Details
Critical
Description
When lf_delete() — for example — searches in the hash it may need to allocate and insert a new dummy node, the latter may fail with OOM. The caller normally doesn't expect OOM from searches or deletes and it's not always easy to handle OOM there.
Instead lf_delete() and lf_search() can be changed never to return OOM. Indeed, when a search cannot insert a new dummy node, it simply needs to start searching from the previous dummy node, which is guaranteed to exist.
It's even easier in lf_hash_iterate() — it only ever looks for the very first bucket, dummy node 0. If that node does not exists, the hash is most certainly empty.
Gliffy Diagrams
Attachments
Activity
- All
- Comments
- Work Log
- History
- Activity
- Transitions
It looks very good, thanks for implementing!
I just have one more wish (better to be implemented separately): make lf_hash_get_pins() never to return OOM.
One way to solve this is to allow to preallocate max possible number of pins (e.g. table cache doesn't need more pins than max_connections?).
Another way is to have static pins protected by mutex. They will be returned on OOM only.
Sergey Vojtovich, please, take a look at the patch.
https://github.com/MariaDB/server/tree/bb-lf-no-oom