X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Flog.c;h=66b4edb9131da2ad5151ae0af83158b49ca3aa64;hb=b8885660a453f0f1f30c9fed196b01a397482a19;hp=ee96976ff1865606e963cba9563bd02f548bdca9;hpb=3bf92d26e3256d76766e77106502a68c9e30c14b;p=bluesky.git

diff --git a/bluesky/log.c b/bluesky/log.c
index ee96976..66b4edb 100644
--- a/bluesky/log.c
+++ b/bluesky/log.c
@@ -19,6 +19,7 @@
 #include <sys/stat.h>
 #include <fcntl.h>
 #include <unistd.h>
+#include <sys/mman.h>
 
 #include "bluesky-private.h"
 
@@ -35,7 +36,7 @@
 
 // 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 << 23)
+#define LOG_SEGMENT_SIZE (1 << 24)
 
 #define HEADER_MAGIC 0x676f4c0a
 #define FOOTER_MAGIC 0x2e435243
@@ -43,8 +44,8 @@
 struct log_header {
     uint32_t magic;             // HEADER_MAGIC
     uint64_t offset;            // Starting byte offset of the log header
-    uint32_t keysize;           // Size of the log key (bytes)
     uint32_t size;              // Size of the data item (bytes)
+    BlueSkyCloudID id;          // Object identifier
 } __attribute__((packed));
 
 struct log_footer {
@@ -65,6 +66,73 @@ 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);
+        bluesky_cloudlog_stats_update(item, -1);
+        item->pending_write &= ~CLOUDLOG_JOURNAL;
+        item->location_flags |= CLOUDLOG_JOURNAL;
+        bluesky_cloudlog_stats_update(item, 1);
+        g_cond_signal(item->cond);
+        g_mutex_unlock(item->lock);
+        log->committed = g_slist_delete_link(log->committed, log->committed);
+        bluesky_cloudlog_unref(item);
+        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;
+    }
+
+    if (log->current_log != NULL) {
+        bluesky_cachefile_unref(log->current_log);
+        log->current_log = NULL;
+    }
+
+    while (log->fd < 0) {
+        g_snprintf(logname, sizeof(logname), "journal-%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;
+        }
+    }
+
+    log->current_log = bluesky_cachefile_lookup(log->fs, -1, log->seq_num);
+    g_assert(log->current_log != NULL);
+    g_mutex_unlock(log->current_log->lock);
+
+    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).
@@ -81,46 +149,49 @@ static gpointer log_thread(gpointer d)
 {
     BlueSkyLog *log = (BlueSkyLog *)d;
 
-    /* If there are multiple log items to write, we may write more than one
-     * before calling fsync().  The committed list is used to track all the
-     * items that should be marked as committed once that final fsync() is
-     * done. */
-    GSList *committed = NULL;
-
-    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);
+        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);
+            g_atomic_int_add(&item->data_lock_count, -1);
+            continue;
+        }
+
+        bluesky_cloudlog_stats_update(item, -1);
+        item->pending_write |= CLOUDLOG_JOURNAL;
+        bluesky_cloudlog_stats_update(item, 1);
 
-        off_t logsize = lseek(log->fd, 0, SEEK_CUR);
         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(logsize);
-        header.keysize = GUINT32_TO_LE(strlen(item->key));
+        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;
@@ -128,9 +199,6 @@ static gpointer log_thread(gpointer d)
         writebuf(log->fd, (const char *)&header, sizeof(header));
         crc = crc32c(crc, (const char *)&header, sizeof(header));
 
-        writebuf(log->fd, item->key, strlen(item->key));
-        crc = crc32c(crc, item->key, strlen(item->key));
-
         writebuf(log->fd, item->data->data, item->data->len);
         crc = crc32c(crc, item->data->data, item->data->len);
 
@@ -139,35 +207,35 @@ static gpointer log_thread(gpointer d)
         footer.crc = crc32c_finalize(crc);
         writebuf(log->fd, (const char *)&footer, sizeof(footer));
 
