linux-kernel - Re: [PATCH] KVM: arm64: Select SCHED

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date:   Thu, 24 Oct 2019 12:22:08 +0100
From:   Steven Price <steven.price@....com>
To:     Marc Zyngier <maz@...nel.org>, Mao Wenan <maowenan@...wei.com>
Cc:     catalin.marinas@....com, kernel-janitors@...r.kernel.org,
        linux-kernel@...r.kernel.org, will@...nel.org,
        kvmarm@...ts.cs.columbia.edu, linux-arm-kernel@...ts.infradead.org
Subject: Re: [PATCH] KVM: arm64: Select SCHED_INFO before SCHEDSTATS

On 23/10/2019 17:51, Marc Zyngier wrote:
> On 2019-10-23 04:22, Mao Wenan wrote:
>> If KVM=y, it will select SCHEDSTATS, below erros can
>> be seen:
>> kernel/sched/stats.h: In function rq_sched_info_arrive:
>> kernel/sched/stats.h:12:20: error: struct sched_info
>> has no member named run_delay
>>    rq->rq_sched_info.run_delay += delta;
>>                     ^
>> kernel/sched/stats.h:13:20: error: struct sched_info
>> has no member named pcount
>>    rq->rq_sched_info.pcount++;
>>                     ^
>> kernel/sched/stats.h: In function rq_sched_info_dequeued:
>> kernel/sched/stats.h:31:20: error: struct sched_info has
>> no member named run_delay
>>    rq->rq_sched_info.run_delay += delta;
>>
>> These are because CONFIG_SCHED_INFO is not set, This patch
>> is to select SCHED_INFO before SCHEDSTATS.
>>
>> Fixes: 8564d6372a7d ("KVM: arm64: Support stolen time reporting via
>> shared structure")
>> Signed-off-by: Mao Wenan <maowenan@...wei.com>
>> ---
>>  arch/arm64/kvm/Kconfig | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/arch/arm64/kvm/Kconfig b/arch/arm64/kvm/Kconfig
>> index d8b88e4..3c46eac 100644
>> --- a/arch/arm64/kvm/Kconfig
>> +++ b/arch/arm64/kvm/Kconfig
>> @@ -39,6 +39,7 @@ config KVM
>>      select IRQ_BYPASS_MANAGER
>>      select HAVE_KVM_IRQ_BYPASS
>>      select HAVE_KVM_VCPU_RUN_PID_CHANGE
>> +    select SCHED_INFO
>>      select SCHEDSTATS
>>      ---help---
>>        Support hosting virtualized guest machines.
> 
> SCHEDSTATS is really an odd choice. Here's what I get after disabling
> DEBUG_KERNEL (from defconfig):
> 
> WARNING: unmet direct dependencies detected for SCHEDSTATS
>   Depends on [n]: DEBUG_KERNEL [=n] && PROC_FS [=y]
>   Selected by [y]:
>   - KVM [=y] && VIRTUALIZATION [=y] && OF [=y]
> 
> WARNING: unmet direct dependencies detected for SCHEDSTATS
>   Depends on [n]: DEBUG_KERNEL [=n] && PROC_FS [=y]
>   Selected by [y]:
>   - KVM [=y] && VIRTUALIZATION [=y] && OF [=y]
> 
> WARNING: unmet direct dependencies detected for SCHEDSTATS
>   Depends on [n]: DEBUG_KERNEL [=n] && PROC_FS [=y]
>   Selected by [y]:
>   - KVM [=y] && VIRTUALIZATION [=y] && OF [=y]
> 
> So clearly SCHEDSTATS isn't meant to be selected on its own.
> 
> We can either just select SCHED_INFO (which *nobody else does*), or go
> the full x86 way which selects TASK_DELAY_ACCT (and thus depends on
> NET && MULTIUSER). My gut feeling is that we shouldn't deviate too much
> from x86...
> 
> Thoughts?

I suspect you're right - TASK_DELAY_ACCT seems to be the closest to what
we need. SCHEDSTATS has the "advantage" of forcing sched_info_on() to
return true - preventing it from being disabled. But we clearly don't
want to require CONFIG_DEBUG_KERNEL for CONFIG_KVM.

The next best is CONFIG_TASK_DELAY_ACCT which enables sched_info_on()
unless "nodelayacct" is specified on the cmdline. It seems reasonable
that the cmdline option might break stolen time.

So let's just copy x86:

-----8<-----
>From 915893f5c57241cc29d90769b3f720a6135277d7 Mon Sep 17 00:00:00 2001
From: Steven Price <steven.price@....com>
Date: Thu, 24 Oct 2019 12:14:36 +0100
Subject: [PATCH] KVM: arm64: Select TASK_DELAY_ACCT rather than SCHEDSTATS

SCHEDSTATS requires DEBUG_KERNEL (and PROC_FS) and therefore isn't a
good choice for enabling the scheduling statistics required for stolen
time.

Instead match the x86 configuration and select TASK_DELAY_ACCT. This
adds the dependencies of NET && MULTIUSER for arm64 KVM.

Suggested-by: Marc Zyngier <maz@...nel.org>
Fixes: 8564d6372a7d ("KVM: arm64: Support stolen time reporting via shared structure")
Signed-off-by: Steven Price <steven.price@....com>
---
 arch/arm64/kvm/Kconfig | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/kvm/Kconfig b/arch/arm64/kvm/Kconfig
index d8b88e40d223..1ffb300e2d92 100644
--- a/arch/arm64/kvm/Kconfig
+++ b/arch/arm64/kvm/Kconfig
@@ -21,6 +21,8 @@ if VIRTUALIZATION
 config KVM
 	bool "Kernel-based Virtual Machine (KVM) support"
 	depends on OF
+	# for TASKSTATS/TASK_DELAY_ACCT:
+	depends on NET && MULTIUSER
 	select MMU_NOTIFIER
 	select PREEMPT_NOTIFIERS
 	select HAVE_KVM_CPU_RELAX_INTERCEPT
@@ -39,7 +41,7 @@ config KVM
 	select IRQ_BYPASS_MANAGER
 	select HAVE_KVM_IRQ_BYPASS
 	select HAVE_KVM_VCPU_RUN_PID_CHANGE
-	select SCHEDSTATS
+	select TASK_DELAY_ACCT
 	---help---
 	  Support hosting virtualized guest machines.
 	  We don't support KVM with 16K page tables yet, due to the multiple
-- 
2.20.1