This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH 1/3] Tunables: Add tunables of spin count for adaptive spin mutex
- From: Adhemerval Zanella <adhemerval dot zanella at linaro dot org>
- To: kemi <kemi dot wang at intel dot com>
- Cc: Dave Hansen <dave dot hansen at linux dot intel dot com>, Tim Chen <tim dot c dot chen at intel dot com>, Andi Kleen <andi dot kleen at intel dot com>, Ying Huang <ying dot huang at intel dot com>, Aaron Lu <aaron dot lu at intel dot com>, Glibc alpha <libc-alpha at sourceware dot org>, aubrey <aubrey dot li at intel dot com>
- Date: Wed, 4 Apr 2018 14:16:59 -0300
- Subject: Re: [PATCH 1/3] Tunables: Add tunables of spin count for adaptive spin mutex
- Autocrypt: addr=adhemerval dot zanella at linaro dot org; prefer-encrypt=mutual; keydata= xsFNBFcVGkoBEADiQU2x/cBBmAVf5C2d1xgz6zCnlCefbqaflUBw4hB/bEME40QsrVzWZ5Nq 8kxkEczZzAOKkkvv4pRVLlLn/zDtFXhlcvQRJ3yFMGqzBjofucOrmdYkOGo0uCaoJKPT186L NWp53SACXguFJpnw4ODI64ziInzXQs/rUJqrFoVIlrPDmNv/LUv1OVPKz20ETjgfpg8MNwG6 iMizMefCl+RbtXbIEZ3TE/IaDT/jcOirjv96lBKrc/pAL0h/O71Kwbbp43fimW80GhjiaN2y WGByepnkAVP7FyNarhdDpJhoDmUk9yfwNuIuESaCQtfd3vgKKuo6grcKZ8bHy7IXX1XJj2X/ BgRVhVgMHAnDPFIkXtP+SiarkUaLjGzCz7XkUn4XAGDskBNfbizFqYUQCaL2FdbW3DeZqNIa nSzKAZK7Dm9+0VVSRZXP89w71Y7JUV56xL/PlOE+YKKFdEw+gQjQi0e+DZILAtFjJLoCrkEX w4LluMhYX/X8XP6/C3xW0yOZhvHYyn72sV4yJ1uyc/qz3OY32CRy+bwPzAMAkhdwcORA3JPb kPTlimhQqVgvca8m+MQ/JFZ6D+K7QPyvEv7bQ7M+IzFmTkOCwCJ3xqOD6GjX3aphk8Sr0dq3 4Awlf5xFDAG8dn8Uuutb7naGBd/fEv6t8dfkNyzj6yvc4jpVxwARAQABzUlBZGhlbWVydmFs IFphbmVsbGEgTmV0dG8gKExpbmFybyBWUE4gS2V5KSA8YWRoZW1lcnZhbC56YW5lbGxhQGxp bmFyby5vcmc+wsF3BBMBCAAhBQJXFRpKAhsDBQsJCAcDBRUKCQgLBRYCAwEAAh4BAheAAAoJ EKqx7BSnlIjv0e8P/1YOYoNkvJ+AJcNUaM5a2SA9oAKjSJ/M/EN4Id5Ow41ZJS4lUA0apSXW NjQg3VeVc2RiHab2LIB4MxdJhaWTuzfLkYnBeoy4u6njYcaoSwf3g9dSsvsl3mhtuzm6aXFH /Qsauav77enJh99tI4T+58rp0EuLhDsQbnBic/ukYNv7sQV8dy9KxA54yLnYUFqH6pfH8Lly sTVAMyi5Fg5O5/hVV+Z0Kpr+ZocC1YFJkTsNLAW5EIYSP9ftniqaVsim7MNmodv/zqK0IyDB GLLH1kjhvb5+6ySGlWbMTomt/or/uvMgulz0bRS+LUyOmlfXDdT+t38VPKBBVwFMarNuREU2 69M3a3jdTfScboDd2ck1u7l+QbaGoHZQ8ZNUrzgObltjohiIsazqkgYDQzXIMrD9H19E+8fw kCNUlXxjEgH/Kg8DlpoYJXSJCX0fjMWfXywL6ZXc2xyG/hbl5hvsLNmqDpLpc1CfKcA0BkK+ k8R57fr91mTCppSwwKJYO9T+8J+o4ho/CJnK/jBy1pWKMYJPvvrpdBCWq3MfzVpXYdahRKHI ypk8m4QlRlbOXWJ3TDd/SKNfSSrWgwRSg7XCjSlR7PNzNFXTULLB34sZhjrN6Q8NQZsZnMNs TX8nlGOVrKolnQPjKCLwCyu8PhllU8OwbSMKskcD1PSkG6h3r0AqzsFNBFcVGkoBEACgAdbR Ck+fsfOVwT8zowMiL3l9a2DP3Eeak23ifdZG+8Avb/SImpv0UMSbRfnw/N81IWwlbjkjbGTu oT37iZHLRwYUFmA8fZX0wNDNKQUUTjN6XalJmvhdz9l71H3WnE0wneEM5ahu5V1L1utUWTyh VUwzX1lwJeV3vyrNgI1kYOaeuNVvq7npNR6t6XxEpqPsNc6O77I12XELic2+36YibyqlTJIQ V1SZEbIy26AbC2zH9WqaKyGyQnr/IPbTJ2Lv0dM3RaXoVf+CeK7gB2B+w1hZummD21c1Laua +VIMPCUQ+EM8W9EtX+0iJXxI+wsztLT6vltQcm+5Q7tY+HFUucizJkAOAz98YFucwKefbkTp eKvCfCwiM1bGatZEFFKIlvJ2QNMQNiUrqJBlW9nZp/k7pbG3oStOjvawD9ZbP9e0fnlWJIsj 6c7pX354Yi7kxIk/6gREidHLLqEb/otuwt1aoMPg97iUgDV5mlNef77lWE8vxmlY0FBWIXuZ yv0XYxf1WF6dRizwFFbxvUZzIJp3spAao7jLsQj1DbD2s5+S1BW09A0mI/1DjB6EhNN+4bDB SJCOv/ReK3tFJXuj/HbyDrOdoMt8aIFbe7YFLEExHpSk+HgN05Lg5TyTro8oW7TSMTk+8a5M kzaH4UGXTTBDP/g5cfL3RFPl79ubXwARAQABwsFfBBgBCAAJBQJXFRpKAhsMAAoJEKqx7BSn lIjvI/8P/jg0jl4Tbvg3B5kT6PxJOXHYu9OoyaHLcay6Cd+ZrOd1VQQCbOcgLFbf4Yr+rE9l mYsY67AUgq2QKmVVbn9pjvGsEaz8UmfDnz5epUhDxC6yRRvY4hreMXZhPZ1pbMa6A0a/WOSt AgFj5V6Z4dXGTM/lNManr0HjXxbUYv2WfbNt3/07Db9T+GZkpUotC6iknsTA4rJi6u2ls0W9 1UIvW4o01vb4nZRCj4rni0g6eWoQCGoVDk/xFfy7ZliR5B+3Z3EWRJcQskip/QAHjbLa3pml xAZ484fVxgeESOoaeC9TiBIp0NfH8akWOI0HpBCiBD5xaCTvR7ujUWMvhsX2n881r/hNlR9g fcE6q00qHSPAEgGr1bnFv74/1vbKtjeXLCcRKk3Ulw0bY1OoDxWQr86T2fZGJ/HIZuVVBf3+ gaYJF92GXFynHnea14nFFuFgOni0Mi1zDxYH/8yGGBXvo14KWd8JOW0NJPaCDFJkdS5hu0VY 7vJwKcyHJGxsCLU+Et0mryX8qZwqibJIzu7kUJQdQDljbRPDFd/xmGUFCQiQAncSilYOcxNU EMVCXPAQTteqkvA+gNqSaK1NM9tY0eQ4iJpo+aoX8HAcn4sZzt2pfUB9vQMTBJ2d4+m/qO6+ cFTAceXmIoFsN8+gFN3i8Is3u12u8xGudcBPvpoy4OoG
- Openpgp: preference=signencrypt
- References: <1522394093-9835-1-git-send-email-kemi.wang@intel.com> <03d60e93-a167-1401-b67e-22d7d2443486@intel.com>
On 04/04/2018 07:25, kemi wrote:
> Hi, Adhemerval
> Thanks for your review. Could you please help to review the other two patches in this series
> if available? Also, please keep guys in the cc list in case someone may not subscribe glibc
> mail list. Thanks:)
Yeah, I am checking the patch along with the referenced benchmark on a aarch64
machine as well.
