Re: [PATCH v4] libgfortran: Replace mutex with rwlock

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

On 1/1/1970 8:00 AM, Thomas Koenig wrote:
> Hi Lipeng,
> 
>> May I know any comment or concern on this patch, thanks for your time :)
>>
> 
> Thanks for your patience in getting this reviewed.
> 
> A few remarks / questions.
> 
> Which strategy is used in this implementation, read-preferring or write-preferring?  And if read-
> preferring is used, is there a danger of deadlock if people do unreasonable things?
> Maybe you could explain that, also in a comment in the code >
> Can you add some sort of torture test case(s) which does a lot of opening/closing/reading/writing,
> possibly with asynchronous I/O and/or pthreads, to catch possible problems?  If there is a system
> dependency or some race condition, chances are that regression testers will catch this.

Hi Thomas,

Thanks for your time for the review.
Sure, I will add test case according to your suggestions and update the 
comment based on the implementation of "read-preferring" strategy.

Thanks,
Lipeng Zhu
> 
> With this, the libgfortran parts are OK, unless somebody else has more comments, so give this a couple
> of days.  I cannot approve the libgcc parts, that would be somebody else (Jakub?)
> 
> Best regards
> 
> 	Thomas
> 
> 

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> > [CCing Ian as libgcc maintainer]
> >
> > On Wed, 1 Nov 2023 10:14:37 +0000
> > "Zhu, Lipeng" <lipeng.zhu@intel.com> wrote:
> >
> > > > >
> > > > > Hi Lipeng,
> > > > >
> > > > > >>> Sure, as your comments, in the patch V6, I added 3 test
> > > > > >>> cases with OpenMP to test different cases in concurrency
> respectively:
> > > > > >>> 1. find and create unit very frequently to stress read lock
> > > > > >>> and write
> > lock.
> > > > > >>> 2. only access the unit which exist in cache to stress read lock.
> > > > > >>> 3. access the same unit in concurrency.
> > > > > >>> For the third test case, it also help to find a bug:  When
> > > > > >>> unit can't be found in cache nor unit list in read phase,
> > > > > >>> then threads will try to acquire write lock to insert the
> > > > > >>> same unit, this will cause duplicate key
> > > > > >> error.
> > > > > >>> To fix this bug, I get the unit from unit list once again
> > > > > >>> before insert in write
> > > > > >> lock.
> > > > > >>> More details you can refer the patch v6.
> > > > > >>>
> > > > > >>
> > > > > >> Could you help to review this update? I really appreciate
> > > > > >> your
> > assistance.
> > > > > >>
> > > > >
> > > > > > Could you help to review this update?  Any concern will be
> > appreciated.
> > > > >
> > > > > Fortran parts are OK (I think I wrote that already), we need
> > > > > somebody for the non-Fortran parts.
> > > > >
> > > > Hi Thomas,
> > > >
> > > > Thanks for your response. Very appreciate for your patience and help.
> > > >
> > > > > Jakub, could you maybe take a look?
> > > > >
> > > > > Best regards
> > > > >
> > > > > 	Thomas
> > > >
> > > > Hi Jakub,
> > > >
> > > > Can you help to take a look at the change for libgcc part that
> > > > added several rwlock macros in libgcc/gthr-posix.h?
> > > >
> > >
> > > Hi Jakub,
> > >
> > > Could you help to review this, any comment will be greatly appreciated.
> >
> > Latest version is at
> > https://inbox.sourceware.org/gcc-patches/20230818031818.2161842-1-
> > lipeng.zhu@intel.com/
> >
> Thanks Bernhard.
> 
> Hi Ian,
> Could you help to review the changes for libgcc part?
> Very looking forward to your help.
> 
> > >
> > > > Best Regards,
> > > > Lipeng Zhu
> > >

Hi Jakub, 

Could you help to review this patch for the changes in libgcc/gthr-posix.h.
Just as Thomas commented: "Fortran parts are OK".  We need your 
comments for the non-fortran parts.  Very appreciated for your help.
Latest version is at https://inbox.sourceware.org/gcc-patches/20230818031818.2161842-1-lipeng.zhu@intel.com/

Lipeng Zhu,
Best Regards.

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> [CCing Ian as libgcc maintainer]
> 
> On Wed, 1 Nov 2023 10:14:37 +0000
> "Zhu, Lipeng" <lipeng.zhu@intel.com> wrote:
> 
> > > >
> > > > Hi Lipeng,
> > > >
> > > > >>> Sure, as your comments, in the patch V6, I added 3 test cases
> > > > >>> with OpenMP to test different cases in concurrency respectively:
> > > > >>> 1. find and create unit very frequently to stress read lock and write
> lock.
> > > > >>> 2. only access the unit which exist in cache to stress read lock.
> > > > >>> 3. access the same unit in concurrency.
> > > > >>> For the third test case, it also help to find a bug:  When
> > > > >>> unit can't be found in cache nor unit list in read phase, then
> > > > >>> threads will try to acquire write lock to insert the same
> > > > >>> unit, this will cause duplicate key
> > > > >> error.
> > > > >>> To fix this bug, I get the unit from unit list once again
> > > > >>> before insert in write
> > > > >> lock.
> > > > >>> More details you can refer the patch v6.
> > > > >>>
> > > > >>
> > > > >> Could you help to review this update? I really appreciate your
> assistance.
> > > > >>
> > > >
> > > > > Could you help to review this update?  Any concern will be
> appreciated.
> > > >
> > > > Fortran parts are OK (I think I wrote that already), we need
> > > > somebody for the non-Fortran parts.
> > > >
> > > Hi Thomas,
> > >
> > > Thanks for your response. Very appreciate for your patience and help.
> > >
> > > > Jakub, could you maybe take a look?
> > > >
> > > > Best regards
> > > >
> > > > 	Thomas
> > >
> > > Hi Jakub,
> > >
> > > Can you help to take a look at the change for libgcc part that added
> > > several rwlock macros in libgcc/gthr-posix.h?
> > >
> >
> > Hi Jakub,
> >
> > Could you help to review this, any comment will be greatly appreciated.
> 
> Latest version is at
> https://inbox.sourceware.org/gcc-patches/20230818031818.2161842-1-
> lipeng.zhu@intel.com/
> 
Thanks Bernhard.

Hi Ian, 
Could you help to review the changes for libgcc part?  
Very looking forward to your help.

> >
> > > Best Regards,
> > > Lipeng Zhu
> >

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Bernhard Reutner-Fischer]

[CCing Ian as libgcc maintainer]

On Wed, 1 Nov 2023 10:14:37 +0000
"Zhu, Lipeng" <lipeng.zhu@intel.com> wrote:

> > >
> > > Hi Lipeng,
> > >  
> > > >>> Sure, as your comments, in the patch V6, I added 3 test cases with
> > > >>> OpenMP to test different cases in concurrency respectively:
> > > >>> 1. find and create unit very frequently to stress read lock and write lock.
> > > >>> 2. only access the unit which exist in cache to stress read lock.
> > > >>> 3. access the same unit in concurrency.
> > > >>> For the third test case, it also help to find a bug:  When unit
> > > >>> can't be found in cache nor unit list in read phase, then threads
> > > >>> will try to acquire write lock to insert the same unit, this will
> > > >>> cause duplicate key  
> > > >> error.  
> > > >>> To fix this bug, I get the unit from unit list once again before
> > > >>> insert in write  
> > > >> lock.  
> > > >>> More details you can refer the patch v6.
> > > >>>  
> > > >>
> > > >> Could you help to review this update? I really appreciate your assistance.
> > > >>  
> > >  
> > > > Could you help to review this update?  Any concern will be appreciated.  
> > >
> > > Fortran parts are OK (I think I wrote that already), we need somebody
> > > for the non-Fortran parts.
> > >  
> > Hi Thomas,
> > 
> > Thanks for your response. Very appreciate for your patience and help.
> >   
> > > Jakub, could you maybe take a look?
> > >
> > > Best regards
> > >
> > > 	Thomas  
> > 
> > Hi Jakub,
> > 
> > Can you help to take a look at the change for libgcc part that added several
> > rwlock macros in libgcc/gthr-posix.h?
> >   
> 
> Hi Jakub,
> 
> Could you help to review this, any comment will be greatly appreciated.

