4 * Compress chunks of data (parallel version).
8 * Copyright (C) 2013 Eric Biggers
10 * This file is free software; you can redistribute it and/or modify it under
11 * the terms of the GNU Lesser General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option) any
15 * This file is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
17 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
20 * You should have received a copy of the GNU Lesser General Public License
21 * along with this file; if not, see http://www.gnu.org/licenses/.
28 #ifdef ENABLE_MULTITHREADED_COMPRESSION
36 #ifdef HAVE_SYS_SYSCTL_H
37 # include <sys/sysctl.h>
40 #include "wimlib/assert.h"
41 #include "wimlib/chunk_compressor.h"
42 #include "wimlib/error.h"
43 #include "wimlib/list.h"
44 #include "wimlib/util.h"
46 # include "wimlib/win32.h" /* win32_get_number_of_processors() */
49 struct message_queue {
50 struct list_head list;
52 pthread_cond_t msg_avail_cond;
53 pthread_cond_t space_avail_cond;
57 struct compressor_thread_data {
59 struct message_queue *chunks_to_compress_queue;
60 struct message_queue *compressed_chunks_queue;
61 struct wimlib_compressor *compressor;
64 #define MAX_CHUNKS_PER_MSG 16
67 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
68 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
69 u32 uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
70 u32 compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
71 size_t num_filled_chunks;
72 size_t num_alloc_chunks;
73 struct list_head list;
75 struct list_head submission_list;
78 struct parallel_chunk_compressor {
79 struct chunk_compressor base;
81 struct message_queue chunks_to_compress_queue;
82 struct message_queue compressed_chunks_queue;
83 struct compressor_thread_data *thread_data;
84 unsigned num_thread_data;
85 unsigned num_started_threads;
90 struct list_head available_msgs;
91 struct list_head submitted_msgs;
92 struct message *next_submit_msg;
93 struct message *next_ready_msg;
94 size_t next_chunk_idx;
98 get_default_num_threads(void)
102 n = win32_get_number_of_processors();
104 n = sysconf(_SC_NPROCESSORS_ONLN);
106 if (n < 1 || n >= UINT_MAX) {
107 WARNING("Failed to determine number of processors; assuming 1.");
114 get_avail_memory(void)
117 u64 phys_bytes = win32_get_avail_memory();
121 #elif defined(_SC_PAGESIZE) && defined(_SC_PHYS_PAGES)
122 long page_size = sysconf(_SC_PAGESIZE);
123 long num_pages = sysconf(_SC_PHYS_PAGES);
124 if (page_size <= 0 || num_pages <= 0)
126 return ((u64)page_size * (u64)num_pages);
128 int mib[2] = {CTL_HW, HW_MEMSIZE};
130 size_t len = sizeof(memsize);
131 if (sysctl(mib, ARRAY_LEN(mib), &memsize, &len, NULL, 0) < 0 || len != 8)
137 WARNING("Failed to determine available memory; assuming 1 GiB");
142 message_queue_init(struct message_queue *q)
144 if (pthread_mutex_init(&q->lock, NULL)) {
145 ERROR_WITH_ERRNO("Failed to initialize mutex");
148 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
149 ERROR_WITH_ERRNO("Failed to initialize condition variable");
150 goto err_destroy_lock;
152 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
153 ERROR_WITH_ERRNO("Failed to initialize condition variable");
154 goto err_destroy_msg_avail_cond;
156 INIT_LIST_HEAD(&q->list);
159 err_destroy_msg_avail_cond:
