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 2008 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 #include <stdio.h>
175 #include <stdlib.h>
176 #include <unistd.h>
177 #include <pthread.h>
178 #include <assert.h>
179 #include <synch.h>
180 #include <signal.h>
181 #include <libgen.h>
182 #include <string.h>
183 #include <errno.h>
184 #include <alloca.h>
185 #include <sys/param.h>
186 #include <sys/types.h>
187 #include <sys/mman.h>
188 #include <sys/sysconf.h>
189
190 #include "ctf_headers.h"
191 #include "ctftools.h"
192 #include "ctfmerge.h"
193 #include "traverse.h"
194 #include "memory.h"
195 #include "fifo.h"
196 #include "barrier.h"
197
198 #pragma init(bigheap)
199
200 #define MERGE_PHASE1_BATCH_SIZE 8
201 #define MERGE_PHASE1_MAX_SLOTS 5
202 #define MERGE_INPUT_THROTTLE_LEN 10
203
204 const char *progname;
205 static char *outfile = NULL;
206 static char *tmpname = NULL;
207 static int dynsym;
208 int debug_level = DEBUG_LEVEL;
209 static size_t maxpgsize = 0x400000;
210
211
212 void
213 usage(void)
214 {
215 (void) fprintf(stderr,
216 "Usage: %s [-fgstv] -l label | -L labelenv -o outfile file ...\n"
217 " %s [-fgstv] -l label | -L labelenv -o outfile -d uniqfile\n"
218 " %*s [-g] [-D uniqlabel] file ...\n"
219 " %s [-fgstv] -l label | -L labelenv -o outfile -w withfile "
220 "file ...\n"
221 " %s [-g] -c srcfile destfile\n"
222 "\n"
223 " Note: if -L labelenv is specified and labelenv is not set in\n"
224 " the environment, a default value is used.\n",
225 progname, progname, strlen(progname), " ",
226 progname, progname);
227 }
228
229 static void
230 bigheap(void)
231 {
232 size_t big, *size;
233 int sizes;
234 struct memcntl_mha mha;
235
236 /*
237 * First, get the available pagesizes.
238 */
239 if ((sizes = getpagesizes(NULL, 0)) == -1)
240 return;
241
242 if (sizes == 1 || (size = alloca(sizeof (size_t) * sizes)) == NULL)
243 return;
244
245 if (getpagesizes(size, sizes) == -1)
246 return;
247
248 while (size[sizes - 1] > maxpgsize)
249 sizes--;
250
251 /* set big to the largest allowed page size */
252 big = size[sizes - 1];
253 if (big & (big - 1)) {
254 /*
255 * The largest page size is not a power of two for some
256 * inexplicable reason; return.
257 */
258 return;
259 }
260
261 /*
262 * Now, align our break to the largest page size.
263 */
264 if (brk((void *)((((uintptr_t)sbrk(0) - 1) & ~(big - 1)) + big)) != 0)
265 return;
266
267 /*
268 * set the preferred page size for the heap
269 */
270 mha.mha_cmd = MHA_MAPSIZE_BSSBRK;
271 mha.mha_flags = 0;
272 mha.mha_pagesize = big;
273
274 (void) memcntl(NULL, 0, MC_HAT_ADVISE, (caddr_t)&mha, 0, 0);
275 }
276
277 static void
278 finalize_phase_one(workqueue_t *wq)
279 {
280 int startslot, i;
281
282 /*
283 * wip slots are cleared out only when maxbatchsz td's have been merged
284 * into them. We're not guaranteed that the number of files we're
285 * merging is a multiple of maxbatchsz, so there will be some partial
286 * groups in the wip array. Move them to the done queue in batch ID
287 * order, starting with the slot containing the next batch that would
288 * have been placed on the done queue, followed by the others.
289 * One thread will be doing this while the others wait at the barrier
290 * back in worker_thread(), so we don't need to worry about pesky things
291 * like locks.
