/*
* __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, uint32_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t result_len;
btree = session->btree;
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->mem;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->mem;
}
/*
* 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_init(session, buf, dsk->mem_size));
buf->size = 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->mem, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(btree->compressor->decompress(
btree->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->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 (result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(
F_ISSET(session, WT_SESSION_SALVAGE_QUIET_ERR) ?
WT_ERROR :
__wt_illegal_value(session, btree->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_CSTAT_INCR(session, cache_read);
WT_DSTAT_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_DSTAT_INCR(session, compress_read);
WT_CSTAT_INCRV(session, cache_bytes_read, addr_size);
WT_DSTAT_INCRV(session, cache_bytes_read, addr_size);
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __conn_config_file --
* Read in any WiredTiger_config file in the home directory.
*/
static int
__conn_config_file(WT_SESSION_IMPL *session, const char **cfg, WT_ITEM **cbufp)
{
WT_DECL_ITEM(cbuf);
WT_DECL_RET;
WT_FH *fh;
off_t size;
uint32_t len;
int exist, quoted;
uint8_t *p, *t;
*cbufp = NULL; /* Returned buffer */
fh = NULL;
/* Check for an optional configuration file. */
#define WT_CONFIGFILE "WiredTiger.config"
WT_RET(__wt_exist(session, WT_CONFIGFILE, &exist));
if (!exist)
return (0);
/* Open the configuration file. */
WT_RET(__wt_open(session, WT_CONFIGFILE, 0, 0, 0, &fh));
WT_ERR(__wt_filesize(session, fh, &size));
if (size == 0)
goto err;
/*
* Sanity test: a 100KB configuration file would be insane. (There's
* no practical reason to limit the file size, but I can either limit
* the file size to something rational, or I can add code to test if
* the off_t size is larger than a uint32_t, which is more complicated
* and a waste of time.)
*/
if (size > 100 * 1024)
WT_ERR_MSG(session, EFBIG, WT_CONFIGFILE);
len = (uint32_t)size;
/*
* Copy the configuration file into memory, with a little slop, I'm not
* interested in debugging off-by-ones.
*
* The beginning of a file is the same as if we run into an unquoted
* newline character, simplify the parsing loop by pretending that's
* what we're doing.
*/
WT_ERR(__wt_scr_alloc(session, len + 10, &cbuf));
WT_ERR(
__wt_read(session, fh, (off_t)0, len, ((uint8_t *)cbuf->mem) + 1));
((uint8_t *)cbuf->mem)[0] = '\n';
cbuf->size = len + 1;
/*
* Collapse the file's lines into a single string: newline characters
* are replaced with commas unless the newline is quoted or backslash
* escaped. Comment lines (an unescaped newline where the next non-
* white-space character is a hash), are discarded.
*/
for (quoted = 0, p = t = cbuf->mem; len > 0;) {
/*
* Backslash pairs pass through untouched, unless immediately
* preceding a newline, in which case both the backslash and
* the newline are discarded. Backslash characters escape
* quoted characters, too, that is, a backslash followed by a
* quote doesn't start or end a quoted string.
*/
if (*p == '\\' && len > 1) {
if (p[1] != '\n') {
*t++ = p[0];
*t++ = p[1];
}
p += 2;
len -= 2;
continue;
}
/*
* If we're in a quoted string, or starting a quoted string,
* take all characters, including white-space and newlines.
*/
if (quoted || *p == '"') {
if (*p == '"')
quoted = !quoted;
*t++ = *p++;
--len;
continue;
}
/* Everything else gets taken, except for newline characters. */
if (*p != '\n') {
*t++ = *p++;
--len;
continue;
}
/*
* Replace any newline characters with commas (and strings of
* commas are safe).
*
* After any newline, skip to a non-white-space character; if
* the next character is a hash mark, skip to the next newline.
*/
for (;;) {
for (*t++ = ','; --len > 0 && isspace(*++p);)
;
if (len == 0)
break;
if (*p != '#')
break;
while (--len > 0 && *++p != '\n')
;
if (len == 0)
break;
}
}
*t = '\0';
#if 0
fprintf(stderr, "file config: {%s}\n", (const char *)cbuf->data);
#endif
/* Check the configuration string. */
WT_ERR(__wt_config_check(
session, __wt_confchk_wiredtiger_open, cbuf->data, 0));
/*
* The configuration file falls between the default configuration and
* the wiredtiger_open() configuration, overriding the defaults but not
* overriding the wiredtiger_open() configuration.
*/
while (cfg[1] != NULL)
++cfg;
cfg[1] = cfg[0];
cfg[0] = cbuf->data;
*cbufp = cbuf;
if (0) {
err: if (cbuf != NULL)
__wt_buf_free(session, cbuf);
}
if (fh != NULL)
WT_TRET(__wt_close(session, fh));
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);
}
/*重命名一个索引或者colgroup*/
static int __rename_tree(WT_SESSION_IMPL* session, WT_TABLE* table, const char *newuri, const char *name, const char *cfg[])
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(nn);
WT_DECL_ITEM(ns);
WT_DECL_ITEM(nv);
WT_DECL_ITEM(os);
WT_DECL_RET;
const char *newname, *olduri, *suffix;
char *value;
int is_colgroup;
olduri = table->name;
value = NULL;
newname = newuri;
(void)WT_PREFIX_SKIP(newname, "table:");
/*判断是重命名一个colgroup*/
is_colgroup = WT_PREFIX_MATCH(name, "colgroup:");
if (!is_colgroup && !WT_PREFIX_MATCH(name, "index:"))
WT_ERR_MSG(session, EINVAL, "expected a 'colgroup:' or 'index:' source: '%s'", name);
suffix = strchr(name, ':');
/* An existing table should have a well formed name. */
WT_ASSERT(session, suffix != NULL);
suffix = strchr(suffix + 1, ':');
WT_ERR(__wt_scr_alloc(session, 0, &nn));
WT_ERR(__wt_buf_fmt(session, nn, "%s%s%s", is_colgroup ? "colgroup:" : "index:", newname, (suffix == NULL) ? "" : suffix));
/* Skip the colon, if any. */
if (suffix != NULL)
++suffix;
WT_ERR(__wt_metadata_search(session, name, &value));
/*
* Calculate the new data source URI. Use the existing table structure
* and substitute the new name temporarily.
*/
WT_ERR(__wt_scr_alloc(session, 0, &ns));
table->name = newuri;
if (is_colgroup)
WT_ERR(__wt_schema_colgroup_source(session, table, suffix, value, ns));
else
WT_ERR(__wt_schema_index_source(session, table, suffix, value, ns));
if ((ret = __wt_config_getones(session, value, "source", &cval)) != 0)
WT_ERR_MSG(session, EINVAL, "index or column group has no data source: %s", value);
/* Take a copy of the old data source. */
WT_ERR(__wt_scr_alloc(session, 0, &os));
WT_ERR(__wt_buf_fmt(session, os, "%.*s", (int)cval.len, cval.str));
/* Overwrite it with the new data source. */
WT_ERR(__wt_scr_alloc(session, 0, &nv));
WT_ERR(__wt_buf_fmt(session, nv, "%.*s%s%s", (int)WT_PTRDIFF(cval.str, value), value, (const char *)ns->data, cval.str + cval.len));
/*
* Remove the old metadata entry.
* Insert the new metadata entry.
*/
WT_ERR(__wt_metadata_remove(session, name));
WT_ERR(__wt_metadata_insert(session, nn->data, nv->data));
/* Rename the file. */
WT_ERR(__wt_schema_rename(session, os->data, ns->data, cfg));
err:
__wt_scr_free(session, &nn);
__wt_scr_free(session, &ns);
__wt_scr_free(session, &nv);
__wt_scr_free(session, &os);
__wt_free(session, value);
table->name = olduri;
return ret;
}
/*
* __wt_bt_write --
* Write a buffer into a block, returning the block's addr/size and
* checksum.
*/
int
__wt_bt_write(WT_SESSION_IMPL *session, WT_ITEM *buf,
uint8_t *addr, size_t *addr_sizep, int checkpoint, int compressed)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(ctmp);
WT_DECL_ITEM(etmp);
WT_DECL_RET;
WT_KEYED_ENCRYPTOR *kencryptor;
WT_ITEM *ip;
WT_PAGE_HEADER *dsk;
size_t dst_len, len, result_len, size, src_len;
int compression_failed, data_cksum, encrypted;
uint8_t *dst, *src;
btree = S2BT(session);
bm = btree->bm;
encrypted = 0;
/* Checkpoint calls are different than standard calls. */
WT_ASSERT(session,
(checkpoint == 0 && addr != NULL && addr_sizep != NULL) ||
(checkpoint == 1 && addr == NULL && addr_sizep == NULL));
#ifdef HAVE_DIAGNOSTIC
/*
* We're passed a table's disk image. Decompress if necessary and
* verify the image. Always check the in-memory length for accuracy.
*/
dsk = buf->mem;
if (compressed) {
WT_ERR(__wt_scr_alloc(session, dsk->mem_size, &ctmp));
memcpy(ctmp->mem, buf->data, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)buf->data + WT_BLOCK_COMPRESS_SKIP,
buf->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)ctmp->data + WT_BLOCK_COMPRESS_SKIP,
ctmp->memsize - WT_BLOCK_COMPRESS_SKIP,
&result_len));
WT_ASSERT(session,
dsk->mem_size == result_len + WT_BLOCK_COMPRESS_SKIP);
ctmp->size = (uint32_t)result_len + WT_BLOCK_COMPRESS_SKIP;
ip = ctmp;
} else {
WT_ASSERT(session, dsk->mem_size == buf->size);
ip = buf;
}
WT_ERR(__wt_verify_dsk(session, "[write-check]", ip));
__wt_scr_free(session, &ctmp);
#endif
/*
* Optionally stream-compress the data, but don't compress blocks that
* are already as small as they're going to get.
*/
if (btree->compressor == NULL ||
btree->compressor->compress == NULL || compressed)
ip = buf;
else if (buf->size <= btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_too_small);
} else {
/* Skip the header bytes of the source data. */
src = (uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP;
src_len = buf->size - WT_BLOCK_COMPRESS_SKIP;
/*
* Compute the size needed for the destination buffer. We only
* allocate enough memory for a copy of the original by default,
* if any compressed version is bigger than the original, we
* won't use it. However, some compression engines (snappy is
* one example), may need more memory because they don't stop
* just because there's no more memory into which to compress.
*/
if (btree->compressor->pre_size == NULL)
len = src_len;
else
WT_ERR(btree->compressor->pre_size(btree->compressor,
&session->iface, src, src_len, &len));
size = len + WT_BLOCK_COMPRESS_SKIP;
WT_ERR(bm->write_size(bm, session, &size));
WT_ERR(__wt_scr_alloc(session, size, &ctmp));
/* Skip the header bytes of the destination data. */
dst = (uint8_t *)ctmp->mem + WT_BLOCK_COMPRESS_SKIP;
dst_len = len;
compression_failed = 0;
WT_ERR(btree->compressor->compress(btree->compressor,
&session->iface,
src, src_len,
dst, dst_len,
&result_len, &compression_failed));
result_len += WT_BLOCK_COMPRESS_SKIP;
/*
* If compression fails, or doesn't gain us at least one unit of
* allocation, fallback to the original version. This isn't
* unexpected: if compression doesn't work for some chunk of
* data for some reason (noting likely additional format/header
* information which compressed output requires), it just means
* the uncompressed version is as good as it gets, and that's
* what we use.
