Re: C++20 report error at file rte_spinlock.h

[Stephen Hemminger <stephen@networkplumber.org>]

On Tue, 20 Dec 2022 02:11:42 +0000
"Zhou, Xiangyun" <xiangyun.zhou@intel.com> wrote:

> Thanks very much  for Konstantin and Tyler's analyzing.
> 
> I agree that removal of 'volatile' is enough. A spinlock has already used to protect these variables.
> 
> -----Original Message-----
> From: Konstantin Ananyev <konstantin.ananyev@huawei.com> 
> Sent: Tuesday, December 20, 2022 12:51 AM
> To: Tyler Retzlaff <roretzla@linux.microsoft.com>; Zhou, Xiangyun <xiangyun.zhou@intel.com>
> Cc: dev@dpdk.org; Xu, Bowen <bowen.xu@intel.com>
> Subject: RE: C++20 report error at file rte_spinlock.h
> 
> 
> > On Tue, Dec 13, 2022 at 06:11:06AM +0000, Zhou, Xiangyun wrote:  
> > > Dear dpdk dev,
> > >
> > > I'm using dpdk 21.11 LTS, when compile my program with CPP flag 
> > > "-std=c++20", the compiler report below errors. After checking file  
> > rte_spinlock.h, I think the error report by compiler is valid, there 
> > should be a potential issue when using functions 
> > rte_spinlock_recursive_lock, rte_spinlock_recursive_unlock and 
> > rte_spinlock_recursive_trylock in multi-thread, we could either remove "volatile" definition to ask users to handle the multi-thread issue, or using atomic operatings instead of self-increment and self-decrement.  
> > >
> > >
> > > /home/dpdk/lib/eal/include/generic/rte_spinlock.h:221:12: error: 
> > > increment of object of volatile-qualified type 'volatile int' is  
> > deprecated [-Werror,-Wdeprecated-volatile]  
> > >         slr->count++;
> > >                   ^
> > > /home/dpdk/lib/eal/include/generic/rte_spinlock.h:231:6: error: 
> > > decrement of object of volatile-qualified type 'volatile int' is  
> > deprecated [-Werror,-Wdeprecated-volatile]  
> > >         if (--(slr->count) == 0) {
> > >             ^
> > > /home/dpdk/lib/eal/include/generic/rte_spinlock.h:255:12: error: 
> > > increment of object of volatile-qualified type 'volatile int' is  
> > deprecated [-Werror,-Wdeprecated-volatile]  
> > >         slr->count++;
> > >  
> > 
> > i have work in progress to optionally use standard atomics but in the 
> > meantime the correct thing to do here is to use the gcc builtins that 
> > match the requirements of the c++11 memory model.
> > 
> > the code should be converted to use __atomic_fetch_{add,sub} or 
> > __atomic_{add,sub}_fetch as appropriate.
> > 
> > ty.  
> 
> From looking at the code, I don't think it is necessary:
> both 'user' and 'count' supposed to be protected by 'sl'.
> In fact, it looks safe just to remove 'volatile' qualifier here.
>  
> 

I noticed a couple of other documentation errors here.

The intro comment says this a read-write lock (and it isn't).
Probably copy/paste error.

The user field says is is "core id" but it really is thread id.