/*
* __truncate_table --
* WT_SESSION::truncate for a table.
*/
static int
__truncate_table(WT_SESSION_IMPL *session, const char *name)
{
WT_BTREE *btree;
WT_DECL_ITEM(namebuf);
WT_DECL_RET;
WT_TABLE *table;
int i;
WT_RET(__wt_schema_get_table(session, name, strlen(name), 0, &table));
WT_RET(__wt_scr_alloc(session, 0, &namebuf));
/* Truncate the column groups. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->cgroups[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
/* Truncate the indices. */
WT_ERR(__wt_schema_open_indices(session, table));
for (i = 0; i < table->nindices; i++) {
/*
* Get an exclusive lock on the handle: it will be released by
* __wt_conn_btree_close_all.
*/
WT_ERR(__wt_session_get_btree(session,
table->indices[i]->source, NULL, NULL, WT_BTREE_EXCLUSIVE));
btree = session->btree;
WT_ERR(__wt_buf_set(
session, namebuf, btree->name, strlen(btree->name) + 1));
WT_ERR(__truncate_file(session, namebuf->data));
}
table->idx_complete = 0;
/* Reopen the column groups. */
ret = __wt_schema_open_colgroups(session, table);
err: __wt_scr_free(&namebuf);
return (ret);
}
/*
* __wt_txn_truncate_log --
* Begin truncating a range of a file.
*/
int
__wt_txn_truncate_log(
WT_SESSION_IMPL *session, WT_CURSOR_BTREE *start, WT_CURSOR_BTREE *stop)
{
WT_BTREE *btree;
WT_ITEM *item;
WT_TXN_OP *op;
btree = S2BT(session);
WT_RET(__txn_next_op(session, &op));
if (btree->type == BTREE_ROW) {
op->type = WT_TXN_OP_TRUNCATE_ROW;
op->u.truncate_row.mode = WT_TXN_TRUNC_ALL;
WT_CLEAR(op->u.truncate_row.start);
WT_CLEAR(op->u.truncate_row.stop);
if (start != NULL) {
op->u.truncate_row.mode = WT_TXN_TRUNC_START;
item = &op->u.truncate_row.start;
WT_RET(__wt_cursor_get_raw_key(&start->iface, item));
WT_RET(__wt_buf_set(
session, item, item->data, item->size));
}
if (stop != NULL) {
op->u.truncate_row.mode =
(op->u.truncate_row.mode == WT_TXN_TRUNC_ALL) ?
WT_TXN_TRUNC_STOP : WT_TXN_TRUNC_BOTH;
item = &op->u.truncate_row.stop;
WT_RET(__wt_cursor_get_raw_key(&stop->iface, item));
WT_RET(__wt_buf_set(
session, item, item->data, item->size));
}
} else {
op->type = WT_TXN_OP_TRUNCATE_COL;
op->u.truncate_col.start =
(start == NULL) ? WT_RECNO_OOB : start->recno;
op->u.truncate_col.stop =
(stop == NULL) ? WT_RECNO_OOB : stop->recno;
}
/* Write that operation into the in-memory log. */
WT_RET(__wt_txn_log_op(session, NULL));
WT_ASSERT(session, !F_ISSET(session, WT_SESSION_LOGGING_INMEM));
F_SET(session, WT_SESSION_LOGGING_INMEM);
return (0);
}
/*
* __curlog_op_read --
* Read out any key/value from an individual operation record
* in the log. We're only interested in put and remove operations
* since truncate is not a cursor operation. All successful
* returns from this function will have set up the cursor copy of
* key and value to give the user.
*/
static int
__curlog_op_read(WT_SESSION_IMPL *session,
WT_CURSOR_LOG *cl, uint32_t optype, uint32_t opsize, uint32_t *fileid)
{
WT_ITEM key, value;
uint64_t recno;
const uint8_t *end, *pp;
pp = cl->stepp;
end = pp + opsize;
switch (optype) {
case WT_LOGOP_COL_PUT:
WT_RET(__wt_logop_col_put_unpack(session, &pp, end,
fileid, &recno, &value));
WT_RET(__wt_buf_set(session, cl->opkey, &recno, sizeof(recno)));
WT_RET(__wt_buf_set(session,
cl->opvalue, value.data, value.size));
break;
case WT_LOGOP_COL_REMOVE:
WT_RET(__wt_logop_col_remove_unpack(session, &pp, end,
fileid, &recno));
WT_RET(__wt_buf_set(session, cl->opkey, &recno, sizeof(recno)));
WT_RET(__wt_buf_set(session, cl->opvalue, NULL, 0));
break;
case WT_LOGOP_ROW_PUT:
WT_RET(__wt_logop_row_put_unpack(session, &pp, end,
fileid, &key, &value));
WT_RET(__wt_buf_set(session, cl->opkey, key.data, key.size));
WT_RET(__wt_buf_set(session,
cl->opvalue, value.data, value.size));
break;
case WT_LOGOP_ROW_REMOVE:
WT_RET(__wt_logop_row_remove_unpack(session, &pp, end,
fileid, &key));
WT_RET(__wt_buf_set(session, cl->opkey, key.data, key.size));
WT_RET(__wt_buf_set(session, cl->opvalue, NULL, 0));
break;
default:
/*
* Any other operations return the record in the value
* and an empty key.
*/
*fileid = 0;
WT_RET(__wt_buf_set(session, cl->opkey, NULL, 0));
WT_RET(__wt_buf_set(session, cl->opvalue, cl->stepp, opsize));
}
return (0);
}
/*
* __curfile_remove --
* WT_CURSOR->remove method for the btree cursor type.
*/
static int
__curfile_remove(WT_CURSOR *cursor)
{
WT_CURSOR_BTREE *cbt;
WT_DECL_RET;
WT_SESSION_IMPL *session;
cbt = (WT_CURSOR_BTREE *)cursor;
CURSOR_UPDATE_API_CALL(cursor, session, remove, cbt->btree);
WT_CURSOR_NEEDKEY(cursor);
WT_CURSOR_NOVALUE(cursor);
WT_BTREE_CURSOR_SAVE_AND_RESTORE(cursor, __wt_btcur_remove(cbt), ret);
/*
* After a successful remove, copy the key: the value is not available.
*/
if (ret == 0) {
if (F_ISSET(cursor, WT_CURSTD_KEY_INT) &&
!WT_DATA_IN_ITEM(&(cursor)->key)) {
WT_ERR(__wt_buf_set(session, &cursor->key,
cursor->key.data, cursor->key.size));
F_CLR(cursor, WT_CURSTD_KEY_INT);
F_SET(cursor, WT_CURSTD_KEY_EXT);
}
F_CLR(cursor, WT_CURSTD_VALUE_SET);
}
err: CURSOR_UPDATE_API_END(session, ret);
return (ret);
}
/*
* __curmetadata_setkv --
* Copy key/value into the public cursor, stripping internal metadata for
* "create-only" cursors.
*/
static int
__curmetadata_setkv(WT_CURSOR_METADATA *mdc, WT_CURSOR *fc)
{
WT_CURSOR *c;
WT_DECL_RET;
WT_SESSION_IMPL *session;
char *value;
c = &mdc->iface;
session = (WT_SESSION_IMPL *)c->session;
c->key.data = fc->key.data;
c->key.size = fc->key.size;
if (F_ISSET(mdc, WT_MDC_CREATEONLY)) {
WT_RET(__schema_create_strip(session, fc->value.data, &value));
ret = __wt_buf_set(
session, &c->value, value, strlen(value) + 1);
__wt_free(session, value);
WT_RET(ret);
} else {
c->value.data = fc->value.data;
c->value.size = fc->value.size;
}
F_SET(c, WT_CURSTD_KEY_EXT | WT_CURSTD_VALUE_EXT);
F_CLR(mdc, WT_MDC_ONMETADATA);
F_SET(mdc, WT_MDC_POSITIONED);
return (0);
}
/*
* __curmetadata_metadata_search --
* Retrieve the metadata for the metadata table
*/
static int
__curmetadata_metadata_search(WT_SESSION_IMPL *session, WT_CURSOR *cursor)
{
WT_CURSOR_METADATA *mdc;
WT_DECL_RET;
const char *value;
mdc = (WT_CURSOR_METADATA *)cursor;
/* The metadata search interface allocates a new string in value. */
WT_RET(__wt_metadata_search(session, WT_METADATA_URI, &value));
/*
* Copy the value to the underlying btree cursor's tmp item which will
* be freed when the cursor is closed.
*/
if (F_ISSET(mdc, WT_MDC_TMP_USED))
__wt_buf_free(session, &mdc->tmp_val);
ret = __wt_buf_set(session, &mdc->tmp_val, value, strlen(value));
__wt_free(session, value);
WT_RET(ret);
cursor->key.data = WT_METADATA_URI;
cursor->key.size = strlen(WT_METADATA_URI);
cursor->value.data = mdc->tmp_val.data;
cursor->value.size = mdc->tmp_val.size;
F_SET(mdc, WT_MDC_ONMETADATA | WT_MDC_POSITIONED | WT_MDC_TMP_USED);
F_SET(cursor, WT_CURSTD_KEY_EXT | WT_CURSTD_VALUE_EXT);
return (0);
}
/*
* __ckpt_load --
* Load a single checkpoint's information into a WT_CKPT structure.