*/
if (compression_failed ||
buf->size / btree->allocsize <=
result_len / btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_fail);
} else {
compressed = 1;
WT_STAT_FAST_DATA_INCR(session, compress_write);
/*
* Copy in the skipped header bytes, set the final data
* size.
*/
memcpy(ctmp->mem, buf->mem, WT_BLOCK_COMPRESS_SKIP);
ctmp->size = result_len;
ip = ctmp;
}
}
/*
* Optionally encrypt the data. We need to add in the original
* length, in case both compression and encryption are done.
*/
if ((kencryptor = btree->kencryptor) != NULL) {
/*
* Get size needed for encrypted buffer.
*/
__wt_encrypt_size(session, kencryptor, ip->size, &size);
WT_ERR(bm->write_size(bm, session, &size));
WT_ERR(__wt_scr_alloc(session, size, &etmp));
WT_ERR(__wt_encrypt(session,
kencryptor, WT_BLOCK_ENCRYPT_SKIP, ip, etmp));
encrypted = 1;
ip = etmp;
}
dsk = ip->mem;
/* If the buffer is compressed, set the flag. */
if (compressed)
F_SET(dsk, WT_PAGE_COMPRESSED);
if (encrypted)
F_SET(dsk, WT_PAGE_ENCRYPTED);
/*
* We increment the block's write generation so it's easy to identify
* newer versions of blocks during salvage. (It's common in WiredTiger,
* at least for the default block manager, for multiple blocks to be
* internally consistent with identical first and last keys, so we need
* a way to know the most recent state of the block. We could check
* which leaf is referenced by a valid internal page, but that implies
* salvaging internal pages, which I don't want to do, and it's not
* as good anyway, because the internal page may not have been written
* after the leaf page was updated. So, write generations it is.
*
* Nothing is locked at this point but two versions of a page with the
* same generation is pretty unlikely, and if we did, they're going to
* be roughly identical for the purposes of salvage, anyway.
*/
dsk->write_gen = ++btree->write_gen;
/*
* Checksum the data if the buffer isn't compressed or checksums are
* configured.
*/
switch (btree->checksum) {
case CKSUM_ON:
data_cksum = 1;
break;
case CKSUM_OFF:
data_cksum = 0;
break;
case CKSUM_UNCOMPRESSED:
default:
data_cksum = !compressed;
break;
}
/* Call the block manager to write the block. */
WT_ERR(checkpoint ?
bm->checkpoint(bm, session, ip, btree->ckpt, data_cksum) :
bm->write(bm, session, ip, addr, addr_sizep, data_cksum));
WT_STAT_FAST_CONN_INCR(session, cache_write);
WT_STAT_FAST_DATA_INCR(session, cache_write);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_write, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_write, dsk->mem_size);
err: __wt_scr_free(session, &ctmp);
__wt_scr_free(session, &etmp);
return (ret);
}
/*
* __wt_verify --
* Verify a file.
*/
int
__wt_verify(WT_SESSION_IMPL *session, const char *cfg[])
{
WT_BM *bm;
WT_BTREE *btree;
WT_CKPT *ckptbase, *ckpt;
WT_DECL_RET;
WT_VSTUFF *vs, _vstuff;
size_t root_addr_size;
uint8_t root_addr[WT_BTREE_MAX_ADDR_COOKIE];
bool bm_start, quit;
btree = S2BT(session);
bm = btree->bm;
ckptbase = NULL;
bm_start = false;
WT_CLEAR(_vstuff);
vs = &_vstuff;
WT_ERR(__wt_scr_alloc(session, 0, &vs->max_key));
WT_ERR(__wt_scr_alloc(session, 0, &vs->max_addr));
WT_ERR(__wt_scr_alloc(session, 0, &vs->tmp1));
WT_ERR(__wt_scr_alloc(session, 0, &vs->tmp2));
WT_ERR(__wt_scr_alloc(session, 0, &vs->tmp3));
WT_ERR(__wt_scr_alloc(session, 0, &vs->tmp4));
/* Check configuration strings. */
WT_ERR(__verify_config(session, cfg, vs));
/* Optionally dump specific block offsets. */
WT_ERR(__verify_config_offsets(session, cfg, &quit));
if (quit)
goto done;
/* Get a list of the checkpoints for this file. */
WT_ERR(
__wt_meta_ckptlist_get(session, btree->dhandle->name, &ckptbase));
/* Inform the underlying block manager we're verifying. */
WT_ERR(bm->verify_start(bm, session, ckptbase, cfg));
bm_start = true;
/* Loop through the file's checkpoints, verifying each one. */
WT_CKPT_FOREACH(ckptbase, ckpt) {
WT_ERR(__wt_verbose(session, WT_VERB_VERIFY,
"%s: checkpoint %s", btree->dhandle->name, ckpt->name));
/* Fake checkpoints require no work. */
if (F_ISSET(ckpt, WT_CKPT_FAKE))
continue;
/* House-keeping between checkpoints. */
__verify_checkpoint_reset(vs);
if (WT_VRFY_DUMP(vs))
WT_ERR(__wt_msg(session, "%s: checkpoint %s",
btree->dhandle->name, ckpt->name));
/* Load the checkpoint. */
WT_ERR(bm->checkpoint_load(bm, session,
ckpt->raw.data, ckpt->raw.size,
root_addr, &root_addr_size, true));
/*
* Ignore trees with no root page.
* Verify, then discard the checkpoint from the cache.
*/
if (root_addr_size != 0 &&
(ret = __wt_btree_tree_open(
session, root_addr, root_addr_size)) == 0) {
if (WT_VRFY_DUMP(vs))
WT_ERR(__wt_msg(session, "Root: %s %s",
__wt_addr_string(session,
root_addr, root_addr_size, vs->tmp1),
__wt_page_type_string(
btree->root.page->type)));
WT_WITH_PAGE_INDEX(session,
ret = __verify_tree(session, &btree->root, vs));
WT_TRET(__wt_cache_op(session, WT_SYNC_DISCARD));
}
/* Unload the checkpoint. */
WT_TRET(bm->checkpoint_unload(bm, session));
WT_ERR(ret);
/* Display the tree shape. */
if (vs->dump_shape)
WT_ERR(__verify_tree_shape(session, vs));
}
/*
* __wt_txn_checkpoint_log --
* Write a log record for a checkpoint operation.
*/
int
__wt_txn_checkpoint_log(
WT_SESSION_IMPL *session, int full, uint32_t flags, WT_LSN *lsnp)
{
WT_DECL_RET;
WT_DECL_ITEM(logrec);
WT_LSN *ckpt_lsn;
WT_TXN *txn;
const char *fmt = WT_UNCHECKED_STRING(IIQIU);
uint8_t *end, *p;
size_t recsize;
uint32_t i, rectype = WT_LOGREC_CHECKPOINT;
txn = &session->txn;
ckpt_lsn = &txn->ckpt_lsn;
/*
* If this is a file sync, log it unless there is a full checkpoint in
* progress.
*/
if (!full) {
if (txn->full_ckpt) {
if (lsnp != NULL)
*lsnp = *ckpt_lsn;
return (0);
} else
return (__txn_log_file_sync(session, flags, lsnp));
}
switch (flags) {
case WT_TXN_LOG_CKPT_PREPARE:
txn->full_ckpt = 1;
*ckpt_lsn = S2C(session)->log->alloc_lsn;
break;
case WT_TXN_LOG_CKPT_START:
/* Take a copy of the transaction snapshot. */
txn->ckpt_nsnapshot = txn->snapshot_count;
recsize = txn->ckpt_nsnapshot * WT_INTPACK64_MAXSIZE;
WT_ERR(__wt_scr_alloc(session, recsize, &txn->ckpt_snapshot));
p = txn->ckpt_snapshot->mem;
end = p + recsize;
for (i = 0; i < txn->snapshot_count; i++)
WT_ERR(__wt_vpack_uint(
&p, WT_PTRDIFF(end, p), txn->snapshot[i]));
break;
case WT_TXN_LOG_CKPT_STOP:
/*
* During a clean connection close, we get here without the
* prepare or start steps. In that case, log the current LSN
* as the checkpoint LSN.
*/
if (!txn->full_ckpt) {
txn->ckpt_nsnapshot = 0;
*ckpt_lsn = S2C(session)->log->alloc_lsn;
}
/* Write the checkpoint log record. */
WT_ERR(__wt_struct_size(session, &recsize, fmt,
rectype, ckpt_lsn->file, ckpt_lsn->offset,
txn->ckpt_nsnapshot, &txn->ckpt_snapshot));
WT_ERR(__wt_logrec_alloc(session, recsize, &logrec));
WT_ERR(__wt_struct_pack(session,
(uint8_t *)logrec->data + logrec->size, recsize, fmt,
rectype, ckpt_lsn->file, ckpt_lsn->offset,
txn->ckpt_nsnapshot, &txn->ckpt_snapshot));
logrec->size += (uint32_t)recsize;
WT_ERR(__wt_log_write(session, logrec, lsnp, 0));
/*
* If this full checkpoint completed successfully and there is
* no hot backup in progress, tell the logging subsystem the
* checkpoint LSN so that it can archive.
*/
if (!S2C(session)->hot_backup)
WT_ERR(__wt_log_ckpt(session, ckpt_lsn));
/* FALLTHROUGH */
case WT_TXN_LOG_CKPT_FAIL:
/* Cleanup any allocated resources */
INIT_LSN(ckpt_lsn);
txn->ckpt_nsnapshot = 0;
__wt_scr_free(&txn->ckpt_snapshot);
txn->full_ckpt = 0;
break;
}
err: __wt_logrec_free(session, &logrec);
return (ret);
}
/*
* __wt_lsm_tree_create --
* Create an LSM tree structure for the given name.
*/
int
__wt_lsm_tree_create(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_CONFIG_ITEM cval;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_TREE *lsm_tree;
const char *cfg[] =
{ WT_CONFIG_BASE(session, session_create), config, NULL };
char *tmpconfig;
/* If the tree is open, it already exists. */
if ((ret = __wt_lsm_tree_get(session, uri, 0, &lsm_tree)) == 0) {
__wt_lsm_tree_release(session, lsm_tree);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
/*
* If the tree has metadata, it already exists.
*
* !!!
* Use a local variable: we don't care what the existing configuration
* is, but we don't want to overwrite the real config.
