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/.
33 #include "wimlib/assert.h"
34 #include "wimlib/chunk_compressor.h"
35 #include "wimlib/error.h"
36 #include "wimlib/list.h"
37 #include "wimlib/util.h"
39 struct message_queue {
40 struct list_head list;
42 pthread_cond_t msg_avail_cond;
43 pthread_cond_t space_avail_cond;
47 struct compressor_thread_data {
49 struct message_queue *chunks_to_compress_queue;
50 struct message_queue *compressed_chunks_queue;
51 struct wimlib_compressor *compressor;
54 #define MAX_CHUNKS_PER_MSG 16
57 u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
58 u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
59 u32 uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
60 u32 compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
61 size_t num_filled_chunks;
62 size_t num_alloc_chunks;
63 struct list_head list;
65 struct list_head submission_list;
68 struct parallel_chunk_compressor {
69 struct chunk_compressor base;
71 struct message_queue chunks_to_compress_queue;
72 struct message_queue compressed_chunks_queue;
73 struct compressor_thread_data *thread_data;
74 unsigned num_thread_data;
75 unsigned num_started_threads;
80 struct list_head available_msgs;
81 struct list_head submitted_msgs;
82 struct message *next_submit_msg;
83 struct message *next_ready_msg;
84 size_t next_chunk_idx;
90 message_queue_init(struct message_queue *q)
92 if (pthread_mutex_init(&q->lock, NULL)) {
93 ERROR_WITH_ERRNO("Failed to initialize mutex");
96 if (pthread_cond_init(&q->msg_avail_cond, NULL)) {
97 ERROR_WITH_ERRNO("Failed to initialize condition variable");
98 goto err_destroy_lock;
100 if (pthread_cond_init(&q->space_avail_cond, NULL)) {
101 ERROR_WITH_ERRNO("Failed to initialize condition variable");
102 goto err_destroy_msg_avail_cond;
104 INIT_LIST_HEAD(&q->list);
107 err_destroy_msg_avail_cond:
108 pthread_cond_destroy(&q->msg_avail_cond);
110 pthread_mutex_destroy(&q->lock);
112 return WIMLIB_ERR_NOMEM;
116 message_queue_destroy(struct message_queue *q)
118 if (q->list.next != NULL) {
119 pthread_mutex_destroy(&q->lock);
120 pthread_cond_destroy(&q->msg_avail_cond);
121 pthread_cond_destroy(&q->space_avail_cond);
126 message_queue_put(struct message_queue *q, struct message *msg)
128 pthread_mutex_lock(&q->lock);
129 list_add_tail(&msg->list, &q->list);
130 pthread_cond_signal(&q->msg_avail_cond);
131 pthread_mutex_unlock(&q->lock);
134 static struct message *
135 message_queue_get(struct message_queue *q)
139 pthread_mutex_lock(&q->lock);
140 while (list_empty(&q->list) && !q->terminating)
141 pthread_cond_wait(&q->msg_avail_cond, &q->lock);
142 if (!q->terminating) {
143 msg = list_entry(q->list.next, struct message, list);
144 list_del(&msg->list);
147 pthread_mutex_unlock(&q->lock);
152 message_queue_terminate(struct message_queue *q)
154 pthread_mutex_lock(&q->lock);
155 q->terminating = true;
156 pthread_cond_broadcast(&q->msg_avail_cond);
157 pthread_mutex_unlock(&q->lock);
161 init_message(struct message *msg, size_t num_chunks, u32 out_chunk_size)
163 msg->num_alloc_chunks = num_chunks;
164 for (size_t i = 0; i < num_chunks; i++) {
165 msg->compressed_chunks[i] = MALLOC(out_chunk_size - 1);
166 msg->uncompressed_chunks[i] = MALLOC(out_chunk_size);
167 if (msg->compressed_chunks[i] == NULL ||
168 msg->uncompressed_chunks[i] == NULL)
169 return WIMLIB_ERR_NOMEM;
175 destroy_message(struct message *msg)
177 for (size_t i = 0; i < msg->num_alloc_chunks; i++) {
178 FREE(msg->compressed_chunks[i]);
179 FREE(msg->uncompressed_chunks[i]);
184 free_messages(struct message *msgs, size_t num_messages)
187 for (size_t i = 0; i < num_messages; i++)
188 destroy_message(&msgs[i]);
193 static struct message *
194 allocate_messages(size_t count, size_t chunks_per_msg, u32 out_chunk_size)
