ham_status_t
blob_free(ham_env_t *env, ham_db_t *db, ham_offset_t blobid, ham_u32_t flags)
{
ham_status_t st;
blob_t hdr;
/*
* in-memory-database: the blobid is actually a pointer to the memory
* buffer, in which the blob is stored
*/
if (env_get_rt_flags(env)&HAM_IN_MEMORY_DB) {
allocator_free(env_get_allocator(env), (void *)U64_TO_PTR(blobid));
return (0);
}
ham_assert(blobid%DB_CHUNKSIZE==0, (0));
/*
* fetch the blob header
*/
st=__read_chunk(env, 0, 0, blobid, (ham_u8_t *)&hdr, sizeof(hdr));
if (st)
return (st);
ham_assert(blob_get_alloc_size(&hdr)%DB_CHUNKSIZE==0, (0));
/*
* sanity check
*/
ham_assert(blob_get_self(&hdr)==blobid,
("invalid blobid %llu != %llu", blob_get_self(&hdr), blobid));
if (blob_get_self(&hdr)!=blobid)
return (HAM_BLOB_NOT_FOUND);
/*
* move the blob to the freelist
*/
st = freel_mark_free(env, db, blobid,
(ham_size_t)blob_get_alloc_size(&hdr), HAM_FALSE);
ham_assert(!st, ("unexpected error, at least not covered in the old code"));
return st;
}
ham_status_t
blob_overwrite(ham_env_t *env, ham_db_t *db, ham_offset_t old_blobid,
ham_record_t *record, ham_u32_t flags, ham_offset_t *new_blobid)
{
ham_status_t st;
ham_size_t alloc_size;
blob_t old_hdr;
blob_t new_hdr;
ham_page_t *page;
/*
* PARTIAL WRITE
*
* if offset+partial_size equals the full record size, then we won't
* have any gaps. In this case we just write the full record and ignore
* the partial parameters.
*/
if (flags&HAM_PARTIAL) {
if (record->partial_offset==0
&& record->partial_offset+record->partial_size==record->size)
flags&=~HAM_PARTIAL;
}
/*
* inmemory-databases: free the old blob,
* allocate a new blob (but if both sizes are equal, just overwrite
* the data)
*/
if (env_get_rt_flags(env)&HAM_IN_MEMORY_DB)
{
blob_t *nhdr, *phdr=(blob_t *)U64_TO_PTR(old_blobid);
if (blob_get_size(phdr)==record->size) {
ham_u8_t *p=(ham_u8_t *)phdr;
if (flags&HAM_PARTIAL) {
memmove(p+sizeof(blob_t)+record->partial_offset,
record->data, record->partial_size);
}
else {
memmove(p+sizeof(blob_t), record->data, record->size);
}
*new_blobid=(ham_offset_t)PTR_TO_U64(phdr);
}
else {
st=blob_allocate(env, db, record, flags, new_blobid);
if (st)
return (st);
nhdr=(blob_t *)U64_TO_PTR(*new_blobid);
blob_set_flags(nhdr, blob_get_flags(phdr));
allocator_free(env_get_allocator(env), phdr);
}
return (HAM_SUCCESS);
}
ham_assert(old_blobid%DB_CHUNKSIZE==0, (0));
/*
* blobs are CHUNKSIZE-allocated
*/
alloc_size=sizeof(blob_t)+record->size;
alloc_size += DB_CHUNKSIZE - 1;
alloc_size -= alloc_size % DB_CHUNKSIZE;
/*
* first, read the blob header; if the new blob fits into the
* old blob, we overwrite the old blob (and add the remaining
* space to the freelist, if there is any)
*/
st=__read_chunk(env, 0, &page, old_blobid, (ham_u8_t *)&old_hdr,
sizeof(old_hdr));
if (st)
return (st);
ham_assert(blob_get_alloc_size(&old_hdr)%DB_CHUNKSIZE==0, (0));
/*
* sanity check
*/
ham_verify(blob_get_self(&old_hdr)==old_blobid,
("invalid blobid %llu != %llu", blob_get_self(&old_hdr),
old_blobid));
if (blob_get_self(&old_hdr)!=old_blobid)
return (HAM_BLOB_NOT_FOUND);
/*
* now compare the sizes; does the new data fit in the old allocated
* space?
