1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2007 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 /* 29 * Given several files containing CTF data, merge and uniquify that data into 30 * a single CTF section in an output file. 31 * 32 * Merges can proceed independently. As such, we perform the merges in parallel 33 * using a worker thread model. A given glob of CTF data (either all of the CTF 34 * data from a single input file, or the result of one or more merges) can only 35 * be involved in a single merge at any given time, so the process decreases in 36 * parallelism, especially towards the end, as more and more files are 37 * consolidated, finally resulting in a single merge of two large CTF graphs. 38 * Unfortunately, the last merge is also the slowest, as the two graphs being 39 * merged are each the product of merges of half of the input files. 40 * 41 * The algorithm consists of two phases, described in detail below. The first 42 * phase entails the merging of CTF data in groups of eight. The second phase 43 * takes the results of Phase I, and merges them two at a time. This disparity 44 * is due to an observation that the merge time increases at least quadratically 45 * with the size of the CTF data being merged. As such, merges of CTF graphs 46 * newly read from input files are much faster than merges of CTF graphs that 47 * are themselves the results of prior merges. 48 * 49 * A further complication is the need to ensure the repeatability of CTF merges. 50 * That is, a merge should produce the same output every time, given the same 51 * input. In both phases, this consistency requirement is met by imposing an 52 * ordering on the merge process, thus ensuring that a given set of input files 53 * are merged in the same order every time. 54 * 55 * Phase I 56 * 57 * The main thread reads the input files one by one, transforming the CTF 58 * data they contain into tdata structures. When a given file has been read 59 * and parsed, it is placed on the work queue for retrieval by worker threads. 60 * 61 * Central to Phase I is the Work In Progress (wip) array, which is used to 62 * merge batches of files in a predictable order. Files are read by the main 63 * thread, and are merged into wip array elements in round-robin order. When 64 * the number of files merged into a given array slot equals the batch size, 65 * the merged CTF graph in that array is added to the done slot in order by 66 * array slot. 67 * 68 * For example, consider a case where we have five input files, a batch size 69 * of two, a wip array size of two, and two worker threads (T1 and T2). 70 * 71 * 1. The wip array elements are assigned initial batch numbers 0 and 1. 72 * 2. T1 reads an input file from the input queue (wq_queue). This is the 73 * first input file, so it is placed into wip[0]. The second file is 74 * similarly read and placed into wip[1]. The wip array slots now contain 75 * one file each (wip_nmerged == 1). 76 * 3. T1 reads the third input file, which it merges into wip[0]. The 77 * number of files in wip[0] is equal to the batch size. 78 * 4. T2 reads the fourth input file, which it merges into wip[1]. wip[1] 79 * is now full too. 80 * 5. T2 attempts to place the contents of wip[1] on the done queue 81 * (wq_done_queue), but it can't, since the batch ID for wip[1] is 1. 82 * Batch 0 needs to be on the done queue before batch 1 can be added, so 83 * T2 blocks on wip[1]'s cv. 84 * 6. T1 attempts to place the contents of wip[0] on the done queue, and 85 * succeeds, updating wq_lastdonebatch to 0. It clears wip[0], and sets 86 * its batch ID to 2. T1 then signals wip[1]'s cv to awaken T2. 87 * 7. T2 wakes up, notices that wq_lastdonebatch is 0, which means that 88 * batch 1 can now be added. It adds wip[1] to the done queue, clears 89 * wip[1], and sets its batch ID to 3. It signals wip[0]'s cv, and 90 * restarts. 