*/
if (__wt_metadata_search(session, uri, &tmpconfig) == 0) {
__wt_free(session, tmpconfig);
return (exclusive ? EEXIST : 0);
}
WT_RET_NOTFOUND_OK(ret);
WT_RET(__wt_config_gets(session, cfg, "key_format", &cval));
if (WT_STRING_MATCH("r", cval.str, cval.len))
WT_RET_MSG(session, EINVAL,
"LSM trees cannot be configured as column stores");
WT_RET(__wt_calloc_def(session, 1, &lsm_tree));
WT_ERR(__lsm_tree_set_name(session, lsm_tree, uri));
WT_ERR(__wt_config_gets(session, cfg, "key_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->key_format));
WT_ERR(__wt_config_gets(session, cfg, "value_format", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->value_format));
WT_ERR(__wt_config_gets(session, cfg, "collator", &cval));
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->collator_name));
WT_ERR(__wt_config_gets(session, cfg, "lsm.auto_throttle", &cval));
if (cval.val)
F_SET(lsm_tree, WT_LSM_TREE_THROTTLE);
else
F_CLR(lsm_tree, WT_LSM_TREE_THROTTLE);
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom", &cval));
FLD_SET(lsm_tree->bloom,
(cval.val == 0 ? WT_LSM_BLOOM_OFF : WT_LSM_BLOOM_MERGED));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_oldest", &cval));
if (cval.val != 0)
FLD_SET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST);
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OFF) &&
FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST))
WT_ERR_MSG(session, EINVAL,
"Bloom filters can only be created on newest and oldest "
"chunks if bloom filters are enabled");
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_config", &cval));
if (cval.type == WT_CONFIG_ITEM_STRUCT) {
cval.str++;
cval.len -= 2;
}
WT_ERR(__wt_strndup(
session, cval.str, cval.len, &lsm_tree->bloom_config));
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_bit_count", &cval));
lsm_tree->bloom_bit_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.bloom_hash_count", &cval));
lsm_tree->bloom_hash_count = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_max", &cval));
lsm_tree->chunk_max = (uint64_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.chunk_size", &cval));
lsm_tree->chunk_size = (uint64_t)cval.val;
if (lsm_tree->chunk_size > lsm_tree->chunk_max)
WT_ERR_MSG(session, EINVAL,
"Chunk size (chunk_size) must be smaller than or equal to "
"the maximum chunk size (chunk_max)");
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_max", &cval));
lsm_tree->merge_max = (uint32_t)cval.val;
WT_ERR(__wt_config_gets(session, cfg, "lsm.merge_min", &cval));
lsm_tree->merge_min = (uint32_t)cval.val;
if (lsm_tree->merge_min > lsm_tree->merge_max)
WT_ERR_MSG(session, EINVAL,
"LSM merge_min must be less than or equal to merge_max");
/*
* Set up the config for each chunk.
*
* Make the memory_page_max double the chunk size, so application
* threads don't immediately try to force evict the chunk when the
* worker thread clears the NO_EVICTION flag.
*/
WT_ERR(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"%s,key_format=u,value_format=u,memory_page_max=%" PRIu64,
config, 2 * lsm_tree->chunk_max));
WT_ERR(__wt_strndup(
session, buf->data, buf->size, &lsm_tree->file_config));
/* Create the first chunk and flush the metadata. */
WT_ERR(__wt_lsm_meta_write(session, lsm_tree));
/* Discard our partially populated handle. */
ret = __lsm_tree_discard(session, lsm_tree);
lsm_tree = NULL;
/*
* Open our new tree and add it to the handle cache. Don't discard on
* error: the returned handle is NULL on error, and the metadata
* tracking macros handle cleaning up on failure.
*/
if (ret == 0)
ret = __lsm_tree_open(session, uri, &lsm_tree);
if (ret == 0)
__wt_lsm_tree_release(session, lsm_tree);
if (0) {
err: WT_TRET(__lsm_tree_discard(session, lsm_tree));
}
__wt_scr_free(&buf);
return (ret);
}
/*
* __wt_curindex_open --
* WT_SESSION->open_cursor method for index cursors.
*/
int
__wt_curindex_open(WT_SESSION_IMPL *session,
const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__curindex_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__curindex_set_value, /* set-value */
__curindex_compare, /* compare */
__wt_cursor_equals, /* equals */
__curindex_next, /* next */
__curindex_prev, /* prev */
__curindex_reset, /* reset */
__curindex_search, /* search */
__curindex_search_near, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curindex_close); /* close */
WT_CURSOR_INDEX *cindex;
WT_CURSOR *cursor;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_INDEX *idx;
WT_TABLE *table;
const char *columns, *idxname, *tablename;
size_t namesize;
tablename = uri;
if (!WT_PREFIX_SKIP(tablename, "index:") ||
(idxname = strchr(tablename, ':')) == NULL)
WT_RET_MSG(session, EINVAL, "Invalid cursor URI: '%s'", uri);
namesize = (size_t)(idxname - tablename);
++idxname;
if ((ret = __wt_schema_get_table(session,
tablename, namesize, false, &table)) != 0) {
if (ret == WT_NOTFOUND)
WT_RET_MSG(session, EINVAL,
"Cannot open cursor '%s' on unknown table", uri);
return (ret);
}
columns = strchr(idxname, '(');
if (columns == NULL)
namesize = strlen(idxname);
else
namesize = (size_t)(columns - idxname);
if ((ret = __wt_schema_open_index(
session, table, idxname, namesize, &idx)) != 0) {
__wt_schema_release_table(session, table);
return (ret);
}
WT_RET(__wt_calloc_one(session, &cindex));
cursor = &cindex->iface;
*cursor = iface;
cursor->session = &session->iface;
cindex->table = table;
cindex->index = idx;
cindex->key_plan = idx->key_plan;
cindex->value_plan = idx->value_plan;
cursor->internal_uri = idx->name;
cursor->key_format = idx->idxkey_format;
cursor->value_format = table->value_format;
/*
* XXX
* A very odd corner case is an index with a recno key.
* The only way to get here is by creating an index on a column store
* using only the primary's recno as the index key. Disallow that for
* now.
*/
if (WT_CURSOR_RECNO(cursor))
WT_ERR_MSG(session, WT_ERROR,
"Column store indexes based on a record number primary "
"key are not supported");
/* Handle projections. */
if (columns != NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__wt_struct_reformat(session, table,
columns, strlen(columns), NULL, false, tmp));
WT_ERR(__wt_strndup(
session, tmp->data, tmp->size, &cursor->value_format));
WT_ERR(__wt_buf_init(session, tmp, 0));
WT_ERR(__wt_struct_plan(session, table,
columns, strlen(columns), false, tmp));
WT_ERR(__wt_strndup(
session, tmp->data, tmp->size, &cindex->value_plan));
}
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_ERR(__wt_open_cursor(
session, idx->source, cursor, cfg, &cindex->child));
/* Open the column groups needed for this index cursor. */
WT_ERR(__curindex_open_colgroups(session, cindex, cfg));
if (F_ISSET(cursor, WT_CURSTD_DUMP_JSON))
__wt_json_column_init(
cursor, table->key_format, &idx->colconf, &table->colconf);
if (0) {
err: WT_TRET(__curindex_close(cursor));
*cursorp = NULL;
}
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_schema_open_indices --
* Open the indices for a table.
*/
int
__wt_schema_open_index(WT_SESSION_IMPL *session,
WT_TABLE *table, const char *idxname, size_t len, WT_INDEX **indexp)
{
WT_CURSOR *cursor;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_INDEX *idx;
u_int i;
int cmp, match;
const char *idxconf, *name, *tablename, *uri;
/* Check if we've already done the work. */
if (idxname == NULL && table->idx_complete)
return (0);
cursor = NULL;
idx = NULL;
/* Build a search key. */
tablename = table->name;
(void)WT_PREFIX_SKIP(tablename, "table:");
WT_ERR(__wt_scr_alloc(session, 512, &tmp));
WT_ERR(__wt_buf_fmt(session, tmp, "index:%s:", tablename));
/* Find matching indices. */
WT_ERR(__wt_metadata_cursor(session, NULL, &cursor));
cursor->set_key(cursor, tmp->data);
if ((ret = cursor->search_near(cursor, &cmp)) == 0 && cmp < 0)
ret = cursor->next(cursor);
for (i = 0; ret == 0; i++, ret = cursor->next(cursor)) {
WT_ERR(cursor->get_key(cursor, &uri));
name = uri;
if (!WT_PREFIX_SKIP(name, tmp->data))
break;
/* Is this the index we are looking for? */
match = idxname == NULL || WT_STRING_MATCH(name, idxname, len);
/*
* Ensure there is space, including if we have to make room for
* a new entry in the middle of the list.
*/
WT_ERR(__wt_realloc_def(session, &table->idx_alloc,
WT_MAX(i, table->nindices) + 1, &table->indices));
/* Keep the in-memory list in sync with the metadata. */
cmp = 0;
while (table->indices[i] != NULL &&
(cmp = strcmp(uri, table->indices[i]->name)) > 0) {
/* Index no longer exists, remove it. */
__wt_free(session, table->indices[i]);
memmove(&table->indices[i], &table->indices[i + 1],
(table->nindices - i) * sizeof(WT_INDEX *));
table->indices[--table->nindices] = NULL;
}
if (cmp < 0) {
/* Make room for a new index. */
memmove(&table->indices[i + 1], &table->indices[i],
(table->nindices - i) * sizeof(WT_INDEX *));
table->indices[i] = NULL;
++table->nindices;
}
if (!match)
continue;
if (table->indices[i] == NULL) {
WT_ERR(cursor->get_value(cursor, &idxconf));
WT_ERR(__wt_calloc_def(session, 1, &idx));
WT_ERR(__wt_strdup(session, uri, &idx->name));
WT_ERR(__wt_strdup(session, idxconf, &idx->config));
WT_ERR(__open_index(session, table, idx));
table->indices[i] = idx;
idx = NULL;
}
/* If we were looking for a single index, we're done. */
if (indexp != NULL)
*indexp = table->indices[i];
if (idxname != NULL)
break;
}
WT_ERR_NOTFOUND_OK(ret);
/* If we did a full pass, we won't need to do it again. */
if (idxname == NULL) {
table->nindices = i;
table->idx_complete = 1;
}
err: __wt_scr_free(&tmp);
if (idx != NULL)
__wt_schema_destroy_index(session, idx);
if (cursor != NULL)
WT_TRET(cursor->close(cursor));
return (ret);
}
/*
* __wt_schema_open_colgroups --
* Open the column groups for a table.
*/
int
__wt_schema_open_colgroups(WT_SESSION_IMPL *session, WT_TABLE *table)
{
WT_COLGROUP *colgroup;
WT_CONFIG cparser;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_RET;
WT_DECL_ITEM(buf);
const char *cgconfig;
u_int i;
if (table->cg_complete)
return (0);
WT_RET(__wt_scr_alloc(session, 0, &buf));
colgroup = NULL;
WT_RET(__wt_config_subinit(session, &cparser, &table->cgconf));
/* Open each column group. */
for (i = 0; i < WT_COLGROUPS(table); i++) {
if (table->ncolgroups > 0)
WT_ERR(__wt_config_next(&cparser, &ckey, &cval));
else
WT_CLEAR(ckey);
/*
* Always open from scratch: we may have failed part of the way
* through opening a table, or column groups may have changed.
