/*
* __wt_compact_page_skip --
* Return if compaction requires we read this page.
*/
int
__wt_compact_page_skip(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_BM *bm;
size_t addr_size;
u_int type;
const uint8_t *addr;
*skipp = false; /* Default to reading. */
type = 0; /* Keep compiler quiet. */
bm = S2BT(session)->bm;
/*
* We aren't holding a hazard pointer, so we can't look at the page
* itself, all we can look at is the WT_REF information. If there's no
* address, the page isn't on disk, but we have to read internal pages
* to walk the tree regardless; throw up our hands and read it.
*/
__wt_ref_info(ref, &addr, &addr_size, &type);
if (addr == NULL)
return (0);
/*
* Internal pages must be read to walk the tree; ask the block-manager
* if it's useful to rewrite leaf pages, don't do the I/O if a rewrite
* won't help.
*/
return (type == WT_CELL_ADDR_INT ? 0 :
bm->compact_page_skip(bm, session, addr, addr_size, skipp));
}
/*
* __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));
}
/*
* __ref_is_leaf --
* Check if a reference is for a leaf page.
*/
static inline bool
__ref_is_leaf(WT_REF *ref)
{
size_t addr_size;
const uint8_t *addr;
u_int type;
/*
* If the page has a disk address, we can crack it to figure out if
* this page is a leaf page or not. If there's no address, the page
* isn't on disk and we don't know the page type.
*/
__wt_ref_info(ref, &addr, &addr_size, &type);
return (addr == NULL ?
false : type == WT_CELL_ADDR_LEAF || type == WT_CELL_ADDR_LEAF_NO);
}
/*
* __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_ITEM *buf, WT_PAGE *page)
{
uint32_t size;
const uint8_t *addr;
if (WT_PAGE_IS_ROOT(page)) {
buf->data = "[Root]";
buf->size = WT_STORE_SIZE(strlen("[Root]"));
return (buf->data);
}
(void)__wt_ref_info(
session, page->parent, page->ref, &addr, &size, NULL);
return (__wt_addr_string(session, buf, addr, size));
}
/*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;
}
/*在读取ref对应的page时，检查它是否需要compact*/
int __wt_compact_page_skip(WT_SESSION_IMPL* session, WT_REF* ref, int* skipp)
{
WT_BM *bm;
size_t addr_size;
u_int type;
const uint8_t *addr;
*skipp = 0;
type = 0;
bm = S2BT(session)->bm;
/*
* We aren't holding a hazard pointer, so we can't look at the page
* itself, all we can look at is the WT_REF information. If there's no
* address, the page isn't on disk, but we have to read internal pages
* to walk the tree regardless; throw up our hands and read it.
*/
WT_RET(__wt_ref_info(session, ref, &addr, &addr_size, &type));
if (addr == NULL)
return 0;
return (type == WT_CELL_ADDR_INT ? 0 : bm->compact_page_skip(bm, session, addr, addr_size, skipp));
}
/*
* __page_read --
* Read a page from the file.
*/
static int
__page_read(WT_SESSION_IMPL *session, WT_REF *ref)
{
const WT_PAGE_HEADER *dsk;
WT_BTREE *btree;
WT_DECL_RET;
WT_ITEM tmp;
WT_PAGE *page;
size_t addr_size;
uint32_t previous_state;
const uint8_t *addr;
btree = S2BT(session);
page = NULL;
/*
* Don't pass an allocated buffer to the underlying block read function,
* force allocation of new memory of the appropriate size.
*/
WT_CLEAR(tmp);
/*
* Attempt to set the state to WT_REF_READING for normal reads, or
* WT_REF_LOCKED, for deleted pages. If successful, we've won the
* race, read the page.
*/
if (__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_READING))
previous_state = WT_REF_DISK;
else if (__wt_atomic_casv32(&ref->state, WT_REF_DELETED, WT_REF_LOCKED))
previous_state = WT_REF_DELETED;
else
return (0);
/*
* Get the address: if there is no address, the page was deleted, but a
* subsequent search or insert is forcing re-creation of the name space.
*/
WT_ERR(__wt_ref_info(session, ref, &addr, &addr_size, NULL));
if (addr == NULL) {
WT_ASSERT(session, previous_state == WT_REF_DELETED);
WT_ERR(__wt_btree_new_leaf_page(session, &page));
ref->page = page;
goto done;
}
/*
* There's an address, read or map the backing disk page and build an
* in-memory version of the page.
*/
WT_ERR(__wt_bt_read(session, &tmp, addr, addr_size));
WT_ERR(__wt_page_inmem(session, ref, tmp.data, tmp.memsize,
WT_DATA_IN_ITEM(&tmp) ?
WT_PAGE_DISK_ALLOC : WT_PAGE_DISK_MAPPED, &page));
/*
* Clear the local reference to an allocated copy of the disk image on
* return; the page steals it, errors in this code should not free it.
*/
tmp.mem = NULL;
/*
* If reading for a checkpoint, there's no additional work to do, the
* page on disk is correct as written.
*/
if (session->dhandle->checkpoint != NULL)
goto done;
/* If the page was deleted, instantiate that information. */
if (previous_state == WT_REF_DELETED)
WT_ERR(__wt_delete_page_instantiate(session, ref));
/*
* Instantiate updates from the database's lookaside table. The page
* flag was set when the page was written, potentially a long time ago.