>
> See my reply below.
>
> On 2018年03月30日 15:14, Kemi Wang wrote:
>> On 30/03/2018 04:14, Kemi Wang wrote:
>>> This patch does not have any functionality change, we only provide a spin
>>> count tunes for pthread adaptive spin mutex. The tunable
>>> glibc.mutex.spin_count tunes can be used by system adminstrator to squeeze
>>> system performance according to different hardware capability and workload
>>> model.
>>>
>>> This is the preparation work for the next patch, in which the way of
>>> adaptive spin would be changed from an expensive cmpxchg to read while
>>> spinning.
>>>
>>> * elf/dl-tunables.list: Add glibc.mutex.spin_count entry.
>>> * manual/tunables.texi: Add glibc.mutex.spin_count description.
>>> * nptl/Makefile: Add mutex-conf.c for compilation.
>>> * nptl/mutex-conf.h: New file.
>>> * nptl/mutex-conf.c: New file.
>>>
>>> Suggested-by: Andi Kleen <andi.kleen@intel.com>
>>> Signed-off-by: Kemi Wang <kemi.wang@intel.com>
>>
>>> ---
>>> ChangeLog | 10 ++++++-
>>> elf/dl-tunables.list | 10 +++++++
>>> manual/tunables.texi | 17 ++++++++++++
>>> nptl/Makefile | 3 +-
>>> nptl/mutex-conf.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++++++++
>>> nptl/mutex-conf.h | 31 +++++++++++++++++++++
>>> 6 files changed, 147 insertions(+), 2 deletions(-)
>>> create mode 100644 nptl/mutex-conf.c
>>> create mode 100644 nptl/mutex-conf.h
>>>
>>> diff --git a/ChangeLog b/ChangeLog
>>> index 1f98425..472657c 100644
>>> --- a/ChangeLog
>>> +++ b/ChangeLog
>>> @@ -1,5 +1,13 @@
>>> -2018-03-29 Florian Weimer <fweimer@redhat.com>
>>> +2018-03-30 Kemi Wang <kemi.wang@intel.com>
>>> +
>>> + * elf/dl-tunables.list: Add glibc.mutex.spin_count entry.
>>> + * manual/tunables.texi: Add glibc.mutex.spin_count description.
>>> + * nptl/Makefile: Add mutex-conf.c for compilation.
>>> + * nptl/mutex-conf.h: New file.
>>> + * nptl/mutex-conf.c: New file.
>>> + * nptl/Makefile: Add new file compilation.
>>>
>>> +2018-03-29 Florian Weimer <fweimer@redhat.com>
>>> * sysdeps/unix/sysv/linux/i386/tst-bz21269.c (do_test): Also
>>> capture SIGBUS.
>>>
>>> diff --git a/elf/dl-tunables.list b/elf/dl-tunables.list
>>> index 1f8ecb8..0c27c14 100644
>>> --- a/elf/dl-tunables.list
>>> +++ b/elf/dl-tunables.list
>>> @@ -121,4 +121,14 @@ glibc {
>>> default: 3
>>> }
>>> }
>>> +
>>> + mutex {
>>> + spin_count {
>>> + type: INT_32
>>> + minval: 0
>>> + maxval: 30000
>>> + env_alias: LD_SPIN_COUNT
>>> + default: 1000
>>> + }
>>> + }
>>
>> Indentation seems off here, rest the file uses double space while your
>> patch uses tabs.
>>
>
> Thanks for catching it.
>
>> Also I not sure if it worth to add environment variable for this tunable,
>> I would rather avoid adding newer ones (and naming seems off, since afaik
>> LD_* meaning some parameters that affects the loader).
>>
>
> AFAIK, environment variable is what I can think of to give people the possibility of
> tunes in shell for that, maybe you have better idea?
>
> Yes, we probably use other name like MUTEX_SPIN_COUNT to avoid confusion, agree?
The tunables framework already provides a environment variable to this [1],
the 'env_alias' is mainly to provide compatibility and to use the same logic
internally.
So I think it is better to use the default tunable env var.
[1] https://www.gnu.org/software/libc/manual/html_node/Tunables.html
>
>>> }
>>> diff --git a/manual/tunables.texi b/manual/tunables.texi
>>> index be33c9f..9c6a9f1 100644
>>> --- a/manual/tunables.texi
>>> +++ b/manual/tunables.texi
>>> @@ -281,6 +281,23 @@ of try lock attempts.