*/
if (table->cgroups[i] != NULL) {
__wt_schema_destroy_colgroup(
session, table->cgroups[i]);
table->cgroups[i] = NULL;
}
WT_ERR(__wt_schema_colgroup_name(session, table,
ckey.str, ckey.len, buf));
if ((ret = __wt_metadata_search(
session, buf->data, &cgconfig)) != 0) {
/* It is okay if the table is incomplete. */
if (ret == WT_NOTFOUND)
ret = 0;
goto err;
}
WT_ERR(__wt_calloc_def(session, 1, &colgroup));
colgroup->name = __wt_buf_steal(session, buf, NULL);
colgroup->config = cgconfig;
WT_ERR(__wt_config_getones(session,
colgroup->config, "columns", &colgroup->colconf));
WT_ERR(__wt_config_getones(
session, colgroup->config, "source", &cval));
WT_ERR(__wt_buf_fmt(session, buf, "%.*s",
(int)cval.len, cval.str));
colgroup->source = __wt_buf_steal(session, buf, NULL);
table->cgroups[i] = colgroup;
colgroup = NULL;
}
if (!table->is_simple) {
WT_ERR(__wt_table_check(session, table));
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, 1, buf));
table->plan = __wt_buf_steal(session, buf, NULL);
}
table->cg_complete = 1;
err: __wt_scr_free(&buf);
if (colgroup != NULL)
__wt_schema_destroy_colgroup(session, colgroup);
return (ret);
}
/*
* ___open_index --
* Open an index.
*/
static int
__open_index(WT_SESSION_IMPL *session, WT_TABLE *table, WT_INDEX *idx)
{
WT_CONFIG colconf;
WT_CONFIG_ITEM ckey, cval;
WT_DECL_ITEM(buf);
WT_DECL_ITEM(plan);
WT_DECL_RET;
u_int cursor_key_cols, i;
WT_ERR(__wt_scr_alloc(session, 0, &buf));
/* Get the data source from the index config. */
WT_ERR(__wt_config_getones(session, idx->config, "source", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->source = __wt_buf_steal(session, buf, NULL);
idx->need_value = WT_PREFIX_MATCH(idx->source, "lsm:");
WT_ERR(__wt_config_getones(session, idx->config, "key_format", &cval));
WT_ERR(__wt_buf_fmt(
session, buf, "%.*s", (int)cval.len, cval.str));
idx->key_format = __wt_buf_steal(session, buf, NULL);
/*
* The key format for an index is somewhat subtle: the application
* specifies a set of columns that it will use for the key, but the
* engine usually adds some hidden columns in order to derive the
* primary key. These hidden columns are part of the file's key.
*
* The file's key_format is stored persistently, we need to calculate
* the index cursor key format (which will usually omit some of those
* keys).
*/
WT_ERR(__wt_config_getones(
session, idx->config, "columns", &idx->colconf));
/* Start with the declared index columns. */
WT_ERR(__wt_config_subinit(session, &colconf, &idx->colconf));
cursor_key_cols = 0;
while ((ret = __wt_config_next(&colconf, &ckey, &cval)) == 0) {
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
++cursor_key_cols;
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
/*
* Now add any primary key columns from the table that are not
* already part of the index key.
*/
WT_ERR(__wt_config_subinit(session, &colconf, &table->colconf));
for (i = 0; i < table->nkey_columns &&
(ret = __wt_config_next(&colconf, &ckey, &cval)) == 0;
i++) {
/*
* If the primary key column is already in the secondary key,
* don't add it again.
*/
if (__wt_config_subgetraw(
session, &idx->colconf, &ckey, &cval) == 0)
continue;
WT_ERR(__wt_buf_catfmt(
session, buf, "%.*s,", (int)ckey.len, ckey.str));
}
if (ret != 0 && ret != WT_NOTFOUND)
goto err;
WT_ERR(__wt_scr_alloc(session, 0, &plan));
WT_ERR(__wt_struct_plan(session, table, buf->data, buf->size, 0, plan));
idx->key_plan = __wt_buf_steal(session, plan, NULL);
/* Set up the cursor key format (the visible columns). */
WT_ERR(__wt_buf_init(session, buf, 0));
WT_ERR(__wt_struct_truncate(session,
idx->key_format, cursor_key_cols, buf));
idx->idxkey_format = __wt_buf_steal(session, buf, NULL);
/* By default, index cursor values are the table value columns. */
/* TODO Optimize to use index columns in preference to table lookups. */
WT_ERR(__wt_struct_plan(session,
table, table->colconf.str, table->colconf.len, 1, plan));
idx->value_plan = __wt_buf_steal(session, plan, NULL);
err: __wt_scr_free(&buf);
__wt_scr_free(&plan);
return (ret);
}
/*
* __verify_dsk_row --
* Walk a WT_PAGE_ROW_INT or WT_PAGE_ROW_LEAF disk page and verify it.
*/
static int
__verify_dsk_row(
WT_SESSION_IMPL *session, const char *addr, const WT_PAGE_HEADER *dsk)
{
WT_BM *bm;
WT_BTREE *btree;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_DECL_ITEM(current);
WT_DECL_ITEM(last_ovfl);
WT_DECL_ITEM(last_pfx);
WT_DECL_RET;
WT_ITEM *last;
enum { FIRST, WAS_KEY, WAS_VALUE } last_cell_type;
void *huffman;
uint32_t cell_num, cell_type, i, key_cnt, prefix;
uint8_t *end;
int cmp;
btree = S2BT(session);
bm = btree->bm;
unpack = &_unpack;
huffman = dsk->type == WT_PAGE_ROW_INT ? NULL : btree->huffman_key;
WT_ERR(__wt_scr_alloc(session, 0, ¤t));
WT_ERR(__wt_scr_alloc(session, 0, &last_pfx));
WT_ERR(__wt_scr_alloc(session, 0, &last_ovfl));
last = last_ovfl;
end = (uint8_t *)dsk + dsk->mem_size;
last_cell_type = FIRST;
cell_num = 0;
key_cnt = 0;
WT_CELL_FOREACH(btree, dsk, cell, unpack, i) {
++cell_num;
/* Carefully unpack the cell. */
if (__wt_cell_unpack_safe(cell, unpack, end) != 0) {
ret = __err_cell_corrupted(session, cell_num, addr);
goto err;
}
/* Check the raw and collapsed cell types. */
WT_ERR(__err_cell_type(
session, cell_num, addr, unpack->raw, dsk->type));
WT_ERR(__err_cell_type(
session, cell_num, addr, unpack->type, dsk->type));
cell_type = unpack->type;
/*
* Check ordering relationships between the WT_CELL entries.
* For row-store internal pages, check for:
* two values in a row,
* two keys in a row,
* a value as the first cell on a page.
* For row-store leaf pages, check for:
* two values in a row,
* a value as the first cell on a page.
*/
switch (cell_type) {
case WT_CELL_KEY:
case WT_CELL_KEY_OVFL:
++key_cnt;
switch (last_cell_type) {
case FIRST:
case WAS_VALUE:
break;
case WAS_KEY:
if (dsk->type == WT_PAGE_ROW_LEAF)
break;
WT_ERR_VRFY(session,
"cell %" PRIu32 " on page at %s is the "
"first of two adjacent keys",
cell_num - 1, addr);
}
last_cell_type = WAS_KEY;
break;
case WT_CELL_ADDR_DEL:
case WT_CELL_ADDR_INT:
case WT_CELL_ADDR_LEAF:
case WT_CELL_ADDR_LEAF_NO:
case WT_CELL_VALUE:
case WT_CELL_VALUE_OVFL:
switch (last_cell_type) {
case FIRST:
WT_ERR_VRFY(session,
"page at %s begins with a value", addr);
case WAS_KEY:
break;
case WAS_VALUE:
WT_ERR_VRFY(session,
"cell %" PRIu32 " on page at %s is the "
"first of two adjacent values",
cell_num - 1, addr);
}
last_cell_type = WAS_VALUE;
break;
}
/* Check if any referenced item has a valid address. */
switch (cell_type) {
case WT_CELL_ADDR_DEL:
case WT_CELL_ADDR_INT:
case WT_CELL_ADDR_LEAF:
case WT_CELL_ADDR_LEAF_NO:
case WT_CELL_KEY_OVFL:
case WT_CELL_VALUE_OVFL:
if (!bm->addr_valid(bm,
session, unpack->data, unpack->size))
goto eof;
break;
}
/*
* Remaining checks are for key order and prefix compression.
* If this cell isn't a key, we're done, move to the next cell.
* If this cell is an overflow item, instantiate the key and
* compare it with the last key. Otherwise, we have to deal with
* prefix compression.
*/
switch (cell_type) {
case WT_CELL_KEY:
break;
case WT_CELL_KEY_OVFL:
WT_ERR(__wt_dsk_cell_data_ref(
session, dsk->type, unpack, current));
goto key_compare;
default:
/* Not a key -- continue with the next cell. */
continue;
}
/*
* Prefix compression checks.
*
* Confirm the first non-overflow key on a page has a zero
* prefix compression count.
*/
prefix = unpack->prefix;
if (last_pfx->size == 0 && prefix != 0)
WT_ERR_VRFY(session,
"the %" PRIu32 " key on page at %s is the first "
"non-overflow key on the page and has a non-zero "
"prefix compression value",
cell_num, addr);
/* Confirm the prefix compression count is possible. */
if (cell_num > 1 && prefix > last->size)
WT_ERR_VRFY(session,
"key %" PRIu32 " on page at %s has a prefix "
"compression count of %" PRIu32 ", larger than "
"the length of the previous key, %" WT_SIZET_FMT,
cell_num, addr, prefix, last->size);
/*
* If Huffman decoding required, unpack the cell to build the
* key, then resolve the prefix. Else, we can do it faster
* internally because we don't have to shuffle memory around as
* much.
*/
if (huffman != NULL) {
WT_ERR(__wt_dsk_cell_data_ref(
session, dsk->type, unpack, current));
/*
* If there's a prefix, make sure there's enough buffer
* space, then shift the decoded data past the prefix
* and copy the prefix into place. Take care with the
* pointers: current->data may be pointing inside the
* buffer.
*/
if (prefix != 0) {
WT_ERR(__wt_buf_grow(
session, current, prefix + current->size));
memmove((uint8_t *)current->mem + prefix,
current->data, current->size);
memcpy(current->mem, last->data, prefix);
current->data = current->mem;
current->size += prefix;
}
} else {
/*
* Get the cell's data/length and make sure we have
* enough buffer space.
*/
WT_ERR(__wt_buf_init(
session, current, prefix + unpack->size));
/* Copy the prefix then the data into place. */
if (prefix != 0)
memcpy(current->mem, last->data, prefix);
memcpy((uint8_t *)current->mem + prefix, unpack->data,
unpack->size);
current->size = prefix + unpack->size;
}
key_compare: /*
* Compare the current key against the last key.
*
* Be careful about the 0th key on internal pages: we only store
* the first byte and custom collators may not be able to handle
* truncated keys.
