X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Ffile.c;h=94677f6520c5b0f314375500565d4f1d0ce8a306;hb=83fd6b61a6e092a22d4d5e59ed95f05f5e287f11;hp=66f84e7bf2e139dbea7a893648e575b4a7516c1c;hpb=7d9e06cc3154631da14e573648940c4d6176ae9a;p=bluesky.git

diff --git a/bluesky/file.c b/bluesky/file.c
index 66f84e7..94677f6 100644
--- a/bluesky/file.c
+++ b/bluesky/file.c
@@ -31,24 +31,30 @@ void bluesky_block_touch(BlueSkyInode *inode, uint64_t i)
 
     switch (block->type) {
     case BLUESKY_BLOCK_ZERO:
-        g_print("Allocating zero block of size %zd\n", block_len);
-        block->data = bluesky_string_new(g_malloc0(block_len), block_len);
+        block->dirty = bluesky_string_new(g_malloc0(block_len), block_len);
         break;
     case BLUESKY_BLOCK_REF:
-        bluesky_block_fetch(inode->fs, block, NULL);
-        g_assert(block->type == BLUESKY_BLOCK_CACHED);
-        /* Fall through */
-    case BLUESKY_BLOCK_CACHED:
+        // FIXME: locking on the cloudlog?
+        bluesky_block_fetch(inode, block, NULL);
+        bluesky_string_ref(block->ref->data);
+        block->dirty = bluesky_string_dup(block->ref->data);
+        break;
     case BLUESKY_BLOCK_DIRTY:
-        block->data = bluesky_string_dup(block->data);
+        block->dirty = bluesky_string_dup(block->dirty);
         break;
     }
 
+    if (block->type != BLUESKY_BLOCK_DIRTY)
+        g_atomic_int_add(&inode->fs->cache_dirty, 1);
+
     block->type = BLUESKY_BLOCK_DIRTY;
+    bluesky_cloudlog_unref(block->ref);
+    block->ref = NULL;
 }
 
 /* Set the size of a file.  This will truncate or extend the file as needed.
  * Newly-allocated bytes are zeroed. */
+// FIXME
 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
 {
     g_return_if_fail(size <= BLUESKY_MAX_FILE_SIZE);
@@ -56,7 +62,10 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
     if (size == inode->size)
         return;
 
-    g_print("Truncating file to %"PRIi64" bytes\n", size);
+    if (bluesky_verbose) {
+        g_log("bluesky/file", G_LOG_LEVEL_DEBUG,
+              "Truncating file to %"PRIi64" bytes", size);
+    }
 
     uint64_t blocks = (size + BLUESKY_BLOCK_SIZE - 1) / BLUESKY_BLOCK_SIZE;
 
@@ -72,11 +81,11 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
 
             if (b->type != BLUESKY_BLOCK_ZERO
                     && (b->type == BLUESKY_BLOCK_REF
-                        || b->data->len < BLUESKY_BLOCK_SIZE)) {
+                        || b->dirty->len < BLUESKY_BLOCK_SIZE)) {
                 bluesky_block_touch(inode, inode->blocks->len - 1);
-                gsize old_size = b->data->len;
-                bluesky_string_resize(b->data, BLUESKY_BLOCK_SIZE);
-                memset(&b->data->data[old_size], 0,
+                gsize old_size = b->dirty->len;
+                bluesky_string_resize(b->dirty, BLUESKY_BLOCK_SIZE);
+                memset(&b->dirty->data[old_size], 0,
                        BLUESKY_BLOCK_SIZE - old_size);
             }
         }
@@ -86,8 +95,10 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
         /* Delete blocks from a file.  Must reclaim memory. */
         for (guint i = inode->blocks->len; i < blocks; i++) {
             BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
-            g_free(b->ref);
-            bluesky_string_unref(b->data);
+            if (b->type == BLUESKY_BLOCK_DIRTY)
+                g_atomic_int_add(&inode->fs->cache_dirty, -1);
+            bluesky_string_unref(b->dirty);
+            bluesky_cloudlog_unref(b->ref);
         }
         g_array_set_size(inode->blocks, blocks);
     }
@@ -100,13 +111,13 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
 
         if (b->type != BLUESKY_BLOCK_ZERO) {
             bluesky_block_touch(inode, blocks - 1);
-            gsize old_size = b->data->len;
+            gsize old_size = b->dirty->len;
             gsize new_size = size - (blocks - 1) * BLUESKY_BLOCK_SIZE;
 
