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.
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.
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]
22 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #pragma ident "%Z%%M% %I% %E% SMI"
29 * Given several files containing CTF data, merge and uniquify that data into
30 * a single CTF section in an output file.
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.
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.
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.
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.
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
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).
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]
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
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
98 * Locking Semantics (Phase I)
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.
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.
114 * When Phase II begins, the queue consists of the merged batches from the
115 * first phase. Assume we have five batches:
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:
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
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.
136 * Locking Semantics (Phase II)
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.
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.
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
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.
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.
189 #include <sys/param.h>
190 #include <sys/types.h>
191 #include <sys/mman.h>
193 #include <sys/sysconf.h>
196 #include "ctf_headers.h"
197 #include "ctftools.h"
198 #include "ctfmerge.h"
199 #include "traverse.h"
204 #pragma init(bigheap)
206 #define MERGE_PHASE1_BATCH_SIZE 8
207 #define MERGE_PHASE1_MAX_SLOTS 5
208 #define MERGE_INPUT_THROTTLE_LEN 10
210 const char *progname;
211 static char *outfile = NULL;
212 static char *tmpname = NULL;
214 int debug_level = DEBUG_LEVEL;
215 static size_t maxpgsize = 0x400000;
221 (void) fprintf(stderr,
222 "Usage: %s [-fgstv] -l label | -L labelenv -o outfile file ...\n"
223 " %s [-fgstv] -l label | -L labelenv -o outfile -d uniqfile\n"
224 " %*s [-g] [-D uniqlabel] file ...\n"
225 " %s [-fgstv] -l label | -L labelenv -o outfile -w withfile "
227 " %s [-g] -c srcfile destfile\n"
229 " Note: if -L labelenv is specified and labelenv is not set in\n"
230 " the environment, a default value is used.\n",
231 progname, progname, (int)strlen(progname), " ",
241 struct memcntl_mha mha;
244 * First, get the available pagesizes.
246 if ((sizes = getpagesizes(NULL, 0)) == -1)
249 if (sizes == 1 || (size = alloca(sizeof (size_t) * sizes)) == NULL)
252 if (getpagesizes(size, sizes) == -1)
255 while (size[sizes - 1] > maxpgsize)
258 /* set big to the largest allowed page size */
259 big = size[sizes - 1];
260 if (big & (big - 1)) {
262 * The largest page size is not a power of two for some
263 * inexplicable reason; return.
269 * Now, align our break to the largest page size.
271 if (brk((void *)((((uintptr_t)sbrk(0) - 1) & ~(big - 1)) + big)) != 0)
275 * set the preferred page size for the heap
277 mha.mha_cmd = MHA_MAPSIZE_BSSBRK;
279 mha.mha_pagesize = big;
281 (void) memcntl(NULL, 0, MC_HAT_ADVISE, (caddr_t)&mha, 0, 0);
286 finalize_phase_one(workqueue_t *wq)
291 * wip slots are cleared out only when maxbatchsz td's have been merged
292 * into them. We're not guaranteed that the number of files we're
293 * merging is a multiple of maxbatchsz, so there will be some partial
294 * groups in the wip array. Move them to the done queue in batch ID
295 * order, starting with the slot containing the next batch that would
296 * have been placed on the done queue, followed by the others.
297 * One thread will be doing this while the others wait at the barrier
298 * back in worker_thread(), so we don't need to worry about pesky things
302 for (startslot = -1, i = 0; i < wq->wq_nwipslots; i++) {
303 if (wq->wq_wip[i].wip_batchid == wq->wq_lastdonebatch + 1) {
309 assert(startslot != -1);
311 for (i = startslot; i < startslot + wq->wq_nwipslots; i++) {
312 int slotnum = i % wq->wq_nwipslots;
313 wip_t *wipslot = &wq->wq_wip[slotnum];
315 if (wipslot->wip_td != NULL) {
316 debug(2, "clearing slot %d (%d) (saving %d)\n",
317 slotnum, i, wipslot->wip_nmerged);
319 debug(2, "clearing slot %d (%d)\n", slotnum, i);
321 if (wipslot->wip_td != NULL) {
322 fifo_add(wq->wq_donequeue, wipslot->wip_td);
323 wq->wq_wip[slotnum].wip_td = NULL;
327 wq->wq_lastdonebatch = wq->wq_next_batchid++;
329 debug(2, "phase one done: donequeue has %d items\n",
330 fifo_len(wq->wq_donequeue));
334 init_phase_two(workqueue_t *wq)
339 * We're going to continually merge the first two entries on the queue,
340 * placing the result on the end, until there's nothing left to merge.
