static ham_status_t
__insert_cursor(ham_btree_t *be, ham_key_t *key, ham_record_t *record,
ham_bt_cursor_t *cursor, insert_hints_t *hints)
{
ham_status_t st;
ham_page_t *root;
ham_db_t *db=be_get_db(be);
ham_env_t *env = db_get_env(db);
insert_scratchpad_t scratchpad;
ham_assert(hints->force_append == HAM_FALSE, (0));
ham_assert(hints->force_prepend == HAM_FALSE, (0));
/*
* initialize the scratchpad
*/
memset(&scratchpad, 0, sizeof(scratchpad));
scratchpad.be=be;
scratchpad.record=record;
scratchpad.cursor=cursor;
/*
* get the root-page...
*/
ham_assert(btree_get_rootpage(be)!=0, ("btree has no root page"));
st=db_fetch_page(&root, db, btree_get_rootpage(be), 0);
ham_assert(st ? root == NULL : 1, (0));
if (st)
return st;
/*
* ... and start the recursion
*/
st=__insert_recursive(root, key, 0, &scratchpad, hints);
/*
* if the root page was split, we have to create a new
* root page.
*/
if (st==SPLIT) {
ham_page_t *newroot;
btree_node_t *node;
/*
* the root-page will be changed...
*/
st=ham_log_add_page_before(root);
if (st)
return (st);
/*
* allocate a new root page
*/
st=db_alloc_page(&newroot, db, PAGE_TYPE_B_ROOT, 0);
ham_assert(st ? newroot == NULL : 1, (0));
if (st)
return (st);
ham_assert(page_get_owner(newroot), (""));
/* clear the node header */
memset(page_get_payload(newroot), 0, sizeof(btree_node_t));
stats_page_is_nuked(db, root, HAM_TRUE);
/*
* insert the pivot element and the ptr_left
*/
node=ham_page_get_btree_node(newroot);
btree_node_set_ptr_left(node, btree_get_rootpage(be));
st=__insert_nosplit(newroot, &scratchpad.key,
scratchpad.rid, scratchpad.record, scratchpad.cursor,
hints);
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
scratchpad.cursor=0; /* don't overwrite cursor if __insert_nosplit
is called again */
if (st) {
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
if (scratchpad.key.data)
allocator_free(env_get_allocator(env), scratchpad.key.data);
return (st);
}
/*
* set the new root page
*
* !!
* do NOT delete the old root page - it's still in use!
*
* also don't forget to flush the backend - otherwise the header
* page of the database will not contain the updated information.
* The backend is flushed when the database is closed, but if
* recovery is enabled then the flush here is critical.
*/
btree_set_rootpage(be, page_get_self(newroot));
be_set_dirty(be, HAM_TRUE);
be->_fun_flush(be);
/*
* As we re-purpose a page, we will reset its pagecounter
* as well to signal its first use as the new type assigned
* here.
*/
if (env_get_cache(env) && (page_get_type(root)!=PAGE_TYPE_B_INDEX))
cache_update_page_access_counter(root, env_get_cache(env), 0);
page_set_type(root, PAGE_TYPE_B_INDEX);
page_set_dirty(root, env);
page_set_dirty(newroot, env);
/* the root page was modified (btree_set_rootpage) - make sure that
* it's logged */
if (env_get_rt_flags(env)&HAM_ENABLE_RECOVERY) {
st=txn_add_page(env_get_txn(env), env_get_header_page(env),
HAM_TRUE);
if (st)
return (st);
}
}
/*
* release the scratchpad-memory and return to caller
*/
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
if (scratchpad.key.data)
allocator_free(env_get_allocator(env), scratchpad.key.data);
return (st);
}
/**
* write a series of data chunks to storage at file offset 'addr'.
*
* The chunks are assumed to be stored in sequential order, adjacent
* to each other, i.e. as one long data strip.
