Re: [PATCH] KVM: MMU: Fix softlockup due to mmu_lock is held too long

[Paolo Bonzini]

On 10/08/2017 15:55, Wanpeng Li wrote:
> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> 
> watchdog: BUG: soft lockup - CPU#5 stuck for 22s! [warn_test:3089]
>  irq event stamp: 20532
>  hardirqs last  enabled at (20531): [<ffffffff8e9b6908>] restore_regs_and_iret+0x0/0x1d
>  hardirqs last disabled at (20532): [<ffffffff8e9b7ae8>] apic_timer_interrupt+0x98/0xb0
>  softirqs last  enabled at (8266): [<ffffffff8e9badc6>] __do_softirq+0x206/0x4c1
>  softirqs last disabled at (8253): [<ffffffff8e083918>] irq_exit+0xf8/0x100
>  CPU: 5 PID: 3089 Comm: warn_test Tainted: G           OE   4.13.0-rc3+ #8
>  RIP: 0010:kvm_mmu_prepare_zap_page+0x72/0x4b0 [kvm]
>  Call Trace:
>   make_mmu_pages_available.isra.120+0x71/0xc0 [kvm]
>   kvm_mmu_load+0x1cf/0x410 [kvm]
>   kvm_arch_vcpu_ioctl_run+0x1316/0x1bf0 [kvm]
>   kvm_vcpu_ioctl+0x340/0x700 [kvm]
>   ? kvm_vcpu_ioctl+0x340/0x700 [kvm]
>   ? __fget+0xfc/0x210
>   do_vfs_ioctl+0xa4/0x6a0
>   ? __fget+0x11d/0x210
>   SyS_ioctl+0x79/0x90
>   entry_SYSCALL_64_fastpath+0x23/0xc2
>   ? __this_cpu_preempt_check+0x13/0x20
> 
> This can be reproduced readily by ept=N and running syzkaller tests since 
> many syzkaller testcases don't setup any memory regions. However, if ept=Y
> rmode identity map will be created, then kvm_mmu_calculate_mmu_pages() will 
> extend the number of VM's mmu pages to at least KVM_MIN_ALLOC_MMU_PAGES 
> which just hide the issue.
> 
> I saw the scenario kvm->arch.n_max_mmu_pages == 0 && kvm->arch.n_used_mmu_pages == 1,
> so there is one active mmu page on the list, kvm_mmu_prepare_zap_page() fails 
> to zap any pages, however prepare_zap_oldest_mmu_page() always returns true.
> It incurs infinite loop in make_mmu_pages_available() which causes mmu->lock 
> softlockup.
> 
> This patch fixes it by setting the return value of prepare_zap_oldest_mmu_page()
> according to whether or not there is mmu page zapped. In addition, we bail out 
> immediately if there is no available mmu page to alloc root page.

Nice!

But I think all callers of make_mmu_pages_available should be handled
the same way.  I'm committing the first hunk for now.  In the meanwhile,
can you look into returning -ENOSPC from make_mmu_pages_available if
!kvm_mmu_available_pages after zapping the pages?

Thanks,

Paolo
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> ---
>  arch/x86/kvm/mmu.c | 16 +++++++++++++---
>  1 file changed, 13 insertions(+), 3 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index 9b1dd11..b9897e8 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -2608,9 +2608,7 @@ static bool prepare_zap_oldest_mmu_page(struct kvm *kvm,
>  
>  	sp = list_last_entry(&kvm->arch.active_mmu_pages,
>  			     struct kvm_mmu_page, link);
> -	kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
> -
> -	return true;
> +	return kvm_mmu_prepare_zap_page(kvm, sp, invalid_list);
>  }
>  
>  /*
> @@ -3379,6 +3377,10 @@ static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
>  	if (vcpu->arch.mmu.shadow_root_level == PT64_ROOT_LEVEL) {
>  		spin_lock(&vcpu->kvm->mmu_lock);
>  		make_mmu_pages_available(vcpu);
> +		if (!kvm_mmu_available_pages(vcpu->kvm)) {
> +			spin_unlock(&vcpu->kvm->mmu_lock);
> +			return 1;
> +		}
>  		sp = kvm_mmu_get_page(vcpu, 0, 0, PT64_ROOT_LEVEL, 1, ACC_ALL);
>  		++sp->root_count;
>  		spin_unlock(&vcpu->kvm->mmu_lock);
> @@ -3390,6 +3392,10 @@ static int mmu_alloc_direct_roots(struct kvm_vcpu *vcpu)
>  			MMU_WARN_ON(VALID_PAGE(root));
>  			spin_lock(&vcpu->kvm->mmu_lock);
>  			make_mmu_pages_available(vcpu);
> +			if (!kvm_mmu_available_pages(vcpu->kvm)) {
> +				spin_unlock(&vcpu->kvm->mmu_lock);
> +				return 1;
> +			}
>  			sp = kvm_mmu_get_page(vcpu, i << (30 - PAGE_SHIFT),
>  					i << 30, PT32_ROOT_LEVEL, 1, ACC_ALL);
>  			root = __pa(sp->spt);
> @@ -3427,6 +3433,10 @@ static int mmu_alloc_shadow_roots(struct kvm_vcpu *vcpu)
>  
>  		spin_lock(&vcpu->kvm->mmu_lock);
>  		make_mmu_pages_available(vcpu);
> +		if (!kvm_mmu_available_pages(vcpu->kvm)) {
> +			spin_unlock(&vcpu->kvm->mmu_lock);
> +			return 1;
> +		}
>  		sp = kvm_mmu_get_page(vcpu, root_gfn, 0, PT64_ROOT_LEVEL,
>  				      0, ACC_ALL);
>  		root = __pa(sp->spt);
>