341 * At that point, everything will have been merged into one. The
342 * initial value of ninqueue needs to be equal to the total number of
343 * entries that will show up on the queue, both at the start of the
344 * phase and as generated by merges during the phase.
346 wq->wq_ninqueue = num = fifo_len(wq->wq_donequeue);
348 wq->wq_ninqueue += num / 2;
349 num = num / 2 + num % 2;
353 * Move the done queue to the work queue. We won't be using the done
356 assert(fifo_len(wq->wq_queue) == 0);
357 fifo_free(wq->wq_queue, NULL);
358 wq->wq_queue = wq->wq_donequeue;
362 wip_save_work(workqueue_t *wq, wip_t *slot, int slotnum)
364 pthread_mutex_lock(&wq->wq_donequeue_lock);
366 while (wq->wq_lastdonebatch + 1 < slot->wip_batchid)
367 pthread_cond_wait(&slot->wip_cv, &wq->wq_donequeue_lock);
368 assert(wq->wq_lastdonebatch + 1 == slot->wip_batchid);
370 fifo_add(wq->wq_donequeue, slot->wip_td);
371 wq->wq_lastdonebatch++;
372 pthread_cond_signal(&wq->wq_wip[(slotnum + 1) %
373 wq->wq_nwipslots].wip_cv);
375 /* reset the slot for next use */
377 slot->wip_batchid = wq->wq_next_batchid++;
379 pthread_mutex_unlock(&wq->wq_donequeue_lock);
383 wip_add_work(wip_t *slot, tdata_t *pow)
385 if (slot->wip_td == NULL) {
387 slot->wip_nmerged = 1;
389 debug(2, "%d: merging %p into %p\n", pthread_self(),
390 (void *)pow, (void *)slot->wip_td);
392 merge_into_master(pow, slot->wip_td, NULL, 0);
400 worker_runphase1(workqueue_t *wq)
404 int wipslotnum, pownum;
407 pthread_mutex_lock(&wq->wq_queue_lock);
409 while (fifo_empty(wq->wq_queue)) {
410 if (wq->wq_nomorefiles == 1) {
411 pthread_cond_broadcast(&wq->wq_work_avail);
412 pthread_mutex_unlock(&wq->wq_queue_lock);
414 /* on to phase 2 ... */
418 pthread_cond_wait(&wq->wq_work_avail,
422 /* there's work to be done! */
423 pow = fifo_remove(wq->wq_queue);
424 pownum = wq->wq_nextpownum++;
425 pthread_cond_broadcast(&wq->wq_work_removed);
429 /* merge it into the right slot */
430 wipslotnum = pownum % wq->wq_nwipslots;
431 wipslot = &wq->wq_wip[wipslotnum];
433 pthread_mutex_lock(&wipslot->wip_lock);
435 pthread_mutex_unlock(&wq->wq_queue_lock);
437 wip_add_work(wipslot, pow);
439 if (wipslot->wip_nmerged == wq->wq_maxbatchsz)
440 wip_save_work(wq, wipslot, wipslotnum);
442 pthread_mutex_unlock(&wipslot->wip_lock);
447 worker_runphase2(workqueue_t *wq)
449 tdata_t *pow1, *pow2;
453 pthread_mutex_lock(&wq->wq_queue_lock);
455 if (wq->wq_ninqueue == 1) {
456 pthread_cond_broadcast(&wq->wq_work_avail);
457 pthread_mutex_unlock(&wq->wq_queue_lock);
459 debug(2, "%d: entering p2 completion barrier\n",
461 if (barrier_wait(&wq->wq_bar1)) {
462 pthread_mutex_lock(&wq->wq_queue_lock);
464 pthread_cond_signal(&wq->wq_alldone_cv);
465 pthread_mutex_unlock(&wq->wq_queue_lock);
471 if (fifo_len(wq->wq_queue) < 2) {
472 pthread_cond_wait(&wq->wq_work_avail,
474 pthread_mutex_unlock(&wq->wq_queue_lock);
478 /* there's work to be done! */
479 pow1 = fifo_remove(wq->wq_queue);
480 pow2 = fifo_remove(wq->wq_queue);
481 wq->wq_ninqueue -= 2;
483 batchid = wq->wq_next_batchid++;
485 pthread_mutex_unlock(&wq->wq_queue_lock);
