/* Wraps pages current wbuf in an io and submits to IO thread.
* Called with p locked, locks e.
*/
static void _submit_wbuf(store_engine *e, store_page *p) {
_store_wbuf *w;
pthread_mutex_lock(&e->mutex);
obj_io *io = e->io_stack;
e->io_stack = io->next;
pthread_mutex_unlock(&e->mutex);
w = p->wbuf;
// zero out the end of the wbuf to allow blind readback of data.
memset(w->buf + (w->size - w->free), 0, w->free);
io->next = NULL;
io->mode = OBJ_IO_WRITE;
io->page_id = p->id;
io->data = w;
io->offset = w->offset;
io->len = w->size;
io->buf = w->buf;
io->cb = _wbuf_cb;
extstore_submit(e, io);
}
// TODO: hoist the storage bits from lru_maintainer_thread in here.
// would be nice if they could avoid hammering the same locks though?
// I guess it's only COLD. that's probably fine.
static void *storage_compact_thread(void *arg) {
void *storage = arg;
useconds_t to_sleep = MAX_STORAGE_COMPACT_SLEEP;
bool compacting = false;
uint64_t page_version = 0;
uint64_t page_size = 0;
uint64_t page_offset = 0;
uint32_t page_id = 0;
bool drop_unread = false;
char *readback_buf = NULL;
struct storage_compact_wrap wrap;
logger *l = logger_create();
if (l == NULL) {
fprintf(stderr, "Failed to allocate logger for storage compaction thread\n");
abort();
}
readback_buf = malloc(settings.ext_wbuf_size);
if (readback_buf == NULL) {
fprintf(stderr, "Failed to allocate readback buffer for storage compaction thread\n");
abort();
}
pthread_mutex_init(&wrap.lock, NULL);
wrap.done = false;
wrap.submitted = false;
wrap.io.data = &wrap;
wrap.io.buf = (void *)readback_buf;
wrap.io.len = settings.ext_wbuf_size;
wrap.io.mode = OBJ_IO_READ;
wrap.io.cb = _storage_compact_cb;
pthread_mutex_lock(&storage_compact_plock);
while (1) {
pthread_mutex_unlock(&storage_compact_plock);
if (to_sleep) {
extstore_run_maint(storage);
usleep(to_sleep);
}
pthread_mutex_lock(&storage_compact_plock);
if (!compacting && storage_compact_check(storage, l,
&page_id, &page_version, &page_size, &drop_unread)) {
page_offset = 0;
compacting = true;
LOGGER_LOG(l, LOG_SYSEVENTS, LOGGER_COMPACT_START,
NULL, page_id, page_version);
}
if (compacting) {
pthread_mutex_lock(&wrap.lock);
if (page_offset < page_size && !wrap.done && !wrap.submitted) {
wrap.io.page_version = page_version;
wrap.io.page_id = page_id;
wrap.io.offset = page_offset;
// FIXME: should be smarter about io->next (unlink at use?)
wrap.io.next = NULL;
wrap.submitted = true;
wrap.miss = false;
extstore_submit(storage, &wrap.io);
} else if (wrap.miss) {
LOGGER_LOG(l, LOG_SYSEVENTS, LOGGER_COMPACT_ABORT,
NULL, page_id);
wrap.done = false;
wrap.submitted = false;
compacting = false;
} else if (wrap.submitted && wrap.done) {
LOGGER_LOG(l, LOG_SYSEVENTS, LOGGER_COMPACT_READ_START,
NULL, page_id, page_offset);
storage_compact_readback(storage, l, drop_unread,
readback_buf, page_id, page_version, settings.ext_wbuf_size);
page_offset += settings.ext_wbuf_size;
wrap.done = false;
wrap.submitted = false;
} else if (page_offset >= page_size) {
compacting = false;
wrap.done = false;
wrap.submitted = false;
extstore_close_page(storage, page_id, page_version);
LOGGER_LOG(l, LOG_SYSEVENTS, LOGGER_COMPACT_END,
NULL, page_id);
}
pthread_mutex_unlock(&wrap.lock);
if (to_sleep > MIN_STORAGE_COMPACT_SLEEP)
to_sleep /= 2;
} else {
if (to_sleep < MAX_STORAGE_COMPACT_SLEEP)
to_sleep += MIN_STORAGE_COMPACT_SLEEP;
}
}
free(readback_buf);
return NULL;
}
