X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Finode.c;h=e15d5b74667f16fb1187fbb964ef907db810f31b;hb=966a43ffed492d387e0d75e56c3984d9001c15b1;hp=8faafa412b9861e5dfabca5258407c7e44828369;hpb=3ffb892cbb0b96a57595b9962d018191c5b3d5af;p=bluesky.git

diff --git a/bluesky/inode.c b/bluesky/inode.c
index 8faafa4..e15d5b7 100644
--- a/bluesky/inode.c
+++ b/bluesky/inode.c
@@ -8,10 +8,11 @@
 
 #include <stdio.h>
 #include <stdint.h>
+#include <inttypes.h>
 #include <glib.h>
 #include <string.h>
 
-#include "bluesky.h"
+#include "bluesky-private.h"
 
 /* Core filesystem.  Different proxies, such as the NFSv3 one, interface to
  * this, but the core actually tracks the data which is stored.  So far we just
@@ -28,7 +29,8 @@ int64_t bluesky_get_current_time()
 
 /* Update an inode to indicate that a modification was made.  This increases
  * the change counter, updates the ctime to the current time, and optionally
- * updates the mtime.  inode must already be locked. */
+ * updates the mtime.  This also makes the inode contents subject to writeback
+ * to storage in the future.  inode must already be locked. */
 void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
 {
     int64_t now = bluesky_get_current_time();
@@ -36,6 +38,19 @@ void bluesky_inode_update_ctime(BlueSkyInode *inode, gboolean update_mtime)
     inode->ctime = now;
     if (update_mtime)
         inode->mtime = now;
+
+    if (inode->change_time == 0)
+        inode->change_time = now;
+
+    if (bluesky_options.writethrough_cache)
+        bluesky_file_flush(inode, NULL);
+
+    g_mutex_lock(inode->fs->lock);
+    bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
+    inode->dirty_list = bluesky_list_prepend(&inode->fs->dirty_list, inode);
+    bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+    inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
+    g_mutex_unlock(inode->fs->lock);
 }
 
 /* Unfortunately a glib hash table is only guaranteed to be able to store
@@ -68,6 +83,39 @@ 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;
+}
+
+BlueSkyFS *bluesky_init_fs(gchar *name, BlueSkyStore *store)
+{
+    BlueSkyRCStr *data = bluesky_store_get(store, "superblock");
+    if (data != NULL) {
+        BlueSkyFS *fs = bluesky_deserialize_superblock(data->data);
+        if (fs != NULL) {
+            fs->store = store;
+            g_print("Loaded filesystem superblock\n");
+            g_free(fs->name);
+            fs->name = g_strdup(name);
+            return fs;
+        }
+        bluesky_string_unref(data);
+    }
+
+    g_print("Initializing fresh filesystem\n");
+    BlueSkyFS *fs = bluesky_new_fs(name);
+    fs->store = store;
+
+    BlueSkyInode *root = bluesky_new_inode(BLUESKY_ROOT_INUM, fs,
+                                           BLUESKY_DIRECTORY);
+    root->nlink = 1;
+    root->mode = 0755;
+    bluesky_insert_inode(fs, root);
+    bluesky_inode_update_ctime(root, TRUE);
+
+    bluesky_inode_flush(fs, root);
+    bluesky_superblock_flush(fs);
 
     return fs;
 }
@@ -81,8 +129,60 @@ void bluesky_inode_ref(BlueSkyInode *inode)
 void bluesky_inode_unref(BlueSkyInode *inode)
 {
     if (g_atomic_int_dec_and_test(&inode->refcount)) {
-        g_error("Reference count for inode %lld dropped to zero!\n",
-                inode->inum);
+        if (bluesky_verbose) {
+            g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+                  "Reference count for inode %"PRIu64" dropped to zero.",
+                  inode->inum);
+        }
+
+        /* Sanity check: Is the inode clean? */
+        if (inode->change_commit < inode->change_count
+                || inode->accessed_list != NULL
+                || inode->dirty_list != NULL) {
+            g_warning("Dropping inode which is not clean (commit %"PRIi64" < change %"PRIi64"; accessed_list = %p; dirty_list = %p)\n", inode->change_commit, inode->change_count, inode->accessed_list, inode->dirty_list);
+        }
+
+        /* These shouldn't be needed, but in case the above warning fires and
+         * we delete the inode anyway, we ought to be sure the inode is not on
+         * any LRU list. */
+        g_mutex_lock(inode->fs->lock);
+        bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+        bluesky_list_unlink(&inode->fs->dirty_list, inode->dirty_list);
+        g_mutex_unlock(inode->fs->lock);
+
+        /* Free file type specific data.  It should be an error for there to be
+         * dirty data to commit when the reference count has reaches zero. */
+        switch (inode->type) {
+        case BLUESKY_REGULAR:
+            for (int i = 0; i < inode->blocks->len; i++) {
+                BlueSkyBlock *b = &g_array_index(inode->blocks,
+                                                 BlueSkyBlock, i);
+                if (b->type == BLUESKY_BLOCK_DIRTY) {
+                    g_error("Deleting an inode with dirty file data!");
+                }
+                g_free(b->ref);
+                bluesky_string_unref(b->data);
+            }
+            g_array_unref(inode->blocks);
+            break;
+
+        case BLUESKY_DIRECTORY:
+            g_hash_table_destroy(inode->dirhash);
+            g_hash_table_destroy(inode->dirhash_folded);
+            g_sequence_free(inode->dirents);
+            break;
+
+        case BLUESKY_SYMLINK:
+            g_free(inode->symlink_contents);
+            break;
+
+        default:
+            break;
+        }
+
+        g_mutex_free(inode->lock);
+
+        g_free(inode);
     }
 }
 