487 debug(2, "%d: merging %p into %p\n", pthread_self(),
488 (void *)pow1, (void *)pow2);
489 merge_into_master(pow1, pow2, NULL, 0);
493 * merging is complete. place at the tail of the queue in
496 pthread_mutex_lock(&wq->wq_queue_lock);
497 while (wq->wq_lastdonebatch + 1 != batchid) {
498 pthread_cond_wait(&wq->wq_done_cv,
502 wq->wq_lastdonebatch = batchid;
504 fifo_add(wq->wq_queue, pow2);
505 debug(2, "%d: added %p to queue, len now %d, ninqueue %d\n",
506 pthread_self(), (void *)pow2, fifo_len(wq->wq_queue),
508 pthread_cond_broadcast(&wq->wq_done_cv);
509 pthread_cond_signal(&wq->wq_work_avail);
510 pthread_mutex_unlock(&wq->wq_queue_lock);
515 * Main loop for worker threads.
518 worker_thread(workqueue_t *wq)
520 worker_runphase1(wq);
522 debug(2, "%d: entering first barrier\n", pthread_self());
524 if (barrier_wait(&wq->wq_bar1)) {
526 debug(2, "%d: doing work in first barrier\n", pthread_self());
528 finalize_phase_one(wq);
532 debug(2, "%d: ninqueue is %d, %d on queue\n", pthread_self(),
533 wq->wq_ninqueue, fifo_len(wq->wq_queue));
536 debug(2, "%d: entering second barrier\n", pthread_self());
538 (void) barrier_wait(&wq->wq_bar2);
540 debug(2, "%d: phase 1 complete\n", pthread_self());
542 worker_runphase2(wq);
546 * Pass a tdata_t tree, built from an input file, off to the work queue for
547 * consumption by worker threads.
550 merge_ctf_cb(tdata_t *td, char *name, void *arg)
552 workqueue_t *wq = arg;
554 debug(3, "Adding tdata %p for processing\n", (void *)td);
556 pthread_mutex_lock(&wq->wq_queue_lock);
557 while (fifo_len(wq->wq_queue) > wq->wq_ithrottle) {
558 debug(2, "Throttling input (len = %d, throttle = %d)\n",
559 fifo_len(wq->wq_queue), wq->wq_ithrottle);
560 pthread_cond_wait(&wq->wq_work_removed, &wq->wq_queue_lock);
563 fifo_add(wq->wq_queue, td);
564 debug(1, "Thread %d announcing %s\n", pthread_self(), name);
565 pthread_cond_broadcast(&wq->wq_work_avail);
566 pthread_mutex_unlock(&wq->wq_queue_lock);
572 * This program is intended to be invoked from a Makefile, as part of the build.
573 * As such, in the event of a failure or user-initiated interrupt (^C), we need
574 * to ensure that a subsequent re-make will cause ctfmerge to be executed again.
575 * Unfortunately, ctfmerge will usually be invoked directly after (and as part
576 * of the same Makefile rule as) a link, and will operate on the linked file
577 * in place. If we merely exit upon receipt of a SIGINT, a subsequent make
578 * will notice that the *linked* file is newer than the object files, and thus
579 * will not reinvoke ctfmerge. The only way to ensure that a subsequent make
580 * reinvokes ctfmerge, is to remove the file to which we are adding CTF
581 * data (confusingly named the output file). This means that the link will need
582 * to happen again, but links are generally fast, and we can't allow the merge
585 * Another possibility would be to block SIGINT entirely - to always run to
586 * completion. The run time of ctfmerge can, however, be measured in minutes
587 * in some cases, so this is not a valid option.