Latest version is at
https://inbox.sourceware.org/gcc-patches/20230818031818.2161842-1-lipeng.zhu@intel.com/

> 
> > Best Regards,
> > Lipeng Zhu  
> 

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> >
> > Hi Lipeng,
> >
> > >>> Sure, as your comments, in the patch V6, I added 3 test cases with
> > >>> OpenMP to test different cases in concurrency respectively:
> > >>> 1. find and create unit very frequently to stress read lock and write lock.
> > >>> 2. only access the unit which exist in cache to stress read lock.
> > >>> 3. access the same unit in concurrency.
> > >>> For the third test case, it also help to find a bug:  When unit
> > >>> can't be found in cache nor unit list in read phase, then threads
> > >>> will try to acquire write lock to insert the same unit, this will
> > >>> cause duplicate key
> > >> error.
> > >>> To fix this bug, I get the unit from unit list once again before
> > >>> insert in write
> > >> lock.
> > >>> More details you can refer the patch v6.
> > >>>
> > >>
> > >> Could you help to review this update? I really appreciate your assistance.
> > >>
> >
> > > Could you help to review this update?  Any concern will be appreciated.
> >
> > Fortran parts are OK (I think I wrote that already), we need somebody
> > for the non-Fortran parts.
> >
> Hi Thomas,
> 
> Thanks for your response. Very appreciate for your patience and help.
> 
> > Jakub, could you maybe take a look?
> >
> > Best regards
> >
> > 	Thomas
> 
> Hi Jakub,
> 
> Can you help to take a look at the change for libgcc part that added several
> rwlock macros in libgcc/gthr-posix.h?
> 

Hi Jakub,

Could you help to review this, any comment will be greatly appreciated.

> Best Regards,
> Lipeng Zhu

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> 
> Hi Lipeng,
> 
> >>> Sure, as your comments, in the patch V6, I added 3 test cases with
> >>> OpenMP to test different cases in concurrency respectively:
> >>> 1. find and create unit very frequently to stress read lock and write lock.
> >>> 2. only access the unit which exist in cache to stress read lock.
> >>> 3. access the same unit in concurrency.
> >>> For the third test case, it also help to find a bug:  When unit
> >>> can't be found in cache nor unit list in read phase, then threads
> >>> will try to acquire write lock to insert the same unit, this will
> >>> cause duplicate key
> >> error.
> >>> To fix this bug, I get the unit from unit list once again before
> >>> insert in write
> >> lock.
> >>> More details you can refer the patch v6.
> >>>
> >>
> >> Could you help to review this update? I really appreciate your assistance.
> >>
> 
> > Could you help to review this update?  Any concern will be appreciated.
> 
> Fortran parts are OK (I think I wrote that already), we need somebody for the
> non-Fortran parts.
> 
Hi Thomas,

Thanks for your response. Very appreciate for your patience and help.

> Jakub, could you maybe take a look?
> 
> Best regards
> 
> 	Thomas

Hi Jakub,

Can you help to take a look at the change for libgcc part that added 
several rwlock macros in libgcc/gthr-posix.h?

Best Regards,
Lipeng Zhu

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Thomas Koenig]

Hi Lipeng,

>>> Sure, as your comments, in the patch V6, I added 3 test cases with
>>> OpenMP to test different cases in concurrency respectively:
>>> 1. find and create unit very frequently to stress read lock and write lock.
>>> 2. only access the unit which exist in cache to stress read lock.
>>> 3. access the same unit in concurrency.
>>> For the third test case, it also help to find a bug:  When unit can't
>>> be found in cache nor unit list in read phase, then threads will try
>>> to acquire write lock to insert the same unit, this will cause duplicate key
>> error.
>>> To fix this bug, I get the unit from unit list once again before insert in write
>> lock.
>>> More details you can refer the patch v6.
>>>
>>
>> Could you help to review this update? I really appreciate your assistance.
>>

> Could you help to review this update?  Any concern will be appreciated.

Fortran parts are OK (I think I wrote that already), we need somebody
for the non-Fortran parts.

Jakub, could you maybe take a look?

Best regards

	Thomas

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> 
> > Hi Thomas,
> >
> > >
> > > Hi Lipeng,
> > >
> > > > May I know any comment or concern on this patch, thanks for your
> > > > time
> > > > 😄
> > >
> > > Thanks for your patience in getting this reviewed.
> > >
> > > A few remarks / questions.
> > >
> > > Which strategy is used in this implementation, read-preferring or
> > > write- preferring?  And if read-preferring is used, is there a
> > > danger of deadlock if people do unreasonable things?
> > > Maybe you could explain that, also in a comment in the code.
> > >
> >
> > Yes, the implementation use the read-preferring strategy, and comments
> > in code.
> > When adding the test cases, I didn’t meet the situation which may
> > cause the deadlock.
> > Maybe you can give more guidance about that.
> >
> > > Can you add some sort of torture test case(s) which does a lot of
> > > opening/closing/reading/writing, possibly with asynchronous I/O
> > > and/or pthreads, to catch possible problems?  If there is a system
> > > dependency or some race condition, chances are that regression testers
> will catch this.
> > >
> >
> > Sure, as your comments, in the patch V6, I added 3 test cases with
> > OpenMP to test different cases in concurrency respectively:
> > 1. find and create unit very frequently to stress read lock and write lock.
> > 2. only access the unit which exist in cache to stress read lock.
> > 3. access the same unit in concurrency.
> > For the third test case, it also help to find a bug:  When unit can't
> > be found in cache nor unit list in read phase, then threads will try
> > to acquire write lock to insert the same unit, this will cause duplicate key
> error.
> > To fix this bug, I get the unit from unit list once again before insert in write
> lock.
> > More details you can refer the patch v6.
> >
> 
> Could you help to review this update? I really appreciate your assistance.
> 

Hi Thomas, Bernhard,

Could you help to review this update?  Any concern will be appreciated.

Regards,
Lipeng Zhu
> > > With this, the libgfortran parts are OK, unless somebody else has
> > > more comments, so give this a couple of days.  I cannot approve the
> > > libgcc parts, that would be somebody else (Jakub?)
> > >
> > > Best regards
> > >
> > > 	Thomas
> > >
> >
> > Best Regards,
> > Lipeng Zhu
> 
> Best Regards,
> Lipeng Zhu

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

> Hi Thomas,
> 
> >
> > Hi Lipeng,
> >
> > > May I know any comment or concern on this patch, thanks for your
> > > time
> > > 😄
> >
> > Thanks for your patience in getting this reviewed.
> >
> > A few remarks / questions.
> >
> > Which strategy is used in this implementation, read-preferring or
> > write- preferring?  And if read-preferring is used, is there a danger
> > of deadlock if people do unreasonable things?
> > Maybe you could explain that, also in a comment in the code.
> >
> 
> Yes, the implementation use the read-preferring strategy, and comments in
> code.
> When adding the test cases, I didn’t meet the situation which may cause the
> deadlock.
> Maybe you can give more guidance about that.
> 
> > Can you add some sort of torture test case(s) which does a lot of
> > opening/closing/reading/writing, possibly with asynchronous I/O and/or
> > pthreads, to catch possible problems?  If there is a system dependency
> > or some race condition, chances are that regression testers will catch this.
> >
> 
> Sure, as your comments, in the patch V6, I added 3 test cases with OpenMP to
> test different cases in concurrency respectively:
> 1. find and create unit very frequently to stress read lock and write lock.
> 2. only access the unit which exist in cache to stress read lock.
> 3. access the same unit in concurrency.
> For the third test case, it also help to find a bug:  When unit can't be found in
> cache nor unit list in read phase, then threads will try to acquire write lock to
> insert the same unit, this will cause duplicate key error.
> To fix this bug, I get the unit from unit list once again before insert in write lock.
> More details you can refer the patch v6.
> 

Could you help to review this update? I really appreciate your assistance.