-            bluesky_string_resize(b->data, new_size);
+            bluesky_string_resize(b->dirty, new_size);
 
             if (new_size > old_size) {
-                memset(&b->data->data[old_size], 0, new_size - old_size);
+                memset(&b->dirty->data[old_size], 0, new_size - old_size);
             }
         }
     }
@@ -118,8 +129,6 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
                         const char *data, gint len)
 {
-    g_print("Write %d bytes at offset %"PRIi64"\n", len, offset);
-
     g_return_if_fail(inode->type == BLUESKY_REGULAR);
     g_return_if_fail(offset < inode->size);
     g_return_if_fail(len <= inode->size - offset);
@@ -127,6 +136,8 @@ void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
     if (len == 0)
         return;
 
+    // TODO: Optimization: If we are entirely overwriting a block we don't need
+    // to fetch it frm storage first.
     while (len > 0) {
         uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
         gint block_offset = offset % BLUESKY_BLOCK_SIZE;
@@ -135,7 +146,7 @@ void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
         bluesky_block_touch(inode, block_num);
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
                                          block_num);
-        memcpy(&b->data->data[block_offset], data, bytes);
+        memcpy(&b->dirty->data[block_offset], data, bytes);
 
         offset += bytes;
         data += bytes;
@@ -148,8 +159,6 @@ void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
                        char *buf, gint len)
 {
-    g_print("Read %d bytes at offset %"PRIi64"\n", len, offset);
-
     if (len == 0 && offset <= inode->size)
         return;
 
@@ -157,24 +166,31 @@ void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
     g_return_if_fail(offset < inode->size);
     g_return_if_fail(len <= inode->size - offset);
 
+#if 0
     /* Start fetches on any data blocks that we will need for this read. */
     BlueSkyStoreAsync *barrier = bluesky_store_async_new(inode->fs->store);
     barrier->op = STORE_OP_BARRIER;
     uint64_t start_block, end_block;
     start_block = offset / BLUESKY_BLOCK_SIZE;
     end_block = (offset + len - 1) / BLUESKY_BLOCK_SIZE;
-    g_print("Start prefetch on blocks %"PRIi64" .. %"PRIi64"\n",
-            start_block, end_block);
+    if (bluesky_verbose) {
+        g_log("bluesky/file", G_LOG_LEVEL_DEBUG,
+              "Start prefetch on blocks %"PRIi64" .. %"PRIi64,
+              start_block, end_block);
+    }
     for (uint64_t i = start_block; i <= end_block; i++) {
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
                                          i);
         if (b->type == BLUESKY_BLOCK_REF)
-            bluesky_block_fetch(inode->fs, b, barrier);
+            bluesky_block_fetch(inode, b, barrier);
     }
     bluesky_store_async_submit(barrier);
     bluesky_store_async_wait(barrier);
     bluesky_store_async_unref(barrier);
-    g_print("Prefetch complete.\n");
+    if (bluesky_verbose) {
+        g_log("bluesky/file", G_LOG_LEVEL_DEBUG, "Prefetch complete.");
+    }
+#endif
 
     while (len > 0) {
         uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
@@ -183,17 +199,17 @@ void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
 
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
                                          block_num);
-        switch (b->type) {
-        case BLUESKY_BLOCK_ZERO:
+        if (b->type == BLUESKY_BLOCK_ZERO) {
             memset(buf, 0, bytes);
-            break;
-        case BLUESKY_BLOCK_REF:
-            bluesky_block_fetch(inode->fs, b, NULL);
-            /* Fall through */
-        case BLUESKY_BLOCK_CACHED:
-        case BLUESKY_BLOCK_DIRTY:
-            memcpy(buf, &b->data->data[block_offset], bytes);
-            break;
+        } else {
+            BlueSkyRCStr *data = NULL;
+            if (b->type == BLUESKY_BLOCK_REF) {
+                bluesky_block_fetch(inode, b, NULL);
+                data = b->ref->data;
+            } else if (b->type == BLUESKY_BLOCK_DIRTY) {
+                data = b->dirty;
+            }
+            memcpy(buf, &data->data[block_offset], bytes);
         }
 
         offset += bytes;
@@ -202,77 +218,54 @@ void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
     }
 }
 
-/* Read the given block from cloud-backed storage if the data is not already
- * cached. */
-static void block_fetch_completion(BlueSkyStoreAsync *async, gpointer data)
-{
-    BlueSkyBlock *block = (BlueSkyBlock *)data;
-
-    bluesky_string_unref(block->data);
-    block->data = async->data;
-    bluesky_string_ref(block->data);
-    block->type = BLUESKY_BLOCK_CACHED;
-}
-
-void bluesky_block_fetch(BlueSkyFS *fs, BlueSkyBlock *block,
+void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
                          BlueSkyStoreAsync *barrier)
 {
     if (block->type != BLUESKY_BLOCK_REF)
         return;
 
