled and disabled at 132 * any time, the MP state is copied to the tokref when the token is acquired 133 * and will not race against sysctl changes. 134 */ 135 struct lwkt_token mp_token = LWKT_TOKEN_INITIALIZER(mp_token); 136 struct lwkt_token pmap_token = LWKT_TOKEN_INITIALIZER(pmap_token); 137 struct lwkt_token dev_token = LWKT_TOKEN_INITIALIZER(dev_token); 138 struct lwkt_token vm_token = LWKT_TOKEN_INITIALIZER(vm_token); 139 struct lwkt_token vmspace_token = LWKT_TOKEN_INITIALIZER(vmspace_token); 140 struct lwkt_token kvm_token = LWKT_TOKEN_INITIALIZER(kvm_token); 141 struct lwkt_token sigio_token = LWKT_TOKEN_INITIALIZER(sigio_token); 142 struct lwkt_token tty_token = LWKT_TOKEN_INITIALIZER(tty_token); 143 struct lwkt_token vnode_token = LWKT_TOKEN_INITIALIZER(vnode_token); 144 struct lwkt_token vga_token = LWKT_TOKEN_INITIALIZER(vga_token); 145 struct lwkt_token kbd_token = LWKT_TOKEN_INITIALIZER(kbd_token); 146 147 /* 148 * Exponential backoff (exclusive tokens) and TSC windowing (shared tokens) 149 * parameters. Remember that tokens backoff to the scheduler. This is a bit 150 * of trade-off. Smaller values like 128 work better in some situations, 151 * but under extreme loads larger values like 4096 seem to provide the most 152 * determinism. 153 */ 154 static int token_backoff_max __cachealign = 4096; 155 SYSCTL_INT(_lwkt, OID_AUTO, token_backoff_max, CTLFLAG_RW, 156 &token_backoff_max, 0, "Tokens exponential backoff"); 157 static int token_window_shift __cachealign = 8; 158 SYSCTL_INT(_lwkt, OID_AUTO, token_window_shift, CTLFLAG_RW, 159 &token_window_shift, 0, "Tokens TSC windowing shift"); 160 161 /* 162 * The collision count is bumped every time the LWKT scheduler fails 163 * to acquire needed tokens in addition to a normal lwkt_gettoken() 164 * stall. 165 */ 166 SYSCTL_LONG(_lwkt, OID_AUTO, mp_collisions, CTLFLAG_RW, 167 &mp_token.t_collisions, 0, "Collision counter of mp_token"); 168 SYSCTL_LONG(_lwkt, OID_AUTO, pmap_collisions, CTLFLAG_RW, 169 &pmap_token.t_collisions, 0, "Collision counter of pmap_token"); 170 SYSCTL_LONG(_lwkt, OID_AUTO, dev_collisions, CTLFLAG_RW, 171 &dev_token.t_collisions, 0, "Collision counter of dev_token"); 172 SYSCTL_LONG(_lwkt, OID_AUTO, vm_collisions, CTLFLAG_RW, 173 &vm_token.t_collisions, 0, "Collision counter of vm_token"); 174 SYSCTL_LONG(_lwkt, OID_AUTO, vmspace_collisions, CTLFLAG_RW, 175 &vmspace_token.t_collisions, 0, "Collision counter of vmspace_token"); 176 SYSCTL_LONG(_lwkt, OID_AUTO, kvm_collisions, CTLFLAG_RW, 177 &kvm_token.t_collisions, 0, "Collision counter of kvm_token"); 178 SYSCTL_LONG(_lwkt, OID_AUTO, sigio_collisions, CTLFLAG_RW, 179 &sigio_token.t_collisions, 0, "Collision counter of sigio_token"); 180 SYSCTL_LONG(_lwkt, OID_AUTO, tty_collisions, CTLFLAG_RW, 181 &tty_token.t_collisions, 0, "Collision counter of tty_token"); 182 SYSCTL_LONG(_lwkt, OID_AUTO, vnode_collisions, CTLFLAG_RW, 183 &vnode_token.t_collisions, 0, "Collision counter of vnode_token"); 184 185 static int tokens_debug_output; 186 SYSCTL_INT(_lwkt, OID_AUTO, tokens_debug_output, CTLFLAG_RW, 