*/
if ((dsk->type == WT_PAGE_ROW_INT && cell_num > 3) ||
(dsk->type != WT_PAGE_ROW_INT && cell_num > 1)) {
WT_ERR(__wt_compare(
session, btree->collator, last, current, &cmp));
if (cmp >= 0)
WT_ERR_VRFY(session,
"the %" PRIu32 " and %" PRIu32 " keys on "
"page at %s are incorrectly sorted",
cell_num - 2, cell_num, addr);
}
/*
* Swap the buffers: last always references the last key entry,
* last_pfx and last_ovfl reference the last prefix-compressed
* and last overflow key entries. Current gets pointed to the
* buffer we're not using this time around, which is where the
* next key goes.
*/
last = current;
if (cell_type == WT_CELL_KEY) {
current = last_pfx;
last_pfx = last;
} else {
current = last_ovfl;
last_ovfl = last;
}
WT_ASSERT(session, last != current);
}
/*
* __wt_txn_checkpoint_log --
* Write a log record for a checkpoint operation.
*/
int
__wt_txn_checkpoint_log(
WT_SESSION_IMPL *session, bool full, uint32_t flags, WT_LSN *lsnp)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(logrec);
WT_DECL_RET;
WT_ITEM *ckpt_snapshot, empty;
WT_LSN *ckpt_lsn;
WT_TXN *txn;
WT_TXN_GLOBAL *txn_global;
uint8_t *end, *p;
size_t recsize;
uint32_t i, rectype;
const char *fmt;
conn = S2C(session);
txn_global = &conn->txn_global;
txn = &session->txn;
ckpt_lsn = &txn->ckpt_lsn;
/*
* If this is a file sync, log it unless there is a full checkpoint in
* progress.
*/
if (!full) {
if (txn->full_ckpt) {
if (lsnp != NULL)
*lsnp = *ckpt_lsn;
return (0);
}
return (__txn_log_file_sync(session, flags, lsnp));
}
switch (flags) {
case WT_TXN_LOG_CKPT_PREPARE:
txn->full_ckpt = true;
if (conn->compat_major >= WT_LOG_V2) {
/*
* Write the system log record containing a checkpoint
* start operation.
*/
rectype = WT_LOGREC_SYSTEM;
fmt = WT_UNCHECKED_STRING(I);
WT_ERR(__wt_struct_size(
session, &recsize, fmt, rectype));
WT_ERR(__wt_logrec_alloc(session, recsize, &logrec));
WT_ERR(__wt_struct_pack(session,
(uint8_t *)logrec->data + logrec->size, recsize,
fmt, rectype));
logrec->size += (uint32_t)recsize;
WT_ERR(__wt_logop_checkpoint_start_pack(
session, logrec));
WT_ERR(__wt_log_write(session, logrec, ckpt_lsn, 0));
} else {
WT_ERR(__wt_log_printf(session,
"CHECKPOINT: Starting record"));
WT_ERR(__wt_log_flush_lsn(session, ckpt_lsn, true));
}
/*
* We take and immediately release the visibility lock.
* Acquiring the write lock guarantees that any transaction
* that has written to the log has also made its transaction
* visible at this time.
*/
__wt_writelock(session, &txn_global->visibility_rwlock);
__wt_writeunlock(session, &txn_global->visibility_rwlock);
/*
* We need to make sure that the log records in the checkpoint
* LSN are on disk. In particular to make sure that the
* current log file exists.
*/
WT_ERR(__wt_log_force_sync(session, ckpt_lsn));
break;
case WT_TXN_LOG_CKPT_START:
/* Take a copy of the transaction snapshot. */
txn->ckpt_nsnapshot = txn->snapshot_count;
recsize = (size_t)txn->ckpt_nsnapshot * WT_INTPACK64_MAXSIZE;
WT_ERR(__wt_scr_alloc(session, recsize, &txn->ckpt_snapshot));
p = txn->ckpt_snapshot->mem;
end = p + recsize;
for (i = 0; i < txn->snapshot_count; i++)
WT_ERR(__wt_vpack_uint(
&p, WT_PTRDIFF(end, p), txn->snapshot[i]));
break;
case WT_TXN_LOG_CKPT_STOP:
/*
* During a clean connection close, we get here without the
* prepare or start steps. In that case, log the current LSN
* as the checkpoint LSN.
*/
if (!txn->full_ckpt) {
txn->ckpt_nsnapshot = 0;
WT_CLEAR(empty);
ckpt_snapshot = ∅
WT_ERR(__wt_log_flush_lsn(session, ckpt_lsn, true));
} else
ckpt_snapshot = txn->ckpt_snapshot;
/* Write the checkpoint log record. */
rectype = WT_LOGREC_CHECKPOINT;
fmt = WT_UNCHECKED_STRING(IIIIu);
WT_ERR(__wt_struct_size(session, &recsize,
fmt, rectype, ckpt_lsn->l.file, ckpt_lsn->l.offset,
txn->ckpt_nsnapshot, ckpt_snapshot));
WT_ERR(__wt_logrec_alloc(session, recsize, &logrec));
WT_ERR(__wt_struct_pack(session,
(uint8_t *)logrec->data + logrec->size, recsize,
fmt, rectype, ckpt_lsn->l.file, ckpt_lsn->l.offset,
txn->ckpt_nsnapshot, ckpt_snapshot));
logrec->size += (uint32_t)recsize;
WT_ERR(__wt_log_write(session, logrec, lsnp,
F_ISSET(conn, WT_CONN_CKPT_SYNC) ?
WT_LOG_FSYNC : 0));
/*
* If this full checkpoint completed successfully and there is
* no hot backup in progress and this is not an unclean
* recovery, tell the logging subsystem the checkpoint LSN so
* that it can archive. Do not update the logging checkpoint
* LSN if this is during a clean connection close, only during
* a full checkpoint. A clean close may not update any
* metadata LSN and we do not want to archive in that case.
*/
if (!conn->hot_backup &&
(!FLD_ISSET(conn->log_flags, WT_CONN_LOG_RECOVER_DIRTY) ||
FLD_ISSET(conn->log_flags, WT_CONN_LOG_FORCE_DOWNGRADE)) &&
txn->full_ckpt)
__wt_log_ckpt(session, ckpt_lsn);
/* FALLTHROUGH */
case WT_TXN_LOG_CKPT_CLEANUP:
/* Cleanup any allocated resources */
WT_INIT_LSN(ckpt_lsn);
txn->ckpt_nsnapshot = 0;
__wt_scr_free(session, &txn->ckpt_snapshot);
txn->full_ckpt = false;
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: __wt_logrec_free(session, &logrec);
return (ret);
}
/*
* __wt_log_newfile --
* Create the next log file and write the file header record into it.
*/
int
__wt_log_newfile(WT_SESSION_IMPL *session, int conn_create)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LOG *log;
WT_LOG_DESC *desc;
WT_LOG_RECORD *logrec;
WT_LOGSLOT tmp;
WT_MYSLOT myslot;
conn = S2C(session);
log = conn->log;
/*
* Set aside the log file handle to be closed later. Other threads
* may still be using it to write to the log.
*/
WT_ASSERT(session, log->log_close_fh == NULL);
log->log_close_fh = log->log_fh;
log->fileid++;
WT_RET(__log_openfile(session, 1, &log->log_fh, log->fileid));
log->alloc_lsn.file = log->fileid;
log->alloc_lsn.offset = log->log_fh->size;
/*
* Set up the log descriptor record. Use a scratch buffer to
* get correct alignment for direct I/O.
*/
WT_ASSERT(session, sizeof(WT_LOG_DESC) < log->allocsize);
WT_RET(__wt_scr_alloc(session, log->allocsize, &buf));
memset(buf->mem, 0, log->allocsize);
logrec = (WT_LOG_RECORD *)buf->mem;
desc = (WT_LOG_DESC *)logrec->record;
desc->log_magic = WT_LOG_MAGIC;
desc->majorv = WT_LOG_MAJOR_VERSION;
desc->minorv = WT_LOG_MINOR_VERSION;
desc->log_size = (uint64_t)conn->log_file_max;
/*
* Now that the record is set up, initialize the record header.
*/
logrec->len = log->allocsize;
logrec->checksum = 0;
logrec->checksum = __wt_cksum(logrec, log->allocsize);
WT_CLEAR(tmp);
myslot.slot = &tmp;
myslot.offset = 0;
/*
* Recursively call __log_acquire to allocate log space for the
* log descriptor record. Call __log_fill to write it, but we
* do not need to call __log_release because we're not waiting for
* earlier operations to complete.
*/
WT_ERR(__log_acquire(session, logrec->len, &tmp));
WT_ERR(__log_fill(session, &myslot, 1, buf, NULL));
/*
* If we're called from connection creation code, we need to update
* the LSNs since we're the only write in progress.
*/
if (conn_create) {
WT_ERR(__wt_fsync(session, log->log_fh));
log->sync_lsn = tmp.slot_end_lsn;
log->write_lsn = tmp.slot_end_lsn;
}
err: __wt_scr_free(&buf);
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);
}
/*
* __open_verbose --
* Optionally output a verbose message on handle open.
*/
static inline int
__open_verbose(
WT_SESSION_IMPL *session, const char *name, int file_type, u_int flags)
{
#ifdef HAVE_VERBOSE
WT_DECL_RET;
WT_DECL_ITEM(tmp);
const char *file_type_tag, *sep;
if (!WT_VERBOSE_ISSET(session, WT_VERB_FILEOPS))
return (0);
/*
* It's useful to track file opens when debugging platforms, take some
* effort to output good tracking information.
*/
switch (file_type) {
case WT_FS_OPEN_FILE_TYPE_CHECKPOINT:
file_type_tag = "checkpoint";
break;
case WT_FS_OPEN_FILE_TYPE_DATA:
file_type_tag = "data";
break;
case WT_FS_OPEN_FILE_TYPE_DIRECTORY:
file_type_tag = "directory";
break;
case WT_FS_OPEN_FILE_TYPE_LOG:
file_type_tag = "log";
break;
case WT_FS_OPEN_FILE_TYPE_REGULAR:
file_type_tag = "regular";
break;
default:
file_type_tag = "unknown open type";
break;
}
WT_RET(__wt_scr_alloc(session, 0, &tmp));
sep = " (";
#define WT_FS_OPEN_VERBOSE_FLAG(f, name) \
if (LF_ISSET(f)) { \
WT_ERR(__wt_buf_catfmt( \
session, tmp, "%s%s", sep, name)); \
sep = ", "; \
}
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_CREATE, "create");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_DIRECTIO, "direct-IO");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_EXCLUSIVE, "exclusive");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_FIXED, "fixed");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_READONLY, "readonly");
if (tmp->size != 0)
WT_ERR(__wt_buf_catfmt(session, tmp, ")"));
__wt_verbose(session, WT_VERB_FILEOPS,
"%s: file-open: type %s%s",
name, file_type_tag, tmp->size == 0 ? "" : (char *)tmp->data);
err: __wt_scr_free(session, &tmp);
return (ret);
#else
WT_UNUSED(session);
WT_UNUSED(name);
WT_UNUSED(file_type);
WT_UNUSED(flags);
return (0);
#endif
}
/*
* __wt_lsm_meta_write --
* Write the metadata for an LSM tree.
