Re: [dpdk-dev] [PATCH 1/1] eal: add 128-bit cmpset (x86-64 only)

[Ola Liljedahl <Ola.Liljedahl@arm.com>]

On Fri, 2019-02-01 at 17:06 +0000, Eads, Gage wrote:
> 
> > 
> > -----Original Message-----
> > From: Ola Liljedahl [mailto:Ola.Liljedahl@arm.com]
> > Sent: Monday, January 28, 2019 5:02 PM
> > To: Eads, Gage <gage.eads@intel.com>; dev@dpdk.org
> > Cc: arybchenko@solarflare.com; jerinj@marvell.com;
> > chaozhu@linux.vnet.ibm.com; nd <nd@arm.com>; Richardson, Bruce
> > <bruce.richardson@intel.com>; Ananyev, Konstantin
> > <konstantin.ananyev@intel.com>; hemant.agrawal@nxp.com;
> > olivier.matz@6wind.com; Honnappa Nagarahalli
> > <Honnappa.Nagarahalli@arm.com>; Gavin Hu (Arm Technology China)
> > <Gavin.Hu@arm.com>
> > Subject: Re: [dpdk-dev] [PATCH 1/1] eal: add 128-bit cmpset (x86-64 only)
> > 
> > On Mon, 2019-01-28 at 11:29 -0600, Gage Eads wrote:
> > > 
> > > This operation can be used for non-blocking algorithms, such as a
> > > non-blocking stack or ring.
> > > 
> > > Signed-off-by: Gage Eads <gage.eads@intel.com>
> > > ---
> > >  .../common/include/arch/x86/rte_atomic_64.h        | 31 +++++++++++
> > >  lib/librte_eal/common/include/generic/rte_atomic.h | 65
> > > ++++++++++++++++++++++
> > >  2 files changed, 96 insertions(+)
> > > 
> > > diff --git a/lib/librte_eal/common/include/arch/x86/rte_atomic_64.h
> > > b/lib/librte_eal/common/include/arch/x86/rte_atomic_64.h
> > > index fd2ec9c53..b7b90b83e 100644
> > > --- a/lib/librte_eal/common/include/arch/x86/rte_atomic_64.h
> > > +++ b/lib/librte_eal/common/include/arch/x86/rte_atomic_64.h
> > > @@ -34,6 +34,7 @@
> > >  /*
> > >   * Inspired from FreeBSD src/sys/amd64/include/atomic.h
> > >   * Copyright (c) 1998 Doug Rabson
> > > + * Copyright (c) 2019 Intel Corporation
> > >   * All rights reserved.
> > >   */
> > > 
> > > @@ -46,6 +47,7 @@
> > > 
> > >  #include <stdint.h>
> > >  #include <rte_common.h>
> > > +#include <rte_compat.h>
> > >  #include <rte_atomic.h>
> > > 
> > >  /*------------------------- 64 bit atomic operations
> > > ------------------------ -*/ @@ -208,4 +210,33 @@ static inline void
> > > rte_atomic64_clear(rte_atomic64_t *v)
> > >  }
> > >  #endif
> > > 
> > > +static inline int __rte_experimental
> > __rte_always_inline?
> > 
> > > 
> > > +rte_atomic128_cmpset(volatile rte_int128_t *dst,
> > No need to declare the location volatile. Volatile doesn't do what you think
> > it
> > does.
> > https://youtu.be/lkgszkPnV8g?t=1027
> > 
> I made this volatile to match the existing rte_atomicN_cmpset definitions,
> which presumably have a good reason for using the keyword. Maintainers, any
> input here?
> 
> > 
> > 
> > > 
> > > +		     rte_int128_t *exp,
> > I would declare 'exp' const as well and document that 'exp' is not updated
> > (with
> > the old value) for a failure. The reason being that ARMv8.0/AArch64 cannot
> > atomically read the old value without also writing the location and that is
> > bad
> > for performance (unnecessary writes leads to unnecessary contention and
> > worse scalability). And the user must anyway read the location (in the start
> > of
> > the critical section) using e.g. non-atomic 64-bit reads so there isn't
> > actually any
> > requirement for an atomic 128-bit read of the location.
> > 
> Will change in v2.
> 
> > 
> > > 
> > >  rte_int128_t *src,
> > const rte_int128_t *src?
> Sure, I don't see any harm in using const.
> 
> > 
> > 
> > But why are we not passing 'exp' and 'src' by value? That works great, even
> > with
> > structs. Passing by value simplifies the compiler's life, especially if the
> > call is
> > inlined. Ask a compiler developer.
> I ran objdump on the nb_stack code with both approaches, and pass-by-reference 
> resulted in fewer overall x86_64 assembly ops.
> PBV: 100 ops for push, 97 ops for pop
> PBR: 92 ops for push, 84 ops for pop
OK I have never checked x86_64 code generation... I have good experiences with
ARM/AArch64, everything seems to be done using registers. I am surprised there
is a difference.