187 &tokens_debug_output, 0, "Generate stack trace N times"); 188 189 static int _lwkt_getalltokens_sorted(thread_t td); 190 191 /* 192 * Acquire the initial mplock 193 * 194 * (low level boot only) 195 */ 196 void 197 cpu_get_initial_mplock(void) 198 { 199 KKASSERT(mp_token.t_ref == NULL); 200 if (lwkt_trytoken(&mp_token) == FALSE) 201 panic("cpu_get_initial_mplock"); 202 } 203 204 /* 205 * Return a pool token given an address. Use a prime number to reduce 206 * overlaps. 207 */ 208 #define POOL_HASH_PRIME1 66555444443333333ULL 209 #define POOL_HASH_PRIME2 989042931893ULL 210 211 static __inline 212 lwkt_token_t 213 _lwkt_token_pool_lookup(void *ptr) 214 { 215 uintptr_t hash1; 216 uintptr_t hash2; 217 218 hash1 = (uintptr_t)ptr + ((uintptr_t)ptr >> 18); 219 hash1 %= POOL_HASH_PRIME1; 220 hash2 = ((uintptr_t)ptr >> 8) + ((uintptr_t)ptr >> 24); 221 hash2 %= POOL_HASH_PRIME2; 222 return (&pool_tokens[(hash1 ^ hash2) & LWKT_POOL_MASK].token); 223 } 224 225 /* 226 * Initialize a tokref_t prior to making it visible in the thread's 227 * token array. 228 */ 229 static __inline 230 void 231 _lwkt_tokref_init(lwkt_tokref_t ref, lwkt_token_t tok, thread_t td, long excl) 232 { 233 ref->tr_tok = tok; 234 ref->tr_count = excl; 235 ref->tr_owner = td; 236 } 237 238 /* 239 * Attempt to acquire a shared or exclusive token. Returns TRUE on success, 240 * FALSE on failure. 241 * 242 * If TOK_EXCLUSIVE is set in mode we are attempting to get an exclusive 243 * token, otherwise are attempting to get a shared token. 244 * 245 * If TOK_EXCLREQ is set in mode this is a blocking operation, otherwise 246 * it is a non-blocking operation (for both exclusive or shared acquisions). 247 */ 248 static __inline 249 int 250 _lwkt_trytokref(lwkt_tokref_t ref, thread_t td, long mode) 251 { 252 lwkt_token_t tok; 253 lwkt_tokref_t oref; 254 long count; 255 256 tok = ref->tr_tok; 257 KASSERT(((mode & TOK_EXCLREQ) == 0 || /* non blocking */ 258 td->td_gd->gd_intr_nesting_level == 0 || 259 panic_cpu_gd == mycpu), 260 ("Attempt to acquire token %p not already " 261 "held in hard code section", tok)); 262 263 if (mode & TOK_EXCLUSIVE) { 264 /* 265 * Attempt to get an exclusive token 266 */ 267 count = tok->t_count; 268 269 for (;;) { 270 oref = tok->t_ref; /* can be NULL */ 271 cpu_ccfence(); 272 if ((count & ~TOK_EXCLREQ) == 0) { 273 /* 274 * It is possible to get the exclusive bit. 275 * We must clear TOK_EXCLREQ on successful 276 * acquisition. 277 */ 278 if (atomic_fcmpset_long(&tok->t_count, &count, 279 (count & ~TOK_EXCLREQ) | 280 TOK_EXCLUSIVE)) { 281 KKASSERT(tok->t_ref == NULL); 282 tok->t_ref = ref; 283 return TRUE; 284 } 285 /* retry */ 286 } else if ((count & TOK_EXCLUSIVE) && 287 oref >= &td->td_toks_base && 288 oref < td->td_toks_stop) { 289 /* 290 * Our thread already holds the exclusive 291 * bit, we treat this tokref as a shared 292 * token (sorta) to make the token release 293 * code easier. Treating this as a shared 294 * token allows us to simply increment the 295 * count field. 