1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
14 #include "bluesky-private.h"
16 #define WRITEBACK_DELAY (20 * 1000000)
17 #define CACHE_DROP_DELAY (20 * 1000000)
19 /* Filesystem caching and cache coherency. There are actually a couple of
20 * different tasks that are performed here:
21 * - Forcing data to the log if needed to reclaim memory or simply if the
22 * data has been dirty in memory long enough.
23 * - Writing batches of data to the cloud.
26 static void flushd_dirty_inode(BlueSkyInode *inode)
28 BlueSkyFS *fs = inode->fs;
30 g_mutex_lock(fs->lock);
31 bluesky_list_unlink(&fs->unlogged_list, inode->unlogged_list);
32 inode->unlogged_list = NULL;
33 g_mutex_unlock(fs->lock);
35 /* Inode is clean; nothing to do. */
36 if (inode->change_count == inode->change_commit)
39 if (bluesky_verbose) {
40 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
41 "Starting flush of inode %"PRIu64, inode->inum);
44 bluesky_inode_start_sync(inode);
47 /* Check whether memory usage may have dropped below critical thresholds for
48 * waking up waiting threads. */
49 void flushd_check_wakeup(BlueSkyFS *fs)
51 int dirty = g_atomic_int_get(&fs->cache_dirty);
52 dirty += g_atomic_int_get(&fs->cache_log_dirty);
54 if (dirty <= bluesky_watermark_high_dirty)
55 g_cond_broadcast(fs->flushd_cond);
58 /* Try to flush dirty data to disk, either due to memory pressure or due to
60 static void flushd_dirty(BlueSkyFS *fs)
62 int64_t start_time = bluesky_get_current_time();
63 g_mutex_lock(fs->lock);
67 if (fs->unlogged_list.prev == NULL)
69 inode = fs->unlogged_list.prev->data;
71 if (bluesky_verbose) {
72 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
73 "Considering flushing inode %"PRIu64, inode->inum);
76 /* Stop processing dirty inodes if we both have enough memory available
77 * and the oldest inode is sufficiently new that it need not be flushed
79 uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
80 if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty
81 && elapsed < WRITEBACK_DELAY)
83 if (inode->change_time > start_time)
86 bluesky_inode_ref(inode);
88 g_mutex_unlock(fs->lock);
90 g_mutex_lock(inode->lock);
91 flushd_dirty_inode(inode);
92 g_mutex_unlock(inode->lock);
93 bluesky_inode_unref(inode);
95 g_mutex_lock(fs->lock);
96 flushd_check_wakeup(fs);
99 g_cond_broadcast(fs->flushd_cond);
101 g_mutex_unlock(fs->lock);
104 /* Try to flush dirty data to the cloud. This will take a snapshot of the
105 * entire filesystem (though only point-in-time consistent for isolated inodes
106 * and not the filesystem as a whole) and ensure all data is written to the
107 * cloud. When the write completes, we will allow old journal segments (those
108 * that were fully written _before_ the snapshot process started) to be garbage
109 * collected. Newer journal segments can't be collected yet since they may
110 * still contain data which has not been written persistently to the cloud. */
111 static void flushd_cloud(BlueSkyFS *fs)
113 g_mutex_lock(fs->lock);
115 /* TODO: Locking? Since we're reading a single variable this is probably
116 * atomic but a lock could be safer. */
117 BlueSkyCloudLog *marker = bluesky_log_get_commit_point(fs);
118 int journal_seq_start = fs->log->seq_num;
122 if (fs->dirty_list.prev == NULL)
124 inode = fs->dirty_list.prev->data;
126 if (bluesky_verbose) {
127 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
128 "Flushing inode %"PRIu64" to cloud", inode->inum);
131 bluesky_inode_ref(inode);
133 g_mutex_unlock(fs->lock);
135 g_mutex_lock(inode->lock);
136 g_assert(inode->change_cloud == inode->change_commit);
137 g_mutex_lock(fs->lock);
138 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
139 inode->dirty_list = NULL;
140 g_mutex_unlock(fs->lock);
142 BlueSkyCloudLog *log = inode->committed_item;
143 inode->committed_item = NULL;
144 g_mutex_unlock(inode->lock);
147 bluesky_cloudlog_serialize(log, fs);
148 bluesky_inode_unref(inode);
149 bluesky_cloudlog_unref(log);
151 g_mutex_lock(fs->lock);
154 /* Write out any updated inode map entries, so that all inodes just written
155 * can be located, and then a final commit record. */
156 BlueSkyCloudLog *commit_record = bluesky_inode_map_serialize(fs);
157 bluesky_cloudlog_serialize(commit_record, fs);
159 g_mutex_unlock(fs->lock);