> > With this, the libgfortran parts are OK, unless somebody else has more
> > comments, so give this a couple of days.  I cannot approve the libgcc
> > parts, that would be somebody else (Jakub?)
> >
> > Best regards
> >
> > 	Thomas
> >
> 
> Best Regards,
> Lipeng Zhu

Best Regards,
Lipeng Zhu

RE: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

Hi Thomas,

> 
> Hi Lipeng,
> 
> > May I know any comment or concern on this patch, thanks for your time
> > 😄
> 
> Thanks for your patience in getting this reviewed.
> 
> A few remarks / questions.
> 
> Which strategy is used in this implementation, read-preferring or write-
> preferring?  And if read-preferring is used, is there a danger of deadlock if
> people do unreasonable things?
> Maybe you could explain that, also in a comment in the code.
> 

Yes, the implementation use the read-preferring strategy, and comments in code.
When adding the test cases, I didn’t meet the situation which may cause the deadlock.
Maybe you can give more guidance about that.

> Can you add some sort of torture test case(s) which does a lot of
> opening/closing/reading/writing, possibly with asynchronous I/O and/or
> pthreads, to catch possible problems?  If there is a system dependency or
> some race condition, chances are that regression testers will catch this.
> 

Sure, as your comments, in the patch V6, I added 3 test cases with OpenMP to test different cases in concurrency respectively:
1. find and create unit very frequently to stress read lock and write lock.
2. only access the unit which exist in cache to stress read lock.
3. access the same unit in concurrency.
For the third test case, it also help to find a bug:  When unit can't be found in cache nor unit list in read phase, then threads will try to acquire write lock to insert the same unit, this will cause duplicate key error.  
To fix this bug, I get the unit from unit list once again before insert in write lock. More details you can refer the patch v6.

> With this, the libgfortran parts are OK, unless somebody else has more
> comments, so give this a couple of days.  I cannot approve the libgcc parts,
> that would be somebody else (Jakub?)
> 
> Best regards
> 
> 	Thomas
> 

Best Regards,
Lipeng Zhu

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Thomas Koenig]

Hi Lipeng,

> May I know any comment or concern on this patch, thanks for your time 😄

Thanks for your patience in getting this reviewed.

A few remarks / questions.

Which strategy is used in this implementation, read-preferring or
write-preferring?  And if read-preferring is used, is there
a danger of deadlock if people do unreasonable things?
Maybe you could explain that, also in a comment in the code.

Can you add some sort of torture test case(s) which does a lot of
opening/closing/reading/writing, possibly with asynchronous
I/O and/or pthreads, to catch possible problems?  If there is a
system dependency or some race condition, chances are that regression
testers will catch this.

With this, the libgfortran parts are OK, unless somebody else has more
comments, so give this a couple of days.  I cannot approve the libgcc
parts, that would be somebody else (Jakub?)

Best regards

	Thomas

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

On 5/16/2023 3:08 PM, Zhu, Lipeng wrote:
> 
> 
> On 5/9/2023 10:32 AM, Zhu, Lipeng wrote:
>>
>>
>> On 1/1/1970 8:00 AM, Bernhard Reutner-Fischer wrote:
>>> On Mon,  8 May 2023 17:44:43 +0800
>>> Lipeng Zhu <lipeng.zhu@intel.com> wrote:
>>>
>>>> This patch try to introduce the rwlock and split the read/write to
>>>> unit_root tree and unit_cache with rwlock instead of the mutex to
>>>> increase CPU efficiency. In the get_gfc_unit function, the percentage
>>>> to step into the insert_unit function is around 30%, in most
>>>> instances, we can get the unit in the phase of reading the unit_cache
>>>> or unit_root tree. So split the read/write phase by rwlock would be an
>>>> approach to make it more parallel.
>>>>
>>>> BTW, the IPC metrics can gain around 9x in our test server with 220
>>>> cores. The benchmark we used is https://github.com/rwesson/NEAT
>>>
>>> See commentary typos below.
>>> You did not state if you regression tested the patch?
>> I use valgrind --tool=helgrind or --tool=drd to test 'make 
>> check-fortran'. Is it necessary to add an additional unit test for 
>> this patch?
>>
>>> Other than that it LGTM but i cannot approve it.
>> Thank you for your kind help for this patch, is there anything that I 
>> can do or can you help to push this patch forward?
>>
> Hi Bernhard,
> 
> Is there any other refinement that need I to do for this patch?
> 
> Thanks.
> 
> 

May I know any comment or concern on this patch, thanks for your time :)

>>>
>>>> diff --git a/libgfortran/io/async.h b/libgfortran/io/async.h index
>>>> ad226c8e856..0033cc74252 100644
>>>> --- a/libgfortran/io/async.h
>>>> +++ b/libgfortran/io/async.h
>>>> @@ -210,6 +210,128 @@
>>>>       DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s 
>>>> %78p\n", aio_prefix, #mutex,
>>> mutex); \
>> Thanks, corrected in Patch v5.
>>
>>>>     } while (0)
>>>> +#ifdef __GTHREAD_RWLOCK_INIT
>>>> +#define RWLOCK_DEBUG_ADD(rwlock) do {        \
>>>> +    aio_rwlock_debug *n;                \
>>>> +    n = xmalloc (sizeof(aio_rwlock_debug));    \
>>>
>>> Missing space before the open brace: sizeof (
>>>
>> Thanks, corrected in Patch v5.
>>
>>>> diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c index
>>>> 82664dc5f98..62f1db21d34 100644
>>>> --- a/libgfortran/io/unit.c
>>>> +++ b/libgfortran/io/unit.c
>>>> @@ -33,34 +33,36 @@ see the files COPYING3 and COPYING.RUNTIME
>>>> respectively.  If not, see
>>>>   /* IO locking rules:
>>>> -   UNIT_LOCK is a master lock, protecting UNIT_ROOT tree and 
>>>> UNIT_CACHE.
>>>> +   UNIT_RWLOCK is a master lock, protecting UNIT_ROOT tree and 
>>>> UNIT_CACHE.
>>>> +   And use the rwlock to spilt read and write phase to UNIT_ROOT tree
>>>> +   and UNIT_CACHE to increase CPU efficiency.
>>>
>>> s/spilt/split. Maybe:
>>>
>>> Using an rwlock improves efficiency by allowing us to separate 
>>> readers and writers of both UNIT_ROOT
>>> and UNIT_CACHE.
>>>
>> Thanks, corrected in Patch v5.
>>
>>>> @@ -350,6 +356,17 @@ retry:
>>>>         if (c == 0)
>>>>       break;
>>>>       }
>>>> +  /* We did not find a unit in the cache nor in the unit list, 
>>>> create a new
>>>> +    (locked) unit and insert into the unit list and cache.
>>>> +    Manipulating either or both the unit list and the unit cache 
>>>> requires to
>>>> +    hold a write-lock [for obvious reasons]:
>>>> +    1. By separating the read/write lock, it will greatly reduce 
>>>> the contention
>>>> +       at the read part, while write part is not always necessary 
>>>> or most
>>>> +       unlikely once the unit hit in cache.
>>>
>>> +    By separating the read/write lock, we will greatly reduce the 
>>> contention
>>> +    on the read part, while the write part is unlikely once the unit 
>>> hits
>>> +    the cache.
>>>
>>>> +    2. We try to balance the implementation complexity and the 
>>>> performance
>>>> +       gains that fit into current cases we observed by just using a
>>>> +       pthread_rwlock. */
>>>
>>> Let's drop 2.
>>
>> Got it, thanks!
>>> thanks,