-    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
-    async->op = STORE_OP_GET;
-    async->key = g_strdup(block->ref);
-    bluesky_store_async_add_notifier(async, (GFunc)block_fetch_completion, block);
-    bluesky_store_async_submit(async);
-
-    if (barrier != NULL)
-        bluesky_store_add_barrier(barrier, async);
-    else
-        bluesky_store_async_wait(async);
+    g_mutex_lock(block->ref->lock);
+    bluesky_cloudlog_fetch(block->ref);
+    g_mutex_unlock(block->ref->lock);
+    block->type = BLUESKY_BLOCK_REF;
 }
 
 /* Write the given block to cloud-backed storage and mark it clean. */
-void bluesky_block_flush(BlueSkyFS *fs, BlueSkyBlock *block,
-                         BlueSkyStoreAsync *barrier)
+void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
+                         GList **log_items)
 {
+    BlueSkyFS *fs = inode->fs;
+
     if (block->type != BLUESKY_BLOCK_DIRTY)
         return;
 
-    BlueSkyRCStr *data = block->data;
-
-    GChecksum *csum = g_checksum_new(G_CHECKSUM_SHA256);
-    g_checksum_update(csum, (const guchar *)data->data, data->len);
-    gchar *name = g_strdup(g_checksum_get_string(csum));
+    g_assert(block->ref == NULL);
 
-    /* Store the file data asynchronously, and don't bother waiting for a
-     * response. */
-    BlueSkyStoreAsync *async = bluesky_store_async_new(fs->store);
-    async->op = STORE_OP_PUT;
-    async->key = g_strdup(name);
-    bluesky_string_ref(data);
-    async->data = data;
-    bluesky_store_async_submit(async);
-    if (barrier != NULL)
-        bluesky_store_add_barrier(barrier, async);
-    bluesky_store_async_unref(async);
+    BlueSkyCloudLog *cloudlog = bluesky_cloudlog_new(fs);
+    cloudlog->type = LOGTYPE_DATA;
+    cloudlog->inum = inode->inum;
+    cloudlog->data = block->dirty;      // String ownership is transferred
+    bluesky_cloudlog_stats_update(cloudlog, 1);
+    bluesky_cloudlog_sync(cloudlog);
+    bluesky_cloudlog_ref(cloudlog);     // Reference for log_items list
+    *log_items = g_list_prepend(*log_items, cloudlog);
+    bluesky_cloudlog_insert(cloudlog);
 
-    g_free(block->ref);
-    block->ref = name;
+    block->ref = cloudlog;              // Uses initial reference from _new()
 
-    block->type = BLUESKY_BLOCK_CACHED;
-
-    g_checksum_free(csum);
+    block->type = BLUESKY_BLOCK_REF;
+    block->dirty = NULL;
+    g_atomic_int_add(&fs->cache_dirty, -1);
 }
 
 /* Flush all blocks in a file to stable storage. */
-void bluesky_file_flush(BlueSkyInode *inode, BlueSkyStoreAsync *barrier)
+void bluesky_file_flush(BlueSkyInode *inode, GList **log_items)
 {
     g_return_if_fail(inode->type == BLUESKY_REGULAR);
 
     for (int i = 0; i < inode->blocks->len; i++) {
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
-        bluesky_block_flush(inode->fs, b, barrier);
+        bluesky_block_flush(inode, b, log_items);
     }
 }
 
@@ -283,14 +276,18 @@ void bluesky_file_drop_cached(BlueSkyInode *inode)
 
     for (int i = 0; i < inode->blocks->len; i++) {
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
-        if (b->type == BLUESKY_BLOCK_CACHED) {
-            g_log("bluesky/cache", G_LOG_LEVEL_DEBUG,
-                  "Dropping block %d of inode %"PRIu64" from cache",
-                  i, inode->inum);
-
-            bluesky_string_unref(b->data);
-            b->data = NULL;
-            b->type = BLUESKY_BLOCK_REF;
+        if (b->type == BLUESKY_BLOCK_REF) {
+            g_mutex_lock(b->ref->lock);
+            if (b->ref->data != NULL
+                && g_atomic_int_get(&b->ref->data_lock_count) == 0
+                && (b->ref->location_flags != 0))
+            {
+                bluesky_cloudlog_stats_update(b->ref, -1);
+                bluesky_string_unref(b->ref->data);
+                b->ref->data = NULL;
+                bluesky_cloudlog_stats_update(b->ref, 1);
+            }
+            g_mutex_unlock(b->ref->lock);
         }
     }
 }