296 * 297 * NOTE: oref cannot race above if it 298 * happens to be ours, so we're good. 299 * But we must still have a stable 300 * variable for both parts of the 301 * comparison. 302 * 303 * NOTE: Since we already have an exclusive 304 * lock and don't need to check EXCLREQ 305 * we can just use an atomic_add here 306 */ 307 atomic_add_long(&tok->t_count, TOK_INCR); 308 ref->tr_count &= ~TOK_EXCLUSIVE; 309 return TRUE; 310 } else if ((mode & TOK_EXCLREQ) && 311 (count & TOK_EXCLREQ) == 0) { 312 /* 313 * Unable to get the exclusive bit but being 314 * asked to set the exclusive-request bit. 315 * Since we are going to retry anyway just 316 * set the bit unconditionally. 317 */ 318 atomic_set_long(&tok->t_count, TOK_EXCLREQ); 319 return FALSE; 320 } else { 321 /* 322 * Unable to get the exclusive bit and not 323 * being asked to set the exclusive-request 324 * (aka lwkt_trytoken()), or EXCLREQ was 325 * already set. 326 */ 327 cpu_pause(); 328 return FALSE; 329 } 330 /* retry */ 331 } 332 } else { 333 /* 334 * Attempt to get a shared token. Note that TOK_EXCLREQ 335 * for shared tokens simply means the caller intends to 336 * block. We never actually set the bit in tok->t_count. 337 * 338 * Due to the token's no-deadlock guarantee, and complications 339 * created by the sorted reacquisition code, we can only 340 * give exclusive requests priority over shared requests 341 * in situations where the thread holds only one token. 342 */ 343 count = tok->t_count; 344 345 for (;;) { 346 oref = tok->t_ref; /* can be NULL */ 347 cpu_ccfence(); 348 if ((count & (TOK_EXCLUSIVE|mode)) == 0 || 349 ((count & TOK_EXCLUSIVE) == 0 && 350 td->td_toks_stop != &td->td_toks_base + 1) 351 ) { 352 /* 353 * It may be possible to get the token shared. 354 */ 355 if ((atomic_fetchadd_long(&tok->t_count, TOK_INCR) & TOK_EXCLUSIVE) == 0) { 356 return TRUE; 357 } 358 count = atomic_fetchadd_long(&tok->t_count, 359 -TOK_INCR); 360 count -= TOK_INCR; 361 /* retry */ 362 } else if ((count & TOK_EXCLUSIVE) && 363 oref >= &td->td_toks_base && 364 oref < td->td_toks_stop) { 365 /* 366 * We own the exclusive bit on the token so 367 * we can in fact also get it shared. 368 */ 369 atomic_add_long(&tok->t_count, TOK_INCR); 370 return TRUE; 371 } else { 372 /* 373 * We failed to get the token shared 374 */ 375 return FALSE; 376 } 377 /* retry */ 378 } 379 } 380 } 381 382 static __inline 383 int 384 _lwkt_trytokref_spin(lwkt_tokref_t ref, thread_t td, long mode) 385 { 386 if (_lwkt_trytokref(ref, td, mode)) 387 return TRUE; 388 389 if (mode & TOK_EXCLUSIVE) { 390 /* 391 * Contested exclusive token, use exponential backoff 392 * algorithm. 393 */ 394 long expbackoff; 395 long loop; 396 397 expbackoff = 0; 398 while (expbackoff < 6 + token_backoff_max) { 399 expbackoff = (expbackoff + 1) * 3 / 2; 400 if ((rdtsc() >> token_window_shift) % ncpus != mycpuid) { 401 for (loop = expbackoff; loop; --loop) 402 cpu_pause(); 403 } 404 if (_lwkt_trytokref(ref, td, mode)) 405 return TRUE; 406 } 407 } else { 408 /* 409 * Contested shared token, use TSC windowing. Note that 410 * exclusive tokens have priority over shared tokens only 411 * for the first token. 