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

On 5/9/2023 10:32 AM, Zhu, Lipeng wrote:
> 
> 
> On 1/1/1970 8:00 AM, Bernhard Reutner-Fischer wrote:
>> On Mon,  8 May 2023 17:44:43 +0800
>> Lipeng Zhu <lipeng.zhu@intel.com> wrote:
>>
>>> This patch try to introduce the rwlock and split the read/write to
>>> unit_root tree and unit_cache with rwlock instead of the mutex to
>>> increase CPU efficiency. In the get_gfc_unit function, the percentage
>>> to step into the insert_unit function is around 30%, in most
>>> instances, we can get the unit in the phase of reading the unit_cache
>>> or unit_root tree. So split the read/write phase by rwlock would be an
>>> approach to make it more parallel.
>>>
>>> BTW, the IPC metrics can gain around 9x in our test server with 220
>>> cores. The benchmark we used is https://github.com/rwesson/NEAT
>>
>> See commentary typos below.
>> You did not state if you regression tested the patch?
> I use valgrind --tool=helgrind or --tool=drd to test 'make 
> check-fortran'. Is it necessary to add an additional unit test for this 
> patch?
> 
>> Other than that it LGTM but i cannot approve it.
> Thank you for your kind help for this patch, is there anything that I 
> can do or can you help to push this patch forward?
> 
Hi Bernhard,

Is there any other refinement that need I to do for this patch?

Thanks.


>>
>>> diff --git a/libgfortran/io/async.h b/libgfortran/io/async.h index
>>> ad226c8e856..0033cc74252 100644
>>> --- a/libgfortran/io/async.h
>>> +++ b/libgfortran/io/async.h
>>> @@ -210,6 +210,128 @@
>>>       DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", 
>>> aio_prefix, #mutex,
>> mutex); \
> Thanks, corrected in Patch v5.
> 
>>>     } while (0)
>>> +#ifdef __GTHREAD_RWLOCK_INIT
>>> +#define RWLOCK_DEBUG_ADD(rwlock) do {        \
>>> +    aio_rwlock_debug *n;                \
>>> +    n = xmalloc (sizeof(aio_rwlock_debug));    \
>>
>> Missing space before the open brace: sizeof (
>>
> Thanks, corrected in Patch v5.
> 
>>> diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c index
>>> 82664dc5f98..62f1db21d34 100644
>>> --- a/libgfortran/io/unit.c
>>> +++ b/libgfortran/io/unit.c
>>> @@ -33,34 +33,36 @@ see the files COPYING3 and COPYING.RUNTIME
>>> respectively.  If not, see
>>>   /* IO locking rules:
>>> -   UNIT_LOCK is a master lock, protecting UNIT_ROOT tree and 
>>> UNIT_CACHE.
>>> +   UNIT_RWLOCK is a master lock, protecting UNIT_ROOT tree and 
>>> UNIT_CACHE.
>>> +   And use the rwlock to spilt read and write phase to UNIT_ROOT tree
>>> +   and UNIT_CACHE to increase CPU efficiency.
>>
>> s/spilt/split. Maybe:
>>
>> Using an rwlock improves efficiency by allowing us to separate readers 
>> and writers of both UNIT_ROOT
>> and UNIT_CACHE.
>>
> Thanks, corrected in Patch v5.
> 
>>> @@ -350,6 +356,17 @@ retry:
>>>         if (c == 0)
>>>       break;
>>>       }
>>> +  /* We did not find a unit in the cache nor in the unit list, 
>>> create a new
>>> +    (locked) unit and insert into the unit list and cache.
>>> +    Manipulating either or both the unit list and the unit cache 
>>> requires to
>>> +    hold a write-lock [for obvious reasons]:
>>> +    1. By separating the read/write lock, it will greatly reduce the 
>>> contention
>>> +       at the read part, while write part is not always necessary or 
>>> most
>>> +       unlikely once the unit hit in cache.
>>
>> +    By separating the read/write lock, we will greatly reduce the 
>> contention
>> +    on the read part, while the write part is unlikely once the unit 
>> hits
>> +    the cache.
>>
>>> +    2. We try to balance the implementation complexity and the 
>>> performance
>>> +       gains that fit into current cases we observed by just using a
>>> +       pthread_rwlock. */
>>
>> Let's drop 2.
> 
> Got it, thanks!
>> thanks,

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Zhu, Lipeng]

On 1/1/1970 8:00 AM, Bernhard Reutner-Fischer wrote:
> On Mon,  8 May 2023 17:44:43 +0800
> Lipeng Zhu <lipeng.zhu@intel.com> wrote:
> 
>> This patch try to introduce the rwlock and split the read/write to
>> unit_root tree and unit_cache with rwlock instead of the mutex to
>> increase CPU efficiency. In the get_gfc_unit function, the percentage
>> to step into the insert_unit function is around 30%, in most
>> instances, we can get the unit in the phase of reading the unit_cache
>> or unit_root tree. So split the read/write phase by rwlock would be an
>> approach to make it more parallel.
>>
>> BTW, the IPC metrics can gain around 9x in our test server with 220
>> cores. The benchmark we used is https://github.com/rwesson/NEAT
> 
> See commentary typos below.
> You did not state if you regression tested the patch?
I use valgrind --tool=helgrind or --tool=drd to test 'make 
check-fortran'. Is it necessary to add an additional unit test for this 
patch?

> Other than that it LGTM but i cannot approve it.
Thank you for your kind help for this patch, is there anything that I 
can do or can you help to push this patch forward?

> 
>> diff --git a/libgfortran/io/async.h b/libgfortran/io/async.h index
>> ad226c8e856..0033cc74252 100644
>> --- a/libgfortran/io/async.h
>> +++ b/libgfortran/io/async.h
>> @@ -210,6 +210,128 @@
>>       DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", aio_prefix, #mutex,
> mutex); \
Thanks, corrected in Patch v5.

>>     } while (0)
>>   
>> +#ifdef __GTHREAD_RWLOCK_INIT
>> +#define RWLOCK_DEBUG_ADD(rwlock) do {		\
>> +    aio_rwlock_debug *n;				\
>> +    n = xmalloc (sizeof(aio_rwlock_debug));	\
> 
> Missing space before the open brace: sizeof (
> 
Thanks, corrected in Patch v5.

>> diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c index
>> 82664dc5f98..62f1db21d34 100644
>> --- a/libgfortran/io/unit.c
>> +++ b/libgfortran/io/unit.c
>> @@ -33,34 +33,36 @@ see the files COPYING3 and COPYING.RUNTIME
>> respectively.  If not, see
>>   
>>   
>>   /* IO locking rules:
>> -   UNIT_LOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
>> +   UNIT_RWLOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
>> +   And use the rwlock to spilt read and write phase to UNIT_ROOT tree
>> +   and UNIT_CACHE to increase CPU efficiency.
> 
> s/spilt/split. Maybe:
> 
> Using an rwlock improves efficiency by allowing us to separate readers and writers of both UNIT_ROOT
> and UNIT_CACHE.
> 
Thanks, corrected in Patch v5.

