From: Michael Vrable Date: Mon, 29 Mar 2010 19:48:03 +0000 (-0700) Subject: Rework cache flushing logic--this version should work much better. X-Git-Url: https://git.vrable.net/?a=commitdiff_plain;h=641411b0a7ce9afa309bea28c01320eda913b2d2;p=bluesky.git Rework cache flushing logic--this version should work much better. --- diff --git a/bluesky/bluesky-private.h b/bluesky/bluesky-private.h index 8d2e3fc..ba9df02 100644 --- a/bluesky/bluesky-private.h +++ b/bluesky/bluesky-private.h @@ -24,6 +24,10 @@ extern int bluesky_watermark_low_dirty; extern int bluesky_watermark_medium_dirty; extern int bluesky_watermark_high_dirty; +extern int bluesky_watermark_low_total; +extern int bluesky_watermark_medium_total; +extern int bluesky_watermark_high_total; + /* TODO: Make this go away entirely. */ BlueSkyFS *bluesky_new_fs(gchar *name); diff --git a/bluesky/bluesky.h b/bluesky/bluesky.h index 59a33c5..714ec55 100644 --- a/bluesky/bluesky.h +++ b/bluesky/bluesky.h @@ -129,6 +129,9 @@ typedef struct { * is held for list editing purposes. Items at the head of the list are * most recently accessed/modified. */ GList dirty_list, accessed_list; + + /* Mutex for the flush daemon, to prevent concurrent execution. */ + GMutex *flushd_lock; } BlueSkyFS; /* Inode number of the root directory. */ @@ -283,6 +286,7 @@ gint bluesky_dirent_compare(gconstpointer a, gconstpointer b, gpointer unused); void bluesky_flushd_invoke(BlueSkyFS *fs); +void bluesky_flushd_invoke_conditional(BlueSkyFS *fs); void bluesky_inode_do_sync(BlueSkyInode *inode); void bluesky_debug_dump(BlueSkyFS *fs); diff --git a/bluesky/cache.c b/bluesky/cache.c index 211c2b9..5f16ab0 100644 --- a/bluesky/cache.c +++ b/bluesky/cache.c @@ -13,8 +13,7 @@ #include "bluesky-private.h" -#define WRITEBACK_DELAY (5 * 1000000) -#define CACHE_CLEAN_DELAY (30 * 1000000) +#define WRITEBACK_DELAY (20 * 1000000) /* Filesystem caching and cache coherency. */ @@ -41,13 +40,7 @@ static void writeback_complete(gpointer a, gpointer i) g_mutex_unlock(inode->lock); } -/* Drop cached data for a given inode, if it is clean. inode must be locked. */ -static void drop_caches(BlueSkyInode *inode) -{ - if (inode->type == BLUESKY_REGULAR) - bluesky_file_drop_cached(inode); -} - +#if 0 static void flushd_inode(gpointer value, gpointer user_data) { BlueSkyFS *fs = (BlueSkyFS *)user_data; @@ -125,33 +118,162 @@ static void flushd_inode(gpointer value, gpointer user_data) g_mutex_unlock(inode->lock); bluesky_inode_unref(inode); } +#endif -/* Scan through the cache for dirty data and start flushing it to stable - * storage. This does not guarantee that data is committed when it returns. - * Instead, this can be called occasionally to ensure that dirty data is - * gradually flushed. - * - * We do not want to hold the filesystem lock while flushing individual inodes, - * a that could lead to deadlock. So first scan through the inode table to get - * a reference to all inodes, then process that queue of inodes after dropping - * the filesystem lock. */ -static void gather_inodes(gpointer key, gpointer value, gpointer user_data) +static void flushd_dirty_inode(BlueSkyInode *inode) { - GSList **list = (GSList **)user_data; - *list = g_slist_prepend(*list, value); - bluesky_inode_ref((BlueSkyInode *)value); + BlueSkyFS *fs = inode->fs; + + /* Inode is clean; nothing to do. */ + if (inode->change_count == inode->change_commit) + return; + + /* Inode