1 /* Blue Sky: File Systems in the Cloud
3 * Copyright (C) 2009 The Regents of the University of California
4 * Written by Michael Vrable <mvrable@cs.ucsd.edu>
14 #include "bluesky-private.h"
16 /* Core filesystem: handling of regular files and caching of file data. */
18 /* Mark a given block dirty and make sure that data is faulted in so that it
21 * If preserve is set to false, this is a hint that the block is about to be
22 * entirely overwritten. In this case, a dirty block is made available but any
23 * prior contents might be lost. A value of preserve = TRUE is always safe. */
24 void bluesky_block_touch(BlueSkyInode *inode, uint64_t i, gboolean preserve)
26 g_return_if_fail(i < inode->blocks->len);
27 BlueSkyBlock *block = &g_array_index(inode->blocks, BlueSkyBlock, i);
30 if (i < inode->blocks->len - 1) {
31 block_len = BLUESKY_BLOCK_SIZE;
33 block_len = inode->size - i * BLUESKY_BLOCK_SIZE;
36 switch (block->type) {
37 case BLUESKY_BLOCK_ZERO:
38 block->dirty = bluesky_string_new(g_malloc0(block_len), block_len);
40 case BLUESKY_BLOCK_REF:
42 // FIXME: locking on the cloudlog?
43 bluesky_block_fetch(inode, block, NULL);
44 bluesky_string_ref(block->ref->data);
45 block->dirty = bluesky_string_dup(block->ref->data);
47 block->dirty = bluesky_string_new(g_malloc0(block_len), block_len);
50 case BLUESKY_BLOCK_DIRTY:
51 block->dirty = bluesky_string_dup(block->dirty);
55 if (block->type != BLUESKY_BLOCK_DIRTY)
56 g_atomic_int_add(&inode->fs->cache_dirty, 1);
58 block->type = BLUESKY_BLOCK_DIRTY;
59 bluesky_cloudlog_unref(block->ref);
63 /* Set the size of a file. This will truncate or extend the file as needed.
64 * Newly-allocated bytes are zeroed. */
66 void bluesky_file_truncate(BlueSkyInode *inode, uint64_t size)
68 g_return_if_fail(size <= BLUESKY_MAX_FILE_SIZE);
70 if (size == inode->size)
73 if (bluesky_verbose) {
74 g_log("bluesky/file", G_LOG_LEVEL_DEBUG,
75 "Truncating file to %"PRIi64" bytes", size);
78 uint64_t blocks = (size + BLUESKY_BLOCK_SIZE - 1) / BLUESKY_BLOCK_SIZE;
80 /* Calculate number of bytes in the last block of the file */
81 int lastblock_old, lastblock_new;
82 lastblock_old = inode->size % BLUESKY_BLOCK_SIZE;
83 if (lastblock_old == 0 && inode->size > 0)
84 lastblock_old = BLUESKY_BLOCK_SIZE;
85 lastblock_new = size % BLUESKY_BLOCK_SIZE;
86 if (lastblock_new == 0 && size > 0)
87 lastblock_new = BLUESKY_BLOCK_SIZE;
89 if (blocks > inode->blocks->len) {
90 /* Need to add new blocks to the end of a file. New block structures
91 * are automatically zeroed, which initializes them to be pointers to
92 * zero blocks so we don't need to do any more work. If the
93 * previously-last block in the file is smaller than
94 * BLUESKY_BLOCK_SIZE, extend it to full size. */
95 if (inode->blocks->len > 0) {
96 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
97 inode->blocks->len - 1);
99 if (b->type != BLUESKY_BLOCK_ZERO
100 && lastblock_old < BLUESKY_BLOCK_SIZE) {
101 bluesky_block_touch(inode, inode->blocks->len - 1, TRUE);
102 gsize old_size = b->dirty->len;
103 g_warn_if_fail(lastblock_old != old_size);
104 bluesky_string_resize(b->dirty, BLUESKY_BLOCK_SIZE);
105 memset(&b->dirty->data[old_size], 0,
106 BLUESKY_BLOCK_SIZE - old_size);
110 g_array_set_size(inode->blocks, blocks);
111 } else if (blocks < inode->blocks->len) {
112 /* Delete blocks from a file. Must reclaim memory. */
113 for (guint i = blocks; i < inode->blocks->len; i++) {
114 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
115 if (b->type == BLUESKY_BLOCK_DIRTY)
116 g_atomic_int_add(&inode->fs->cache_dirty, -1);
117 bluesky_string_unref(b->dirty);
118 bluesky_cloudlog_unref(b->ref);
120 g_array_set_size(inode->blocks, blocks);
123 /* Ensure the new last block of the file is properly sized. If the block
124 * is extended, newly-added bytes must be zeroed. */
126 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
129 gboolean need_resize = TRUE;
130 if (b->type == BLUESKY_BLOCK_ZERO)
132 else if (size < inode->size && lastblock_new == BLUESKY_BLOCK_SIZE)
136 bluesky_block_touch(inode, blocks - 1, TRUE);
137 gsize old_size = b->dirty->len;
138 gsize new_size = size - (blocks - 1) * BLUESKY_BLOCK_SIZE;
140 bluesky_string_resize(b->dirty, new_size);
142 if (new_size > old_size) {
143 memset(&b->dirty->data[old_size], 0, new_size - old_size);
149 bluesky_inode_update_ctime(inode, 1);
152 void bluesky_file_write(BlueSkyInode *inode, uint64_t offset,
153 const char *data, gint len)
155 g_return_if_fail(inode->type == BLUESKY_REGULAR);
156 g_return_if_fail(offset < inode->size);
157 g_return_if_fail(len <= inode->size - offset);
162 // TODO: Optimization: If we are entirely overwriting a block we don't need
163 // to fetch it frm storage first. We don't yet handle the case where the
164 // partial last block of a file is entirely overwritten.
