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.
105 * TODO: Rewrite this to work on cloud log items rather than inodes, so we can
106 * better track which logs are fully synchronized to the cloud and can be
107 * garbage collected if needed? */
108 static void flushd_cloud(BlueSkyFS *fs)
110 int64_t start_time = bluesky_get_current_time();
111 g_mutex_lock(fs->lock);
115 if (fs->dirty_list.prev == NULL)
117 inode = fs->dirty_list.prev->data;
119 if (bluesky_verbose) {
120 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
121 "Flushing inode %"PRIu64" to cloud", inode->inum);
124 /* Stop processing dirty inodes if we both have enough memory available
125 * and the oldest inode is sufficiently new that it need not be flushed
127 uint64_t elapsed = bluesky_get_current_time() - inode->change_time;
128 if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty
129 && elapsed < WRITEBACK_DELAY)
131 if (inode->change_time > start_time)
134 bluesky_inode_ref(inode);
136 g_mutex_unlock(fs->lock);
138 g_mutex_lock(inode->lock);
139 flushd_dirty_inode(inode);
140 g_mutex_lock(fs->lock);
141 bluesky_list_unlink(&fs->dirty_list, inode->dirty_list);
142 inode->dirty_list = NULL;
143 g_mutex_unlock(fs->lock);
145 BlueSkyCloudLog *log = inode->committed_item;
146 bluesky_cloudlog_ref(log);
147 g_mutex_unlock(inode->lock);
150 bluesky_cloudlog_serialize(log, fs);
151 bluesky_inode_unref(inode);
152 bluesky_cloudlog_unref(log);
154 g_mutex_lock(fs->lock);
157 g_mutex_unlock(fs->lock);
158 bluesky_cloudlog_flush(fs);
161 /* Drop cached data for a given inode, if it is clean. inode must be locked. */
162 static void drop_caches(BlueSkyInode *inode)
164 if (inode->type == BLUESKY_REGULAR)
165 bluesky_file_drop_cached(inode);
167 BlueSkyCloudLog *log = inode->committed_item;
169 g_mutex_lock(log->lock);
170 if (log->data != NULL
171 && g_atomic_int_get(&log->data_lock_count) == 0
172 && (log->location_flags != 0))
174 bluesky_cloudlog_stats_update(log, -1);
175 bluesky_string_unref(log->data);
177 bluesky_cloudlog_stats_update(log, 1);
179 if (log->location_flags & CLOUDLOG_CLOUD) {
180 log->location_flags &= ~CLOUDLOG_JOURNAL;
182 g_mutex_unlock(log->lock);
186 /* Drop clean data from the cache if needed. Clean data should generally be
187 * memory-mapped from log file or similar, so the kernel can drop this clean
188 * data from memory for us and hence memory management isn't too important.
189 * Mainly, we'll want to drop references to data that hasn't been accessed in a
190 * while so that it is possible to reclaim log segments on disk. */
191 static void flushd_clean(BlueSkyFS *fs)
193 g_mutex_lock(fs->lock);
195 size_t inode_count = g_hash_table_size(fs->inodes);
199 while (inode_count-- > 0) {
201 if (fs->accessed_list.prev == NULL)
203 inode = fs->accessed_list.prev->data;
205 uint64_t elapsed = bluesky_get_current_time() - inode->access_time;
206 if (elapsed < CACHE_DROP_DELAY)
209 if (bluesky_verbose) {
210 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
211 "Considering dropping cached data for inode %"PRIu64,
215 bluesky_inode_ref(inode);
217 g_mutex_unlock(fs->lock);
219 g_mutex_lock(inode->lock);
221 g_mutex_lock(fs->lock);
222 bluesky_list_unlink(&fs->accessed_list, inode->accessed_list);
223 inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode);
224 g_mutex_unlock(fs->lock);
228 g_mutex_unlock(inode->lock);
229 bluesky_inode_unref(inode);
231 g_mutex_lock(fs->lock);
234 g_mutex_unlock(fs->lock);
237 /* Run the flush daemon for a single iteration, though if it is already
238 * executing returns immediately. */
239 static gpointer flushd_task(BlueSkyFS *fs)
241 if (!g_mutex_trylock(fs->flushd_lock))
246 g_mutex_unlock(fs->flushd_lock);
251 void bluesky_flushd_invoke(BlueSkyFS *fs)
253 g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL);
256 void bluesky_flushd_invoke_conditional(BlueSkyFS *fs)
258 if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty)
261 if (bluesky_verbose) {
262 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
263 "Too much data; invoking flushd: dirty=%d",
264 g_atomic_int_get(&fs->cache_dirty));
267 bluesky_flushd_invoke(fs);
269 /* If the system is under heavy memory pressure, actually delay execution
270 * so the flush daemon can catch up. */
271 while (g_atomic_int_get(&fs->cache_dirty)
272 + g_atomic_int_get(&fs->cache_log_dirty)
273 > bluesky_watermark_high_dirty) {
274 g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG,
275 "Waiting due to memory pressure, dirty=%d + %d",
276 g_atomic_int_get(&fs->cache_dirty),
277 g_atomic_int_get(&fs->cache_log_dirty));
278 g_mutex_lock(fs->lock);
279 g_cond_wait(fs->flushd_cond, fs->lock);
280 g_mutex_unlock(fs->lock);
284 /* Start a perpetually-running thread that flushes the cache occasionally. */
285 static gpointer flushd_thread(BlueSkyFS *fs)
288 bluesky_flushd_invoke(fs);
289 struct timespec delay;
292 nanosleep(&delay, NULL);
298 void bluesky_flushd_thread_launch(BlueSkyFS *fs)
300 g_thread_create((GThreadFunc)flushd_thread, fs, FALSE, NULL);