292 */
293
294 for (startslot = -1, i = 0; i < wq->wq_nwipslots; i++) {
295 if (wq->wq_wip[i].wip_batchid == wq->wq_lastdonebatch + 1) {
296 startslot = i;
297 break;
298 }
299 }
300
301 assert(startslot != -1);
302
303 for (i = startslot; i < startslot + wq->wq_nwipslots; i++) {
304 int slotnum = i % wq->wq_nwipslots;
305 wip_t *wipslot = &wq->wq_wip[slotnum];
306
307 if (wipslot->wip_td != NULL) {
308 debug(2, "clearing slot %d (%d) (saving %d)\n",
309 slotnum, i, wipslot->wip_nmerged);
310 } else
311 debug(2, "clearing slot %d (%d)\n", slotnum, i);
312
313 if (wipslot->wip_td != NULL) {
314 fifo_add(wq->wq_donequeue, wipslot->wip_td);
315 wq->wq_wip[slotnum].wip_td = NULL;
316 }
317 }
318
319 wq->wq_lastdonebatch = wq->wq_next_batchid++;
320
321 debug(2, "phase one done: donequeue has %d items\n",
322 fifo_len(wq->wq_donequeue));
323 }
324
325 static void
326 init_phase_two(workqueue_t *wq)
327 {
328 int num;
329
330 /*
331 * We're going to continually merge the first two entries on the queue,
332 * placing the result on the end, until there's nothing left to merge.
333 * At that point, everything will have been merged into one. The
334 * initial value of ninqueue needs to be equal to the total number of
335 * entries that will show up on the queue, both at the start of the
336 * phase and as generated by merges during the phase.
337 */
338 wq->wq_ninqueue = num = fifo_len(wq->wq_donequeue);
339 while (num != 1) {
340 wq->wq_ninqueue += num / 2;
341 num = num / 2 + num % 2;
342 }
343
344 /*
345 * Move the done queue to the work queue. We won't be using the done
346 * queue in phase 2.
347 */
348 assert(fifo_len(wq->wq_queue) == 0);
349 fifo_free(wq->wq_queue, NULL);
350 wq->wq_queue = wq->wq_donequeue;
351 }
352
353 static void
354 wip_save_work(workqueue_t *wq, wip_t *slot, int slotnum)
355 {
356 pthread_mutex_lock(&wq->wq_donequeue_lock);
357
358 while (wq->wq_lastdonebatch + 1 < slot->wip_batchid)
359 pthread_cond_wait(&slot->wip_cv, &wq->wq_donequeue_lock);
360 assert(wq->wq_lastdonebatch + 1 == slot->wip_batchid);
361
362 fifo_add(wq->wq_donequeue, slot->wip_td);
363 wq->wq_lastdonebatch++;
364 pthread_cond_signal(&wq->wq_wip[(slotnum + 1) %
365 wq->wq_nwipslots].wip_cv);
366
367 /* reset the slot for next use */
368 slot->wip_td = NULL;
369 slot->wip_batchid = wq->wq_next_batchid++;
370
371 pthread_mutex_unlock(&wq->wq_donequeue_lock);
372 }
373
374 static void
375 wip_add_work(wip_t *slot, tdata_t *pow)
376 {
377 if (slot->wip_td == NULL) {
378 slot->wip_td = pow;
379 slot->wip_nmerged = 1;
380 } else {
381 debug(2, "%d: merging %p into %p\n", pthread_self(),
382 (void *)pow, (void *)slot->wip_td);
383
384 merge_into_master(pow, slot->wip_td, NULL, 0);
385 tdata_free(pow);
386
387 slot->wip_nmerged++;
388 }
389 }
390
391 static void
392 worker_runphase1(workqueue_t *wq)
393 {
394 wip_t *wipslot;
395 tdata_t *pow;
396 int wipslotnum, pownum;
397
398 for (;;) {
399 pthread_mutex_lock(&wq->wq_queue_lock);
400
401 while (fifo_empty(wq->wq_queue)) {
402 if (wq->wq_nomorefiles == 1) {
403 pthread_cond_broadcast(&wq->wq_work_avail);
404 pthread_mutex_unlock(&wq->wq_queue_lock);
405
406 /* on to phase 2 ... */
407 return;
408 }
409
410 pthread_cond_wait(&wq->wq_work_avail,
411 &wq->wq_queue_lock);