*
* Writing is performed on a per-page basis, where special conditions
* will decide whether or not the write operation is performed
* through the page cache or directly to device; such is determined
* on a per-page basis.
*/
static ham_status_t
__write_chunks(ham_env_t *env, ham_page_t *page, ham_offset_t addr,
ham_bool_t allocated, ham_bool_t freshly_created,
ham_u8_t **chunk_data, ham_size_t *chunk_size,
ham_size_t chunks)
{
ham_size_t i;
ham_status_t st;
ham_offset_t pageid;
ham_device_t *device=env_get_device(env);
ham_size_t pagesize = env_get_pagesize(env);
ham_assert(freshly_created ? allocated : 1, (0));
/*
* for each chunk...
*/
for (i=0; i<chunks; i++) {
while (chunk_size[i]) {
/*
* get the page-ID from this chunk
*/
pageid = addr - (addr % pagesize);
/*
* is this the current page?
*/
if (page && page_get_self(page)!=pageid)
page=0;
/*
* fetch the page from the cache, if it's in the cache
* (unless we're logging - in this case always go through
* the buffered routines)
*/
if (!page) {
/*
* keep pages in cache when they are located at the 'edges' of
* the blob, as they MAY be accessed for different data.
* Of course, when a blob is small, there's only one (partial)
* page accessed anyhow, so that one should end up in cache
* then.
*
* When transaction logging is turned on, it's the same story,
* really. We _could_ keep all those pages in cache now,
* but this would be thrashing the cache with blob data that's
* accessed once only and for transaction abort (or commit)
* the amount of effort does not change.
*
* THOUGHT:
*
* Do we actually care what was in that page, which is going
* to be overwritten in its entirety, BEFORE we do this, i.e.
* before the transaction?
*
* Answer: NO (and YES in special circumstances).
*
* Elaboration: As this would have been free space before, the
* actual content does not matter, so it's not required to add
* the FULL pages written by the blob write action here to the
* transaction log: even on transaction abort, that lingering
* data is marked as 'bogus'/free as it was before anyhow.
*
* And then, assuming a longer running transaction, where this
* page was freed during a previous action WITHIN
* the transaction, well, than the transaction log should
* already carry this page's previous content as instructed
* by the erase operation. HOWEVER, the erase operation would
* not have a particular NEED to edit this page, as an erase op
* is complete by just marking this space as free in the
* freelist, resulting in the freelist pages (and the btree
* pages) being the only ones being edited and ending up in
* the transaction log then.
*
* Which means we'll have to log the previous content of these
* pages to the transaction log anyhow. UNLESS, that is, when
* WE allocated these pages in the first place: then there
* cannot be any 'pre-transaction' state of these pages
* except that of 'not existing', i.e. 'free'. In which case,
* their actual content doesn't matter! (freshly_created)
*
* And what if we have recovery logging turned on, but it's
* not about an active transaction here?
* In that case, the recovery log would only log the OLD page
* content, which we've concluded is insignificant, ever. Of
* course, that's assuming (again!) that we're writing to
* freshly created pages, which no-one has seen before.
*
* Just as long as we can prevent this section from thrashing
* the page cache, thank you very much...