160 pthread_cond_destroy(&q->msg_avail_cond);
162 pthread_mutex_destroy(&q->lock);
164 return WIMLIB_ERR_NOMEM;
168 message_queue_destroy(struct message_queue *q)
170 if (q->list.next != NULL) {
171 pthread_mutex_destroy(&q->lock);
172 pthread_cond_destroy(&q->msg_avail_cond);
173 pthread_cond_destroy(&q->space_avail_cond);
178 message_queue_put(struct message_queue *q, struct message *msg)
180 pthread_mutex_lock(&q->lock);
181 list_add_tail(&msg->list, &q->list);
182 pthread_cond_signal(&q->msg_avail_cond);
183 pthread_mutex_unlock(&q->lock);
186 static struct message *
187 message_queue_get(struct message_queue *q)
191 pthread_mutex_lock(&q->lock);
192 while (list_empty(&q->list) && !q->terminating)
193 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
194 if (!q->terminating) {
195 msg = list_entry(q->list.next, struct message, list);
196 list_del(&msg->list);
199 pthread_mutex_unlock(&q->lock);
204 message_queue_terminate(struct message_queue *q)
206 pthread_mutex_lock(&q->lock);
207 q->terminating = true;
208 pthread_cond_broadcast(&q->msg_avail_cond);
209 pthread_mutex_unlock(&q->lock);
213 init_message(struct message *msg, size_t num_chunks, u32 out_chunk_size)
215 msg->num_alloc_chunks = num_chunks;
216 for (size_t i = 0; i < num_chunks; i++) {
217 msg->compressed_chunks[i] = MALLOC(out_chunk_size - 1);
218 msg->uncompressed_chunks[i] = MALLOC(out_chunk_size);
219 if (msg->compressed_chunks[i] == NULL ||
220 msg->uncompressed_chunks[i] == NULL)
221 return WIMLIB_ERR_NOMEM;
227 destroy_message(struct message *msg)
229 for (size_t i = 0; i < msg->num_alloc_chunks; i++) {
230 FREE(msg->compressed_chunks[i]);
231 FREE(msg->uncompressed_chunks[i]);
236 free_messages(struct message *msgs, size_t num_messages)
239 for (size_t i = 0; i < num_messages; i++)
240 destroy_message(&msgs[i]);
245 static struct message *
246 allocate_messages(size_t count, size_t chunks_per_msg, u32 out_chunk_size)
248 struct message *msgs;
250 msgs = CALLOC(count, sizeof(struct message));
253 for (size_t i = 0; i < count; i++) {
254 if (init_message(&msgs[i], chunks_per_msg, out_chunk_size)) {
255 free_messages(msgs, count);
263 compress_chunks(struct message *msg, struct wimlib_compressor *compressor)
266 for (size_t i = 0; i < msg->num_filled_chunks; i++) {
267 wimlib_assert(msg->uncompressed_chunk_sizes[i] != 0);
268 msg->compressed_chunk_sizes[i] =
269 wimlib_compress(msg->uncompressed_chunks[i],
270 msg->uncompressed_chunk_sizes[i],
271 msg->compressed_chunks[i],
272 msg->uncompressed_chunk_sizes[i] - 1,
278 compressor_thread_proc(void *arg)
280 struct compressor_thread_data *params = arg;
283 while ((msg = message_queue_get(params->chunks_to_compress_queue)) != NULL) {
284 compress_chunks(msg, params->compressor);
285 message_queue_put(params->compressed_chunks_queue, msg);
291 parallel_chunk_compressor_destroy(struct chunk_compressor *_ctx)
293 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
299 if (ctx->num_started_threads != 0) {
300 DEBUG("Terminating %u compressor threads", ctx->num_started_threads);
301 message_queue_terminate(&ctx->chunks_to_compress_queue);
303 for (i = 0; i < ctx->num_started_threads; i++)
304 pthread_join(ctx->thread_data[i].thread, NULL);
307 message_queue_destroy(&ctx->chunks_to_compress_queue);