412 }
413
414 /* there's work to be done! */
415 pow = fifo_remove(wq->wq_queue);
416 pownum = wq->wq_nextpownum++;
417 pthread_cond_broadcast(&wq->wq_work_removed);
418
419 assert(pow != NULL);
420
421 /* merge it into the right slot */
422 wipslotnum = pownum % wq->wq_nwipslots;
423 wipslot = &wq->wq_wip[wipslotnum];
424
425 pthread_mutex_lock(&wipslot->wip_lock);
426
427 pthread_mutex_unlock(&wq->wq_queue_lock);
428
429 wip_add_work(wipslot, pow);
430
431 if (wipslot->wip_nmerged == wq->wq_maxbatchsz)
432 wip_save_work(wq, wipslot, wipslotnum);
433
434 pthread_mutex_unlock(&wipslot->wip_lock);
435 }
436 }
437
438 static void
439 worker_runphase2(workqueue_t *wq)
440 {
441 tdata_t *pow1, *pow2;
442 int batchid;
443
444 for (;;) {
445 pthread_mutex_lock(&wq->wq_queue_lock);
446
447 if (wq->wq_ninqueue == 1) {
448 pthread_cond_broadcast(&wq->wq_work_avail);
449 pthread_mutex_unlock(&wq->wq_queue_lock);
450
451 debug(2, "%d: entering p2 completion barrier\n",
452 pthread_self());
453 if (barrier_wait(&wq->wq_bar1)) {
454 pthread_mutex_lock(&wq->wq_queue_lock);
455 wq->wq_alldone = 1;
456 pthread_cond_signal(&wq->wq_alldone_cv);
457 pthread_mutex_unlock(&wq->wq_queue_lock);
458 }
459
460 return;
461 }
462
463 if (fifo_len(wq->wq_queue) < 2) {
464 pthread_cond_wait(&wq->wq_work_avail,
465 &wq->wq_queue_lock);
466 pthread_mutex_unlock(&wq->wq_queue_lock);
467 continue;
468 }
469
470 /* there's work to be done! */
471 pow1 = fifo_remove(wq->wq_queue);
472 pow2 = fifo_remove(wq->wq_queue);
473 wq->wq_ninqueue -= 2;
474
475 batchid = wq->wq_next_batchid++;
476
477 pthread_mutex_unlock(&wq->wq_queue_lock);
478
479 debug(2, "%d: merging %p into %p\n", pthread_self(),
480 (void *)pow1, (void *)pow2);
481 merge_into_master(pow1, pow2, NULL, 0);
482 tdata_free(pow1);
483
484 /*
485 * merging is complete. place at the tail of the queue in
486 * proper order.
487 */
488 pthread_mutex_lock(&wq->wq_queue_lock);
489 while (wq->wq_lastdonebatch + 1 != batchid) {
490 pthread_cond_wait(&wq->wq_done_cv,
491 &wq->wq_queue_lock);
492 }
493
494 wq->wq_lastdonebatch = batchid;
495
496 fifo_add(wq->wq_queue, pow2);
497 debug(2, "%d: added %p to queue, len now %d, ninqueue %d\n",
498 pthread_self(), (void *)pow2, fifo_len(wq->wq_queue),
499 wq->wq_ninqueue);
500 pthread_cond_broadcast(&wq->wq_done_cv);
501 pthread_cond_signal(&wq->wq_work_avail);
502 pthread_mutex_unlock(&wq->wq_queue_lock);
503 }
504 }
505
506 /*
507 * Main loop for worker threads.
508 */
509 static void
510 worker_thread(workqueue_t *wq)
511 {
512 worker_runphase1(wq);
513
514 debug(2, "%d: entering first barrier\n", pthread_self());
515
516 if (barrier_wait(&wq->wq_bar1)) {
517
518 debug(2, "%d: doing work in first barrier\n", pthread_self());
519
520 finalize_phase_one(wq);
521
522 init_phase_two(wq);
523
524 debug(2, "%d: ninqueue is %d, %d on queue\n", pthread_self(),
525 wq->wq_ninqueue, fifo_len(wq->wq_queue));
526 }
527
528 debug(2, "%d: entering second barrier\n", pthread_self());
529
530 (void) barrier_wait(&wq->wq_bar2);
531
532 debug(2, "%d: phase 1 complete\n", pthread_self());
533
534 worker_runphase2(wq);
535 }
536
537 /*
538 * Pass a tdata_t tree, built from an input file, off to the work queue for
539 * consumption by worker threads.
