Re: Support for "pshared" POSIX semaphores

[Jason Thorpe <thorpej%me.com@localhost>]

> On Feb 2, 2019, at 10:33 AM, Mindaugas Rasiukevicius <rmind%netbsd.org@localhost> wrote:
> 
> Jason Thorpe <thorpej%me.com@localhost> wrote:
>> Patch is here:
>> https://patch-diff.githubusercontent.com/raw/thorpej/netbsd-src/pull/5.diff
> 
> Thanks for working on this.
> 
> - Why not just add a new syscall, instead of the KSEM_PSHARED stuff?

Mainly because a new syscall is more intrusive, and harder to back-port.  It’s a shame the existing system call didn’t have a way to pass flags up.

> - If you add the 'fd' parameter to ksem_release(), then you can move
>  fd_putfile() there too (with fd != -1 case) and simplify a little bit.

I’ll take a look.  There are a few spots where I think the pattern still needs to “leak” out of ksem_release().

> - ksem_alloc_pshared_id: can ~KSEM_MARKER_MASK range be exhausted as an
>  attach vector?

You’d have to allocate a ton of pshared semaphores … it’s a 23-bit range, and that definitely exceeds the system max semaphores limit.  I assign it randomly only to make it harder to guess that the in-use semaphore IDs are (mainly as a way to avoid bad behavior caused by buggy programs).

(As an aside, the POSIX semaphore API is … annoyingly underspecified — there’s a lot of room for implementation-defined behavior that the spec doesn’t even bother to call out as “implementation-defined”).

> - Can you eliminate ksem_insert_pshared_locked(), ksem_remove_pshared_locked()
>  and ksem_remove_pshared() wrappers?  They serve no purpose.

I initially did that just to hide the “what kind of data structure contains the objects” away from the main logic.    I’ll go ahead and simplify, though.

> - ksem_lookup_pshared_locked: it makes sense to use a hash table here.
>  Using hashinit() is clumsy nowadays.. personally, I would replace most
>  of the subr_hash.c use cases with something like rhashmap [1] and much
>  more convenient get/put/del semantics.  Anyway, that is off-topic, so
>  up to you if you want to bother.

It does make sense to use something else .. I thought about using an r-b tree.  Given the default system limits, we’re talking about a small number of objects, but yah, I’ll use something better there.

-- thorpej