/*
* __compact_rewrite --
* Return if a page needs to be re-written.
*/
static int
__compact_rewrite(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_BM *bm;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
size_t addr_size;
const uint8_t *addr;
*skipp = true; /* Default skip. */
bm = S2BT(session)->bm;
page = ref->page;
mod = page->modify;
/*
* Ignore the root: it may not have a replacement address, and besides,
* if anything else gets written, so will it.
*/
if (__wt_ref_is_root(ref))
return (0);
/* Ignore currently dirty pages, they will be written regardless. */
if (__wt_page_is_modified(page))
return (0);
/*
* If the page is clean, test the original addresses.
* If the page is a 1-to-1 replacement, test the replacement addresses.
* Ignore empty pages, they get merged into the parent.
*/
if (mod == NULL || mod->rec_result == 0) {
WT_RET(__wt_ref_info(session, ref, &addr, &addr_size, NULL));
if (addr == NULL)
return (0);
WT_RET(
bm->compact_page_skip(bm, session, addr, addr_size, skipp));
} else if (mod->rec_result == WT_PM_REC_REPLACE) {
/*
* The page's modification information can change underfoot if
* the page is being reconciled, serialize with reconciliation.
*/
WT_RET(__wt_fair_lock(session, &page->page_lock));
ret = bm->compact_page_skip(bm, session,
mod->mod_replace.addr, mod->mod_replace.size, skipp);
WT_TRET(__wt_fair_unlock(session, &page->page_lock));
WT_RET(ret);
}
return (0);
}
/*
* __wt_page_addr_string --
* Figure out a page's "address" and load a buffer with a printable,
* nul-terminated representation of that address.
*/
const char *
__wt_page_addr_string(WT_SESSION_IMPL *session, WT_REF *ref, WT_ITEM *buf)
{
size_t addr_size;
const uint8_t *addr;
if (__wt_ref_is_root(ref)) {
buf->data = "[Root]";
buf->size = strlen("[Root]");
return (buf->data);
}
(void)__wt_ref_info(session, ref, &addr, &addr_size, NULL);
return (__wt_addr_string(session, addr, addr_size, buf));
}
/*
* __split_should_deepen --
* Return if we should deepen the tree.
*/
static bool
__split_should_deepen(WT_SESSION_IMPL *session, WT_REF *ref)
{
WT_BTREE *btree;
WT_PAGE *page;
WT_PAGE_INDEX *pindex;
btree = S2BT(session);
page = ref->page;
/*
* Our caller is holding the parent page locked to single-thread splits,
* which means we can safely look at the page's index without setting a
* split generation.
*/
pindex = WT_INTL_INDEX_GET_SAFE(page);
/*
* Deepen the tree if the page's memory footprint is larger than the
* maximum size for a page in memory (presumably putting eviction
* pressure on the cache).
*/
if (page->memory_footprint < btree->maxmempage)
return (false);
/*
* Ensure the page has enough entries to make it worth splitting and
* we get a significant payback (in the case of a set of large keys,
* splitting won't help).
*/
if (pindex->entries > btree->split_deepen_min_child)
return (true);
/*
* Don't allow a single page to put pressure on cache usage. The root
* page cannot be evicted, so if it's larger than the maximum, or if
* and page has a quarter of the cache, let it split, a deep tree is
* better than making no progress at all. Sanity check for 100 on-page
* keys, nothing helps in the case of large keys and a too-small cache.
*/
if (pindex->entries >= 100 &&
(__wt_ref_is_root(ref) ||
page->memory_footprint >= S2C(session)->cache_size / 4))
return (true);
return (false);
}
/*
* __sync_dup_walk --
* Duplicate a tree walk point.
*/
static inline int
__sync_dup_walk(
WT_SESSION_IMPL *session, WT_REF *walk, uint32_t flags, WT_REF **dupp)
{
WT_REF *old;
bool busy;
if ((old = *dupp) != NULL) {
*dupp = NULL;
WT_RET(__wt_page_release(session, old, flags));
}
/* It is okay to duplicate a walk before it starts. */
if (walk == NULL || __wt_ref_is_root(walk)) {
*dupp = walk;
return (0);
}
/* Get a duplicate hazard pointer. */
for (;;) {
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, walk, &busy, __func__, __LINE__));
#else
WT_RET(__wt_hazard_set(session, walk, &busy));
#endif
/*
* We already have a hazard pointer, we should generally be able
* to get another one. We can get spurious busy errors (e.g., if
* eviction is attempting to lock the page. Keep trying: we have
* one hazard pointer so we should be able to get another one.
*/
if (!busy)
break;
__wt_yield();
}
*dupp = walk;
return (0);
}
/*将多余的文件空间compact到合适的位置,如果ref在compact范围内，返回skip = 1,表示文件空间不能进行compact*/
static int __compact_rewrite(WT_SESSION_IMPL* session, WT_REF* ref, int* skipp)
{
WT_BM *bm;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
size_t addr_size;
const uint8_t *addr;
*skipp = 1;
bm = S2BT(session)->bm;
page = ref->page;
mod = page->modify;
/*root page是不能被compact*/
if (__wt_ref_is_root(ref))
return 0;
/*ref指向的是个脏页，不进行compact*/
if (__wt_page_is_modified(page))
return (0);
/*假如page一已经被清空的，直接判断是否可以它的block空间compact*/
if (mod == NULL || F_ISSET(mod, WT_PM_REC_MASK) == 0) {
WT_RET(__wt_ref_info(session, ref, &addr, &addr_size, NULL));
if (addr == NULL)
return (0);
WT_RET(bm->compact_page_skip(bm, session, addr, addr_size, skipp));
}
else if (F_ISSET(mod, WT_PM_REC_MASK) == WT_PM_REC_REPLACE){ /*如果page空间是替换，那么进行替换block的compact操作判断*/
WT_PAGE_LOCK(session, page);
ret = bm->compact_page_skip(bm, session, mod->mod_replace.addr, mod->mod_replace.size, skipp);
WT_PAGE_UNLOCK(session, page);
WT_RET(ret);
}
return 0;
}
/*
* __tree_walk_internal --
* Move to the next/previous page in the tree.