412 */ 413 if ((rdtsc() >> token_window_shift) % ncpus == mycpuid) { 414 if (_lwkt_trytokref(ref, td, mode & ~TOK_EXCLREQ)) 415 return TRUE; 416 } else { 417 if (_lwkt_trytokref(ref, td, mode)) 418 return TRUE; 419 } 420 421 } 422 ++mycpu->gd_cnt.v_lock_colls; 423 424 return FALSE; 425 } 426 427 /* 428 * Release a token that we hold. 429 * 430 * Since tokens are polled, we don't have to deal with wakeups and releasing 431 * is really easy. 432 */ 433 static __inline 434 void 435 _lwkt_reltokref(lwkt_tokref_t ref, thread_t td) 436 { 437 lwkt_token_t tok; 438 long count; 439 440 tok = ref->tr_tok; 441 if (tok->t_ref == ref) { 442 /* 443 * We are an exclusive holder. We must clear tr_ref 444 * before we clear the TOK_EXCLUSIVE bit. If we are 445 * unable to clear the bit we must restore 446 * tok->t_ref. 447 */ 448 #if 0 449 KKASSERT(count & TOK_EXCLUSIVE); 450 #endif 451 tok->t_ref = NULL; 452 atomic_clear_long(&tok->t_count, TOK_EXCLUSIVE); 453 } else { 454 /* 455 * We are a shared holder 456 */ 457 count = atomic_fetchadd_long(&tok->t_count, -TOK_INCR); 458 KKASSERT(count & TOK_COUNTMASK); /* count prior */ 459 } 460 } 461 462 /* 463 * Obtain all the tokens required by the specified thread on the current 464 * cpu, return 0 on failure and non-zero on success. If a failure occurs 465 * any partially acquired tokens will be released prior to return. 466 * 467 * lwkt_getalltokens is called by the LWKT scheduler to re-acquire all 468 * tokens that the thread had to release when it switched away. 469 * 470 * If spinning is non-zero this function acquires the tokens in a particular 471 * order to deal with potential deadlocks. We simply use address order for 472 * the case. 473 * 474 * Called from a critical section. 475 */ 476 int 477 lwkt_getalltokens(thread_t td, int spinning) 478 { 479 lwkt_tokref_t scan; 480 lwkt_token_t tok; 481 482 if (spinning) 483 return(_lwkt_getalltokens_sorted(td)); 484 485 /* 486 * Acquire tokens in forward order, assign or validate tok->t_ref. 487 */ 488 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 489 tok = scan->tr_tok; 490 for (;;) { 491 /* 492 * Only try really hard on the last token 493 */ 494 if (scan == td->td_toks_stop - 1) { 495 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 496 break; 497 } else { 498 if (_lwkt_trytokref(scan, td, scan->tr_count)) 499 break; 500 } 501 502 /* 503 * Otherwise we failed to acquire all the tokens. 504 * Release whatever we did get. 505 */ 506 KASSERT(tok->t_desc, 507 ("token %p is not initialized", tok)); 508 td->td_gd->gd_cnt.v_lock_name[0] = 't'; 509 strncpy(td->td_gd->gd_cnt.v_lock_name + 1, 510 tok->t_desc, 511 sizeof(td->td_gd->gd_cnt.v_lock_name) - 2); 512 if (lwkt_sched_debug > 0) { 513 --lwkt_sched_debug; 514 kprintf("toka %p %s %s\n", 515 tok, tok->t_desc, td->td_comm); 516 } 517 td->td_wmesg = tok->t_desc; 518 ++tok->t_collisions; 519 while (--scan >= &td->td_toks_base) 520 _lwkt_reltokref(scan, td); 521 return(FALSE); 522 } 523 } 524 return (TRUE); 525 } 526 527 /* 528 * Release all tokens owned by the specified thread on the current cpu. 529 * 530 * This code is really simple. Even in cases where we own all the tokens 531 * note that t_ref may not match the scan for recursively held tokens which 532 * are held deeper in the stack, or for the case where a lwkt_getalltokens() 533 * failed. 