91 * 92 * The above process continues until all input files have been consumed. At 93 * this point, a pair of barriers are used to allow a single thread to move 94 * any partial batches from the wip array to the done array in batch ID order. 95 * When this is complete, wq_done_queue is moved to wq_queue, and Phase II 96 * begins. 97 * 98 * Locking Semantics (Phase I) 99 * 100 * The input queue (wq_queue) and the done queue (wq_done_queue) are 101 * protected by separate mutexes - wq_queue_lock and wq_done_queue. wip 102 * array slots are protected by their own mutexes, which must be grabbed 103 * before releasing the input queue lock. The wip array lock is dropped 104 * when the thread restarts the loop. If the array slot was full, the 105 * array lock will be held while the slot contents are added to the done 106 * queue. The done queue lock is used to protect the wip slot cv's. 107 * 108 * The pow number is protected by the queue lock. The master batch ID 109 * and last completed batch (wq_lastdonebatch) counters are protected *in 110 * Phase I* by the done queue lock. 111 * 112 * Phase II 113 * 114 * When Phase II begins, the queue consists of the merged batches from the 115 * first phase. Assume we have five batches: 116 * 117 * Q: a b c d e 118 * 119 * Using the same batch ID mechanism we used in Phase I, but without the wip 120 * array, worker threads remove two entries at a time from the beginning of 121 * the queue. These two entries are merged, and are added back to the tail 122 * of the queue, as follows: 123 * 124 * Q: a b c d e # start 125 * Q: c d e ab # a, b removed, merged, added to end 126 * Q: e ab cd # c, d removed, merged, added to end 127 * Q: cd eab # e, ab removed, merged, added to end 128 * Q: cdeab # cd, eab removed, merged, added to end 129 * 130 * When one entry remains on the queue, with no merges outstanding, Phase II 131 * finishes. We pre-determine the stopping point by pre-calculating the 132 * number of nodes that will appear on the list. In the example above, the 133 * number (wq_ninqueue) is 9. When ninqueue is 1, we conclude Phase II by 134 * signaling the main thread via wq_done_cv. 135 * 136 * Locking Semantics (Phase II) 137 * 138 * The queue (wq_queue), ninqueue, and the master batch ID and last 139 * completed batch counters are protected by wq_queue_lock. The done 140 * queue and corresponding lock are unused in Phase II as is the wip array. 141 * 142 * Uniquification 143 * 144 * We want the CTF data that goes into a given module to be as small as 145 * possible. For example, we don't want it to contain any type data that may 146 * be present in another common module. As such, after creating the master 147 * tdata_t for a given module, we can, if requested by the user, uniquify it 148 * against the tdata_t from another module (genunix in the case of the SunOS 149 * kernel). We perform a merge between the tdata_t for this module and the 150 * tdata_t from genunix. Nodes found in this module that are not present in 151 * genunix are added to a third tdata_t - the uniquified tdata_t. 152 * 153 * Additive Merges 154 * 155 * In some cases, for example if we are issuing a new version of a common 156 * module in a patch, we need to make sure that the CTF data already present 157 * in that module does not change. Changes to this data would void the CTF 158 * data in any module that uniquified against the common module. To preserve 159 * the existing data, we can perform what is known as an additive merge. In 160 * this case, a final uniquification is performed against the CTF data in the 161 * previous version of the module. The result will be the placement of new 162 * and changed data after the existing data, thus preserving the existing type 163 * ID space. 