166 uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
167 gint block_offset = offset % BLUESKY_BLOCK_SIZE;
168 gint bytes = MIN(BLUESKY_BLOCK_SIZE - block_offset, len);
170 gboolean preserve = TRUE;
171 if (block_offset == 0 && bytes == BLUESKY_BLOCK_SIZE) {
174 bluesky_block_touch(inode, block_num, preserve);
175 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
177 memcpy(&b->dirty->data[block_offset], data, bytes);
184 bluesky_inode_update_ctime(inode, 1);
187 void bluesky_file_read(BlueSkyInode *inode, uint64_t offset,
190 if (len == 0 && offset <= inode->size)
193 g_return_if_fail(inode->type == BLUESKY_REGULAR);
194 g_return_if_fail(offset < inode->size);
195 g_return_if_fail(len <= inode->size - offset);
197 BlueSkyProfile *profile = bluesky_profile_get();
199 bluesky_profile_add_event(profile,
200 g_strdup_printf("Start file read prefetch"));
201 uint64_t start_block, end_block;
202 start_block = offset / BLUESKY_BLOCK_SIZE;
203 end_block = (offset + len - 1) / BLUESKY_BLOCK_SIZE;
204 for (uint64_t i = start_block; i <= end_block; i++) {
205 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
207 if (b->type == BLUESKY_BLOCK_REF)
208 bluesky_cloudlog_prefetch(b->ref);
211 bluesky_profile_add_event(profile,
212 g_strdup_printf("End file read prefetch"));
215 uint64_t block_num = offset / BLUESKY_BLOCK_SIZE;
216 gint block_offset = offset % BLUESKY_BLOCK_SIZE;
217 gint bytes = MIN(BLUESKY_BLOCK_SIZE - block_offset, len);
219 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock,
221 if (b->type == BLUESKY_BLOCK_ZERO) {
222 memset(buf, 0, bytes);
224 BlueSkyRCStr *data = NULL;
225 if (b->type == BLUESKY_BLOCK_REF) {
226 bluesky_block_fetch(inode, b, NULL);
228 } else if (b->type == BLUESKY_BLOCK_DIRTY) {
231 memcpy(buf, &data->data[block_offset], bytes);
239 bluesky_profile_add_event(profile,
240 g_strdup_printf("BlueSky read complete"));
243 void bluesky_block_fetch(BlueSkyInode *inode, BlueSkyBlock *block,
244 BlueSkyStoreAsync *barrier)
246 if (block->type != BLUESKY_BLOCK_REF)
249 g_mutex_lock(block->ref->lock);
250 bluesky_cloudlog_fetch(block->ref);
251 g_mutex_unlock(block->ref->lock);
252 block->type = BLUESKY_BLOCK_REF;
255 /* Write the given block to cloud-backed storage and mark it clean. */
256 void bluesky_block_flush(BlueSkyInode *inode, BlueSkyBlock *block,
259 BlueSkyFS *fs = inode->fs;
261 if (block->type != BLUESKY_BLOCK_DIRTY)
264 g_assert(block->ref == NULL);
266 BlueSkyCloudLog *cloudlog = bluesky_cloudlog_new(fs, NULL);
267 cloudlog->type = LOGTYPE_DATA;
268 cloudlog->inum = inode->inum;
269 cloudlog->data = block->dirty; // String ownership is transferred
270 bluesky_cloudlog_stats_update(cloudlog, 1);
271 bluesky_cloudlog_sync(cloudlog);
272 bluesky_cloudlog_ref(cloudlog); // Reference for log_items list
273 *log_items = g_list_prepend(*log_items, cloudlog);
274 bluesky_cloudlog_insert(cloudlog);
276 block->ref = cloudlog; // Uses initial reference from _new()
278 block->type = BLUESKY_BLOCK_REF;
280 g_atomic_int_add(&fs->cache_dirty, -1);
283 /* Flush all blocks in a file to stable storage. */
284 void bluesky_file_flush(BlueSkyInode *inode, GList **log_items)
286 g_return_if_fail(inode->type == BLUESKY_REGULAR);
288 for (int i = 0; i < inode->blocks->len; i++) {
289 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
290 bluesky_block_flush(inode, b, log_items);
294 /* Drop clean data blocks for a file from cache. */
295 void bluesky_file_drop_cached(BlueSkyInode *inode)
297 g_return_if_fail(inode->type == BLUESKY_REGULAR);
299 for (int i = 0; i < inode->blocks->len; i++) {
300 BlueSkyBlock *b = &g_array_index(inode->blocks, BlueSkyBlock, i);
301 if (b->type == BLUESKY_BLOCK_REF) {
302 g_mutex_lock(b->ref->lock);
303 if (b->ref->data != NULL
304 && g_atomic_int_get(&b->ref->data_lock_count) == 0
305 && (b->ref->location_flags != 0))
307 bluesky_cloudlog_stats_update(b->ref, -1);
308 bluesky_string_unref(b->ref->data);
310 bluesky_cloudlog_stats_update(b->ref, 1);
312 g_mutex_unlock(b->ref->lock);