534 * 535 * Tokens are released in reverse order to reduce chasing race failures. 536 * 537 * Called from a critical section. 538 */ 539 void 540 lwkt_relalltokens(thread_t td) 541 { 542 lwkt_tokref_t scan; 543 544 /* 545 * Weird order is to try to avoid a panic loop 546 */ 547 if (td->td_toks_have) { 548 scan = td->td_toks_have; 549 td->td_toks_have = NULL; 550 } else { 551 scan = td->td_toks_stop; 552 } 553 while (--scan >= &td->td_toks_base) 554 _lwkt_reltokref(scan, td); 555 } 556 557 /* 558 * This is the decontention version of lwkt_getalltokens(). The tokens are 559 * acquired in address-sorted order to deal with any deadlocks. Ultimately 560 * token failures will spin into the scheduler and get here. 561 * 562 * Called from critical section 563 */ 564 static 565 int 566 _lwkt_getalltokens_sorted(thread_t td) 567 { 568 lwkt_tokref_t sort_array[LWKT_MAXTOKENS]; 569 lwkt_tokref_t scan; 570 lwkt_token_t tok; 571 int i; 572 int j; 573 int n; 574 575 /* 576 * Sort the token array. Yah yah, I know this isn't fun. 577 * 578 * NOTE: Recursively acquired tokens are ordered the same as in the 579 * td_toks_array so we can always get the earliest one first. 580 * This is particularly important when a token is acquired 581 * exclusively multiple times, as only the first acquisition 582 * is treated as an exclusive token. 583 */ 584 i = 0; 585 scan = &td->td_toks_base; 586 while (scan < td->td_toks_stop) { 587 for (j = 0; j < i; ++j) { 588 if (scan->tr_tok < sort_array[j]->tr_tok) 589 break; 590 } 591 if (j != i) { 592 bcopy(sort_array + j, sort_array + j + 1, 593 (i - j) * sizeof(lwkt_tokref_t)); 594 } 595 sort_array[j] = scan; 596 ++scan; 597 ++i; 598 } 599 n = i; 600 601 /* 602 * Acquire tokens in forward order, assign or validate tok->t_ref. 603 */ 604 for (i = 0; i < n; ++i) { 605 scan = sort_array[i]; 606 tok = scan->tr_tok; 607 for (;;) { 608 /* 609 * Only try really hard on the last token 610 */ 611 if (scan == td->td_toks_stop - 1) { 612 if (_lwkt_trytokref_spin(scan, td, scan->tr_count)) 613 break; 614 } else { 615 if (_lwkt_trytokref(scan, td, scan->tr_count)) 616 break; 617 } 618 619 /* 620 * Otherwise we failed to acquire all the tokens. 621 * Release whatever we did get. 622 */ 623 td->td_gd->gd_cnt.v_lock_name[0] = 't'; 624 strncpy(td->td_gd->gd_cnt.v_lock_name + 1, 625 tok->t_desc, 626 sizeof(td->td_gd->gd_cnt.v_lock_name) - 2); 627 if (lwkt_sched_debug > 0) { 628 --lwkt_sched_debug; 629 kprintf("tokb %p %s %s\n", 630 tok, tok->t_desc, td->td_comm); 631 } 632 td->td_wmesg = tok->t_desc; 633 ++tok->t_collisions; 634 while (--i >= 0) { 635 scan = sort_array[i]; 636 _lwkt_reltokref(scan, td); 637 } 638 return(FALSE); 639 } 640 } 641 642 /* 643 * We were successful, there is no need for another core to signal 644 * us. 645 */ 646 return (TRUE); 647 } 648 649 /* 650 * Get a serializing token. This routine can block. 