164 * 165 * Saving the result 166 * 167 * When the merges are complete, the resulting tdata_t is placed into the 168 * output file, replacing the .SUNW_ctf section (if any) already in that file. 169 * 170 * The person who changes the merging thread code in this file without updating 171 * this comment will not live to see the stock hit five. 172 */ 173 174 #if HAVE_NBTOOL_CONFIG_H 175 # include "nbtool_config.h" 176 #endif 177 178 #include <stdio.h> 179 #include <stdlib.h> 180 #ifndef _NETBSD_SOURCE 181 #define _NETBSD_SOURCE /* XXX TBD fix this */ 182 #include <unistd.h> 183 #undef _NETBSD_SOURCE 184 #else 185 #include <unistd.h> 186 #endif 187 #include <pthread.h> 188 #include <assert.h> 189 #if defined(sun) 190 #include <synch.h> 191 #endif 192 #include <signal.h> 193 #include <libgen.h> 194 #include <string.h> 195 #include <errno.h> 196 #if defined(sun) 197 #include <alloca.h> 198 #endif 199 #include <sys/param.h> 200 #include <sys/types.h> 201 #include <sys/mman.h> 202 #if defined(sun) 203 #include <sys/sysconf.h> 204 #endif 205 206 #include "ctf_headers.h" 207 #include "ctftools.h" 208 #include "ctfmerge.h" 209 #include "traverse.h" 210 #include "memory.h" 211 #include "fifo.h" 212 #include "barrier.h" 213 214 #pragma init(bigheap) 215 216 #define MERGE_PHASE1_BATCH_SIZE 8 217 #define MERGE_PHASE1_MAX_SLOTS 5 218 #define MERGE_INPUT_THROTTLE_LEN 10 219 220 const char *progname; 221 static char *outfile = NULL; 222 static char *tmpname = NULL; 223 static int dynsym; 224 int debug_level = DEBUG_LEVEL; 225 static size_t maxpgsize = 0x400000; 226 227 228 void 229 usage(void) 230 { 231 (void) fprintf(stderr, 232 "Usage: %s [-fgstv] -l label | -L labelenv -o outfile file ...\n" 233 " %s [-fgstv] -l label | -L labelenv -o outfile -d uniqfile\n" 234 " %*s [-g] [-D uniqlabel] file ...\n" 235 " %s [-fgstv] -l label | -L labelenv -o outfile -w withfile " 236 "file ...\n" 237 " %s [-g] -c srcfile destfile\n" 238 "\n" 239 " Note: if -L labelenv is specified and labelenv is not set in\n" 240 " the environment, a default value is used.\n", 241 progname, progname, strlen(progname), " ", 242 progname, progname); 243 } 244 245 #if defined(sun) 246 static void 247 bigheap(void) 248 { 249 size_t big, *size; 250 int sizes; 251 struct memcntl_mha mha; 252 253 /* 254 * First, get the available pagesizes. 255 */ 256 if ((sizes = getpagesizes(NULL, 0)) == -1) 257 return; 258 259 if (sizes == 1 || (size = alloca(sizeof (size_t) * sizes)) == NULL) 260 return; 261 262 if (getpagesizes(size, sizes) == -1) 263 return; 264 265 while (size[sizes - 1] > maxpgsize) 266 sizes--; 267 268 /* set big to the largest allowed page size */ 269 big = size[sizes - 1]; 270 if (big & (big - 1)) { 271 /* 272 * The largest page size is not a power of two for some 273 * inexplicable reason; return. 274 */ 275 return; 276 } 277 278 /* 279 * Now, align our break to the largest page size. 280 */ 281 if (brk((void *)((((uintptr_t)sbrk(0) - 1) & ~(big - 1)) + big)) != 0) 282 return; 283 284 /* 285 * set the preferred page size for the heap 286 */ 287 mha.mha_cmd = MHA_MAPSIZE_BSSBRK; 288 mha.mha_flags = 0; 289 mha.mha_pagesize = big; 290 291 (void) memcntl(NULL, 0, MC_HAT_ADVISE, (caddr_t)&mha, 0, 0); 292 } 293 #endif 294 295 static void 296 finalize_phase_one(workqueue_t *wq) 297 { 298 int startslot, i; 299 300 /* 301 * wip slots are cleared out only when maxbatchsz td's have been merged 302 * into them. We're not guaranteed that the number of files we're 303 * merging is a multiple of maxbatchsz, so there will be some partial 304 * groups in the wip array. Move them to the done queue in batch ID 305 * order, starting with the slot containing the next batch that would 306 * have been placed on the done queue, followed by the others. 307 * One thread will be doing this while the others wait at the barrier 308 * back in worker_thread(), so we don't need to worry about pesky things 309 * like locks. 