xref: /freebsd/cddl/contrib/opensolaris/tools/ctf/cvt/ctfmerge.c (revision 95ca89cda1a6c4e0ef0b3f765c6563f1db0d23fa)
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 #ifdef illumos
180 #include <synch.h>
181 #endif
182 #include <signal.h>
183 #include <libgen.h>
184 #include <string.h>
185 #include <errno.h>
186 #ifdef illumos
187 #include <alloca.h>
188 #endif
189 #include <sys/param.h>
190 #include <sys/types.h>
191 #include <sys/mman.h>
192 #ifdef illumos
193 #include <sys/sysconf.h>
194 #endif
195 
196 #include "ctf_headers.h"
197 #include "ctftools.h"
198 #include "ctfmerge.h"
199 #include "traverse.h"
200 #include "memory.h"
201 #include "fifo.h"
202 #include "barrier.h"
203 
204 #pragma init(bigheap)
205 
206 #define	MERGE_PHASE1_BATCH_SIZE		8
207 #define	MERGE_PHASE1_MAX_SLOTS		5
208 #define	MERGE_INPUT_THROTTLE_LEN	10
209 
210 const char *progname;
211 static char *outfile = NULL;
212 static char *tmpname = NULL;
213 static int dynsym;
214 int debug_level = DEBUG_LEVEL;
215 static size_t maxpgsize = 0x400000;
216 
217 
218 void
usage(void)219 usage(void)
220 {
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 "
226 	    "file ...\n"
227 	    "       %s [-g] -c srcfile destfile\n"
228 	    "\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), " ",
232 	    progname, progname);
233 }
234 
235 #ifdef illumos
236 static void
bigheap(void)237 bigheap(void)
238 {
239 	size_t big, *size;
240 	int sizes;
241 	struct memcntl_mha mha;
242 
243 	/*
244 	 * First, get the available pagesizes.
245 	 */
246 	if ((sizes = getpagesizes(NULL, 0)) == -1)
247 		return;
248 
249 	if (sizes == 1 || (size = alloca(sizeof (size_t) * sizes)) == NULL)
250 		return;
251 
252 	if (getpagesizes(size, sizes) == -1)
253 		return;
254 
255 	while (size[sizes - 1] > maxpgsize)
256 		sizes--;
257 
258 	/* set big to the largest allowed page size */
259 	big = size[sizes - 1];
260 	if (big & (big - 1)) {
261 		/*
262 		 * The largest page size is not a power of two for some
263 		 * inexplicable reason; return.
264 		 */
265 		return;
266 	}
267 
268 	/*
269 	 * Now, align our break to the largest page size.
270 	 */
271 	if (brk((void *)((((uintptr_t)sbrk(0) - 1) & ~(big - 1)) + big)) != 0)
272 		return;
273 
274 	/*
275 	 * set the preferred page size for the heap
276 	 */
277 	mha.mha_cmd = MHA_MAPSIZE_BSSBRK;
278 	mha.mha_flags = 0;
279 	mha.mha_pagesize = big;
280 
281 	(void) memcntl(NULL, 0, MC_HAT_ADVISE, (caddr_t)&mha, 0, 0);
282 }
283 #endif	/* illumos */
284 
285 static void
finalize_phase_one(workqueue_t * wq)286 finalize_phase_one(workqueue_t *wq)
287 {
288 	int startslot, i;
289 
290 	/*
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
299 	 * like locks.
