/*
* __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);
}
/*
* __wt_merge_tree --
* Attempt to collapse a stack of split-merge pages in memory into a
* shallow tree. If enough keys are found, create a real internal node
* that can be evicted (and, if necessary, split further).
*
* This code is designed to deal with workloads that otherwise create
* arbitrarily deep (and slow) trees in memory.
*/
int
__wt_merge_tree(WT_SESSION_IMPL *session, WT_PAGE *top)
{
WT_DECL_RET;
WT_PAGE *lchild, *newtop, *rchild;
WT_REF *newref;
WT_VISIT_STATE visit_state;
uint32_t refcnt, split;
int promote;
u_int levels;
uint8_t page_type;
WT_CLEAR(visit_state);
visit_state.session = session;
lchild = newtop = rchild = NULL;
page_type = top->type;
WT_ASSERT(session, __wt_btree_mergeable(top));
WT_ASSERT(session, top->ref->state == WT_REF_LOCKED);
/*
* Walk the subtree, count the references at the bottom level and
* calculate the maximum depth.
*/
WT_RET(__merge_walk(session, top, 1, __merge_count, &visit_state));
/* If there aren't enough useful levels, give up. */
if (visit_state.maxdepth < WT_MERGE_STACK_MIN)
return (EBUSY);
/*
* Don't allow split merges to generate arbitrarily large pages.
* Ideally we would choose a size based on the internal_page_max
* setting for the btree, but we don't have the correct btree handle
* available.
*/
if (visit_state.refcnt > WT_MERGE_MAX_REFS)
return (EBUSY);
/*
* Now we either collapse the internal pages into one split-merge page,
* or if there are "enough" keys, we split into two equal internal
* pages, each of which can be evicted independently.
*
* We set a flag (WT_PM_REC_SPLIT_MERGE) on the created page if it
* isn't big enough to justify the cost of evicting it. If splits
* continue, it will be merged again until it gets over this limit.
*/
promote = 0;
refcnt = (uint32_t)visit_state.refcnt;
if (refcnt >= WT_MERGE_FULL_PAGE && visit_state.seen_live) {
/*
* In the normal case where there are live children spread
* through the subtree, create two child pages.
*
* Handle the case where the only live child is first / last
* specially: put the live child into the top-level page.
*
* Set SPLIT_MERGE on the internal pages if there are any live
* children: they can't be evicted, so there is no point
* permanently deepening the tree.
*/
if (visit_state.first_live == visit_state.last_live &&
(visit_state.first_live == 0 ||
visit_state.first_live == refcnt - 1))
split = (visit_state.first_live == 0) ? 1 : refcnt - 1;
else
split = (refcnt + 1) / 2;
/* Only promote if we can create a real page. */
if (split == 1 || split == refcnt - 1)
promote = 1;
else if (split >= WT_MERGE_FULL_PAGE &&
visit_state.first_live >= split)
promote = 1;
else if (refcnt - split >= WT_MERGE_FULL_PAGE &&
visit_state.last_live < split)
promote = 1;
}
if (promote) {
/* Create a new top-level split-merge page with two entries. */
WT_ERR(__merge_new_page(session, page_type, 2, 1, &newtop));
visit_state.split = split;
/* Left split. */
if (split == 1)
visit_state.first = newtop;
else {
WT_ERR(__merge_new_page(session, page_type, split,
visit_state.first_live < split, &lchild));
visit_state.first = lchild;
}
/* Right split. */
if (split == refcnt - 1) {
visit_state.second = newtop;
visit_state.second_ref = &newtop->u.intl.t[1];
} else {
WT_ERR(__merge_new_page(session, page_type,
refcnt - split, visit_state.last_live >= split,
&rchild));
visit_state.second = rchild;
visit_state.second_ref =
&visit_state.second->u.intl.t[0];
}
} else {
/*
* Create a new split-merge page for small merges, or if the
* page above is a split merge page. When we do a big enough
* merge, we create a real page at the top and don't consider
* it as a merge candidate again. Over time with an insert
* workload the tree will grow deeper, but that's inevitable,
* and this keeps individual merges small.
*/
WT_ERR(__merge_new_page(session, page_type, refcnt,
refcnt < WT_MERGE_FULL_PAGE ||
__wt_btree_mergeable(top->parent),
&newtop));
visit_state.first = newtop;
}
/*
* Copy the references into the new tree, but don't update anything in
* the locked tree in case there is an error and we need to back out.
* We do this in a separate pass so that we can figure out the key for
* the split point: that allocates memory and so it could still fail.
*/
visit_state.page = visit_state.first;
visit_state.ref = visit_state.page->u.intl.t;
visit_state.refcnt = 0;
WT_ERR(__merge_walk(session, top, 0, __merge_copy_ref, &visit_state));
if (promote) {
/* Promote keys into the top-level page. */
if (lchild != NULL) {
newref = &newtop->u.intl.t[0];
WT_LINK_PAGE(newtop, newref, lchild);
newref->state = WT_REF_MEM;
WT_ERR(__merge_promote_key(session, newref));
}
if (rchild != NULL) {
newref = &newtop->u.intl.t[1];
WT_LINK_PAGE(newtop, newref, rchild);
newref->state = WT_REF_MEM;
WT_ERR(__merge_promote_key(session, newref));
}
}
/*
* We have copied everything into place and allocated all of the memory
* we need. Now link all pages into the new tree and unlock them.