*/
static inline int
__tree_walk_internal(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp,
int (*skip_func)(WT_SESSION_IMPL *, WT_REF *, void *, bool *),
void *func_cookie, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
uint32_t slot;
bool empty_internal, initial_descent, prev, skip;
btree = S2BT(session);
pindex = NULL;
empty_internal = initial_descent = false;
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/* Walk should never instantiate deleted pages. */
LF_SET(WT_READ_NO_EMPTY);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* Clear the returned value, it makes future error handling easier.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start/end of the tree. */
if (ref == NULL) {
restart: /*
* We can be here with a NULL or root WT_REF; the page release
* function handles them internally, don't complicate this code
* by calling them out.
*/
WT_ERR(__wt_page_release(session, couple, flags));
/*
* We're not supposed to walk trees without root pages. As this
* has not always been the case, assert to debug that change.
*/
WT_ASSERT(session, btree->root.page != NULL);
couple = couple_orig = ref = &btree->root;
initial_descent = true;
goto descend;
}
/*
* If the active page was the root, we've reached the walk's end; we
* only get here if we've returned the root to our caller, so we're
* holding no hazard pointers.
*/
if (__wt_ref_is_root(ref))
goto done;
/* Figure out the current slot in the WT_REF array. */
__ref_index_slot(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the internal page, return
* it in post-order traversal. Otherwise move to the next/prev
* slot and left/right-most element in that subtree.
*/
while ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
/* Ascend to the parent. */
__ref_ascend(session, &ref, &pindex, &slot);
/*
* If at the root and returning internal pages, return
* the root page, otherwise we're done. Regardless, no
* hazard pointer is required, release the one we hold.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(
session, couple, flags));
if (!LF_ISSET(WT_READ_SKIP_INTL))
*refp = ref;
goto done;
}
/*
* If we got all the way through an internal page and
* all of the child pages were deleted, mark it for
* eviction.
*/
if (empty_internal && pindex->entries > 1) {
__wt_page_evict_soon(session, ref);
empty_internal = false;
}
/*
* Optionally return internal pages. Swap our previous
* hazard pointer for the page we'll return. We don't
* handle restart or not-found returns, it would require
* additional complexity and is not a possible return:
* we're moving to the parent of the current child page,
* the parent can't have been evicted.
*/
if (!LF_ISSET(WT_READ_SKIP_INTL)) {
WT_ERR(__wt_page_swap(
session, couple, ref, flags));
*refp = ref;
goto done;
}
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
/*
* Move to the next slot, and set the reference hint if
* it's wrong (used when we continue the walk). We don't
* always update the hints when splitting, it's expected
* for them to be incorrect in some workloads.
*/
ref = pindex->index[slot];
if (ref->pindex_hint != slot)
ref->pindex_hint = slot;
/*
* If we see any child states other than deleted, the
* page isn't empty.
*/
if (ref->state != WT_REF_DELETED &&
!LF_ISSET(WT_READ_TRUNCATE))
empty_internal = false;
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
/* Skip lookaside pages if not requested. */
if (ref->state == WT_REF_LOOKASIDE &&
!LF_ISSET(WT_READ_LOOKASIDE))
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
empty_internal = false;
} else if (skip_func != NULL) {
WT_ERR(skip_func(session,
ref, func_cookie, &skip));
if (skip)
break;
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
}
ret = __wt_page_swap(session, couple, ref,
WT_READ_NOTFOUND_OK | WT_READ_RESTART_OK | flags);
/*
* Not-found is an expected return when only walking
* in-cache pages, or if we see a deleted page.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a cursor is setting up at the end of the
* tree, we can't use our parent page's index,
* because it may have already split; restart
* the walk.
*/
if (prev && initial_descent)
goto restart;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root)
goto restart;
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__ref_index_slot(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
couple = ref;
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
if (WT_PAGE_IS_INTERNAL(ref->page)) {
descend: empty_internal = true;
/*
* There's a split race when a cursor is setting
* up at the end of the tree or moving backwards
* through the tree and descending a level. When
* splitting an internal page into its parent,
* we move the WT_REF structures and update the
* parent's page index before updating the split
* page's page index, and it's not an atomic
* update. A thread can read the parent page's
* replacement page index, then read the split
* page's original index, or the parent page's
* original and the split page's replacement.
*
* This isn't a problem for a cursor setting up
* at the start of the tree or moving forwards
* through the tree because we do right-hand
* splits on internal pages and the initial part
* of the split page's namespace won't change as
* part of a split. A thread reading the parent
* page's and split page's indexes will move to
* the same slot no matter what order of indexes
* are read.
*
* Handle a cursor setting up at the end of the
* tree or moving backwards through the tree.
*/
if (!prev) {
WT_INTL_INDEX_GET(
session, ref->page, pindex);
slot = 0;
} else if (initial_descent) {
if (!__ref_initial_descent_prev(
session, ref, &pindex))
goto restart;
slot = pindex->entries - 1;
} else {
__ref_descend_prev(
session, ref, &pindex);
slot = pindex->entries - 1;
}
continue;
}
/*
* The tree-walk restart code knows we return any leaf
* page we acquire (never hazard-pointer coupling on
* after acquiring a leaf page), and asserts no restart
* happens while holding a leaf page. This page must be
* returned to our caller.
