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;
MIME-Version: 1.0
References: <CAPm50a+aygp3T1mNjzGXtL2nyNm-mHFZ3YO8F7eO0gCxZDuQsA@mail.gmail.com>
 <Y0BeEVxkDkctmTIX@google.com> <CAPm50aLzOLyURhvhYkCyp1hpRagAczFXg9jYbFg_86Qaf5usbg@mail.gmail.com>
 <Y0RYp7CZO5u1Eg/s@google.com>
In-Reply-To: <Y0RYp7CZO5u1Eg/s@google.com>
From:   Hao Peng <flyingpenghao@gmail.com>
Date:   Mon, 24 Oct 2022 11:27:38 +0800
Message-ID: <CAPm50aJUjCumrMmEGCw2eNc6nPi5y=ZzTMi8MC2aJwkatQGnww@mail.gmail.com>
Subject: Re: [PATCH] kvm: x86: keep srcu writer side operation mutually exclusive
To:     Sean Christopherson <seanjc@google.com>
Cc:     pbonzini@redhat.com, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, Oct 11, 2022 at 1:38 AM Sean Christopherson <seanjc@google.com> wrote:
>
> On Sun, Oct 09, 2022, Hao Peng wrote:
> > On Sat, Oct 8, 2022 at 1:12 AM Sean Christopherson <seanjc@google.com> wrote:
> > >
> > > On Sat, Oct 08, 2022, Hao Peng wrote:
> > > > From: Peng Hao <flyingpeng@tencent.com>
> > > >
> > > > Synchronization operations on the writer side of SRCU should be
> > > > invoked within the mutex.
> > >
> > > Why?  Synchronizing SRCU is necessary only to ensure that all previous readers go
> > > away before the old filter is freed.  There's no need to serialize synchronization
> > > between writers.  The mutex ensures each writer operates on the "new" filter that's
> > > set by the previous writer, i.e. there's no danger of a double-free.  And the next
> > > writer will wait for readers to _its_ "new" filter.
> > >
> > Array srcu_lock_count/srcu_unlock_count[] in srcu_data, which is used
> > alternately to determine
> > which readers need to wait to get out of the critical area. If  two
> > synchronize_srcu are initiated concurrently,
> > there may be a problem with the judgment of gp. But if it is confirmed
> > that there will be no writer concurrency,
> > it is not necessary to ensure that synchronize_srcu is executed within
> > the scope of the mutex lock.
>
> I don't see anything in the RCU documentation or code that suggests that callers
> need to serialize synchronization calls.  E.g. the "tree" SRCU implementation uses
> a dedicated mutex to serialize grace period work
>
>         struct mutex srcu_gp_mutex;             /* Serialize GP work. */
>
> static void srcu_advance_state(struct srcu_struct *ssp)
> {
>         int idx;
>
>         mutex_lock(&ssp->srcu_gp_mutex);
>
>         <magic>
> }
>
>
> and its state machine explicitly accounts for "Someone else" starting a grace
> period
>
>                 if (idx != SRCU_STATE_IDLE) {
>                         mutex_unlock(&ssp->srcu_gp_mutex);
>                         return; /* Someone else started the grace period. */
>                 }
>
> and srcu_gp_end() also guards against creating more than 2 grace periods.
>
>         /* Prevent more than one additional grace period. */
>         mutex_lock(&ssp->srcu_cb_mutex);
>
> And if this is a subtle requirement, there is a lot of broken kernel code, e.g.
> mmu_notifier, other KVM code, srcu_notifier_chain_unregister(), etc...

srcu_gp_mutex is meaningless because the workqueue already guarantees
that the same work_struct will not be reentrant.
If synchronize_srcu is not mutually exclusive on the update side, it may cause
a GP to fail for a long time. I will continue to analyze when I have time.