*/
static int
__ckpt_load(WT_SESSION_IMPL *session,
WT_CONFIG_ITEM *k, WT_CONFIG_ITEM *v, WT_CKPT *ckpt)
{
WT_CONFIG_ITEM a;
char timebuf[64];
/*
* Copy the name, address (raw and hex), order and time into the slot.
* If there's no address, it's a fake.
*/
WT_RET(__wt_strndup(session, k->str, k->len, &ckpt->name));
WT_RET(__wt_config_subgets(session, v, "addr", &a));
WT_RET(__wt_buf_set(session, &ckpt->addr, a.str, a.len));
if (a.len == 0)
F_SET(ckpt, WT_CKPT_FAKE);
else
WT_RET(__wt_nhex_to_raw(session, a.str, a.len, &ckpt->raw));
WT_RET(__wt_config_subgets(session, v, "order", &a));
if (a.len == 0)
goto format;
ckpt->order = a.val;
WT_RET(__wt_config_subgets(session, v, "time", &a));
if (a.len == 0 || a.len > sizeof(timebuf) - 1)
goto format;
memcpy(timebuf, a.str, a.len);
timebuf[a.len] = '\0';
if (sscanf(timebuf, "%" SCNuMAX, &ckpt->sec) != 1)
goto format;
WT_RET(__wt_config_subgets(session, v, "size", &a));
ckpt->ckpt_size = (uint64_t)a.val;
WT_RET(__wt_config_subgets(session, v, "write_gen", &a));
if (a.len == 0)
goto format;
/*
* The largest value a WT_CONFIG_ITEM can handle is signed: this value
* appears on disk and I don't want to sign it there, so I'm casting it
* here instead.
*/
ckpt->write_gen = (uint64_t)a.val;
return (0);
format:
WT_RET_MSG(session, WT_ERROR, "corrupted checkpoint list");
}
/*
* __async_set_key --
* WT_ASYNC_OP->set_key implementation for op handles.
*/
static void
__async_set_key(WT_ASYNC_OP *asyncop, ...)
{
WT_CURSOR *c;
va_list ap;
c = &asyncop->c;
va_start(ap, asyncop);
__wt_cursor_set_keyv(c, c->flags, ap);
if (!WT_DATA_IN_ITEM(&c->key) && !WT_CURSOR_RECNO(c))
c->saved_err = __wt_buf_set(
O2S((WT_ASYNC_OP_IMPL *)asyncop),
&c->key, c->key.data, c->key.size);
va_end(ap);
}
/*
* __async_set_value --
* WT_ASYNC_OP->set_value implementation for op handles.
*/
static void
__async_set_value(WT_ASYNC_OP *asyncop, ...)
{
WT_CURSOR *c;
va_list ap;
c = &asyncop->c;
va_start(ap, asyncop);
__wt_cursor_set_valuev(c, ap);
/* Copy the data, if it is pointing at data elsewhere. */
if (!WT_DATA_IN_ITEM(&c->value))
c->saved_err = __wt_buf_set(
O2S((WT_ASYNC_OP_IMPL *)asyncop),
&c->value, c->value.data, c->value.size);
va_end(ap);
}
/*
* __curlog_logrec --
* Callback function from log_scan to get a log record.
*/
static int
__curlog_logrec(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_CURSOR_LOG *cl;
cl = cookie;
WT_UNUSED(firstrecord);
/* Set up the LSNs and take a copy of the log record for the cursor. */
*cl->cur_lsn = *lsnp;
*cl->next_lsn = *next_lsnp;
WT_RET(__wt_buf_set(session, cl->logrec, logrec->data, logrec->size));
/*
* Read the log header. Set up the step pointers to walk the
* operations inside the record. Get the record type.
*/
cl->stepp = WT_LOG_SKIP_HEADER(cl->logrec->data);
cl->stepp_end = (uint8_t *)cl->logrec->data + logrec->size;
WT_RET(__wt_logrec_read(session, &cl->stepp, cl->stepp_end,
&cl->rectype));
/* A step count of 0 means the entire record. */
cl->step_count = 0;
/*
* Unpack the txnid so that we can return each
* individual operation for this txnid.
*/
if (cl->rectype == WT_LOGREC_COMMIT)
WT_RET(__wt_vunpack_uint(&cl->stepp,
WT_PTRDIFF(cl->stepp_end, cl->stepp), &cl->txnid));
else {
/*
* Step over anything else.
* Setting stepp to NULL causes the next()
* method to read a new record on the next call.
*/
cl->stepp = NULL;
cl->txnid = 0;
}
return (0);
}
/*
* __curbulk_insert_row --
* Row-store bulk cursor insert, with key-sort checks.
*/
static int
__curbulk_insert_row(WT_CURSOR *cursor)
{
WT_BTREE *btree;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
WT_SESSION_IMPL *session;
int cmp;
cbulk = (WT_CURSOR_BULK *)cursor;
btree = cbulk->cbt.btree;
/*
* Bulk cursor inserts are updates, but don't need auto-commit
* transactions because they are single-threaded and not visible
* until the bulk cursor is closed.
*/
CURSOR_API_CALL(cursor, session, insert, btree);
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
WT_CURSOR_CHECKKEY(cursor);
WT_CURSOR_CHECKVALUE(cursor);
/*
* If this isn't the first key inserted, compare it against the last key
* to ensure the application doesn't accidentally corrupt the table.
*/
if (!cbulk->first_insert) {
WT_ERR(__wt_compare(session,
btree->collator, &cursor->key, &cbulk->last, &cmp));
if (cmp <= 0)
WT_ERR(__bulk_row_keycmp_err(cbulk));
} else
cbulk->first_insert = false;
/* Save a copy of the key for the next comparison. */
WT_ERR(__wt_buf_set(session,
&cbulk->last, cursor->key.data, cursor->key.size));
ret = __wt_bulk_insert_row(session, cbulk);
err: API_END_RET(session, ret);
}
/*
* __wt_cursor_key_order_init --
* Initialize key ordering checks for cursor movements after a successful
* search.
*/
int
__wt_cursor_key_order_init(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
{
/*
* Cursor searches set the position for cursor movements, set the
* last-key value for diagnostic checking.
*/
switch (cbt->ref->page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
cbt->lastrecno = cbt->recno;
return (0);
case WT_PAGE_ROW_LEAF:
return (__wt_buf_set(session,
cbt->lastkey, cbt->iface.key.data, cbt->iface.key.size));
WT_ILLEGAL_VALUE(session);
}
/* NOTREACHED */
}
/*
* __cursor_key_order_check_row --
* Check key ordering for row-store cursor movements.
*/
static int
__cursor_key_order_check_row(
WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, bool next)
{
WT_BTREE *btree;
WT_ITEM *key;
WT_DECL_RET;
WT_DECL_ITEM(a);
WT_DECL_ITEM(b);
int cmp;
btree = S2BT(session);
key = &cbt->iface.key;
cmp = 0; /* -Werror=maybe-uninitialized */
if (cbt->lastkey->size != 0)
WT_RET(__wt_compare(
session, btree->collator, cbt->lastkey, key, &cmp));
if (cbt->lastkey->size == 0 || (next && cmp < 0) || (!next && cmp > 0))
return (__wt_buf_set(session,
cbt->lastkey, cbt->iface.key.data, cbt->iface.key.size));
WT_ERR(__wt_scr_alloc(session, 512, &a));
WT_ERR(__wt_scr_alloc(session, 512, &b));
WT_PANIC_ERR(session, EINVAL,
"WT_CURSOR.%s out-of-order returns: returned key %s then key %s",
next ? "next" : "prev",
__wt_buf_set_printable(
session, cbt->lastkey->data, cbt->lastkey->size, a),
__wt_buf_set_printable(session, key->data, key->size, b));
err: __wt_scr_free(session, &a);
__wt_scr_free(session, &b);
return (ret);
}
/*
* __wt_bulk_insert --
* Bulk insert, called once per item.
*/
int
__wt_bulk_insert(WT_CURSOR_BULK *cbulk)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_SESSION_IMPL *session;
int cmp;
session = (WT_SESSION_IMPL *)cbulk->cbt.iface.session;
btree = S2BT(session);
cursor = &cbulk->cbt.iface;
switch (btree->type) {
case BTREE_COL_FIX:
WT_RET(__wt_rec_col_fix_bulk_insert(cbulk));
break;
case BTREE_COL_VAR:
/*
* If this isn't the first value inserted, compare it against
* the last value and increment the RLE count.
*
* Instead of a "first time" variable, I'm using the RLE count,
* because it is set to 0 exactly once, the first time through
* the code.
*/
if (cbulk->rle != 0) {
if (cbulk->cmp.size == cursor->value.size &&
memcmp(cbulk->cmp.data,
cursor->value.data, cursor->value.size) == 0) {
++cbulk->rle;
break;
}
WT_RET(__wt_rec_col_var_bulk_insert(cbulk));
}
WT_RET(__wt_buf_set(session,
&cbulk->cmp, cursor->value.data, cursor->value.size));
cbulk->rle = 1;
break;
case BTREE_ROW:
/*
* If this isn't the first value inserted, compare it against
* the last key to ensure the application doesn't accidentally
* corrupt the table.