*/
*refp = ref;
goto done;
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
return (ret);
}
/*
* __ref_ascend --
* Ascend the tree one level.
*/
static inline void
__ref_ascend(WT_SESSION_IMPL *session,
WT_REF **refp, WT_PAGE_INDEX **pindexp, uint32_t *slotp)
{
WT_REF *parent_ref, *ref;
/*
* Ref points to the first/last slot on an internal page from which we
* are ascending the tree, moving to the parent page. This is tricky
* because the internal page we're on may be splitting into its parent.
* Find a stable configuration where the page we start from and the
* page we're moving to are connected. The tree eventually stabilizes
* into that configuration, keep trying until we succeed.
*/
for (ref = *refp;;) {
/*
* Find our parent slot on the next higher internal page, the
* slot from which we move to a next/prev slot, checking that
* we haven't reached the root.
*/
parent_ref = ref->home->pg_intl_parent_ref;
if (__wt_ref_is_root(parent_ref))
break;
__ref_index_slot(session, parent_ref, pindexp, slotp);
/*
* There's a split race when a cursor moving forwards through
* the tree ascends the tree. If we're splitting an internal
* page into its parent, we move the WT_REF structures and
* then update the parent's page index before updating the split
* page's page index, and it's not an atomic update. A thread
* can read the split page's original page index and then read
* the parent page's replacement index.
*
* This can create a race for next-cursor movements.
*
* For example, imagine an internal page with 3 child pages,
* with the namespaces a-f, g-h and i-j; the first child page
* splits. The parent starts out with the following page-index:
*
* | ... | a | g | i | ... |
*
* which changes to this:
*
* | ... | a | c | e | g | i | ... |
*
* The split page starts out with the following page-index:
*
* | a | b | c | d | e | f |
*
* Imagine a cursor finishing the 'f' part of the namespace that
* starts its ascent to the parent's 'a' slot. Then the page
* splits and the parent page's page index is replaced. If the
* cursor then searches the parent's replacement page index for
* the 'a' slot, it finds it and then increments to the slot
* after the 'a' slot, the 'c' slot, and then it incorrectly
* repeats its traversal of part of the namespace.
*
* This function takes a WT_REF argument which is the page from
* which we start our ascent. If the parent's slot we find in
* our search doesn't point to the same page as that initial
* WT_REF, there's a race and we start over again.
*/
if (ref->home == parent_ref->page)
break;
}
*refp = parent_ref;
}
/*
* __wt_tree_walk --
* Move to the next/previous page in the tree.
*/
int
__wt_tree_walk(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
int prev, skip;
uint32_t slot;
btree = S2BT(session);
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* We may be passed a pointer to btree->evict_page that we are clearing
* here. We check when discarding pages that we're not discarding that
* page, so this clear must be done before the page is released.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start of the tree. */
if (ref == NULL) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
ascend: /*
* If the active page was the root, we've reached the walk's end.
* Release any hazard-pointer we're holding.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(session, couple, flags));
goto done;
}
/* Figure out the current slot in the WT_REF array. */
__wt_page_refp(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the page, return this page
* in post-order traversal. Otherwise we move to the next/prev
* slot and left/right-most element in its subtree.
*/
if ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
ref = ref->home->pg_intl_parent_ref;
/* Optionally skip internal pages. */
if (LF_ISSET(WT_READ_SKIP_INTL))
goto ascend;
/*
* We've ascended the tree and are returning an internal
* page. If it's the root, discard our hazard pointer,
* otherwise, swap our hazard pointer for the page we'll
* return.
*/
if (__wt_ref_is_root(ref))
WT_ERR(__wt_page_release(
session, couple, flags));
else {
/*
* Locate the reference to our parent page then
* swap our child hazard pointer for the parent.
* We don't handle restart or not-found returns.
* It would require additional complexity and is
* not a possible return: we're moving to the
* parent of the current child page, our parent
* reference can't have split or been evicted.
*/
__wt_page_refp(session, ref, &pindex, &slot);
if ((ret = __wt_page_swap(
session, couple, ref, flags)) != 0) {
WT_TRET(__wt_page_release(
session, couple, flags));
WT_ERR(ret);
}
}
*refp = ref;
goto done;
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
ref = pindex->index[slot];
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
} else if (LF_ISSET(WT_READ_COMPACT)) {
/*
* Skip deleted pages, rewriting them doesn't
* seem useful.
*/
if (ref->state == WT_REF_DELETED)
break;
/*
* If the page is in-memory, we want to look at
* it (it may have been modified and written,
* and the current location is the interesting
* one in terms of compaction, not the original
* location). If the page isn't in-memory, test
* if the page will help with compaction, don't
* read it if we don't have to.
*/
if (ref->state == WT_REF_DISK) {
WT_ERR(__wt_compact_page_skip(
session, ref, &skip));
if (skip)
break;
}
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref))
break;
}
ret = __wt_page_swap(session, couple, ref, flags);
/*
* Not-found is an expected return when only walking
* in-cache pages.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__wt_page_refp(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
descend: couple = ref;
page = ref->page;
if (page->type == WT_PAGE_ROW_INT ||
page->type == WT_PAGE_COL_INT) {
WT_INTL_INDEX_GET(session, page, pindex);
slot = prev ? pindex->entries - 1 : 0;
} else {
*refp = ref;
goto done;
}
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
return (ret);
}
/*
* __verify_tree --
* Verify a tree, recursively descending through it in depth-first fashion.
* The page argument was physically verified (so we know it's correctly formed),
* and the in-memory version built. Our job is to check logical relationships
* in the page and in the tree.