308 message_queue_destroy(&ctx->compressed_chunks_queue);
310 if (ctx->thread_data != NULL)
311 for (i = 0; i < ctx->num_thread_data; i++)
312 wimlib_free_compressor(ctx->thread_data[i].compressor);
314 FREE(ctx->thread_data);
316 free_messages(ctx->msgs, ctx->num_messages);
322 submit_compression_msg(struct parallel_chunk_compressor *ctx)
324 struct message *msg = ctx->next_submit_msg;
326 msg->complete = false;
327 list_add_tail(&msg->submission_list, &ctx->submitted_msgs);
328 message_queue_put(&ctx->chunks_to_compress_queue, msg);
329 ctx->next_submit_msg = NULL;
333 parallel_chunk_compressor_get_chunk_buffer(struct chunk_compressor *_ctx)
335 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
338 if (ctx->next_submit_msg) {
339 msg = ctx->next_submit_msg;
341 if (list_empty(&ctx->available_msgs))
344 msg = list_entry(ctx->available_msgs.next, struct message, list);
345 list_del(&msg->list);
346 ctx->next_submit_msg = msg;
347 msg->num_filled_chunks = 0;
350 return msg->uncompressed_chunks[msg->num_filled_chunks];
354 parallel_chunk_compressor_signal_chunk_filled(struct chunk_compressor *_ctx, u32 usize)
356 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
359 wimlib_assert(usize > 0);
360 wimlib_assert(usize <= ctx->base.out_chunk_size);
361 wimlib_assert(ctx->next_submit_msg);
363 msg = ctx->next_submit_msg;
364 msg->uncompressed_chunk_sizes[msg->num_filled_chunks] = usize;
365 if (++msg->num_filled_chunks == msg->num_alloc_chunks)
366 submit_compression_msg(ctx);
370 parallel_chunk_compressor_get_compression_result(struct chunk_compressor *_ctx,
371 const void **cdata_ret, u32 *csize_ret,
374 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
377 if (ctx->next_submit_msg)
378 submit_compression_msg(ctx);
380 if (ctx->next_ready_msg) {
381 msg = ctx->next_ready_msg;
383 if (list_empty(&ctx->submitted_msgs))
386 while (!(msg = list_entry(ctx->submitted_msgs.next,
388 submission_list))->complete)
389 message_queue_get(&ctx->compressed_chunks_queue)->complete = true;
391 ctx->next_ready_msg = msg;
392 ctx->next_chunk_idx = 0;
395 if (msg->compressed_chunk_sizes[ctx->next_chunk_idx]) {
396 *cdata_ret = msg->compressed_chunks[ctx->next_chunk_idx];
397 *csize_ret = msg->compressed_chunk_sizes[ctx->next_chunk_idx];
399 *cdata_ret = msg->uncompressed_chunks[ctx->next_chunk_idx];
400 *csize_ret = msg->uncompressed_chunk_sizes[ctx->next_chunk_idx];
402 *usize_ret = msg->uncompressed_chunk_sizes[ctx->next_chunk_idx];
404 if (++ctx->next_chunk_idx == msg->num_filled_chunks) {
405 list_del(&msg->submission_list);
406 list_add_tail(&msg->list, &ctx->available_msgs);
407 ctx->next_ready_msg = NULL;
413 new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
414 unsigned num_threads, u64 max_memory,
415 struct chunk_compressor **compressor_ret)
417 u64 approx_mem_required;
418 size_t chunks_per_msg;
419 size_t msgs_per_thread;
420 struct parallel_chunk_compressor *ctx;
423 unsigned desired_num_threads;
425 wimlib_assert(out_chunk_size > 0);
427 if (num_threads == 0)
428 num_threads = get_default_num_threads();
430 if (num_threads == 1) {
431 DEBUG("Only 1 thread; Not bothering with "
432 "parallel chunk compressor.");
437 max_memory = get_avail_memory();
439 desired_num_threads = num_threads;
441 if (out_chunk_size < ((u32)1 << 23)) {
