]> wimlib.net Git - wimlib/blobdiff - src/compress_parallel.c
header cleanups
[wimlib] / src / compress_parallel.c
index 6cd306c2fcb96d5aa6d8b7e64c08c27022c3ff12..acf4ea58afc24b4299f7848d22301486c6be5bbe 100644 (file)
@@ -7,19 +7,18 @@
 /*
  * Copyright (C) 2013 Eric Biggers
  *
- * This file is part of wimlib, a library for working with WIM files.
+ * This file is free software; you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the Free
+ * Software Foundation; either version 3 of the License, or (at your option) any
+ * later version.
  *
- * wimlib is free software; you can redistribute it and/or modify it under the
- * terms of the GNU General Public License as published by the Free Software
- * Foundation; either version 3 of the License, or (at your option) any later
- * version.
+ * This file is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
+ * details.
  *
- * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
- * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
- * A PARTICULAR PURPOSE. See the GNU General Public License for more details.
- *
- * You should have received a copy of the GNU General Public License along with
- * wimlib; if not, see http://www.gnu.org/licenses/.
+ * You should have received a copy of the GNU Lesser General Public License
+ * along with this file; if not, see http://www.gnu.org/licenses/.
  */
 
 #ifdef HAVE_CONFIG_H
 
 #ifdef ENABLE_MULTITHREADED_COMPRESSION
 
+#include <errno.h>
+#include <limits.h>
+#include <pthread.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#ifdef HAVE_SYS_SYSCTL_H
+#  include <sys/sysctl.h>
+#endif
+
 #include "wimlib/assert.h"
 #include "wimlib/chunk_compressor.h"
 #include "wimlib/error.h"
 #  include "wimlib/win32.h" /* win32_get_number_of_processors() */
 #endif
 
-#include <errno.h>
-#include <limits.h>
-#include <pthread.h>
-#include <stdlib.h>
-#include <string.h>
-#include <unistd.h>
-
 struct message_queue {
        struct list_head list;
        pthread_mutex_t lock;
@@ -59,13 +61,13 @@ struct compressor_thread_data {
        struct wimlib_compressor *compressor;
 };
 
-#define MAX_CHUNKS_PER_MSG 2
+#define MAX_CHUNKS_PER_MSG 16
 
 struct message {
        u8 *uncompressed_chunks[MAX_CHUNKS_PER_MSG];
        u8 *compressed_chunks[MAX_CHUNKS_PER_MSG];
-       unsigned uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
-       unsigned compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
+       u32 uncompressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
+       u32 compressed_chunk_sizes[MAX_CHUNKS_PER_MSG];
        size_t num_filled_chunks;
        size_t num_alloc_chunks;
        struct list_head list;
@@ -79,7 +81,7 @@ struct parallel_chunk_compressor {
        struct message_queue chunks_to_compress_queue;
        struct message_queue compressed_chunks_queue;
        struct compressor_thread_data *thread_data;
-       unsigned num_threads;
+       unsigned num_thread_data;
        unsigned num_started_threads;
 
        struct message *msgs;
@@ -116,17 +118,24 @@ get_avail_memory(void)
        if (phys_bytes == 0)
                goto default_size;
        return phys_bytes;
-#else
+#elif defined(_SC_PAGESIZE) && defined(_SC_PHYS_PAGES)
        long page_size = sysconf(_SC_PAGESIZE);
        long num_pages = sysconf(_SC_PHYS_PAGES);
        if (page_size <= 0 || num_pages <= 0)
                goto default_size;
        return ((u64)page_size * (u64)num_pages);
+#else
+       int mib[2] = {CTL_HW, HW_MEMSIZE};
+       u64 memsize;
+       size_t len = sizeof(memsize);
+       if (sysctl(mib, ARRAY_LEN(mib), &memsize, &len, NULL, 0) < 0 || len != 8)
+               goto default_size;
+       return memsize;
 #endif
 
 default_size:
        WARNING("Failed to determine available memory; assuming 1 GiB");
-       return 1U << 30;
+       return 1ULL << 30;
 }
 
 static int
@@ -299,7 +308,7 @@ parallel_chunk_compressor_destroy(struct chunk_compressor *_ctx)
        message_queue_destroy(&ctx->compressed_chunks_queue);
 
        if (ctx->thread_data != NULL)
-               for (i = 0; i < ctx->num_threads; i++)
+               for (i = 0; i < ctx->num_thread_data; i++)
                        wimlib_free_compressor(ctx->thread_data[i].compressor);
 
