1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2010 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
10 #define _ATFILE_SOURCE
18 #include <sys/types.h>
23 #include "bluesky-private.h"
25 /* The logging layer for BlueSky. This is used to write filesystem changes
26 * durably to disk so that they can be recovered in the event of a system
29 /* The logging layer takes care out writing out a sequence of log records to
30 * disk. On disk, each record consists of a header, a data payload, and a
31 * footer. The footer contains a checksum of the record, meant to help with
32 * identifying corrupt log records (we would assume because the log record was
33 * only incompletely written out before a crash, which should only happen for
34 * log records that were not considered committed). */
36 // Rough size limit for a log segment. This is not a firm limit and there are
37 // no absolute guarantees on the size of a log segment.
38 #define LOG_SEGMENT_SIZE (1 << 23)
40 #define HEADER_MAGIC 0x676f4c0a
41 #define FOOTER_MAGIC 0x2e435243
44 uint32_t magic; // HEADER_MAGIC
45 uint64_t offset; // Starting byte offset of the log header
46 uint32_t keysize; // Size of the log key (bytes)
47 uint32_t size; // Size of the data item (bytes)
48 } __attribute__((packed));
51 uint32_t magic; // FOOTER_MAGIC
52 uint32_t crc; // Computed from log_header to log_footer.magic
53 } __attribute__((packed));
55 static void writebuf(int fd, const char *buf, size_t len)
59 written = write(fd, buf, len);
60 if (written < 0 && errno == EINTR)
62 g_assert(written >= 0);
68 /* All log writes (at least for a single log) are made by one thread, so we
69 * don't need to worry about concurrent access to the log file. Log items to
70 * write are pulled off a queue (and so may be posted by any thread).
71 * fdatasync() is used to ensure the log items are stable on disk.
73 * The log is broken up into separate files, roughly of size LOG_SEGMENT_SIZE
74 * each. If a log segment is not currently open (log->fd is negative), a new
75 * one is created. Log segment filenames are assigned sequentially.
77 * Log replay ought to be implemented later, and ought to set the initial
78 * sequence number appropriately.
80 static gpointer log_thread(gpointer d)
82 BlueSkyLog *log = (BlueSkyLog *)d;
84 /* If there are multiple log items to write, we may write more than one
85 * before calling fsync(). The committed list is used to track all the
86 * items that should be marked as committed once that final fsync() is
88 GSList *committed = NULL;
90 int dirfd = open(log->log_directory, O_DIRECTORY);
92 fprintf(stderr, "Unable to open logging directory: %m\n");
99 g_snprintf(logfile, sizeof(logfile), "log-%08d", log->seq_num);
100 log->fd = openat(dirfd, logfile, O_CREAT|O_WRONLY|O_EXCL, 0600);
101 if (log->fd < 0 && errno == EEXIST) {
102 fprintf(stderr, "Log file %s already exists...\n", logfile);
105 } else if (log->fd < 0) {
106 fprintf(stderr, "Error opening logfile %s: %m\n", logfile);
113 BlueSkyLogItem *item = (BlueSkyLogItem *)g_async_queue_pop(log->queue);
114 g_mutex_lock(item->lock);
116 off_t logsize = lseek(log->fd, 0, SEEK_CUR);
117 struct log_header header;
118 struct log_footer footer;
120 header.magic = GUINT32_TO_LE(HEADER_MAGIC);
121 header.offset = GUINT64_TO_LE(logsize);
122 header.keysize = GUINT32_TO_LE(strlen(item->key));
123 header.size = GUINT32_TO_LE(item->data->len);
124 footer.magic = GUINT32_TO_LE(FOOTER_MAGIC);
126 uint32_t crc = BLUESKY_CRC32C_SEED;
128 writebuf(log->fd, (const char *)&header, sizeof(header));
129 crc = crc32c(crc, (const char *)&header, sizeof(header));
131 writebuf(log->fd, item->key, strlen(item->key));
132 crc = crc32c(crc, item->key, strlen(item->key));
134 writebuf(log->fd, item->data->data, item->data->len);
135 crc = crc32c(crc, item->data->data, item->data->len);
137 crc = crc32c(crc, (const char *)&footer,
138 sizeof(footer) - sizeof(uint32_t));
139 footer.crc = crc32c_finalize(crc);
140 writebuf(log->fd, (const char *)&footer, sizeof(footer));
142 committed = g_slist_prepend(committed, item);
144 /* Force an fsync either if we will be closing this log segment and
145 * opening a new file, or if there are no other log items currently
146 * waiting to be written. */
147 logsize += strlen(item->key) + item->data->len + sizeof(header) + sizeof(footer);
149 if (logsize >= LOG_SEGMENT_SIZE
150 || g_async_queue_length(log->queue) <= 0)
154 while (committed != NULL) {
155 item = (BlueSkyLogItem *)committed->data;
156 item->committed = TRUE;
157 g_cond_signal(item->cond);
158 g_mutex_unlock(item->lock);
159 committed = g_slist_delete_link(committed, committed);
162 /* if (batchsize > 1)
163 g_print("Log batch size: %d\n", batchsize); */
166 if (logsize < 0 || logsize >= LOG_SEGMENT_SIZE) {
176 BlueSkyLog *bluesky_log_new(const char *log_directory)
178 BlueSkyLog *log = g_new0(BlueSkyLog, 1);
180 log->log_directory = g_strdup(log_directory);
183 log->queue = g_async_queue_new();
185 g_thread_create(log_thread, log, FALSE, NULL);
190 BlueSkyLogItem *bluesky_log_item_new()
192 BlueSkyLogItem *item = g_new(BlueSkyLogItem, 1);
193 item->committed = FALSE;
194 item->lock = g_mutex_new();
195 item->cond = g_cond_new();
201 void bluesky_log_item_submit(BlueSkyLogItem *item, BlueSkyLog *log)
203 g_async_queue_push(log->queue, item);
206 static void bluesky_log_item_free(BlueSkyLogItem *item)
209 bluesky_string_unref(item->data);
210 g_mutex_free(item->lock);
211 g_cond_free(item->cond);
215 void bluesky_log_item_finish(BlueSkyLogItem *item)
217 g_mutex_lock(item->lock);
218 while (!item->committed)
219 g_cond_wait(item->cond, item->lock);
220 g_mutex_unlock(item->lock);
221 bluesky_log_item_free(item);