X-Git-Url: http://git.vrable.net/?a=blobdiff_plain;f=bluesky%2Ffile.c;h=19363756b17e3fc32ede50a6c4b2f4faa539909a;hb=388030970805a70cb4fad34ade5e3de7a3607a57;hp=57193dd9ee69fefdc3595a6a1df1c27c2990aabe;hpb=ed6f1f04bd4d7a73ed26b3bb0d5e0df1cc0acf42;p=bluesky.git

diff --git a/bluesky/file.c b/bluesky/file.c
index 57193dd..1936375 100644
--- a/bluesky/file.c
+++ b/bluesky/file.c
@@ -16,8 +16,12 @@
 /* Core filesystem: handling of regular files and caching of file data. */
 
 /* Mark a given block dirty and make sure that data is faulted in so that it
- * can be written to. */
-void bluesky_block_touch(BlueSkyInode *inode, uint64_t i)
+ * can be written to.
+ *
+ * If preserve is set to false, this is a hint that the block is about to be
+ * entirely overwritten.  In this case, a dirty block is made available but any
+ * prior contents might be lost.  A value of preserve = TRUE is always safe. */
+void bluesky_block_touch(BlueSkyInode *inode, uint64_t i, gboolean preserve)
 {
     g_return_if_fail(i < inode->blocks->len);
     BlueSkyBlock *block = &g_array_index(inode->blocks, BlueSkyBlock, i);
@@ -31,32 +35,34 @@ void bluesky_block_touch(BlueSkyInode *inode, uint64_t i)
 
     switch (block->type) {
     case BLUESKY_BLOCK_ZERO:
-        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, block, NULL);
-        g_assert(block->type == BLUESKY_BLOCK_CACHED);
-        /* Fall through */
-    case BLUESKY_BLOCK_CACHED:
+        if (preserve) {
+            // 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);
+        } else {
+            block->dirty = bluesky_string_new(g_malloc0(block_len), block_len);
+        }
+        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_CACHED
-            && block->type != BLUESKY_BLOCK_DIRTY)
-        g_atomic_int_add(&inode->fs->cache_total, 1);
     if (block->type != BLUESKY_BLOCK_DIRTY)
         g_atomic_int_add(&inode->fs->cache_dirty, 1);
 
     block->type = BLUESKY_BLOCK_DIRTY;
-    if (block->cloudref != NULL)
-        bluesky_cloudlog_unref(block->cloudref);
-    block->cloudref = NULL;
+    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);
@@ -71,6 +77,15 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
 
     uint64_t blocks = (size + BLUESKY_BLOCK_SIZE - 1) / BLUESKY_BLOCK_SIZE;
 
+    /* Calculate number of bytes in the last block of the file */
+    int lastblock_old, lastblock_new;
+    lastblock_old = inode->size % BLUESKY_BLOCK_SIZE;
+    if (lastblock_old == 0 && inode->size > 0)
+        lastblock_old = BLUESKY_BLOCK_SIZE;
+    lastblock_new = size % BLUESKY_BLOCK_SIZE;
+    if (lastblock_new == 0 && size > 0)
+        lastblock_new = BLUESKY_BLOCK_SIZE;
+
     if (blocks > inode->blocks->len) {
         /* Need to add new blocks to the end of a file.  New block structures
          * are automatically zeroed, which initializes them to be pointers to
@@ -82,12 +97,16 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
                                              inode->blocks->len - 1);
 
             if (b->type != BLUESKY_BLOCK_ZERO
-                    && (b->type == BLUESKY_BLOCK_REF
-                        || b->data->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,
+                    && lastblock_old < BLUESKY_BLOCK_SIZE) {
+                bluesky_block_touch(inode, inode->blocks->len - 1, TRUE);
+                gsize old_size = b->dirty->len;
+                if (lastblock_old != old_size) {
+                    fprintf(stderr,
+                            "Warning: last block size = %zd, expected %d\n",
+                            old_size, lastblock_old);
+                }
+                bluesky_string_resize(b->dirty, BLUESKY_BLOCK_SIZE);
+                memset(&b->dirty->data[old_size], 0,
                        BLUESKY_BLOCK_SIZE - old_size);
             }
         }
@@ -95,15 +114,12 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
         g_array_set_size(inode->blocks, blocks);
     } else if (blocks < inode->blocks->len) {
         /* Delete blocks from a file.  Must reclaim memory. */
-        for (guint i = inode->blocks->len; i < blocks; i++) {
+        for (guint i = blocks; i < inode->blocks->len; i++) {
             BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
-            if (b->type == BLUESKY_BLOCK_CACHED
-                    || b->type == BLUESKY_BLOCK_DIRTY)
-                g_atomic_int_add(&inode->fs->cache_total, -1);
             if (b->type == BLUESKY_BLOCK_DIRTY)
                 g_atomic_int_add(&inode->fs->cache_dirty, -1);
-            bluesky_string_unref(b->data);
-            bluesky_cloudlog_unref(b->cloudref);
+            bluesky_string_unref(b->dirty);
+            bluesky_cloudlog_unref(b->ref);
         }
         g_array_set_size(inode->blocks, blocks);
     }
@@ -114,15 +130,21 @@ void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
         BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
                                          blocks - 1);
 