*/
static int
__verify_tree(WT_SESSION_IMPL *session, WT_REF *ref, WT_VSTUFF *vs)
{
WT_BM *bm;
WT_CELL *cell;
WT_CELL_UNPACK *unpack, _unpack;
WT_COL *cip;
WT_DECL_RET;
WT_PAGE *page;
WT_REF *child_ref;
uint64_t recno;
uint32_t entry, i;
bool found;
bm = S2BT(session)->bm;
page = ref->page;
unpack = &_unpack;
WT_CLEAR(*unpack); /* -Wuninitialized */
WT_RET(__wt_verbose(session, WT_VERB_VERIFY, "%s %s",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type)));
/* Optionally dump the address. */
if (vs->dump_address)
WT_RET(__wt_msg(session, "%s %s",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type)));
/* Track the shape of the tree. */
if (WT_PAGE_IS_INTERNAL(page))
++vs->depth_internal[
WT_MIN(vs->depth, WT_ELEMENTS(vs->depth_internal) - 1)];
else
++vs->depth_leaf[
WT_MIN(vs->depth, WT_ELEMENTS(vs->depth_internal) - 1)];
/*
* The page's physical structure was verified when it was read into
* memory by the read server thread, and then the in-memory version
* of the page was built. Now we make sure the page and tree are
* logically consistent.
*
* !!!
* The problem: (1) the read server has to build the in-memory version
* of the page because the read server is the thread that flags when
* any thread can access the page in the tree; (2) we can't build the
* in-memory version of the page until the physical structure is known
* to be OK, so the read server has to verify at least the physical
* structure of the page; (3) doing complete page verification requires
* reading additional pages (for example, overflow keys imply reading
* overflow pages in order to test the key's order in the page); (4)
* the read server cannot read additional pages because it will hang
* waiting on itself. For this reason, we split page verification
* into a physical verification, which allows the in-memory version
* of the page to be built, and then a subsequent logical verification
* which happens here.
*
* Report progress occasionally.
*/
#define WT_VERIFY_PROGRESS_INTERVAL 100
if (++vs->fcnt % WT_VERIFY_PROGRESS_INTERVAL == 0)
WT_RET(__wt_progress(session, NULL, vs->fcnt));
#ifdef HAVE_DIAGNOSTIC
/* Optionally dump the blocks or page in debugging mode. */
if (vs->dump_blocks)
WT_RET(__wt_debug_disk(session, page->dsk, NULL));
if (vs->dump_pages)
WT_RET(__wt_debug_page(session, page, NULL));
#endif
/*
* Column-store key order checks: check the page's record number and
* then update the total record count.
*/
switch (page->type) {
case WT_PAGE_COL_FIX:
recno = page->pg_fix_recno;
goto recno_chk;
case WT_PAGE_COL_INT:
recno = page->pg_intl_recno;
goto recno_chk;
case WT_PAGE_COL_VAR:
recno = page->pg_var_recno;
recno_chk: if (recno != vs->record_total + 1)
WT_RET_MSG(session, WT_ERROR,
"page at %s has a starting record of %" PRIu64
" when the expected starting record is %" PRIu64,
__wt_page_addr_string(session, ref, vs->tmp1),
recno, vs->record_total + 1);
break;
}
switch (page->type) {
case WT_PAGE_COL_FIX:
vs->record_total += page->pg_fix_entries;
break;
case WT_PAGE_COL_VAR:
recno = 0;
WT_COL_FOREACH(page, cip, i)
if ((cell = WT_COL_PTR(page, cip)) == NULL)
++recno;
else {
__wt_cell_unpack(cell, unpack);
recno += __wt_cell_rle(unpack);
}
vs->record_total += recno;
break;
}
/*
* Row-store leaf page key order check: it's a depth-first traversal,
* the first key on this page should be larger than any key previously
* seen.
*/
switch (page->type) {
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_row_leaf_key_order(session, ref, vs));
break;
}
/* If it's not the root page, unpack the parent cell. */
if (!__wt_ref_is_root(ref)) {
__wt_cell_unpack(ref->addr, unpack);
/* Compare the parent cell against the page type. */
switch (page->type) {
case WT_PAGE_COL_FIX:
if (unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_COL_VAR:
if (unpack->raw != WT_CELL_ADDR_LEAF &&
unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_ROW_LEAF:
if (unpack->raw != WT_CELL_ADDR_DEL &&
unpack->raw != WT_CELL_ADDR_LEAF &&
unpack->raw != WT_CELL_ADDR_LEAF_NO)
goto celltype_err;
break;
case WT_PAGE_COL_INT:
case WT_PAGE_ROW_INT:
if (unpack->raw != WT_CELL_ADDR_INT)
celltype_err: WT_RET_MSG(session, WT_ERROR,
"page at %s, of type %s, is referenced in "
"its parent by a cell of type %s",
__wt_page_addr_string(
session, ref, vs->tmp1),
__wt_page_type_string(page->type),
__wt_cell_type_string(unpack->raw));
break;
}
}
/*
* Check overflow pages. We check overflow cells separately from other
* tests that walk the page as it's simpler, and I don't care much how
* fast table verify runs.
*/
switch (page->type) {
case WT_PAGE_COL_VAR:
case WT_PAGE_ROW_INT:
case WT_PAGE_ROW_LEAF:
WT_RET(__verify_overflow_cell(session, ref, &found, vs));
if (__wt_ref_is_root(ref) || page->type == WT_PAGE_ROW_INT)
break;
/*
* Object if a leaf-no-overflow address cell references a page
* with overflow keys, but don't object if a leaf address cell
* references a page without overflow keys. Reconciliation
* doesn't guarantee every leaf page without overflow items will
* be a leaf-no-overflow type.