>>> The default value of this tunable is @samp{3}.
>>> @end deftp
>>>
>>> +@node Pthread Mutex Tunables
>>> +@section Pthread Mutex Tunables
>>> +@cindex pthread mutex tunables
>>> +
>>> +@deftp {Tunable namespace} glibc.mutex
>>> +Behavior of ptherad mutex can be tuned to acquire performance improvement
>>> +according to specific hardware capablity and workload character by setting
>>> +the following tunables in the @code{mutex} namespace.
>>> +@end deftp
>>
>> There is a typo (s/ptherad/pthread)
>
> thanks for catching it.
>
>> and I think 'acquire' is not the best
>> word here, I would use 'increase' (I am not a native speaker, so someone
>> might suggest a better wording here).
>>
>
> All right, I will ask for some native speakers to help check it.
>
>>> +
>>> +@deftp Tunable glibc.mutex.spin_count
>>> +The @code{glibc.mutex.spin_count} tunable set the maximum times the thread
>>> +should spin on the lock before going to sleep.
>>> +
>>> +The default value of this tunable is @samp{1000}.
>>> +@end deftp
>>
>> I think we need to expand it to specify:
>>
>> * spinning is only used for the mutex initialized with PTHREAD_MUTEX_ADAPTIVE_NP
>> GNU extension(current wording is ambiguous).
>>> * It affects both pthread_mutex_lock and pthread_mutex_timedlock.
>>
>> * The spinning is done by first issuing an atomic operation similar to trylock
>> followed by a arch-specific no operation (not sure if we need to expand
>> how spinning is backoff value is incremented).
>>
> How about this?
> The spinning is done in case of either the maximum spin count
> is reached or lock is acquired during spinning.
It is better for this specific part.
>
>> * Not sure 'sleep' is right terminology here, since for Linux 'futex' can
>> return right away when maximum spin_count it reached.
>>
> 'block' here may be more precise, agree?
> e.g. use "before calling into the kernel to block" or something like that
>
>>> +
>>> @node Hardware Capability Tunables
>>> @section Hardware Capability Tunables
>>> @cindex hardware capability tunables
>>> diff --git a/nptl/Makefile b/nptl/Makefile
>>> index 94be92c..5edacea 100644
>>> --- a/nptl/Makefile
>>> +++ b/nptl/Makefile
>>> @@ -139,7 +139,8 @@ libpthread-routines = nptl-init vars events version pt-interp \
>>> pthread_mutex_getprioceiling \
>>> pthread_mutex_setprioceiling \
>>> pthread_setname pthread_getname \
>>> - pthread_setattr_default_np pthread_getattr_default_np
>>> + pthread_setattr_default_np pthread_getattr_default_np \
>>> + mutex-conf
>>> # pthread_setuid pthread_seteuid pthread_setreuid \
>>> # pthread_setresuid \
>>> # pthread_setgid pthread_setegid pthread_setregid \
>>
>> I think a better name for the file would be pthread_mutex_conf.c.
>>
>
> Agree, will send V2 to fix it.
>
>>> diff --git a/nptl/mutex-conf.c b/nptl/mutex-conf.c
>>> new file mode 100644
>>> index 0000000..f4ffd6d
>>> --- /dev/null
>>> +++ b/nptl/mutex-conf.c
>>> @@ -0,0 +1,78 @@
>>> +/* mutex-conf.c: Pthread mutex tunable parameters.
>>> + Copyright (C) 2013-2018 Free Software Foundation, Inc.
>>> + This file is part of the GNU C Library.
>>> +
>>
>> Copyright should start at 2018 for newer implementations.
>>
>
> I am not aware of that before, do you mean "Copyright (C) 2018-2023"?
No, just "Copyright (C) 2018".
>
>>> + The GNU C Library is free software; you can redistribute it and/or
>>> + modify it under the terms of the GNU Lesser General Public
>>> + License as published by the Free Software Foundation; either
>>> + version 2.1 of the License, or (at your option) any later version.
>>> +
>>> + The GNU C Library is distributed in the hope that it will be useful,
>>> + but WITHOUT ANY WARRANTY; without even the implied warranty of
>>> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
>>> + Lesser General Public License for more details.