160 bluesky_cloudlog_flush(fs);
162 /* Wait until all segments have been written to the cloud, so that it
163 * becomes safe to free up journal segments. */
164 while (fs->log_state->pending_segments != NULL) {
165 SerializedRecord *segment
166 = (SerializedRecord *)fs->log_state->pending_segments->data;
167 g_mutex_lock(segment->lock);
168 while (!segment->complete)
169 g_cond_wait(segment->cond, segment->lock);
170 g_mutex_unlock(segment->lock);
172 g_mutex_free(segment->lock);
173 g_cond_free(segment->cond);
176 fs->log_state->pending_segments
177 = g_list_delete_link(fs->log_state->pending_segments,
178 fs->log_state->pending_segments);
181 bluesky_log_write_commit_point(fs, marker);
183 g_print("All segments have been flushed, journal < %d is clean\n",
186 fs->log->journal_watermark = journal_seq_start;
189 /* Drop cached data for a given inode, if it is clean. inode must be locked. */
190 static void drop_caches(BlueSkyInode *inode)
192 if (inode->type == BLUESKY_REGULAR)
193 bluesky_file_drop_cached(inode);
195 BlueSkyCloudLog *log = inode->committed_item;
197 g_mutex_lock(log->lock);
198 if (log->data != NULL
199 && g_atomic_int_get(&log->data_lock_count) == 0
200 && (log->location_flags != 0))
202 bluesky_cloudlog_stats_update(log, -1);
203 bluesky_string_unref(log->data);
205 bluesky_cloudlog_stats_update(log, 1);
207 if (log->location_flags & CLOUDLOG_CLOUD) {
208 log->location_flags &= ~CLOUDLOG_JOURNAL;
210 g_mutex_unlock(log->lock);
214 /* Drop clean data from the cache if needed. Clean data should generally be
215 * memory-mapped from log file or similar, so the kernel can drop this clean
216 * data from memory for us and hence memory management isn't too important.
217 * Mainly, we'll want to drop references to data that hasn't been accessed in a
218 * while so that it is possible to reclaim log segments on disk. */
219 static void flushd_clean(BlueSkyFS *fs)
221 g_mutex_lock(fs->lock);
223 size_t inode_count = g_hash_table_size(fs->inodes);
227 while (inode_count-- > 0) {
229 if (fs->accessed_list.prev == NULL)
231 inode = fs->accessed_list.prev->data;
233 uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
234 if (elapsed < CACHE_DROP_DELAY)
237 if (bluesky_verbose) {
238 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
239 "Considering dropping cached data for inode %"PRIu64,
243 bluesky_inode_ref(inode);
245 g_mutex_unlock(fs->lock);
247 g_mutex_lock(inode->lock);
249 g_mutex_lock(fs->lock);
250 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
251 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
252 g_mutex_unlock(fs->lock);
256 g_mutex_unlock(inode->lock);
257 bluesky_inode_unref(inode);
259 g_mutex_lock(fs->lock);
262 g_mutex_unlock(fs->lock);
265 /* Run the flush daemon for a single iteration, though if it is already
266 * executing returns immediately. */
267 static gpointer flushd_task(BlueSkyFS *fs)
269 if (!g_mutex_trylock(fs->flushd_lock))
274 bluesky_cachefile_gc(fs);
275 g_mutex_unlock(fs->flushd_lock);
280 void bluesky_flushd_invoke(BlueSkyFS *fs)
282 g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
285 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
287 if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
290 if (bluesky_verbose) {
291 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
292 "Too much data; invoking flushd: dirty=%d",
293 g_atomic_int_get(&fs->cache_dirty));
296 bluesky_flushd_invoke(fs);
298 /* If the system is under heavy memory pressure, actually delay execution
299 * so the flush daemon can catch up. */
300 while (g_atomic_int_get(&fs->cache_dirty)
301 + g_atomic_int_get(&fs->cache_log_dirty)
302 > bluesky_watermark_high_dirty) {
303 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
304 "Waiting due to memory pressure, dirty=%d + %d",
305 g_atomic_int_get(&fs->cache_dirty),
306 g_atomic_int_get(&fs->cache_log_dirty));
307 g_mutex_lock(fs->lock);
308 g_cond_wait(fs->flushd_cond, fs->lock);
309 g_mutex_unlock(fs->lock);
313 /* Start a perpetually-running thread that flushes the cache occasionally. */
314 static gpointer flushd_thread(BlueSkyFS *fs)
317 bluesky_flushd_invoke(fs);
318 struct timespec delay;
321 nanosleep(&delay, NULL);
327 void bluesky_flushd_thread_launch(BlueSkyFS *fs)
329 g_thread_create((GThreadFunc)flushd_thread, fs, FALSE, NULL);