*/
if (found && unpack->raw == WT_CELL_ADDR_LEAF_NO)
WT_RET_MSG(session, WT_ERROR,
"page at %s, of type %s and referenced in its "
"parent by a cell of type %s, contains overflow "
"items",
__wt_page_addr_string(session, ref, vs->tmp1),
__wt_page_type_string(page->type),
__wt_cell_type_string(WT_CELL_ADDR_LEAF_NO));
break;
}
/* Check tree connections and recursively descend the tree. */
switch (page->type) {
case WT_PAGE_COL_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_INTL_FOREACH_BEGIN(session, page, child_ref) {
/*
* It's a depth-first traversal: this entry's starting
* record number should be 1 more than the total records
* reviewed to this point.
*/
++entry;
if (child_ref->key.recno != vs->record_total + 1) {
WT_RET_MSG(session, WT_ERROR,
"the starting record number in entry %"
PRIu32 " of the column internal page at "
"%s is %" PRIu64 " and the expected "
"starting record number is %" PRIu64,
entry,
__wt_page_addr_string(
session, child_ref, vs->tmp1),
child_ref->key.recno,
vs->record_total + 1);
}
/* Verify the subtree. */
++vs->depth;
WT_RET(__wt_page_in(session, child_ref, 0));
ret = __verify_tree(session, child_ref, vs);
WT_TRET(__wt_page_release(session, child_ref, 0));
--vs->depth;
WT_RET(ret);
__wt_cell_unpack(child_ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
} WT_INTL_FOREACH_END;
break;
case WT_PAGE_ROW_INT:
/* For each entry in an internal page, verify the subtree. */
entry = 0;
WT_INTL_FOREACH_BEGIN(session, page, child_ref) {
/*
* It's a depth-first traversal: this entry's starting
* key should be larger than the largest key previously
* reviewed.
*
* The 0th key of any internal page is magic, and we
* can't test against it.
*/
++entry;
if (entry != 1)
WT_RET(__verify_row_int_key_order(
session, page, child_ref, entry, vs));
/* Verify the subtree. */
++vs->depth;
WT_RET(__wt_page_in(session, child_ref, 0));
ret = __verify_tree(session, child_ref, vs);
WT_TRET(__wt_page_release(session, child_ref, 0));
--vs->depth;
WT_RET(ret);
__wt_cell_unpack(child_ref->addr, unpack);
WT_RET(bm->verify_addr(
bm, session, unpack->data, unpack->size));
} WT_INTL_FOREACH_END;
/*
* __compact_rewrite --
* Return if a page needs to be re-written.
*/
static int
__compact_rewrite(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_BM *bm;
WT_DECL_RET;
WT_MULTI *multi;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
size_t addr_size;
uint32_t i;
const uint8_t *addr;
*skipp = true; /* Default skip. */
bm = S2BT(session)->bm;
page = ref->page;
mod = page->modify;
/*
* Ignore the root: it may not have a replacement address, and besides,
* if anything else gets written, so will it.
*/
if (__wt_ref_is_root(ref))
return (0);
/* Ignore currently dirty pages, they will be written regardless. */
if (__wt_page_is_modified(page))
return (0);
/*
* If the page is clean, test the original addresses.
* If the page is a replacement, test the replacement addresses.
* Ignore empty pages, they get merged into the parent.
*/
if (mod == NULL || mod->rec_result == 0) {
__wt_ref_info(ref, &addr, &addr_size, NULL);
if (addr == NULL)
return (0);
return (
bm->compact_page_skip(bm, session, addr, addr_size, skipp));
}
/*
* The page's modification information can change underfoot if the page
* is being reconciled, serialize with reconciliation.
*/
if (mod->rec_result == WT_PM_REC_REPLACE ||
mod->rec_result == WT_PM_REC_MULTIBLOCK)
WT_RET(__wt_fair_lock(session, &page->page_lock));
if (mod->rec_result == WT_PM_REC_REPLACE)
ret = bm->compact_page_skip(bm, session,
mod->mod_replace.addr, mod->mod_replace.size, skipp);
if (mod->rec_result == WT_PM_REC_MULTIBLOCK)
for (multi = mod->mod_multi,
i = 0; i < mod->mod_multi_entries; ++multi, ++i) {
if (multi->disk_image != NULL)
continue;
if ((ret = bm->compact_page_skip(bm, session,
multi->addr.addr, multi->addr.size, skipp)) != 0)
break;
if (!*skipp)
break;
}
if (mod->rec_result == WT_PM_REC_REPLACE ||
mod->rec_result == WT_PM_REC_MULTIBLOCK)
WT_TRET(__wt_fair_unlock(session, &page->page_lock));
return (ret);
}
/*
* __wt_evict_file --
* Discard pages for a specific file.
*/
int
__wt_evict_file(WT_SESSION_IMPL *session, int syncop)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_REF *next_ref, *ref;
int eviction_enabled;
btree = S2BT(session);
eviction_enabled = !F_ISSET(btree, WT_BTREE_NO_EVICTION);
/*
* We need exclusive access to the file -- disable ordinary eviction
* and drain any blocks already queued.
*/
if (eviction_enabled)
WT_RET(__wt_evict_file_exclusive_on(session));
/* Make sure the oldest transaction ID is up-to-date. */
__wt_txn_update_oldest(session);
/* Walk the tree, discarding pages. */
next_ref = NULL;
WT_ERR(__wt_tree_walk(
session, &next_ref, WT_READ_CACHE | WT_READ_NO_EVICT));
while ((ref = next_ref) != NULL) {
page = ref->page;
/*
* Eviction can fail when a page in the evicted page's subtree
* switches state. For example, if we don't evict a page marked
* empty, because we expect it to be merged into its parent, it
* might no longer be empty after it's reconciled, in which case
* eviction of its parent would fail. We can either walk the
* tree multiple times (until it's finally empty), or reconcile
* each page to get it to its final state before considering if
* it's an eviction target or will be merged into its parent.