651 */ 652 void 653 lwkt_gettoken(lwkt_token_t tok) 654 { 655 thread_t td = curthread; 656 lwkt_tokref_t ref; 657 658 ref = td->td_toks_stop; 659 KKASSERT(ref < &td->td_toks_end); 660 ++td->td_toks_stop; 661 cpu_ccfence(); 662 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 663 664 #ifdef DEBUG_LOCKS 665 /* 666 * Taking an exclusive token after holding it shared will 667 * livelock. Scan for that case and assert. 668 */ 669 lwkt_tokref_t tk; 670 int found = 0; 671 for (tk = &td->td_toks_base; tk < ref; tk++) { 672 if (tk->tr_tok != tok) 673 continue; 674 675 found++; 676 if (tk->tr_count & TOK_EXCLUSIVE) 677 goto good; 678 } 679 /* We found only shared instances of this token if found >0 here */ 680 KASSERT((found == 0), ("Token %p s/x livelock", tok)); 681 good: 682 #endif 683 684 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLUSIVE|TOK_EXCLREQ)) 685 return; 686 687 /* 688 * Give up running if we can't acquire the token right now. 689 * 690 * Since the tokref is already active the scheduler now 691 * takes care of acquisition, so we need only call 692 * lwkt_switch(). 693 * 694 * Since we failed this was not a recursive token so upon 695 * return tr_tok->t_ref should be assigned to this specific 696 * ref. 697 */ 698 td->td_wmesg = tok->t_desc; 699 ++tok->t_collisions; 700 logtoken(fail, ref); 701 td->td_toks_have = td->td_toks_stop - 1; 702 703 if (tokens_debug_output > 0) { 704 --tokens_debug_output; 705 spin_lock(&tok_debug_spin); 706 kprintf("Excl Token %p thread %p %s %s\n", 707 tok, td, tok->t_desc, td->td_comm); 708 print_backtrace(6); 709 kprintf("\n"); 710 spin_unlock(&tok_debug_spin); 711 } 712 713 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 714 lwkt_switch(); 715 logtoken(succ, ref); 716 KKASSERT(tok->t_ref == ref); 717 } 718 719 /* 720 * Similar to gettoken but we acquire a shared token instead of an exclusive 721 * token. 722 */ 723 void 724 lwkt_gettoken_shared(lwkt_token_t tok) 725 { 726 thread_t td = curthread; 727 lwkt_tokref_t ref; 728 729 ref = td->td_toks_stop; 730 KKASSERT(ref < &td->td_toks_end); 731 ++td->td_toks_stop; 732 cpu_ccfence(); 733 _lwkt_tokref_init(ref, tok, td, TOK_EXCLREQ); 734 735 #ifdef DEBUG_LOCKS 736 /* 737 * Taking a pool token in shared mode is a bad idea; other 738 * addresses deeper in the call stack may hash to the same pool 739 * token and you may end up with an exclusive-shared livelock. 740 * Warn in this condition. 741 */ 742 if ((tok >= &pool_tokens[0].token) && 743 (tok < &pool_tokens[LWKT_POOL_TOKENS].token)) 744 kprintf("Warning! Taking pool token %p in shared mode\n", tok); 745 #endif 746 747 748 if (_lwkt_trytokref_spin(ref, td, TOK_EXCLREQ)) 749 return; 750 751 /* 752 * Give up running if we can't acquire the token right now. 753 * 754 * Since the tokref is already active the scheduler now 755 * takes care of acquisition, so we need only call 756 * lwkt_switch(). 