Spill pages, take 3

vmware
Howard Chu 12 years ago
parent d7bc4baf63
commit e9ed4d75f7
  1. 366
      libraries/liblmdb/mdb.c
  2. 4
      libraries/liblmdb/midl.c
  3. 12
      libraries/liblmdb/midl.h

@ -585,6 +585,7 @@ typedef struct MDB_page {
#define P_DIRTY 0x10 /**< dirty page */
#define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
#define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
#define P_KEEP 0x8000 /**< leave this page alone during spill */
/** @} */
uint16_t mp_flags; /**< @ref mdb_page */
#define mp_lower mp_pb.pb.pb_lower
@ -824,6 +825,10 @@ struct MDB_txn {
/** The list of pages that became unused during this transaction.
*/
MDB_IDL mt_free_pgs;
/** The list of dirty pages we temporarily wrote to disk
* because the dirty list was full.
*/
MDB_IDL mt_spill_pgs;
union {
MDB_ID2L dirty_list; /**< for write txns: modified pages */
MDB_reader *reader; /**< this thread's reader table slot or NULL */
@ -857,6 +862,7 @@ struct MDB_txn {
#define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
#define MDB_TXN_ERROR 0x02 /**< an error has occurred */
#define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
#define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
/** @} */
unsigned int mt_flags; /**< @ref mdb_txn */
/** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
@ -1306,7 +1312,7 @@ mdb_dpage_free(MDB_env *env, MDB_page *dp)
}
}
/* Return all dirty pages to dpage list */
/** Return all dirty pages to dpage list */
static void
mdb_dlist_free(MDB_txn *txn)
{
@ -1320,6 +1326,148 @@ mdb_dlist_free(MDB_txn *txn)
dl[0].mid = 0;
}
static int mdb_page_flush(MDB_txn *txn);
/** Spill pages from the dirty list back to disk.
* This is intended to prevent running into #MDB_TXN_FULL situations,
* but note that they may still occur in a few cases:
* 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
* are too many of these dirtied in one txn, the txn may still get
* too full.
* 2) child txns may run out of space if their parents dirtied a
* lot of pages and never spilled them. TODO: we probably should do
* a preemptive spill during #mdb_txn_begin() of a child txn, if
* the parent's dirty_room is below a given threshold.
* 3) our estimate of the txn size could be too small. At the
* moment this seems unlikely.
*
* Otherwise, if not using nested txns, it is expected that apps will
* not run into #MDB_TXN_FULL any more. The pages are flushed to disk
* the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
* If the txn never references them again, they can be left alone.
* If the txn only reads them, they can be used without any fuss.
* If the txn writes them again, they can be dirtied immediately without
* going thru all of the work of #mdb_page_touch(). Such references are
* handled by #mdb_page_unspill().
*
* Also note, we never spill DB root pages, nor pages of active cursors,
* because we'll need these back again soon anyway. And in nested txns,
* we can't spill a page in a child txn if it was already spilled in a
* parent txn. That would alter the parent txns' data even though
* the child hasn't committed yet, and we'd have no way to undo it if
* the child aborted.
*
* @param[in] m0 cursor A cursor handle identifying the transaction and
* database for which we are checking space.
* @param[in] key For a put operation, the key being stored.
* @param[in] data For a put operation, the data being stored.
* @return 0 on success, non-zero on failure.