196 struct message *msgs;
198 msgs = CALLOC(count, sizeof(struct message));
201 for (size_t i = 0; i < count; i++) {
202 if (init_message(&msgs[i], chunks_per_msg, out_chunk_size)) {
203 free_messages(msgs, count);
211 compress_chunks(struct message *msg, struct wimlib_compressor *compressor)
214 for (size_t i = 0; i < msg->num_filled_chunks; i++) {
215 wimlib_assert(msg->uncompressed_chunk_sizes[i] != 0);
216 msg->compressed_chunk_sizes[i] =
217 wimlib_compress(msg->uncompressed_chunks[i],
218 msg->uncompressed_chunk_sizes[i],
219 msg->compressed_chunks[i],
220 msg->uncompressed_chunk_sizes[i] - 1,
226 compressor_thread_proc(void *arg)
228 struct compressor_thread_data *params = arg;
231 while ((msg = message_queue_get(params->chunks_to_compress_queue)) != NULL) {
232 compress_chunks(msg, params->compressor);
233 message_queue_put(params->compressed_chunks_queue, msg);
239 parallel_chunk_compressor_destroy(struct chunk_compressor *_ctx)
241 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
247 if (ctx->num_started_threads != 0) {
248 message_queue_terminate(&ctx->chunks_to_compress_queue);
250 for (i = 0; i < ctx->num_started_threads; i++)
251 pthread_join(ctx->thread_data[i].thread, NULL);
254 message_queue_destroy(&ctx->chunks_to_compress_queue);
255 message_queue_destroy(&ctx->compressed_chunks_queue);
257 if (ctx->thread_data != NULL)
258 for (i = 0; i < ctx->num_thread_data; i++)
259 wimlib_free_compressor(ctx->thread_data[i].compressor);
261 FREE(ctx->thread_data);
263 free_messages(ctx->msgs, ctx->num_messages);
269 submit_compression_msg(struct parallel_chunk_compressor *ctx)
271 struct message *msg = ctx->next_submit_msg;
273 msg->complete = false;
274 list_add_tail(&msg->submission_list, &ctx->submitted_msgs);
275 message_queue_put(&ctx->chunks_to_compress_queue, msg);
276 ctx->next_submit_msg = NULL;
280 parallel_chunk_compressor_get_chunk_buffer(struct chunk_compressor *_ctx)
282 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
285 if (ctx->next_submit_msg) {
286 msg = ctx->next_submit_msg;
288 if (list_empty(&ctx->available_msgs))
291 msg = list_entry(ctx->available_msgs.next, struct message, list);
292 list_del(&msg->list);
293 ctx->next_submit_msg = msg;
294 msg->num_filled_chunks = 0;
297 return msg->uncompressed_chunks[msg->num_filled_chunks];
301 parallel_chunk_compressor_signal_chunk_filled(struct chunk_compressor *_ctx, u32 usize)
303 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
306 wimlib_assert(usize > 0);
307 wimlib_assert(usize <= ctx->base.out_chunk_size);
308 wimlib_assert(ctx->next_submit_msg);
310 msg = ctx->next_submit_msg;
311 msg->uncompressed_chunk_sizes[msg->num_filled_chunks] = usize;
312 if (++msg->num_filled_chunks == msg->num_alloc_chunks)
313 submit_compression_msg(ctx);
317 parallel_chunk_compressor_get_compression_result(struct chunk_compressor *_ctx,
318 const void **cdata_ret, u32 *csize_ret,
321 struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
324 if (ctx->next_submit_msg)
325 submit_compression_msg(ctx);
327 if (ctx->next_ready_msg) {
328 msg = ctx->next_ready_msg;
330 if (list_empty(&ctx->submitted_msgs))
333 while (!(msg = list_entry(ctx->submitted_msgs.next,
335 submission_list))->complete)
336 message_queue_get(&ctx->compressed_chunks_queue)->complete = true;
338 ctx->next_ready_msg = msg;
339 ctx->next_chunk_idx = 0;
342 if (msg->compressed_chunk_sizes[ctx->next_chunk_idx]) {
343 *cdata_ret = msg->compressed_chunks[ctx->next_chunk_idx];
344 *csize_ret = msg->compressed_chunk_sizes[ctx->next_chunk_idx];
346 *cdata_ret = msg->uncompressed_chunks[ctx->next_chunk_idx];
347 *csize_ret = msg->uncompressed_chunk_sizes[ctx->next_chunk_idx];
349 *usize_ret = msg->uncompressed_chunk_sizes[ctx->next_chunk_idx];