757 * 758 * Since we failed this was not a recursive token so upon 759 * return tr_tok->t_ref should be assigned to this specific 760 * ref. 761 */ 762 td->td_wmesg = tok->t_desc; 763 ++tok->t_collisions; 764 logtoken(fail, ref); 765 td->td_toks_have = td->td_toks_stop - 1; 766 767 if (tokens_debug_output > 0) { 768 --tokens_debug_output; 769 spin_lock(&tok_debug_spin); 770 kprintf("Shar Token %p thread %p %s %s\n", 771 tok, td, tok->t_desc, td->td_comm); 772 print_backtrace(6); 773 kprintf("\n"); 774 spin_unlock(&tok_debug_spin); 775 } 776 777 atomic_set_int(&td->td_mpflags, TDF_MP_DIDYIELD); 778 lwkt_switch(); 779 logtoken(succ, ref); 780 } 781 782 /* 783 * Attempt to acquire a token, return TRUE on success, FALSE on failure. 784 * 785 * We setup the tokref in case we actually get the token (if we switch later 786 * it becomes mandatory so we set TOK_EXCLREQ), but we call trytokref without 787 * TOK_EXCLREQ in case we fail. 788 */ 789 int 790 lwkt_trytoken(lwkt_token_t tok) 791 { 792 thread_t td = curthread; 793 lwkt_tokref_t ref; 794 795 ref = td->td_toks_stop; 796 KKASSERT(ref < &td->td_toks_end); 797 ++td->td_toks_stop; 798 cpu_ccfence(); 799 _lwkt_tokref_init(ref, tok, td, TOK_EXCLUSIVE|TOK_EXCLREQ); 800 801 if (_lwkt_trytokref(ref, td, TOK_EXCLUSIVE)) 802 return TRUE; 803 804 /* 805 * Failed, unpend the request 806 */ 807 cpu_ccfence(); 808 --td->td_toks_stop; 809 ++tok->t_collisions; 810 return FALSE; 811 } 812 813 lwkt_token_t 814 lwkt_getpooltoken(void *ptr) 815 { 816 lwkt_token_t tok; 817 818 tok = _lwkt_token_pool_lookup(ptr); 819 lwkt_gettoken(tok); 820 return (tok); 821 } 822 823 /* 824 * Release a serializing token. 825 * 826 * WARNING! All tokens must be released in reverse order. This will be 827 * asserted. 828 */ 829 void 830 lwkt_reltoken(lwkt_token_t tok) 831 { 832 thread_t td = curthread; 833 lwkt_tokref_t ref; 834 835 /* 836 * Remove ref from thread token list and assert that it matches 837 * the token passed in. Tokens must be released in reverse order. 838 */ 839 ref = td->td_toks_stop - 1; 840 if (__predict_false(ref < &td->td_toks_base || ref->tr_tok != tok)) { 841 kprintf("LWKT_RELTOKEN ASSERTION td %p tok %p ref %p/%p\n", 842 td, tok, &td->td_toks_base, ref); 843 kprintf("REF CONTENT: tok=%p count=%016lx owner=%p\n", 844 ref->tr_tok, ref->tr_count, ref->tr_owner); 845 if (ref < &td->td_toks_base) { 846 kprintf("lwkt_reltoken: no tokens to release\n"); 847 } else { 848 kprintf("lwkt_reltoken: release wants %s and got %s\n", 849 tok->t_desc, ref->tr_tok->t_desc); 850 } 851 panic("lwkt_reltoken: illegal release"); 852 } 853 _lwkt_reltokref(ref, td); 854 cpu_sfence(); 855 td->td_toks_stop = ref; 856 } 857 858 /* 859 * It is faster for users of lwkt_getpooltoken() to use the returned 860 * token and just call lwkt_reltoken(), but for convenience we provide 861 * this function which looks the token up based on the ident. 