300 	 */
301 
302 	for (startslot = -1, i = 0; i < wq->wq_nwipslots; i++) {
303 		if (wq->wq_wip[i].wip_batchid == wq->wq_lastdonebatch + 1) {
304 			startslot = i;
305 			break;
306 		}
307 	}
308 
309 	assert(startslot != -1);
310 
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];
314 
315 		if (wipslot->wip_td != NULL) {
316 			debug(2, "clearing slot %d (%d) (saving %d)\n",
317 			    slotnum, i, wipslot->wip_nmerged);
318 		} else
319 			debug(2, "clearing slot %d (%d)\n", slotnum, i);
320 
321 		if (wipslot->wip_td != NULL) {
322 			fifo_add(wq->wq_donequeue, wipslot->wip_td);
323 			wq->wq_wip[slotnum].wip_td = NULL;
324 		}
325 	}
326 
327 	wq->wq_lastdonebatch = wq->wq_next_batchid++;
328 
329 	debug(2, "phase one done: donequeue has %d items\n",
330 	    fifo_len(wq->wq_donequeue));
331 }
332 
333 static void
init_phase_two(workqueue_t * wq)334 init_phase_two(workqueue_t *wq)
335 {
336 	int num;
337 
338 	/*
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.
345 	 */
346 	wq->wq_ninqueue = num = fifo_len(wq->wq_donequeue);
347 	while (num != 1) {
348 		wq->wq_ninqueue += num / 2;
349 		num = num / 2 + num % 2;
350 	}
351 
352 	/*
353 	 * Move the done queue to the work queue.  We won't be using the done
354 	 * queue in phase 2.
355 	 */
356 	assert(fifo_len(wq->wq_queue) == 0);
357 	fifo_free(wq->wq_queue, NULL);
358 	wq->wq_queue = wq->wq_donequeue;
359 }
360 
361 static void
wip_save_work(workqueue_t * wq,wip_t * slot,int slotnum)362 wip_save_work(workqueue_t *wq, wip_t *slot, int slotnum)
363 {
364 	pthread_mutex_lock(&wq->wq_donequeue_lock);
365 
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);
369 
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);
374 
375 	/* reset the slot for next use */
376 	slot->wip_td = NULL;
377 	slot->wip_batchid = wq->wq_next_batchid++;
378 
379 	pthread_mutex_unlock(&wq->wq_donequeue_lock);
380 }
381 
382 static void
wip_add_work(wip_t * slot,tdata_t * pow)383 wip_add_work(wip_t *slot, tdata_t *pow)
384 {
385 	if (slot->wip_td == NULL) {
386 		slot->wip_td = pow;
387 		slot->wip_nmerged = 1;
388 	} else {
389 		debug(2, "%d: merging %p into %p\n", pthread_self(),
390 		    (void *)pow, (void *)slot->wip_td);
391 
392 		merge_into_master(pow, slot->wip_td, NULL, 0);
393 		tdata_free(pow);
394 
395 		slot->wip_nmerged++;
396 	}
397 }
398 
399 static void
worker_runphase1(workqueue_t * wq)400 worker_runphase1(workqueue_t *wq)
401 {
402 	wip_t *wipslot;
403 	tdata_t *pow;
404 	int wipslotnum, pownum;
405 
406 	for (;;) {
407 		pthread_mutex_lock(&wq->wq_queue_lock);
408 
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);
413 
414 				/* on to phase 2 ... */
415 				return;
416 			}
417 
418 			pthread_cond_wait(&wq->wq_work_avail,
419 			    &wq->wq_queue_lock);
420 		}
421 
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);
426 
427 		assert(pow != NULL);
428 
429 		/* merge it into the right slot */
430 		wipslotnum = pownum % wq->wq_nwipslots;
431 		wipslot = &wq->wq_wip[wipslotnum];
432 
433 		pthread_mutex_lock(&wipslot->wip_lock);
434 
435 		pthread_mutex_unlock(&wq->wq_queue_lock);
436 
437 		wip_add_work(wipslot, pow);
438 
439 		if (wipslot->wip_nmerged == wq->wq_maxbatchsz)
440 			wip_save_work(wq, wipslot, wipslotnum);
441 
