Skip to content

Commit

Permalink
Track range of free-list updates and only save those
Browse files Browse the repository at this point in the history
  • Loading branch information
@kriszyp
kriszyp committed Mar 7, 2024
1 parent 075c092 commit 23cdfee
Show file tree
Hide file tree
Showing 2 changed files with 104 additions and 46 deletions.
147 changes: 102 additions & 45 deletions dependencies/lmdb/libraries/liblmdb/mdb.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1581,13 +1581,15 @@ typedef struct MDB_xcursor {
/** State of FreeDB old pages, stored in the MDB_env */
typedef struct MDB_pgstate {
pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
txnid_t mf_freelist_start; /**< ID of last used record, or 0 if !mf_pghead */
txnid_t mf_freelist_end; /**< ID of last used record, or 0 if !mf_pghead */
pgno_t mf_freelist_start; /**< Page number of beginning of active freelist (in-memory) */
pgno_t mf_freelist_end; /**< Page number of end of active freelist (in-memory) */
/** Position in the free page list, so that we can keep trying to write to the same block
* if possible. We use sign here, a negative number means we are in middle of contiguous
* (in a transaction) and we really want to use the next page, a positive number is
* just the next page number */
int mf_position;
int mf_written_start; /* start of modified entries in free-list */
int mf_written_end; /* start of modified entries in free-list */
} MDB_pgstate;
/*<lmdb-js>*/
struct MDB_last_map {
Expand Down Expand Up @@ -1650,6 +1652,8 @@ struct MDB_env {
# define me_freelist_end me_pgstate.mf_freelist_end
# define me_pghead me_pgstate.mf_pghead
# define me_freelist_position me_pgstate.mf_position
# define me_freelist_written_start me_pgstate.mf_written_start
# define me_freelist_written_end me_pgstate.mf_written_end
unsigned int me_maxfreepgs_to_load; /**< max freelist entries to load into memory */
unsigned int me_maxfreepgs_to_retain; /**< max freelist entries to load into memory */
MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
Expand Down Expand Up @@ -2808,9 +2812,12 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
if (block_size >= num) {
if (block_size == num) {
// we found a block of the right size
mop[i] = 0;
if (entry < 1) mop[i - 1] = 0;
env->me_freelist_position = i + 1;
mop[i] = 0;
if (env->me_freelist_written_end < i) env->me_freelist_written_end = i;
if (entry < 1)
mop[--i] = 0;
if (env->me_freelist_written_start > i || !env->me_freelist_written_start) env->me_freelist_written_start = i;
goto search_done;
} else if (block_size < best_fit_size || best_fit_size == 0) {
best_fit_start = i - 1;
Expand All @@ -2829,6 +2836,9 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
mdb_midl_respread(&env->me_pghead);
mop = env->me_pghead;
mop_len = mop[0];
// consider the whole free-list to be updated now
env->me_freelist_written_end = 0x7fffffff;
env->me_freelist_written_start = 1;
//fprintf(stderr, "resized from %u to %u\n", old_length, mop_len);
goto restart_search;
}
Expand Down Expand Up @@ -2902,6 +2912,7 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
if (!last) break; // should be no zero entry, break out
key.mv_data = &last; // start at the end of the freelist and read newer transactions free pages
key.mv_size = sizeof(last);
mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
rc = mdb_cursor_get(&m2, &key, NULL, op);
op = MDB_PREV;
if (rc) {
Expand All @@ -2915,6 +2926,11 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
goto fail;
}
last = *(txnid_t*)key.mv_data;
if (!last) {
fprintf(stderr, "Invalid null txn id\n");
rc = MDB_NOTFOUND;
goto fail;
}
if (last >= env->me_freelist_start) {
// go to previous entry, through next iteration
rc = mdb_cursor_get(&m2, &key, NULL, op);
Expand All @@ -2931,6 +2947,12 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
if (rc == MDB_NOTFOUND) break;
env->me_freelist_start = last;
}
if (!last) {
fprintf(stderr, "Invalid null txn id\n");
rc = MDB_NOTFOUND;
goto fail;
}