*/
int
__wt_lsm_meta_write(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree)
{
WT_DECL_ITEM(buf);
WT_DECL_RET;
WT_LSM_CHUNK *chunk;
u_int i;
int first;
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_buf_fmt(session, buf,
"key_format=%s,value_format=%s,bloom_config=(%s),file_config=(%s)",
lsm_tree->key_format, lsm_tree->value_format,
lsm_tree->bloom_config, lsm_tree->file_config));
if (lsm_tree->collator_name != NULL)
WT_ERR(__wt_buf_catfmt(
session, buf, ",collator=%s", lsm_tree->collator_name));
WT_ERR(__wt_buf_catfmt(session, buf,
",last=%" PRIu32
",chunk_count_limit=%" PRIu32
",chunk_max=%" PRIu64
",chunk_size=%" PRIu64
",auto_throttle=%" PRIu32
",merge_max=%" PRIu32
",merge_min=%" PRIu32
",bloom=%" PRIu32
",bloom_bit_count=%" PRIu32
",bloom_hash_count=%" PRIu32,
lsm_tree->last, lsm_tree->chunk_count_limit,
lsm_tree->chunk_max, lsm_tree->chunk_size,
F_ISSET(lsm_tree, WT_LSM_TREE_THROTTLE) ? 1 : 0,
lsm_tree->merge_max, lsm_tree->merge_min, lsm_tree->bloom,
lsm_tree->bloom_bit_count, lsm_tree->bloom_hash_count));
WT_ERR(__wt_buf_catfmt(session, buf, ",chunks=["));
for (i = 0; i < lsm_tree->nchunks; i++) {
chunk = lsm_tree->chunk[i];
if (i > 0)
WT_ERR(__wt_buf_catfmt(session, buf, ","));
WT_ERR(__wt_buf_catfmt(session, buf, "id=%" PRIu32, chunk->id));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(__wt_buf_catfmt(session, buf, ",bloom"));
if (chunk->size != 0)
WT_ERR(__wt_buf_catfmt(session, buf,
",chunk_size=%" PRIu64, chunk->size));
if (chunk->count != 0)
WT_ERR(__wt_buf_catfmt(
session, buf, ",count=%" PRIu64, chunk->count));
WT_ERR(__wt_buf_catfmt(
session, buf, ",generation=%" PRIu32, chunk->generation));
}
WT_ERR(__wt_buf_catfmt(session, buf, "]"));
WT_ERR(__wt_buf_catfmt(session, buf, ",old_chunks=["));
first = 1;
for (i = 0; i < lsm_tree->nold_chunks; i++) {
chunk = lsm_tree->old_chunks[i];
WT_ASSERT(session, chunk != NULL);
if (first)
first = 0;
else
WT_ERR(__wt_buf_catfmt(session, buf, ","));
WT_ERR(__wt_buf_catfmt(session, buf, "\"%s\"", chunk->uri));
if (F_ISSET(chunk, WT_LSM_CHUNK_BLOOM))
WT_ERR(__wt_buf_catfmt(
session, buf, ",bloom=\"%s\"", chunk->bloom_uri));
}
WT_ERR(__wt_buf_catfmt(session, buf, "]"));
ret = __wt_metadata_update(session, lsm_tree->name, buf->data);
WT_ERR(ret);
err: __wt_scr_free(session, &buf);
return (ret);
}
/*
* __wt_huffman_read --
* Read a Huffman table from a file.
*/
static int
__wt_huffman_read(WT_SESSION_IMPL *session, WT_CONFIG_ITEM *ip,
struct __wt_huffman_table **tablep, u_int *entriesp, u_int *numbytesp)
{
struct __wt_huffman_table *table, *tp;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_FSTREAM *fs;
int64_t symbol, frequency;
u_int entries, lineno;
int n;
bool is_utf8;
*tablep = NULL;
*entriesp = *numbytesp = 0;
fs = NULL;
table = NULL;
/*
* Try and open the backing file.
*/
WT_RET(__huffman_confchk_file(session, ip, &is_utf8, &fs));
/*
* UTF-8 table is 256 bytes, with a range of 0-255.
* UTF-16 is 128KB (2 * 65536) bytes, with a range of 0-65535.
*/
if (is_utf8) {
entries = UINT8_MAX;
*numbytesp = 1;
WT_ERR(__wt_calloc_def(session, entries, &table));
} else {
entries = UINT16_MAX;
*numbytesp = 2;
WT_ERR(__wt_calloc_def(session, entries, &table));
}
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
for (tp = table, lineno = 1;; ++tp, ++lineno) {
WT_ERR(__wt_getline(session, fs, tmp));
if (tmp->size == 0)
break;
n = sscanf(
tmp->data, "%" SCNi64 " %" SCNi64, &symbol, &frequency);
/*
* Entries is 0-based, that is, there are (entries +1) possible
* values that can be configured. The line number is 1-based, so
* adjust the test for too many entries, and report (entries +1)
* in the error as the maximum possible number of entries.
*/
if (lineno > entries + 1)
WT_ERR_MSG(session, EINVAL,
"Huffman table file %.*s is corrupted, "
"more than %" PRIu32 " entries",
(int)ip->len, ip->str, entries + 1);
if (n != 2)
WT_ERR_MSG(session, EINVAL,
"line %u of Huffman table file %.*s is corrupted: "
"expected two unsigned integral values",
lineno, (int)ip->len, ip->str);
if (symbol < 0 || symbol > entries)
WT_ERR_MSG(session, EINVAL,
"line %u of Huffman file %.*s is corrupted; "
"symbol %" PRId64 " not in range, maximum "
"value is %u",
lineno, (int)ip->len, ip->str, symbol, entries);
if (frequency < 0 || frequency > UINT32_MAX)
WT_ERR_MSG(session, EINVAL,
"line %u of Huffman file %.*s is corrupted; "
"frequency %" PRId64 " not in range, maximum "
"value is %" PRIu32,
lineno, (int)ip->len, ip->str, frequency,
(uint32_t)UINT32_MAX);
tp->symbol = (uint32_t)symbol;
tp->frequency = (uint32_t)frequency;
}
*entriesp = lineno - 1;
*tablep = table;
if (0) {
err: __wt_free(session, table);
}
(void)__wt_fclose(session, &fs);
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_conn_optrack_setup --
* Set up operation logging.
*/
int
__wt_conn_optrack_setup(WT_SESSION_IMPL *session,
const char *cfg[], bool reconfig)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_ITEM(buf);
WT_DECL_RET;
conn = S2C(session);
/* Once an operation tracking path has been set it can't be changed. */
if (!reconfig) {
WT_RET(__wt_config_gets(session,
cfg, "operation_tracking.path", &cval));
WT_RET(__wt_strndup(session,
cval.str, cval.len, &conn->optrack_path));
}
WT_RET(__wt_config_gets(session,
cfg, "operation_tracking.enabled", &cval));
if (cval.val == 0) {
if (F_ISSET(conn, WT_CONN_OPTRACK)) {
WT_RET(__wt_conn_optrack_teardown(session, reconfig));
F_CLR(conn, WT_CONN_OPTRACK);
}
return (0);
}
if (F_ISSET(conn, WT_CONN_READONLY))
/* Operation tracking isn't supported in read-only mode */
WT_RET_MSG(session, EINVAL,
"Operation tracking is incompatible with read only "
"configuration.");
if (F_ISSET(conn, WT_CONN_OPTRACK))
/* Already enabled, nothing else to do */
return (0);
/*
* Operation tracking files will include the ID of the creating process
* in their name, so we can distinguish between log files created by
* different WiredTiger processes in the same directory. We cache the
* process id for future use.
*/
conn->optrack_pid = __wt_process_id();
/*
* Open the file in the same directory that will hold a map of
* translations between function names and function IDs. If the file
* exists, remove it.
*/
WT_RET(__wt_scr_alloc(session, 0, &buf));
WT_ERR(__wt_filename_construct(session, conn->optrack_path,
"optrack-map", conn->optrack_pid, UINT32_MAX, buf));
WT_ERR(__wt_open(session,
(const char *)buf->data, WT_FS_OPEN_FILE_TYPE_REGULAR,
WT_FS_OPEN_CREATE, &conn->optrack_map_fh));
WT_ERR(__wt_spin_init(session,
&conn->optrack_map_spinlock, "optrack map spinlock"));
WT_ERR(__wt_malloc(session, WT_OPTRACK_BUFSIZE,
&conn->dummy_session.optrack_buf));
/* Set operation tracking on */
F_SET(conn, WT_CONN_OPTRACK);
err: __wt_scr_free(session, &buf);
return (ret);
}
/*
* __wt_curlog_open --
* Initialize a log cursor.
*/
int
__wt_curlog_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curlog_compare, /* compare */
__wt_cursor_equals, /* equals */
__curlog_next, /* next */
__wt_cursor_notsup, /* prev */
__curlog_reset, /* reset */
__curlog_search, /* search */
__wt_cursor_search_near_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__curlog_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_LOG *log;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_LOG, iface) == 0);
conn = S2C(session);
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED))
WT_RET_MSG(session, EINVAL,
"Cannot open a log cursor without logging enabled");
log = conn->log;
cl = NULL;
WT_RET(__wt_calloc_one(session, &cl));
cursor = &cl->iface;
*cursor = iface;
cursor->session = &session->iface;
WT_ERR(__wt_calloc_one(session, &cl->cur_lsn));
WT_ERR(__wt_calloc_one(session, &cl->next_lsn));
WT_ERR(__wt_scr_alloc(session, 0, &cl->logrec));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opkey));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opvalue));
cursor->key_format = WT_LOGC_KEY_FORMAT;
cursor->value_format = WT_LOGC_VALUE_FORMAT;
WT_INIT_LSN(cl->cur_lsn);
WT_INIT_LSN(cl->next_lsn);
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
/*
* The user may be trying to read a log record they just wrote.
* Log records may be buffered, so force out any now.
*/
WT_ERR(__wt_log_force_write(session, 1));
/* Log cursors block archiving. */
WT_ERR(__wt_readlock(session, log->log_archive_lock));
if (0) {
err: if (F_ISSET(cursor, WT_CURSTD_OPEN))
WT_TRET(cursor->close(cursor));
else {
__wt_free(session, cl->cur_lsn);
__wt_free(session, cl->next_lsn);
__wt_scr_free(session, &cl->logrec);
__wt_scr_free(session, &cl->opkey);
__wt_scr_free(session, &cl->opvalue);
/*
* NOTE: We cannot get on the error path with the
* readlock held. No need to unlock it unless that
* changes above.
*/
__wt_free(session, cl);
}
*cursorp = NULL;
}
return (ret);
}
/*
* __wt_curjoin_open --
* Initialize a join cursor.
*
* Join cursors are read-only.