*/
static int
mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
{
MDB_txn *txn = m0->mc_txn;
MDB_page *dp;
MDB_ID2L dl = txn->mt_u.dirty_list;
unsigned int i, j;
int rc;
if (m0->mc_flags & C_SUB)
return MDB_SUCCESS;
/* Estimate how much space this op will take */
i = m0->mc_db->md_depth;
/* Named DBs also dirty the main DB */
if (m0->mc_dbi > MAIN_DBI)
i += txn->mt_dbs[MAIN_DBI].md_depth;
/* For puts, roughly factor in the key+data size */
if (key)
i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
i += i; /* double it for good measure */
if (txn->mt_dirty_room > i)
return MDB_SUCCESS;
if (!txn->mt_spill_pgs) {
txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
if (!txn->mt_spill_pgs)
return ENOMEM;
}
/* Mark all the dirty root pages we want to preserve */
for (i=0; i<txn->mt_numdbs; i++) {
if (txn->mt_dbflags[i] & DB_DIRTY) {
j = mdb_mid2l_search(dl, txn->mt_dbs[i].md_root);
if (j <= dl[0].mid) {
dp = dl[j].mptr;
dp->mp_flags |= P_KEEP;
}
}
}
/* Mark all the pages of active cursors we want to preserve */
for (i=0; i<txn->mt_numdbs; i++) {
MDB_cursor *mc = txn->mt_cursors[i];
/* See if m0 is tracked or not */
if (i == m0->mc_dbi && !(m0->mc_flags & C_UNTRACK)) {
/* nope. tack it on in front */
m0->mc_next = mc;
mc = m0;
}
for (; mc; mc=mc->mc_next) {
if (mc->mc_flags & C_INITIALIZED) {
for (j=0; j<mc->mc_snum; j++) {
if (mc->mc_pg[j]->mp_flags & P_DIRTY)
mc->mc_pg[j]->mp_flags |= P_KEEP;
}
if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
MDB_cursor *mx = &mc->mc_xcursor->mx_cursor;
if (mx->mc_flags & C_INITIALIZED) {
for (j=0; j<mx->mc_snum; j++) {
if ((mx->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY))
== P_DIRTY)
mx->mc_pg[j]->mp_flags |= P_KEEP;
}
}
}
}
}
}
/* Save the page IDs of all the pages we're flushing */
for (i=1; i<=dl[0].mid; i++) {
dp = dl[i].mptr;
if (dp->mp_flags & P_KEEP)
continue;
/* Can't spill twice, make sure it's not already in a parent's
* spill list.
*/
if (txn->mt_parent) {
MDB_txn *tx2;
for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
if (tx2->mt_spill_pgs) {
j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
dp->mp_flags |= P_KEEP;
break;
}
}
}
if (tx2)
continue;
}
if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
return rc;
}
mdb_midl_sort(txn->mt_spill_pgs);
rc = mdb_page_flush(txn);
if (rc == 0) {
txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
txn->mt_flags |= MDB_TXN_SPILLS;
}
return rc;
}
/** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
static txnid_t
mdb_find_oldest(MDB_txn *txn)
@ -1533,6 +1681,61 @@ mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
}
}
/** Pull a page off the txn's spill list, if present.
* If a page being referenced was spilled to disk in this txn, bring
* it back and make it dirty/writable again.
* @param[in] tx0 the transaction handle.
* @param[in] mp the page being referenced.
* @param[out] ret the writable page, if any. ret is unchanged if
* mp wasn't spilled.
*/
static int
mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
{
MDB_env *env = tx0->mt_env;
MDB_txn *txn;
unsigned x;
pgno_t pgno = mp->mp_pgno;
for (txn = tx0; txn; txn=txn->mt_parent) {
if (!txn->mt_spill_pgs)
continue;
x = mdb_midl_search(txn->mt_spill_pgs, pgno);
if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
MDB_page *np;
int num;
if (IS_OVERFLOW(mp))
num = mp->mp_pages;
else
num = 1;
if (env->me_flags & MDB_WRITEMAP) {
np = mp;
} else {
np = mdb_page_malloc(txn, num);
if (!np)
return ENOMEM;
if (num > 1)
memcpy(np, mp, num * env->me_psize);
else
mdb_page_copy(np, mp, env->me_psize);
}
if (txn == tx0) {
/* If in current txn, this page is no longer spilled */
for (; x < txn->mt_spill_pgs[0]; x++)
txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
txn->mt_spill_pgs[0]--;
} /* otherwise, if belonging to a parent txn, the
* page remains spilled until child commits
*/
mdb_page_dirty(tx0, np);
np->mp_flags |= P_DIRTY;
*ret = np;
break;
}
}
return MDB_SUCCESS;
}
/** Touch a page: make it dirty and re-insert into tree with updated pgno.