*/
if (alloc_size<=blob_get_alloc_size(&old_hdr))
{
ham_u8_t *chunk_data[2];
ham_size_t chunk_size[2];
/*
* setup the new blob header
*/
blob_set_self(&new_hdr, blob_get_self(&old_hdr));
blob_set_size(&new_hdr, record->size);
blob_set_flags(&new_hdr, blob_get_flags(&old_hdr));
if (blob_get_alloc_size(&old_hdr)-alloc_size>SMALLEST_CHUNK_SIZE)
blob_set_alloc_size(&new_hdr, alloc_size);
else
blob_set_alloc_size(&new_hdr, blob_get_alloc_size(&old_hdr));
/*
* PARTIAL WRITE
*
* if we have a gap at the beginning, then we have to write the
* blob header and the blob data in two steps; otherwise we can
* write both immediately
*/
if ((flags&HAM_PARTIAL) && (record->partial_offset)) {
chunk_data[0]=(ham_u8_t *)&new_hdr;
chunk_size[0]=sizeof(new_hdr);
st=__write_chunks(env, page, blob_get_self(&new_hdr), HAM_FALSE,
HAM_FALSE, chunk_data, chunk_size, 1);
if (st)
return (st);
chunk_data[0]=record->data;
chunk_size[0]=record->partial_size;
st=__write_chunks(env, page,
blob_get_self(&new_hdr)+sizeof(new_hdr)
+record->partial_offset,
HAM_FALSE, HAM_FALSE, chunk_data, chunk_size, 1);
if (st)
return (st);
}
else {
chunk_data[0]=(ham_u8_t *)&new_hdr;
chunk_size[0]=sizeof(new_hdr);
chunk_data[1]=record->data;
chunk_size[1]=(flags&HAM_PARTIAL)
? record->partial_size
: record->size;
st=__write_chunks(env, page, blob_get_self(&new_hdr), HAM_FALSE,
HAM_FALSE, chunk_data, chunk_size, 2);
if (st)
return (st);
}
/*
* move remaining data to the freelist
*/
if (blob_get_alloc_size(&old_hdr)!=blob_get_alloc_size(&new_hdr)) {
(void)freel_mark_free(env, db,
blob_get_self(&new_hdr)+blob_get_alloc_size(&new_hdr),
(ham_size_t)(blob_get_alloc_size(&old_hdr)-
blob_get_alloc_size(&new_hdr)), HAM_FALSE);
}
/*
* the old rid is the new rid
*/
*new_blobid=blob_get_self(&new_hdr);
return (HAM_SUCCESS);
}
else {
/*
* when the new data is larger, allocate a fresh space for it
* and discard the old;
'overwrite' has become (delete + insert) now.
*/
st=blob_allocate(env, db, record, flags, new_blobid);
if (st)
return (st);
(void)freel_mark_free(env, db, old_blobid,
(ham_size_t)blob_get_alloc_size(&old_hdr), HAM_FALSE);
}
return (HAM_SUCCESS);
}
/**
* Allocate space in storage for and write the content references by 'data'
* (and length 'size') to storage.
*
* Conditions will apply whether the data is written through cache or direct
* to device.
*
* The content is, of course, prefixed by a BLOB header.
*
* Partial writes are handled in this function.
*/
ham_status_t
blob_allocate(ham_env_t *env, ham_db_t *db, ham_record_t *record,
ham_u32_t flags, ham_offset_t *blobid)
{
ham_status_t st;
ham_page_t *page=0;
ham_offset_t addr;
blob_t hdr;
ham_u8_t *chunk_data[2];
ham_size_t alloc_size;
ham_size_t chunk_size[2];
ham_device_t *device=env_get_device(env);
ham_bool_t freshly_created = HAM_FALSE;
*blobid=0;
/*
* PARTIAL WRITE
*
* if offset+partial_size equals the full record size, then we won't
* have any gaps. In this case we just write the full record and ignore
* the partial parameters.