* We only care if the lookaside table is currently active, check that
* before doing any work.
*/
dsk = tmp.data;
if (F_ISSET(dsk, WT_PAGE_LAS_UPDATE) && __wt_las_is_written(session)) {
WT_STAT_FAST_CONN_INCR(session, cache_read_lookaside);
WT_STAT_FAST_DATA_INCR(session, cache_read_lookaside);
WT_ERR(__las_page_instantiate(
session, ref, btree->id, addr, addr_size));
}
done: WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
err: /*
* If the function building an in-memory version of the page failed,
* it discarded the page, but not the disk image. Discard the page
* and separately discard the disk image in all cases.
*/
if (ref->page != NULL)
__wt_ref_out(session, ref);
WT_PUBLISH(ref->state, previous_state);
__wt_buf_free(session, &tmp);
return (ret);
}
/*
* __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_cache_read --
* Read a page from the file.
*/
int
__wt_cache_read(WT_SESSION_IMPL *session, WT_REF *ref)
{
WT_DECL_RET;
WT_ITEM tmp;
WT_PAGE *page;
WT_PAGE_STATE previous_state;
size_t addr_size;
const uint8_t *addr;
page = NULL;
/*
* Don't pass an allocated buffer to the underlying block read function,
* force allocation of new memory of the appropriate size.
*/
WT_CLEAR(tmp);
/*
* Attempt to set the state to WT_REF_READING for normal reads, or
* WT_REF_LOCKED, for deleted pages. If successful, we've won the
* race, read the page.
*/
if (WT_ATOMIC_CAS4(ref->state, WT_REF_DISK, WT_REF_READING))
previous_state = WT_REF_DISK;
else if (WT_ATOMIC_CAS4(ref->state, WT_REF_DELETED, WT_REF_LOCKED))
previous_state = WT_REF_DELETED;
else
return (0);
/*
* Get the address: if there is no address, the page was deleted, but a
* subsequent search or insert is forcing re-creation of the name space.
* Otherwise, there's an address, read the backing disk page and build
* an in-memory version of the page.
*/
WT_ERR(__wt_ref_info(session, ref, &addr, &addr_size, NULL));
if (addr == NULL) {
WT_ASSERT(session, previous_state == WT_REF_DELETED);
WT_ERR(__wt_btree_new_leaf_page(session, &page));
ref->page = page;
} else {
/* Read the backing disk page. */
WT_ERR(__wt_bt_read(session, &tmp, addr, addr_size));
/* Build the in-memory version of the page. */
WT_ERR(__wt_page_inmem(session, ref, tmp.data,
WT_DATA_IN_ITEM(&tmp) ?
WT_PAGE_DISK_ALLOC : WT_PAGE_DISK_MAPPED, &page));
/* If the page was deleted, instantiate that information. */
if (previous_state == WT_REF_DELETED)
WT_ERR(__wt_delete_page_instantiate(session, ref));
}
WT_ERR(__wt_verbose(session, WT_VERB_READ,
"page %p: %s", page, __wt_page_type_string(page->type)));
WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
err: /*
* If the function building an in-memory version of the page failed,
* it discarded the page, but not the disk image. Discard the page
* and separately discard the disk image in all cases.
*/
if (ref->page != NULL)
__wt_ref_out(session, ref);
WT_PUBLISH(ref->state, previous_state);
__wt_buf_free(session, &tmp);
return (ret);
}
/*
* __wt_compact_page_skip --
* Return if compaction requires we read this page.
*/
int
__wt_compact_page_skip(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_BM *bm;
WT_DECL_RET;
size_t addr_size;
u_int type;
const uint8_t *addr;
/*
* Skip deleted pages, rewriting them doesn't seem useful; in a better
* world we'd write the parent to delete the page.
*/
if (ref->state == WT_REF_DELETED) {
*skipp = true;
return (0);
}
*skipp = false; /* Default to reading */
/*
* 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).
*
* This test could be combined with the next one, but this is a cheap
* test and the next one is expensive.
*/
if (ref->state != WT_REF_DISK)
return (0);
/*
* There's nothing to prevent the WT_REF state from changing underfoot,
* which can change its address. For example, the WT_REF address might
* reference an on-page cell, and page eviction can free that memory.
* Lock the WT_REF so we can look at its address.
*/
if (!__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_LOCKED))
return (0);
/*
* The page is on disk, so there had better be an address; assert that
* fact, test at run-time to avoid the core dump.
*
* Internal pages must be read to walk the tree; ask the block-manager
* if it's useful to rewrite leaf pages, don't do the I/O if a rewrite
* won't help.
*/
__wt_ref_info(ref, &addr, &addr_size, &type);
WT_ASSERT(session, addr != NULL);
if (addr != NULL && type != WT_CELL_ADDR_INT) {
bm = S2BT(session)->bm;
ret = bm->compact_page_skip(
bm, session, addr, addr_size, skipp);
}
/*
* Reset the WT_REF state and push the change. The full-barrier isn't
* necessary, but it's better to keep pages in circulation than not.
*/
ref->state = WT_REF_DISK;
WT_FULL_BARRIER();
return (ret);
}