-        committed  = g_slist_prepend(committed, item);
-
-        /* Force an fsync either if we will be closing this log segment and
-         * opening a new file, or if there are no other log items currently
-         * waiting to be written. */
-        logsize += strlen(item->key) + item->data->len + sizeof(header) + sizeof(footer);
-
-        if (logsize >= LOG_SEGMENT_SIZE
-            || g_async_queue_length(log->queue) <= 0)
-        {
-            int batchsize = 0;
-            fdatasync(log->fd);
-            while (committed != NULL) {
-                item = (BlueSkyLogItem *)committed->data;
-                item->committed = TRUE;
-                g_cond_signal(item->cond);
-                g_mutex_unlock(item->lock);
-                committed = g_slist_delete_link(committed, committed);
-                batchsize++;
-            }
-            /* if (batchsize > 1)
-                g_print("Log batch size: %d\n", batchsize); */
-        }
-
-        if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) {
-            close(log->fd);
-            log->fd = -1;
-            log->seq_num++;
-        }
+        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;
+
+        /* Since we have just written a new dirty object to the journal,
+         * increment the count of live dirty objects in that journal file.  The
+         * count will be decremented when objects are deleted or written to the
+         * cloud. */
+        g_atomic_int_add(&log->current_log->dirty_refs, 1);
+        item->dirty_journal = log->current_log;
+
+        /* Replace the log item's string data with a memory-mapped copy of the
+         * data, now that it has been written to the log file.  (Even if it
+         * isn't yet on disk, it should at least be in the page cache and so
+         * available to memory map.) */
+        bluesky_string_unref(item->data);
+        item->data = NULL;
+        bluesky_cloudlog_fetch(item);
+
+        log->committed  = g_slist_prepend(log->committed, item);
+        g_atomic_int_add(&item->data_lock_count, -1);
+        g_mutex_unlock(item->lock);
+
+        /* 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;
@@ -181,42 +249,284 @@ 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(g_str_hash, g_str_equal);
+
+    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()
+void bluesky_log_item_submit(BlueSkyCloudLog *item, BlueSkyLog *log)
 {
-    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;
+    bluesky_cloudlog_ref(item);
+    g_atomic_int_add(&item->data_lock_count, 1);
+    g_async_queue_push(log->queue, item);
 }
 