442 		pthread_mutex_unlock(&wipslot->wip_lock);
443 	}
444 }
445 
446 static void
worker_runphase2(workqueue_t * wq)447 worker_runphase2(workqueue_t *wq)
448 {
449 	tdata_t *pow1, *pow2;
450 	int batchid;
451 
452 	for (;;) {
453 		pthread_mutex_lock(&wq->wq_queue_lock);
454 
455 		if (wq->wq_ninqueue == 1) {
456 			pthread_cond_broadcast(&wq->wq_work_avail);
457 			pthread_mutex_unlock(&wq->wq_queue_lock);
458 
459 			debug(2, "%d: entering p2 completion barrier\n",
460 			    pthread_self());
461 			if (barrier_wait(&wq->wq_bar1)) {
462 				pthread_mutex_lock(&wq->wq_queue_lock);
463 				wq->wq_alldone = 1;
464 				pthread_cond_signal(&wq->wq_alldone_cv);
465 				pthread_mutex_unlock(&wq->wq_queue_lock);
466 			}
467 
468 			return;
469 		}
470 
471 		if (fifo_len(wq->wq_queue) < 2) {
472 			pthread_cond_wait(&wq->wq_work_avail,
473 			    &wq->wq_queue_lock);
474 			pthread_mutex_unlock(&wq->wq_queue_lock);
475 			continue;
476 		}
477 
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;
482 
483 		batchid = wq->wq_next_batchid++;
484 
485 		pthread_mutex_unlock(&wq->wq_queue_lock);
486 
487 		debug(2, "%d: merging %p into %p\n", pthread_self(),
488 		    (void *)pow1, (void *)pow2);
489 		merge_into_master(pow1, pow2, NULL, 0);
490 		tdata_free(pow1);
491 
492 		/*
493 		 * merging is complete.  place at the tail of the queue in
494 		 * proper order.
495 		 */
496 		pthread_mutex_lock(&wq->wq_queue_lock);
497 		while (wq->wq_lastdonebatch + 1 != batchid) {
498 			pthread_cond_wait(&wq->wq_done_cv,
499 			    &wq->wq_queue_lock);
500 		}
501 
502 		wq->wq_lastdonebatch = batchid;
503 
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),
507 		    wq->wq_ninqueue);
508 		pthread_cond_broadcast(&wq->wq_done_cv);
509 		pthread_cond_signal(&wq->wq_work_avail);
510 		pthread_mutex_unlock(&wq->wq_queue_lock);
511 	}
512 }
513 
514 /*
515  * Main loop for worker threads.
516  */
517 static void
worker_thread(workqueue_t * wq)518 worker_thread(workqueue_t *wq)
519 {
520 	worker_runphase1(wq);
521 
522 	debug(2, "%d: entering first barrier\n", pthread_self());
523 
524 	if (barrier_wait(&wq->wq_bar1)) {
525 
526 		debug(2, "%d: doing work in first barrier\n", pthread_self());
527 
528 		finalize_phase_one(wq);
529 
530 		init_phase_two(wq);
531 
532 		debug(2, "%d: ninqueue is %d, %d on queue\n", pthread_self(),
533 		    wq->wq_ninqueue, fifo_len(wq->wq_queue));
534 	}
535 
536 	debug(2, "%d: entering second barrier\n", pthread_self());
537 
538 	(void) barrier_wait(&wq->wq_bar2);
539 
540 	debug(2, "%d: phase 1 complete\n", pthread_self());
541 
542 	worker_runphase2(wq);
543 }
544 
545 /*
546  * Pass a tdata_t tree, built from an input file, off to the work queue for
547  * consumption by worker threads.
548  */
549 static int
merge_ctf_cb(tdata_t * td,char * name,void * arg)550 merge_ctf_cb(tdata_t *td, char *name, void *arg)
551 {
552 	workqueue_t *wq = arg;
553 
554 	debug(3, "Adding tdata %p for processing\n", (void *)td);
555 
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);
561 	}
562 
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);
567 
568 	return (1);
569 }
570 
571 /*
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
583  * to be skipped.
584  *
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.
