Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <9241d8a5-d341-aa84-9aee-6e84377dbe9d@cfa.harvard.edu>
Date:   Thu, 21 Jul 2022 15:19:24 -0400
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Thunderbird/91.11.0
Subject: Re: Thread safe client destruction
Content-Language: en-US
To:     Chuck Lever III <chuck.lever@oracle.com>
Cc:     libtirpc List <libtirpc-devel@lists.sourceforge.net>,
        Linux NFS Mailing List <linux-nfs@vger.kernel.org>
References: <53af8ec6-4ece-09b2-9499-d46d0fdfaa15@cfa.harvard.edu>
 <4D490A86-E873-40E7-A30E-9189B127C65C@oracle.com>
From:   Attila Kovacs <attila.kovacs@cfa.harvard.edu>
In-Reply-To: <4D490A86-E873-40E7-A30E-9189B127C65C@oracle.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hi Chuck,


I'm not adding any new functionality here. I'm merely trying to fix 
what's already there but broken (in my opinion). :-)

-- A.


On 7/21/22 14:45, Chuck Lever III wrote:
> 
> 
>> On Jul 21, 2022, at 2:41 PM, Attila Kovacs <attila.kovacs@cfa.harvard.edu> wrote:
>>
>> Hi again,
>>
>>
>> I found yet more potential MT flaws in clnt_dg.c and clnt_vg.c.
> 
> IIRC libtirpc is not MT-enabled because it was forked from the
> Solaris libtirpc before MT-enablement was completed. I could
> be remembering incorrectly.
> 
> Therefore I think we would need some unit tests along with the
> improvements you propose. There's a test suite somewhere, but
> someone else will need to provide details.
> 
> (and also, please inline your patch submissions rather than
> attaching them, to make it easier for us to review. thanks!)
> 
> 
>> 1. In clnt_dg.c in clnt_dg_freeres(), cu_fd_lock->active is set to TRUE, with no corresponding clearing when the operation (*xdr_res() call) is completed. This would leave other waiting operations blocked indefinitely, effectively deadlocked. For comparison, clnt_vd_freeres() in clnt_vc.c does not set the active state to TRUE. I believe the vc behavior is correct, while the dg behavior is a bug.
>>
>> 2. If clnt_dg_destroy() or clnt_vc_destroy() is awoken with other blocked operations pending (such as clnt_*_call(), clnt_*_control(), or clnt_*_freeres()) but no active operation currently being executed, then the client gets destroyed. Then, as the other blocked operations get subsequently awoken, they will try operate on an invalid client handle, potentially causing unpredictable behavior and stack corruption.
>>
>> The proposed fix is to introduce a simple mutexed counting variable into the client lock structure, which keeps track of the number of pending blocking operations on the client. This way, clnt_*_destroy() can check if there are any operations pending on a client, and keep waiting until all pending operations are completed, before the client is destroyed safely and its resources are freed.
>>
>> Attached is a patch with the above fixes.
>>
>> -- A.<clnt_mt_safe_destroy.patch>
> 
> --
> Chuck Lever
> 
> 
>