RE: [RFC v7 3/6] eal: add exactly-once bit access functions

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Mattias Rönnblom [mailto:hofors@lysator.liu.se]
> Sent: Wednesday, 8 May 2024 08.47
> 
> On 2024-05-07 21:17, Morten Brørup wrote:
> >> From: Mattias Rönnblom [mailto:mattias.ronnblom@ericsson.com]
> >> Sent: Sunday, 5 May 2024 10.38
> >>
> >> Add test/set/clear/assign/flip functions which prevents certain
> >> compiler optimizations and guarantees that program-level memory loads
> >> and/or stores will actually occur.
> >>
> >> These functions are useful when interacting with memory-mapped
> >> hardware devices.
> >>
> >> The "once" family of functions does not promise atomicity and provides
> >> no memory ordering guarantees beyond the C11 relaxed memory model.
> >
> > In another thread, Stephen referred to the extended discussion on memory
> models in Linux kernel documentation:
> > https://www.kernel.org/doc/html/latest/core-api/wrappers/memory-
> barriers.html
> >
> > Unlike the "once" family of functions in this RFC, the "once" family of
> functions in the kernel also guarantee memory ordering, specifically for
> memory-mapped hardware devices. The document describes the rationale with
> examples.
> >
> 
> What more specifically did you have in mind? READ_ONCE() and
> WRITE_ONCE()? They give almost no guarantees. Very much relaxed.

The way I read it, they do provide memory ordering guarantees.

Ignore that the kernel's "once" functions operates on words and this RFC operates on bits, the behavior is the same. Either there are memory ordering guarantees, or there are not.

> 
> I've read that document.
> 
> What you should keep in mind if you read that document, is that DPDK
> doesn't use the kernel's memory model, and doesn't have the kernel's
> barrier and atomics APIs. What we have are an obsolete, miniature
> look-alike in <rte_atomic.h> and something C11-like in <rte_stdatomic.h>.
> 
> My general impression is that DPDK was moving in the C11 direction
> memory model-wise, which is not the model the kernel uses.

I think you and I agree that using legacy methods only because "the kernel does it that way" would not be the optimal roadmap for DPDK.

We should keep moving in the C11 direction memory model-wise.
I consider it more descriptive, and thus expect compilers to eventually produce better optimized code.

> 
> > It makes me think that DPDK "once" family of functions should behave
> similarly.
> 
> I think they do already.

I haven't looked deep into it, but the RFC's documentation says otherwise:
The "once" family of functions does not promise atomicity and provides *no memory ordering* guarantees beyond the C11 relaxed memory model.

> 
> Also, rte_bit_once_set() works as the kernel's __set_bit().
> 
> > Alternatively, if the "once" family of functions cannot be generically
> implemented with a memory ordering that is optimal for all use cases, drop
> this family of functions, and instead rely on the "atomic" family of functions
> for interacting with memory-mapped hardware devices.
> >
> >>
> >> RFC v7:
> >>   * Fix various minor issues in documentation.
> >>
> >> RFC v6:
> >>   * Have rte_bit_once_test() accept const-marked bitsets.
> >>
> >> RFC v3:
> >>   * Work around lack of C++ support for _Generic (Tyler Retzlaff).
> >>
> >> Signed-off-by: Mattias Rönnblom <mattias.ronnblom@ericsson.com>
> >> Acked-by: Morten Brørup <mb@smartsharesystems.com>
> >> Acked-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
> >> ---
> >