Message-ID: <20c85633-b559-c299-3e57-ae136b201526@oracle.com>
Date:   Mon, 12 Oct 2020 13:14:50 -0600
From:   Khalid Aziz <khalid.aziz@...cle.com>
To:     Catalin Marinas <catalin.marinas@....com>
Cc:     Jann Horn <jannh@...gle.com>,
        "David S. Miller" <davem@...emloft.net>,
        sparclinux@...r.kernel.org,
        Andrew Morton <akpm@...ux-foundation.org>, linux-mm@...ck.org,
        linux-kernel@...r.kernel.org,
        Christoph Hellwig <hch@...radead.org>,
        Anthony Yznaga <anthony.yznaga@...cle.com>,
        Will Deacon <will@...nel.org>,
        linux-arm-kernel@...ts.infradead.org,
        Michael Ellerman <mpe@...erman.id.au>,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Paul Mackerras <paulus@...ba.org>,
        linuxppc-dev@...ts.ozlabs.org
Subject: Re: [PATCH 1/2] mm/mprotect: Call arch_validate_prot under mmap_lock
 and with length

On 10/12/20 11:22 AM, Catalin Marinas wrote:
> On Mon, Oct 12, 2020 at 11:03:33AM -0600, Khalid Aziz wrote:
>> On 10/10/20 5:09 AM, Catalin Marinas wrote:
>>> On Wed, Oct 07, 2020 at 02:14:09PM -0600, Khalid Aziz wrote:
>>>> On 10/7/20 1:39 AM, Jann Horn wrote:
>>>>> arch_validate_prot() is a hook that can validate whether a given set of
>>>>> protection flags is valid in an mprotect() operation. It is given the set
>>>>> of protection flags and the address being modified.
>>>>>
>>>>> However, the address being modified can currently not actually be used in
>>>>> a meaningful way because:
>>>>>
>>>>> 1. Only the address is given, but not the length, and the operation can
>>>>>    span multiple VMAs. Therefore, the callee can't actually tell which
>>>>>    virtual address range, or which VMAs, are being targeted.
>>>>> 2. The mmap_lock is not held, meaning that if the callee were to check
>>>>>    the VMA at @addr, that VMA would be unrelated to the one the
>>>>>    operation is performed on.
>>>>>
>>>>> Currently, custom arch_validate_prot() handlers are defined by
>>>>> arm64, powerpc and sparc.
>>>>> arm64 and powerpc don't care about the address range, they just check the
>>>>> flags against CPU support masks.
>>>>> sparc's arch_validate_prot() attempts to look at the VMA, but doesn't take
>>>>> the mmap_lock.
>>>>>
>>>>> Change the function signature to also take a length, and move the
>>>>> arch_validate_prot() call in mm/mprotect.c down into the locked region.
>>> [...]
>>>> As Chris pointed out, the call to arch_validate_prot() from do_mmap2()
>>>> is made without holding mmap_lock. Lock is not acquired until
>>>> vm_mmap_pgoff(). This variance is uncomfortable but I am more
>>>> uncomfortable forcing all implementations of validate_prot to require
>>>> mmap_lock be held when non-sparc implementations do not have such need
>>>> yet. Since do_mmap2() is in powerpc specific code, for now this patch
>>>> solves a current problem.
>>>
>>> I still think sparc should avoid walking the vmas in
>>> arch_validate_prot(). The core code already has the vmas, though not
>>> when calling arch_validate_prot(). That's one of the reasons I added
>>> arch_validate_flags() with the MTE patches. For sparc, this could be
>>> (untested, just copied the arch_validate_prot() code):
>>
>> I am little uncomfortable with the idea of validating protection bits
>> inside the VMA walk loop in do_mprotect_pkey(). When ADI is being
>> enabled across multiple VMAs and arch_validate_flags() fails on a VMA
>> later, do_mprotect_pkey() will bail out with error leaving ADI enabled
>> on earlier VMAs. This will apply to protection bits other than ADI as
>> well of course. This becomes a partial failure of mprotect() call. I
>> think it should be all or nothing with mprotect() - when one calls
>> mprotect() from userspace, either the entire address range passed in
>> gets its protection bits updated or none of it does. That requires
>> validating protection bits upfront or undoing what earlier iterations of
>> VMA walk loop might have done.
> 
> I thought the same initially but mprotect() already does this with the
> VM_MAY* flag checking. If you ask it for an mprotect() that crosses
> multiple vmas and one of them fails, it doesn't roll back the changes to
> the prior ones. I considered that a similar approach is fine for MTE
> (it's most likely a user error).
> 

You are right about the current behavior with VM_MAY* flags, but that is
not the right behavior. Adding more cases to this just perpetuates
incorrect behavior. It is not easy to roll back changes after VMAs have
potentially been split/merged which is probably why the current code
simply throws in the towel and returns with partially modified address
space. It is lot easier to do all the checks upfront and then proceed or
not proceed with modifying VMAs. One approach might be to call
arch_validate_flags() in a loop before modifying VMAs and walk all VMAs
with a read lock held. Current code also bails out with ENOMEM if it
finds a hole in the address range and leaves any modifications already
made in place. This is another case where a hole could have been
detected earlier.

--
Khalid

Powered by blists - more mailing lists