* @param[in] mc cursor pointing to the page to be touched
* @return 0 on success, non-zero on failure.
@ -1548,6 +1751,14 @@ mdb_page_touch(MDB_cursor *mc)
int rc;
if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
if (txn->mt_flags & MDB_TXN_SPILLS) {
np = NULL;
rc = mdb_page_unspill(txn, mp, &np);
if (rc)
return rc;
if (np)
goto done;
}
if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
(rc = mdb_page_alloc(mc, 1, &np)))
return rc;
@ -1595,6 +1806,7 @@ mdb_page_touch(MDB_cursor *mc)
np->mp_pgno = pgno;
np->mp_flags |= P_DIRTY;
done:
/* Adjust cursors pointing to mp */
mc->mc_pg[mc->mc_top] = np;
dbi = mc->mc_dbi;
@ -1801,6 +2013,7 @@ mdb_txn_renew0(MDB_txn *txn)
txn->mt_u.dirty_list[0].mid = 0;
txn->mt_free_pgs = env->me_free_pgs;
txn->mt_free_pgs[0] = 0;
txn->mt_spill_pgs = NULL;
env->me_txn = txn;
}
@ -1906,6 +2119,7 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
txn->mt_toggle = parent->mt_toggle;
txn->mt_dirty_room = parent->mt_dirty_room;
txn->mt_u.dirty_list[0].mid = 0;
txn->mt_spill_pgs = NULL;
txn->mt_next_pgno = parent->mt_next_pgno;
parent->mt_child = txn;
txn->mt_parent = parent;
@ -2008,6 +2222,7 @@ mdb_txn_reset0(MDB_txn *txn, const char *act)
txn->mt_parent->mt_child = NULL;
env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
mdb_midl_free(txn->mt_free_pgs);
mdb_midl_free(txn->mt_spill_pgs);
free(txn->mt_u.dirty_list);
return;
}
@ -2210,7 +2425,7 @@ mdb_page_flush(MDB_txn *txn)
{
MDB_env *env = txn->mt_env;
MDB_ID2L dl = txn->mt_u.dirty_list;
unsigned psize = env->me_psize;
unsigned psize = env->me_psize, j;
int i, pagecount = dl[0].mid, rc;
size_t size = 0, pos = 0;
pgno_t pgno = 0;
@ -2224,13 +2439,20 @@ mdb_page_flush(MDB_txn *txn)
int n = 0;
#endif
j = 0;
if (env->me_flags & MDB_WRITEMAP) {
/* Clear dirty flags */
for (i = pagecount; i; i--) {
dp = dl[i].mptr;
/* Don't flush this page yet */
if (dp->mp_flags & P_KEEP) {
dp->mp_flags ^= P_KEEP;
dl[++j] = dl[i];
continue;
}
dp->mp_flags &= ~P_DIRTY;
}
dl[0].mid = 0;
dl[0].mid = j;
return MDB_SUCCESS;
}
@ -2238,6 +2460,12 @@ mdb_page_flush(MDB_txn *txn)
for (i = 1;; i++) {
if (i <= pagecount) {
dp = dl[i].mptr;
/* Don't flush this page yet */
if (dp->mp_flags & P_KEEP) {
dp->mp_flags ^= P_KEEP;
dl[i].mid = 0;
continue;
}
pgno = dl[i].mid;
/* clear dirty flag */
dp->mp_flags &= ~P_DIRTY;
@ -2309,7 +2537,18 @@ mdb_page_flush(MDB_txn *txn)
#endif /* _WIN32 */
}
mdb_dlist_free(txn);
j = 0;
for (i=1; i<=pagecount; i++) {
dp = dl[i].mptr;
/* This is a page we skipped above */
if (!dl[i].mid) {
dl[++j] = dl[i];
dl[j].mid = dp->mp_pgno;
continue;
}
mdb_dpage_free(env, dp);
}
dl[0].mid = j;
return MDB_SUCCESS;
}
@ -2378,6 +2617,37 @@ mdb_txn_commit(MDB_txn *txn)
dst = parent->mt_u.dirty_list;
src = txn->mt_u.dirty_list;
/* Remove anything in our dirty list from parent's spill list */
if (parent->mt_spill_pgs) {
x = parent->mt_spill_pgs[0];
len = x;
/* zero out our dirty pages in parent spill list */
for (i=1; i<=src[0].mid; i++) {
if (src[i].mid < parent->mt_spill_pgs[x])
continue;
if (src[i].mid > parent->mt_spill_pgs[x]) {
if (x <= 1)
break;
x--;
continue;
}
parent->mt_spill_pgs[x] = 0;
len--;
}
/* OK, we had a few hits, squash zeros from the spill list */
if (len < parent->mt_spill_pgs[0]) {
x=1;
for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
if (parent->mt_spill_pgs[y]) {
if (y != x) {
parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
}
x++;
}
}
parent->mt_spill_pgs[0] = len;
}
}
/* Find len = length of merging our dirty list with parent's */
x = dst[0].mid;
dst[0].mid = 0; /* simplify loops */
@ -2409,6 +2679,15 @@ mdb_txn_commit(MDB_txn *txn)
dst[0].mid = len;
free(txn->mt_u.dirty_list);
parent->mt_dirty_room = txn->mt_dirty_room;
if (txn->mt_spill_pgs) {
if (parent->mt_spill_pgs) {
mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
mdb_midl_free(txn->mt_spill_pgs);
mdb_midl_sort(parent->mt_spill_pgs);
} else {
parent->mt_spill_pgs = txn->mt_spill_pgs;
}
}
parent->mt_child = NULL;
mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
@ -3991,6 +4270,19 @@ mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
level = 1;
do {
MDB_ID2L dl = tx2->mt_u.dirty_list;
unsigned x;
/* Spilled pages were dirtied in this txn and flushed
* because the dirty list got full. Bring this page
* back in from the map (but don't unspill it here,
* leave that unless page_touch happens again).
*/
if (tx2->mt_spill_pgs) {
x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
goto done;
}
}
if (dl[0].mid) {
unsigned x = mdb_mid2l_search(dl, pgno);
if (x <= dl[0].mid && dl[x].mid == pgno) {
@ -4091,6 +4383,8 @@ mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
key ? DKEY(key) : NULL);
mc->mc_flags |= C_INITIALIZED;
mc->mc_flags &= ~C_EOF;
return MDB_SUCCESS;
}
@ -4218,11 +4512,21 @@ mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
int rc;
DPRINTF("free ov page %zu (%d)", pg, ovpages);
/* If the page is dirty we just acquired it, so we should
* give it back to our current free list, if any.
/* If the page is dirty or on the spill list we just acquired it,
* so we should give it back to our current free list, if any.
* Not currently supported in nested txns.
* Otherwise put it onto the list of pages we freed in this txn.
*/
if (!(mp->mp_flags & P_DIRTY) && txn->mt_spill_pgs) {
unsigned x = mdb_midl_search(txn->mt_spill_pgs, pg);
if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pg) {
/* This page is no longer spilled */
for (; x < txn->mt_spill_pgs[0]; x++)
txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
txn->mt_spill_pgs[0]--;
goto release;
}
}
if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
unsigned j, x;
pgno_t *mop;
@ -4248,6 +4552,7 @@ mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
}
if (!(env->me_flags & MDB_WRITEMAP))
mdb_dpage_free(env, mp);
release:
/* Insert in me_pghead */
mop = env->me_pghead;
j = mop[0] + ovpages;
@ -4964,6 +5269,9 @@ mdb_cursor_touch(MDB_cursor *mc)
return MDB_SUCCESS;
}
/** Do not spill pages to disk if txn is getting full, may fail instead */
#define MDB_NOSPILL 0x8000
int
mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
unsigned int flags)
@ -4974,7 +5282,7 @@ mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
MDB_page *fp;
MDB_db dummy;
int do_sub = 0, insert = 0;
unsigned int mcount = 0, dcount = 0;
unsigned int mcount = 0, dcount = 0, nospill;
size_t nsize;
int rc, rc2;
MDB_pagebuf pbuf;
@ -4992,6 +5300,9 @@ mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
return EINVAL;
}
nospill = flags & MDB_NOSPILL;
flags &= ~MDB_NOSPILL;
if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
return EACCES;
@ -5048,7 +5359,17 @@ mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
return rc;
}
/* Cursor is positioned */