Did a quick check with lfring, passing 'src' (third param) by reference and by
value. No difference in code generation on x86_64.

But if you insist let's go with PBR.

> 
> (Using the in-progress v5 nb_stack code)
> 
> Another factor -- though much less compelling -- is that with pass-by-
> reference, the user can create a 16B structure and cast it to rte_int128_t
> when they call rte_atomic128_cmpset, whereas with pass-by-value they need to
> put that struct in a union with rte_int128_t.
Which is what I always do nowadays... Trying to use as few casts as possible and
lie to the compiler as seldom as possible. But I can see the freedom provided by
taking a pointer to something and cast it it rte_int128_t ptr in the call
to rte_atomic128_cmpset().

Would prefer a name that is more similar to __atomic_compare_exchange(). E.g.
rte_atomic128_compare_exchange() (or perhaps just rte_atomic128_cmpxchg)? All
the rte_atomicXX_cmpset() functions do not take any memory order parameters.
From an Arm perspective, we are not happy with that.

> 
> > 
> > 
> > > 
> > > +		     unsigned int weak,
> > > +		     enum rte_atomic_memmodel_t success,
> > > +		     enum rte_atomic_memmodel_t failure) {
> > > +	RTE_SET_USED(weak);
> > > +	RTE_SET_USED(success);
> > > +	RTE_SET_USED(failure);
> > > +	uint8_t res;
> > > +
> > > +	asm volatile (
> > > +		      MPLOCKED
> > > +		      "cmpxchg16b %[dst];"
> > > +		      " sete %[res]"
> > > +		      : [dst] "=m" (dst->val[0]),
> > > +			"=A" (exp->val[0]),
> > > +			[res] "=r" (res)
> > > +		      : "c" (src->val[1]),
> > > +			"b" (src->val[0]),
> > > +			"m" (dst->val[0]),
> > > +			"d" (exp->val[1]),
> > > +			"a" (exp->val[0])
> > > +		      : "memory");
> > > +
> > > +	return res;
> > > +}
> > > +
> > >  #endif /* _RTE_ATOMIC_X86_64_H_ */
> > > diff --git a/lib/librte_eal/common/include/generic/rte_atomic.h
> > > b/lib/librte_eal/common/include/generic/rte_atomic.h
> > > index b99ba4688..8d612d566 100644
> > > --- a/lib/librte_eal/common/include/generic/rte_atomic.h
> > > +++ b/lib/librte_eal/common/include/generic/rte_atomic.h
> > > @@ -14,6 +14,7 @@
> > > 
> > >  #include <stdint.h>
> > >  #include <rte_common.h>
> > > +#include <rte_compat.h>
> > > 
> > >  #ifdef __DOXYGEN__
> > > 
> > > @@ -1082,4 +1083,68 @@ static inline void
> > > rte_atomic64_clear(rte_atomic64_t
> > > *v)
> > >  }
> > >  #endif
> > > 
> > > +/*------------------------ 128 bit atomic operations
> > > +------------------------
> > > -*/
> > > +
> > > +/**
> > > + * 128-bit integer structure.
> > > + */
> > > +typedef struct {
> > > +	uint64_t val[2];
> > > +} __rte_aligned(16) rte_int128_t;
> > So we can't use __int128?
> > 
> I'll put it in a union with val[2], in case any implementations want to use
> it.
Thinking on this one more time, since the inline asm functions (e.g. for x86_64
cmpxchg16b and for AArch64 LDXP/STXP) anyway will use 64-bit registers, it makes
most sense to make rte_int128_t a struct of 2x64b. The question is whether to
use an array like above or a struct with two elements (which I normally do
internally). Can you compare code generation with the following definition?
typedef struct {
        uint64_t lo, hi;
} __rte_aligned(16) rte_int128_t;

> 
> Thanks,
> Gage
> 
> [snip]
-- 
Ola Liljedahl, Networking System Architect, Arm
Phone +46706866373, Skype ola.liljedahl