*
* The only way this could fail is if a reference state has been
* changed by another thread since they were locked. Panic in that
* case: that should never happen.
*/
visit_state.page = visit_state.first;
visit_state.ref = visit_state.page->u.intl.t;
visit_state.refcnt = 0;
ret = __merge_walk(session, top, 0, __merge_switch_page, &visit_state);
if (ret != 0)
WT_ERR(__wt_illegal_value(session, "__wt_merge_tree"));
newtop->u.intl.recno = top->u.intl.recno;
newtop->parent = top->parent;
newtop->ref = top->ref;
#ifdef HAVE_DIAGNOSTIC
/*
* Before swapping in the new tree, walk the pages we are discarding,
* check that everything looks right.
*/
__merge_check_discard(session, top);
#endif
/*
* Set up the new top-level page as a split so that it will be swapped
* into place by our caller.
*/
top->modify->flags = WT_PM_REC_SPLIT;
top->modify->u.split = newtop;
WT_VERBOSE_ERR(session, evict,
"Successfully %s %" PRIu32
" split-merge pages containing %" PRIu32 " keys\n",
promote ? "promoted" : "merged", visit_state.maxdepth, refcnt);
/* Evict new child pages as soon as possible. */
if (lchild != NULL && !F_ISSET(lchild->modify, WT_PM_REC_SPLIT_MERGE))
lchild->read_gen = WT_READ_GEN_OLDEST;
if (rchild != NULL && !F_ISSET(rchild->modify, WT_PM_REC_SPLIT_MERGE))
rchild->read_gen = WT_READ_GEN_OLDEST;
/* Update statistics. */
WT_CSTAT_INCR(session, cache_eviction_merge);
WT_DSTAT_INCR(session, cache_eviction_merge);
/* How many levels did we remove? */
levels = visit_state.maxdepth - (promote ? 2 : 1);
WT_CSTAT_INCRV(session, cache_eviction_merge_levels, levels);
WT_DSTAT_INCRV(session, cache_eviction_merge_levels, levels);
return (0);
err: WT_VERBOSE_TRET(session, evict,
"Failed to merge %" PRIu32
" split-merge pages containing %" PRIu32 " keys\n",
visit_state.maxdepth, refcnt);
WT_CSTAT_INCR(session, cache_eviction_merge_fail);
WT_DSTAT_INCR(session, cache_eviction_merge_fail);
if (newtop != NULL)
__wt_page_out(session, &newtop);
if (lchild != NULL)
__wt_page_out(session, &lchild);
if (rchild != NULL)
__wt_page_out(session, &rchild);
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, uint32_t *addr_size, int checkpoint, int compressed)
{
WT_BM *bm;
WT_BTREE *btree;
WT_ITEM *ip;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t len, src_len, dst_len, result_len, size;
int data_cksum, compression_failed;
uint8_t *src, *dst;
btree = session->btree;
bm = btree->bm;
/* Checkpoint calls are different than standard calls. */
WT_ASSERT(session,
(checkpoint == 0 && addr != NULL && addr_size != NULL) ||
(checkpoint == 1 && addr == NULL && addr_size == 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, &tmp));
memcpy(tmp->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 *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->memsize - WT_BLOCK_COMPRESS_SKIP,
&result_len));
WT_ASSERT(session,
dsk->mem_size == result_len + WT_BLOCK_COMPRESS_SKIP);
tmp->size = (uint32_t)result_len + WT_BLOCK_COMPRESS_SKIP;
ip = tmp;
} else {
WT_ASSERT(session, dsk->mem_size == buf->size);
ip = buf;
}
WT_ERR(__wt_verify_dsk(session, "[write-check]", ip));
__wt_scr_free(&tmp);
#endif
/*
* Optionally stream-compress the data, but don't compress blocks that
* are already as small as they're going to get.
*/
if (buf->size <= btree->allocsize ||
btree->compressor == NULL ||
btree->compressor->compress == NULL || compressed) {
ip = buf;
WT_DSTAT_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, &tmp));
/* Skip the header bytes of the destination data. */
dst = (uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP;
dst_len = len;
/*
* If compression fails, fallback to the original version. This
* isn't unexpected: if compression doesn't work for some chunk
* of bytes for some reason (noting there's 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.
*/
compression_failed = 0;
WT_ERR(btree->compressor->compress(btree->compressor,
&session->iface,
src, src_len,
dst, dst_len,
&result_len, &compression_failed));
if (compression_failed) {
ip = buf;
WT_DSTAT_INCR(session, compress_write_fail);
} else {
compressed = 1;
WT_DSTAT_INCR(session, compress_write);
/*
* Copy in the skipped header bytes, set the final data
* size.
*/
memcpy(tmp->mem, buf->mem, WT_BLOCK_COMPRESS_SKIP);
tmp->size =
(uint32_t)result_len + WT_BLOCK_COMPRESS_SKIP;
ip = tmp;
}
}
dsk = ip->mem;
/* If the buffer is compressed, set the flag. */
if (compressed)
F_SET(dsk, WT_PAGE_COMPRESSED);
/*
* 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_size, data_cksum));
WT_CSTAT_INCR(session, cache_write);
WT_DSTAT_INCR(session, cache_write);
WT_CSTAT_INCRV(session, cache_bytes_write, ip->size);
WT_DSTAT_INCRV(session, cache_bytes_write, ip->size);
err: __wt_scr_free(&tmp);
return (ret);
}