592 terminate("Caught signal %d - exiting\n", sig);
596 terminate_cleanup(void)
598 int dounlink = getenv("CTFMERGE_TERMINATE_NO_UNLINK") ? 0 : 1;
600 if (tmpname != NULL && dounlink)
606 #if !defined(__FreeBSD__)
608 fprintf(stderr, "Removing %s\n", outfile);
615 copy_ctf_data(char *srcfile, char *destfile, int keep_stabs)
619 if (read_ctf(&srcfile, 1, NULL, read_ctf_save_cb, &srctd, 1) == 0)
620 terminate("No CTF data found in source file %s\n", srcfile);
622 tmpname = mktmpname(destfile, ".ctf");
623 write_ctf(srctd, destfile, tmpname, CTF_COMPRESS | CTF_SWAP_BYTES | keep_stabs);
624 if (rename(tmpname, destfile) != 0) {
625 terminate("Couldn't rename temp file %s to %s", tmpname,
633 wq_init(workqueue_t *wq, int nfiles)
635 int throttle, nslots, i;
637 if (getenv("CTFMERGE_MAX_SLOTS"))
638 nslots = atoi(getenv("CTFMERGE_MAX_SLOTS"));
640 nslots = MERGE_PHASE1_MAX_SLOTS;
642 if (getenv("CTFMERGE_PHASE1_BATCH_SIZE"))
643 wq->wq_maxbatchsz = atoi(getenv("CTFMERGE_PHASE1_BATCH_SIZE"));
645 wq->wq_maxbatchsz = MERGE_PHASE1_BATCH_SIZE;
647 nslots = MIN(nslots, (nfiles + wq->wq_maxbatchsz - 1) /
650 wq->wq_wip = xcalloc(sizeof (wip_t) * nslots);
651 wq->wq_nwipslots = nslots;
652 wq->wq_nthreads = MIN(sysconf(_SC_NPROCESSORS_ONLN) * 3 / 2, nslots);
653 wq->wq_thread = xmalloc(sizeof (pthread_t) * wq->wq_nthreads);
655 if (getenv("CTFMERGE_INPUT_THROTTLE"))
656 throttle = atoi(getenv("CTFMERGE_INPUT_THROTTLE"));
658 throttle = MERGE_INPUT_THROTTLE_LEN;
659 wq->wq_ithrottle = throttle * wq->wq_nthreads;
661 debug(1, "Using %d slots, %d threads\n", wq->wq_nwipslots,
664 wq->wq_next_batchid = 0;
666 for (i = 0; i < nslots; i++) {
667 pthread_mutex_init(&wq->wq_wip[i].wip_lock, NULL);
668 wq->wq_wip[i].wip_batchid = wq->wq_next_batchid++;
671 pthread_mutex_init(&wq->wq_queue_lock, NULL);
672 wq->wq_queue = fifo_new();
673 pthread_cond_init(&wq->wq_work_avail, NULL);
674 pthread_cond_init(&wq->wq_work_removed, NULL);
675 wq->wq_ninqueue = nfiles;
676 wq->wq_nextpownum = 0;
678 pthread_mutex_init(&wq->wq_donequeue_lock, NULL);
679 wq->wq_donequeue = fifo_new();
680 wq->wq_lastdonebatch = -1;
682 pthread_cond_init(&wq->wq_done_cv, NULL);
684 pthread_cond_init(&wq->wq_alldone_cv, NULL);
687 barrier_init(&wq->wq_bar1, wq->wq_nthreads);
688 barrier_init(&wq->wq_bar2, wq->wq_nthreads);
690 wq->wq_nomorefiles = 0;
694 start_threads(workqueue_t *wq)
700 sigaddset(&sets, SIGINT);
701 sigaddset(&sets, SIGQUIT);
702 sigaddset(&sets, SIGTERM);
703 pthread_sigmask(SIG_BLOCK, &sets, NULL);
705 for (i = 0; i < wq->wq_nthreads; i++) {
706 pthread_create(&wq->wq_thread[i], NULL,
707 (void *(*)(void *))worker_thread, wq);
711 sigset(SIGINT, handle_sig);
712 sigset(SIGQUIT, handle_sig);
713 sigset(SIGTERM, handle_sig);
715 signal(SIGINT, handle_sig);
716 signal(SIGQUIT, handle_sig);
717 signal(SIGTERM, handle_sig);
719 pthread_sigmask(SIG_UNBLOCK, &sets, NULL);
723 join_threads(workqueue_t *wq)
727 for (i = 0; i < wq->wq_nthreads; i++) {
728 pthread_join(wq->wq_thread[i], NULL);
733 strcompare(const void *p1, const void *p2)
735 char *s1 = *((char **)p1);
736 char *s2 = *((char **)p2);
738 return (strcmp(s1, s2));
742 * Core work queue structure; passed to worker threads on thread creation
743 * as the main point of coordination. Allocate as a static structure; we
744 * could have put this into a local variable in main, but passing a pointer
745 * into your stack to another thread is fragile at best and leads to some
746 * hard-to-debug failure modes.
