High priority MDL_BACKUP_COMMIT_HIGH_PRIO and Slave-side BACKUP-COMMIT tracking to prevent parallel-slave/backup deadlock

[hemantdangi-gc]

A parallel-slave worker that has acquired MDL_BACKUP_COMMIT and is
waiting for a prior commit can be deadlocked by an incoming
BACKUP STAGE BLOCK_COMMIT: the prior commit needs MDL_BACKUP_COMMIT too
but the standard fairness rule denies it while the backup X-request is
in the waiting queue.

Design

Add MDL_BACKUP_COMMIT_HIGH_PRIO, a sibling lock mode of
MDL_BACKUP_COMMIT. The only matrix difference is one cell of
m_waiting_incompatible: HIGH_PRIO is granted even when a BACKUP
X-request is waiting.

A committing parallel-slave worker:

1. registers its sub_id in backup_team,
2. queries has_higher() — true iff a teammate in the same domain has
   a strictly greater sub_id,
3. requests MDL_BACKUP_COMMIT_HIGH_PRIO if true, otherwise
   MDL_BACKUP_COMMIT,
4. unregisters after release_lock().

Register-before-decide ordering ensures that any teammate arriving
later observes our sub_id, and we observe at least our own back.

How it solves the problem

When backup's X-request is in the waiting queue and a worker that is
needed to unblock a prior commit arrives, that worker observes a
later-numbered teammate already holding the BACKUP lock, requests
HIGH_PRIO, and is granted past the X-request via the matrix. The
prior-commit chain drains; the backup X-request is then granted.

The teammate approach made the MDL scheduler call a slave-supplied callback during lock acquisition: when MDL_BACKUP_COMMIT was blocked, MDL itself walked m_granted.
This new approach inverts the ownership: the slave maintains its own per-(rli, domain) multiset of in-flight gtid_sub_ids on Relay_log_info, and before requesting the lock, the worker reads this slave-owned state to pick its lock mode — MDL_BACKUP_COMMIT_HIGH_PRIO if a teammate with greater sub_id is registered, else MDL_BACKUP_COMMIT. MDL then just consults its compatibility/priority matrices like for any other lock; the leapfrog is a pure matrix property of the new mode, not a scheduler-side override.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ hemantdangi-gc
❌ emoonrain
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[gemini-code-assist]

Code Review

This pull request addresses deadlocks between parallel replication and backup operations (MDEV-36025) by introducing a high-priority Metadata Lock (MDL_BACKUP_COMMIT_HIGH_PRIO) and a Backup_commit_team mechanism. Slave workers now register their gtid_sub_id and can escalate their lock priority to bypass waiting backup locks when a teammate with a higher ID is already in flight, replacing previous logic that released locks during waits. Feedback was provided to optimize memory usage in the Backup_commit_team class by erasing empty domain entries from the internal map and simplifying the is_empty function accordingly.

[hemantdangi-gc]

@svoj
Please review:
High priority MDL_BACKUP_COMMIT_HIGH_PRIO and Slave-side BACKUP-COMMIT tracking to prevent parallel-slave/backup deadlock
a20023b