>> @@ -350,6 +356,17 @@ retry:
>>         if (c == 0)
>>   	break;
>>       }
>> +  /* We did not find a unit in the cache nor in the unit list, create a new
>> +    (locked) unit and insert into the unit list and cache.
>> +    Manipulating either or both the unit list and the unit cache requires to
>> +    hold a write-lock [for obvious reasons]:
>> +    1. By separating the read/write lock, it will greatly reduce the contention
>> +       at the read part, while write part is not always necessary or most
>> +       unlikely once the unit hit in cache.
> 
> +    By separating the read/write lock, we will greatly reduce the contention
> +    on the read part, while the write part is unlikely once the unit hits
> +    the cache.
> 
>> +    2. We try to balance the implementation complexity and the performance
>> +       gains that fit into current cases we observed by just using a
>> +       pthread_rwlock. */
> 
> Let's drop 2.

Got it, thanks!
> thanks,

Re: [PATCH v4] libgfortran: Replace mutex with rwlock

[Bernhard Reutner-Fischer]

On Mon,  8 May 2023 17:44:43 +0800
Lipeng Zhu <lipeng.zhu@intel.com> wrote:

> This patch try to introduce the rwlock and split the read/write to
> unit_root tree and unit_cache with rwlock instead of the mutex to
> increase CPU efficiency. In the get_gfc_unit function, the percentage
> to step into the insert_unit function is around 30%, in most instances,
> we can get the unit in the phase of reading the unit_cache or unit_root
> tree. So split the read/write phase by rwlock would be an approach to
> make it more parallel.
> 
> BTW, the IPC metrics can gain around 9x in our test
> server with 220 cores. The benchmark we used is
> https://github.com/rwesson/NEAT

See commentary typos below.
You did not state if you regression tested the patch?
Other than that it LGTM but i cannot approve it.

> diff --git a/libgfortran/io/async.h b/libgfortran/io/async.h
> index ad226c8e856..0033cc74252 100644
> --- a/libgfortran/io/async.h
> +++ b/libgfortran/io/async.h
> @@ -210,6 +210,128 @@
>      DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", aio_prefix, #mutex, mutex); \
>    } while (0)
>  
> +#ifdef __GTHREAD_RWLOCK_INIT
> +#define RWLOCK_DEBUG_ADD(rwlock) do {		\
> +    aio_rwlock_debug *n;				\
> +    n = xmalloc (sizeof(aio_rwlock_debug));	\

Missing space before the open brace: sizeof (

> diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c
> index 82664dc5f98..62f1db21d34 100644
> --- a/libgfortran/io/unit.c
> +++ b/libgfortran/io/unit.c
> @@ -33,34 +33,36 @@ see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
>  
>  
>  /* IO locking rules:
> -   UNIT_LOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
> +   UNIT_RWLOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
> +   And use the rwlock to spilt read and write phase to UNIT_ROOT tree
> +   and UNIT_CACHE to increase CPU efficiency.

s/spilt/split. Maybe:

Using an rwlock improves efficiency by allowing us to separate readers
and writers of both UNIT_ROOT and UNIT_CACHE.

> @@ -350,6 +356,17 @@ retry:
>        if (c == 0)
>  	break;
>      }
> +  /* We did not find a unit in the cache nor in the unit list, create a new
> +    (locked) unit and insert into the unit list and cache.
> +    Manipulating either or both the unit list and the unit cache requires to
> +    hold a write-lock [for obvious reasons]:
> +    1. By separating the read/write lock, it will greatly reduce the contention
> +       at the read part, while write part is not always necessary or most
> +       unlikely once the unit hit in cache.

+    By separating the read/write lock, we will greatly reduce the contention
+    on the read part, while the write part is unlikely once the unit hits
+    the cache.

> +    2. We try to balance the implementation complexity and the performance
> +       gains that fit into current cases we observed by just using a
> +       pthread_rwlock. */

Let's drop 2.
thanks,

[PATCH v4] libgfortran: Replace mutex with rwlock

[Lipeng Zhu]

This patch try to introduce the rwlock and split the read/write to
unit_root tree and unit_cache with rwlock instead of the mutex to
increase CPU efficiency. In the get_gfc_unit function, the percentage
to step into the insert_unit function is around 30%, in most instances,
we can get the unit in the phase of reading the unit_cache or unit_root
tree. So split the read/write phase by rwlock would be an approach to
make it more parallel.

BTW, the IPC metrics can gain around 9x in our test
server with 220 cores. The benchmark we used is
https://github.com/rwesson/NEAT

libgcc/ChangeLog:

    * gthr-posix.h (__GTHREAD_RWLOCK_INIT): New macro
    (__gthrw): New function
    (__gthread_rwlock_rdlock): New function
    (__gthread_rwlock_tryrdlock): New function
    (__gthread_rwlock_wrlock): New function
    (__gthread_rwlock_trywrlock): New function
    (__gthread_rwlock_unlock): New function

libgfortran/ChangeLog:

    * io/async.c (DEBUG_LINE): New
    * io/async.h (RWLOCK_DEBUG_ADD): New macro
    (CHECK_RDLOCK): New macro
    (CHECK_WRLOCK): New macro
    (TAIL_RWLOCK_DEBUG_QUEUE): New macro
    (IN_RWLOCK_DEBUG_QUEUE): New macro
    (RDLOCK): New macro
    (WRLOCK): New macro
    (RWUNLOCK): New macro
    (RD_TO_WRLOCK): New macro
    (INTERN_RDLOCK): New macro
    (INTERN_WRLOCK): New macro
    (INTERN_RWUNLOCK): New macro
    * io/io.h (internal_proto): Define unit_rwlock
    * io/transfer.c (st_read_done_worker): Relace unit_lock with unit_rwlock
    (st_write_done_worker): Relace unit_lock with unit_rwlock
    * io/unit.c (get_gfc_unit): Relace unit_lock with unit_rwlock
    (if): Relace unit_lock with unit_rwlock
    (close_unit_1): Relace unit_lock with unit_rwlock
    (close_units): Relace unit_lock with unit_rwlock
    (newunit_alloc): Relace unit_lock with unit_rwlock
    * io/unix.c (flush_all_units): Relace unit_lock with unit_rwlock

---
v1 -> v2:
Limit the pthread_rwlock usage in libgcc only when __cplusplus isn't defined.

v2 -> v3:
Rebase the patch with trunk branch.

v3 -> v4:
Update the comments.

Reviewed-by: Hongjiu Lu <hongjiu.lu@intel.com>
Signed-off-by: Lipeng Zhu <lipeng.zhu@intel.com>
---
 libgcc/gthr-posix.h       |  60 +++++++++++++++
 libgfortran/io/async.c    |   4 +
 libgfortran/io/async.h    | 151 ++++++++++++++++++++++++++++++++++++++
 libgfortran/io/io.h       |  15 ++--
 libgfortran/io/transfer.c |   8 +-
 libgfortran/io/unit.c     |  75 +++++++++++--------
 libgfortran/io/unix.c     |  16 ++--
 7 files changed, 283 insertions(+), 46 deletions(-)

diff --git a/libgcc/gthr-posix.h b/libgcc/gthr-posix.h
index aebcfdd9f4c..73283082997 100644
--- a/libgcc/gthr-posix.h
+++ b/libgcc/gthr-posix.h
@@ -48,6 +48,9 @@ typedef pthread_t __gthread_t;
 typedef pthread_key_t __gthread_key_t;
 typedef pthread_once_t __gthread_once_t;
 typedef pthread_mutex_t __gthread_mutex_t;
+#ifndef __cplusplus
+typedef pthread_rwlock_t __gthread_rwlock_t;
+#endif
 typedef pthread_mutex_t __gthread_recursive_mutex_t;
 typedef pthread_cond_t __gthread_cond_t;
 typedef struct timespec __gthread_time_t;
@@ -58,6 +61,9 @@ typedef struct timespec __gthread_time_t;
 