442 /* Relatively small chunks. Use 2 messages per thread, each
443 * with at least 2 chunks. Use more chunks per message if there
444 * are lots of threads and/or the chunks are very small. */
446 chunks_per_msg += num_threads * (65536 / out_chunk_size) / 16;
447 chunks_per_msg = max(chunks_per_msg, 2);
448 chunks_per_msg = min(chunks_per_msg, MAX_CHUNKS_PER_MSG);
451 /* Big chunks: Just have one buffer per thread --- more would
452 * just waste memory. */
457 approx_mem_required =
458 (u64)chunks_per_msg *
459 (u64)msgs_per_thread *
464 + num_threads * wimlib_get_compressor_needed_memory(out_ctype,
467 if (approx_mem_required <= max_memory)
470 if (chunks_per_msg > 1)
472 else if (msgs_per_thread > 1)
474 else if (num_threads > 1)
480 if (num_threads < desired_num_threads) {
481 WARNING("Wanted to use %u threads, but limiting to %u "
482 "to fit in available memory!",
483 desired_num_threads, num_threads);
486 if (num_threads == 1) {
487 DEBUG("Only 1 thread; Not bothering with "
488 "parallel chunk compressor.");
492 ret = WIMLIB_ERR_NOMEM;
493 ctx = CALLOC(1, sizeof(*ctx));
497 ctx->base.out_ctype = out_ctype;
498 ctx->base.out_chunk_size = out_chunk_size;
499 ctx->base.destroy = parallel_chunk_compressor_destroy;
500 ctx->base.get_chunk_buffer = parallel_chunk_compressor_get_chunk_buffer;
501 ctx->base.signal_chunk_filled = parallel_chunk_compressor_signal_chunk_filled;
502 ctx->base.get_compression_result = parallel_chunk_compressor_get_compression_result;
504 ctx->num_thread_data = num_threads;
506 ret = message_queue_init(&ctx->chunks_to_compress_queue);
510 ret = message_queue_init(&ctx->compressed_chunks_queue);
514 ret = WIMLIB_ERR_NOMEM;
515 ctx->thread_data = CALLOC(num_threads, sizeof(ctx->thread_data[0]));
516 if (ctx->thread_data == NULL)
519 for (i = 0; i < num_threads; i++) {
520 struct compressor_thread_data *dat;
522 dat = &ctx->thread_data[i];
524 dat->chunks_to_compress_queue = &ctx->chunks_to_compress_queue;
525 dat->compressed_chunks_queue = &ctx->compressed_chunks_queue;
526 ret = wimlib_create_compressor(out_ctype, out_chunk_size,
527 WIMLIB_COMPRESSOR_FLAG_DESTRUCTIVE,
533 for (ctx->num_started_threads = 0;
534 ctx->num_started_threads < num_threads;
535 ctx->num_started_threads++)
537 DEBUG("pthread_create thread %u of %u",
538 ctx->num_started_threads + 1, num_threads);
539 ret = pthread_create(&ctx->thread_data[ctx->num_started_threads].thread,
541 compressor_thread_proc,
542 &ctx->thread_data[ctx->num_started_threads]);
545 WARNING_WITH_ERRNO("Failed to create compressor thread %u of %u",
546 ctx->num_started_threads + 1,
548 ret = WIMLIB_ERR_NOMEM;
549 if (ctx->num_started_threads >= 2)
555 ctx->base.num_threads = ctx->num_started_threads;
557 ret = WIMLIB_ERR_NOMEM;
558 ctx->num_messages = ctx->num_started_threads * msgs_per_thread;
559 ctx->msgs = allocate_messages(ctx->num_messages,
560 chunks_per_msg, out_chunk_size);
561 if (ctx->msgs == NULL)
564 INIT_LIST_HEAD(&ctx->available_msgs);
565 for (size_t i = 0; i < ctx->num_messages; i++)
566 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
568 INIT_LIST_HEAD(&ctx->submitted_msgs);
570 *compressor_ret = &ctx->base;
574 parallel_chunk_compressor_destroy(&ctx->base);
578 #endif /* ENABLE_MULTITHREADED_COMPRESSION */