np = m2.mc_pg[m2.mc_top];
leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
if ((rc = mdb_node_read(&m2, leaf, &data)) != MDB_SUCCESS)
Expand All @@ -2948,16 +2970,18 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
goto fail;
mop = env->me_pghead;
}
//env->me_freelist_start = last;
// consider the whole free-list to be updated now
env->me_freelist_written_end = 0x7fffffff;
env->me_freelist_written_start = 1;
#if (MDB_DEBUG) > 1
DPRINTF(("IDL read txn %"Yu" root %"Yu" num %u",
last, txn->mt_dbs[FREE_DBI].md_root, i));
for (j = i; j; j--)
DPRINTF(("IDL %"Yu, idl[j]));
#endif
/* Merge in descending sorted order */
//fprintf(stderr, "Merging %u: ", last);
//mdb_midl_print(stderr, idl);
// fprintf(stderr, "Merging %u: ", last);
// mdb_midl_print(stderr, idl);
if ((rc = mdb_midl_xmerge(&mop, idl)) != 0)
goto fail;
if (mop != env->me_pghead) env->me_pghead = mop;
Expand All @@ -2967,9 +2991,13 @@ mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
if (best_fit_start > 0) {
mop[best_fit_start] += num; // block length is a negative, so we add to it in order to subtract the amount we are using
if (mop[best_fit_start] == -1) mop[best_fit_start] = 0;
pgno = mop[best_fit_start + 1];
mop[best_fit_start + 1] += num;
if (env->me_freelist_written_start > best_fit_start || !env->me_freelist_written_start) env->me_freelist_written_start = best_fit_start;
// we use a negative value as an indicator that this position is in the middle of a block, and we can continue it
// for sequential writing (faster)
env->me_freelist_position = -best_fit_start;
pgno = mop[++best_fit_start];
mop[best_fit_start] += num; // update position
if (env->me_freelist_written_end < best_fit_start) env->me_freelist_written_end = best_fit_start;
//fprintf(stderr, "using best fit at %u size %u of %u\n", pgno, num, best_fit_size);
//env->me_block_size_cache[best_fit_size] = 0; // clear this out of the cache (TODO: could move it)

Expand Down Expand Up @@ -4038,13 +4066,15 @@ mdb_freelist_save(MDB_txn *txn)
* env->me_freelist_start and txn->mt_free_pgs[] can only grow.
* Page numbers cannot disappear from txn->mt_free_pgs[].
*/
char fl_writes[100000];
MDB_cursor mc;
MDB_env *env = txn->mt_env;
int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1, lost_loose = 0;
txnid_t pglast = 0;
txnid_t head_id = 0;
pgno_t freecnt = 0, *free_pgs, *mop;
ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
unsigned entry_size;

mdb_cursor_init(&mc, txn, FREE_DBI, NULL);

Expand Down Expand Up @@ -4170,20 +4200,32 @@ mdb_freelist_save(MDB_txn *txn)
}
if (mop_len) {
// determine the size of the blocks we need
unsigned entry_size = ((mop_len / ((env->me_freelist_end - start_written) * maxfree_1pg)) + 1) * maxfree_1pg;
pglast = mop_len / entry_size + head_id;
entry_size = ((mop_len / ((env->me_freelist_end - start_written) * maxfree_1pg)) + 1) * maxfree_1pg;
pglast = (mop_len - 1) / entry_size + head_id;
//fprintf(stderr, "Reserving in-memory freelist, length: %u, ids: ", mop_len);
int i = 0;
while (head_id <= pglast) {
key.mv_size = sizeof(head_id);
key.mv_data = &head_id;
int len = mop_len > entry_size ? entry_size : mop_len;
unsigned end = mop_len;
mop_len -= len;
data.mv_size = (len + 2) * sizeof(pgno_t);
//fprintf(stderr, "id: %u size: %u", head_id, data.mv_size);
rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
head_id++;
if (rc)
return rc;
unsigned start = mop_len > 0 ? mop_len - 1 : 0;
char do_write = env->me_freelist_written_start <= end && env->me_freelist_written_end >= start;
fl_writes[i++] = do_write;
// determine if it is in the range of actively written pages
if (env->me_freelist_written_start <= end && env->me_freelist_written_end >= start) {
// we will reserve one entry for size and potentially one extra entry in case we are splitting an entry
data.mv_size = (len + (head_id < pglast ? 2 : 1)) * sizeof(pgno_t);
//fprintf(stderr, "id: %u size: %u", head_id, data.mv_size);
rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
head_id++;
if (rc)
return rc;
} else {
//fprintf(stderr, "Skipping reservation of freelist %u from %u-%u", head_id, mop_len, end);
head_id++;
}
}
} else pglast = head_id - 1;

