Re: [PATCH 1/2] LoongArch: KVM: Protect kvm_check_requests() with SRCU

[Huacai Chen]

On Tue, Dec 3, 2024 at 4:27 PM bibo mao <maobibo@xxxxxxxxxxx> wrote:
>
>
>
> On 2024/12/3 下午2:50, Huacai Chen wrote:
> > When we enable lockdep we get such a warning:
> >
> >   =============================
> >   WARNING: suspicious RCU usage
> >   6.12.0-rc7+ #1891 Tainted: G        W
> >   -----------------------------
> >   include/linux/kvm_host.h:1043 suspicious rcu_dereference_check() usage!
> >   other info that might help us debug this:
> >   rcu_scheduler_active = 2, debug_locks = 1
> >   1 lock held by qemu-system-loo/948:
> >    #0: 90000001184a00a8 (&vcpu->mutex){+.+.}-{4:4}, at: kvm_vcpu_ioctl+0xf4/0xe20 [kvm]
> >   stack backtrace:
> >   CPU: 0 UID: 0 PID: 948 Comm: qemu-system-loo Tainted: G        W          6.12.0-rc7+ #1891
> >   Tainted: [W]=WARN
> >   Hardware name: Loongson Loongson-3A5000-7A1000-1w-CRB/Loongson-LS3A5000-7A1000-1w-CRB, BIOS vUDK2018-LoongArch-V2.0.0-prebeta9 10/21/2022
> >   Stack : 0000000000000089 9000000005a0db9c 90000000071519c8 900000012c578000
> >           900000012c57b920 0000000000000000 900000012c57b928 9000000007e53788
> >           900000000815bcc8 900000000815bcc0 900000012c57b790 0000000000000001
> >           0000000000000001 4b031894b9d6b725 0000000004dec000 90000001003299c0
> >           0000000000000414 0000000000000001 000000000000002d 0000000000000003
> >           0000000000000030 00000000000003b4 0000000004dec000 90000001184a0000
> >           900000000806d000 9000000007e53788 00000000000000b4 0000000000000004
> >           0000000000000004 0000000000000000 0000000000000000 9000000107baf600
> >           9000000008916000 9000000007e53788 9000000005924778 0000000010000044
> >           00000000000000b0 0000000000000004 0000000000000000 0000000000071c1d
> >           ...
> >   Call Trace:
> >   [<9000000005924778>] show_stack+0x38/0x180
> >   [<90000000071519c4>] dump_stack_lvl+0x94/0xe4
> >   [<90000000059eb754>] lockdep_rcu_suspicious+0x194/0x240
> >   [<ffff8000022143bc>] kvm_gfn_to_hva_cache_init+0xfc/0x120 [kvm]
> >   [<ffff80000222ade4>] kvm_pre_enter_guest+0x3a4/0x520 [kvm]
> >   [<ffff80000222b3dc>] kvm_handle_exit+0x23c/0x480 [kvm]
> >
> > Fix it by protecting kvm_check_requests() with SRCU.
> >
> > Cc: stable@xxxxxxxxxxxxxxx
> > Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
> > ---
> >   arch/loongarch/kvm/vcpu.c | 4 +++-
> >   1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
> > index cab1818be68d..d18a4a270415 100644
> > --- a/arch/loongarch/kvm/vcpu.c
> > +++ b/arch/loongarch/kvm/vcpu.c
> > @@ -240,7 +240,7 @@ static void kvm_late_check_requests(struct kvm_vcpu *vcpu)
> >    */
> >   static int kvm_enter_guest_check(struct kvm_vcpu *vcpu)
> >   {
> > -     int ret;
> > +     int idx, ret;
> >
> >       /*
> >        * Check conditions before entering the guest
> > @@ -249,7 +249,9 @@ static int kvm_enter_guest_check(struct kvm_vcpu *vcpu)
> >       if (ret < 0)
> >               return ret;
> >
> > +     idx = srcu_read_lock(&vcpu->kvm->srcu);
> >       ret = kvm_check_requests(vcpu);
> > +     srcu_read_unlock(&vcpu->kvm->srcu, idx);
> >
> >       return ret;
> >   }
> >
> How about adding rcu readlock with closest function
> kvm_update_stolen_time()?
I have considered this method before. But then I read vcpu_run() of
x86, it protect the whole vcpu_run() except  the subroutine
xfer_to_guest_mode_handle_work(), so I think protect the whole
kvm_check_requests() is more like x86.

Huacai

>
>   static int kvm_check_requests(struct kvm_vcpu *vcpu)
>   {
> +       int idx;
> +
>          if (!kvm_request_pending(vcpu))
>                  return RESUME_GUEST;
>
> @@ -213,8 +215,11 @@ static int kvm_check_requests(struct kvm_vcpu *vcpu)
>          if (kvm_dirty_ring_check_request(vcpu))
>                  return RESUME_HOST;
>
> -       if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu))
> +       if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu)) {
> +               idx = srcu_read_lock(&vcpu->kvm->srcu);
>                  kvm_update_stolen_time(vcpu);
> +               srcu_read_unlock(&vcpu->kvm->srcu, idx);
> +       }
>
>          return RESUME_GUEST;
>   }
>
> Both method look good to me.
>
> Regards
> Bibo Mao
>