310 */ 311 312 for (startslot = -1, i = 0; i < wq->wq_nwipslots; i++) { 313 if (wq->wq_wip[i].wip_batchid == wq->wq_lastdonebatch + 1) { 314 startslot = i; 315 break; 316 } 317 } 318 319 assert(startslot != -1); 320 321 for (i = startslot; i < startslot + wq->wq_nwipslots; i++) { 322 int slotnum = i % wq->wq_nwipslots; 323 wip_t *wipslot = &wq->wq_wip[slotnum]; 324 325 if (wipslot->wip_td != NULL) { 326 debug(2, "clearing slot %d (%d) (saving %d)\n", 327 slotnum, i, wipslot->wip_nmerged); 328 } else 329 debug(2, "clearing slot %d (%d)\n", slotnum, i); 330 331 if (wipslot->wip_td != NULL) { 332 fifo_add(wq->wq_donequeue, wipslot->wip_td); 333 wq->wq_wip[slotnum].wip_td = NULL; 334 } 335 } 336 337 wq->wq_lastdonebatch = wq->wq_next_batchid++; 338 339 debug(2, "phase one done: donequeue has %d items\n", 340 fifo_len(wq->wq_donequeue)); 341 } 342 343 static void 344 init_phase_two(workqueue_t *wq) 345 { 346 int num; 347 348 /* 349 * We're going to continually merge the first two entries on the queue, 350 * placing the result on the end, until there's nothing left to merge. 351 * At that point, everything will have been merged into one. The 352 * initial value of ninqueue needs to be equal to the total number of 353 * entries that will show up on the queue, both at the start of the 354 * phase and as generated by merges during the phase. 355 */ 356 wq->wq_ninqueue = num = fifo_len(wq->wq_donequeue); 357 while (num != 1) { 358 wq->wq_ninqueue += num / 2; 359 num = num / 2 + num % 2; 360 } 361 362 /* 363 * Move the done queue to the work queue. We won't be using the done 364 * queue in phase 2. 365 */ 366 assert(fifo_len(wq->wq_queue) == 0); 367 fifo_free(wq->wq_queue, NULL); 368 wq->wq_queue = wq->wq_donequeue; 369 } 370 371 static void 372 wip_save_work(workqueue_t *wq, wip_t *slot, int slotnum) 373 { 374 pthread_mutex_lock(&wq->wq_donequeue_lock); 375 376 while (wq->wq_lastdonebatch + 1 < slot->wip_batchid) 377 pthread_cond_wait(&slot->wip_cv, &wq->wq_donequeue_lock); 378 assert(wq->wq_lastdonebatch + 1 == slot->wip_batchid); 379 380 fifo_add(wq->wq_donequeue, slot->wip_td); 381 wq->wq_lastdonebatch++; 382 pthread_cond_signal(&wq->wq_wip[(slotnum + 1) % 383 wq->wq_nwipslots].wip_cv); 384 385 /* reset the slot for next use */ 386 slot->wip_td = NULL; 387 slot->wip_batchid = wq->wq_next_batchid++; 388 389 pthread_mutex_unlock(&wq->wq_donequeue_lock); 390 } 391 392 static void 393 wip_add_work(wip_t *slot, tdata_t *pow) 394 { 395 if (slot->wip_td == NULL) { 396 slot->wip_td = pow; 397 slot->wip_nmerged = 1; 398 } else { 399 debug(2, "%d: merging %p into %p\n", pthread_self(), 400 (void *)pow, (void *)slot->wip_td); 401 402 merge_into_master(pow, slot->wip_td, NULL, 0); 403 tdata_free(pow); 404 405 slot->wip_nmerged++; 406 } 407 } 408 409 static void 410 worker_runphase1(workqueue_t *wq) 411 { 412 wip_t *wipslot; 413 tdata_t *pow; 414 int wipslotnum, pownum; 415 416 for (;;) { 417 pthread_mutex_lock(&wq->wq_queue_lock); 418 419 while (fifo_empty(wq->wq_queue)) { 420 if (wq->wq_nomorefiles == 1) { 421 pthread_cond_broadcast(&wq->wq_work_avail); 422 pthread_mutex_unlock(&wq->wq_queue_lock); 423 424 /* on to phase 2 ... */ 425 return; 426 } 427 428 pthread_cond_wait(&wq->wq_work_avail, 429 &wq->wq_queue_lock); 430 } 431 432 /* there's work to be done! */ 433 pow = fifo_remove(wq->wq_queue); 434 pownum = wq->wq_nextpownum++; 435 pthread_cond_broadcast(&wq->wq_work_removed); 436 437 assert(pow != NULL); 438 439 /* merge it into the right slot */ 440 wipslotnum = pownum % wq->wq_nwipslots; 441 wipslot = &wq->wq_wip[wipslotnum]; 442 443 pthread_mutex_lock(&wipslot->wip_lock); 444 445 pthread_mutex_unlock(&wq->wq_queue_lock); 446 447 wip_add_work(wipslot, pow); 448 449 if (wipslot->wip_nmerged == wq->wq_maxbatchsz) 450 wip_save_work(wq, wipslot, wipslotnum); 451 452 pthread_mutex_unlock(&wipslot->wip_lock); 453 } 454 } 455 456 static void 457 worker_runphase2(workqueue_t *wq) 458 { 459 tdata_t *pow1, *pow2; 460 int batchid; 461 462 for (;;) { 463 pthread_mutex_lock(&wq->wq_queue_lock); 464 465 if (wq->wq_ninqueue == 1) { 466 pthread_cond_broadcast(&wq->wq_work_avail); 467 pthread_mutex_unlock(&wq->wq_queue_lock); 468 469 debug(2, "%d: entering p2 completion barrier\n", 