*
* Don't limit this test to any particular page type, that tends
* to introduce bugs when the reconciliation of other page types
* changes, and there's no advantage to doing so.
*
* Eviction can also fail because an update cannot be written.
* If sessions have disjoint sets of files open, updates in a
* no-longer-referenced file may not yet be globally visible,
* and the write will fail with EBUSY. Our caller handles that
* error, retrying later.
*/
if (syncop == WT_SYNC_CLOSE && __wt_page_is_modified(page))
WT_ERR(__wt_reconcile(session, ref, NULL, WT_EVICTING));
/*
* We can't evict the page just returned to us (it marks our
* place in the tree), so move the walk to one page ahead of
* the page being evicted. Note, we reconciled the returned
* page first: if reconciliation of that page were to change
* the shape of the tree, and we did the next walk call before
* the reconciliation, the next walk call could miss a page in
* the tree.
*/
WT_ERR(__wt_tree_walk(
session, &next_ref, WT_READ_CACHE | WT_READ_NO_EVICT));
switch (syncop) {
case WT_SYNC_CLOSE:
/*
* Evict the page.
* Do not attempt to evict pages expected to be merged
* into their parents, with the exception that the root
* page can't be merged, it must be written.
*/
if (__wt_ref_is_root(ref) ||
page->modify == NULL ||
!F_ISSET(page->modify, WT_PM_REC_EMPTY))
WT_ERR(__wt_evict(session, ref, 1));
break;
case WT_SYNC_DISCARD:
/*
* Ordinary discard of the page, whether clean or dirty.
* If we see a dirty page in an ordinary discard (e.g.,
* from sweep), give up: an update must have happened
* since the file was selected for sweeping.
*/
if (__wt_page_is_modified(page))
WT_ERR(EBUSY);
/*
* If the page contains an update that is too recent to
* evict, stop. This should never happen during
* connection close, but in other paths our caller
* should be prepared to deal with this case.
*/
if (page->modify != NULL &&
!__wt_txn_visible_all(session,
page->modify->rec_max_txn))
WT_ERR(EBUSY);
__wt_evict_page_clean_update(session, ref);
break;
case WT_SYNC_DISCARD_FORCE:
/*
* Forced discard of the page, whether clean or dirty.
* If we see a dirty page in a forced discard, clean
* the page, both to keep statistics correct, and to
* let the page-discard function assert no dirty page
* is ever discarded.
*/
if (__wt_page_is_modified(page)) {
page->modify->write_gen = 0;
__wt_cache_dirty_decr(session, page);
}
F_SET(session, WT_SESSION_DISCARD_FORCE);
__wt_evict_page_clean_update(session, ref);
F_CLR(session, WT_SESSION_DISCARD_FORCE);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
}
if (0) {
err: /* On error, clear any left-over tree walk. */
if (next_ref != NULL)
WT_TRET(__wt_page_release(
session, next_ref, WT_READ_NO_EVICT));
}
if (eviction_enabled)
__wt_evict_file_exclusive_off(session);
return (ret);
}
/*
* __evict_stat_walk --
* Walk all the pages in cache for a dhandle gathering stats information
*/
static void
__evict_stat_walk(WT_SESSION_IMPL *session)
{
WT_BTREE *btree;
WT_CACHE *cache;
WT_PAGE *page;
WT_REF *next_walk;
uint64_t dsk_size, gen_gap, gen_gap_max, gen_gap_sum, max_pagesize;
uint64_t min_written_size, num_memory, num_not_queueable, num_queued;
uint64_t num_smaller_allocsz, pages_clean, pages_dirty, pages_internal;
uint64_t pages_leaf, seen_count, size, visited_count;
uint64_t visited_age_gap_sum, unvisited_count, unvisited_age_gap_sum;
uint64_t walk_count, written_size_cnt, written_size_sum;
btree = S2BT(session);
cache = S2C(session)->cache;
next_walk = NULL;
gen_gap_max = gen_gap_sum = max_pagesize = 0;
num_memory = num_not_queueable = num_queued = 0;
num_smaller_allocsz = pages_clean = pages_dirty = pages_internal = 0;
pages_leaf = seen_count = size = visited_count = 0;
visited_age_gap_sum = unvisited_count = unvisited_age_gap_sum = 0;
walk_count = written_size_cnt = written_size_sum = 0;
min_written_size = UINT64_MAX;
while (__wt_tree_walk_count(session, &next_walk, &walk_count,
WT_READ_CACHE | WT_READ_NO_EVICT |
WT_READ_NO_GEN | WT_READ_NO_WAIT) == 0 &&
next_walk != NULL) {
++seen_count;
page = next_walk->page;
size = page->memory_footprint;
if (__wt_page_is_modified(page))
++pages_dirty;
else
++pages_clean;
if (!__wt_ref_is_root(next_walk) &&
!__wt_page_can_evict(session, next_walk, NULL))
++num_not_queueable;
if (F_ISSET_ATOMIC(page, WT_PAGE_EVICT_LRU))
++num_queued;
if (size > max_pagesize)
max_pagesize = size;
dsk_size = page->dsk != NULL ? page->dsk->mem_size : 0;
if (dsk_size != 0) {
if (dsk_size < btree->allocsize)
++num_smaller_allocsz;
if (dsk_size < min_written_size)
min_written_size = dsk_size;
++written_size_cnt;
written_size_sum += dsk_size;
} else
++num_memory;
if (WT_PAGE_IS_INTERNAL(page))
++pages_internal;
else
++pages_leaf;
/* Skip root pages since they are never considered */
if (__wt_ref_is_root(next_walk))
continue;
if (page->evict_pass_gen == 0) {
unvisited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
++unvisited_count;
} else {
visited_age_gap_sum +=
(cache->evict_pass_gen - page->cache_create_gen);
gen_gap = cache->evict_pass_gen - page->evict_pass_gen;
if (gen_gap > gen_gap_max)
gen_gap_max = gen_gap;
gen_gap_sum += gen_gap;