588  */
589 static void
handle_sig(int sig)590 handle_sig(int sig)
591 {
592 	terminate("Caught signal %d - exiting\n", sig);
593 }
594 
595 static void
terminate_cleanup(void)596 terminate_cleanup(void)
597 {
598 	int dounlink = getenv("CTFMERGE_TERMINATE_NO_UNLINK") ? 0 : 1;
599 
600 	if (tmpname != NULL && dounlink)
601 		unlink(tmpname);
602 
603 	if (outfile == NULL)
604 		return;
605 
606 #if !defined(__FreeBSD__)
607 	if (dounlink) {
608 		fprintf(stderr, "Removing %s\n", outfile);
609 		unlink(outfile);
610 	}
611 #endif
612 }
613 
614 static void
copy_ctf_data(char * srcfile,char * destfile,int keep_stabs)615 copy_ctf_data(char *srcfile, char *destfile, int keep_stabs)
616 {
617 	tdata_t *srctd;
618 
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);
621 
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,
626 		    destfile);
627 	}
628 	free(tmpname);
629 	tdata_free(srctd);
630 }
631 
632 static void
wq_init(workqueue_t * wq,int nfiles)633 wq_init(workqueue_t *wq, int nfiles)
634 {
635 	int throttle, nslots, i;
636 
637 	if (getenv("CTFMERGE_MAX_SLOTS"))
638 		nslots = atoi(getenv("CTFMERGE_MAX_SLOTS"));
639 	else
640 		nslots = MERGE_PHASE1_MAX_SLOTS;
641 
642 	if (getenv("CTFMERGE_PHASE1_BATCH_SIZE"))
643 		wq->wq_maxbatchsz = atoi(getenv("CTFMERGE_PHASE1_BATCH_SIZE"));
644 	else
645 		wq->wq_maxbatchsz = MERGE_PHASE1_BATCH_SIZE;
646 
647 	nslots = MIN(nslots, (nfiles + wq->wq_maxbatchsz - 1) /
648 	    wq->wq_maxbatchsz);
649 
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);
654 
655 	if (getenv("CTFMERGE_INPUT_THROTTLE"))
656 		throttle = atoi(getenv("CTFMERGE_INPUT_THROTTLE"));
657 	else
658 		throttle = MERGE_INPUT_THROTTLE_LEN;
659 	wq->wq_ithrottle = throttle * wq->wq_nthreads;
660 
661 	debug(1, "Using %d slots, %d threads\n", wq->wq_nwipslots,
662 	    wq->wq_nthreads);
663 
664 	wq->wq_next_batchid = 0;
665 
666 	for (i = 0; i < nslots; i++) {
667 		pthread_mutex_init(&wq->wq_wip[i].wip_lock, NULL);
668 		pthread_cond_init(&wq->wq_wip[i].wip_cv, NULL);
669 		wq->wq_wip[i].wip_batchid = wq->wq_next_batchid++;
670 	}
671 
672 	pthread_mutex_init(&wq->wq_queue_lock, NULL);
673 	wq->wq_queue = fifo_new();
674 	pthread_cond_init(&wq->wq_work_avail, NULL);
675 	pthread_cond_init(&wq->wq_work_removed, NULL);
676 	wq->wq_ninqueue = nfiles;
677 	wq->wq_nextpownum = 0;
678 
679 	pthread_mutex_init(&wq->wq_donequeue_lock, NULL);
680 	wq->wq_donequeue = fifo_new();
681 	wq->wq_lastdonebatch = -1;
682 
683 	pthread_cond_init(&wq->wq_done_cv, NULL);
684 
685 	pthread_cond_init(&wq->wq_alldone_cv, NULL);
686 	wq->wq_alldone = 0;
687 
688 	barrier_init(&wq->wq_bar1, wq->wq_nthreads);
689 	barrier_init(&wq->wq_bar2, wq->wq_nthreads);
690 
691 	wq->wq_nomorefiles = 0;
692 }
693 
694 static void
start_threads(workqueue_t * wq)695 start_threads(workqueue_t *wq)
696 {
697 	sigset_t sets;
698 	int i;
699 
700 	sigemptyset(&sets);