470 pthread_self()); 471 if (barrier_wait(&wq->wq_bar1)) { 472 pthread_mutex_lock(&wq->wq_queue_lock); 473 wq->wq_alldone = 1; 474 pthread_cond_signal(&wq->wq_alldone_cv); 475 pthread_mutex_unlock(&wq->wq_queue_lock); 476 } 477 478 return; 479 } 480 481 if (fifo_len(wq->wq_queue) < 2) { 482 pthread_cond_wait(&wq->wq_work_avail, 483 &wq->wq_queue_lock); 484 pthread_mutex_unlock(&wq->wq_queue_lock); 485 continue; 486 } 487 488 /* there's work to be done! */ 489 pow1 = fifo_remove(wq->wq_queue); 490 pow2 = fifo_remove(wq->wq_queue); 491 wq->wq_ninqueue -= 2; 492 493 batchid = wq->wq_next_batchid++; 494 495 pthread_mutex_unlock(&wq->wq_queue_lock); 496 497 debug(2, "%d: merging %p into %p\n", pthread_self(), 498 (void *)pow1, (void *)pow2); 499 merge_into_master(pow1, pow2, NULL, 0); 500 tdata_free(pow1); 501 502 /* 503 * merging is complete. place at the tail of the queue in 504 * proper order. 505 */ 506 pthread_mutex_lock(&wq->wq_queue_lock); 507 while (wq->wq_lastdonebatch + 1 != batchid) { 508 pthread_cond_wait(&wq->wq_done_cv, 509 &wq->wq_queue_lock); 510 } 511 512 wq->wq_lastdonebatch = batchid; 513 514 fifo_add(wq->wq_queue, pow2); 515 debug(2, "%d: added %p to queue, len now %d, ninqueue %d\n", 516 pthread_self(), (void *)pow2, fifo_len(wq->wq_queue), 517 wq->wq_ninqueue); 518 pthread_cond_broadcast(&wq->wq_done_cv); 519 pthread_cond_signal(&wq->wq_work_avail); 520 pthread_mutex_unlock(&wq->wq_queue_lock); 521 } 522 } 523 524 /* 525 * Main loop for worker threads. 526 */ 527 static void 528 worker_thread(workqueue_t *wq) 529 { 530 worker_runphase1(wq); 531 532 debug(2, "%d: entering first barrier\n", pthread_self()); 533 534 if (barrier_wait(&wq->wq_bar1)) { 535 536 debug(2, "%d: doing work in first barrier\n", pthread_self()); 537 538 finalize_phase_one(wq); 539 540 init_phase_two(wq); 541 542 debug(2, "%d: ninqueue is %d, %d on queue\n", pthread_self(), 543 wq->wq_ninqueue, fifo_len(wq->wq_queue)); 544 } 545 546 debug(2, "%d: entering second barrier\n", pthread_self()); 547 548 (void) barrier_wait(&wq->wq_bar2); 549 550 debug(2, "%d: phase 1 complete\n", pthread_self()); 551 552 worker_runphase2(wq); 553 } 554 555 /* 556 * Pass a tdata_t tree, built from an input file, off to the work queue for 557 * consumption by worker threads. 558 */ 559 static int 560 merge_ctf_cb(tdata_t *td, char *name, void *arg) 561 { 562 workqueue_t *wq = arg; 563 564 debug(3, "Adding tdata %p for processing\n", (void *)td); 565 566 pthread_mutex_lock(&wq->wq_queue_lock); 567 while (fifo_len(wq->wq_queue) > wq->wq_ithrottle) { 568 debug(2, "Throttling input (len = %d, throttle = %d)\n", 569 fifo_len(wq->wq_queue), wq->wq_ithrottle); 570 pthread_cond_wait(&wq->wq_work_removed, &wq->wq_queue_lock); 571 } 572 573 fifo_add(wq->wq_queue, td); 574 debug(1, "Thread %d announcing %s\n", pthread_self(), name); 575 pthread_cond_broadcast(&wq->wq_work_avail); 576 pthread_mutex_unlock(&wq->wq_queue_lock); 577 578 return (1); 579 } 580 581 /* 582 * This program is intended to be invoked from a Makefile, as part of the build. 583 * As such, in the event of a failure or user-initiated interrupt (^C), we need 584 * to ensure that a subsequent re-make will cause ctfmerge to be executed again. 585 * Unfortunately, ctfmerge will usually be invoked directly after (and as part 586 * of the same Makefile rule as) a link, and will operate on the linked file 587 * in place. If we merely exit upon receipt of a SIGINT, a subsequent make 588 * will notice that the *linked* file is newer than the object files, and thus 589 * will not reinvoke ctfmerge. The only way to ensure that a subsequent make 590 * reinvokes ctfmerge, is to remove the file to which we are adding CTF 591 * data (confusingly named the output file). This means that the link will need 592 * to happen again, but links are generally fast, and we can't allow the merge 593 * to be skipped. 