@@ -95,19 +195,17 @@ uint64_t bluesky_fs_alloc_inode(BlueSkyFS *fs)
     inum = fs->next_inum;
     fs->next_inum++;
 
+    bluesky_superblock_flush(fs);
+
     return inum;
 }
 
-BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
-                                BlueSkyFileType type)
+/* Perform type-specification initialization of an inode.  Normally performed
+ * in bluesky_new_inode, but can be separated if an inode is created first,
+ * then deserialized. */
+void bluesky_init_inode(BlueSkyInode *i, BlueSkyFileType type)
 {
-    BlueSkyInode *i = g_new0(BlueSkyInode, 1);
-
-    i->lock = g_mutex_new();
-    i->refcount = 1;
     i->type = type;
-    i->fs = fs;
-    i->inum = inum;
 
     switch (type) {
     case BLUESKY_REGULAR:
@@ -116,26 +214,40 @@ BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
     case BLUESKY_DIRECTORY:
         i->dirents = g_sequence_new(bluesky_dirent_destroy);
         i->dirhash = g_hash_table_new(g_str_hash, g_str_equal);
+        i->dirhash_folded = g_hash_table_new(g_str_hash, g_str_equal);
         break;
-    case BLUESKY_BLOCK:
-    case BLUESKY_CHARACTER:
-    case BLUESKY_SYMLINK:
-    case BLUESKY_SOCKET:
-    case BLUESKY_FIFO:
+    default:
         break;
     }
+}
+
+BlueSkyInode *bluesky_new_inode(uint64_t inum, BlueSkyFS *fs,
+                                BlueSkyFileType type)
+{
+    BlueSkyInode *i = g_new0(BlueSkyInode, 1);
+
+    i->lock = g_mutex_new();
+    i->refcount = 1;
+    i->fs = fs;
+    i->inum = inum;
+    i->change_count = 1;
+    bluesky_init_inode(i, type);
 
     return i;
 }
 
 /* Retrieve an inode from the filesystem.  Eventually this will be a cache and
  * so we might need to go fetch the inode from elsewhere; for now all
- * filesystem state is stored here.  inode is returned locked with a reference
- * held. */
+ * filesystem state is stored here.  inode is returned with a reference held
+ * but not locked. */
 BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
 {
     BlueSkyInode *inode = NULL;
 
+    if (inum == 0) {
+        return NULL;
+    }
+
     g_mutex_lock(fs->lock);
     inode = (BlueSkyInode *)g_hash_table_lookup(fs->inodes, &inum);
 
@@ -145,8 +257,11 @@ BlueSkyInode *bluesky_get_inode(BlueSkyFS *fs, uint64_t inum)
     }
 
     if (inode != NULL) {
-        g_mutex_lock(inode->lock);
         bluesky_inode_ref(inode);
+
+        /* FIXME: We assume we can atomically update the in-memory access time
+         * without a lock. */
+        inode->access_time = bluesky_get_current_time();
     }
 
     g_mutex_unlock(fs->lock);
@@ -160,33 +275,145 @@ void bluesky_insert_inode(BlueSkyFS *fs, BlueSkyInode *inode)
     g_hash_table_insert(fs->inodes, &inode->inum, inode);
 }
 
-/* Synchronize an inode to stable storage. */
+/* Deprecated: Synchronize an inode to stable storage. */
 void bluesky_inode_flush(BlueSkyFS *fs, BlueSkyInode *inode)
 {
     GString *buf = g_string_new("");
     bluesky_serialize_inode(buf, inode);
+    BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
+
+    char key[64];
+    sprintf(key, "inode-%016"PRIx64, inode->inum);
+
+    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+    async->op = STORE_OP_PUT;
+    async->key = g_strdup(key);
+    async->data = data;
+    bluesky_store_async_submit(async);
+    bluesky_store_async_unref(async);
+}
+
+/* Start writeback of an inode and all associated data. */
+void bluesky_inode_start_sync(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
+{
+    BlueSkyFS *fs = inode->fs;
 