*/
if (flags&HAM_PARTIAL) {
if (record->partial_offset==0
&& record->partial_offset+record->partial_size==record->size)
flags&=~HAM_PARTIAL;
}
/*
* in-memory-database: the blobid is actually a pointer to the memory
* buffer, in which the blob (with the blob-header) is stored
*/
if (env_get_rt_flags(env)&HAM_IN_MEMORY_DB) {
blob_t *hdr;
ham_u8_t *p=(ham_u8_t *)allocator_alloc(env_get_allocator(env),
record->size+sizeof(blob_t));
if (!p) {
return HAM_OUT_OF_MEMORY;
}
/* initialize the header */
hdr=(blob_t *)p;
memset(hdr, 0, sizeof(*hdr));
blob_set_self(hdr, (ham_offset_t)PTR_TO_U64(p));
blob_set_alloc_size(hdr, record->size+sizeof(blob_t));
blob_set_size(hdr, record->size);
/* do we have gaps? if yes, fill them with zeroes */
if (flags&HAM_PARTIAL) {
ham_u8_t *s=p+sizeof(blob_t);
if (record->partial_offset)
memset(s, 0, record->partial_offset);
memcpy(s+record->partial_offset,
record->data, record->partial_size);
if (record->partial_offset+record->partial_size<record->size)
memset(s+record->partial_offset+record->partial_size, 0,
record->size-(record->partial_offset+record->partial_size));
}
else {
memcpy(p+sizeof(blob_t), record->data, record->size);
}
*blobid=(ham_offset_t)PTR_TO_U64(p);
return (0);
}
memset(&hdr, 0, sizeof(hdr));
/*
* blobs are CHUNKSIZE-allocated
*/
alloc_size=sizeof(blob_t)+record->size;
alloc_size += DB_CHUNKSIZE - 1;
alloc_size -= alloc_size % DB_CHUNKSIZE;
/*
* check if we have space in the freelist
*/
st = freel_alloc_area(&addr, env, db, alloc_size);
if (!addr)
{
if (st)
return st;
/*
* if the blob is small AND if logging is disabled: load the page
* through the cache
*/
if (__blob_from_cache(env, alloc_size)) {
st = db_alloc_page(&page, db, PAGE_TYPE_BLOB,
PAGE_IGNORE_FREELIST);
ham_assert(st ? page == NULL : 1, (0));
ham_assert(!st ? page != NULL : 1, (0));
if (st)
return st;
/* blob pages don't have a page header */
page_set_npers_flags(page,
page_get_npers_flags(page)|PAGE_NPERS_NO_HEADER);
addr=page_get_self(page);
/* move the remaining space to the freelist */
(void)freel_mark_free(env, db, addr+alloc_size,
env_get_pagesize(env)-alloc_size, HAM_FALSE);
blob_set_alloc_size(&hdr, alloc_size);
}
else {
/*
* otherwise use direct IO to allocate the space
*/
ham_size_t aligned=alloc_size;
aligned += env_get_pagesize(env) - 1;
aligned -= aligned % env_get_pagesize(env);
st=device->alloc(device, aligned, &addr);
if (st)
return (st);
/* if aligned!=size, and the remaining chunk is large enough:
* move it to the freelist */
{
ham_size_t diff=aligned-alloc_size;
if (diff > SMALLEST_CHUNK_SIZE) {