writeback is in progress; put back on the dirty list. */ + if (inode->change_pending) { + /* Waiting for an earlier writeback to finish, so don't start a new + * writeback yet. */ + g_mutex_lock(fs->lock); + inode->change_time = bluesky_get_current_time(); + bluesky_list_unlink(&fs->dirty_list, inode->dirty_list); + inode->dirty_list = bluesky_list_prepend(&fs->dirty_list, inode); + g_mutex_unlock(fs->lock); + return; + } + + g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG, + "Starting flush of inode %"PRIu64, inode->inum); + inode->change_pending = inode->change_count; + + /* Create a store barrier. All operations part of the writeback will be + * added to this barrier, so when the barrier completes we know that the + * writeback is finished. */ + BlueSkyStoreAsync *barrier = bluesky_store_async_new(fs->store); + barrier->op = STORE_OP_BARRIER; + + bluesky_inode_start_sync(inode, barrier); + + bluesky_store_async_add_notifier(barrier, writeback_complete, inode); + bluesky_store_async_submit(barrier); + bluesky_store_async_unref(barrier); } -void bluesky_flushd_invoke(BlueSkyFS *fs) +/* Try to flush dirty data to disk, either due to memory pressure or due to + * timeouts. */ +static void flushd_dirty(BlueSkyFS *fs) { - GSList *list = NULL; + int64_t start_time = bluesky_get_current_time(); + g_mutex_lock(fs->lock); + + while (1) { + BlueSkyInode *inode; + if (fs->dirty_list.prev == NULL) + break; + inode = fs->dirty_list.prev->data; + + g_print("Considering flushing inode %"PRIu64"\n", inode->inum); + + /* Stop processing dirty inodes if we both have enough memory available + * and the oldest inode is sufficiently new that it need not be flushed + * out. */ + uint64_t elapsed = bluesky_get_current_time() - inode->change_time; + if (g_atomic_int_get(&fs->cache_dirty) < bluesky_watermark_low_dirty + && elapsed < WRITEBACK_DELAY) + break; + if (inode->change_time > start_time) + break; + bluesky_inode_ref(inode); + + bluesky_list_unlink(&fs->dirty_list, fs->dirty_list.prev); + inode->dirty_list = NULL; + + g_mutex_unlock(fs->lock); + + g_mutex_lock(inode->lock); + flushd_dirty_inode(inode); + g_mutex_unlock(inode->lock); + bluesky_inode_unref(inode); + + g_mutex_lock(fs->lock); + } + + g_mutex_unlock(fs->lock); +} + +/* Drop cached data for a given inode, if it is clean. inode must be locked. */ +static void drop_caches(BlueSkyInode *inode) +{ + if (inode->type == BLUESKY_REGULAR) + bluesky_file_drop_cached(inode); +} + +/* Drop clean data fromt the cache if needed due to memory pressure. */ +static void flushd_clean(BlueSkyFS *fs) +{ g_mutex_lock(fs->lock); - g_hash_table_foreach(fs->inodes, gather_inodes, &list); + + size_t inode_count = g_hash_table_size(fs->inodes); + if (!inode_count) + inode_count = 1; + + while (inode_count-- > 0) { + if (g_atomic_int_get(&fs->cache_total) < bluesky_watermark_medium_total) + break; + + BlueSkyInode *inode; + if (fs->accessed_list.prev == NULL) + break; + inode = fs->accessed_list.prev->data; + + g_print("Considering dropping cached data for inode %"PRIu64"\n", + inode->inum); + + bluesky_inode_ref(inode); + + bluesky_list_unlink(&fs->accessed_list, fs->accessed_list.prev); + inode->accessed_list = bluesky_list_prepend(&fs->accessed_list, inode); + + g_mutex_unlock(fs->lock); + + g_mutex_lock(inode->lock); + drop_caches(inode); + g_mutex_unlock(inode->lock); + bluesky_inode_unref(inode); + + g_mutex_lock(fs->lock); + } + g_mutex_unlock(fs->lock); +} + +/* Run