/* Cursor is positioned, check for room in the dirty list */
if (!nospill) {
if (flags & MDB_MULTIPLE) {
rdata = &xdata;
xdata.mv_size = data->mv_size * dcount;
} else {
rdata = data;
}
if ((rc2 = mdb_page_spill(mc, key, rdata)))
return rc2;
}
if (rc == MDB_NO_ROOT) {
MDB_page *np;
@ -5227,8 +5548,18 @@ current:
return rc2;
ovpages = omp->mp_pages;
/* Is the ov page writable and large enough? */
if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
/* Is the ov page large enough? */
if (ovpages >= dpages) {
if (!(omp->mp_flags & P_DIRTY) &&
(level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
{
rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
if (rc)
return rc;
level = 0; /* dirty in this txn or clean */
}
/* Is it dirty? */
if (omp->mp_flags & P_DIRTY) {
/* yes, overwrite it. Note in this case we don't
* bother to try shrinking the page if the new data
* is smaller than the overflow threshold.
@ -5261,10 +5592,10 @@ current:
else
memcpy(METADATA(omp), data->mv_data, data->mv_size);
goto done;
} else {
}
}
if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
return rc2;
}
} else if (NODEDSZ(leaf) == data->mv_size) {
/* same size, just replace it. Note that we could
* also reuse this node if the new data is smaller,
@ -5337,10 +5668,11 @@ put_sub:
xdata.mv_data = "";
leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
if (flags & MDB_CURRENT) {
xflags = MDB_CURRENT;
xflags = MDB_CURRENT|MDB_NOSPILL;
} else {
mdb_xcursor_init1(mc, leaf);
xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
xflags = (flags & MDB_NODUPDATA) ?
MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
}
/* converted, write the original data first */
if (dkey.mv_size) {
@ -5411,6 +5743,10 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
if (!(mc->mc_flags & C_INITIALIZED))
return EINVAL;
if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
return rc;
flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
rc = mdb_cursor_touch(mc);
if (rc)
return rc;
@ -5422,7 +5758,7 @@ mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
}
rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
/* If sub-DB still has entries, we're done */
if (mc->mc_xcursor->mx_db.md_entries) {
if (leaf->mn_flags & F_SUBDATA) {

@ -31,8 +31,7 @@
*/
#define CMP(x,y) ( (x) < (y) ? -1 : (x) > (y) )
#if 0 /* superseded by append/sort */
static unsigned mdb_midl_search( MDB_IDL ids, MDB_ID id )
unsigned mdb_midl_search( MDB_IDL ids, MDB_ID id )
{
/*
* binary search of id in ids
@ -67,6 +66,7 @@ static unsigned mdb_midl_search( MDB_IDL ids, MDB_ID id )
return cursor;
}
#if 0 /* superseded by append/sort */
int mdb_midl_insert( MDB_IDL ids, MDB_ID id )
{
unsigned x, i;

@ -74,14 +74,12 @@ typedef MDB_ID *MDB_IDL;
xidl[xlen] = (id); \
} while (0)
#if 0 /* superseded by append/sort */
/** Insert an ID into an IDL.
* @param[in,out] ids The IDL to insert into.
* @param[in] id The ID to insert.
* @return 0 on success, -1 if ID was already present, -2 on error.
/** Search for an ID in an IDL.
* @param[in] ids The IDL to search.
* @param[in] id The ID to search for.
* @return The index of the first ID greater than or equal to \b id.
*/
int mdb_midl_insert( MDB_IDL ids, MDB_ID id );
#endif
unsigned mdb_midl_search( MDB_IDL ids, MDB_ID id );
/** Allocate an IDL.
* Allocates memory for an IDL of the given size.

Loading…
Cancel
Save