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:   Wed, 2 Nov 2022 19:47:20 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Ben Gardon <bgardon@google.com>
Cc:     linux-kernel@vger.kernel.org, kvm@vger.kernel.org,
        Paolo Bonzini <pbonzini@redhat.com>,
        David Matlack <dmatlack@google.com>,
        Anish Ghulati <aghulati@google.com>
Subject: Re: [PATCH] KVM: x86: Use SRCU to protect zap in
 __kvm_set_or_clear_apicv_inhibit
Message-ID: <Y2LJSE5nuHZJV7fF@google.com>
References: <20221102193020.1091939-1-bgardon@google.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20221102193020.1091939-1-bgardon@google.com>
Precedence: bulk

On Wed, Nov 02, 2022, Ben Gardon wrote:
> kvm_zap_gfn_range must be called in an SRCU read-critical section, but

Please add parantheses when referencing functions, i.e. kvm_zap_gfn_range().

> there is no SRCU annotation in __kvm_set_or_clear_apicv_inhibit.

__kvm_set_or_clear_apicv_inhibit()

> Add the needed SRCU annotation.

It's not an annotation, acquiring SRCU is very much functional code.

> Tested: ran tools/testing/selftests/kvm/x86_64/debug_regs on a DBG
> 	build. This patch causes the suspicious RCU warning to disappear.
> 	Note that the warning is hit in __kvm_zap_rmaps, so
> 	kvm_memslots_have_rmaps must return true in order for this to
> 	repro (i.e. the TDP MMU must be off or nesting in use.)

Please provide the stack trace or at least a verbal description of what paths
can reach __kvm_set_or_clear_apicv_inhibit() without holding SRCU, i.e. explain
why this bug isn't being hit left and right.

E.g.

  Unconditionally take KVM's SRCU lock in __kvm_set_or_clear_apicv_inhibit()
  when zapping virtual APIC SPTEs.  SRCU must be held when zapping SPTEs in
  shadow MMUs to protect the gfn=>memslot translation (the TDP MMU walks all
  roots and so doesn't dereference memslots).

  In most cases, the inhibits are updated during KVM_RUN and so SRCU is
  already held, but other ioctls() can also modify inhibits and don't
  acquire SRCU, e.g. KVM_SET_GUEST_DEBUG and KVM_SET_LAPIC.  Acquire SRCU
  unconditionally to avoid playing whack-a-mole, as nesting SRCU locks is
  safe and this is not a hot path.

> Fixes: 36222b117e36 ("KVM: x86: don't disable APICv memslot when inhibited")

Reported-by?  IIRC this originated in a syzkaller report?