540 */
541 static int
542 merge_ctf_cb(tdata_t *td, char *name, void *arg)
543 {
544 workqueue_t *wq = arg;
545
546 debug(3, "Adding tdata %p for processing\n", (void *)td);
547
548 pthread_mutex_lock(&wq->wq_queue_lock);
549 while (fifo_len(wq->wq_queue) > wq->wq_ithrottle) {
550 debug(2, "Throttling input (len = %d, throttle = %d)\n",
551 fifo_len(wq->wq_queue), wq->wq_ithrottle);
552 pthread_cond_wait(&wq->wq_work_removed, &wq->wq_queue_lock);
553 }
554
555 fifo_add(wq->wq_queue, td);
556 debug(1, "Thread %d announcing %s\n", pthread_self(), name);
557 pthread_cond_broadcast(&wq->wq_work_avail);
558 pthread_mutex_unlock(&wq->wq_queue_lock);
559
560 return (1);
561 }
562
563 /*
564 * This program is intended to be invoked from a Makefile, as part of the build.
565 * As such, in the event of a failure or user-initiated interrupt (^C), we need
566 * to ensure that a subsequent re-make will cause ctfmerge to be executed again.
567 * Unfortunately, ctfmerge will usually be invoked directly after (and as part
568 * of the same Makefile rule as) a link, and will operate on the linked file
569 * in place. If we merely exit upon receipt of a SIGINT, a subsequent make
570 * will notice that the *linked* file is newer than the object files, and thus
571 * will not reinvoke ctfmerge. The only way to ensure that a subsequent make
572 * reinvokes ctfmerge, is to remove the file to which we are adding CTF
573 * data (confusingly named the output file). This means that the link will need
574 * to happen again, but links are generally fast, and we can't allow the merge
575 * to be skipped.
576 *
577 * Another possibility would be to block SIGINT entirely - to always run to
578 * completion. The run time of ctfmerge can, however, be measured in minutes
579 * in some cases, so this is not a valid option.
580 */
581 static void
582 handle_sig(int sig)
583 {
584 terminate("Caught signal %d - exiting\n", sig);
585 }
586
587 static void
588 terminate_cleanup(void)
589 {
590 int dounlink = getenv("CTFMERGE_TERMINATE_NO_UNLINK") ? 0 : 1;
591
592 if (tmpname != NULL && dounlink)
593 unlink(tmpname);
594
595 if (outfile == NULL)
596 return;
597
598 if (dounlink) {
599 fprintf(stderr, "Removing %s\n", outfile);
600 unlink(outfile);
601 }
602 }
603
604 static void
605 copy_ctf_data(char *srcfile, char *destfile, int keep_stabs)
606 {
607 tdata_t *srctd;
608
609 if (read_ctf(&srcfile, 1, NULL, read_ctf_save_cb, &srctd, 1) == 0)
610 terminate("No CTF data found in source file %s\n", srcfile);
611
612 tmpname = mktmpname(destfile, ".ctf");
613 write_ctf(srctd, destfile, tmpname, CTF_COMPRESS | keep_stabs);
614 if (rename(tmpname, destfile) != 0) {
615 terminate("Couldn't rename temp file %s to %s", tmpname,
616 destfile);
617 }
618 free(tmpname);
619 tdata_free(srctd);
620 }
621
622 static void
623 wq_init(workqueue_t *wq, int nfiles)
624 {
625 int throttle, nslots, i;
626
627 if (getenv("CTFMERGE_MAX_SLOTS"))