 #define __GTHREAD_MUTEX_INIT PTHREAD_MUTEX_INITIALIZER
 #define __GTHREAD_MUTEX_INIT_FUNCTION __gthread_mutex_init_function
+#ifndef __cplusplus
+#define __GTHREAD_RWLOCK_INIT PTHREAD_RWLOCK_INITIALIZER
+#endif
 #define __GTHREAD_ONCE_INIT PTHREAD_ONCE_INIT
 #if defined(PTHREAD_RECURSIVE_MUTEX_INITIALIZER)
 #define __GTHREAD_RECURSIVE_MUTEX_INIT PTHREAD_RECURSIVE_MUTEX_INITIALIZER
@@ -135,6 +141,13 @@ __gthrw(pthread_mutexattr_init)
 __gthrw(pthread_mutexattr_settype)
 __gthrw(pthread_mutexattr_destroy)
 
+#ifndef __cplusplus
+__gthrw(pthread_rwlock_rdlock)
+__gthrw(pthread_rwlock_tryrdlock)
+__gthrw(pthread_rwlock_wrlock)
+__gthrw(pthread_rwlock_trywrlock)
+__gthrw(pthread_rwlock_unlock)
+#endif
 
 #if defined(_LIBOBJC) || defined(_LIBOBJC_WEAK)
 /* Objective-C.  */
@@ -885,6 +898,53 @@ __gthread_cond_destroy (__gthread_cond_t* __cond)
   return __gthrw_(pthread_cond_destroy) (__cond);
 }
 
+#ifndef __cplusplus
+static inline int
+__gthread_rwlock_rdlock (__gthread_rwlock_t *__rwlock)
+{
+  if (__gthread_active_p ())
+    return __gthrw_(pthread_rwlock_rdlock) (__rwlock);
+  else
+    return 0;
+}
+
+static inline int
+__gthread_rwlock_tryrdlock (__gthread_rwlock_t *__rwlock)
+{
+  if (__gthread_active_p ())
+    return __gthrw_(pthread_rwlock_tryrdlock) (__rwlock);
+  else
+    return 0;
+}
+
+static inline int
+__gthread_rwlock_wrlock (__gthread_rwlock_t *__rwlock)
+{
+  if (__gthread_active_p ())
+    return __gthrw_(pthread_rwlock_wrlock) (__rwlock);
+  else
+    return 0;
+}
+
+static inline int
+__gthread_rwlock_trywrlock (__gthread_rwlock_t *__rwlock)
+{
+  if (__gthread_active_p ())
+    return __gthrw_(pthread_rwlock_trywrlock) (__rwlock);
+  else
+    return 0;
+}
+
+static inline int
+__gthread_rwlock_unlock (__gthread_rwlock_t *__rwlock)
+{
+  if (__gthread_active_p ())
+    return __gthrw_(pthread_rwlock_unlock) (__rwlock);
+  else
+    return 0;
+}
+#endif
+
 #endif /* _LIBOBJC */
 
 #endif /* ! GCC_GTHR_POSIX_H */
diff --git a/libgfortran/io/async.c b/libgfortran/io/async.c
index 81d1d8175aa..a97272ce5e4 100644
--- a/libgfortran/io/async.c
+++ b/libgfortran/io/async.c
@@ -42,6 +42,10 @@ DEBUG_LINE (__thread const char *aio_prefix = MPREFIX);
 
 DEBUG_LINE (__gthread_mutex_t debug_queue_lock = __GTHREAD_MUTEX_INIT;)
 DEBUG_LINE (aio_lock_debug *aio_debug_head = NULL;)
+#ifdef __GTHREAD_RWLOCK_INIT
+DEBUG_LINE (aio_rwlock_debug *aio_rwlock_debug_head = NULL;)
+DEBUG_LINE (__gthread_rwlock_t debug_queue_rwlock = __GTHREAD_RWLOCK_INIT;)
+#endif
 
 /* Current unit for asynchronous I/O.  Needed for error reporting.  */
 
diff --git a/libgfortran/io/async.h b/libgfortran/io/async.h
index ad226c8e856..0033cc74252 100644
--- a/libgfortran/io/async.h
+++ b/libgfortran/io/async.h
@@ -210,6 +210,128 @@
     DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", aio_prefix, #mutex, mutex); \
   } while (0)
 
+#ifdef __GTHREAD_RWLOCK_INIT
+#define RWLOCK_DEBUG_ADD(rwlock) do {		\
+    aio_rwlock_debug *n;				\
+    n = xmalloc (sizeof(aio_rwlock_debug));	\
+    n->prev = TAIL_RWLOCK_DEBUG_QUEUE;			\
+    if (n->prev)				\
+      n->prev->next = n;			\
+    n->next = NULL;				\
+    n->line = __LINE__;				\
+    n->func = __FUNCTION__;			\
+    n->rw = rwlock;				\
+    if (!aio_rwlock_debug_head) {			\
+      aio_rwlock_debug_head = n;			\
+    }						\
+  } while (0)
+
+#define CHECK_RDLOCK(rwlock, status) do {					\
+    aio_rwlock_debug *curr;						\
+    INTERN_WRLOCK (&debug_queue_rwlock);					\
+    if (__gthread_rwlock_tryrdlock (rwlock)) {				\
+      if ((curr = IN_RWLOCK_DEBUG_QUEUE (rwlock))) {				\
+	sprintf (status, DEBUG_RED "%s():%d" DEBUG_NORM, curr->func, curr->line); \
+      } else								\
+	sprintf (status, DEBUG_RED "unknown" DEBUG_NORM);			\
+    }									\
+    else {								\
+      __gthread_rwlock_unlock (rwlock);					\
+      sprintf (status, DEBUG_GREEN "rwunlocked" DEBUG_NORM);			\
+    }									\
+    INTERN_RWUNLOCK (&debug_queue_rwlock);					\
+  }while (0)
+
+#define CHECK_WRLOCK(rwlock, status) do {					\
+    aio_rwlock_debug *curr;						\
+    INTERN_WRLOCK (&debug_queue_rwlock);					\
+    if (__gthread_rwlock_trywrlock (rwlock)) {				\
+      if ((curr = IN_RWLOCK_DEBUG_QUEUE (rwlock))) {				\
+	sprintf (status, DEBUG_RED "%s():%d" DEBUG_NORM, curr->func, curr->line); \
+      } else								\
+	sprintf (status, DEBUG_RED "unknown" DEBUG_NORM);			\
+    }									\
+    else {								\
+      __gthread_rwlock_unlock (rwlock);					\
+      sprintf (status, DEBUG_GREEN "rwunlocked" DEBUG_NORM);			\
+    }									\
+    INTERN_RWUNLOCK (&debug_queue_rwlock);					\
+  }while (0)
+
+#define TAIL_RWLOCK_DEBUG_QUEUE ({			\
+      aio_rwlock_debug *curr = aio_rwlock_debug_head;	\
+      while (curr && curr->next) {		\
+	curr = curr->next;			\
+      }						\
+      curr;					\
+    })
+
+#define IN_RWLOCK_DEBUG_QUEUE(rwlock) ({		\
+      __label__ end;				\
+      aio_rwlock_debug *curr = aio_rwlock_debug_head;	\
+      while (curr) {				\
+	if (curr->rw == rwlock) {			\
+	  goto end;				\
+	}					\
+	curr = curr->next;			\
+      }						\
+    end:;					\
+      curr;					\
+    })
+
+#define RDLOCK(rwlock) do {						\
+    char status[200];							\
+    CHECK_RDLOCK (rwlock, status);						\
+    DEBUG_PRINTF ("%s%-42s prev: %-35s %20s():%-5d %18p\n", aio_prefix,	\
+		 DEBUG_RED "RDLOCK: " DEBUG_NORM #rwlock, status, __FUNCTION__, __LINE__, (void *) rwlock); \
+    INTERN_RDLOCK (rwlock);							\
+    INTERN_WRLOCK (&debug_queue_rwlock);					\
+    RWLOCK_DEBUG_ADD (rwlock);						\
+    INTERN_RWUNLOCK (&debug_queue_rwlock);					\
+    DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", aio_prefix, #rwlock, rwlock); \
+  } while (0)
+
+#define WRLOCK(rwlock) do {						\
+    char status[200];							\
+    CHECK_WRLOCK (rwlock, status);						\
+    DEBUG_PRINTF ("%s%-42s prev: %-35s %20s():%-5d %18p\n", aio_prefix,	\
+		 DEBUG_RED "WRLOCK: " DEBUG_NORM #rwlock, status, __FUNCTION__, __LINE__, (void *) rwlock); \
+    INTERN_WRLOCK (rwlock);							\
+    INTERN_WRLOCK (&debug_queue_rwlock);					\
+    RWLOCK_DEBUG_ADD (rwlock);						\
+    INTERN_RWUNLOCK (&debug_queue_rwlock);					\
+    DEBUG_PRINTF ("%s" DEBUG_RED "ACQ:" DEBUG_NORM " %-30s %78p\n", aio_prefix, #rwlock, rwlock); \
+  } while (0)
+
+#define RWUNLOCK(rwlock) do {						\
+    aio_rwlock_debug *curr;						\
+    DEBUG_PRINTF ("%s%-75s %20s():%-5d %18p\n", aio_prefix, DEBUG_GREEN "RWUNLOCK: " DEBUG_NORM #rwlock, \
+		 __FUNCTION__, __LINE__, (void *) rwlock);		\
+    INTERN_WRLOCK (&debug_queue_rwlock);					\
+    curr = IN_RWLOCK_DEBUG_QUEUE (rwlock);					\
+    if (curr)								\
+      {									\
+	if (curr->prev)							\
+	  curr->prev->next = curr->next;				\
+	if (curr->next) {						\
+	  curr->next->prev = curr->prev;				\
+	  if (curr == aio_rwlock_debug_head)					\
+	    aio_rwlock_debug_head = curr->next;				\
+	} else {							\
+	  if (curr == aio_rwlock_debug_head)					\
+	    aio_rwlock_debug_head = NULL;					\
+	}								\
+	free (curr);							\
+      }									\
+    INTERN_RWUNLOCK (&debug_queue_rwlock);					\
+    INTERN_RWUNLOCK (rwlock);						\
+  } while (0)
+
+#define RD_TO_WRLOCK(rwlock) \
+  RWUNLOCK (rwlock);\
+  WRLOCK (rwlock);
+#endif
+
 #define DEBUG_LINE(...) __VA_ARGS__
 
