]> wimlib.net Git - wimlib/blobdiff - src/compress_parallel.c
Merge compression updates
[wimlib] / src / compress_parallel.c
index 25403423bd35af84f516908336cee520fa1193d6..e624819be77bb432a6c685ece3e2e039537c56b7 100644 (file)
 #  include "config.h"
 #endif
 
+#ifdef ENABLE_MULTITHREADED_COMPRESSION
+
 #include "wimlib/assert.h"
-#include "wimlib/compress_chunks.h"
+#include "wimlib/chunk_compressor.h"
 #include "wimlib/error.h"
 #include "wimlib/list.h"
 #include "wimlib/util.h"
@@ -41,6 +43,9 @@
 #include <stdlib.h>
 #include <string.h>
 #include <unistd.h>
+#ifdef HAVE_SYS_SYSCTL_H
+#  include <sys/sysctl.h>
+#endif
 
 struct message_queue {
        struct list_head list;
@@ -54,9 +59,7 @@ struct compressor_thread_data {
        pthread_t thread;
        struct message_queue *chunks_to_compress_queue;
        struct message_queue *compressed_chunks_queue;
-       int out_ctype;
-       u32 out_chunk_size;
-       struct wimlib_lzx_context *comp_ctx;
+       struct wimlib_compressor *compressor;
 };
 
 #define MAX_CHUNKS_PER_MSG 2
@@ -64,8 +67,8 @@ struct compressor_thread_data {
 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 +82,7 @@ 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 +119,24 @@ 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
@@ -251,17 +261,17 @@ allocate_messages(size_t count, size_t chunks_per_msg, u32 out_chunk_size)
 }
 
 static void
-compress_chunks(struct message *msg, int out_ctype,
-               struct wimlib_lzx_context *comp_ctx)
+compress_chunks(struct message *msg, struct wimlib_compressor *compressor)
 {
 
        for (size_t i = 0; i < msg->num_filled_chunks; i++) {
+               wimlib_assert(msg->uncompressed_chunk_sizes[i] != 0);
                msg->compressed_chunk_sizes[i] =
-                       compress_chunk(msg->uncompressed_chunks[i],
-                                      msg->uncompressed_chunk_sizes[i],
-                                      msg->compressed_chunks[i],
-                                      out_ctype,
-                                      comp_ctx);
+                       wimlib_compress(msg->uncompressed_chunks[i],
+                                       msg->uncompressed_chunk_sizes[i],
+                                       msg->compressed_chunks[i],
+                                       msg->uncompressed_chunk_sizes[i] - 1,
+                                       compressor);
        }
 }
 
@@ -272,7 +282,7 @@ compressor_thread_proc(void *arg)
        struct message *msg;
 
        while ((msg = message_queue_get(params->chunks_to_compress_queue)) != NULL) {
-               compress_chunks(msg, params->out_ctype, params->comp_ctx);
+               compress_chunks(msg, params->compressor);
                message_queue_put(params->compressed_chunks_queue, msg);
        }
        return NULL;
@@ -298,10 +308,9 @@ parallel_chunk_compressor_destroy(struct chunk_compressor *_ctx)
        message_queue_destroy(&ctx->chunks_to_compress_queue);
        message_queue_destroy(&ctx->compressed_chunks_queue);
 
-       if (ctx->base.out_ctype == WIMLIB_COMPRESSION_TYPE_LZX &&
-           ctx->thread_data != NULL)
-               for (i = 0; i < ctx->num_threads; i++)
-                       wimlib_lzx_free_context(ctx->thread_data[i].comp_ctx);
+       if (ctx->thread_data != NULL)
+               for (i = 0; i < ctx->num_thread_data; i++)
+                       wimlib_free_compressor(ctx->thread_data[i].compressor);
 
        FREE(ctx->thread_data);
 
@@ -323,7 +332,7 @@ 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;
@@ -352,8 +361,8 @@ 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;
@@ -408,7 +417,6 @@ 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();
@@ -424,16 +432,26 @@ new_parallel_chunk_compressor(int out_ctype, u32 out_chunk_size,
 
        desired_num_threads = num_threads;
 
-       chunks_per_msg = MAX_CHUNKS_PER_MSG;
-       msgs_per_thread = 2;
+       if (out_chunk_size < ((u32)1 << 23)) {
+               chunks_per_msg = MAX_CHUNKS_PER_MSG;
+               msgs_per_thread = 2;
+       } else {
+               /* Big chunks: Just have one buffer per thread --- more would
+                * just waste memory.  */
+               chunks_per_msg = 1;
+               msgs_per_thread = 1;
+       }
        for (;;) {
                approx_mem_required =
                        (u64)chunks_per_msg *
                        (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;
 
@@ -466,12 +484,11 @@ 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)
@@ -493,14 +510,10 @@ 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;
-               dat->out_ctype = out_ctype;
-               dat->out_chunk_size = out_chunk_size;
-               if (out_ctype == WIMLIB_COMPRESSION_TYPE_LZX) {
-                       ret = wimlib_lzx_alloc_context(out_chunk_size, NULL,
-                                                      &dat->comp_ctx);
-                       if (ret)
-                               goto err;
-               }
+               ret = wimlib_create_compressor(out_ctype, out_chunk_size, 0,
+                                              &dat->compressor);
+               if (ret)
+                       goto err;
        }
 
        for (ctx->num_started_threads = 0;
@@ -515,14 +528,20 @@ 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)
@@ -541,3 +560,5 @@ err:
        parallel_chunk_compressor_destroy(&ctx->base);
        return ret;
 }
+
+#endif /* ENABLE_MULTITHREADED_COMPRESSION */