628 nslots = atoi(getenv("CTFMERGE_MAX_SLOTS"));
629 else
630 nslots = MERGE_PHASE1_MAX_SLOTS;
631
632 if (getenv("CTFMERGE_PHASE1_BATCH_SIZE"))
633 wq->wq_maxbatchsz = atoi(getenv("CTFMERGE_PHASE1_BATCH_SIZE"));
634 else
635 wq->wq_maxbatchsz = MERGE_PHASE1_BATCH_SIZE;
636
637 nslots = MIN(nslots, (nfiles + wq->wq_maxbatchsz - 1) /
638 wq->wq_maxbatchsz);
639
640 wq->wq_wip = xcalloc(sizeof (wip_t) * nslots);
641 wq->wq_nwipslots = nslots;
642 wq->wq_nthreads = MIN(sysconf(_SC_NPROCESSORS_ONLN) * 3 / 2, nslots);
643 wq->wq_thread = xmalloc(sizeof (pthread_t) * wq->wq_nthreads);
644
645 if (getenv("CTFMERGE_INPUT_THROTTLE"))
646 throttle = atoi(getenv("CTFMERGE_INPUT_THROTTLE"));
647 else
648 throttle = MERGE_INPUT_THROTTLE_LEN;
649 wq->wq_ithrottle = throttle * wq->wq_nthreads;
650
651 debug(1, "Using %d slots, %d threads\n", wq->wq_nwipslots,
652 wq->wq_nthreads);
653
654 wq->wq_next_batchid = 0;
655
656 for (i = 0; i < nslots; i++) {
657 pthread_mutex_init(&wq->wq_wip[i].wip_lock, NULL);
658 wq->wq_wip[i].wip_batchid = wq->wq_next_batchid++;
659 }
660
661 pthread_mutex_init(&wq->wq_queue_lock, NULL);
662 wq->wq_queue = fifo_new();
663 pthread_cond_init(&wq->wq_work_avail, NULL);
664 pthread_cond_init(&wq->wq_work_removed, NULL);
665 wq->wq_ninqueue = nfiles;
666 wq->wq_nextpownum = 0;
667
668 pthread_mutex_init(&wq->wq_donequeue_lock, NULL);
669 wq->wq_donequeue = fifo_new();
670 wq->wq_lastdonebatch = -1;
671
672 pthread_cond_init(&wq->wq_done_cv, NULL);
673
674 pthread_cond_init(&wq->wq_alldone_cv, NULL);
675 wq->wq_alldone = 0;
676
677 barrier_init(&wq->wq_bar1, wq->wq_nthreads);
678 barrier_init(&wq->wq_bar2, wq->wq_nthreads);
679
680 wq->wq_nomorefiles = 0;
681 }
682
683 static void
684 start_threads(workqueue_t *wq)
685 {
686 sigset_t sets;
687 int i;
688
689 sigemptyset(&sets);
690 sigaddset(&sets, SIGINT);
691 sigaddset(&sets, SIGQUIT);
692 sigaddset(&sets, SIGTERM);
693 pthread_sigmask(SIG_BLOCK, &sets, NULL);
694
695 for (i = 0; i < wq->wq_nthreads; i++) {
696 pthread_create(&wq->wq_thread[i], NULL,
697 (void *(*)(void *))worker_thread, wq);
698 }
699
700 sigset(SIGINT, handle_sig);
701 sigset(SIGQUIT, handle_sig);
702 sigset(SIGTERM, handle_sig);
703 pthread_sigmask(SIG_UNBLOCK, &sets, NULL);
704 }
705
706 static void
707 join_threads(workqueue_t *wq)
708 {
709 int i;
710
711 for (i = 0; i < wq->wq_nthreads; i++) {
712 pthread_join(wq->wq_thread[i], NULL);
713 }
714 }
715
716 static int
717 strcompare(const void *p1, const void *p2)
718 {
719 char *s1 = *((char **)p1);
720 char *s2 = *((char **)p2);
721
722 return (strcmp(s1, s2));
723 }
724
725 /*
726 * Core work queue structure; passed to worker threads on thread creation
727 * as the main point of coordination. Allocate as a static structure; we
728 * could have put this into a local variable in main, but passing a pointer
729 * into your stack to another thread is fragile at best and leads to some
730 * hard-to-debug failure modes.