594 * 595 * Another possibility would be to block SIGINT entirely - to always run to 596 * completion. The run time of ctfmerge can, however, be measured in minutes 597 * in some cases, so this is not a valid option. 598 */ 599 static void 600 handle_sig(int sig) 601 { 602 terminate("Caught signal %d - exiting\n", sig); 603 } 604 605 static void 606 terminate_cleanup(void) 607 { 608 int dounlink = getenv("CTFMERGE_TERMINATE_NO_UNLINK") ? 0 : 1; 609 610 if (tmpname != NULL && dounlink) 611 unlink(tmpname); 612 613 if (outfile == NULL) 614 return; 615 616 #if !defined(__FreeBSD__) 617 if (dounlink) { 618 fprintf(stderr, "Removing %s\n", outfile); 619 unlink(outfile); 620 } 621 #endif 622 } 623 624 static void 625 copy_ctf_data(char *srcfile, char *destfile, int keep_stabs) 626 { 627 tdata_t *srctd; 628 629 if (read_ctf(&srcfile, 1, NULL, read_ctf_save_cb, &srctd, 1) == 0) 630 terminate("No CTF data found in source file %s\n", srcfile); 631 632 tmpname = mktmpname(destfile, ".ctf"); 633 write_ctf(srctd, destfile, tmpname, CTF_COMPRESS | keep_stabs); 634 if (rename(tmpname, destfile) != 0) { 635 terminate("Couldn't rename temp file %s to %s", tmpname, 636 destfile); 637 } 638 free(tmpname); 639 tdata_free(srctd); 640 } 641 642 static void 643 wq_init(workqueue_t *wq, int nfiles) 644 { 645 int throttle, nslots, i; 646 647 if (getenv("CTFMERGE_MAX_SLOTS")) 648 nslots = atoi(getenv("CTFMERGE_MAX_SLOTS")); 649 else 650 nslots = MERGE_PHASE1_MAX_SLOTS; 651 652 if (getenv("CTFMERGE_PHASE1_BATCH_SIZE")) 653 wq->wq_maxbatchsz = atoi(getenv("CTFMERGE_PHASE1_BATCH_SIZE")); 654 else 655 wq->wq_maxbatchsz = MERGE_PHASE1_BATCH_SIZE; 656 657 nslots = MIN(nslots, (nfiles + wq->wq_maxbatchsz - 1) / 658 wq->wq_maxbatchsz); 659 660 wq->wq_wip = xcalloc(sizeof (wip_t) * nslots); 661 wq->wq_nwipslots = nslots; 662 wq->wq_nthreads = MIN(sysconf(_SC_NPROCESSORS_ONLN) * 3 / 2, nslots); 663 wq->wq_thread = xmalloc(sizeof (pthread_t) * wq->wq_nthreads); 664 665 if (getenv("CTFMERGE_INPUT_THROTTLE")) 666 throttle = atoi(getenv("CTFMERGE_INPUT_THROTTLE")); 667 else 668 throttle = MERGE_INPUT_THROTTLE_LEN; 669 wq->wq_ithrottle = throttle * wq->wq_nthreads; 670 671 debug(1, "Using %d slots, %d threads\n", wq->wq_nwipslots, 672 wq->wq_nthreads); 673 674 wq->wq_next_batchid = 0; 675 676 for (i = 0; i < nslots; i++) { 677 pthread_mutex_init(&wq->wq_wip[i].wip_lock, NULL); 678 pthread_cond_init(&wq->wq_wip[i].wip_cv, NULL); 679 wq->wq_wip[i].wip_batchid = wq->wq_next_batchid++; 680 } 681 682 pthread_mutex_init(&wq->wq_queue_lock, NULL); 683 wq->wq_queue = fifo_new(); 684 pthread_cond_init(&wq->wq_work_avail, NULL); 685 pthread_cond_init(&wq->wq_work_removed, NULL); 686 wq->wq_ninqueue = nfiles; 687 wq->wq_nextpownum = 0; 688 689 pthread_mutex_init(&wq->wq_donequeue_lock, NULL); 690 wq->wq_donequeue = fifo_new(); 691 wq->wq_lastdonebatch = -1; 692 693 pthread_cond_init(&wq->wq_done_cv, NULL); 694 695 pthread_cond_init(&wq->wq_alldone_cv, NULL); 696 wq->wq_alldone = 0; 697 698 barrier_init(&wq->wq_bar1, wq->wq_nthreads); 699 barrier_init(&wq->wq_bar2, wq->wq_nthreads); 700 701 wq->wq_nomorefiles = 0; 702 } 703 704 static void 705 start_threads(workqueue_t *wq) 706 { 707 pthread_t thrid; 708 sigset_t sets; 709 int i; 710 711 sigemptyset(&sets); 712 sigaddset(&sets, SIGINT); 713 sigaddset(&sets, SIGQUIT); 714 sigaddset(&sets, SIGTERM); 715 pthread_sigmask(SIG_BLOCK, &sets, NULL); 716 717 for (i = 0; i < wq->wq_nthreads; i++) { 718 pthread_create(&wq->wq_thread[i], NULL, 719 (void *(*)(void *))worker_thread, wq); 720 } 721 722 #if defined(sun) 723 sigset(SIGINT, handle_sig); 724 sigset(SIGQUIT, handle_sig); 725 sigset(SIGTERM, handle_sig); 726 #else 727 signal(SIGINT, handle_sig); 728 signal(SIGQUIT, handle_sig); 729 signal(SIGTERM, handle_sig); 730 #endif 731 pthread_sigmask(SIG_UNBLOCK, &sets, NULL); 732 } 733 734 static void 735 join_threads(workqueue_t *wq) 736 { 737 int i; 738 739 for (i = 0; i < wq->wq_nthreads; i++) { 740 pthread_join(wq->wq_thread[i], NULL); 741 } 742 } 743 744 745 static int 746 strcompare(const void *p1, const void *p2) 747 { 748 char *s1 = *((char **)p1); 749 char *s2 = *((char **)p2); 750 751 return (strcmp(s1, s2)); 752 } 753 754 /* 755 * Core work queue structure; passed to worker threads on thread creation 756 * as the main point of coordination. Allocate as a static structure; we 757 * could have put this into a local variable in main, but passing a pointer 758 * into your stack to another thread is fragile at best and leads to some 759 * hard-to-debug failure modes. 