-        if (b->type != BLUESKY_BLOCK_ZERO) {
-            bluesky_block_touch(inode, blocks - 1);
-            gsize old_size = b->data->len;
+        gboolean need_resize = TRUE;
+        if (b->type == BLUESKY_BLOCK_ZERO)
+            need_resize = FALSE;
+        else if (size < inode->size && lastblock_new == BLUESKY_BLOCK_SIZE)
+            need_resize = FALSE;
+
+        if (need_resize) {
+            bluesky_block_touch(inode, blocks - 1, TRUE);
+            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);
             }
         }
     }
@@ -146,10 +168,18 @@ void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
         gint block_offset = offset % BLUESKY_BLOCK_SIZE;
         gint bytes = MIN(BLUESKY_BLOCK_SIZE - block_offset, len);
 
-        bluesky_block_touch(inode, block_num);
+        gboolean preserve = TRUE;
+        gsize block_size = BLUESKY_BLOCK_SIZE;
+        if (block_num == inode->blocks->len - 1) {
+            block_size = inode->size - block_num * BLUESKY_BLOCK_SIZE;
+        }
+        if (block_offset == 0 && bytes == block_size) {
+            preserve = FALSE;
+        }
+        bluesky_block_touch(inode, block_num, preserve);
         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;
@@ -169,30 +199,23 @@ void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
     g_return_if_fail(offset < inode->size);
     g_return_if_fail(len <= inode->size - offset);
 
-    /* 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;
+    BlueSkyProfile *profile = bluesky_profile_get();
+
+    bluesky_profile_add_event(profile,
+                              g_strdup_printf("Start file read prefetch"));
     uint64_t start_block, end_block;
     start_block = offset / BLUESKY_BLOCK_SIZE;
     end_block = (offset + len - 1) / BLUESKY_BLOCK_SIZE;
-    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, b, barrier);
-    }
-    bluesky_store_async_submit(barrier);
-    bluesky_store_async_wait(barrier);
-    bluesky_store_async_unref(barrier);
-    if (bluesky_verbose) {
-        g_log("bluesky/file", G_LOG_LEVEL_DEBUG, "Prefetch complete.");
+            bluesky_cloudlog_prefetch(b->ref);
     }
 
+    bluesky_profile_add_event(profile,
+                              g_strdup_printf("End file read prefetch"));
+
     while (len > 0) {
         uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
         gint block_offset = offset % BLUESKY_BLOCK_SIZE;
@@ -200,23 +223,26 @@ 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, 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;
         buf += bytes;
         len -= bytes;
     }
+
+    bluesky_profile_add_event(profile,
+                              g_strdup_printf("BlueSky read complete"));
 }
 
 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
@@ -225,13 +251,10 @@ void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
     if (block->type != BLUESKY_BLOCK_REF)
         return;
 
-    g_mutex_lock(block->cloudref->lock);
-    bluesky_cloudlog_fetch(block->cloudref);
-    block->data = block->cloudref->data;
-    bluesky_string_ref(block->data);
-    g_mutex_unlock(block->cloudref->lock);
-    block->type = BLUESKY_BLOCK_CACHED;
-    g_atomic_int_add(&inode->fs->cache_total, 1);
+    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. */
@@ -243,24 +266,22 @@ void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
     if (block->type != BLUESKY_BLOCK_DIRTY)
         return;
 
-    if (block->cloudref != NULL)
-        bluesky_cloudlog_unref(block->cloudref);
-
-    BlueSkyRCStr *data = block->data;
+    g_assert(block->ref == NULL);
 
-    BlueSkyCloudLog *cloudlog = bluesky_cloudlog_new(fs);
+    BlueSkyCloudLog *cloudlog = bluesky_cloudlog_new(fs, NULL);
     cloudlog->type = LOGTYPE_DATA;
     cloudlog->inum = inode->inum;
-    cloudlog->data = data;
-    bluesky_string_ref(data);
+    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);
 
-    block->cloudref = cloudlog;
-    bluesky_cloudlog_ref(cloudlog);
+    block->ref = cloudlog;              // Uses initial reference from _new()
 
-    block->type = BLUESKY_BLOCK_CACHED;
+    block->type = BLUESKY_BLOCK_REF;
+    block->dirty = NULL;
     g_atomic_int_add(&fs->cache_dirty, -1);
 }
 
@@ -282,25 +303,18 @@ 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) {
-            if (bluesky_verbose) {
-                g_log("bluesky/cache", G_LOG_LEVEL_DEBUG,
-                      "Dropping block %d of inode %"PRIu64" from cache",
-                      i, inode->inum);
-                g_log("bluesky/cache", G_LOG_LEVEL_DEBUG,
-                      "  (reference count was %d)", b->data->refcount);
-            }
-
-            bluesky_string_unref(b->data);
-            b->data = NULL;
-            b->type = BLUESKY_BLOCK_REF;
-            g_atomic_int_add(&inode->fs->cache_total, -1);
-            g_mutex_lock(b->cloudref->lock);
-            if (b->cloudref->location_flags & CLOUDLOG_JOURNAL) {
-                bluesky_string_unref(b->cloudref->data);
-                b->cloudref->data = NULL;
+        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->cloudref->lock);
+            g_mutex_unlock(b->ref->lock);
         }
     }
 }