The purpose of commit 282d8998 ("srcu: Prevent expedited GPs
and blocking readers from consuming CPU") was to prevent a long
series of never-blocking expedited SRCU grace periods from blocking
kernel-live-patching (KLP) progress.  Although it was successful, it also
resulted in excessive boot times on certain embedded workloads running
under qemu with the "-bios QEMU_EFI.fd" command line.  Here "excessive"
means increasing the boot time up into the three-to-four minute range.
This increase in boot time was due to the more than 6000 back-to-back
invocations of synchronize_rcu_expedited() within the KVM host OS, which
in turn resulted from qemu's emulation of a long series of MMIO accesses.

Commit 640a7d37c3f4 ("srcu: Block less aggressively for expedited grace
periods") did not significantly help this particular use case.

Zhangfei Gao and Shameerali Kolothum Thodi did experiments varying the
value of SRCU_MAX_NODELAY_PHASE with HZ=250 and with various values
of non-sleeping per phase counts on a system with preemption enabled,
and observed the following boot times:

+──────────────────────────+────────────────+
| SRCU_MAX_NODELAY_PHASE   | Boot time (s)  |
+──────────────────────────+────────────────+
| 100                      | 30.053         |
| 150                      | 25.151         |
| 200                      | 20.704         |
| 250                      | 15.748         |
| 500                      | 11.401         |
| 1000                     | 11.443         |
| 10000                    | 11.258         |
| 1000000                  | 11.154         |
+──────────────────────────+────────────────+

Analysis on the experiment results show additional improvements with
CPU-bound delays approaching one jiffy in duration. This improvement was
also seen when number of per-phase iterations were scaled to one jiffy.

This commit therefore scales per-grace-period phase number of non-sleeping
polls so that non-sleeping polls extend for about one jiffy. In addition,
the delay-calculation call to srcu_get_delay() in srcu_gp_end() is
replaced with a simple check for an expedited grace period.  This change
schedules callback invocation immediately after expedited grace periods
complete, which results in greatly improved boot times.  Testing done
by Marc and Zhangfei confirms that this change recovers most of the
performance degradation in boottime; for CONFIG_HZ_250 configuration,
specifically, boot times improve from 3m50s to 41s on Marc's setup;
and from 2m40s to ~9.7s on Zhangfei's setup.

In addition to the changes to default per phase delays, this
change adds 3 new kernel parameters - srcutree.srcu_max_nodelay,
srcutree.srcu_max_nodelay_phase, and srcutree.srcu_retry_check_delay.
This allows users to configure the srcu grace period scanning delays in
order to more quickly react to additional use cases.

Fixes: 640a7d37c3f4 ("srcu: Block less aggressively for expedited grace periods")
Fixes: 282d8998 ("srcu: Prevent expedited GPs and blocking readers from consuming CPU")
Reported-by: Zhangfei Gao <zhangfei.gao@linaro.org>
Reported-by: yueluck <yueluck@163.com>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@quicinc.com>
Tested-by: Marc Zyngier <maz@kernel.org>
Tested-by: Zhangfei Gao <zhangfei.gao@linaro.org>
Link: https://lore.kernel.org/all/20615615-0013-5adc-584f-2b1d5c03ebfc@linaro.org/Signed-off-by: Paul E. McKenney <paulmck@kernel.org>