351 if (++ctx->next_chunk_idx == msg->num_filled_chunks) {
352 list_del(&msg->submission_list);
353 list_add_tail(&msg->list, &ctx->available_msgs);
354 ctx->next_ready_msg = NULL;
360 new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
361 unsigned num_threads, u64 max_memory,
362 struct chunk_compressor **compressor_ret)
364 u64 approx_mem_required;
365 size_t chunks_per_msg;
366 size_t msgs_per_thread;
367 struct parallel_chunk_compressor *ctx;
370 unsigned desired_num_threads;
372 wimlib_assert(out_chunk_size > 0);
374 if (num_threads == 0)
375 num_threads = get_available_cpus();
377 if (num_threads == 1)
381 max_memory = get_available_memory();
383 desired_num_threads = num_threads;
385 if (out_chunk_size < ((u32)1 << 23)) {
386 /* Relatively small chunks. Use 2 messages per thread, each
387 * with at least 2 chunks. Use more chunks per message if there
388 * are lots of threads and/or the chunks are very small. */
390 chunks_per_msg += num_threads * (65536 / out_chunk_size) / 16;
391 chunks_per_msg = max(chunks_per_msg, 2);
392 chunks_per_msg = min(chunks_per_msg, MAX_CHUNKS_PER_MSG);
395 /* Big chunks: Just have one buffer per thread --- more would
396 * just waste memory. */
401 approx_mem_required =
402 (u64)chunks_per_msg *
403 (u64)msgs_per_thread *
408 + num_threads * wimlib_get_compressor_needed_memory(out_ctype,
411 if (approx_mem_required <= max_memory)
414 if (chunks_per_msg > 1)
416 else if (msgs_per_thread > 1)
418 else if (num_threads > 1)
424 if (num_threads < desired_num_threads) {
425 WARNING("Wanted to use %u threads, but limiting to %u "
426 "to fit in available memory!",
427 desired_num_threads, num_threads);
430 if (num_threads == 1)
433 ret = WIMLIB_ERR_NOMEM;
434 ctx = CALLOC(1, sizeof(*ctx));
438 ctx->base.out_ctype = out_ctype;
439 ctx->base.out_chunk_size = out_chunk_size;
440 ctx->base.destroy = parallel_chunk_compressor_destroy;
441 ctx->base.get_chunk_buffer = parallel_chunk_compressor_get_chunk_buffer;
442 ctx->base.signal_chunk_filled = parallel_chunk_compressor_signal_chunk_filled;
443 ctx->base.get_compression_result = parallel_chunk_compressor_get_compression_result;
445 ctx->num_thread_data = num_threads;
447 ret = message_queue_init(&ctx->chunks_to_compress_queue);
451 ret = message_queue_init(&ctx->compressed_chunks_queue);
455 ret = WIMLIB_ERR_NOMEM;
456 ctx->thread_data = CALLOC(num_threads, sizeof(ctx->thread_data[0]));
457 if (ctx->thread_data == NULL)
460 for (i = 0; i < num_threads; i++) {
461 struct compressor_thread_data *dat;
463 dat = &ctx->thread_data[i];
465 dat->chunks_to_compress_queue = &ctx->chunks_to_compress_queue;
466 dat->compressed_chunks_queue = &ctx->compressed_chunks_queue;
467 ret = wimlib_create_compressor(out_ctype, out_chunk_size,
468 WIMLIB_COMPRESSOR_FLAG_DESTRUCTIVE,
474 for (ctx->num_started_threads = 0;
475 ctx->num_started_threads < num_threads;
476 ctx->num_started_threads++)
478 ret = pthread_create(&ctx->thread_data[ctx->num_started_threads].thread,
480 compressor_thread_proc,
481 &ctx->thread_data[ctx->num_started_threads]);
484 WARNING_WITH_ERRNO("Failed to create compressor thread %u of %u",
485 ctx->num_started_threads + 1,
487 ret = WIMLIB_ERR_NOMEM;
488 if (ctx->num_started_threads >= 2)
494 ctx->base.num_threads = ctx->num_started_threads;
496 ret = WIMLIB_ERR_NOMEM;
497 ctx->num_messages = ctx->num_started_threads * msgs_per_thread;
498 ctx->msgs = allocate_messages(ctx->num_messages,
499 chunks_per_msg, out_chunk_size);
500 if (ctx->msgs == NULL)
503 INIT_LIST_HEAD(&ctx->available_msgs);
504 for (size_t i = 0; i < ctx->num_messages; i++)
505 list_add_tail(&ctx->msgs[i].list, &ctx->available_msgs);
507 INIT_LIST_HEAD(&ctx->submitted_msgs);
509 *compressor_ret = &ctx->base;
513 parallel_chunk_compressor_destroy(&ctx->base);