*/
int
__wt_curjoin_open(WT_SESSION_IMPL *session,
const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__curjoin_get_key, /* get-key */
__curjoin_get_value, /* get-value */
__wt_cursor_set_key_notsup, /* set-key */
__wt_cursor_set_value_notsup, /* set-value */
__wt_cursor_compare_notsup, /* compare */
__wt_cursor_equals_notsup, /* equals */
__curjoin_next, /* next */
__wt_cursor_notsup, /* prev */
__curjoin_reset, /* reset */
__wt_cursor_notsup, /* search */
__wt_cursor_search_near_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_modify_notsup, /* modify */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_notsup, /* reserve */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__wt_cursor_notsup, /* cache */
__wt_cursor_reopen_notsup, /* reopen */
__curjoin_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_JOIN *cjoin;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_TABLE *table;
size_t size;
const char *tablename, *columns;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_JOIN, iface) == 0);
if (owner != NULL)
WT_RET_MSG(session, EINVAL,
"unable to initialize a join cursor with existing owner");
tablename = uri;
if (!WT_PREFIX_SKIP(tablename, "join:table:"))
return (
__wt_unexpected_object_type(session, uri, "join:table:"));
columns = strchr(tablename, '(');
if (columns == NULL)
size = strlen(tablename);
else
size = WT_PTRDIFF(columns, tablename);
WT_RET(__wt_schema_get_table(
session, tablename, size, false, 0, &table));
WT_RET(__wt_calloc_one(session, &cjoin));
cursor = (WT_CURSOR *)cjoin;
*cursor = iface;
cursor->session = (WT_SESSION *)session;
cursor->key_format = table->key_format;
cursor->value_format = table->value_format;
cjoin->table = table;
/* Handle projections. */
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
if (columns != NULL) {
WT_ERR(__wt_struct_reformat(session, table,
columns, strlen(columns), NULL, false, tmp));
WT_ERR(__wt_strndup(
session, tmp->data, tmp->size, &cursor->value_format));
WT_ERR(__wt_strdup(session, columns, &cjoin->projection));
}
WT_ERR(__wt_cursor_init(cursor, uri, owner, cfg, cursorp));
if (0) {
err: WT_TRET(__curjoin_close(cursor));
*cursorp = NULL;
}
__wt_scr_free(session, &tmp);
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);
}
/*
* __curjoin_init_bloom --
* Populate Bloom filters
*/
static int
__curjoin_init_bloom(WT_SESSION_IMPL *session, WT_CURSOR_JOIN *cjoin,
WT_CURSOR_JOIN_ENTRY *entry, WT_BLOOM *bloom)
{
WT_COLLATOR *collator;
WT_CURSOR *c;
WT_CURSOR_JOIN_ENDPOINT *end, *endmax;
WT_DECL_ITEM(uribuf);
WT_DECL_RET;
WT_ITEM curkey, curvalue;
size_t size;
u_int skip;
int cmp;
const char *uri;
const char *raw_cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "raw", NULL };
c = NULL;
skip = 0;
if (entry->index != NULL)
/*
* Open the raw index. We're avoiding any references
* to the main table, they may be expensive.
*/
uri = entry->index->source;
else {
/*
* For joins on the main table, we just need the primary
* key for comparison, we don't need any values.
*/
size = strlen(cjoin->table->iface.name) + 3;
WT_ERR(__wt_scr_alloc(session, size, &uribuf));
WT_ERR(__wt_buf_fmt(session, uribuf, "%s()",
cjoin->table->iface.name));
uri = uribuf->data;
}
WT_ERR(__wt_open_cursor(session, uri, &cjoin->iface, raw_cfg, &c));
/* Initially position the cursor if necessary. */
endmax = &entry->ends[entry->ends_next];
if ((end = &entry->ends[0]) < endmax) {
if (F_ISSET(end, WT_CURJOIN_END_GT) ||
WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_EQ) {
WT_ERR(__wt_cursor_dup_position(end->cursor, c));
if (WT_CURJOIN_END_RANGE(end) == WT_CURJOIN_END_GE)
skip = 1;
} else if (F_ISSET(end, WT_CURJOIN_END_LT)) {
if ((ret = c->next(c)) == WT_NOTFOUND)
goto done;
WT_ERR(ret);
} else
WT_PANIC_ERR(session, EINVAL,
"fatal error in join cursor position state");
}
collator = (entry->index == NULL) ? NULL : entry->index->collator;
while (ret == 0) {
WT_ERR(c->get_key(c, &curkey));
entry->stats.iterated++;
if (entry->index != NULL) {
/*
* Repack so it's comparable to the
* reference endpoints.
*/
WT_ERR(__wt_struct_repack(session,
c->key_format,
(entry->repack_format != NULL ?
entry->repack_format : entry->index->idxkey_format),
&c->key, &curkey));
}
for (end = &entry->ends[skip]; end < endmax; end++) {
WT_ERR(__wt_compare(session, collator, &curkey,
&end->key, &cmp));
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION)) {
/* if condition satisfied, insert immediately */
switch (WT_CURJOIN_END_RANGE(end)) {
case WT_CURJOIN_END_EQ:
if (cmp == 0)
goto insert;
break;
case WT_CURJOIN_END_GT:
if (cmp > 0) {
/* skip this check next time */
skip = entry->ends_next;
goto insert;
}
break;
case WT_CURJOIN_END_GE:
if (cmp >= 0)
goto insert;
break;
case WT_CURJOIN_END_LT:
if (cmp < 0)
goto insert;
break;
case WT_CURJOIN_END_LE:
if (cmp <= 0)
goto insert;
break;
}
} else if (!F_ISSET(end, WT_CURJOIN_END_LT)) {
if (cmp < 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto advance;
if (cmp > 0) {
if (F_ISSET(end, WT_CURJOIN_END_GT))
skip = 1;
else
goto done;
}
} else {
if (cmp > 0 || (cmp == 0 &&
!F_ISSET(end, WT_CURJOIN_END_EQ)))
goto done;
}
}
/*
* Either it's a disjunction that hasn't satisfied any
* condition, or it's a conjunction that has satisfied all
* conditions.
*/
if (F_ISSET(entry, WT_CURJOIN_ENTRY_DISJUNCTION))
goto advance;
insert:
if (entry->index != NULL) {
curvalue.data =
(unsigned char *)curkey.data + curkey.size;
WT_ASSERT(session, c->key.size > curkey.size);
curvalue.size = c->key.size - curkey.size;
}
else
WT_ERR(c->get_key(c, &curvalue));
__wt_bloom_insert(bloom, &curvalue);
entry->stats.bloom_insert++;
advance:
if ((ret = c->next(c)) == WT_NOTFOUND)
break;
}
done:
WT_ERR_NOTFOUND_OK(ret);
err: if (c != NULL)
WT_TRET(c->close(c));
__wt_scr_free(session, &uribuf);
return (ret);
}
/*
* __wt_win_directory_list --
* Get a list of files from a directory, MSVC version.
*/
int
__wt_win_directory_list(WT_FILE_SYSTEM *file_system,
WT_SESSION *wt_session, const char *directory,
const char *prefix, char ***dirlistp, uint32_t *countp)
{
DWORD windows_error;
HANDLE findhandle;
WIN32_FIND_DATAW finddata;
WT_DECL_ITEM(pathbuf);
WT_DECL_ITEM(file_utf8);
WT_DECL_ITEM(pathbuf_wide);
WT_DECL_ITEM(prefix_wide);
WT_DECL_RET;
WT_SESSION_IMPL *session;
size_t dirallocsz, pathlen, prefix_widelen;
uint32_t count;
char *dir_copy, **entries;
session = (WT_SESSION_IMPL *)wt_session;
*dirlistp = NULL;
*countp = 0;
findhandle = INVALID_HANDLE_VALUE;
dirallocsz = 0;
entries = NULL;
WT_ERR(__wt_strdup(session, directory, &dir_copy));
pathlen = strlen(dir_copy);
if (dir_copy[pathlen - 1] == '\\')
dir_copy[pathlen - 1] = '\0';
WT_ERR(__wt_scr_alloc(session, pathlen + 3, &pathbuf));
WT_ERR(__wt_buf_fmt(session, pathbuf, "%s\\*", dir_copy));
WT_ERR(__wt_to_utf16_string(session, pathbuf->data, &pathbuf_wide));
WT_ERR(__wt_to_utf16_string(session, prefix, &prefix_wide));
prefix_widelen = wcslen(prefix_wide->data);
findhandle = FindFirstFileW(pathbuf_wide->data, &finddata);
if (findhandle == INVALID_HANDLE_VALUE) {
windows_error = __wt_getlasterror();
__wt_errx(session,
"%s: directory-list: FindFirstFile: %s",
pathbuf->data, __wt_formatmessage(session, windows_error));
WT_ERR(__wt_map_windows_error(windows_error));
}
count = 0;
do {
/*
* Skip . and ..
*/
if (wcscmp(finddata.cFileName, L".") == 0 ||
wcscmp(finddata.cFileName, L"..") == 0)
continue;
/* The list of files is optionally filtered by a prefix. */
if (prefix != NULL &&
wcsncmp(finddata.cFileName, prefix_wide->data,
prefix_widelen) != 0)
continue;
WT_ERR(__wt_realloc_def(
session, &dirallocsz, count + 1, &entries));
WT_ERR(__wt_to_utf8_string(
session, finddata.cFileName, &file_utf8));
WT_ERR(__wt_strdup(session, file_utf8->data, &entries[count]));
++count;
__wt_scr_free(session, &file_utf8);
} while (FindNextFileW(findhandle, &finddata) != 0);
*dirlistp = entries;
*countp = count;
err: if (findhandle != INVALID_HANDLE_VALUE)
if (FindClose(findhandle) == 0) {
windows_error = __wt_getlasterror();
__wt_errx(session,
"%s: directory-list: FindClose: %s",
pathbuf->data,
__wt_formatmessage(session, windows_error));
if (ret == 0)
ret = __wt_map_windows_error(windows_error);
}
__wt_free(session, dir_copy);
__wt_scr_free(session, &pathbuf);
__wt_scr_free(session, &file_utf8);
__wt_scr_free(session, &pathbuf_wide);
__wt_scr_free(session, &prefix_wide);
if (ret == 0)
return (0);
WT_TRET(__wt_win_directory_list_free(
file_system, wt_session, entries, count));
WT_RET_MSG(session, ret,
"%s: directory-list, prefix \"%s\"",
directory, prefix == NULL ? "" : prefix);
}
/*
* __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);
}
/*
* __config_merge_scan --
* Walk a configuration string, inserting entries into the merged array.
