Improve the reading back of objects committed to the journal.
[bluesky.git] / bluesky / log.c
index a4a8150..9b4b62b 100644 (file)
@@ -19,6 +19,7 @@
 #include <sys/stat.h>
 #include <fcntl.h>
 #include <unistd.h>
+#include <sys/mman.h>
 
 #include "bluesky-private.h"
 
 
 // Rough size limit for a log segment.  This is not a firm limit and there are
 // no absolute guarantees on the size of a log segment.
-#define LOG_SEGMENT_SIZE (1 << 20)
+#define LOG_SEGMENT_SIZE (1 << 23)
+
+#define HEADER_MAGIC 0x676f4c0a
+#define FOOTER_MAGIC 0x2e435243
+
+struct log_header {
+    uint32_t magic;             // HEADER_MAGIC
+    uint64_t offset;            // Starting byte offset of the log header
+    uint32_t size;              // Size of the data item (bytes)
+    BlueSkyCloudID id;          // Object identifier
+} __attribute__((packed));
+
+struct log_footer {
+    uint32_t magic;             // FOOTER_MAGIC
+    uint32_t crc;               // Computed from log_header to log_footer.magic
+} __attribute__((packed));
 
 static void writebuf(int fd, const char *buf, size_t len)
 {
@@ -50,6 +66,61 @@ static void writebuf(int fd, const char *buf, size_t len)
     }
 }
 
+static void log_commit(BlueSkyLog *log)
+{
+    int batchsize = 0;
+
+    if (log->fd < 0)
+        return;
+
+    fdatasync(log->fd);
+    while (log->committed != NULL) {
+        BlueSkyCloudLog *item = (BlueSkyCloudLog *)log->committed->data;
+        g_mutex_lock(item->lock);
+        item->pending_write &= ~CLOUDLOG_JOURNAL;
+        item->location_flags |= CLOUDLOG_JOURNAL;
+        g_cond_signal(item->cond);
+        g_mutex_unlock(item->lock);
+        log->committed = g_slist_delete_link(log->committed, log->committed);
+        batchsize++;
+    }
+
+    if (bluesky_verbose && batchsize > 1)
+        g_print("Log batch size: %d\n", batchsize);
+}
+
+static gboolean log_open(BlueSkyLog *log)
+{
+    char logname[64];
+
+    if (log->fd >= 0) {
+        log_commit(log);
+        close(log->fd);
+        log->seq_num++;
+        log->fd = -1;
+    }
+
+    while (log->fd < 0) {
+        g_snprintf(logname, sizeof(logname), "log-%08d", log->seq_num);
+        log->fd = openat(log->dirfd, logname, O_CREAT|O_WRONLY|O_EXCL, 0600);
+        if (log->fd < 0 && errno == EEXIST) {
+            fprintf(stderr, "Log file %s already exists...\n", logname);
+            log->seq_num++;
+            continue;
+        } else if (log->fd < 0) {
+            fprintf(stderr, "Error opening logfile %s: %m\n", logname);
+            return FALSE;
+        }
+    }
+
+    if (ftruncate(log->fd, LOG_SEGMENT_SIZE) < 0) {
+        fprintf(stderr, "Unable to truncate logfile %s: %m\n", logname);
+    }
+    fsync(log->fd);
+    fsync(log->dirfd);
+    return TRUE;
+}
+
 /* All log writes (at least for a single log) are made by one thread, so we
  * don't need to worry about concurrent access to the log file.  Log items to
  * write are pulled off a queue (and so may be posted by any thread).
@@ -66,44 +137,74 @@ static gpointer log_thread(gpointer d)
 {
     BlueSkyLog *log = (BlueSkyLog *)d;
 
-    int dirfd = open(log->log_directory, O_DIRECTORY);
-    if (dirfd < 0) {
-        fprintf(stderr, "Unable to open logging directory: %m\n");
-        return NULL;
-    }
-
     while (TRUE) {
         if (log->fd < 0) {
-            char logfile[64];
-            g_snprintf(logfile, sizeof(logfile), "log-%08d", log->seq_num);
-            log->fd = openat(dirfd, logfile, O_CREAT|O_WRONLY|O_EXCL, 0600);
-            if (log->fd < 0 && errno == EEXIST) {
-                fprintf(stderr, "Log file %s already exists...\n", logfile);
-                log->seq_num++;
-                continue;
-            } else if (log->fd < 0) {
-                fprintf(stderr, "Error opening logfile %s: %m\n", logfile);
+            if (!log_open(log)) {
                 return NULL;
             }
-            fsync(log->fd);
-            fsync(dirfd);
         }
 