748 static workqueue_t wq;
751 main(int argc, char **argv)
753 tdata_t *mstrtd, *savetd;
754 char *uniqfile = NULL, *uniqlabel = NULL;
755 char *withfile = NULL;
757 char **ifiles, **tifiles;
758 int verbose = 0, docopy = 0;
759 int write_fuzzy_match = 0;
762 int nifiles, nielems;
763 int c, i, idx, tidx, err;
765 progname = basename(argv[0]);
767 if (getenv("CTFMERGE_DEBUG_LEVEL"))
768 debug_level = atoi(getenv("CTFMERGE_DEBUG_LEVEL"));
771 while ((c = getopt(argc, argv, ":cd:D:fgl:L:o:tvw:s")) != EOF) {
777 /* Uniquify against `uniqfile' */
781 /* Uniquify against label `uniqlabel' in `uniqfile' */
785 write_fuzzy_match = CTF_FUZZY_MATCH;
788 keep_stabs = CTF_KEEP_STABS;
791 /* Label merged types with `label' */
795 /* Label merged types with getenv(`label`) */
796 if ((label = getenv(optarg)) == NULL)
797 label = CTF_DEFAULT_LABEL;
800 /* Place merged types in CTF section in `outfile' */
804 /* Insist *all* object files built from C have CTF */
808 /* More debugging information */
812 /* Additive merge with data from `withfile' */
816 /* use the dynsym rather than the symtab */
817 dynsym = CTF_USE_DYNSYM;
825 /* Validate arguments */
827 if (uniqfile != NULL || uniqlabel != NULL || label != NULL ||
828 outfile != NULL || withfile != NULL || dynsym != 0)
831 if (argc - optind != 2)
834 if (uniqfile != NULL && withfile != NULL)
837 if (uniqlabel != NULL && uniqfile == NULL)
840 if (outfile == NULL || label == NULL)
843 if (argc - optind == 0)
852 if (getenv("STRIPSTABS_KEEP_STABS") != NULL)
853 keep_stabs = CTF_KEEP_STABS;
855 if (uniqfile && access(uniqfile, R_OK) != 0) {
856 warning("Uniquification file %s couldn't be opened and "
857 "will be ignored.\n", uniqfile);
860 if (withfile && access(withfile, R_OK) != 0) {
861 warning("With file %s couldn't be opened and will be "
862 "ignored.\n", withfile);
865 if (outfile && access(outfile, R_OK|W_OK) != 0)
866 terminate("Cannot open output file %s for r/w", outfile);
869 * This is ugly, but we don't want to have to have a separate tool
870 * (yet) just for copying an ELF section with our specific requirements,
871 * so we shoe-horn a copier into ctfmerge.
