Re: [PATCH 3/3] context_tracking,x86: remove extraneous irq disable & enable from context tracking on syscall entry

[Rik van Riel <riel@redhat.com>]

On 05/04/2015 05:26 AM, Paolo Bonzini wrote:
> 
> 
> On 02/05/2015 07:27, Ingo Molnar wrote:
>>
>>   - synchronize_rcu() avoids having to send an IPI by taking a 
>>     peak at rq->curr's pt_regs::flag, and if:
>>
>>      - the flag is 0 then it has observed a quiescent state.
>>
>>      - the flag is 1, then it would set TIF_NOHZ and wait for a 
>>        completion from a TIF_NOHZ callback.
> 
> Isn't this racy?
> 
> 	synchronize_rcu CPU		nohz CPU
> 	---------------------------------------------------------
> 					set flag = 0
> 	read flag = 0
> 					return to userspace
> 	set TIF_NOHZ
> 
> and there's no guarantee that TIF_NOHZ is ever processed by the nohz CPU.

Actually, the "race" in this direction is fine. If flag==0, then
the nohz CPU is not accessing any RCU protected data structures,
and the synhcronize_rcu CPU will not be setting TIF_NOHZ.

The race is only a concern if the synchronize_rcu CPU reads
flag==1 (nohz CPU is in kernel space), and sets TIF_NOHZ after
the nohz CPU has cleared flag (and is unable to handle RCU
stuff). An atomic compare and swap prevents that issue.

The other race, of the synchronize_rcu CPU reading 0, followed
by the nohz CPU going into kernel space, and setting the flag
to 1, should be fine. After all, this means the nohz_full CPU
just went into a new RCU grace period, which is just what the
synchronize_rcu CPU was waiting for.

-- 
All rights reversed