-        BlueSkyLogItem *item = (BlueSkyLogItem *)g_async_queue_pop(log->queue);
+        BlueSkyCloudLog *item
+            = (BlueSkyCloudLog *)g_async_queue_pop(log->queue);
         g_mutex_lock(item->lock);
-        writebuf(log->fd, item->key, strlen(item->key));
+        g_assert(item->data != NULL);
+
+        /* The item may have already been written to the journal... */
+        if ((item->location_flags | item->pending_write) & CLOUDLOG_JOURNAL) {
+            g_mutex_unlock(item->lock);
+            bluesky_cloudlog_unref(item);
+            continue;
+        }
+
+        item->pending_write |= CLOUDLOG_JOURNAL;
+
+        struct log_header header;
+        struct log_footer footer;
+        size_t size = sizeof(header) + sizeof(footer) + item->data->len;
+        off_t offset = 0;
+        if (log->fd >= 0)
+            offset = lseek(log->fd, 0, SEEK_CUR);
+
+        /* Check whether the item would overflow the allocated journal size.
+         * If so, start a new log segment.  We only allow oversized log
+         * segments if they contain a single log entry. */
+        if (offset + size >= LOG_SEGMENT_SIZE && offset > 0) {
+            log_open(log);
+            offset = 0;
+        }
+
+        header.magic = GUINT32_TO_LE(HEADER_MAGIC);
+        header.offset = GUINT64_TO_LE(offset);
+        header.size = GUINT32_TO_LE(item->data->len);
+        header.id = item->id;
+        footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
+
+        uint32_t crc = BLUESKY_CRC32C_SEED;
+
+        writebuf(log->fd, (const char *)&header, sizeof(header));
+        crc = crc32c(crc, (const char *)&header, sizeof(header));
+
         writebuf(log->fd, item->data->data, item->data->len);
-        fdatasync(log->fd);
-        item->committed = TRUE;
-        g_cond_signal(item->cond);
+        crc = crc32c(crc, item->data->data, item->data->len);
+
+        crc = crc32c(crc, (const char *)&footer,
+                     sizeof(footer) - sizeof(uint32_t));
+        footer.crc = crc32c_finalize(crc);
+        writebuf(log->fd, (const char *)&footer, sizeof(footer));
+
+        item->log_seq = log->seq_num;
+        item->log_offset = offset + sizeof(header);
+        item->log_size = item->data->len;
+
+        offset += sizeof(header) + sizeof(footer) + item->data->len;
+
+        log->committed  = g_slist_prepend(log->committed, item);
         g_mutex_unlock(item->lock);
 
-        off_t logsize = lseek(log->fd, 0, SEEK_CUR);
-        if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) {
-            close(log->fd);
-            log->fd = -1;
-            log->seq_num++;
-        }
+        /* Force an if there are no other log items currently waiting to be
+         * written. */
+        if (g_async_queue_length(log->queue) <= 0)
+            log_commit(log);
     }
 
     return NULL;
@@ -117,42 +218,81 @@ BlueSkyLog *bluesky_log_new(const char *log_directory)
     log->fd = -1;
     log->seq_num = 0;
     log->queue = g_async_queue_new();
+    log->mmap_lock = g_mutex_new();
+    log->mmap_cache = g_hash_table_new(NULL, NULL);
+
+    log->dirfd = open(log->log_directory, O_DIRECTORY);
+    if (log->dirfd < 0) {
+        fprintf(stderr, "Unable to open logging directory: %m\n");
+        return NULL;
+    }
 
     g_thread_create(log_thread, log, FALSE, NULL);
 
     return log;
 }
 
-BlueSkyLogItem *bluesky_log_item_new()
-{
-    BlueSkyLogItem *item = g_new(BlueSkyLogItem, 1);
-    item->committed = FALSE;
-    item->lock = g_mutex_new();
-    item->cond = g_cond_new();
-    item->key = NULL;
-    item->data = NULL;
-    return item;
-}
-
-void bluesky_log_item_submit(BlueSkyLogItem *item, BlueSkyLog *log)
+void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
 {
+    bluesky_cloudlog_ref(item);
     g_async_queue_push(log->queue, item);
 }
 
-static void bluesky_log_item_free(BlueSkyLogItem *item)
+void bluesky_log_finish_all(GList *log_items)
 {
-    g_free(item->key);
-    bluesky_string_unref(item->data);
-    g_mutex_free(item->lock);
-    g_cond_free(item->cond);
-    g_free(item);
+    while (log_items != NULL) {
+        BlueSkyCloudLog *item = (BlueSkyCloudLog *)log_items->data;
+
+        g_mutex_lock(item->lock);
+        while ((item->pending_write & CLOUDLOG_JOURNAL))
+            g_cond_wait(item->cond, item->lock);
+        g_mutex_unlock(item->lock);
+        bluesky_cloudlog_unref(item);
+
+        log_items = g_list_delete_link(log_items, log_items);
+    }
 }
 
-void bluesky_log_item_finish(BlueSkyLogItem *item)
+/* Memory-map the given log object into memory (read-only) and return a pointer
+ * to it. */
+static int page_size = 0;
+
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyLog *log,
+                                     int log_seq, int log_offset, int log_size)
 {
-    g_mutex_lock(item->lock);
-    while (!item->committed)
-        g_cond_wait(item->cond, item->lock);
-    g_mutex_unlock(item->lock);
-    bluesky_log_item_free(item);
+    if (page_size == 0) {
+        page_size = getpagesize();
+    }
+
+    BlueSkyMmap *map;
+    g_mutex_lock(log->mmap_lock);
+    map = g_hash_table_lookup(log->mmap_cache, GINT_TO_POINTER(log_seq));
+
+    if (map == NULL) {
+        char logname[64];
+        g_snprintf(logname, sizeof(logname), "log-%08d", log_seq);
+        int fd = openat(log->dirfd, logname, O_RDONLY);
+
+        if (fd < 0) {
+            fprintf(stderr, "Error opening logfile %s: %m\n", logname);
+            g_mutex_unlock(log->mmap_lock);
+            return NULL;
+        }
+
+        map = g_new0(BlueSkyMmap, 1);
+
+        off_t length = lseek(fd, 0, SEEK_END);
+        map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
+                                       fd, 0);
+        map->len = length;
+        g_atomic_int_set(&map->refcount, 1);
+
+        g_hash_table_insert(log->mmap_cache, GINT_TO_POINTER(log_seq), map);
+
+        close(fd);
+    }
+
+    g_mutex_unlock(log->mmap_lock);
+
+    return bluesky_string_new_from_mmap(map, log_offset, log_size);
 }