*
* Instead of a "first time" variable, I'm using the RLE count,
* because it is set to 0 exactly once, the first time through
* the code.
*/
if (cbulk->rle != 0) {
WT_RET(WT_LEX_CMP(session,
btree->collator, &cursor->key, &cbulk->cmp, cmp));
if (cmp <= 0)
return (__bulk_row_keycmp_err(cbulk));
}
WT_RET(__wt_buf_set(session,
&cbulk->cmp, cursor->key.data, cursor->key.size));
cbulk->rle = 1;
WT_RET(__wt_rec_row_bulk_insert(cbulk));
break;
WT_ILLEGAL_VALUE(session);
}
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
return (0);
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_PAGE *page;
WT_ROW *rip;
WT_UPDATE *upd;
uint8_t v;
btree = S2BT(session);
page = cbt->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
cell = WT_COL_PTR(page, &page->u.col_var.d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->u.row.d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take the key and
* related WT_UPDATE item. Otherwise, take the key from the
* original page, and the value from any related WT_UPDATE item,
* or the page if the key was never updated.
*/
if (cbt->ins != NULL &&
(upd = __wt_txn_read(session, cbt->ins->upd)) != NULL) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
} else {
WT_RET(
__wt_row_key(session, page, rip, &cursor->key, 0));
upd = __wt_txn_read(session, WT_ROW_UPDATE(page, rip));
}
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the original cell (which may be empty). */
if ((cell = __wt_row_value(page, rip)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* The value is an on-page cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, &unpack);
ret = __wt_cell_unpack_ref(session, &unpack, &cursor->value);
/*
* Restart for a variable-length column-store. We could catch restart
* higher up the call-stack but there's no point to it: unlike row-store
* (where a normal search path finds cached overflow values), we have to
* access the page's reconciliation structures, and that's as easy here
* as higher up the stack.
*/
if (ret == WT_RESTART && page->type == WT_PAGE_COL_VAR)
ret = __wt_ovfl_cache_col_restart(
session, page, &unpack, &cursor->value);
return (ret);
}
/*
* __las_page_instantiate --
* Instantiate lookaside update records in a recently read page.
*/
static int
__las_page_instantiate(WT_SESSION_IMPL *session,
WT_REF *ref, uint32_t read_id, const uint8_t *addr, size_t addr_size)
{
WT_CURSOR *cursor;
WT_CURSOR_BTREE cbt;
WT_DECL_ITEM(current_key);
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_ITEM(las_value);
WT_DECL_RET;
WT_PAGE *page;
WT_UPDATE *first_upd, *last_upd, *upd;
size_t incr, total_incr;
uint64_t current_recno, las_counter, las_txnid, recno, upd_txnid;
uint32_t las_id, upd_size, session_flags;
int exact;
const uint8_t *p;
cursor = NULL;
page = ref->page;
first_upd = last_upd = upd = NULL;
total_incr = 0;
current_recno = recno = WT_RECNO_OOB;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
__wt_btcur_init(session, &cbt);
__wt_btcur_open(&cbt);
WT_ERR(__wt_scr_alloc(session, 0, ¤t_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_scr_alloc(session, 0, &las_value));
/* Open a lookaside table cursor. */
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* The lookaside records are in key and update order, that is, there
* will be a set of in-order updates for a key, then another set of
* in-order updates for a subsequent key. We process all of the updates
* for a key and then insert those updates into the page, then all the
* updates for the next key, and so on.
*
* Search for the block's unique prefix, stepping through any matching
* records.
*/
las_addr->data = addr;
las_addr->size = addr_size;
las_key->size = 0;
cursor->set_key(
cursor, read_id, las_addr, (uint64_t)0, (uint32_t)0, las_key);
if ((ret = cursor->search_near(cursor, &exact)) == 0 && exact < 0)
ret = cursor->next(cursor);
for (; ret == 0; ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* Confirm the search using the unique prefix; if not a match,
* we're done searching for records for this page.
*/
if (las_id != read_id ||
las_addr->size != addr_size ||
memcmp(las_addr->data, addr, addr_size) != 0)
break;
/*
* If the on-page value has become globally visible, this record
* is no longer needed.
*/
if (__wt_txn_visible_all(session, las_txnid))
continue;
/* Allocate the WT_UPDATE structure. */
WT_ERR(cursor->get_value(
cursor, &upd_txnid, &upd_size, las_value));
WT_ERR(__wt_update_alloc(session,
(upd_size == WT_UPDATE_DELETED_VALUE) ? NULL : las_value,
&upd, &incr));
total_incr += incr;
upd->txnid = upd_txnid;
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
p = las_key->data;
WT_ERR(__wt_vunpack_uint(&p, 0, &recno));
if (current_recno == recno)
break;
WT_ASSERT(session, current_recno < recno);
if (first_upd != NULL) {
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
}
current_recno = recno;
break;
case WT_PAGE_ROW_LEAF:
if (current_key->size == las_key->size &&
memcmp(current_key->data,
las_key->data, las_key->size) == 0)
break;
if (first_upd != NULL) {
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
}
WT_ERR(__wt_buf_set(session,
current_key, las_key->data, las_key->size));
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Append the latest update to the list. */
if (first_upd == NULL)
first_upd = last_upd = upd;
else {
last_upd->next = upd;
last_upd = upd;
}
upd = NULL;
}
WT_ERR_NOTFOUND_OK(ret);
/* Insert the last set of updates, if any. */
if (first_upd != NULL)
switch (page->type) {
case WT_PAGE_COL_FIX:
case WT_PAGE_COL_VAR:
WT_ERR(__col_instantiate(session,
current_recno, ref, &cbt, first_upd));
first_upd = NULL;
break;
case WT_PAGE_ROW_LEAF:
WT_ERR(__row_instantiate(session,
current_key, ref, &cbt, first_upd));
first_upd = NULL;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
/* Discard the cursor. */
WT_ERR(__wt_las_cursor_close(session, &cursor, session_flags));
if (total_incr != 0) {
__wt_cache_page_inmem_incr(session, page, total_incr);
/*
* We've modified/dirtied the page, but that's not necessary and
* if we keep the page clean, it's easier to evict. We leave the
* lookaside table updates in place, so if we evict this page
* without dirtying it, any future instantiation of it will find
* the records it needs. If the page is dirtied before eviction,
* then we'll write any needed lookaside table records for the
* new location of the page.
*/
__wt_page_modify_clear(session, page);
}
err: WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
WT_TRET(__wt_btcur_close(&cbt, 1));
/*
* On error, upd points to a single unlinked WT_UPDATE structure,
* first_upd points to a list.
*/
if (upd != NULL)
__wt_free(session, upd);
if (first_upd != NULL)
__wt_free_update_list(session, first_upd);
__wt_scr_free(session, ¤t_key);
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
__wt_scr_free(session, &las_value);
return (ret);
}
/*
* __clsm_open_cursors --
* Open cursors for the current set of files.
*/
static int
__clsm_open_cursors(
WT_CURSOR_LSM *clsm, int update, u_int start_chunk, uint32_t start_id)
{
WT_CURSOR *c, **cp, *primary;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_LSM_TREE *lsm_tree;
WT_SESSION_IMPL *session;
WT_TXN *txn;
const char *checkpoint, *ckpt_cfg[3];
uint64_t saved_gen;
u_int i, nchunks, ngood, nupdates;
int locked;
c = &clsm->iface;
session = (WT_SESSION_IMPL *)c->session;
txn = &session->txn;
lsm_tree = clsm->lsm_tree;
chunk = NULL;
ckpt_cfg[0] = WT_CONFIG_BASE(session, session_open_cursor);
ckpt_cfg[1] = "checkpoint=" WT_CHECKPOINT ",raw";
ckpt_cfg[2] = NULL;
/* Copy the key, so we don't lose the cursor position. */
if (F_ISSET(c, WT_CURSTD_KEY_INT) && !WT_DATA_IN_ITEM(&c->key))
WT_RET(__wt_buf_set(
session, &c->key, c->key.data, c->key.size));
F_CLR(clsm, WT_CLSM_ITERATE_NEXT | WT_CLSM_ITERATE_PREV);
if (update) {
if (txn->isolation == TXN_ISO_SNAPSHOT)
F_SET(clsm, WT_CLSM_OPEN_SNAPSHOT);
} else
F_SET(clsm, WT_CLSM_OPEN_READ);
WT_RET(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
/*
* If there is no in-memory chunk in the tree for an update operation,
* create one.
*
* !!!
* It is exceeding unlikely that we get here at all, but if there is a
* transaction in progress and it rolls back, it would leave the
* metadata inconsistent.
