mbox series

[net-next,0/9] Try to simplify the gnet_stats and remove qdisc->running sequence counter.

Message ID 20211016084910.4029084-1-bigeasy@linutronix.de (mailing list archive)
Headers show
Series Try to simplify the gnet_stats and remove qdisc->running sequence counter. | expand

Message

Sebastian Andrzej Siewior Oct. 16, 2021, 8:49 a.m. UTC 
The first few patches is a follow up to
    https://lore.kernel.org/all/20211007175000.2334713-1-bigeasy@linutronix.de/

The remaining patches (#5+) remove the seqcount_t (Qdisc::running) from
the Qdisc. The statistics (Qdisc::bstats and Qdisc::cpu_bstats) use
u64_stats_t and the "running state" is now represented by a bit in
Qdisc::state.

By removing the seqcount_t from Qdisc and decoupling the bstats
statistics from the seqcount_t it is possible to query the statistics
even if the Qdisc is running instead of waiting until it is idle again.

The try-lock like usage of the seqcount_t in qdisc_run_begin() is
problematic on PREEMPT_RT. Inside the qdisc_run_begin/end() qdisc->running
sequence counter write sections, at sch_direct_xmit(), the seqcount write
serialization lock is released then re-acquired. This is fine for !RT, because
the writer is in a BH disabled region and there is a no in-IRQ reader. For RT
though, BH sections are preemptible. The earlier introduced seqcount_LOCKNAME_t
mechanism, which for RT the reader acquires then relesaes the write
serailization lock to avoid infinite spinning if it preempts a seqcount write
section, cannot work: the qdisc->running write serialization lock is already
intermittingly released inside the seqcount write section.

Sebastian