701 	sigaddset(&sets, SIGINT);
702 	sigaddset(&sets, SIGQUIT);
703 	sigaddset(&sets, SIGTERM);
704 	pthread_sigmask(SIG_BLOCK, &sets, NULL);
705 
706 	for (i = 0; i < wq->wq_nthreads; i++) {
707 		pthread_create(&wq->wq_thread[i], NULL,
708 		    (void *(*)(void *))worker_thread, wq);
709 	}
710 
711 #ifdef illumos
712 	sigset(SIGINT, handle_sig);
713 	sigset(SIGQUIT, handle_sig);
714 	sigset(SIGTERM, handle_sig);
715 #else
716 	signal(SIGINT, handle_sig);
717 	signal(SIGQUIT, handle_sig);
718 	signal(SIGTERM, handle_sig);
719 #endif
720 	pthread_sigmask(SIG_UNBLOCK, &sets, NULL);
721 }
722 
723 static void
join_threads(workqueue_t * wq)724 join_threads(workqueue_t *wq)
725 {
726 	int i;
727 
728 	for (i = 0; i < wq->wq_nthreads; i++) {
729 		pthread_join(wq->wq_thread[i], NULL);
730 	}
731 }
732 
733 static int
strcompare(const void * p1,const void * p2)734 strcompare(const void *p1, const void *p2)
735 {
736 	char *s1 = *((char **)p1);
737 	char *s2 = *((char **)p2);
738 
739 	return (strcmp(s1, s2));
740 }
741 
742 /*
743  * Core work queue structure; passed to worker threads on thread creation
744  * as the main point of coordination.  Allocate as a static structure; we
745  * could have put this into a local variable in main, but passing a pointer
746  * into your stack to another thread is fragile at best and leads to some
747  * hard-to-debug failure modes.
748  */
749 static workqueue_t wq;
750 
751 int
main(int argc,char ** argv)752 main(int argc, char **argv)
753 {
754 	tdata_t *mstrtd, *savetd;
755 	char *uniqfile = NULL, *uniqlabel = NULL;
756 	char *withfile = NULL;
757 	char *label = NULL;
758 	char **ifiles, **tifiles;
759 	int verbose = 0, docopy = 0;
760 	int write_fuzzy_match = 0;
761 	int keep_stabs = 0;
762 	int require_ctf = 0;
763 	int nifiles, nielems;
764 	int c, i, idx, tidx, err;
765 
766 	progname = basename(argv[0]);
767 
768 	if (getenv("CTFMERGE_DEBUG_LEVEL"))
769 		debug_level = atoi(getenv("CTFMERGE_DEBUG_LEVEL"));
770 
771 	err = 0;
772 	while ((c = getopt(argc, argv, ":cd:D:fgl:L:o:tvw:s")) != EOF) {
773 		switch (c) {
774 		case 'c':
775 			docopy = 1;
776 			break;
777 		case 'd':
778 			/* Uniquify against `uniqfile' */
779 			uniqfile = optarg;
780 			break;
781 		case 'D':
782 			/* Uniquify against label `uniqlabel' in `uniqfile' */
783 			uniqlabel = optarg;
784 			break;
785 		case 'f':
786 			write_fuzzy_match = CTF_FUZZY_MATCH;
787 			break;
788 		case 'g':
789 			keep_stabs = CTF_KEEP_STABS;
790 			break;
791 		case 'l':
792 			/* Label merged types with `label' */
793 			label = optarg;
794 			break;
795 		case 'L':
796 			/* Label merged types with getenv(`label`) */
797 			if ((label = getenv(optarg)) == NULL)
798 				label = CTF_DEFAULT_LABEL;
799 			break;
800 		case 'o':
801 			/* Place merged types in CTF section in `outfile' */
802 			outfile = optarg;
803 			break;
804 		case 't':
805 			/* Insist *all* object files built from C have CTF */