*/
if (update && (lsm_tree->nchunks == 0 ||
(chunk = lsm_tree->chunk[lsm_tree->nchunks - 1]) == NULL ||
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK))) {
/* Release our lock because switch will get a write lock. */
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
WT_ERR(__wt_lsm_tree_switch(session, lsm_tree));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
}
F_SET(session, WT_SESSION_NO_CACHE_CHECK);
/* Merge cursors have already figured out how many chunks they need. */
retry: if (F_ISSET(clsm, WT_CLSM_MERGE)) {
nchunks = clsm->nchunks;
ngood = 0;
/*
* We may have raced with another merge completing. Check that
* we're starting at the right offset in the chunk array.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id) {
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++) {
chunk = lsm_tree->chunk[start_chunk];
if (chunk->id == start_id)
break;
}
/* We have to find the start chunk: merge locked it. */
WT_ASSERT(session, start_chunk < lsm_tree->nchunks);
}
WT_ASSERT(session, start_chunk + nchunks <= lsm_tree->nchunks);
} else {
nchunks = lsm_tree->nchunks;
/*
* If we are only opening the cursor for updates, only open the
* primary chunk, plus any other chunks that might be required
* to detect snapshot isolation conflicts.
*/
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
WT_ERR(__wt_realloc_def(session,
&clsm->txnid_alloc, nchunks,
&clsm->txnid_max));
if (F_ISSET(clsm, WT_CLSM_OPEN_READ))
ngood = nupdates = 0;
else if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
/*
* Keep going until all updates in the next
* chunk are globally visible. Copy the maximum
* transaction IDs into the cursor as we go.
*/
for (ngood = nchunks - 1, nupdates = 1;
ngood > 0;
ngood--, nupdates++) {
chunk = lsm_tree->chunk[ngood - 1];
clsm->txnid_max[ngood - 1] =
chunk->txnid_max;
if (__wt_txn_visible_all(
session, chunk->txnid_max))
break;
}
} else {
nupdates = 1;
ngood = nchunks - 1;
}
/* Check how many cursors are already open. */
for (cp = clsm->cursors + ngood;
ngood < clsm->nchunks && ngood < nchunks;
cp++, ngood++) {
chunk = lsm_tree->chunk[ngood];
/* If the cursor isn't open yet, we're done. */
if (*cp == NULL)
break;
/* Easy case: the URIs don't match. */
if (strcmp((*cp)->uri, chunk->uri) != 0)
break;
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
if (checkpoint == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty)
break;
/* Make sure the Bloom config matches. */
if (clsm->blooms[ngood] == NULL &&
F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
break;
}
/* Spurious generation bump? */
if (ngood == clsm->nchunks && clsm->nchunks == nchunks) {
clsm->dsk_gen = lsm_tree->dsk_gen;
goto err;
}
/*
* Close any cursors we no longer need.
*
* Drop the LSM tree lock while we do this: if the cache is
* full, we may block while closing a cursor. Save the
* generation number and retry if it has changed under us.
*/
if (clsm->cursors != NULL && (ngood < clsm->nchunks ||
(!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0))) {
saved_gen = lsm_tree->dsk_gen;
locked = 0;
WT_ERR(__wt_lsm_tree_unlock(session, lsm_tree));
if (!F_ISSET(clsm, WT_CLSM_OPEN_READ) && nupdates > 0)
WT_ERR(__clsm_close_cursors(
clsm, 0, nchunks - nupdates));
WT_ERR(__clsm_close_cursors(
clsm, ngood, clsm->nchunks));
WT_ERR(__wt_lsm_tree_lock(session, lsm_tree, 0));
locked = 1;
if (lsm_tree->dsk_gen != saved_gen)
goto retry;
}
/* Detach from our old primary. */
clsm->primary_chunk = NULL;
clsm->current = NULL;
}
WT_ERR(__wt_realloc_def(session,
&clsm->bloom_alloc, nchunks, &clsm->blooms));
WT_ERR(__wt_realloc_def(session,
&clsm->cursor_alloc, nchunks, &clsm->cursors));
clsm->nchunks = nchunks;
/* Open the cursors for chunks that have changed. */
for (i = ngood, cp = clsm->cursors + i; i != nchunks; i++, cp++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Copy the maximum transaction ID. */
if (F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT))
clsm->txnid_max[i] = chunk->txnid_max;
/*
* Read from the checkpoint if the file has been written.
* Once all cursors switch, the in-memory tree can be evicted.
*/
WT_ASSERT(session, *cp == NULL);
ret = __wt_open_cursor(session, chunk->uri, c,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) && !chunk->empty) ?
ckpt_cfg : NULL, cp);
/*
* XXX kludge: we may have an empty chunk where no checkpoint
* was written. If so, try to open the ordinary handle on that
* chunk instead.
*/
if (ret == WT_NOTFOUND &&
F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
ret = __wt_open_cursor(
session, chunk->uri, c, NULL, cp);
if (ret == 0)
chunk->empty = 1;
}
WT_ERR(ret);
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE))
WT_ERR(__wt_bloom_open(session, chunk->bloom_uri,
lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count,
c, &clsm->blooms[i]));
/* Child cursors always use overwrite and raw mode. */
F_SET(*cp, WT_CURSTD_OVERWRITE | WT_CURSTD_RAW);
}
/* The last chunk is our new primary. */
if (chunk != NULL && !F_ISSET(chunk, WT_LSM_CHUNK_ONDISK)) {
clsm->primary_chunk = chunk;
primary = clsm->cursors[clsm->nchunks - 1];
WT_WITH_BTREE(session, ((WT_CURSOR_BTREE *)(primary))->btree,
__wt_btree_evictable(session, 0));
}
clsm->dsk_gen = lsm_tree->dsk_gen;
err: F_CLR(session, WT_SESSION_NO_CACHE_CHECK);
#ifdef HAVE_DIAGNOSTIC
/* Check that all cursors are open as expected. */
if (ret == 0 && F_ISSET(clsm, WT_CLSM_OPEN_READ)) {
for (i = 0, cp = clsm->cursors; i != clsm->nchunks; cp++, i++) {
chunk = lsm_tree->chunk[i + start_chunk];
/* Make sure the cursor is open. */
WT_ASSERT(session, *cp != NULL);
/* Easy case: the URIs should match. */
WT_ASSERT(session, strcmp((*cp)->uri, chunk->uri) == 0);
/* Make sure the checkpoint config matches. */
checkpoint = ((WT_CURSOR_BTREE *)*cp)->
btree->dhandle->checkpoint;
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_ONDISK) &&
!chunk->empty) ?
checkpoint != NULL : checkpoint == NULL);
/* Make sure the Bloom config matches. */
WT_ASSERT(session,
(F_ISSET(chunk, WT_LSM_CHUNK_BLOOM) &&
!F_ISSET(clsm, WT_CLSM_MERGE)) ?
clsm->blooms[i] != NULL : clsm->blooms[i] == NULL);
}
}
#endif
if (locked)
WT_TRET(__wt_lsm_tree_unlock(session, lsm_tree));
return (ret);
}
/*
* __wt_bt_read --
* Read a cookie referenced block into a buffer.
*/
int
__wt_bt_read(WT_SESSION_IMPL *session,
WT_ITEM *buf, const uint8_t *addr, size_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(etmp);
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_ENCRYPTOR *encryptor;
WT_ITEM *ip;
const WT_PAGE_HEADER *dsk;
const char *fail_msg;
size_t result_len;
btree = S2BT(session);
bm = btree->bm;
fail_msg = NULL; /* -Wuninitialized */
/*
* If anticipating a compressed or encrypted block, read into a scratch
* buffer and decompress into the caller's buffer. Else, read directly
* into the caller's buffer.
*/
if (btree->compressor == NULL && btree->kencryptor == NULL) {
WT_RET(bm->read(bm, session, buf, addr, addr_size));
dsk = buf->data;
ip = NULL;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->data;
ip = tmp;
}
/*
* If the block is encrypted, copy the skipped bytes of the original
* image into place, then decrypt.
*/
if (F_ISSET(dsk, WT_PAGE_ENCRYPTED)) {
if (btree->kencryptor == NULL ||
(encryptor = btree->kencryptor->encryptor) == NULL ||
encryptor->decrypt == NULL) {
fail_msg =
"encrypted block in file for which no encryption "
"configured";
goto corrupt;
}
WT_ERR(__wt_scr_alloc(session, 0, &etmp));
if ((ret = __wt_decrypt(session,
encryptor, WT_BLOCK_ENCRYPT_SKIP, ip, etmp)) != 0) {
fail_msg = "block decryption failed";
goto corrupt;
}
ip = etmp;
dsk = ip->data;
} else if (btree->kencryptor != NULL) {
fail_msg =
"unencrypted block in file for which encryption configured";
goto corrupt;
}
if (F_ISSET(dsk, WT_PAGE_COMPRESSED)) {
if (btree->compressor == NULL ||
btree->compressor->decompress == NULL) {
fail_msg =
"compressed block in file for which no compression "
"configured";
goto corrupt;
}
/*
* Size the buffer based on the in-memory bytes we're expecting
* from decompression.
*/
WT_ERR(__wt_buf_initsize(session, buf, dsk->mem_size));
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, ip->data, WT_BLOCK_COMPRESS_SKIP);
ret = btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)ip->data + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->mem_size - WT_BLOCK_COMPRESS_SKIP, &result_len);
/*
* If checksums were turned off because we're depending on the
* decompression to fail on any corrupted data, we'll end up
* here after corruption happens. If we're salvaging the file,
* it's OK, otherwise it's really, really bad.
*/
if (ret != 0 ||
result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP) {
fail_msg = "block decryption failed";
goto corrupt;
}
} else
/*
* If we uncompressed above, the page is in the correct buffer.
