Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Fri, 14 Oct 2022 22:47:34 +0000
From:   Joel Fernandes <joel@joelfernandes.org>
To:     "Paul E. McKenney" <paulmck@kernel.org>
Cc:     rcu@vger.kernel.org, linux-kernel@vger.kernel.org,
        kernel-team@fb.com, rostedt@goodmis.org, tglx@linutronix.de,
        john.ogness@linutronix.de, pmladek@suse.com
Subject: Re: [PATCH v2 rcu 0/8] NMI-safe SRCU reader API
Message-ID: <Y0nnBjTHgoIVYMrJ@google.com>
References: <20220921144620.GA1200846@paulmck-ThinkPad-P17-Gen-1>
 <20220929180714.GA2874192@paulmck-ThinkPad-P17-Gen-1>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220929180714.GA2874192@paulmck-ThinkPad-P17-Gen-1>
Precedence: bulk

On Thu, Sep 29, 2022 at 11:07:14AM -0700, Paul E. McKenney wrote:
> Hello!
> 
> This RFC series provides the second version of an NMI-safe SRCU reader API
> in the guise of srcu_read_lock_nmisafe() and srcu_read_unlock_nmisafe().

Just a comment about high-level use of the new NMI-safe SRCU api: say if for
certain arch, NMI cannot be interrupted (by breakpoint or something), then
using NMI-safe API will force such arch to potentially use more expensive RMW
atomic than the previously cheap local non-atomic operations that the arch
was able to get away with.

Thoughts? Otherwise, LGTM.

thanks,

 - Joel


> A given srcu_struct structure must use either the traditional
> srcu_read_lock() and srcu_read_unlock() API or the new _nmisafe() API:
> Mixing and matching is not permitted.  So much so that kernels built
> with CONFIG_PROVE_RCU=y will complain if you try it.
> 
> The reason for this restriction is that I have yet to find a use case
> that is not a accident waiting to happen.  And if free intermixing
> were permitted, it is pretty much a given that someone somewhere will
> get confused and use srcu_read_lock_nmisafe() within NMI handlers and
> srcu_read_lock() elsewhere, which will not (repeat, NOT) provide NMI
> safety.
> 
> I do not expect to push this into the v6.1 merge window.  However, if
> the printk() series that needs it goes in, then I will push it as a fix
> for the resulting regression.
> 
> The series is as follows:
> 
> 1.	Convert ->srcu_lock_count and ->srcu_unlock_count to atomic.
> 
> 2.	Create an srcu_read_lock_nmisafe() and srcu_read_unlock_nmisafe().
> 
> 3.	Check for consistent per-CPU per-srcu_struct NMI safety.
> 
> 4.	Check for consistent global per-srcu_struct NMI safety.
> 
> 5.	Add ARCH_HAS_NMI_SAFE_THIS_CPU_OPS Kconfig option.
> 
> 6.	Add ARCH_HAS_NMI_SAFE_THIS_CPU_OPS Kconfig option.
> 
> 7.	Add ARCH_HAS_NMI_SAFE_THIS_CPU_OPS Kconfig option.
> 
> 8.	Add ARCH_HAS_NMI_SAFE_THIS_CPU_OPS Kconfig option.
> 
> Changes since v1 RFC:
> 
> 1.	Added enabling patches for arm64, loongarch, s390, and x86.
> 	These have what appear to me to be NMI-safe this_cpu_inc()
> 	implementations.
> 
> 2.	Fix a build error on !SMP kernels built without SRCU.
> 
> 3.	Fix a build error on !SMP kernels.
> 
> 						Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
>  b/arch/arm64/Kconfig       |    1 
>  b/arch/loongarch/Kconfig   |    1 
>  b/arch/s390/Kconfig        |    1 
>  b/arch/x86/Kconfig         |    1 
>  b/include/linux/srcu.h     |   39 +++++++++++++++++++++
>  b/include/linux/srcutiny.h |   11 ++++++
>  b/include/linux/srcutree.h |    4 +-
>  b/kernel/rcu/Kconfig       |    3 +
>  b/kernel/rcu/rcutorture.c  |   11 ++++--
>  b/kernel/rcu/srcutree.c    |   24 ++++++-------
>  include/linux/srcu.h       |    4 +-
>  include/linux/srcutiny.h   |    4 +-
>  include/linux/srcutree.h   |   12 +++++-
>  kernel/rcu/srcutree.c      |   82 +++++++++++++++++++++++++++++++++++++++------
>  14 files changed, 166 insertions(+), 32 deletions(-)