806 			require_ctf = 1;
807 			break;
808 		case 'v':
809 			/* More debugging information */
810 			verbose = 1;
811 			break;
812 		case 'w':
813 			/* Additive merge with data from `withfile' */
814 			withfile = optarg;
815 			break;
816 		case 's':
817 			/* use the dynsym rather than the symtab */
818 			dynsym = CTF_USE_DYNSYM;
819 			break;
820 		default:
821 			usage();
822 			exit(2);
823 		}
824 	}
825 
826 	/* Validate arguments */
827 	if (docopy) {
828 		if (uniqfile != NULL || uniqlabel != NULL || label != NULL ||
829 		    outfile != NULL || withfile != NULL || dynsym != 0)
830 			err++;
831 
832 		if (argc - optind != 2)
833 			err++;
834 	} else {
835 		if (uniqfile != NULL && withfile != NULL)
836 			err++;
837 
838 		if (uniqlabel != NULL && uniqfile == NULL)
839 			err++;
840 
841 		if (outfile == NULL || label == NULL)
842 			err++;
843 
844 		if (argc - optind == 0)
845 			err++;
846 	}
847 
848 	if (err) {
849 		usage();
850 		exit(2);
851 	}
852 
853 	if (getenv("STRIPSTABS_KEEP_STABS") != NULL)
854 		keep_stabs = CTF_KEEP_STABS;
855 
856 	if (uniqfile && access(uniqfile, R_OK) != 0) {
857 		warning("Uniquification file %s couldn't be opened and "
858 		    "will be ignored.\n", uniqfile);
859 		uniqfile = NULL;
860 	}
861 	if (withfile && access(withfile, R_OK) != 0) {
862 		warning("With file %s couldn't be opened and will be "
863 		    "ignored.\n", withfile);
864 		withfile = NULL;
865 	}
866 	if (outfile && access(outfile, R_OK|W_OK) != 0)
867 		terminate("Cannot open output file %s for r/w", outfile);
868 
869 	/*
870 	 * This is ugly, but we don't want to have to have a separate tool
871 	 * (yet) just for copying an ELF section with our specific requirements,
872 	 * so we shoe-horn a copier into ctfmerge.
873 	 */
874 	if (docopy) {
875 		copy_ctf_data(argv[optind], argv[optind + 1], keep_stabs);
876 
877 		exit(0);
878 	}
879 
880 	set_terminate_cleanup(terminate_cleanup);
881 
882 	/* Sort the input files and strip out duplicates */
883 	nifiles = argc - optind;
884 	ifiles = xmalloc(sizeof (char *) * nifiles);
885 	tifiles = xmalloc(sizeof (char *) * nifiles);
886 
887 	for (i = 0; i < nifiles; i++)
888 		tifiles[i] = argv[optind + i];
889 	qsort(tifiles, nifiles, sizeof (char *), strcompare);
890 
891 	ifiles[0] = tifiles[0];
892 	for (idx = 0, tidx = 1; tidx < nifiles; tidx++) {
893 		if (strcmp(ifiles[idx], tifiles[tidx]) != 0)
894 			ifiles[++idx] = tifiles[tidx];
895 	}
896 	nifiles = idx + 1;
897 
898 	/* Make sure they all exist */
899 	if ((nielems = count_files(ifiles, nifiles)) < 0)
900 		terminate("Some input files were inaccessible\n");
901 
902 	/* Prepare for the merge */
903 	wq_init(&wq, nielems);
904 
905 	start_threads(&wq);
906 
907 	/*
908 	 * Start the merge
909 	 *
910 	 * We're reading everything from each of the object files, so we
911 	 * don't need to specify labels.