*/
ham_bool_t at_blob_edge = (__blob_from_cache(env, chunk_size[i])
|| (addr % pagesize) != 0
|| chunk_size[i] < pagesize);
ham_bool_t cacheonly = (!at_blob_edge
&& (!env_get_log(env)
|| freshly_created));
//ham_assert(db_get_txn(db) ? !!env_get_log(env) : 1, (0));
st=env_fetch_page(&page, env, pageid,
cacheonly ? DB_ONLY_FROM_CACHE :
at_blob_edge ? 0 : DB_NEW_PAGE_DOES_THRASH_CACHE);
ham_assert(st ? !page : 1, (0));
/* blob pages don't have a page header */
if (page)
{
page_set_npers_flags(page,
page_get_npers_flags(page)|PAGE_NPERS_NO_HEADER);
/* if this page was recently allocated by the parent
* function: set a flag */
if (cacheonly
&& allocated
&& addr==page_get_self(page)
&& env_get_txn(env))
page_set_alloc_txn_id(page, txn_get_id(env_get_txn(env)));
}
else if (st) {
return st;
}
}
/*
* if we have a page pointer: use it; otherwise write directly
* to the device
*/
if (page) {
ham_size_t writestart=
(ham_size_t)(addr-page_get_self(page));
ham_size_t writesize =
(ham_size_t)(pagesize - writestart);
if (writesize>chunk_size[i])
writesize=chunk_size[i];
if ((st=ham_log_add_page_before(page)))
return (st);
memcpy(&page_get_raw_payload(page)[writestart], chunk_data[i],
writesize);
page_set_dirty(page, env);
addr+=writesize;
chunk_data[i]+=writesize;
chunk_size[i]-=writesize;
}
else {
ham_size_t s = chunk_size[i];
/* limit to the next page boundary */
if (s > pageid+pagesize-addr)
s = (ham_size_t)(pageid+pagesize-addr);
ham_assert(env_get_log(env) ? freshly_created : 1, (0));
st=device->write(device, addr, chunk_data[i], s);
if (st)
return st;
addr+=s;
chunk_data[i]+=s;
chunk_size[i]-=s;
}
}
}
return (0);
}
static ham_status_t
__append_key(ham_btree_t *be, ham_key_t *key, ham_record_t *record,
ham_bt_cursor_t *cursor, insert_hints_t *hints)
{
ham_status_t st=0;
ham_page_t *page;
btree_node_t *node;
ham_db_t *db;
#ifdef HAM_DEBUG
if (cursor && !bt_cursor_is_nil(cursor)) {
ham_assert(be_get_db(be) == bt_cursor_get_db(cursor), (0));
}
#endif
db = be_get_db(be);
/*
* see if we get this btree leaf; if not, revert to regular scan
*
* As this is a speed-improvement hint re-using recent material, the page
* should still sit in the cache, or we're using old info, which should be
* discarded.
*/
st = db_fetch_page(&page, db, hints->leaf_page_addr, DB_ONLY_FROM_CACHE);
if (st)
return st;
if (!page) {
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
page_add_ref(page);
node=ham_page_get_btree_node(page);
ham_assert(btree_node_is_leaf(node), ("iterator points to internal node"));
/*
* if the page is already full OR this page is not the right-most page
* when we APPEND or the left-most node when we PREPEND
* OR the new key is not the highest key: perform a normal insert
*/
if ((hints->force_append && btree_node_get_right(node))
|| (hints->force_prepend && btree_node_get_left(node))
|| btree_node_get_count(node) >= btree_get_maxkeys(be)) {
page_release_ref(page);
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
/*
* if the page is not empty: check if we append the key at the end / start
* (depending on force_append/force_prepend),
* or if it's actually inserted in the middle (when neither force_append
* or force_prepend is specified: that'd be SEQUENTIAL insertion
* hinting somewhere in the middle of the total key range.