* If we get here the data may be in the wrong buffer and the
* buffer may be the wrong size. If needed, get the page
* into the destination buffer.
*/
if (ip != NULL)
WT_ERR(__wt_buf_set(
session, buf, ip->data, dsk->mem_size));
/* If the handle is a verify handle, verify the physical page. */
if (F_ISSET(btree, WT_BTREE_VERIFY)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->addr_string(bm, session, tmp, addr, addr_size));
WT_ERR(__wt_verify_dsk(session, tmp->data, buf));
}
WT_STAT_FAST_CONN_INCR(session, cache_read);
WT_STAT_FAST_DATA_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_STAT_FAST_DATA_INCR(session, compress_read);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_read, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_read, dsk->mem_size);
if (0) {
corrupt: if (ret == 0)
ret = WT_ERROR;
if (!F_ISSET(btree, WT_BTREE_VERIFY) &&
!F_ISSET(session, WT_SESSION_QUIET_CORRUPT_FILE)) {
__wt_err(session, ret, "%s", fail_msg);
ret = __wt_illegal_value(session, btree->dhandle->name);
}
}
err: __wt_scr_free(session, &tmp);
__wt_scr_free(session, &etmp);
return (ret);
}
int __wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_UPDATE *upd)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_PAGE *page;
WT_ROW *rip;
uint8_t v;
switch (page->type){
case WT_PAGE_COL_FIX:
cursor->recno = cbt->recno;
/*cursor对应的是一个upd,直接返回value*/
if (upd != NULL){
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return 0;
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return __wt_buf_set(session, &cursor->value, &v, 1);
case WT_PAGE_COL_VAR:
cursor->recno = cbt->recno;
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/*获得对应的cell,并通过cell得到K/V值*/
cell = WT_COL_PTR(page, &page->pg_var_d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->pg_row_d[cbt->slot];
if (cbt->ins != NULL){ /*插入的k/v对*/
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
}
else if (cbt->compare == 0){/*比较器定位到了对应的k/v对*/
cursor->key.data = cbt->search_key.data;
cursor->key.size = cbt->search_key.size;
}
else
WT_RET(__wt_row_leaf_key(session, page, rip, &cursor->key, 0)); /*设置key的值*/
/*值是在append/update list当中，从当中取*/
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/*可以直接通过rip指针获得value,K/V是存储在cell空间之内*/
if (__wt_row_leaf_value(page, rip, &cursor->value))
return 0;
/*不是连续存储的，需要通过解析cell来定位到value*/
if (cell = __wt_row_leaf_value_cell(page, rip, NULL) == NULL){
cursor->value.size = 0;
return 0;
}
break;
WT_ILLEGAL_VALUE(session);
}
/*通过cell解析到对应的value值, ovfl item*/
__wt_cell_unpack(cell, &unpack);
WT_RET(__wt_page_cell_data_ref(session, page, &unpack, &cursor->value));
return 0;
}
/*
* __wt_struct_plan --
* Given a table cursor containing a complete table, build the "projection
* plan" to distribute the columns to dependent stores. A string
* representing the plan will be appended to the plan buffer.
*/
int
__wt_struct_plan(WT_SESSION_IMPL *session, WT_TABLE *table,
const char *columns, size_t len, bool value_only, WT_ITEM *plan)
{
WT_CONFIG conf;
WT_CONFIG_ITEM k, v;
WT_DECL_RET;
u_int cg, col, current_cg, current_col, i, start_cg, start_col;
char coltype, current_coltype;
bool have_it;
start_cg = start_col = UINT_MAX; /* -Wuninitialized */
/* Work through the value columns by skipping over the key columns. */
__wt_config_initn(session, &conf, columns, len);
if (value_only)
for (i = 0; i < table->nkey_columns; i++)
WT_RET(__wt_config_next(&conf, &k, &v));
current_cg = cg = 0;
current_col = col = INT_MAX;
current_coltype = coltype = WT_PROJ_KEY; /* Keep lint quiet. */
for (i = 0; (ret = __wt_config_next(&conf, &k, &v)) == 0; i++) {
have_it = false;
while ((ret = __find_next_col(session, table,
&k, &cg, &col, &coltype)) == 0 &&
(!have_it || cg != start_cg || col != start_col)) {
/*
* First we move to the column. If that is in a
* different column group to the last column we
* accessed, or before the last column in the same
* column group, or moving from the key to the value,
* we need to switch column groups or rewind.
*/
if (current_cg != cg || current_col > col ||
current_coltype != coltype) {
WT_ASSERT(session, !value_only ||
coltype == WT_PROJ_VALUE);
WT_RET(__wt_buf_catfmt(
session, plan, "%u%c", cg, coltype));
/*
* Set the current column group and column
* within the table.
*/
current_cg = cg;
current_col = 0;
current_coltype = coltype;
}
/* Now move to the column we want. */
if (current_col < col) {
if (col - current_col > 1)
WT_RET(__wt_buf_catfmt(session,
plan, "%u", col - current_col));
WT_RET(__wt_buf_catfmt(session,
plan, "%c", WT_PROJ_SKIP));
}
/*
* Now copy the value in / out. In the common case,
* where each value is used in one column, we do a
* "next" operation. If the value is used again, we do
* a "reuse" operation to avoid making another copy.
*/
if (!have_it) {
WT_RET(__wt_buf_catfmt(session,
plan, "%c", WT_PROJ_NEXT));
start_cg = cg;
start_col = col;
have_it = true;
} else
WT_RET(__wt_buf_catfmt(session,
plan, "%c", WT_PROJ_REUSE));
current_col = col + 1;
}
/*
* We may fail to find a column if it is a custom extractor.
* In that case, treat it as the first value column: we only
* ever use such plans to extract the primary key from the
* index.
*/
if (ret == WT_NOTFOUND)
WT_RET(__wt_buf_catfmt(session, plan,
"0%c%c", WT_PROJ_VALUE, WT_PROJ_NEXT));
}
WT_RET_TEST(ret != WT_NOTFOUND, ret);
/* Special case empty plans. */
if (i == 0 && plan->size == 0)
WT_RET(__wt_buf_set(session, plan, "", 1));
return (0);
}
/*
* __wt_struct_reformat --
* Given a table and a list of columns (which could be values in a column
* group or index keys), calculate the resulting new format string.
* The result will be appended to the format buffer.
*/
int
__wt_struct_reformat(WT_SESSION_IMPL *session, WT_TABLE *table,
const char *columns, size_t len, const char *extra_cols, bool value_only,
WT_ITEM *format)
{
WT_CONFIG config;
WT_CONFIG_ITEM k, next_k, next_v;
WT_DECL_PACK_VALUE(pv);
WT_DECL_RET;
bool have_next;
__wt_config_initn(session, &config, columns, len);
/*
* If an empty column list is specified, this will fail with
* WT_NOTFOUND, that's okay.
*/
WT_RET_NOTFOUND_OK(ret = __wt_config_next(&config, &next_k, &next_v));
if (ret == WT_NOTFOUND) {
if (extra_cols != NULL) {
__wt_config_init(session, &config, extra_cols);
WT_RET(__wt_config_next(&config, &next_k, &next_v));
extra_cols = NULL;
} else if (format->size == 0) {
WT_RET(__wt_buf_set(session, format, "", 1));
return (0);
}
}
do {
k = next_k;
ret = __wt_config_next(&config, &next_k, &next_v);
if (ret != 0 && ret != WT_NOTFOUND)
return (ret);
have_next = ret == 0;
if (!have_next && extra_cols != NULL) {
__wt_config_init(session, &config, extra_cols);
WT_RET(__wt_config_next(&config, &next_k, &next_v));
have_next = true;
extra_cols = NULL;
}
if ((ret = __find_column_format(session,
table, &k, value_only, &pv)) != 0) {
if (value_only && ret == EINVAL)
WT_RET_MSG(session, EINVAL,
"A column group cannot store key column "
"'%.*s' in its value", (int)k.len, k.str);
WT_RET_MSG(session, EINVAL,
"Column '%.*s' not found", (int)k.len, k.str);
}
/*
* Check whether we're moving an unsized WT_ITEM from the end
* to the middle, or vice-versa. This determines whether the
* size needs to be prepended. This is the only case where the
* destination size can be larger than the source size.
*/
if (pv.type == 'u' && !pv.havesize && have_next)
pv.type = 'U';
else if (pv.type == 'U' && !have_next)
pv.type = 'u';
if (pv.havesize)
WT_RET(__wt_buf_catfmt(session,
format, "%" PRIu32 "%c", pv.size, pv.type));
else
WT_RET(__wt_buf_catfmt(session, format, "%c", pv.type));
} while (have_next);
return (0);
}
/*
* __wt_meta_ckptlist_get --
* Load all available checkpoint information for a file.
