need a working rw_tryupgrade()

[dweeezil]

The lack of a working rw_tryupgrade() in all but the CONFIG_RWSEM_GENERIC_SPINLOCK case (which is pretty rare these days) can have ill effects if the upstream OpenZFS code uses it in certain performance-sensitive areas.

For example, openzfs/zfs@7f60329 introduced a statement which inhibits creation of new prefetch streams when rw_tryupgrade() returns zero.

The problem stems from the imperfect mapping between the Solaris krwlock facility and Linux's rw_semaphore facility (see d28db80 and others for details). In addition to the differing API, Linux does not natively support an upgrade from a reader lock to a reader/writer lock (but it does support the reverse).

Since the entire Solaris krwlock facility is implemented as macros which expand to statement expressions and also since the krwlock_t structure is a wrapper for rw_semaphore, there's a great deal of flexibility available within the SPL. It seems that, for example, it might be possible to add additional members to krwlock_t to help track such things as pending waiters for write or any other accounting and/or spinlocks which might be useful to implement a semantically equivalent rw_tryupgrade().

[EDIT] Referencing openzfs/zfs#4388.

[dweeezil]

After auditing the other users of rw_tryupgrade() in ZFS, it doesn't seem this is worth doing at the moment. The code mentioned in openzfs/zfs@7f60329 is the only place in which it's used without effectively immediately dropping and re-acquiring a RW lock in the failure case.

[dweeezil]

Closing. As indicated in openzfs/zfs#4388 (comment) and also from my own audit, this isn't worth implementing, but at least it's mentioned here for documentation in the future.