*/
if (btree_node_get_count(node)!=0) {
int cmp_hi;
int cmp_lo;
hints->cost++;
if (!hints->force_prepend) {
cmp_hi = key_compare_pub_to_int(db, page, key,
btree_node_get_count(node)-1);
/* key is in the middle */
if (cmp_hi < -1) {
page_release_ref(page);
return (ham_status_t)cmp_hi;
}
/* key is at the end */
if (cmp_hi > 0) {
if (btree_node_get_right(node)) {
/* not at top end of the btree, so we can't do the
* fast track */
page_release_ref(page);
//hints->flags &= ~HAM_HINT_APPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
hints->force_append = HAM_TRUE;
hints->force_prepend = HAM_FALSE;
}
}
else { /* hints->force_prepend is true */
/* not bigger than the right-most node while we
* were trying to APPEND */
cmp_hi = -1;
}
if (!hints->force_append) {
cmp_lo = key_compare_pub_to_int(db, page, key, 0);
/* in the middle range */
if (cmp_lo < -1) {
page_release_ref(page);
return ((ham_status_t)cmp_lo);
}
/* key is at the start of page */
if (cmp_lo < 0) {
if (btree_node_get_left(node)) {
/* not at bottom end of the btree, so we can't
* do the fast track */
page_release_ref(page);
//hints->flags &= ~HAM_HINT_PREPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_TRUE;
}
}
else { /* hints->force_prepend is true */
/* not smaller than the left-most node while we were
* trying to PREPEND */
cmp_lo = +1;
}
/* handle inserts in the middle range */
if (cmp_lo >= 0 && cmp_hi <= 0) {
/*
* Depending on where we are in the btree, the current key either
* is going to end up in the middle of the given node/page,
* OR the given key is out of range of the given leaf node.
*/
if (hints->force_append || hints->force_prepend) {
/*
* when prepend or append is FORCED, we are expected to
* add keys ONLY at the beginning or end of the btree
* key range. Clearly the current key does not fit that
* criterium.
*/
page_release_ref(page);
//hints->flags &= ~HAM_HINT_PREPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
/*
* we discovered that the key must be inserted in the middle
* of the current leaf.
*
* It does not matter whether the current leaf is at the start or
* end of the btree range; as we need to add the key in the middle
* of the current leaf, that info alone is enough to continue with
* the fast track insert operation.
*/
ham_assert(!hints->force_prepend && !hints->force_append, (0));
}
ham_assert((hints->force_prepend + hints->force_append) < 2,
("Either APPEND or PREPEND flag MAY be set, but not both"));
}
else { /* empty page: force insertion in slot 0 */
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_TRUE;
}
/*
* the page will be changed - write it to the log (if a log exists)
*/
st=ham_log_add_page_before(page);
if (st) {
page_release_ref(page);
return (st);
}
/*
* OK - we're really appending/prepending the new key.
*/
ham_assert(hints->force_append || hints->force_prepend, (0));
st=__insert_nosplit(page, key, 0, record, cursor, hints);
page_release_ref(page);
return (st);
}
ham_status_t
blob_duplicate_insert(ham_db_t *db, ham_offset_t table_id,
ham_record_t *record, ham_size_t position, ham_u32_t flags,
dupe_entry_t *entries, ham_size_t num_entries,
ham_offset_t *rid, ham_size_t *new_position)
{
ham_status_t st=0;
dupe_table_t *table=0;
ham_bool_t alloc_table=0;
ham_bool_t resize=0;
ham_page_t *page=0;
ham_env_t *env=db_get_env(db);
/*
* create a new duplicate table if none existed, and insert
* the first entry
*/
if (!table_id) {
ham_assert(num_entries==2, (""));
/* allocates space for 8 (!) entries */
table=allocator_calloc(env_get_allocator(env),
