On Tue, 16 Mar 2021, Nicholas Piggin wrote:

One request, could you add a comment in place that references
smp_cond_load_relaxed() so this commit can be found again if
someone looks at it? Something like this

/*
* smp_cond_load_relaxed was found to have performance problems if
* implemented with spin_begin()/spin_end().
*/

Sure, let me see where I can fit that in and send out a v2.

Similarly, but unrelated to this patch, is there any chance we could
remove the whole spin_until_cond() machinery and make it specific to
powerpc? This was introduced in 2017 and doesn't really have any users
outside of powerpc, except for these:

drivers/firmware/arm_scmi/driver.c:             spin_until_cond(scmi_xfer_done_no_timeout(cinfo, xfer, stop));
drivers/firmware/arm_scmi/shmem.c:      spin_until_cond(ioread32(&shmem->channel_status) &
drivers/net/ethernet/xilinx/ll_temac_main.c:    spin_until_cond(hard_acs_rdy_or_timeout(lp, timeout));

... which afaict only the xilinx one can actually build on powerpc.
Regardless, these could be converted to smp_cond_load_relaxed(), being
the more standard way to do optimized busy-waiting, caring more about
the family of barriers than ad-hoc SMT priorities. Of course, I have
no way of testing any of these changes.

I wonder if it should have a Fixes: tag to the original commit as
well.

I'm not sure either. I've actually been informed recently of other
workloads that benefit from the revert on large Power9 boxes. So I'll
go ahead and add it.

Otherwise,

Acked-by: Nicholas Piggin <npiggin@gmail.com>

Thanks,
Davidlohr