*/
int
__wt_meta_ckptlist_get(
WT_SESSION_IMPL *session, const char *fname, WT_CKPT **ckptbasep)
{
WT_CKPT *ckpt, *ckptbase;
WT_CONFIG ckptconf;
WT_CONFIG_ITEM a, k, v;
WT_DECL_RET;
WT_ITEM *buf;
size_t allocated, slot;
const char *config;
char timebuf[64];
*ckptbasep = NULL;
buf = NULL;
ckptbase = NULL;
allocated = slot = 0;
config = NULL;
/* Retrieve the metadata information for the file. */
WT_RET(__wt_metadata_read(session, fname, &config));
/* Load any existing checkpoints into the array. */
WT_ERR(__wt_scr_alloc(session, 0, &buf));
if (__wt_config_getones(session, config, "checkpoint", &v) == 0 &&
__wt_config_subinit(session, &ckptconf, &v) == 0)
for (; __wt_config_next(&ckptconf, &k, &v) == 0; ++slot) {
if (slot * sizeof(WT_CKPT) == allocated)
WT_ERR(__wt_realloc(session, &allocated,
(slot + 50) * sizeof(WT_CKPT), &ckptbase));
ckpt = &ckptbase[slot];
/*
* Copy the name, address (raw and hex), order and time
* into the slot. If there's no address, it's a fake.
*/
WT_ERR(
__wt_strndup(session, k.str, k.len, &ckpt->name));
WT_ERR(__wt_config_subgets(session, &v, "addr", &a));
WT_ERR(
__wt_buf_set(session, &ckpt->addr, a.str, a.len));
if (a.len == 0)
F_SET(ckpt, WT_CKPT_FAKE);
else
WT_ERR(__wt_nhex_to_raw(
session, a.str, a.len, &ckpt->raw));
WT_ERR(__wt_config_subgets(session, &v, "order", &a));
if (a.val == 0)
goto format;
ckpt->order = a.val;
WT_ERR(__wt_config_subgets(session, &v, "time", &a));
if (a.len == 0)
goto format;
if (a.len > sizeof(timebuf) - 1)
goto format;
memcpy(timebuf, a.str, a.len);
timebuf[a.len] = '\0';
if (sscanf(timebuf, "%" SCNuMAX, &ckpt->sec) != 1)
goto format;
WT_ERR(__wt_config_subgets(session, &v, "size", &a));
ckpt->ckpt_size = (uint64_t)a.val;
}
/*
* Allocate an extra slot for a new value, plus a slot to mark the end.
*
* This isn't very clean, but there's necessary cooperation between the
* schema layer (that maintains the list of checkpoints), the btree
* layer (that knows when the root page is written, creating a new
* checkpoint), and the block manager (which actually creates the
* checkpoint). All of that cooperation is handled in the WT_CKPT
* structure referenced from the WT_BTREE structure.
*/
if ((slot + 2) * sizeof(WT_CKPT) > allocated)
WT_ERR(__wt_realloc(session, &allocated,
(slot + 2) * sizeof(WT_CKPT), &ckptbase));
/* Sort in creation-order. */
qsort(ckptbase, slot, sizeof(WT_CKPT), __ckpt_compare_order);
/* Return the array to our caller. */
*ckptbasep = ckptbase;
if (0) {
format: WT_ERR_MSG(session, WT_ERROR, "corrupted checkpoint list");
err: __wt_meta_ckptlist_free(session, ckptbase);
}
__wt_free(session, config);
__wt_scr_free(&buf);
return (ret);
}
/*
* __curbulk_insert_var --
* Variable-length column-store bulk cursor insert.
*/
static int
__curbulk_insert_var(WT_CURSOR *cursor)
{
WT_BTREE *btree;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
WT_SESSION_IMPL *session;
uint64_t recno;
cbulk = (WT_CURSOR_BULK *)cursor;
btree = cbulk->cbt.btree;
/*
* Bulk cursor inserts are updates, but don't need auto-commit
* transactions because they are single-threaded and not visible
* until the bulk cursor is closed.
*/
CURSOR_API_CALL(cursor, session, insert, btree);
WT_STAT_FAST_DATA_INCR(session, cursor_insert_bulk);
/*
* If the "append" flag was configured, the application doesn't have to
* supply a key, else require a key.
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
recno = cbulk->recno + 1;
else {
WT_CURSOR_CHECKKEY(cursor);
if ((recno = cursor->recno) <= cbulk->recno)
WT_ERR(__bulk_col_keycmp_err(cbulk));
}
WT_CURSOR_CHECKVALUE(cursor);
if (!cbulk->first_insert) {
/*
* If not the first insert and the key space is sequential,
* compare the current value against the last value; if the
* same, just increment the RLE count.
*/
if (recno == cbulk->recno + 1 &&
cbulk->last.size == cursor->value.size &&
memcmp(cbulk->last.data,
cursor->value.data, cursor->value.size) == 0) {
++cbulk->rle;
++cbulk->recno;
goto duplicate;
}
/* Insert the previous key/value pair. */
WT_ERR(__wt_bulk_insert_var(session, cbulk, false));
} else
cbulk->first_insert = false;
/*
* Insert any skipped records as deleted records, update the current
* record count and RLE counter.
*/
if (recno != cbulk->recno + 1) {
cbulk->rle = (recno - cbulk->recno) - 1;
WT_ERR(__wt_bulk_insert_var(session, cbulk, true));
}
cbulk->rle = 1;
cbulk->recno = recno;
/* Save a copy of the value for the next comparison. */
ret = __wt_buf_set(session,
&cbulk->last, cursor->value.data, cursor->value.size);
duplicate:
err: API_END_RET(session, ret);
}
/*
* __wt_bt_read --
* Read a cookie referenced block into a buffer.
*/
int
__wt_bt_read(WT_SESSION_IMPL *session,
WT_ITEM *buf, const uint8_t *addr, size_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
const WT_PAGE_HEADER *dsk;
size_t result_len;
btree = S2BT(session);
bm = btree->bm;
/*
* If anticipating a compressed block, read into a scratch buffer and
* decompress into the caller's buffer. Else, read directly into the
* caller's buffer.
*/
if (btree->compressor == NULL) {
WT_RET(bm->read(bm, session, buf, addr, addr_size));
dsk = buf->data;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->data;
}
/*
* If the block is compressed, copy the skipped bytes of the original
* image into place, then decompress.
*/
if (F_ISSET(dsk, WT_PAGE_COMPRESSED)) {
if (btree->compressor == NULL ||
btree->compressor->decompress == NULL)
WT_ERR_MSG(session, WT_ERROR,
"read compressed block where no compression engine "
"configured");
/*
* We're allocating the exact number of bytes we're expecting
* from decompression.
*/
WT_ERR(__wt_buf_initsize(session, buf, dsk->mem_size));
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, tmp->data, WT_BLOCK_COMPRESS_SKIP);
ret = btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->mem_size - WT_BLOCK_COMPRESS_SKIP, &result_len);
/*
* If checksums were turned off because we're depending on the
* decompression to fail on any corrupted data, we'll end up
* here after corruption happens. If we're salvaging the file,
* it's OK, otherwise it's really, really bad.
*/
if (ret != 0 ||
result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(
F_ISSET(btree, WT_BTREE_VERIFY) ||
F_ISSET(session, WT_SESSION_SALVAGE_CORRUPT_OK) ?
WT_ERROR :
__wt_illegal_value(session, btree->dhandle->name));
} else
if (btree->compressor == NULL)
buf->size = dsk->mem_size;
else
/*
* We guessed wrong: there was a compressor, but this
* block was not compressed, and now the page is in the
* wrong buffer and the buffer may be of the wrong size.
* This should be rare, but happens with small blocks
* that aren't worth compressing.
*/
WT_ERR(__wt_buf_set(
session, buf, tmp->data, dsk->mem_size));
/* If the handle is a verify handle, verify the physical page. */
if (F_ISSET(btree, WT_BTREE_VERIFY)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->addr_string(bm, session, tmp, addr, addr_size));
WT_ERR(__wt_verify_dsk(session, (const char *)tmp->data, buf));
}
WT_STAT_FAST_CONN_INCR(session, cache_read);
WT_STAT_FAST_DATA_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_STAT_FAST_DATA_INCR(session, compress_read);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_read, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_read, dsk->mem_size);
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_block_read_off --
* Read an addr/size pair referenced block into a buffer.
*/
int
__wt_block_read_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
WT_ITEM *buf, off_t offset, uint32_t size, uint32_t cksum)
{
WT_BLOCK_HEADER *blk;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t result_len;
uint32_t page_cksum;
WT_VERBOSE_RET(session, read,
"off %" PRIuMAX ", size %" PRIu32 ", cksum %" PRIu32,
(uintmax_t)offset, size, cksum);
#ifdef HAVE_DIAGNOSTIC
/*
* In diagnostic mode, verify the block we're about to read isn't on
* either the available or discard lists.
*
* Don't check during salvage, it's possible we're reading an already
* freed overflow page.
*/
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
WT_RET(
__wt_block_misplaced(session, block, "read", offset, size));
#endif
/*
* If we're compressing the file blocks, place the initial read into a
* scratch buffer, we're going to have to re-allocate more memory for
* decompression. Else check the caller's buffer size and grow it as
* necessary, there will only be one buffer.