++visited_count;
}
}
WT_STAT_DATA_SET(session, cache_state_gen_avg_gap,
visited_count == 0 ? 0 : gen_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_unvisited_age,
unvisited_count == 0 ? 0 : unvisited_age_gap_sum / unvisited_count);
WT_STAT_DATA_SET(session, cache_state_avg_visited_age,
visited_count == 0 ? 0 : visited_age_gap_sum / visited_count);
WT_STAT_DATA_SET(session, cache_state_avg_written_size,
written_size_cnt == 0 ? 0 : written_size_sum / written_size_cnt);
WT_STAT_DATA_SET(session, cache_state_gen_max_gap, gen_gap_max);
WT_STAT_DATA_SET(session, cache_state_max_pagesize, max_pagesize);
WT_STAT_DATA_SET(session,
cache_state_min_written_size, min_written_size);
WT_STAT_DATA_SET(session, cache_state_memory, num_memory);
WT_STAT_DATA_SET(session, cache_state_queued, num_queued);
WT_STAT_DATA_SET(session, cache_state_not_queueable, num_not_queueable);
WT_STAT_DATA_SET(session, cache_state_pages, walk_count);
WT_STAT_DATA_SET(session, cache_state_pages_clean, pages_clean);
WT_STAT_DATA_SET(session, cache_state_pages_dirty, pages_dirty);
WT_STAT_DATA_SET(session, cache_state_pages_internal, pages_internal);
WT_STAT_DATA_SET(session, cache_state_pages_leaf, pages_leaf);
WT_STAT_DATA_SET(session,
cache_state_refs_skipped, walk_count - seen_count);
WT_STAT_DATA_SET(session,
cache_state_smaller_alloc_size, num_smaller_allocsz);
WT_STAT_DATA_SET(session,
cache_state_unvisited_count, unvisited_count);
}
/*
* __tree_walk_internal --
* Move to the next/previous page in the tree.
*/
static inline int
__tree_walk_internal(WT_SESSION_IMPL *session,
WT_REF **refp, uint64_t *walkcntp, uint64_t *skipleafcntp, uint32_t flags)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE_INDEX *pindex;
WT_REF *couple, *couple_orig, *ref;
bool empty_internal, prev, skip;
uint32_t slot;
btree = S2BT(session);
empty_internal = false;
/*
* Tree walks are special: they look inside page structures that splits
* may want to free. Publish that the tree is active during this
* window.
*/
WT_ENTER_PAGE_INDEX(session);
/* Walk should never instantiate deleted pages. */
LF_SET(WT_READ_NO_EMPTY);
/*
* !!!
* Fast-truncate currently only works on row-store trees.
*/
if (btree->type != BTREE_ROW)
LF_CLR(WT_READ_TRUNCATE);
prev = LF_ISSET(WT_READ_PREV) ? 1 : 0;
/*
* There are multiple reasons and approaches to walking the in-memory
* tree:
*
* (1) finding pages to evict (the eviction server);
* (2) writing just dirty leaves or internal nodes (checkpoint);
* (3) discarding pages (close);
* (4) truncating pages in a range (fast truncate);
* (5) skipping pages based on outside information (compaction);
* (6) cursor scans (applications).
*
* Except for cursor scans and compaction, the walk is limited to the
* cache, no pages are read. In all cases, hazard pointers protect the
* walked pages from eviction.
*
* Walks use hazard-pointer coupling through the tree and that's OK
* (hazard pointers can't deadlock, so there's none of the usual
* problems found when logically locking up a btree). If the eviction
* thread tries to evict the active page, it fails because of our
* hazard pointer. If eviction tries to evict our parent, that fails
* because the parent has a child page that can't be discarded. We do
* play one game: don't couple up to our parent and then back down to a
* new leaf, couple to the next page to which we're descending, it
* saves a hazard-pointer swap for each cursor page movement.
*
* !!!
* NOTE: we depend on the fact it's OK to release a page we don't hold,
* that is, it's OK to release couple when couple is set to NULL.
*
* Take a copy of any held page and clear the return value. Remember
* the hazard pointer we're currently holding.
*
* We may be passed a pointer to btree->evict_page that we are clearing
* here. We check when discarding pages that we're not discarding that
* page, so this clear must be done before the page is released.
*/
couple = couple_orig = ref = *refp;
*refp = NULL;
/* If no page is active, begin a walk from the start of the tree. */
if (ref == NULL) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If the active page was the root, we've reached the walk's end.
* Release any hazard-pointer we're holding.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(session, couple, flags));
goto done;
}
/* Figure out the current slot in the WT_REF array. */
__ref_index_slot(session, ref, &pindex, &slot);
for (;;) {
/*
* If we're at the last/first slot on the internal page, return
* it in post-order traversal. Otherwise move to the next/prev
* slot and left/right-most element in that subtree.
*/
while ((prev && slot == 0) ||
(!prev && slot == pindex->entries - 1)) {
/* Ascend to the parent. */
__page_ascend(session, &ref, &pindex, &slot);
/*
* If we got all the way through an internal page and
* all of the child pages were deleted, mark it for
* eviction.
*/
if (empty_internal && pindex->entries > 1) {
__wt_page_evict_soon(ref->page);
empty_internal = false;
}
/*
* If at the root and returning internal pages, return
* the root page, otherwise we're done. Regardless, no
* hazard pointer is required, release the one we hold.