862 */ 863 void 864 lwkt_relpooltoken(void *ptr) 865 { 866 lwkt_token_t tok = _lwkt_token_pool_lookup(ptr); 867 lwkt_reltoken(tok); 868 } 869 870 /* 871 * Return a count of the number of token refs the thread has to the 872 * specified token, whether it currently owns the token or not. 873 */ 874 int 875 lwkt_cnttoken(lwkt_token_t tok, thread_t td) 876 { 877 lwkt_tokref_t scan; 878 int count = 0; 879 880 for (scan = &td->td_toks_base; scan < td->td_toks_stop; ++scan) { 881 if (scan->tr_tok == tok) 882 ++count; 883 } 884 return(count); 885 } 886 887 /* 888 * Pool tokens are used to provide a type-stable serializing token 889 * pointer that does not race against disappearing data structures. 890 * 891 * This routine is called in early boot just after we setup the BSP's 892 * globaldata structure. 893 */ 894 void 895 lwkt_token_pool_init(void) 896 { 897 int i; 898 899 for (i = 0; i < LWKT_POOL_TOKENS; ++i) 900 lwkt_token_init(&pool_tokens[i].token, "pool"); 901 } 902 903 lwkt_token_t 904 lwkt_token_pool_lookup(void *ptr) 905 { 906 return (_lwkt_token_pool_lookup(ptr)); 907 } 908 909 /* 910 * Initialize a token. 911 */ 912 void 913 lwkt_token_init(lwkt_token_t tok, const char *desc) 914 { 915 tok->t_count = 0; 916 tok->t_ref = NULL; 917 tok->t_collisions = 0; 918 tok->t_desc = desc; 919 } 920 921 void 922 lwkt_token_uninit(lwkt_token_t tok) 923 { 924 /* empty */ 925 } 926 927 /* 928 * Exchange the two most recent tokens on the tokref stack. This allows 929 * you to release a token out of order. 930 * 931 * We have to be careful about the case where the top two tokens are 932 * the same token. In this case tok->t_ref will point to the deeper 933 * ref and must remain pointing to the deeper ref. If we were to swap 934 * it the first release would clear the token even though a second 935 * ref is still present. 936 * 937 * Only exclusively held tokens contain a reference to the tokref which 938 * has to be flipped along with the swap. 939 */ 940 void 941 lwkt_token_swap(void) 942 { 943 lwkt_tokref_t ref1, ref2; 944 lwkt_token_t tok1, tok2; 945 long count1, count2; 946 thread_t td = curthread; 947 948 crit_enter(); 949 950 ref1 = td->td_toks_stop - 1; 951 ref2 = td->td_toks_stop - 2; 952 KKASSERT(ref1 >= &td->td_toks_base); 953 KKASSERT(ref2 >= &td->td_toks_base); 954 955 tok1 = ref1->tr_tok; 956 tok2 = ref2->tr_tok; 957 count1 = ref1->tr_count; 958 count2 = ref2->tr_count; 959 960 if (tok1 != tok2) { 961 ref1->tr_tok = tok2; 962 ref1->tr_count = count2; 963 ref2->tr_tok = tok1; 964 ref2->tr_count = count1; 965 if (tok1->t_ref == ref1) 966 tok1->t_ref = ref2; 967 if (tok2->t_ref == ref2) 968 tok2->t_ref = ref1; 969 } 970 971 crit_exit(); 972 } 973 974 #ifdef DDB 975 DB_SHOW_COMMAND(tokens, db_tok_all) 976 { 977 struct lwkt_token *tok, **ptr; 978 struct lwkt_token *toklist[16] = { 979 &mp_token, 980 &pmap_token, 981 &dev_token, 982 &vm_token, 983 &vmspace_token, 984 &kvm_token, 985 &sigio_token, 986 &tty_token, 987 &vnode_token, 988 NULL 989 }; 990 991 ptr = toklist; 992 for (tok = *ptr; tok; tok = *(++ptr)) { 993 db_printf("tok=%p tr_owner=%p t_colissions=%ld t_desc=%s\n", tok, 994 (tok->t_ref ? tok->t_ref->tr_owner : NULL), 995 tok->t_collisions, tok->t_desc); 996 } 997 } 998 #endif /* DDB */ 999