*/
if (block->compressor == NULL) {
F_SET(buf, WT_ITEM_ALIGNED);
WT_RET(__wt_buf_init(session, buf, size));
buf->size = size;
dsk = buf->mem;
} else {
WT_RET(__wt_scr_alloc(session, size, &tmp));
tmp->size = size;
dsk = tmp->mem;
}
/* Read. */
WT_ERR(__wt_read(session, block->fh, offset, size, dsk));
blk = WT_BLOCK_HEADER_REF(dsk);
/* Validate the checksum. */
if (block->checksum &&
cksum != WT_BLOCK_CHECKSUM_NOT_SET &&
blk->cksum != WT_BLOCK_CHECKSUM_NOT_SET) {
blk->cksum = 0;
page_cksum = __wt_cksum(dsk, size);
if (page_cksum == WT_BLOCK_CHECKSUM_NOT_SET)
++page_cksum;
if (cksum != page_cksum) {
if (!F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR))
__wt_errx(session,
"read checksum error [%"
PRIu32 "B @ %" PRIuMAX ", %"
PRIu32 " != %" PRIu32 "]",
size, (uintmax_t)offset, cksum, page_cksum);
WT_ERR(WT_ERROR);
}
}
/*
* If the in-memory block size is larger than the on-disk block size,
* the block is compressed. Size the user's buffer, copy the skipped
* bytes of the original image into place, then decompress.
*
* If the in-memory block size is less than or equal to the on-disk
* block size, the block is not compressed.
*/
if (blk->disk_size < dsk->size) {
if (block->compressor == NULL)
WT_ERR(__wt_illegal_value(session, block->name));
WT_ERR(__wt_buf_init(session, buf, dsk->size));
buf->size = dsk->size;
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, tmp->mem, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(block->compressor->decompress(
block->compressor, &session->iface,
(uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->size - WT_BLOCK_COMPRESS_SKIP,
&result_len));
if (result_len != dsk->size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(__wt_illegal_value(session, block->name));
} else
if (block->compressor == NULL)
buf->size = dsk->size;
else
/*
* We guessed wrong: there was a compressor, but this
* block was not compressed, and now the page is in the
* wrong buffer and the buffer may be of the wrong size.
* This should be rare, why configure a compressor that
* doesn't work? Allocate a buffer of the right size
* (we used a scratch buffer which might be large), and
* copy the data into place.
*/
WT_ERR(
__wt_buf_set(session, buf, tmp->data, dsk->size));
WT_BSTAT_INCR(session, page_read);
WT_CSTAT_INCR(session, block_read);
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __wt_las_sweep --
* Sweep the lookaside table.
*/
int
__wt_las_sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *cursor;
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_RET;
WT_ITEM *key;
uint64_t cnt, las_counter, las_txnid;
uint32_t las_id, session_flags;
int notused;
conn = S2C(session);
cursor = NULL;
key = &conn->las_sweep_key;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* If we're not starting a new sweep, position the cursor using the key
* from the last call (we don't care if we're before or after the key,
* just roughly in the same spot is fine).
*/
if (conn->las_sweep_call != 0 && key->data != NULL) {
__wt_cursor_set_raw_key(cursor, key);
if ((ret = cursor->search_near(cursor, ¬used)) != 0)
goto srch_notfound;
}
/*
* The sweep server wakes up every 10 seconds (by default), it's a slow
* moving thread. Try to review the entire lookaside table once every 5
* minutes, or every 30 calls.
*
* The reason is because the lookaside table exists because we're seeing
* cache/eviction pressure (it allows us to trade performance and disk
* space for cache space), and it's likely lookaside blocks are being
* evicted, and reading them back in doesn't help things. A trickier,
* but possibly better, alternative might be to review all lookaside
* blocks in the cache in order to get rid of them, and slowly review
* lookaside blocks that have already been evicted.
*
* We can't know for sure how many records are in the lookaside table,
* the cursor insert and remove statistics aren't updated atomically.
* Start with reviewing 100 rows, and if it takes more than the target
* number of calls to finish, increase the number of rows checked on
* each call; if it takes less than the target calls to finish, then
* decrease the number of rows reviewed on each call (but never less
* than 100).
*/
#define WT_SWEEP_LOOKASIDE_MIN_CNT 100
#define WT_SWEEP_LOOKASIDE_PASS_TARGET 30
++conn->las_sweep_call;
if ((cnt = conn->las_sweep_cnt) < WT_SWEEP_LOOKASIDE_MIN_CNT)
cnt = conn->las_sweep_cnt = WT_SWEEP_LOOKASIDE_MIN_CNT;
/* Walk the file. */
for (; cnt > 0 && (ret = cursor->next(cursor)) == 0; --cnt) {
/*
* If the loop terminates after completing a work unit, we will
* continue the table sweep next time. Get a local copy of the
* sweep key, we're going to reset the cursor; do so before
* calling cursor.remove, cursor.remove can discard our hazard
* pointer and the page could be evicted from underneath us.
*/
if (cnt == 1) {
WT_ERR(__wt_cursor_get_raw_key(cursor, key));
if (!WT_DATA_IN_ITEM(key))
WT_ERR(__wt_buf_set(
session, key, key->data, key->size));
}
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* If the on-page record transaction ID associated with the
* record is globally visible, the record can be discarded.
*
* Cursor opened overwrite=true: won't return WT_NOTFOUND should
* another thread remove the record before we do, and the cursor
* remains positioned in that case.
*/
if (__wt_txn_visible_all(session, las_txnid))
WT_ERR(cursor->remove(cursor));
}
/*
* When reaching the lookaside table end or the target number of calls,
* adjust the row count. Decrease/increase the row count depending on
* if the number of calls is less/more than the target.
*/
if (ret == WT_NOTFOUND ||
conn->las_sweep_call > WT_SWEEP_LOOKASIDE_PASS_TARGET) {
if (conn->las_sweep_call < WT_SWEEP_LOOKASIDE_PASS_TARGET &&
conn->las_sweep_cnt > WT_SWEEP_LOOKASIDE_MIN_CNT)
conn->las_sweep_cnt -= WT_SWEEP_LOOKASIDE_MIN_CNT;
if (conn->las_sweep_call > WT_SWEEP_LOOKASIDE_PASS_TARGET)
conn->las_sweep_cnt += WT_SWEEP_LOOKASIDE_MIN_CNT;
}
srch_notfound:
if (ret == WT_NOTFOUND)
conn->las_sweep_call = 0;
WT_ERR_NOTFOUND_OK(ret);
if (0) {
err: __wt_buf_free(session, key);
}
WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
return (ret);
}
/*
* __clsm_enter --
* Start an operation on an LSM cursor, update if the tree has changed.
*/
static inline int
__clsm_enter(WT_CURSOR_LSM *clsm, int reset, int update)
{
WT_CURSOR *c;
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
WT_SESSION_IMPL *session;
uint64_t *txnid_maxp;
uint64_t id, myid, snap_min;
session = (WT_SESSION_IMPL *)clsm->iface.session;
/* Merge cursors never update. */
if (F_ISSET(clsm, WT_CLSM_MERGE))
return (0);
if (reset) {
c = &clsm->iface;
/* Copy out data before resetting chunk cursors. */
if (F_ISSET(c, WT_CURSTD_KEY_INT) &&
!WT_DATA_IN_ITEM(&c->key))
WT_RET(__wt_buf_set(
session, &c->key, c->key.data, c->key.size));
if (F_ISSET(c, WT_CURSTD_VALUE_INT) &&
!WT_DATA_IN_ITEM(&c->value))
WT_RET(__wt_buf_set(
session, &c->value, c->value.data, c->value.size));
WT_RET(__clsm_reset_cursors(clsm, NULL));
}
for (;;) {
/*
* If the cursor looks up-to-date, check if the cache is full.
* In case this call blocks, the check will be repeated before
* proceeding.
*/
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
WT_RET(__wt_cache_full_check(session));
if (clsm->dsk_gen != clsm->lsm_tree->dsk_gen)
goto open;
/* Update the maximum transaction ID in the primary chunk. */
if (update && (chunk = clsm->primary_chunk) != NULL) {
WT_RET(__wt_txn_autocommit_check(session));
for (id = chunk->txnid_max, myid = session->txn.id;
!TXNID_LE(myid, id);
id = chunk->txnid_max) {
WT_ASSERT(session, myid != WT_TXN_NONE);
(void)WT_ATOMIC_CAS(
chunk->txnid_max, id, myid);
}
}
/*
* Figure out how many updates are required for snapshot
* isolation.
*
* This is not a normal visibility check on the maximum
* transaction ID in each chunk: any transaction ID that
* overlaps with our snapshot is a potential conflict.
*/
clsm->nupdates = 1;
if (session->txn.isolation == TXN_ISO_SNAPSHOT &&
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) {
snap_min = session->txn.snap_min;
for (txnid_maxp = &clsm->txnid_max[clsm->nchunks - 2];
clsm->nupdates < clsm->nchunks;
clsm->nupdates++, txnid_maxp--)
if (TXNID_LT(*txnid_maxp, snap_min))
break;
}
/*
* Stop when we are up-to-date, as long as this is:
* - a snapshot isolation update and the cursor is set up for
* that;
* - an update operation with a primary chunk, or
* - a read operation and the cursor is open for reading.