731 */
732 static workqueue_t wq;
733
734 int
735 main(int argc, char **argv)
736 {
737 tdata_t *mstrtd, *savetd;
738 char *uniqfile = NULL, *uniqlabel = NULL;
739 char *withfile = NULL;
740 char *label = NULL;
741 char **ifiles, **tifiles;
742 int verbose = 0, docopy = 0;
743 int write_fuzzy_match = 0;
744 int keep_stabs = 0;
745 int require_ctf = 0;
746 int nifiles, nielems;
747 int c, i, idx, tidx, err;
748
749 progname = basename(argv[0]);
750
751 if (getenv("CTFMERGE_DEBUG_LEVEL"))
752 debug_level = atoi(getenv("CTFMERGE_DEBUG_LEVEL"));
753
754 err = 0;
755 while ((c = getopt(argc, argv, ":cd:D:fgl:L:o:tvw:s")) != EOF) {
756 switch (c) {
757 case 'c':
758 docopy = 1;
759 break;
760 case 'd':
761 /* Uniquify against `uniqfile' */
762 uniqfile = optarg;
763 break;
764 case 'D':
765 /* Uniquify against label `uniqlabel' in `uniqfile' */
766 uniqlabel = optarg;
767 break;
768 case 'f':
769 write_fuzzy_match = CTF_FUZZY_MATCH;
770 break;
771 case 'g':
772 keep_stabs = CTF_KEEP_STABS;
773 break;
774 case 'l':
775 /* Label merged types with `label' */
776 label = optarg;
777 break;
778 case 'L':
779 /* Label merged types with getenv(`label`) */
780 if ((label = getenv(optarg)) == NULL)
781 label = CTF_DEFAULT_LABEL;
782 break;
783 case 'o':
784 /* Place merged types in CTF section in `outfile' */
785 outfile = optarg;
786 break;
787 case 't':
788 /* Insist *all* object files built from C have CTF */
789 require_ctf = 1;
790 break;
791 case 'v':
792 /* More debugging information */
793 verbose = 1;
794 break;
795 case 'w':
796 /* Additive merge with data from `withfile' */
797 withfile = optarg;
798 break;
799 case 's':
800 /* use the dynsym rather than the symtab */
801 dynsym = CTF_USE_DYNSYM;
802 break;
803 default:
804 usage();
805 exit(2);
806 }
807 }
808
809 /* Validate arguments */
810 if (docopy) {
811 if (uniqfile != NULL || uniqlabel != NULL || label != NULL ||
812 outfile != NULL || withfile != NULL || dynsym != 0)
813 err++;
814
815 if (argc - optind != 2)
816 err++;
817 } else {
818 if (uniqfile != NULL && withfile != NULL)
819 err++;
820
821 if (uniqlabel != NULL && uniqfile == NULL)
822 err++;
823
824 if (outfile == NULL || label == NULL)
825 err++;
826
827 if (argc - optind == 0)
828 err++;
829 }
830
831 if (err) {
832 usage();
833 exit(2);
834 }
835
836 if (getenv("STRIPSTABS_KEEP_STABS") != NULL)
837 keep_stabs = CTF_KEEP_STABS;
838
839 if (uniqfile && access(uniqfile, R_OK) != 0) {
840 warning("Uniquification file %s couldn't be opened and "
841 "will be ignored.\n", uniqfile);
842 uniqfile = NULL;
843 }
844 if (withfile && access(withfile, R_OK) != 0) {
845 warning("With file %s couldn't be opened and will be "
846 "ignored.\n", withfile);
847 withfile = NULL;
848 }
849 if (outfile && access(outfile, R_OK|W_OK) != 0)
850 terminate("Cannot open output file %s for r/w", outfile);
851
852 /*
853 * This is ugly, but we don't want to have to have a separate tool
854 * (yet) just for copying an ELF section with our specific requirements,
855 * so we shoe-horn a copier into ctfmerge.
856 */
857 if (docopy) {
858 copy_ctf_data(argv[optind], argv[optind + 1], keep_stabs);
859
860 exit(0);
861 }
862
863 set_terminate_cleanup(terminate_cleanup);
864
865 /* Sort the input files and strip out duplicates */
866 nifiles = argc - optind;
867 ifiles = xmalloc(sizeof (char *) * nifiles);
868 tifiles = xmalloc(sizeof (char *) * nifiles);
869
870 for (i = 0; i < nifiles; i++)
871 tifiles[i] = argv[optind + i];
872 qsort(tifiles, nifiles, sizeof (char *), (int (*)())strcompare);
873
874 ifiles[0] = tifiles[0];
875 for (idx = 0, tidx = 1; tidx < nifiles; tidx++) {
876 if (strcmp(ifiles[idx], tifiles[tidx]) != 0)
877 ifiles[++idx] = tifiles[tidx];
878 }
879 nifiles = idx + 1;
880
881 /* Make sure they all exist */
882 if ((nielems = count_files(ifiles, nifiles)) < 0)
883 terminate("Some input files were inaccessible\n");
884
885 /* Prepare for the merge */
886 wq_init(&wq, nielems);
887
888 start_threads(&wq);
889
890 /*
891 * Start the merge
892 *
893 * We're reading everything from each of the object files, so we
894 * don't need to specify labels.
