MDS/openide
1
0
Fork 0
mirror of https://gitflic.ru/project/openide/openide.git synced 2026-02-04 23:39:07 +07:00
Code Issues Packages Projects Releases Wiki Activity

[vfs] fixed ExtendibleHashMap bug

Browse Source 
+ incorrect processing of tombstones in highly-filled map, could led to 'Table is full' even though there are slots to use

GitOrigin-RevId: d7ac2e9f2a6f236587bee0022e963fad55b73a2b
This commit is contained in:
Ruslan Cheremin
2024-05-09 17:56:46 +02:00
committed by intellij-monorepo-bot
parent 23c784b02f
commit 83ec59fe3c
2 changed files with 26 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
platform/util/storages/src/com/intellij/platform/util/io/storages/intmultimaps/Int2IntMultimap.java
View File

@@ -149,12 +149,8 @@ public final class Int2IntMultimap {
// But if there are no alive entries anymore -- we _can_ clear everything without breaking anything.
Arrays.fill(table, NO_VALUE);
filledSlots = 0;
put(key, value);
//MAYBE RC: instead of waiting for table to be full of tombstones -- shrink the table earlier, then
// e.g. (aliveEntries < 0.5..0.2 filledSlots)?
// Otherwise performance degradation could be quite significant long before we come to this point!
}
if (firstTombstoneIndex != -1) {

32
platform/util/storages/src/com/intellij/platform/util/io/storages/intmultimaps/extendiblehashmap/ExtendibleHashMap.java
View File

@@ -1014,11 +1014,18 @@ public class ExtendibleHashMap implements DurableIntToMultiIntMap, Unmappable {
}
}
//MAYBE RC: it could be there are not-so-many alive entries, but a lot of tombstones, and no free slots remain.
// This is because in the current design we never clear tombstones: during segment split we copy half of alive
// entries to a new segment, leaving tombstones in old one -- but we never clean the tombstones in the old segment.
// Tombstones are somewhat 'cleaned' by reusing their slots for new records, but this is stochastic, and could be
// not very effective
//probing sequence went through all the table: i.e. table is full -- but maybe there are tombstones to replace?
//TODO RC: In the current design we never clear tombstones: during segment split we copy half of alive entries
// to a new segment, but we leave the tombstones in old one -- hence, we never clean the tombstones in the old
// segment. Tombstones are somewhat 'cleaned' by reusing their slots for new records, but this is stochastic,
// and could be not very effective.
// Hence, in the current design it could be there are not-so-many alive entries, but a lot of tombstones -- which
// deteriorates performance.
// The simplest solution would be to prune the tombstones on segment split in old segment also: currently we copy
// ~1/2 alive entries in the new segment -- instead we should copy _all_ alive entries into memory buffer first,
// clean the old segment entirely, and copy the entries from in-memory buffer into old/new segments afterwards.
if (aliveValues(table) == 0) {
//If there is 0 alive records => it is OK to clear all the tombstones.
// We can't clear all tombstones while alive entries exist because such a cleaning breaks lookup: we treat
@@ -1037,11 +1044,22 @@ public class ExtendibleHashMap implements DurableIntToMultiIntMap, Unmappable {
return put(table, key, value);
}
if (firstTombstoneIndex != -1) {
//Probing sequence is terminated if:
// 1) key is found
// 2) free slot is found
// 3) no more slots to look for key -- all slots was visited (=key is definitely absent in the map)
// So, if we scan all the entries and not found a key, but we have found a tombstone along the way
// => we can replace it with the entry at hand:
table.updateEntry(firstTombstoneIndex, key, value);
incrementAliveValues(table);
return true;
}
//Table must be resized well before such a condition occurs!
throw new AssertionError(
"Table is full: all " + capacity + " items were traversed, but no free slot found " +
"table.aliveEntries: " + table.aliveEntriesCount() + ", table: " + table
"Table is full: all " + capacity + " items were traversed, but no free slot found, " +
"table: " + table
);
}