874 copy_ctf_data(argv[optind], argv[optind + 1], keep_stabs);
879 set_terminate_cleanup(terminate_cleanup);
881 /* Sort the input files and strip out duplicates */
882 nifiles = argc - optind;
883 ifiles = xmalloc(sizeof (char *) * nifiles);
884 tifiles = xmalloc(sizeof (char *) * nifiles);
886 for (i = 0; i < nifiles; i++)
887 tifiles[i] = argv[optind + i];
888 qsort(tifiles, nifiles, sizeof (char *), (int (*)())strcompare);
890 ifiles[0] = tifiles[0];
891 for (idx = 0, tidx = 1; tidx < nifiles; tidx++) {
892 if (strcmp(ifiles[idx], tifiles[tidx]) != 0)
893 ifiles[++idx] = tifiles[tidx];
897 /* Make sure they all exist */
898 if ((nielems = count_files(ifiles, nifiles)) < 0)
899 terminate("Some input files were inaccessible\n");
901 /* Prepare for the merge */
902 wq_init(&wq, nielems);
909 * We're reading everything from each of the object files, so we
910 * don't need to specify labels.
912 if (read_ctf(ifiles, nifiles, NULL, merge_ctf_cb,
913 &wq, require_ctf) == 0) {
915 * If we're verifying that C files have CTF, it's safe to
916 * assume that in this case, we're building only from assembly
921 terminate("No ctf sections found to merge\n");
924 pthread_mutex_lock(&wq.wq_queue_lock);
925 wq.wq_nomorefiles = 1;
926 pthread_cond_broadcast(&wq.wq_work_avail);
927 pthread_mutex_unlock(&wq.wq_queue_lock);
929 pthread_mutex_lock(&wq.wq_queue_lock);
930 while (wq.wq_alldone == 0)
931 pthread_cond_wait(&wq.wq_alldone_cv, &wq.wq_queue_lock);
932 pthread_mutex_unlock(&wq.wq_queue_lock);
937 * All requested files have been merged, with the resulting tree in
938 * mstrtd. savetd is the tree that will be placed into the output file.
940 * Regardless of whether we're doing a normal uniquification or an
941 * additive merge, we need a type tree that has been uniquified
942 * against uniqfile or withfile, as appropriate.
944 * If we're doing a uniquification, we stuff the resulting tree into
945 * outfile. Otherwise, we add the tree to the tree already in withfile.
947 assert(fifo_len(wq.wq_queue) == 1);
948 mstrtd = fifo_remove(wq.wq_queue);
950 if (verbose || debug_level) {
951 debug(2, "Statistics for td %p\n", (void *)mstrtd);
953 iidesc_stats(mstrtd->td_iihash);
956 if (uniqfile != NULL || withfile != NULL) {
957 char *reffile, *reflabel = NULL;
960 if (uniqfile != NULL) {
962 reflabel = uniqlabel;
966 if (read_ctf(&reffile, 1, reflabel, read_ctf_save_cb,
967 &reftd, require_ctf) == 0) {
968 terminate("No CTF data found in reference file %s\n",
972 savetd = tdata_new();
974 if (CTF_TYPE_ISCHILD(reftd->td_nextid))
975 terminate("No room for additional types in master\n");
977 savetd->td_nextid = withfile ? reftd->td_nextid :
978 CTF_INDEX_TO_TYPE(1, TRUE);
979 merge_into_master(mstrtd, reftd, savetd, 0);
981 tdata_label_add(savetd, label, CTF_LABEL_LASTIDX);
985 * savetd holds the new data to be added to the withfile
987 tdata_t *withtd = reftd;
989 tdata_merge(withtd, savetd);
993 char uniqname[MAXPATHLEN];
996 parle = tdata_label_top(reftd);
998 savetd->td_parlabel = xstrdup(parle->le_name);
1000 strncpy(uniqname, reffile, sizeof (uniqname));
1001 uniqname[MAXPATHLEN - 1] = '\0';
1002 savetd->td_parname = xstrdup(basename(uniqname));
1007 * No post processing. Write the merged tree as-is into the
1010 tdata_label_free(mstrtd);
1011 tdata_label_add(mstrtd, label, CTF_LABEL_LASTIDX);
1016 tmpname = mktmpname(outfile, ".ctf");
1017 write_ctf(savetd, outfile, tmpname,
1018 CTF_COMPRESS | CTF_SWAP_BYTES | write_fuzzy_match | dynsym | keep_stabs);
1019 if (rename(tmpname, outfile) != 0)
1020 terminate("Couldn't rename output temp file %s", tmpname);