>>> +
>>> + You should have received a copy of the GNU Lesser General Public
>>> + License along with the GNU C Library; if not, see
>>> + <http://www.gnu.org/licenses/>. */
>>> +
>>> +#include "config.h"
>>> +#include <pthreadP.h>
>>> +#include <init-arch.h>
>>> +#include <mutex-conf.h>
>>> +#include <unistd.h>
>>> +
>>> +#if HAVE_TUNABLES
>>> +# define TUNABLE_NAMESPACE mutex
>>> +#endif
>>> +#include <elf/dl-tunables.h>
>>> +
>>> +
>>> +struct mutex_config __mutex_aconf =
>>> + {
>>> + /* The maximum times a thread spin on the lock before
>>> + * going to sleep */
>>> + .spin_count = 1000,
>>> + };
>>> +
>>
>> Indentation seems off here and in other places in this file.
>>
>
> Will fix it in V2, thanks for pointing it out.
>
>>> +#if HAVE_TUNABLES
>>> +#define TUNABLE_CALLBACK_FNDECL(__name, __type) \
>>
>> Indentation for nested preprocessor is to just one space each level:
>>
>> #if HAVE_TUNABLES
>> # define ...
>>
>>> +static inline void \
>>> +__always_inline \
>>> +do_set_mutex_ ## __name (__type value) \
>>> +{ \
>>> + __mutex_aconf.__name = value; \
>>> +} \
>>> +void \
>>> +TUNABLE_CALLBACK (set_mutex_ ## __name) (tunable_val_t *valp) \
>>> +{ \
>>> + __type value = (__type) (valp)->numval; \
>>> + do_set_mutex_ ## __name (value); \
>>> +}
>>> +
>>> +TUNABLE_CALLBACK_FNDECL (spin_count, int32_t);
>>> +#endif
>>> +
>>> +static void
>>> +mutex_tunables_init (int argc __attribute__ ((unused)),
>>> + char **argv __attribute__ ((unused)),
>>> + char **environ)
>>> +{
>>> +#if HAVE_TUNABLES
>>> +
>>> + TUNABLE_GET (spin_count, int32_t,
>>> + TUNABLE_CALLBACK (set_mutex_spin_count));
>>> +#endif
>>> +}
>>> +
>>> +#ifdef SHARED
>>> +# define INIT_SECTION ".init_array"
>>> +#else
>>> +# define INIT_SECTION ".preinit_array"
>>> +#endif
>>> +
>>> +void (*const __pthread_mutex_tunables_init_array []) (int, char **, char **)
>>> + __attribute__ ((section (INIT_SECTION), aligned (sizeof (void *)))) =
>>> +{
>>> + &mutex_tunables_init
>>> +};> diff --git a/nptl/mutex-conf.h b/nptl/mutex-conf.h
>>> new file mode 100644
>>> index 0000000..babefe3
>>> --- /dev/null
>>> +++ b/nptl/mutex-conf.h
>>> @@ -0,0 +1,31 @@
>>> +/* mutex-conf.h: Pthread mutex tunable parameters.
>>> + Copyright (C) 2013-2018 Free Software Foundation, Inc.
>>
>> Copyright should start at 2018 for newer implementations.
>>
>
> Sure.
>
>>> + This file is part of the GNU C Library.
>>> +
>>> + The GNU C Library is free software; you can redistribute it and/or
>>> + modify it under the terms of the GNU Lesser General Public
>>> + License as published by the Free Software Foundation; either
>>> + version 2.1 of the License, or (at your option) any later version.
>>> +
>>> + The GNU C Library is distributed in the hope that it will be useful,
>>> + but WITHOUT ANY WARRANTY; without even the implied warranty of
>>> + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
>>> + Lesser General Public License for more details.
>>> +
>>> + You should have received a copy of the GNU Lesser General Public
>>> + License along with the GNU C Library; if not, see
>>> + <http://www.gnu.org/licenses/>. */
>>> +#ifndef _MUTEX_CONF_H
>>> +#define _MUTEX_CONF_H 1
>>> +
>>> +#include <pthread.h>
>>> +#include <time.h>
>>> +
>>> +struct mutex_config
>>> +{
>>> + int spin_count;
>>> +};
>>> +
>>> +extern struct mutex_config __mutex_aconf attribute_hidden;
>>> +
>>> +#endif
>>>
>>