912 	 */
913 	if (read_ctf(ifiles, nifiles, NULL, merge_ctf_cb,
914 	    &wq, require_ctf) == 0) {
915 		warning("No ctf sections found to merge\n");
916 		exit(0);
917 	}
918 
919 	pthread_mutex_lock(&wq.wq_queue_lock);
920 	wq.wq_nomorefiles = 1;
921 	pthread_cond_broadcast(&wq.wq_work_avail);
922 	pthread_mutex_unlock(&wq.wq_queue_lock);
923 
924 	pthread_mutex_lock(&wq.wq_queue_lock);
925 	while (wq.wq_alldone == 0)
926 		pthread_cond_wait(&wq.wq_alldone_cv, &wq.wq_queue_lock);
927 	pthread_mutex_unlock(&wq.wq_queue_lock);
928 
929 	join_threads(&wq);
930 
931 	/*
932 	 * All requested files have been merged, with the resulting tree in
933 	 * mstrtd.  savetd is the tree that will be placed into the output file.
934 	 *
935 	 * Regardless of whether we're doing a normal uniquification or an
936 	 * additive merge, we need a type tree that has been uniquified
937 	 * against uniqfile or withfile, as appropriate.
938 	 *
939 	 * If we're doing a uniquification, we stuff the resulting tree into
940 	 * outfile.  Otherwise, we add the tree to the tree already in withfile.
941 	 */
942 	assert(fifo_len(wq.wq_queue) == 1);
943 	mstrtd = fifo_remove(wq.wq_queue);
944 
945 	if (verbose || debug_level) {
946 		debug(2, "Statistics for td %p\n", (void *)mstrtd);
947 
948 		iidesc_stats(mstrtd->td_iihash);
949 	}
950 
951 	if (uniqfile != NULL || withfile != NULL) {
952 		char *reffile, *reflabel = NULL;
953 		tdata_t *reftd;
954 
955 		if (uniqfile != NULL) {
956 			reffile = uniqfile;
957 			reflabel = uniqlabel;
958 		} else
959 			reffile = withfile;
960 
961 		if (read_ctf(&reffile, 1, reflabel, read_ctf_save_cb,
962 		    &reftd, require_ctf) == 0) {
963 			terminate("No CTF data found in reference file %s\n",
964 			    reffile);
965 		}
966 
967 		savetd = tdata_new();
968 
969 		if (CTF_V3_TYPE_ISCHILD(reftd->td_nextid))
970 			terminate("No room for additional types in master\n");
971 
972 		savetd->td_nextid = withfile ? reftd->td_nextid :
973 		    CTF_V3_INDEX_TO_TYPE(1, TRUE);
974 		merge_into_master(mstrtd, reftd, savetd, 0);
975 
976 		tdata_label_add(savetd, label, CTF_LABEL_LASTIDX);
977 
978 		if (withfile) {
979 			/*
980 			 * savetd holds the new data to be added to the withfile
981 			 */
982 			tdata_t *withtd = reftd;
983 
984 			tdata_merge(withtd, savetd);
985 
986 			savetd = withtd;
987 		} else {
988 			char uniqname[MAXPATHLEN];
989 			labelent_t *parle;
990 
991 			parle = tdata_label_top(reftd);
992 
993 			savetd->td_parlabel = xstrdup(parle->le_name);
994 
995 			strncpy(uniqname, reffile, sizeof (uniqname));
996 			uniqname[MAXPATHLEN - 1] = '\0';
997 			savetd->td_parname = xstrdup(basename(uniqname));
998 		}
999 
1000 	} else {
1001 		/*
1002 		 * No post processing.  Write the merged tree as-is into the
1003 		 * output file.
1004 		 */
1005 		tdata_label_free(mstrtd);
1006 		tdata_label_add(mstrtd, label, CTF_LABEL_LASTIDX);
1007 
1008 		savetd = mstrtd;
1009 	}
1010 
1011 	tmpname = mktmpname(outfile, ".ctf");
1012 	write_ctf(savetd, outfile, tmpname,
1013 	    CTF_COMPRESS | CTF_SWAP_BYTES | write_fuzzy_match | dynsym | keep_stabs);
1014 	if (rename(tmpname, outfile) != 0)
1015 		terminate("Couldn't rename output temp file %s", tmpname);
1016 	free(tmpname);
1017 
1018 	return (0);
1019 }
1020