*/
if (__wt_ref_is_root(ref)) {
WT_ERR(__wt_page_release(
session, couple, flags));
if (!LF_ISSET(WT_READ_SKIP_INTL))
*refp = ref;
goto done;
}
/*
* Optionally return internal pages. Swap our previous
* hazard pointer for the page we'll return. We don't
* handle restart or not-found returns, it would require
* additional complexity and is not a possible return:
* we're moving to the parent of the current child page,
* the parent can't have been evicted.
*/
if (!LF_ISSET(WT_READ_SKIP_INTL)) {
WT_ERR(__wt_page_swap(
session, couple, ref, flags));
*refp = ref;
goto done;
}
}
if (prev)
--slot;
else
++slot;
if (walkcntp != NULL)
++*walkcntp;
for (;;) {
/*
* Move to the next slot, and set the reference hint if
* it's wrong (used when we continue the walk). We don't
* update those hints when splitting, so it's common for
* them to be incorrect in some workloads.
*/
ref = pindex->index[slot];
if (ref->pindex_hint != slot)
ref->pindex_hint = slot;
/*
* If we see any child states other than deleted, the
* page isn't empty.
*/
if (ref->state != WT_REF_DELETED &&
!LF_ISSET(WT_READ_TRUNCATE))
empty_internal = false;
if (LF_ISSET(WT_READ_CACHE)) {
/*
* Only look at unlocked pages in memory:
* fast-path some common cases.
*/
if (LF_ISSET(WT_READ_NO_WAIT) &&
ref->state != WT_REF_MEM)
break;
} else if (LF_ISSET(WT_READ_TRUNCATE)) {
/*
* Avoid pulling a deleted page back in to try
* to delete it again.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
/*
* If deleting a range, try to delete the page
* without instantiating it.
*/
WT_ERR(__wt_delete_page(session, ref, &skip));
if (skip)
break;
empty_internal = false;
} else if (LF_ISSET(WT_READ_COMPACT)) {
/*
* Skip deleted pages, rewriting them doesn't
* seem useful.
*/
if (ref->state == WT_REF_DELETED)
break;
/*
* If the page is in-memory, we want to look at
* it (it may have been modified and written,
* and the current location is the interesting
* one in terms of compaction, not the original
* location). If the page isn't in-memory, test
* if the page will help with compaction, don't
* read it if we don't have to.
*/
if (ref->state == WT_REF_DISK) {
WT_ERR(__wt_compact_page_skip(
session, ref, &skip));
if (skip)
break;
}
} else {
/*
* Try to skip deleted pages visible to us.
*/
if (ref->state == WT_REF_DELETED &&
__wt_delete_page_skip(session, ref, false))
break;
}
/*
* Optionally skip leaf pages: skip all leaf pages if
* WT_READ_SKIP_LEAF is set, when the skip-leaf-count
* variable is non-zero, skip some count of leaf pages.
* If this page is disk-based, crack the cell to figure
* out it's a leaf page without reading it.
*
* If skipping some number of leaf pages, decrement the
* count of pages to zero, and then take the next leaf
* page we can. Be cautious around the page decrement,
* if for some reason don't take this particular page,
* we can take the next one, and, there are additional
* tests/decrements when we're about to return a leaf
* page.
*/
if (skipleafcntp != NULL || LF_ISSET(WT_READ_SKIP_LEAF))
if (__ref_is_leaf(ref)) {
if (LF_ISSET(WT_READ_SKIP_LEAF))
break;
if (*skipleafcntp > 0) {
--*skipleafcntp;
break;
}
}
ret = __wt_page_swap(session, couple, ref,
WT_READ_NOTFOUND_OK | WT_READ_RESTART_OK | flags);
/*
* Not-found is an expected return when only walking
* in-cache pages, or if we see a deleted page.
*/
if (ret == WT_NOTFOUND) {
ret = 0;
break;
}
/*
* The page we're moving to might have split, in which
* case move to the last position we held.
*/
if (ret == WT_RESTART) {
ret = 0;
/*
* If a new walk that never coupled from the
* root to a new saved position in the tree,
* restart the walk.
*/
if (couple == &btree->root) {
ref = &btree->root;
if (ref->page == NULL)
goto done;
goto descend;
}
/*
* If restarting from some original position,
* repeat the increment or decrement we made at
* that time. Otherwise, couple is an internal
* page we've acquired after moving from that
* starting position and we can treat it as a
* new page. This works because we never acquire
* a hazard pointer on a leaf page we're not
* going to return to our caller, this will quit
* working if that ever changes.
*/
WT_ASSERT(session,
couple == couple_orig ||
WT_PAGE_IS_INTERNAL(couple->page));
ref = couple;
__ref_index_slot(session, ref, &pindex, &slot);
if (couple == couple_orig)
break;
}
WT_ERR(ret);
/*
* A new page: configure for traversal of any internal
* page's children, else return the leaf page.
*/
if (WT_PAGE_IS_INTERNAL(ref->page)) {
descend: couple = ref;
empty_internal = true;
__page_descend(
session, ref->page, &pindex, &slot, prev);
} else {
/*
* Optionally skip leaf pages, the second half.
* We didn't have an on-page cell to figure out
* if it was a leaf page, we had to acquire the
* hazard pointer and look at the page.
*/
if (skipleafcntp != NULL ||
LF_ISSET(WT_READ_SKIP_LEAF)) {
couple = ref;
if (LF_ISSET(WT_READ_SKIP_LEAF))
break;
if (*skipleafcntp > 0) {
--*skipleafcntp;
break;
}
}
*refp = ref;
goto done;
}
}
}
done:
err: WT_LEAVE_PAGE_INDEX(session);
return (ret);
}