 #else
@@ -221,12 +343,31 @@
 #define LOCK(mutex) INTERN_LOCK (mutex)
 #define UNLOCK(mutex) INTERN_UNLOCK (mutex)
 #define TRYLOCK(mutex) (__gthread_mutex_trylock (mutex))
+#ifdef __GTHREAD_RWLOCK_INIT
+#define RDLOCK(rwlock) INTERN_RDLOCK (rwlock)
+#define WRLOCK(rwlock) INTERN_WRLOCK (rwlock)
+#define RWUNLOCK(rwlock) INTERN_RWUNLOCK (rwlock)
+#define RD_TO_WRLOCK(rwlock) \
+  RWUNLOCK (rwlock);\
+  WRLOCK (rwlock);
+#endif
+#endif
+
+#ifndef __GTHREAD_RWLOCK_INIT
+#define RDLOCK(rwlock) LOCK (rwlock)
+#define WRLOCK(rwlock) LOCK (rwlock)
+#define RWUNLOCK(rwlock) UNLOCK (rwlock)
+#define RD_TO_WRLOCK(rwlock) {}
 #endif
 
 #define INTERN_LOCK(mutex) T_ERROR (__gthread_mutex_lock, mutex);
 
 #define INTERN_UNLOCK(mutex) T_ERROR (__gthread_mutex_unlock, mutex);
 
+#define INTERN_RDLOCK(rwlock) T_ERROR (__gthread_rwlock_rdlock, rwlock);
+#define INTERN_WRLOCK(rwlock) T_ERROR (__gthread_rwlock_wrlock, rwlock);
+#define INTERN_RWUNLOCK(rwlock) T_ERROR (__gthread_rwlock_unlock, rwlock);
+
 #if ASYNC_IO
 
 /* au->lock has to be held when calling this macro.  */
@@ -288,8 +429,18 @@ DEBUG_LINE (typedef struct aio_lock_debug{
   struct aio_lock_debug *prev;
 } aio_lock_debug;)
 
+DEBUG_LINE (typedef struct aio_rwlock_debug{
+  __gthread_rwlock_t *rw;
+  int line;
+  const char *func;
+  struct aio_rwlock_debug *next;
+  struct aio_rwlock_debug *prev;
+} aio_rwlock_debug;)
+
 DEBUG_LINE (extern aio_lock_debug *aio_debug_head;)
 DEBUG_LINE (extern __gthread_mutex_t debug_queue_lock;)
+DEBUG_LINE (extern aio_rwlock_debug *aio_rwlock_debug_head;)
+DEBUG_LINE (extern __gthread_rwlock_t debug_queue_rwlock;)
 
 /* Thread - local storage of the current unit we are looking at. Needed for
    error reporting.  */
diff --git a/libgfortran/io/io.h b/libgfortran/io/io.h
index ecdf1dd3f05..15daa0995b1 100644
--- a/libgfortran/io/io.h
+++ b/libgfortran/io/io.h
@@ -690,7 +690,7 @@ typedef struct gfc_unit
      from the UNIT_ROOT tree, but doesn't free it and the
      last of the waiting threads will do that.
      This must be either atomically increased/decreased, or
-     always guarded by UNIT_LOCK.  */
+     always guarded by UNIT_RWLOCK.  */
   int waiting;
   /* Flag set by close_unit if the unit as been closed.
      Must be manipulated under unit's lock.  */
@@ -769,8 +769,13 @@ internal_proto(default_recl);
 extern gfc_unit *unit_root;
 internal_proto(unit_root);
 
-extern __gthread_mutex_t unit_lock;
-internal_proto(unit_lock);
+#ifdef __GTHREAD_RWLOCK_INIT
+extern __gthread_rwlock_t unit_rwlock;
+internal_proto(unit_rwlock);
+#else
+extern __gthread_mutex_t unit_rwlock;
+internal_proto(unit_rwlock);
+#endif
 
 extern int close_unit (gfc_unit *);
 internal_proto(close_unit);
@@ -1015,9 +1020,9 @@ dec_waiting_unlocked (gfc_unit *u)
 #ifdef HAVE_ATOMIC_FETCH_ADD
   (void) __atomic_fetch_add (&u->waiting, -1, __ATOMIC_RELAXED);
 #else
-  __gthread_mutex_lock (&unit_lock);
+  WRLOCK (&unit_rwlock);
   u->waiting--;
-  __gthread_mutex_unlock (&unit_lock);
+  RWUNLOCK (&unit_rwlock);
 #endif
 }
 
diff --git a/libgfortran/io/transfer.c b/libgfortran/io/transfer.c
index 8bb5d1101ca..b01f45b80f6 100644
--- a/libgfortran/io/transfer.c
+++ b/libgfortran/io/transfer.c
@@ -4539,9 +4539,9 @@ st_read_done_worker (st_parameter_dt *dtp, bool unlock)
    if (free_newunit)
      {
        /* Avoid inverse lock issues by placing after unlock_unit.  */
-       LOCK (&unit_lock);
+       WRLOCK (&unit_rwlock);
        newunit_free (dtp->common.unit);
-       UNLOCK (&unit_lock);
+       RWUNLOCK (&unit_rwlock);
      }
 }
 