-void bluesky_log_item_submit(BlueSkyLogItem *item, BlueSkyLog *log)
+void bluesky_log_finish_all(GList *log_items)
 {
-    g_async_queue_push(log->queue, 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);
+    }
+}
+
+/* Memory-map the given log object into memory (read-only) and return a pointer
+ * to it. */
+static int page_size = 0;
+
+void bluesky_cachefile_unref(BlueSkyCacheFile *cachefile)
+{
+    g_atomic_int_add(&cachefile->refcount, -1);
+}
+
+static void cloudlog_fetch_complete(BlueSkyStoreAsync *async,
+                                    BlueSkyCacheFile *cachefile)
+{
+    g_print("Fetch of %s from cloud complete, status = %d\n",
+            async->key, async->result);
+
+    if (async->result < 0)
+        return;
+
+    g_mutex_lock(cachefile->lock);
+    char *pathname = g_strdup_printf("%s/%s", cachefile->log->log_directory,
+                                     cachefile->filename);
+    if (!g_file_set_contents(pathname, async->data->data, async->data->len,
+                             NULL))
+    {
+        g_print("Error writing out fetched file to cache!\n");
+    }
+    g_free(pathname);
+    cachefile->fetching = FALSE;
+    cachefile->ready = TRUE;
+    bluesky_cachefile_unref(cachefile);
+    g_cond_broadcast(cachefile->cond);
+    g_mutex_unlock(cachefile->lock);
+}
+
+/* Find the BlueSkyCacheFile object for the given journal or cloud log segment.
+ * Returns the object in the locked state and with a reference taken. */
+BlueSkyCacheFile *bluesky_cachefile_lookup(BlueSkyFS *fs,
+                                           int clouddir, int log_seq)
+{
+    if (page_size == 0) {
+        page_size = getpagesize();
+    }
+
+    BlueSkyLog *log = fs->log;
+
+    char logname[64];
+    int type;
+
+    // A request for a local log file
+    if (clouddir < 0) {
+        sprintf(logname, "journal-%08d", log_seq);
+        type = CLOUDLOG_JOURNAL;
+    } else {
+        sprintf(logname, "log-%08d-%08d", clouddir, log_seq);
+        type = CLOUDLOG_CLOUD;
+    }
+
+    BlueSkyCacheFile *map;
+    g_mutex_lock(log->mmap_lock);
+    map = g_hash_table_lookup(log->mmap_cache, logname);
+
+    if (map == NULL) {
+        /* TODO: stat() call */
+        map = g_new0(BlueSkyCacheFile, 1);
+        map->type = CLOUDLOG_JOURNAL;
+        map->lock = g_mutex_new();
+        map->type = type;
+        g_mutex_lock(map->lock);
+        map->cond = g_cond_new();
+        map->filename = g_strdup(logname);
+        map->log_seq = log_seq;
+        map->log = log;
+        g_atomic_int_set(&map->mapcount, 0);
+        g_atomic_int_set(&map->refcount, 0);
+
+        g_hash_table_insert(log->mmap_cache, map->filename, map);
+
+        // If the log file is stored in the cloud, we may need to fetch it
+        if (clouddir >= 0) {
+            g_atomic_int_inc(&map->refcount);
+            map->fetching = TRUE;
+            g_print("Starting fetch of %s from cloud\n", logname);
+            BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+            async->op = STORE_OP_GET;
+            async->key = g_strdup(logname);
+            bluesky_store_async_add_notifier(async,
+                                             (GFunc)cloudlog_fetch_complete,
+                                             map);
+            bluesky_store_async_submit(async);
+            bluesky_store_async_unref(async);
+        }
+    } else {
+        g_mutex_lock(map->lock);
+    }
+
+    g_mutex_unlock(log->mmap_lock);
+    g_atomic_int_inc(&map->refcount);
+    return map;
+}
+
+BlueSkyRCStr *bluesky_log_map_object(BlueSkyFS *fs, int log_dir,
+                                     int log_seq, int log_offset, int log_size)
+{
+    if (page_size == 0) {
+        page_size = getpagesize();
+    }
+
+    BlueSkyLog *log = fs->log;
+    BlueSkyCacheFile *map = bluesky_cachefile_lookup(fs, log_dir, log_seq);
+
+    if (map == NULL) {
+        return NULL;
+    }
+
+    if (map->addr == NULL) {
+        while (!map->ready && map->fetching) {
+            g_print("Waiting for log segment to be fetched from cloud...\n");
+            g_cond_wait(map->cond, map->lock);
+        }
+
+        int fd = openat(log->dirfd, map->filename, O_RDONLY);
+
+        if (fd < 0) {
+            fprintf(stderr, "Error opening logfile %s: %m\n", map->filename);
+            bluesky_cachefile_unref(map);
+            g_mutex_unlock(map->lock);
+            return NULL;
+        }
+
+        off_t length = lseek(fd, 0, SEEK_END);
+        map->addr = (const char *)mmap(NULL, length, PROT_READ, MAP_SHARED,
+                                       fd, 0);
+        g_atomic_int_add(&log->disk_used, -(map->len / 1024));
+        map->len = length;
+        g_atomic_int_add(&log->disk_used, map->len / 1024);
+
+        g_print("Re-mapped log segment %d...\n", log_seq);
+        g_atomic_int_inc(&map->refcount);
+
+        close(fd);
+    }
+
+    g_mutex_unlock(log->mmap_lock);
+
+    BlueSkyRCStr *str;
+    map->atime = bluesky_get_current_time();
+    str = bluesky_string_new_from_mmap(map, log_offset, log_size);
+    bluesky_cachefile_unref(map);
+    g_mutex_unlock(map->lock);
+    return str;
 }
 
