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 int journal_seq_start = fs->log->seq_num;
121 if (fs->dirty_list.prev == NULL)
123 inode = fs->dirty_list.prev->data;
125 if (bluesky_verbose) {
126 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
127 "Flushing inode %"PRIu64" to cloud", inode->inum);
130 bluesky_inode_ref(inode);
132 g_mutex_unlock(fs->lock);
134 g_mutex_lock(inode->lock);
135 g_assert(inode->change_cloud == inode->change_commit);
136 g_mutex_lock(fs->lock);
137 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
138 inode->dirty_list = NULL;
139 g_mutex_unlock(fs->lock);
141 BlueSkyCloudLog *log = inode->committed_item;
142 inode->committed_item = NULL;
143 g_mutex_unlock(inode->lock);
146 bluesky_cloudlog_serialize(log, fs);
147 bluesky_inode_unref(inode);
148 bluesky_cloudlog_unref(log);
150 g_mutex_lock(fs->lock);
153 g_mutex_unlock(fs->lock);
154 bluesky_cloudlog_flush(fs);
156 /* Wait until all segments have been written to the cloud, so that it
157 * becomes safe to free up journal segments. */
158 while (fs->log_state->pending_segments != NULL) {
159 SerializedRecord *segment
160 = (SerializedRecord *)fs->log_state->pending_segments->data;
161 g_mutex_lock(segment->lock);
162 while (!segment->complete)
163 g_cond_wait(segment->cond, segment->lock);
164 g_mutex_unlock(segment->lock);
166 g_mutex_free(segment->lock);
167 g_cond_free(segment->cond);
170 fs->log_state->pending_segments
171 = g_list_delete_link(fs->log_state->pending_segments,
172 fs->log_state->pending_segments);
175 g_print("All segments have been flushed, journal < %d is clean\n",
178 fs->log->journal_watermark = journal_seq_start;
181 /* Drop cached data for a given inode, if it is clean. inode must be locked. */
182 static void drop_caches(BlueSkyInode *inode)
184 if (inode->type == BLUESKY_REGULAR)
185 bluesky_file_drop_cached(inode);
187 BlueSkyCloudLog *log = inode->committed_item;
189 g_mutex_lock(log->lock);
190 if (log->data != NULL
191 && g_atomic_int_get(&log->data_lock_count) == 0
192 && (log->location_flags != 0))
194 bluesky_cloudlog_stats_update(log, -1);
195 bluesky_string_unref(log->data);
197 bluesky_cloudlog_stats_update(log, 1);
199 if (log->location_flags & CLOUDLOG_CLOUD) {
200 log->location_flags &= ~CLOUDLOG_JOURNAL;
202 g_mutex_unlock(log->lock);
206 /* Drop clean data from the cache if needed. Clean data should generally be
207 * memory-mapped from log file or similar, so the kernel can drop this clean
208 * data from memory for us and hence memory management isn't too important.
209 * Mainly, we'll want to drop references to data that hasn't been accessed in a
210 * while so that it is possible to reclaim log segments on disk. */
211 static void flushd_clean(BlueSkyFS *fs)
213 g_mutex_lock(fs->lock);
215 size_t inode_count = g_hash_table_size(fs->inodes);
219 while (inode_count-- > 0) {
221 if (fs->accessed_list.prev == NULL)
223 inode = fs->accessed_list.prev->data;
225 uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
226 if (elapsed < CACHE_DROP_DELAY)
229 if (bluesky_verbose) {
230 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
231 "Considering dropping cached data for inode %"PRIu64,
235 bluesky_inode_ref(inode);
237 g_mutex_unlock(fs->lock);
239 g_mutex_lock(inode->lock);
241 g_mutex_lock(fs->lock);
242 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
243 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
244 g_mutex_unlock(fs->lock);
248 g_mutex_unlock(inode->lock);
249 bluesky_inode_unref(inode);
251 g_mutex_lock(fs->lock);
254 g_mutex_unlock(fs->lock);
257 /* Run the flush daemon for a single iteration, though if it is already
258 * executing returns immediately. */
259 static gpointer flushd_task(BlueSkyFS *fs)
261 if (!g_mutex_trylock(fs->flushd_lock))
266 bluesky_cachefile_gc(fs);
267 g_mutex_unlock(fs->flushd_lock);
272 void bluesky_flushd_invoke(BlueSkyFS *fs)
274 g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
277 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
279 if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
282 if (bluesky_verbose) {
283 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
284 "Too much data; invoking flushd: dirty=%d",
285 g_atomic_int_get(&fs->cache_dirty));
288 bluesky_flushd_invoke(fs);
290 /* If the system is under heavy memory pressure, actually delay execution
291 * so the flush daemon can catch up. */
292 while (g_atomic_int_get(&fs->cache_dirty)
293 + g_atomic_int_get(&fs->cache_log_dirty)
294 > bluesky_watermark_high_dirty) {
295 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
296 "Waiting due to memory pressure, dirty=%d + %d",
297 g_atomic_int_get(&fs->cache_dirty),
298 g_atomic_int_get(&fs->cache_log_dirty));
299 g_mutex_lock(fs->lock);
300 g_cond_wait(fs->flushd_cond, fs->lock);
301 g_mutex_unlock(fs->lock);
305 /* Start a perpetually-running thread that flushes the cache occasionally. */
306 static gpointer flushd_thread(BlueSkyFS *fs)
309 bluesky_flushd_invoke(fs);
310 struct timespec delay;
313 nanosleep(&delay, NULL);
319 void bluesky_flushd_thread_launch(BlueSkyFS *fs)
321 g_thread_create((GThreadFunc)flushd_thread, fs, FALSE, NULL);