(void)freel_mark_free(env, db, addr+alloc_size,
diff, HAM_FALSE);
blob_set_alloc_size(&hdr, aligned-diff);
}
else {
blob_set_alloc_size(&hdr, aligned);
}
}
freshly_created = HAM_TRUE;
}
ham_assert(HAM_SUCCESS == freel_check_area_is_allocated(env, db,
addr, alloc_size), (0));
}
else {
ham_assert(!st, (0));
blob_set_alloc_size(&hdr, alloc_size);
}
blob_set_size(&hdr, record->size);
blob_set_self(&hdr, addr);
/*
* PARTIAL WRITE
*
* are there gaps at the beginning? If yes, then we'll fill with zeros
*/
if ((flags&HAM_PARTIAL) && (record->partial_offset)) {
ham_u8_t *ptr;
ham_size_t gapsize=record->partial_offset;
ptr=allocator_calloc(env_get_allocator(env),
gapsize > env_get_pagesize(env)
? env_get_pagesize(env)
: gapsize);
if (!ptr)
return (HAM_OUT_OF_MEMORY);
/*
* first: write the header
*/
chunk_data[0]=(ham_u8_t *)&hdr;
chunk_size[0]=sizeof(hdr);
st=__write_chunks(env, page, addr, HAM_TRUE, freshly_created,
chunk_data, chunk_size, 1);
if (st)
return (st);
addr+=sizeof(hdr);
/* now fill the gap; if the gap is bigger than a pagesize we'll
* split the gap into smaller chunks
*/
while (gapsize>=env_get_pagesize(env)) {
chunk_data[0]=ptr;
chunk_size[0]=env_get_pagesize(env);
st=__write_chunks(env, page, addr, HAM_TRUE,
freshly_created, chunk_data, chunk_size, 1);
if (st)
break;
gapsize-=env_get_pagesize(env);
addr+=env_get_pagesize(env);
}
/* fill the remaining gap */
if (gapsize) {
chunk_data[0]=ptr;
chunk_size[0]=gapsize;
st=__write_chunks(env, page, addr, HAM_TRUE, freshly_created,
chunk_data, chunk_size, 1);
if (st)
return (st);
addr+=gapsize;
}
allocator_free(env_get_allocator(env), ptr);
/* now write the "real" data */
chunk_data[0]=(ham_u8_t *)record->data;
chunk_size[0]=record->partial_size;
st=__write_chunks(env, page, addr, HAM_TRUE, freshly_created,
chunk_data, chunk_size, 1);
if (st)
return (st);
addr+=record->partial_size;
}
else {
/*
* not writing partially: write header and data, then we're done
*/
chunk_data[0]=(ham_u8_t *)&hdr;
chunk_size[0]=sizeof(hdr);
chunk_data[1]=(ham_u8_t *)record->data;
chunk_size[1]=(flags&HAM_PARTIAL)
? record->partial_size
: record->size;
st=__write_chunks(env, page, addr, HAM_TRUE, freshly_created,
chunk_data, chunk_size, 2);
if (st)
return (st);
addr+=sizeof(hdr)+
((flags&HAM_PARTIAL) ? record->partial_size : record->size);
}
/*
* store the blobid; it will be returned to the caller
*/
*blobid=blob_get_self(&hdr);
/*
* PARTIAL WRITES:
*
* if we have gaps at the end of the blob: just append more chunks to
* fill these gaps. Since they can be pretty large we split them into
* smaller chunks if necessary.
