| Message ID | 20231218020249.3447973-3-i.maximets@ovn.org |
|---|---|
| State | Accepted |
| Commit | 6f11d9daad526bf51ee1538c7451a04572dc7a12 |
| Delegated to: | Ilya Maximets |
| Headers | show |
| Series | ovsdb: transaction: Bug Fixes + Reuse of column diffs. | expand |
| Context | Check | Description |
|---|---|---|
| ovsrobot/apply-robot | success | apply and check: success |
| ovsrobot/github-robot-_Build_and_Test | success | github build: passed |
| ovsrobot/intel-ovs-compilation | fail | test: fail |
On Sun, Dec 17, 2023 at 9:03 PM Ilya Maximets <i.maximets@ovn.org> wrote: > > While counting strong references current code first generates a > difference between old and new datums and only after it checks > the types of the atoms to be strong references. > > Similar thing happens while assessing weak references. First > the added/removed are generated and then we check for atoms > to be weak references. > > Check the type first to avoid unnecessary work. This change > doubles the performance of transactions that modify large sets > of references. > > For example, with this change applied, initial read of OVSDB > file containing 136K transactions of large OVN port groups > and address sets on my laptop takes 24 seconds vs 43 seconds > without. > > Fixes: 4dbff9f0a685 ("ovsdb: transaction: Incremental reassessment of weak refs.") > Fixes: b2712d026eae ("ovsdb: transaction: Use diffs for strong reference counting.") > Signed-off-by: Ilya Maximets <i.maximets@ovn.org> Acked-by: Mike Pattrick <mkp@redhat.com>
diff --git a/lib/ovsdb-types.h b/lib/ovsdb-types.h index 9777efea3..688fe5633 100644 --- a/lib/ovsdb-types.h +++ b/lib/ovsdb-types.h @@ -238,6 +238,18 @@ static inline bool ovsdb_type_is_map(const struct ovsdb_type *type) return type->value.type != OVSDB_TYPE_VOID; } +static inline bool ovsdb_type_has_strong_refs(const struct ovsdb_type *type) +{ + return ovsdb_base_type_is_strong_ref(&type->key) + || ovsdb_base_type_is_strong_ref(&type->value); +} + +static inline bool ovsdb_type_has_weak_refs(const struct ovsdb_type *type) +{ + return ovsdb_base_type_is_weak_ref(&type->key) + || ovsdb_base_type_is_weak_ref(&type->value); +} + #ifdef __cplusplus } #endif diff --git a/ovsdb/transaction.c b/ovsdb/transaction.c index f43533a8c..bbe4cddc1 100644 --- a/ovsdb/transaction.c +++ b/ovsdb/transaction.c @@ -322,7 +322,8 @@ update_row_ref_count(struct ovsdb_txn *txn, struct ovsdb_txn_row *r) const struct ovsdb_column *column = node->data; struct ovsdb_error *error; - if (bitmap_is_set(r->changed, column->index)) { + if (bitmap_is_set(r->changed, column->index) + && ovsdb_type_has_strong_refs(&column->type)) { if (r->old && !r->new) { error = ovsdb_txn_adjust_row_refs( txn, r->old, column, @@ -718,6 +719,10 @@ assess_weak_refs(struct ovsdb_txn *txn, struct ovsdb_txn_row *txn_row) unsigned int orig_n; bool zero = false; + if (!ovsdb_type_has_weak_refs(&column->type)) { + continue; + } + orig_n = datum->n; /* Collecting all key-value pairs that references deleted rows. */
While counting strong references current code first generates a difference between old and new datums and only after it checks the types of the atoms to be strong references. Similar thing happens while assessing weak references. First the added/removed are generated and then we check for atoms to be weak references. Check the type first to avoid unnecessary work. This change doubles the performance of transactions that modify large sets of references. For example, with this change applied, initial read of OVSDB file containing 136K transactions of large OVN port groups and address sets on my laptop takes 24 seconds vs 43 seconds without. Fixes: 4dbff9f0a685 ("ovsdb: transaction: Incremental reassessment of weak refs.") Fixes: b2712d026eae ("ovsdb: transaction: Use diffs for strong reference counting.") Signed-off-by: Ilya Maximets <i.maximets@ovn.org> --- lib/ovsdb-types.h | 12 ++++++++++++ ovsdb/transaction.c | 7 ++++++- 2 files changed, 18 insertions(+), 1 deletion(-)