Subject: Re: [PATCH 3/3] kvm: arm/arm64: Fix locking for kvm_free_stage2_pgd
To: Christoffer Dall <cdall@linaro.org>
References: <1489503154-20705-1-git-send-email-suzuki.poulose@arm.com>
 <1489503154-20705-4-git-send-email-suzuki.poulose@arm.com>
 <20170315092147.GM1277@cbox> <314fbde3-17e6-414b-85e6-326de22bdc1c@arm.com>
 <20170315105639.GA31974@cbox>
Cc: Suzuki K Poulose <suzuki.poulose@arm.com>,
        linux-arm-kernel@lists.infradead.org, andreyknvl@google.com,
        dvyukov@google.com, christoffer.dall@linaro.org,
        kvmarm@lists.cs.columbia.edu, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org, kcc@google.com,
        syzkaller@googlegroups.com, will.deacon@arm.com,
        catalin.marinas@arm.com, pbonzini@redhat.com, mark.rutland@arm.com,
        ard.biesheuvel@linaro.org, stable@vger.kernel.org
From: Marc Zyngier <marc.zyngier@arm.com>
Organization: ARM Ltd
Message-ID: <0e5ff7f7-855c-ea28-fdee-73c062c3d289@arm.com>
Date: Wed, 15 Mar 2017 13:28:07 +0000
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101
 Icedove/45.6.0
MIME-Version: 1.0
In-Reply-To: <20170315105639.GA31974@cbox>
Content-Type: text/plain; charset=windows-1252
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2965
Lines: 77

On 15/03/17 10:56, Christoffer Dall wrote:
> On Wed, Mar 15, 2017 at 09:39:26AM +0000, Marc Zyngier wrote:
>> On 15/03/17 09:21, Christoffer Dall wrote:
>>> On Tue, Mar 14, 2017 at 02:52:34PM +0000, Suzuki K Poulose wrote:
>>>> In kvm_free_stage2_pgd() we don't hold the kvm->mmu_lock while calling
>>>> unmap_stage2_range() on the entire memory range for the guest. This could
>>>> cause problems with other callers (e.g, munmap on a memslot) trying to
>>>> unmap a range.
>>>>
>>>> Fixes: commit d5d8184d35c9 ("KVM: ARM: Memory virtualization setup")
>>>> Cc: stable@vger.kernel.org # v3.10+
>>>> Cc: Marc Zyngier <marc.zyngier@arm.com>
>>>> Cc: Christoffer Dall <christoffer.dall@linaro.org>
>>>> Signed-off-by: Suzuki K Poulose <suzuki.poulose@arm.com>
>>>> ---
>>>>  arch/arm/kvm/mmu.c | 3 +++
>>>>  1 file changed, 3 insertions(+)
>>>>
>>>> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
>>>> index 13b9c1f..b361f71 100644
>>>> --- a/arch/arm/kvm/mmu.c
>>>> +++ b/arch/arm/kvm/mmu.c
>>>> @@ -831,7 +831,10 @@ void kvm_free_stage2_pgd(struct kvm *kvm)
>>>>  	if (kvm->arch.pgd == NULL)
>>>>  		return;
>>>>  
>>>> +	spin_lock(&kvm->mmu_lock);
>>>>  	unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
>>>> +	spin_unlock(&kvm->mmu_lock);
>>>> +
>>>
>>> This ends up holding the spin lock for potentially quite a while, where
>>> we can do things like __flush_dcache_area(), which I think can fault.
>>
>> I believe we're always using the linear mapping (or kmap on 32bit) in
>> order not to fault.
>>
> 
> ok, then there's just the concern that we may be holding a spinlock for
> a very long time.  I seem to recall Mario once added something where he
> unlocked and gave a chance to schedule something else for each PUD or
> something like that, because he ran into the issue during migration.  Am
> I confusing this with something else?

That definitely rings a bell: stage2_wp_range() uses that kind of trick
to give the system a chance to breathe. Maybe we could use a similar
trick in our S2 unmapping code? How about this (completely untested) patch:

diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
index 962616fd4ddd..1786c24212d4 100644
--- a/arch/arm/kvm/mmu.c
+++ b/arch/arm/kvm/mmu.c
@@ -292,8 +292,13 @@ static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
 	phys_addr_t addr = start, end = start + size;
 	phys_addr_t next;
 
+	BUG_ON(!spin_is_locked(&kvm->mmu_lock));
+
 	pgd = kvm->arch.pgd + stage2_pgd_index(addr);
 	do {
+		if (need_resched() || spin_needbreak(&kvm->mmu_lock))
+			cond_resched_lock(&kvm->mmu_lock);
+
 		next = stage2_pgd_addr_end(addr, end);
 		if (!stage2_pgd_none(*pgd))
 			unmap_stage2_puds(kvm, pgd, addr, next);

The additional BUG_ON() is just for my own peace of mind - we seem to
have missed a couple of these lately, and the "breathing" code makes
it imperative that this lock is being taken prior to entering the
function.

Thoughts?

	M.
-- 
Jazz is not dead. It just smells funny...