*/
if (flags&HAM_PARTIAL) {
if (record->partial_offset+record->partial_size < record->size) {
ham_u8_t *ptr;
ham_size_t gapsize=record->size
- (record->partial_offset+record->partial_size);
/* now fill the gap; if the gap is bigger than a pagesize we'll
* split the gap into smaller chunks
*
* we split this loop in two - the outer loop will allocate the
* memory buffer, thus saving some allocations
*/
while (gapsize>env_get_pagesize(env)) {
ham_u8_t *ptr=allocator_calloc(env_get_allocator(env),
env_get_pagesize(env));
if (!ptr)
return (HAM_OUT_OF_MEMORY);
while (gapsize>env_get_pagesize(env)) {
chunk_data[0]=ptr;
chunk_size[0]=env_get_pagesize(env);
st=__write_chunks(env, page, addr, HAM_TRUE,
freshly_created, chunk_data, chunk_size, 1);
if (st)
break;
gapsize-=env_get_pagesize(env);
addr+=env_get_pagesize(env);
}
allocator_free(env_get_allocator(env), ptr);
if (st)
return (st);
}
/* now write the remainder, which is less than a pagesize */
ham_assert(gapsize<env_get_pagesize(env), (""));
chunk_size[0]=gapsize;
ptr=chunk_data[0]=allocator_calloc(env_get_allocator(env), gapsize);
if (!ptr)
return (HAM_OUT_OF_MEMORY);
st=__write_chunks(env, page, addr, HAM_TRUE, freshly_created,
chunk_data, chunk_size, 1);
allocator_free(env_get_allocator(env), ptr);
if (st)
return (st);
}
}
return (0);
}
ham_status_t
txn_abort(ham_txn_t *txn, ham_u32_t flags)
{
ham_status_t st;
ham_env_t *env=txn_get_env(txn);
/*
* are cursors attached to this txn? if yes, fail
*/
if (txn_get_cursor_refcount(txn)) {
ham_trace(("transaction cannot be aborted till all attached "
"cursors are closed"));
return HAM_CURSOR_STILL_OPEN;
}
if (env_get_log(env) && !(txn_get_flags(txn)&HAM_TXN_READ_ONLY)) {
st=ham_log_append_txn_abort(env_get_log(env), txn);
if (st)
return st;
}
env_set_txn(env, 0);
/*
* undo all operations from this transaction
*
* this includes allocated pages (they're moved to the freelist),
* deleted pages (they're un-deleted) and other modifications (will
* re-create the original page from the logfile)
*
* keep txn_get_pagelist(txn) intact during every round, so no
* local var for this one.
*/
while (txn_get_pagelist(txn)) {
ham_page_t *head = txn_get_pagelist(txn);
if (!(flags & DO_NOT_NUKE_PAGE_STATS)) {
/*
* nuke critical statistics, such as tracked outer bounds; imagine,
* for example, a failing erase transaction which, through erasing
* the top-most key, lowers the actual upper bound, after which
* the transaction fails at some later point in life. Now if we
* wouldn't 'rewind' our bounds-statistics, we would have a
* situation where a subsequent out-of-bounds insert (~ append)
* would possibly FAIL due to the hinter using incorrect bounds
* information then!
*
* Hence we 'reverse' our statistics here and the easiest route
* is to just nuke the critical bits; subsequent find/insert/erase
* operations will ensure that the stats will get updated again,
* anyhow. All we loose then is a few subsequent operations, which
* might have been hinted if we had played a smarter game of
* statistics 'reversal'. Soit.
*/
ham_db_t *db = page_get_owner(head);
/*
* only need to do this for index pages anyhow, and those are the
* ones which have their 'ownership' set.
*/
if (db) {
stats_page_is_nuked(db, head, HAM_FALSE);
}
}
ham_assert(page_is_in_list(txn_get_pagelist(txn), head, PAGE_LIST_TXN),
(0));
txn_get_pagelist(txn) = page_list_remove(head, PAGE_LIST_TXN, head);
/* if this page was allocated by this transaction, then we can
* move the whole page to the freelist */
if (page_get_alloc_txn_id(head)==txn_get_id(txn)) {
(void)freel_mark_free(env, 0, page_get_self(head),
env_get_pagesize(env), HAM_TRUE);
}
else {
/* remove the 'delete pending' flag */
page_set_npers_flags(head,
page_get_npers_flags(head)&~PAGE_NPERS_DELETE_PENDING);
/* if the page is dirty, and RECOVERY is enabled: recreate
* the original, unmodified page from the log */
if (env_get_log(env) && page_is_dirty(head)) {
st=ham_log_recreate(env_get_log(env), head);
if (st)
return (st);
/*page_set_undirty(head); */
}
}
/* page is no longer in use */
page_release_ref(head);
}
ham_assert(txn_get_pagelist(txn)==0, (0));
return (0);
}
