Re: [PATCH 11/11] x86,rcu: use percpu rcu_preempt_depth

[Lai Jiangshan]

On 2019/11/1 10:30 äå, Paul E. McKenney wrote:
> On Fri, Nov 01, 2019 at 02:13:15PM +0100, Peter Zijlstra wrote:
> > On Fri, Nov 01, 2019 at 05:58:16AM -0700, Paul E. McKenney wrote:
> > > On Thu, Oct 31, 2019 at 10:08:06AM +0000, Lai Jiangshan wrote:
> > > > +/* We mask the RCU_NEED_SPECIAL bit so that it return real depth */
> > > > +static __always_inline int rcu_preempt_depth(void)
> > > > +{
> > > > +	return raw_cpu_read_4(__rcu_preempt_depth) & ~RCU_NEED_SPECIAL;
> > >
> > > Why not raw_cpu_generic_read()?
> > >
> > > OK, OK, I get that raw_cpu_read_4() translates directly into an "mov"
> > > instruction on x86, but given that x86 percpu_from_op() is able to
> > > adjust based on operand size, why doesn't something like raw_cpu_read()
> > > also have an x86-specific definition that adjusts based on operand size?
> >
> > The reason for preempt.h was header recursion hell.
>
> Fair enough, being as that is also the reason for _rcu_read_lock()
> not being inlined.  :-/
>
> > > > +}
> > > > +
> > > > +static __always_inline void rcu_preempt_depth_set(int pc)
> > > > +{
> > > > +	int old, new;
> > > > +
> > > > +	do {
> > > > +		old = raw_cpu_read_4(__rcu_preempt_depth);
> > > > +		new = (old & RCU_NEED_SPECIAL) |
> > > > +			(pc & ~RCU_NEED_SPECIAL);
> > > > +	} while (raw_cpu_cmpxchg_4(__rcu_preempt_depth, old, new) != old);
> > >
> > > Ummm...
> > >
> > > OK, as you know, I have long wanted _rcu_read_lock() to be inlineable.
> > > But are you -sure- that an x86 cmpxchg is faster than a function call
> > > and return?  I have strong doubts on that score.
> >
> > This is a regular CMPXCHG instruction, not a LOCK prefixed one, and that
> > should make all the difference
>
> Yes, understood, but this is also adding some arithmetic, a comparison,
> and a conditional branch.  Are you -sure- that this is cheaper than
> an unconditional call and return?

rcu_preempt_depth_set() is used only for exit_rcu().
The performance doesn't matter here. And since RCU_NEED_SPECIAL
bit is allowed to lost in exit_rcu(), rcu_preempt_depth_set()
can be a single raw_cpu_write_4() if the performance is matter.

(This complex code is copied from preempt.h and I can't expect
how will rcu_preempt_depth_set() be used in the feture
so I keep it unchanged.)


+static __always_inline void rcu_preempt_depth_inc(void)
+{
+	raw_cpu_add_4(__rcu_preempt_depth, 1);
+}

This one is for read_read_lock(). ONE instruction.

+
+static __always_inline bool rcu_preempt_depth_dec_and_test(void)
+{
+	return GEN_UNARY_RMWcc("decl", __rcu_preempt_depth, e, 
__percpu_arg([var]));
+}

This one is for read_read_unlock() which will be 2 instructions
("decl" and "je"), which is the same as preempt_enable().

In news days, preempt_disable() is discouraged unless it is
really necessary and rcu is always encouraged. Low overhead
read_read_[un]lock() is essential.


> > > Plus multiplying the x86-specific code by 26 doesn't look good.
> > >
> > > And the RCU read-side nesting depth really is a per-task thing.  Copying
> > > it to and from the task at context-switch time might make sense if we
> > > had a serious optimization, but it does not appear that we do.

Once upon a time, __preempt_count is also being copied to and from the
task at context-switch, and worked well.

> > > You original patch some years back, ill-received though it was at the
> > > time, is looking rather good by comparison.  Plus it did not require
> > > architecture-specific code!
> >
> > Right, so the per-cpu preempt_count code relies on the preempt_count
> > being invariant over context switches. That means we never have to
> > save/restore the thing.
> >
> > For (preemptible) rcu, this is 'obviously' not the case.
> >
> > That said, I've not looked over this patch series, I only got 1 actual
> > patch, not the whole series, and I've not had time to go dig out the
> > rest..
>
> I have taken a couple of the earlier patches in the series.
>
> Perhaps inlining these things is instead a job for the long anticipated
> GCC LTO?  ;-)

Adding a kenerl/offset.c and some Mafefile stuff will help inlining
these things. But I don't think Linus will happy with introducing
kenerl/offset.c. There will be 3 instructions for rcu_read_lock()
and 5 for rcu_read_unlock(), which doesn't taste so delicious.

Moving rcu_read_lock_nesting to struct thread_info is another
possible way. The number of instructions is also 3 and 5.

Thanks
Lai