760 */ 761 static workqueue_t wq; 762 763 int 764 main(int argc, char **argv) 765 { 766 tdata_t *mstrtd, *savetd; 767 char *uniqfile = NULL, *uniqlabel = NULL; 768 char *withfile = NULL; 769 char *label = NULL; 770 char **ifiles, **tifiles; 771 int verbose = 0, docopy = 0; 772 int write_fuzzy_match = 0; 773 int keep_stabs = 0; 774 int require_ctf = 0; 775 int nifiles, nielems; 776 int c, i, idx, tidx, err; 777 778 progname = basename(argv[0]); 779 780 if (getenv("CTFMERGE_DEBUG_LEVEL")) 781 debug_level = atoi(getenv("CTFMERGE_DEBUG_LEVEL")); 782 783 err = 0; 784 while ((c = getopt(argc, argv, ":cd:D:fgl:L:o:tvw:s")) != EOF) { 785 switch (c) { 786 case 'c': 787 docopy = 1; 788 break; 789 case 'd': 790 /* Uniquify against `uniqfile' */ 791 uniqfile = optarg; 792 break; 793 case 'D': 794 /* Uniquify against label `uniqlabel' in `uniqfile' */ 795 uniqlabel = optarg; 796 break; 797 case 'f': 798 write_fuzzy_match = CTF_FUZZY_MATCH; 799 break; 800 case 'g': 801 keep_stabs = CTF_KEEP_STABS; 802 break; 803 case 'l': 804 /* Label merged types with `label' */ 805 label = optarg; 806 break; 807 case 'L': 808 /* Label merged types with getenv(`label`) */ 809 if ((label = getenv(optarg)) == NULL) 810 label = CTF_DEFAULT_LABEL; 811 break; 812 case 'o': 813 /* Place merged types in CTF section in `outfile' */ 814 outfile = optarg; 815 break; 816 case 't': 817 /* Insist *all* object files built from C have CTF */ 818 require_ctf = 1; 819 break; 820 case 'v': 821 /* More debugging information */ 822 verbose = 1; 823 break; 824 case 'w': 825 /* Additive merge with data from `withfile' */ 826 withfile = optarg; 827 break; 828 case 's': 829 /* use the dynsym rather than the symtab */ 830 dynsym = CTF_USE_DYNSYM; 831 break; 832 default: 833 usage(); 834 exit(2); 835 } 836 } 837 838 /* Validate arguments */ 839 if (docopy) { 840 if (uniqfile != NULL || uniqlabel != NULL || label != NULL || 841 outfile != NULL || withfile != NULL || dynsym != 0) 842 err++; 843 844 if (argc - optind != 2) 845 err++; 846 } else { 847 if (uniqfile != NULL && withfile != NULL) 848 err++; 849 850 if (uniqlabel != NULL && uniqfile == NULL) 851 err++; 852 853 if (outfile == NULL || label == NULL) 854 err++; 855 856 if (argc - optind == 0) 857 err++; 858 } 859 860 if (err) { 861 usage(); 862 exit(2); 863 } 864 865 if (getenv("STRIPSTABS_KEEP_STABS") != NULL) 866 keep_stabs = CTF_KEEP_STABS; 867 868 if (uniqfile && access(uniqfile, R_OK) != 0) { 869 warning("Uniquification file %s couldn't be opened and " 870 "will be ignored.\n", uniqfile); 871 uniqfile = NULL; 872 } 873 if (withfile && access(withfile, R_OK) != 0) { 874 warning("With file %s couldn't be opened and will be " 875 "ignored.\n", withfile); 876 withfile = NULL; 877 } 878 if (outfile && access(outfile, R_OK|W_OK) != 0) 879 terminate("Cannot open output file %s for r/w", outfile); 880 881 /* 882 * This is ugly, but we don't want to have to have a separate tool 883 * (yet) just for copying an ELF section with our specific requirements, 884 * so we shoe-horn a copier into ctfmerge. 