the flush daemon for a single iteration, though if it is already + * executing returns immediately. */ +static gpointer flushd_task(BlueSkyFS *fs) +{ + if (!g_mutex_trylock(fs->flushd_lock)) + return NULL; + flushd_dirty(fs); + flushd_clean(fs); + g_mutex_unlock(fs->flushd_lock); - list = g_slist_reverse(list); - g_slist_foreach(list, flushd_inode, fs); + return NULL; +} + +void bluesky_flushd_invoke(BlueSkyFS *fs) +{ + g_thread_create((GThreadFunc)flushd_task, fs, FALSE, NULL); +} + +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) + return; + + g_log("bluesky/flushd", G_LOG_LEVEL_DEBUG, + "Too much data; invoking flushd: dirty=%d total=%d", + g_atomic_int_get(&fs->cache_dirty), + g_atomic_int_get(&fs->cache_total)); - g_slist_free(list); + bluesky_flushd_invoke(fs); } diff --git a/bluesky/init.c b/bluesky/init.c index eb2515c..520ed57 100644 --- a/bluesky/init.c +++ b/bluesky/init.c @@ -32,6 +32,10 @@ int bluesky_watermark_low_dirty = (64 << 20) / BLUESKY_BLOCK_SIZE; int bluesky_watermark_medium_dirty = (96 << 20) / BLUESKY_BLOCK_SIZE; int bluesky_watermark_high_dirty = (192 << 20) / BLUESKY_BLOCK_SIZE; +int bluesky_watermark_low_total = (64 << 20) / BLUESKY_BLOCK_SIZE; +int bluesky_watermark_medium_total = (128 << 20) / BLUESKY_BLOCK_SIZE; +int bluesky_watermark_high_total = (256 << 20) / BLUESKY_BLOCK_SIZE; + /* Environment variables that can be used to initialize settings. */ static struct { const char *env; diff --git a/bluesky/inode.c b/bluesky/inode.c index 39feff8..40531cd 100644 --- a/bluesky/inode.c +++ b/bluesky/inode.c @@ -83,6 +83,7 @@ BlueSkyFS *bluesky_new_fs(gchar *name) bluesky_fs_key_equal_func); fs->next_inum = BLUESKY_ROOT_INUM + 1; fs->store = bluesky_store_new("file"); + fs->flushd_lock = g_mutex_new(); return fs; } diff --git a/kvstore/kvservice.cc b/kvstore/kvservice.cc index 59e194a..b620f86 100644 --- a/kvstore/kvservice.cc +++ b/kvstore/kvservice.cc @@ -68,7 +68,7 @@ void KeyValueRpcService::PutValue( response->set_result(kvrpc::FAILURE); } - minimum_delay(&start, 1000000); + //minimum_delay(&start, 1000000); done->Run(); } @@ -92,7 +92,7 @@ void KeyValueRpcService::GetValue( response->set_result(kvrpc::FAILURE); } - minimum_delay(&start, 1000000); + //minimum_delay(&start, 1000000); done->Run(); } diff --git a/nfs3/nfs3.c b/nfs3/nfs3.c index b479107..ec5c736 100644 --- a/nfs3/nfs3.c +++ b/nfs3/nfs3.c @@ -328,6 +328,8 @@ void nfsproc3_read_3_svc(read3args *argp, RPCRequest *req) memset(&result, 0, sizeof(result)); char buf[NFS_MAXSIZE]; + bluesky_flushd_invoke_conditional(fs); + BlueSkyInode *inode = lookup_fh(req, &argp->file); if (inode == NULL) { result.status = NFS3ERR_STALE; @@ -372,6 +374,8 @@ void nfsproc3_write_3_svc(write3args *argp, RPCRequest *req) struct wcc_data wcc; memset(&wcc, 0, sizeof(wcc)); + bluesky_flushd_invoke_conditional(fs); + BlueSkyInode *inode = lookup_fh(req, &argp->file); if (inode == NULL) { result.status = NFS3ERR_STALE; @@ -380,6 +384,7 @@ void nfsproc3_write_3_svc(write3args *argp, RPCRequest *req) return; } +#if 0 /* FIXME: Hack to throttle writes when there is too much dirty data still * to be written out. */ while (g_atomic_int_get(&fs->cache_dirty) > 4096 @@ -392,6 +397,7 @@ void nfsproc3_write_3_svc(write3args *argp, RPCRequest *req) delay.tv_nsec = 0; nanosleep(&delay, NULL); } +#endif g_mutex_lock(inode->lock);