-    gsize len = buf->len;
-    BlueSkyRCStr *data = bluesky_string_new(g_string_free(buf, FALSE), len);
+    if (inode->type == BLUESKY_REGULAR)
+        bluesky_file_flush(inode, barrier);
+
+    GString *buf = g_string_new("");
+    bluesky_serialize_inode(buf, inode);
+    BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
 
     char key[64];
-    sprintf(key, "inode-%016llx", inode->inum);
+    sprintf(key, "inode-%016"PRIx64, inode->inum);
+
+    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+    async->op = STORE_OP_PUT;
+    async->key = g_strdup(key);
+    async->data = data;
+    bluesky_store_async_submit(async);
+    if (barrier != NULL)
+        bluesky_store_add_barrier(barrier, async);
+    bluesky_store_async_unref(async);
+}
+
+/* Write back an inode and all associated data and wait for completion.  Inode
+ * should already be locked. */
+void bluesky_inode_do_sync(BlueSkyInode *inode)
+{
+    BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
+    barrier->op = STORE_OP_BARRIER;
+
+    if (bluesky_verbose) {
+        g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+            "Synchronous writeback for inode %"PRIu64"...", inode->inum);
+    }
+    bluesky_inode_start_sync(inode, barrier);
+    bluesky_store_async_submit(barrier);
+    bluesky_store_async_wait(barrier);
+    bluesky_store_async_unref(barrier);
+    if (bluesky_verbose) {
+        g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+              "Writeback for inode %"PRIu64" complete", inode->inum);
+    }
+}
+
+static void complete_inode_fetch(BlueSkyStoreAsync *async, BlueSkyInode *inode)
+{
+    if (bluesky_verbose) {
+        g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+              "Completing fetch of inode %"PRIu64"...", inode->inum);
+    }
 
-    bluesky_store_put(fs->store, key, data);
+    if (async->result != 0
+        || !bluesky_deserialize_inode(inode, async->data->data))
+    {
+        if (bluesky_verbose) {
+            g_log("bluesky/inode", G_LOG_LEVEL_DEBUG,
+                  "    failed to load inode, cleaning up");
+        }
+        g_mutex_lock(inode->fs->lock);
+        g_hash_table_remove(inode->fs->inodes, &inode->inum);
+        bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+        inode->accessed_list = NULL;
+        g_mutex_unlock(inode->fs->lock);
+        bluesky_inode_unref(inode);
+    }
+
+    inode->access_time = bluesky_get_current_time();
+    g_mutex_lock(inode->fs->lock);
+    bluesky_list_unlink(&inode->fs->accessed_list, inode->accessed_list);
+    inode->accessed_list = bluesky_list_prepend(&inode->fs->accessed_list, inode);
+    g_mutex_unlock(inode->fs->lock);
+
+    g_mutex_unlock(inode->lock);
+    bluesky_inode_unref(inode);
 }
 
-/* Fetch an inode from stable storage. */
+/* Fetch an inode from stable storage.  The fetch can be performed
+ * asynchronously: the in-memory inode is allocated, but not filled with data
+ * immediately.  It is kept locked until it has been filled in, so any users
+ * should try to acquire the lock on the inode before accessing any data.  The
+ * fs lock must be held. */
 void bluesky_inode_fetch(BlueSkyFS *fs, uint64_t inum)
 {
     char key[64];
-    sprintf(key, "inode-%016llx", inum);
-    BlueSkyRCStr *data = bluesky_store_get(fs->store, key);
-    if (data == NULL)
-        return;
+    sprintf(key, "inode-%016"PRIx64, inum);
 
-    BlueSkyInode *inode = bluesky_deserialize_inode(fs, data->data);
-    if (inode != NULL) {
-        bluesky_insert_inode(fs, inode);
-        g_print("Loaded inode %lld\n", (long long)inum);
+    BlueSkyInode *inode = bluesky_new_inode(inum, fs, BLUESKY_PENDING);
+    inode->change_count = 0;
+    bluesky_inode_ref(inode);       // Extra ref held by fetching process
+    g_mutex_lock(inode->lock);
+    bluesky_insert_inode(fs, inode);
+
+    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+    async->op = STORE_OP_GET;
+    async->key = g_strdup(key);
+
+    bluesky_store_async_add_notifier(async, (GFunc)complete_inode_fetch, inode);
+    bluesky_store_async_submit(async);
+
+    if (bluesky_options.sync_inode_fetches) {
+        bluesky_store_async_wait(async);
     }
+
+    bluesky_store_async_unref(async);
+}
+
+/* Synchronize filesystem superblock to stable storage. */
+void bluesky_superblock_flush(BlueSkyFS *fs)
+{
+    GString *buf = g_string_new("");
+    bluesky_serialize_superblock(buf, fs);
+    BlueSkyRCStr *data = bluesky_string_new_from_gstring(buf);
+
+    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
+    async->op = STORE_OP_PUT;
+    async->key = g_strdup("superblock");
+    async->data = data;
+    bluesky_store_async_submit(async);
+    bluesky_store_async_unref(async);
+
+    bluesky_store_sync(fs->store);
 }