        FREE(ctx->thread_data);
@@ -322,7 +331,7 @@ submit_compression_msg(struct parallel_chunk_compressor *ctx)
 
 static bool
 parallel_chunk_compressor_submit_chunk(struct chunk_compressor *_ctx,
-                                      const void *chunk, size_t size)
+                                      const void *chunk, u32 size)
 {
        struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
        struct message *msg;
@@ -351,8 +360,8 @@ parallel_chunk_compressor_submit_chunk(struct chunk_compressor *_ctx,
 
 static bool
 parallel_chunk_compressor_get_chunk(struct chunk_compressor *_ctx,
-                                   const void **cdata_ret, unsigned *csize_ret,
-                                   unsigned *usize_ret)
+                                   const void **cdata_ret, u32 *csize_ret,
+                                   u32 *usize_ret)
 {
        struct parallel_chunk_compressor *ctx = (struct parallel_chunk_compressor *)_ctx;
        struct message *msg;
@@ -407,7 +416,6 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
        unsigned desired_num_threads;
 
        wimlib_assert(out_chunk_size > 0);
-       wimlib_assert(out_ctype != WIMLIB_COMPRESSION_TYPE_NONE);
 
        if (num_threads == 0)
                num_threads = get_default_num_threads();
@@ -424,7 +432,13 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
        desired_num_threads = num_threads;
 
        if (out_chunk_size < ((u32)1 << 23)) {
-               chunks_per_msg = MAX_CHUNKS_PER_MSG;
+               /* Relatively small chunks.  Use 2 messages per thread, each
+                * with at least 2 chunks.  Use more chunks per message if there
+                * are lots of threads and/or the chunks are very small.  */
+               chunks_per_msg = 2;
+               chunks_per_msg += num_threads * (65536 / out_chunk_size) / 16;
+               chunks_per_msg = max(chunks_per_msg, 2);
+               chunks_per_msg = min(chunks_per_msg, MAX_CHUNKS_PER_MSG);
                msgs_per_thread = 2;
        } else {
                /* Big chunks: Just have one buffer per thread --- more would
@@ -438,8 +452,11 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
                        (u64)msgs_per_thread *
                        (u64)num_threads *
                        (u64)out_chunk_size
+                       + out_chunk_size
                        + 1000000
-                       + (out_chunk_size * num_threads * 4);
+                       + num_threads * wimlib_get_compressor_needed_memory(out_ctype,
+                                                                           out_chunk_size,
+                                                                           0);
                if (approx_mem_required <= max_memory)
                        break;
 
@@ -472,12 +489,11 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
 
        ctx->base.out_ctype = out_ctype;
        ctx->base.out_chunk_size = out_chunk_size;
-       ctx->base.num_threads = num_threads;
        ctx->base.destroy = parallel_chunk_compressor_destroy;
        ctx->base.submit_chunk = parallel_chunk_compressor_submit_chunk;
        ctx->base.get_chunk = parallel_chunk_compressor_get_chunk;
 
-       ctx->num_threads = num_threads;
+       ctx->num_thread_data = num_threads;
 
        ret = message_queue_init(&ctx->chunks_to_compress_queue);
        if (ret)
@@ -499,8 +515,8 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
 
                dat->chunks_to_compress_queue = &ctx->chunks_to_compress_queue;
                dat->compressed_chunks_queue = &ctx->compressed_chunks_queue;
-               ret = wimlib_create_compressor(out_ctype, out_chunk_size,
-                                              NULL, &dat->compressor);
+               ret = wimlib_create_compressor(out_ctype, out_chunk_size, 0,
+                                              &dat->compressor);
                if (ret)
                        goto err;
        }
@@ -517,14 +533,20 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
                                     &ctx->thread_data[ctx->num_started_threads]);
                if (ret) {
                        errno = ret;
+                       WARNING_WITH_ERRNO("Failed to create compressor thread %u of %u",
+                                          ctx->num_started_threads + 1,
+                                          num_threads);
                        ret = WIMLIB_ERR_NOMEM;
-                       WARNING_WITH_ERRNO("Failed to create compressor thread %u of %u");
+                       if (ctx->num_started_threads >= 2)
+                               break;
                        goto err;
                }
        }
 
+       ctx->base.num_threads = ctx->num_started_threads;
+
        ret = WIMLIB_ERR_NOMEM;
-       ctx->num_messages = num_threads * msgs_per_thread;
+       ctx->num_messages = ctx->num_started_threads * msgs_per_thread;
        ctx->msgs = allocate_messages(ctx->num_messages,
                                      chunks_per_msg, out_chunk_size);
        if (ctx->msgs == NULL)