*/
if ((!update ||
session->txn.isolation != TXN_ISO_SNAPSHOT ||
F_ISSET(clsm, WT_CLSM_OPEN_SNAPSHOT)) &&
((update && clsm->primary_chunk != NULL) ||
(!update && F_ISSET(clsm, WT_CLSM_OPEN_READ))))
break;
open: WT_WITH_SCHEMA_LOCK(session,
ret = __clsm_open_cursors(clsm, update, 0, 0));
WT_RET(ret);
}
if (!F_ISSET(clsm, WT_CLSM_ACTIVE)) {
WT_RET(__cursor_enter(session));
F_SET(clsm, WT_CLSM_ACTIVE);
}
return (0);
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, int key_ret)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_CURSOR *cursor;
WT_IKEY *ikey;
WT_PAGE *page;
WT_ROW *rip;
WT_UPDATE *upd;
uint8_t v;
btree = session->btree;
unpack = &_unpack;
page = cbt->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
if (key_ret)
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL) {
upd = cbt->ins->upd;
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
if (key_ret)
cursor->recno = cbt->recno;
/*
* If the cursor references a WT_INSERT item, take the related
* WT_UPDATE item.
*/
if (cbt->ins != NULL) {
upd = cbt->ins->upd;
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
cell = WT_COL_PTR(page, &page->u.col_var.d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->u.row.d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take the key and
* related WT_UPDATE item. Otherwise, take the key from the
* original page, and the value from any related WT_UPDATE item,
* or the page if the key was never updated.
*/
if (cbt->ins == NULL) {
if (key_ret) {
if (__wt_off_page(page, rip->key)) {
ikey = rip->key;
cursor->key.data = WT_IKEY_DATA(ikey);
cursor->key.size = ikey->size;
} else
WT_RET(__wt_row_key(
session, page, rip, &cursor->key));
}
upd = WT_ROW_UPDATE(page, rip);
} else {
if (key_ret) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
}
upd = cbt->ins->upd;
}
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the original cell (which may be empty). */
if ((cell = __wt_row_value(page, rip)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* It's a cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, unpack);
if (btree->huffman_value == NULL && unpack->type == WT_CELL_VALUE) {
cursor->value.data = unpack->data;
cursor->value.size = unpack->size;
return (0);
} else
return (__wt_cell_unpack_copy(session, unpack, &cursor->value));
}
/*
* __wt_las_sweep --
* Sweep the lookaside table.
*/
int
__wt_las_sweep(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR *cursor;
WT_DECL_ITEM(las_addr);
WT_DECL_ITEM(las_key);
WT_DECL_RET;
WT_ITEM *key;
uint64_t cnt, las_counter, las_txnid;
int64_t remove_cnt;
uint32_t las_id, session_flags;
int notused;
conn = S2C(session);
cursor = NULL;
key = &conn->las_sweep_key;
remove_cnt = 0;
session_flags = 0; /* [-Werror=maybe-uninitialized] */
WT_ERR(__wt_scr_alloc(session, 0, &las_addr));
WT_ERR(__wt_scr_alloc(session, 0, &las_key));
WT_ERR(__wt_las_cursor(session, &cursor, &session_flags));
/*
* If we're not starting a new sweep, position the cursor using the key
* from the last call (we don't care if we're before or after the key,
* just roughly in the same spot is fine).
*/
if (key->size != 0) {
__wt_cursor_set_raw_key(cursor, key);
ret = cursor->search_near(cursor, ¬used);
/*
* Don't search for the same key twice; if we don't set a new
* key below, it's because we've reached the end of the table
* and we want the next pass to start at the beginning of the
* table. Searching for the same key could leave us stuck at
* the end of the table, repeatedly checking the same rows.
*/
key->size = 0;
if (ret != 0)
goto srch_notfound;
}
/*
* The sweep server wakes up every 10 seconds (by default), it's a slow
* moving thread. Try to review the entire lookaside table once every 5
* minutes, or every 30 calls.
*
* The reason is because the lookaside table exists because we're seeing
* cache/eviction pressure (it allows us to trade performance and disk
* space for cache space), and it's likely lookaside blocks are being
* evicted, and reading them back in doesn't help things. A trickier,
* but possibly better, alternative might be to review all lookaside
* blocks in the cache in order to get rid of them, and slowly review
* lookaside blocks that have already been evicted.
*/
cnt = (uint64_t)WT_MAX(100, conn->las_record_cnt / 30);
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
/* Walk the file. */
for (; cnt > 0 && (ret = cursor->next(cursor)) == 0; --cnt) {
/*
* If the loop terminates after completing a work unit, we will
* continue the table sweep next time. Get a local copy of the
* sweep key, we're going to reset the cursor; do so before
* calling cursor.remove, cursor.remove can discard our hazard
* pointer and the page could be evicted from underneath us.
*/
if (cnt == 1) {
WT_ERR(__wt_cursor_get_raw_key(cursor, key));
if (!WT_DATA_IN_ITEM(key))
WT_ERR(__wt_buf_set(
session, key, key->data, key->size));
}
WT_ERR(cursor->get_key(cursor,
&las_id, las_addr, &las_counter, &las_txnid, las_key));
/*
* If the on-page record transaction ID associated with the
* record is globally visible, the record can be discarded.
*
* Cursor opened overwrite=true: won't return WT_NOTFOUND should
* another thread remove the record before we do, and the cursor
* remains positioned in that case.
*/
if (__wt_txn_visible_all(session, las_txnid)) {
WT_ERR(cursor->remove(cursor));
++remove_cnt;
}
}
srch_notfound:
WT_ERR_NOTFOUND_OK(ret);
if (0) {
err: __wt_buf_free(session, key);
}
WT_TRET(__wt_las_cursor_close(session, &cursor, session_flags));
/*
* If there were races to remove records, we can over-count. All
* arithmetic is signed, so underflow isn't fatal, but check anyway so
* we don't skew low over time.
*/
if (remove_cnt > S2C(session)->las_record_cnt)
S2C(session)->las_record_cnt = 0;
else if (remove_cnt > 0)
(void)__wt_atomic_subi64(&conn->las_record_cnt, remove_cnt);
F_CLR(session, WT_SESSION_NO_CACHE);
__wt_scr_free(session, &las_addr);
__wt_scr_free(session, &las_key);
return (ret);
}
/*
* __wt_kv_return --
* Return a page referenced key/value pair to the application.
*/
int
__wt_kv_return(WT_SESSION_IMPL *session, WT_CURSOR_BTREE *cbt, WT_UPDATE *upd)
{
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK unpack;
WT_CURSOR *cursor;
WT_ITEM *tmp;
WT_PAGE *page;
WT_ROW *rip;
uint8_t v;
btree = S2BT(session);
page = cbt->ref->page;
cursor = &cbt->iface;
switch (page->type) {
case WT_PAGE_COL_FIX:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the value from the original page. */
v = __bit_getv_recno(page, cbt->iface.recno, btree->bitcnt);
return (__wt_buf_set(session, &cursor->value, &v, 1));
case WT_PAGE_COL_VAR:
/*
* The interface cursor's record has usually been set, but that
* isn't universally true, specifically, cursor.search_near may
* call here without first setting the interface cursor.
*/
cursor->recno = cbt->recno;
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Take the value from the original page cell. */
cell = WT_COL_PTR(page, &page->pg_var_d[cbt->slot]);
break;
case WT_PAGE_ROW_LEAF:
rip = &page->pg_row_d[cbt->slot];
/*
* If the cursor references a WT_INSERT item, take its key.
* Else, if we have an exact match, we copied the key in the
* search function, take it from there.
* If we don't have an exact match, take the key from the
* original page.
*/
if (cbt->ins != NULL) {
cursor->key.data = WT_INSERT_KEY(cbt->ins);
cursor->key.size = WT_INSERT_KEY_SIZE(cbt->ins);
} else if (cbt->compare == 0) {
/*
* If not in an insert list and there's an exact match,
* the row-store search function built the key we want
* to return in the cursor's temporary buffer. Swap the
* cursor's search-key and temporary buffers so we can
* return it (it's unsafe to return the temporary buffer
* itself because our caller might do another search in
* this table using the key we return, and we'd corrupt
* the search key during any subsequent search that used
* the temporary buffer.
*/
tmp = cbt->row_key;
cbt->row_key = cbt->tmp;
cbt->tmp = tmp;
cursor->key.data = cbt->row_key->data;
cursor->key.size = cbt->row_key->size;
} else
WT_RET(__wt_row_leaf_key(
session, page, rip, &cursor->key, false));
/* If the cursor references a WT_UPDATE item, return it. */
if (upd != NULL) {
cursor->value.data = WT_UPDATE_DATA(upd);
cursor->value.size = upd->size;
return (0);
}
/* Simple values have their location encoded in the WT_ROW. */
if (__wt_row_leaf_value(page, rip, &cursor->value))
return (0);
/*
* Take the value from the original page cell (which may be
* empty).
*/
if ((cell =
__wt_row_leaf_value_cell(page, rip, NULL)) == NULL) {
cursor->value.size = 0;
return (0);
}
break;
WT_ILLEGAL_VALUE(session);
}
/* The value is an on-page cell, unpack and expand it as necessary. */
__wt_cell_unpack(cell, &unpack);
WT_RET(__wt_page_cell_data_ref(session, page, &unpack, &cursor->value));
return (0);
}