*/
static int
__config_merge_scan(WT_SESSION_IMPL *session,
const char *key, const char *value, WT_CONFIG_MERGE *cp)
{
WT_CONFIG cparser;
WT_CONFIG_ITEM k, v;
WT_DECL_ITEM(kb);
WT_DECL_ITEM(vb);
WT_DECL_RET;
size_t len;
WT_ERR(__wt_scr_alloc(session, 0, &kb));
WT_ERR(__wt_scr_alloc(session, 0, &vb));
WT_ERR(__wt_config_init(session, &cparser, value));
while ((ret = __wt_config_next(&cparser, &k, &v)) == 0) {
if (k.type != WT_CONFIG_ITEM_STRING &&
k.type != WT_CONFIG_ITEM_ID)
WT_ERR_MSG(session, EINVAL,
"Invalid configuration key found: '%s'\n", k.str);
/* Include the quotes around string keys/values. */
if (k.type == WT_CONFIG_ITEM_STRING) {
--k.str;
k.len += 2;
}
if (v.type == WT_CONFIG_ITEM_STRING) {
--v.str;
v.len += 2;
}
/*
* !!!
* We're using a JSON quote character to separate the names we
* create for nested structures. That's not completely safe as
* it's possible to quote characters in JSON such that a quote
* character appears as a literal character in a key name. In
* a few cases, applications can create their own key namespace
* (for example, shared library extension names), and therefore
* it's possible for an application to confuse us. Error if we
* we ever see a key with a magic character.
*/
for (len = 0; len < k.len; ++len)
if (k.str[len] == SEPC)
WT_ERR_MSG(session, EINVAL,
"key %.*s contains a '%c' separator "
"character",
(int)k.len, (char *)k.str, SEPC);
/* Build the key/value strings. */
WT_ERR(__wt_buf_fmt(session,
kb, "%s%s%.*s",
key == NULL ? "" : key,
key == NULL ? "" : SEP,
(int)k.len, k.str));
WT_ERR(__wt_buf_fmt(session,
vb, "%.*s", (int)v.len, v.str));
/*
* If the value is a structure, recursively parse it.
*
* !!!
* Don't merge unless the structure has field names. WiredTiger
* stores checkpoint LSNs in the metadata file using nested
* structures without field names: "checkpoint_lsn=(1,0)", not
* "checkpoint_lsn=(file=1,offset=0)". The value type is still
* WT_CONFIG_ITEM_STRUCT, so we check for a field name in the
* value.
*/
if (v.type == WT_CONFIG_ITEM_STRUCT &&
strchr(vb->data, '=') != NULL) {
WT_ERR(__config_merge_scan(
session, kb->data, vb->data, cp));
continue;
}
/* Insert the value into the array. */
WT_ERR(__wt_realloc_def(session,
&cp->entries_allocated,
cp->entries_next + 1, &cp->entries));
WT_ERR(__wt_strndup(session,
kb->data, kb->size, &cp->entries[cp->entries_next].k));
WT_ERR(__wt_strndup(session,
vb->data, vb->size, &cp->entries[cp->entries_next].v));
cp->entries[cp->entries_next].gen = cp->entries_next;
++cp->entries_next;
}
WT_ERR_NOTFOUND_OK(ret);
err: __wt_scr_free(session, &kb);
__wt_scr_free(session, &vb);
return (ret);
}
/*
* __create_file --
* Create a new 'file:' object.
*/
static int
__create_file(WT_SESSION_IMPL *session,
const char *uri, int exclusive, const char *config)
{
WT_DECL_ITEM(val);
WT_DECL_RET;
uint32_t allocsize;
int is_metadata;
const char *filename, **p, *filecfg[] =
{ WT_CONFIG_BASE(session, file_meta), config, NULL, NULL };
char *fileconf;
fileconf = NULL;
is_metadata = strcmp(uri, WT_METAFILE_URI) == 0;
filename = uri;
if (!WT_PREFIX_SKIP(filename, "file:"))
WT_RET_MSG(session, EINVAL, "Expected a 'file:' URI: %s", uri);
/* Check if the file already exists. */
if (!is_metadata && (ret =
__wt_metadata_search(session, uri, &fileconf)) != WT_NOTFOUND) {
if (exclusive)
WT_TRET(EEXIST);
goto err;
}
/* Sanity check the allocation size. */
WT_RET(__wt_direct_io_size_check(
session, filecfg, "allocation_size", &allocsize));
/* Create the file. */
WT_ERR(__wt_block_manager_create(session, filename, allocsize));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_fileop(session, NULL, uri));
/*
* If creating an ordinary file, append the file ID and current version
* numbers to the passed-in configuration and insert the resulting
* configuration into the metadata.
*/
if (!is_metadata) {
WT_ERR(__wt_scr_alloc(session, 0, &val));
WT_ERR(__wt_buf_fmt(session, val,
"id=%" PRIu32 ",version=(major=%d,minor=%d)",
++S2C(session)->next_file_id,
WT_BTREE_MAJOR_VERSION_MAX, WT_BTREE_MINOR_VERSION_MAX));
for (p = filecfg; *p != NULL; ++p)
;
*p = val->data;
WT_ERR(__wt_config_collapse(session, filecfg, &fileconf));
WT_ERR(__wt_metadata_insert(session, uri, fileconf));
}
/*
* Open the file to check that it was setup correctly. We don't need to
* pass the configuration, we just wrote the collapsed configuration
* into the metadata file, and it's going to be read/used by underlying
* functions.
*
* Keep the handle exclusive until it is released at the end of the
* call, otherwise we could race with a drop.
*/
WT_ERR(__wt_session_get_btree(
session, uri, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
if (WT_META_TRACKING(session))
WT_ERR(__wt_meta_track_handle_lock(session, 1));
else
WT_ERR(__wt_session_release_btree(session));
err: __wt_scr_free(session, &val);
__wt_free(session, fileconf);
return (ret);
}
/*
* __wt_meta_ckptlist_set --
* Set a file's checkpoint value from the WT_CKPT list.
*/
int
__wt_meta_ckptlist_set(WT_SESSION_IMPL *session,
const char *fname, WT_CKPT *ckptbase, WT_LSN *ckptlsn)
{
WT_CKPT *ckpt;
WT_DECL_ITEM(buf);
WT_DECL_RET;
time_t secs;
int64_t maxorder;
const char *sep;
WT_ERR(__wt_scr_alloc(session, 0, &buf));
maxorder = 0;
sep = "";
WT_ERR(__wt_buf_fmt(session, buf, "checkpoint=("));
WT_CKPT_FOREACH(ckptbase, ckpt) {
/*
* Each internal checkpoint name is appended with a generation
* to make it a unique name. We're solving two problems: when
* two checkpoints are taken quickly, the timer may not be
* unique and/or we can even see time travel on the second
* checkpoint if we snapshot the time in-between nanoseconds
* rolling over. Second, if we reset the generational counter
* when new checkpoints arrive, we could logically re-create
* specific checkpoints, racing with cursors open on those
* checkpoints. I can't think of any way to return incorrect
* results by racing with those cursors, but it's simpler not
* to worry about it.
*/
if (ckpt->order > maxorder)
maxorder = ckpt->order;
/* Skip deleted checkpoints. */
if (F_ISSET(ckpt, WT_CKPT_DELETE))
continue;
if (F_ISSET(ckpt, WT_CKPT_ADD | WT_CKPT_UPDATE)) {
/*
* We fake checkpoints for handles in the middle of a
* bulk load. If there is a checkpoint, convert the
* raw cookie to a hex string.
*/
if (ckpt->raw.size == 0)
ckpt->addr.size = 0;
else
WT_ERR(__wt_raw_to_hex(session,
ckpt->raw.data,
ckpt->raw.size, &ckpt->addr));
/* Set the order and timestamp. */
if (F_ISSET(ckpt, WT_CKPT_ADD))
ckpt->order = ++maxorder;
/*
* XXX
* Assumes a time_t fits into a uintmax_t, which isn't
* guaranteed, a time_t has to be an arithmetic type,
* but not an integral type.
*/
__wt_seconds(session, &secs);
ckpt->sec = (uintmax_t)secs;
}
if (strcmp(ckpt->name, WT_CHECKPOINT) == 0)
WT_ERR(__wt_buf_catfmt(session, buf,
"%s%s.%" PRId64 "=(addr=\"%.*s\",order=%" PRId64
",time=%" PRIuMAX ",size=%" PRIu64
",write_gen=%" PRIu64 ")",
sep, ckpt->name, ckpt->order,
(int)ckpt->addr.size, (char *)ckpt->addr.data,
ckpt->order, ckpt->sec, ckpt->ckpt_size,
ckpt->write_gen));
else
WT_ERR(__wt_buf_catfmt(session, buf,
"%s%s=(addr=\"%.*s\",order=%" PRId64
",time=%" PRIuMAX ",size=%" PRIu64
",write_gen=%" PRIu64 ")",
sep, ckpt->name,
(int)ckpt->addr.size, (char *)ckpt->addr.data,
ckpt->order, ckpt->sec, ckpt->ckpt_size,
ckpt->write_gen));
sep = ",";
}
WT_ERR(__wt_buf_catfmt(session, buf, ")"));
if (ckptlsn != NULL)
WT_ERR(__wt_buf_catfmt(session, buf,
",checkpoint_lsn=(%" PRIu32 ",%" PRIuMAX ")",
ckptlsn->l.file, (uintmax_t)ckptlsn->l.offset));
WT_ERR(__ckpt_set(session, fname, buf->mem));
err: __wt_scr_free(session, &buf);
return (ret);
}
/*
* __wt_curlog_open --
* Initialize a log cursor.
*/
int
__wt_curlog_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curlog_compare, /* compare */
__wt_cursor_equals, /* equals */
__curlog_next, /* next */
__wt_cursor_notsup, /* prev */
__curlog_reset, /* reset */
__curlog_search, /* search */
__wt_cursor_search_near_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_modify_notsup, /* modify */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__wt_cursor_notsup, /* reserve */
__wt_cursor_reconfigure_notsup, /* reconfigure */
__wt_cursor_notsup, /* cache */
__wt_cursor_reopen_notsup, /* reopen */
__curlog_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_LOG *log;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_LOG, iface) == 0);
conn = S2C(session);
log = conn->log;
WT_RET(__wt_calloc_one(session, &cl));
cursor = (WT_CURSOR *)cl;
*cursor = iface;
cursor->session = (WT_SESSION *)session;
cursor->key_format = WT_LOGC_KEY_FORMAT;
cursor->value_format = WT_LOGC_VALUE_FORMAT;
WT_ERR(__wt_calloc_one(session, &cl->cur_lsn));
WT_ERR(__wt_calloc_one(session, &cl->next_lsn));
WT_ERR(__wt_scr_alloc(session, 0, &cl->logrec));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opkey));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opvalue));
WT_INIT_LSN(cl->cur_lsn);
WT_INIT_LSN(cl->next_lsn);
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
if (log != NULL) {
/*
* The user may be trying to read a log record they just wrote.
* Log records may be buffered, so force out any now.
*/
WT_ERR(__wt_log_force_write(session, 1, NULL));
/* Log cursors block archiving. */
__wt_readlock(session, &log->log_archive_lock);
F_SET(cl, WT_CURLOG_ARCHIVE_LOCK);
(void)__wt_atomic_add32(&conn->log_cursors, 1);
}
if (0) {
err: WT_TRET(__curlog_close(cursor));
*cursorp = NULL;
}
return (ret);
}
