Looks very good!

Looks good!

@jhb : OK, if you can move lockdep_is_held() under #ifdef INVARIANTS and otherwise define lockdep_is_held() as 1, I think we are in-line.

Ok, can you tell me what the semantic in Linux is for lockdep_is_held() after on a lock locked via read_lock()?

If we look at Linux all the LOCKDEP stuff is under ifdef:

Are we certain that lockep() is not used with read_lock()? Also I don't like that cast to lock_object. It would be better to do this clean by referring to the lock object field in the lock, assuming they are named then same. Or use WITNESS for selected locks, which is much better. The reason WITNESS is currently disabled for LinuxKPI spinlocks is that spinlocks are not destroyed under Linux, so the WITNESS object never goes away.

@hselasky Do you have any thoughts on the LOCK_CLASS-based approach? Is this something you can test? I've only tested it with drm.

Please test!

Can you try this patch and report back?

Rebase patch after latest changes upstream.

Please test.

Can you test this patch?

Can you test this updated patch?

Adding Slava to this review aswell.

Looks good. Maybe you should also bump the __FreeBSD_version when you commit this.

There is no regular (operating on the integer regiter file) instructions on amd64 that could add e.g. three registers once. There is a LEA instruction which allows you to add to registers and a constant, but thats all.

I don't understand this comment. Can you elaborate? IMO, we want to differentiate from Linux when and only when it is possible to do better than Linux does; not gratuitously.

Maybe the following is unclear:

If the functions are not broken there is no reason to fix them from my POV.

OK, I'll try to have a look at your patch.

Can you abandon this differential revision if no longer active?

Ping - is this patch still relevant?

Was patch updated?

Can this device be used via /usr/ports/multimedia/webcamd ?

Ping: Can this change be abandoned now?

Changes look good to me.

@jhibbits : Can you update the patch with the agreed upon changes?

Just remove lockdep_is_held(m) .

@zeising : Do you use CONFIG_LOCKDEP in drm-next ?

Assume that the driver knows not to enable both MSIX and MSI.

I'll have a closer look tomorrow. Most of this looks good.

Looks good!

I'll have a closer look at you patch(es) when I'm back in office. Please stay tuned.

What about lockedep_is_held() ??

Please stay tuned!

Please re-base this patch on top of the latest 13-current.

Looks good. Some minor nits. Please fix before commit.

Patch looks good to me. Did you run with this patch?

Patch looks good to me. Did you run this patch, I.E. force the error case somehow?

Can you upload full context for the diff?

Spell lkpi -> LinuxKPI in the commit message.

Optimize epoch_drain_cb() by using an atomic operation.

Update manual page to describe new epoch(9) APIs.