885 */ 886 if (docopy) { 887 copy_ctf_data(argv[optind], argv[optind + 1], keep_stabs); 888 889 exit(0); 890 } 891 892 set_terminate_cleanup(terminate_cleanup); 893 894 /* Sort the input files and strip out duplicates */ 895 nifiles = argc - optind; 896 ifiles = xmalloc(sizeof (char *) * nifiles); 897 tifiles = xmalloc(sizeof (char *) * nifiles); 898 899 for (i = 0; i < nifiles; i++) 900 tifiles[i] = argv[optind + i]; 901 qsort(tifiles, nifiles, sizeof (char *), (int (*)())strcompare); 902 903 ifiles[0] = tifiles[0]; 904 for (idx = 0, tidx = 1; tidx < nifiles; tidx++) { 905 if (strcmp(ifiles[idx], tifiles[tidx]) != 0) 906 ifiles[++idx] = tifiles[tidx]; 907 } 908 nifiles = idx + 1; 909 910 /* Make sure they all exist */ 911 if ((nielems = count_files(ifiles, nifiles)) < 0) 912 terminate("Some input files were inaccessible\n"); 913 914 /* Prepare for the merge */ 915 wq_init(&wq, nielems); 916 917 start_threads(&wq); 918 919 /* 920 * Start the merge 921 * 922 * We're reading everything from each of the object files, so we 923 * don't need to specify labels. 924 */ 925 if (read_ctf(ifiles, nifiles, NULL, merge_ctf_cb, 926 &wq, require_ctf) == 0) { 927 /* 928 * If we're verifying that C files have CTF, it's safe to 929 * assume that in this case, we're building only from assembly 930 * inputs. 931 */ 932 if (require_ctf) 933 exit(0); 934 terminate("No ctf sections found to merge\n"); 935 } 936 937 pthread_mutex_lock(&wq.wq_queue_lock); 938 wq.wq_nomorefiles = 1; 939 pthread_cond_broadcast(&wq.wq_work_avail); 940 pthread_mutex_unlock(&wq.wq_queue_lock); 941 942 pthread_mutex_lock(&wq.wq_queue_lock); 943 while (wq.wq_alldone == 0) 944 pthread_cond_wait(&wq.wq_alldone_cv, &wq.wq_queue_lock); 945 pthread_mutex_unlock(&wq.wq_queue_lock); 946 947 join_threads(&wq); 948 949 /* 950 * All requested files have been merged, with the resulting tree in 951 * mstrtd. savetd is the tree that will be placed into the output file. 952 * 953 * Regardless of whether we're doing a normal uniquification or an 954 * additive merge, we need a type tree that has been uniquified 955 * against uniqfile or withfile, as appropriate. 956 * 957 * If we're doing a uniquification, we stuff the resulting tree into 958 * outfile. Otherwise, we add the tree to the tree already in withfile. 959 */ 960 assert(fifo_len(wq.wq_queue) == 1); 961 mstrtd = fifo_remove(wq.wq_queue); 962 963 if (verbose || debug_level) { 964 debug(2, "Statistics for td %p\n", (void *)mstrtd); 965 966 iidesc_stats(mstrtd->td_iihash); 967 } 968 969 if (uniqfile != NULL || withfile != NULL) { 970 char *reffile, *reflabel = NULL; 971 tdata_t *reftd; 972 973 if (uniqfile != NULL) { 974 reffile = uniqfile; 975 reflabel = uniqlabel; 976 } else 977 reffile = withfile; 978 979 if (read_ctf(&reffile, 1, reflabel, read_ctf_save_cb, 980 &reftd, require_ctf) == 0) { 981 terminate("No CTF data found in reference file %s\n", 982 reffile); 983 } 984 985 savetd = tdata_new(); 986 987 if (CTF_TYPE_ISCHILD(reftd->td_nextid)) 988 terminate("No room for additional types in master\n"); 989 990 savetd->td_nextid = withfile ? reftd->td_nextid : 991 CTF_INDEX_TO_TYPE(1, TRUE); 992 merge_into_master(mstrtd, reftd, savetd, 0); 993 994 tdata_label_add(savetd, label, CTF_LABEL_LASTIDX); 995 996 if (withfile) { 997 /* 998 * savetd holds the new data to be added to the withfile 999 */ 1000 tdata_t *withtd = reftd; 1001 1002 tdata_merge(withtd, savetd); 1003 1004 savetd = withtd; 1005 } else { 1006 char uniqname[MAXPATHLEN]; 1007 labelent_t *parle; 1008 1009 parle = tdata_label_top(reftd); 1010 1011 savetd->td_parlabel = xstrdup(parle->le_name); 1012 1013 strncpy(uniqname, reffile, sizeof (uniqname)); 1014 uniqname[MAXPATHLEN - 1] = '\0'; 1015 savetd->td_parname = xstrdup(basename(uniqname)); 1016 } 1017 1018 } else { 1019 /* 1020 * No post processing. Write the merged tree as-is into the 1021 * output file. 1022 */ 1023 tdata_label_free(mstrtd); 1024 tdata_label_add(mstrtd, label, CTF_LABEL_LASTIDX); 1025 1026 savetd = mstrtd; 1027 } 1028 1029 tmpname = mktmpname(outfile, ".ctf"); 1030 write_ctf(savetd, outfile, tmpname, 1031 CTF_COMPRESS | write_fuzzy_match | dynsym | keep_stabs); 1032 if (rename(tmpname, outfile) != 0) 1033 terminate("Couldn't rename output temp file %s", tmpname); 1034 free(tmpname); 1035 1036 return (0); 1037 } 1038