-static void bluesky_log_item_free(BlueSkyLogItem *item)
+void bluesky_mmap_unref(BlueSkyCacheFile *mmap)
 {
-    g_free(item->key);
-    bluesky_string_unref(item->data);
-    g_mutex_free(item->lock);
-    g_cond_free(item->cond);
-    g_free(item);
+    if (mmap == NULL)
+        return;
+
+    if (g_atomic_int_dec_and_test(&mmap->mapcount)) {
+        g_mutex_lock(mmap->lock);
+        if (g_atomic_int_get(&mmap->mapcount) == 0) {
+            g_print("Unmapped log segment %d...\n", mmap->log_seq);
+            munmap((void *)mmap->addr, mmap->len);
+            mmap->addr = NULL;
+            g_atomic_int_add(&mmap->refcount, -1);
+        }
+        g_mutex_unlock(mmap->lock);
+    }
+}
+
+/* Scan through all currently-stored files in the journal/cache and garbage
+ * collect old unused ones, if needed. */
+static void gather_cachefiles(gpointer key, gpointer value, gpointer user_data)
+{
+    GList **files = (GList **)user_data;
+    *files = g_list_prepend(*files, value);
+}
+
+static gint compare_cachefiles(gconstpointer a, gconstpointer b)
+{
+    int64_t ta, tb;
+
+    ta = ((BlueSkyCacheFile *)a)->atime;
+    tb = ((BlueSkyCacheFile *)b)->atime;
+    if (ta < tb)
+        return -1;
+    else if (ta > tb)
+        return 1;
+    else
+        return 0;
 }
 
-void bluesky_log_item_finish(BlueSkyLogItem *item)
+void bluesky_cachefile_gc(BlueSkyFS *fs)
 {
-    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);
+    GList *files = NULL;
+
+    g_mutex_lock(fs->log->mmap_lock);
+    g_hash_table_foreach(fs->log->mmap_cache, gather_cachefiles, &files);
+
+    /* Sort based on atime.  The atime should be stable since it shouln't be
+     * updated except by threads which can grab the mmap_lock, which we already
+     * hold. */
+    files = g_list_sort(files, compare_cachefiles);
+
+    /* Walk the list of files, starting with the oldest, deleting files if
+     * possible until enough space has been reclaimed. */
+    g_print("\nScanning cache: (total size = %d kB)\n", fs->log->disk_used);
+    while (files != NULL) {
+        BlueSkyCacheFile *cachefile = (BlueSkyCacheFile *)files->data;
+        /* Try to lock the structure, but if the lock is held by another thread
+         * then we'll just skip the file on this pass. */
+        if (g_mutex_trylock(cachefile->lock)) {
+            int64_t age = bluesky_get_current_time() - cachefile->atime;
+            g_print("%s addr=%p mapcount=%d refcount=%d dirty=%d atime_age=%f",
+                    cachefile->filename, cachefile->addr, cachefile->mapcount,
+                    cachefile->refcount, cachefile->dirty_refs, age / 1e6);
+            if (cachefile->fetching)
+                g_print(" (fetching)");
+            g_print("\n");
+
+            if (g_atomic_int_get(&cachefile->refcount) == 0
+                && g_atomic_int_get(&cachefile->mapcount) == 0
+                && g_atomic_int_get(&cachefile->dirty_refs) == 0)
+            {
+                g_print("   ...deleting\n");
+                if (unlinkat(fs->log->dirfd, cachefile->filename, 0) < 0) {
+                    fprintf(stderr, "Unable to unlink journal %s: %m\n",
+                            cachefile->filename);
+                }
+
+                g_hash_table_remove(fs->log->mmap_cache, cachefile->filename);
+                g_mutex_unlock(cachefile->lock);
+                g_mutex_free(cachefile->lock);
+                g_cond_free(cachefile->cond);
+                g_free(cachefile->filename);
+                g_free(cachefile);
+            } else {
+                g_mutex_unlock(cachefile->lock);
+            }
+        }
+        files = g_list_delete_link(files, files);
+    }
+    g_list_free(files);
+
+    g_mutex_unlock(fs->log->mmap_lock);
 }