Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Subject: Re: [PATCH v2] powerpc/uaccess: simplify the get_fs() set_fs() logic
To:     Michael Ellerman <mpe@ellerman.id.au>,
        Christophe Leroy <christophe.leroy@c-s.fr>,
        Benjamin Herrenschmidt <benh@kernel.crashing.org>,
        Paul Mackerras <paulus@samba.org>
Cc:     linux-kernel@vger.kernel.org, linuxppc-dev@lists.ozlabs.org
References: <dd2876b808ea38eb7b7f760ecd6ce06096c61fb5.1580295551.git.christophe.leroy@c-s.fr>
 <87mu3nyh3w.fsf@mpe.ellerman.id.au>
From:   Christophe Leroy <christophe.leroy@csgroup.eu>
Message-ID: <2d0b777d-f1aa-08a1-f287-47ac68efbd99@csgroup.eu>
Date:   Thu, 6 Aug 2020 10:29:56 +0200
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:68.0) Gecko/20100101
 Thunderbird/68.11.0
MIME-Version: 1.0
In-Reply-To: <87mu3nyh3w.fsf@mpe.ellerman.id.au>
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Language: fr
Content-Transfer-Encoding: 8bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


Le 25/07/2020 à 13:22, Michael Ellerman a écrit :
> Hi Christophe,
> 
> Unfortunately this would collide messily with "uaccess: remove
> segment_eq" in linux-next, so I'll ask you to do a respin based on that,
> some comments below.

Done, sent as v3, together with the 2 patchs from Linux next to get it 
build and boot.

> 
> Christophe Leroy <christophe.leroy@c-s.fr> writes:
>> On powerpc, we only have USER_DS and KERNEL_DS
>>
>> Today, this is managed as an 'unsigned long' data space limit
>> which is used to compare the passed address with, plus a bit
>> in the thread_info flags that is set whenever modifying the limit
>> to enable the verification in addr_limit_user_check()
>>
>> The limit is either the last address of user space when USER_DS is
>> set, and the last address of address space when KERNEL_DS is set.
>> In both cases, the limit is a compiletime constant.
>>
>> get_fs() returns the limit, which is part of thread_info struct
>> set_fs() updates the limit then set the TI_FSCHECK flag.
>> addr_limit_user_check() check the flag, and if it is set it checks
>> the limit is the user limit, then unsets the TI_FSCHECK flag.
>>
>> In addition, when the flag is set the syscall exit work is involved.
>> This exit work is heavy compared to normal syscall exit as it goes
>> through normal exception exit instead of the fast syscall exit.
>>
>> Rename this TI_FSCHECK flag to TIF_KERNEL_DS flag which tells whether
>> KERNEL_DS or USER_DS is set. Get mm_segment_t be redifined as a bool
>> struct that is either false (for USER_DS) or true (for KERNEL_DS).
>> When TIF_KERNEL_DS is set, the limit is ~0UL. Otherwise it is
>> TASK_SIZE_USER (resp TASK_SIZE_USER64 on PPC64). When KERNEL_DS is
>> set, there is no range to check. Define TI_FSCHECK as an alias to
>> TIF_KERNEL_DS.
> 
> I'd rather avoid the "DS" name any more than we have to. Maybe it means
> "data space" but that's not a very common term.

I thought it was a reference to the ds/fs/gs ... segment registers in 
the 8086 ?

> 
> The generic helper these days is called uaccess_kernel(), which returns
> true when uaccess routines are allowed to access the kernel.
> 
> So calling it TIF_UACCESS_KERNEL would work I think?

ok

> 
> The bool could be called uaccess_kernel.
> And END_OF_USER_DS could be USER_ADDR_MAX.

ok

> 
>> On exit, involve exit work when the bit is set, i.e. when KERNEL_DS
>> is set. addr_limit_user_check() will clear the bit and kill the
>> user process.
> 
> I guess this is safe. The check was added to make sure we never return
> to userspace with KERNEL_DS set, but using the actual TIF flag to
> determine the address limit should be equally safe, and avoid the
> overhead of the check in the good case.

That's the purpose indeed, yes.


christophe