895 */
896 if (read_ctf(ifiles, nifiles, NULL, merge_ctf_cb,
897 &wq, require_ctf) == 0) {
898 /*
899 * If we're verifying that C files have CTF, it's safe to
900 * assume that in this case, we're building only from assembly
901 * inputs.
902 */
903 if (require_ctf)
904 exit(0);
905 terminate("No ctf sections found to merge\n");
906 }
907
908 pthread_mutex_lock(&wq.wq_queue_lock);
909 wq.wq_nomorefiles = 1;
910 pthread_cond_broadcast(&wq.wq_work_avail);
911 pthread_mutex_unlock(&wq.wq_queue_lock);
912
913 pthread_mutex_lock(&wq.wq_queue_lock);
914 while (wq.wq_alldone == 0)
915 pthread_cond_wait(&wq.wq_alldone_cv, &wq.wq_queue_lock);
916 pthread_mutex_unlock(&wq.wq_queue_lock);
917
918 join_threads(&wq);
919
920 /*
921 * All requested files have been merged, with the resulting tree in
922 * mstrtd. savetd is the tree that will be placed into the output file.
923 *
924 * Regardless of whether we're doing a normal uniquification or an
925 * additive merge, we need a type tree that has been uniquified
926 * against uniqfile or withfile, as appropriate.
927 *
928 * If we're doing a uniquification, we stuff the resulting tree into
929 * outfile. Otherwise, we add the tree to the tree already in withfile.
930 */
931 assert(fifo_len(wq.wq_queue) == 1);
932 mstrtd = fifo_remove(wq.wq_queue);
933
934 if (verbose || debug_level) {
935 debug(2, "Statistics for td %p\n", (void *)mstrtd);
936
937 iidesc_stats(mstrtd->td_iihash);
938 }
939
940 if (uniqfile != NULL || withfile != NULL) {
941 char *reffile, *reflabel = NULL;
942 tdata_t *reftd;
943
944 if (uniqfile != NULL) {
945 reffile = uniqfile;
946 reflabel = uniqlabel;
947 } else
948 reffile = withfile;
949
950 if (read_ctf(&reffile, 1, reflabel, read_ctf_save_cb,
951 &reftd, require_ctf) == 0) {
952 terminate("No CTF data found in reference file %s\n",
953 reffile);
954 }
955
956 savetd = tdata_new();
957
958 if (CTF_TYPE_ISCHILD(reftd->td_nextid))
959 terminate("No room for additional types in master\n");
960
961 savetd->td_nextid = withfile ? reftd->td_nextid :
962 CTF_INDEX_TO_TYPE(1, TRUE);
963 merge_into_master(mstrtd, reftd, savetd, 0);
964
965 tdata_label_add(savetd, label, CTF_LABEL_LASTIDX);
966
967 if (withfile) {
968 /*
969 * savetd holds the new data to be added to the withfile
970 */
971 tdata_t *withtd = reftd;
972
973 tdata_merge(withtd, savetd);
974
975 savetd = withtd;
976 } else {
977 char uniqname[MAXPATHLEN];
978 labelent_t *parle;
979
980 parle = tdata_label_top(reftd);
981
982 savetd->td_parlabel = xstrdup(parle->le_name);
983
984 strncpy(uniqname, reffile, sizeof (uniqname));
985 uniqname[MAXPATHLEN - 1] = '\0';
986 savetd->td_parname = xstrdup(basename(uniqname));
987 }
988
989 } else {
990 /*
991 * No post processing. Write the merged tree as-is into the
992 * output file.
993 */
994 tdata_label_free(mstrtd);
995 tdata_label_add(mstrtd, label, CTF_LABEL_LASTIDX);
996
997 savetd = mstrtd;
998 }
999
1000 tmpname = mktmpname(outfile, ".ctf");
1001 write_ctf(savetd, outfile, tmpname,
1002 CTF_COMPRESS | write_fuzzy_match | dynsym | keep_stabs);
1003 if (rename(tmpname, outfile) != 0)
1004 terminate("Couldn't rename output temp file %s", tmpname);
1005 free(tmpname);
1006
1007 return (0);
1008 }