Expand Down Expand Up @@ -4227,6 +4269,7 @@ mdb_freelist_save(MDB_txn *txn)
txn->mt_loose_pgs = NULL;
txn->mt_loose_count = 0;
mop = env->me_pghead;
// TODO: Append instead
mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
if (mop_len > reserved_space) {
mop = mdb_midl_pack(env->me_pghead);
Expand All @@ -4247,52 +4290,66 @@ mdb_freelist_save(MDB_txn *txn)
env->me_pghead = NULL;
txn->mt_dirty_room = 0;

//fprintf(stderr, "Writing in-memory freelist, length: %u ids: ", mop_len);
// fprintf(stderr, "Writing in-memory freelist, length: %u ids: ", mop_len);
int i = 0;
for (txnid_t id = start_written; id <= pglast; id++) {
key.mv_size = sizeof(id);
key.mv_data = &id;
mdb_cursor_get(&mc, &key, &data, MDB_SET_KEY);
//fprintf(stderr, "put %u: ", id);
// fprintf(stderr, "put %u: ", id);

int len = (data.mv_size / sizeof(id)) - 1;
ssize_t end = mop_len;
ssize_t save_end = mop[end];
if (save_end < 0) {
// ends with a block length, zero it out, was already handled by previous iteration
mop[end] = 0;
}
ssize_t start = mop_len - len;
if (start < 0) {
// oversized entry, we will copy in our data and fill the rest with zeros
len += start;
// oversized entry, just use the remaining length
len = mop_len;
start = 0;
ssize_t save = mop[0];
mop[0] = len;
//mdb_midl_print(stderr, mop);
rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT|MDB_RESERVE);
if (rc) {
mop[0] = save;
return rc;
}
ssize_t save2 = 0;
ssize_t save = mop[start];
if (len > entry_size) {
// full blocks
// we have an extra overlapping byte to handle block length prefix. But don't use it if it is a page number
// because it could be preceded by a block length prefix, instead zero it out
save2 = mop[start + 1];
if (save2 > 0) {
mop[start + 1] = 0;
}
memcpy(data.mv_data, mop, (len + 1) * sizeof(id));
memset((txnid_t*) data.mv_data + len + 1, 0, data.mv_size - (len + 1) * sizeof(id));
mop[0] = save;
mop_len = 0; // nothing left to save
mop_len -= entry_size;
} else {
ssize_t save = mop[start];
ssize_t save2 = 0;
mdb_tassert(txn, mop_len <= entry_size);
mop_len = 0; // nothing left to save
}
char do_write = env->me_freelist_written_start <= end && env->me_freelist_written_end >= start || !end;
if (fl_writes[i++] != do_write) {
fprintf(stderr, "Do write of page does not match");
}
if (env->me_freelist_written_start <= end && env->me_freelist_written_end >= start || !end) {
// we are in part of the range of the freelist that was written, so we actually need to save it
// if end == 0 it means that the length was probably truncated and we need to rewrite the length byte
if (save_end <= 0 && len > 0) len--; // leave off the last byte if we are ignoring it
mop[start] = len;
mop_len -= len;
if (save < 0) {
// this would split an entry, so don't write the first byte,
// instead zero it out and don't count as many bytes used
mop_len++;
save2 = mop[start + 1];
mop[start + 1] = 0;
}
data.mv_data = &mop[start];
//mdb_midl_print(stderr, &mop[start]);
rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
mop[start] = save;
if (save2) mop[start + 1] = save2;
if (rc)
break;
} else {// else we are just skipping this write
//fprintf(stderr, "Skipping write of freelist %u from %u-%u", id, start, end);
}
if (save_end < 0) mop[end] = save_end;
if (save2 > 0) mop[start + 1] = save2;
if (rc)
break;
}
// reset the start and end to indicate no writes have taken place
env->me_freelist_written_start = 0x7fffffff;
env->me_freelist_written_end = 0;
env->me_pghead = pghead;
/* Restore this so we can check vs. dirty_list after mdb_page_flush() */
if (! (txn->mt_flags & MDB_TXN_WRITEMAP))
Expand Down
3 changes: 2 additions & 1 deletion dependencies/lmdb/libraries/liblmdb/midl.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -123,7 +123,8 @@ int mdb_midl_insert( MDB_IDL* ids_ref, MDB_ID id, int insertion_count )
else next_count = 1;
if (id - next_count <= next_id && next_id > 0) {
if (id - next_count < next_id) {
fprintf(stderr, "overlapping duplicate entry");
fprintf(stderr, "overlapping duplicate entry %u\n", id);
mdb_tassert(ids, 0);
return -1;
}
// connected to next entry
Expand Down

0 comments on commit 23cdfee

Please sign in to comment.