Re: [PATCH rcu 13/17] srcu: Add SRCU-fast readers

[Paul E. McKenney]

On Thu, Jan 16, 2025 at 01:00:24PM -0800, Alexei Starovoitov wrote:
> On Thu, Jan 16, 2025 at 12:21 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > +/*
> > + * Counts the new reader in the appropriate per-CPU element of the
> > + * srcu_struct.  Returns a pointer that must be passed to the matching
> > + * srcu_read_unlock_fast().
> > + *
> > + * Note that this_cpu_inc() is an RCU read-side critical section either
> > + * because it disables interrupts, because it is a single instruction,
> > + * or because it is a read-modify-write atomic operation, depending on
> > + * the whims of the architecture.
> > + */
> > +static inline struct srcu_ctr __percpu *__srcu_read_lock_fast(struct srcu_struct *ssp)
> > +{
> > +       struct srcu_ctr __percpu *scp = READ_ONCE(ssp->srcu_ctrp);
> > +
> > +       RCU_LOCKDEP_WARN(!rcu_is_watching(), "RCU must be watching srcu_read_lock_fast().");
> > +       this_cpu_inc(scp->srcu_locks.counter); /* Y */
> > +       barrier(); /* Avoid leaking the critical section. */
> > +       return scp;
> > +}
> 
> This doesn't look fast.
> If I'm reading this correctly,
> even with debugs off RCU_LOCKDEP_WARN() will still call
> rcu_is_watching() and this doesn't look cheap or fast.

Here is the CONFIG_PROVE_RCU=n definition:

	#define RCU_LOCKDEP_WARN(c, s) do { } while (0 && (c))

The "0" in the "0 && (c)" should prevent that call to rcu_is_watching().

But why not see what the compiler thinks?  I added the following function
to kernel/rcu/srcutree.c:

struct srcu_ctr __percpu *test_srcu_read_lock_fast(struct srcu_struct *ssp)
{
	struct srcu_ctr __percpu *p;

	p = srcu_read_lock_fast(ssp);
	return p;
}

This function compiles to the following code:

Dump of assembler code for function test_srcu_read_lock_fast:
   0xffffffff811220c0 <+0>:     endbr64
   0xffffffff811220c4 <+4>:     sub    $0x8,%rsp
   0xffffffff811220c8 <+8>:     mov    0x8(%rdi),%rax
   0xffffffff811220cc <+12>:    add    %gs:0x7eef3944(%rip),%rax        # 0x15a18 <this_cpu_off>
   0xffffffff811220d4 <+20>:    mov    0x20(%rax),%eax
   0xffffffff811220d7 <+23>:    test   $0x8,%al
   0xffffffff811220d9 <+25>:    je     0xffffffff811220eb <test_srcu_read_lock_fast+43>
   0xffffffff811220db <+27>:    mov    (%rdi),%rax
   0xffffffff811220de <+30>:    incq   %gs:(%rax)
   0xffffffff811220e2 <+34>:    add    $0x8,%rsp
   0xffffffff811220e6 <+38>:    jmp    0xffffffff81f5fe60 <__x86_return_thunk>
   0xffffffff811220eb <+43>:    mov    $0x8,%esi
   0xffffffff811220f0 <+48>:    mov    %rdi,(%rsp)
   0xffffffff811220f4 <+52>:    call   0xffffffff8111fb90 <__srcu_check_read_flavor>
   0xffffffff811220f9 <+57>:    mov    (%rsp),%rdi
   0xffffffff811220fd <+61>:    jmp    0xffffffff811220db <test_srcu_read_lock_fast+27>

The first call to srcu_read_lock_fast() invokes __srcu_check_read_flavor(),
but after that the "je" instruction will fall through.  So the common-case
code path executes only the part of this function up to and including the
"jmp 0xffffffff81f5fe60 <__x86_return_thunk>".

Does that serve?

							Thanx, Paul