sizeof(dupe_table_t)+7*sizeof(dupe_entry_t));
if (!table)
return HAM_OUT_OF_MEMORY;
dupe_table_set_capacity(table, 8);
dupe_table_set_count(table, 1);
memcpy(dupe_table_get_entry(table, 0), &entries[0],
sizeof(entries[0]));
/* skip the first entry */
entries++;
num_entries--;
alloc_table=1;
}
else {
/*
* otherwise load the existing table
*/
st=__get_duplicate_table(&table, &page, env, table_id);
ham_assert(st ? table == NULL : 1, (0));
ham_assert(st ? page == NULL : 1, (0));
if (!table)
return st ? st : HAM_INTERNAL_ERROR;
if (!page && !(env_get_rt_flags(env)&HAM_IN_MEMORY_DB))
alloc_table=1;
}
if (page)
if ((st=ham_log_add_page_before(page)))
return (st);
ham_assert(num_entries==1, (""));
/*
* resize the table, if necessary
*/
if (!(flags & HAM_OVERWRITE)
&& dupe_table_get_count(table)+1>=dupe_table_get_capacity(table)) {
dupe_table_t *old=table;
ham_size_t new_cap=dupe_table_get_capacity(table);
if (new_cap < 3*8)
new_cap += 8;
else
new_cap += new_cap/3;
table=allocator_calloc(env_get_allocator(env), sizeof(dupe_table_t)+
(new_cap-1)*sizeof(dupe_entry_t));
if (!table)
return (HAM_OUT_OF_MEMORY);
dupe_table_set_capacity(table, new_cap);
dupe_table_set_count(table, dupe_table_get_count(old));
memcpy(dupe_table_get_entry(table, 0), dupe_table_get_entry(old, 0),
dupe_table_get_count(old)*sizeof(dupe_entry_t));
if (alloc_table)
allocator_free(env_get_allocator(env), old);
alloc_table=1;
resize=1;
}
/*
* insert sorted, unsorted or overwrite the entry at the requested position
*/
if (flags&HAM_OVERWRITE) {
dupe_entry_t *e=dupe_table_get_entry(table, position);
if (!(dupe_entry_get_flags(e)&(KEY_BLOB_SIZE_SMALL
|KEY_BLOB_SIZE_TINY
|KEY_BLOB_SIZE_EMPTY))) {
(void)blob_free(env, db, dupe_entry_get_rid(e), 0);
}
memcpy(dupe_table_get_entry(table, position),
&entries[0], sizeof(entries[0]));
}
else {
if (db_get_rt_flags(db)&HAM_SORT_DUPLICATES) {
if (page)
page_add_ref(page);
position=__get_sorted_position(db, table, record, flags);
if (page)
page_release_ref(page);
if (position<0)
return ((ham_status_t)position);
}
else if (flags&HAM_DUPLICATE_INSERT_BEFORE) {
/* do nothing, insert at the current position */
}
else if (flags&HAM_DUPLICATE_INSERT_AFTER) {
position++;
if (position > dupe_table_get_count(table))
position=dupe_table_get_count(table);
}
else if (flags&HAM_DUPLICATE_INSERT_FIRST) {
position=0;
}
else if (flags&HAM_DUPLICATE_INSERT_LAST) {
position=dupe_table_get_count(table);
}
else {
position=dupe_table_get_count(table);
}
if (position != dupe_table_get_count(table)) {
memmove(dupe_table_get_entry(table, position+1),
dupe_table_get_entry(table, position),
sizeof(entries[0])*(dupe_table_get_count(table)-position));
}
memcpy(dupe_table_get_entry(table, position),
&entries[0], sizeof(entries[0]));
dupe_table_set_count(table, dupe_table_get_count(table)+1);
}
/*
* write the table back to disk and return the blobid of the table
*/
if ((table_id && !page) || resize)
{
ham_record_t rec= {0};
rec.data=(ham_u8_t *)table;
rec.size=sizeof(dupe_table_t)
+(dupe_table_get_capacity(table)-1)*sizeof(dupe_entry_t);
st=blob_overwrite(env, db, table_id, &rec, 0, rid);
}
else if (!table_id)
{
ham_record_t rec= {0};
rec.data=(ham_u8_t *)table;
rec.size=sizeof(dupe_table_t)
+(dupe_table_get_capacity(table)-1)*sizeof(dupe_entry_t);
st=blob_allocate(env, db, &rec, 0, rid);
}
else if (table_id && page)
{
page_set_dirty(page, env);
}
else
{
ham_assert(!"shouldn't be here", (0));
}
if (alloc_table)
allocator_free(env_get_allocator(env), table);
if (new_position)
*new_position=position;
return (st);
}
