X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Fcache.c;h=a4b9a0cd6cca6f49b91218133ad3edb9add5bbfa;hb=4bd76c48487520c2b1c08ff67009f752e0bde941;hp=9287c978220277bc01a20d491b41e3c1876f559e;hpb=4a207da5eef9f87a702a011d003d5ad17b651085;p=bluesky.git diff --git a/bluesky/cache.c b/bluesky/cache.c index 9287c97..a4b9a0c 100644 --- a/bluesky/cache.c +++ b/bluesky/cache.c @@ -14,6 +14,7 @@ #include "bluesky-private.h" #define WRITEBACK_DELAY (20 * 1000000) +#define CACHE_DROP_DELAY (20 * 1000000) /* Filesystem caching and cache coherency. There are actually a couple of * different tasks that are performed here: @@ -43,6 +44,17 @@ static void flushd_dirty_inode(BlueSkyInode *inode) bluesky_inode_start_sync(inode); } +/* Check whether memory usage may have dropped below critical thresholds for + * waking up waiting threads. */ +void flushd_check_wakeup(BlueSkyFS *fs) +{ + int dirty = g_atomic_int_get(&fs->cache_dirty); + dirty += g_atomic_int_get(&fs->cache_log_dirty); + + if (dirty <= bluesky_watermark_high_dirty) + g_cond_broadcast(fs->flushd_cond); +} + /* Try to flush dirty data to disk, either due to memory pressure or due to * timeouts. */ static void flushd_dirty(BlueSkyFS *fs) @@ -81,12 +93,18 @@ static void flushd_dirty(BlueSkyFS *fs) bluesky_inode_unref(inode); g_mutex_lock(fs->lock); + flushd_check_wakeup(fs); } + g_cond_broadcast(fs->flushd_cond); + g_mutex_unlock(fs->lock); } -/* Try to flush dirty data to the cloud. */ +/* Try to flush dirty data to the cloud. + * TODO: Rewrite this to work on cloud log items rather than inodes, so we can + * better track which logs are fully synchronized to the cloud and can be + * garbage collected if needed? */ static void flushd_cloud(BlueSkyFS *fs) { int64_t start_time = bluesky_get_current_time(); @@ -131,7 +149,7 @@ static void flushd_cloud(BlueSkyFS *fs) if (log != NULL) bluesky_cloudlog_serialize(log, fs); bluesky_inode_unref(inode); - bluesky_cloudlog_ref(log); + bluesky_cloudlog_unref(log); g_mutex_lock(fs->lock); } @@ -145,9 +163,28 @@ static void drop_caches(BlueSkyInode *inode) { if (inode->type == BLUESKY_REGULAR) bluesky_file_drop_cached(inode); + + BlueSkyCloudLog *log = inode->committed_item; + if (log != NULL) { + g_mutex_lock(log->lock); + if (log->data != NULL + && g_atomic_int_get(&log->data_lock_count) == 0 + && (log->location_flags != 0)) + { + bluesky_cloudlog_stats_update(log, -1); + bluesky_string_unref(log->data); + log->data = NULL; + bluesky_cloudlog_stats_update(log, 1); + } + g_mutex_unlock(log->lock); + } } -/* Drop clean data from the cache if needed due to memory pressure. */ +/* Drop clean data from the cache if needed. Clean data should generally be + * memory-mapped from log file or similar, so the kernel can drop this clean + * data from memory for us and hence memory management isn't too important. + * Mainly, we'll want to drop references to data that hasn't been accessed in a + * while so that it is possible to reclaim log segments on disk. */ static void flushd_clean(BlueSkyFS *fs) { g_mutex_lock(fs->lock); @@ -157,16 +194,15 @@ static void flushd_clean(BlueSkyFS *fs) inode_count = 1; while (inode_count-- > 0) { -#if 0 - if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total) - break; -#endif - BlueSkyInode *inode; if (fs->accessed_list.prev == NULL) break; inode = fs->accessed_list.prev->data; + uint64_t elapsed = bluesky_get_current_time() - inode->access_time; + if (elapsed < CACHE_DROP_DELAY) + break; + if (bluesky_verbose) { g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG, "Considering dropping cached data for inode %"PRIu64, @@ -216,8 +252,7 @@ void bluesky_flushd_invoke(BlueSkyFS *fs) void bluesky_flushd_invoke_conditional(BlueSkyFS *fs) { - if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_high_dirty - /*&& g_atomic_int_get(&fs->cache_total) < bluesky_watermark_high_total*/) + if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_medium_dirty) return; if (bluesky_verbose) { @@ -227,6 +262,20 @@ void bluesky_flushd_invoke_conditional(BlueSkyFS *fs) } bluesky_flushd_invoke(fs); + + /* If the system is under heavy memory pressure, actually delay execution + * so the flush daemon can catch up. */ + while (g_atomic_int_get(&fs->cache_dirty) + + g_atomic_int_get(&fs->cache_log_dirty) + > bluesky_watermark_high_dirty) { + g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG, + "Waiting due to memory pressure, dirty=%d + %d", + g_atomic_int_get(&fs->cache_dirty), + g_atomic_int_get(&fs->cache_log_dirty)); + g_mutex_lock(fs->lock); + g_cond_wait(fs->flushd_cond, fs->lock); + g_mutex_unlock(fs->lock); + } } /* Start a perpetually-running thread that flushes the cache occasionally. */