@@ -4636,9 +4636,9 @@ st_write_done_worker (st_parameter_dt *dtp, bool unlock)
    if (free_newunit)
      {
        /* Avoid inverse lock issues by placing after unlock_unit.  */
-       LOCK (&unit_lock);
+       WRLOCK (&unit_rwlock);
        newunit_free (dtp->common.unit);
-       UNLOCK (&unit_lock);
+       RWUNLOCK (&unit_rwlock);
      }
 }
 
diff --git a/libgfortran/io/unit.c b/libgfortran/io/unit.c
index 82664dc5f98..62f1db21d34 100644
--- a/libgfortran/io/unit.c
+++ b/libgfortran/io/unit.c
@@ -33,34 +33,36 @@ see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
 
 
 /* IO locking rules:
-   UNIT_LOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
+   UNIT_RWLOCK is a master lock, protecting UNIT_ROOT tree and UNIT_CACHE.
+   And use the rwlock to spilt read and write phase to UNIT_ROOT tree
+   and UNIT_CACHE to increase CPU efficiency.
    Concurrent use of different units should be supported, so
    each unit has its own lock, LOCK.
    Open should be atomic with its reopening of units and list_read.c
    in several places needs find_unit another unit while holding stdin
-   unit's lock, so it must be possible to acquire UNIT_LOCK while holding
+   unit's lock, so it must be possible to acquire UNIT_RWLOCK while holding
    some unit's lock.  Therefore to avoid deadlocks, it is forbidden
-   to acquire unit's private locks while holding UNIT_LOCK, except
+   to acquire unit's private locks while holding UNIT_RWLOCK, except
    for freshly created units (where no other thread can get at their
    address yet) or when using just trylock rather than lock operation.
    In addition to unit's private lock each unit has a WAITERS counter
    and CLOSED flag.  WAITERS counter must be either only
    atomically incremented/decremented in all places (if atomic builtins
-   are supported), or protected by UNIT_LOCK in all places (otherwise).
+   are supported), or protected by UNIT_RWLOCK in all places (otherwise).
    CLOSED flag must be always protected by unit's LOCK.
-   After finding a unit in UNIT_CACHE or UNIT_ROOT with UNIT_LOCK held,
+   After finding a unit in UNIT_CACHE or UNIT_ROOT with UNIT_RWLOCK held,
    WAITERS must be incremented to avoid concurrent close from freeing
-   the unit between unlocking UNIT_LOCK and acquiring unit's LOCK.
-   Unit freeing is always done under UNIT_LOCK.  If close_unit sees any
+   the unit between unlocking UNIT_RWLOCK and acquiring unit's LOCK.
+   Unit freeing is always done under UNIT_RWLOCK.  If close_unit sees any
    WAITERS, it doesn't free the unit but instead sets the CLOSED flag
    and the thread that decrements WAITERS to zero while CLOSED flag is
-   set is responsible for freeing it (while holding UNIT_LOCK).
+   set is responsible for freeing it (while holding UNIT_RWLOCK).
    flush_all_units operation is iterating over the unit tree with
-   increasing UNIT_NUMBER while holding UNIT_LOCK and attempting to
+   increasing UNIT_NUMBER while holding UNIT_RWLOCK and attempting to
    flush each unit (and therefore needs the unit's LOCK held as well).
    To avoid deadlocks, it just trylocks the LOCK and if unsuccessful,
-   remembers the current unit's UNIT_NUMBER, unlocks UNIT_LOCK, acquires
-   unit's LOCK and after flushing reacquires UNIT_LOCK and restarts with
+   remembers the current unit's UNIT_NUMBER, unlocks UNIT_RWLOCK, acquires
+   unit's LOCK and after flushing reacquires UNIT_RWLOCK and restarts with
    the smallest UNIT_NUMBER above the last one flushed.
 
    If find_unit/find_or_create_unit/find_file/get_unit routines return
@@ -101,10 +103,14 @@ gfc_offset max_offset;
 gfc_offset default_recl;
 
 gfc_unit *unit_root;
+#ifdef __GTHREAD_RWLOCK_INIT
+__gthread_rwlock_t unit_rwlock = __GTHREAD_RWLOCK_INIT;
+#else
 #ifdef __GTHREAD_MUTEX_INIT
-__gthread_mutex_t unit_lock = __GTHREAD_MUTEX_INIT;
+__gthread_mutex_t unit_rwlock = __GTHREAD_MUTEX_INIT;
 #else
-__gthread_mutex_t unit_lock;
+__gthread_mutex_t unit_rwlock;
+#endif
 #endif
 
 /* We use these filenames for error reporting.  */
@@ -329,7 +335,7 @@ get_gfc_unit (int n, int do_create)
   int c, created = 0;
 
   NOTE ("Unit n=%d, do_create = %d", n, do_create);
-  LOCK (&unit_lock);
+  RDLOCK (&unit_rwlock);
 
 retry:
   for (c = 0; c < CACHE_SIZE; c++)
@@ -350,6 +356,17 @@ retry:
       if (c == 0)
 	break;
     }
+  /* We did not find a unit in the cache nor in the unit list, create a new
+    (locked) unit and insert into the unit list and cache.
+    Manipulating either or both the unit list and the unit cache requires to
+    hold a write-lock [for obvious reasons]:
+    1. By separating the read/write lock, it will greatly reduce the contention
+       at the read part, while write part is not always necessary or most
+       unlikely once the unit hit in cache.
+    2. We try to balance the implementation complexity and the performance
+       gains that fit into current cases we observed by just using a
+       pthread_rwlock. */
+  RD_TO_WRLOCK (&unit_rwlock);
 
   if (p == NULL && do_create)
     {
@@ -368,8 +385,8 @@ retry:
   if (created)
     {
       /* Newly created units have their lock held already
-	 from insert_unit.  Just unlock UNIT_LOCK and return.  */
-      UNLOCK (&unit_lock);
+	 from insert_unit.  Just unlock UNIT_RWLOCK and return.  */
+      RWUNLOCK (&unit_rwlock);
       return p;
     }
 
@@ -380,7 +397,7 @@ found:
       if (! TRYLOCK (&p->lock))
 	{
 	  /* assert (p->closed == 0); */
-	  UNLOCK (&unit_lock);
+	  RWUNLOCK (&unit_rwlock);
 	  return p;
 	}
 
@@ -388,14 +405,14 @@ found:
     }
 
 
-  UNLOCK (&unit_lock);
+  RWUNLOCK (&unit_rwlock);
 
   if (p != NULL && (p->child_dtio == 0))
     {
       LOCK (&p->lock);
       if (p->closed)
 	{
-	  LOCK (&unit_lock);
+	  WRLOCK (&unit_rwlock);
 	  UNLOCK (&p->lock);
 	  if (predec_waiting_locked (p) == 0)
 	    destroy_unit_mutex (p);
@@ -593,8 +610,8 @@ init_units (void)
 #endif
 #endif
 
-#ifndef __GTHREAD_MUTEX_INIT
-  __GTHREAD_MUTEX_INIT_FUNCTION (&unit_lock);
+#if (!defined(__GTHREAD_RWLOCK_INIT) && !defined(__GTHREAD_MUTEX_INIT))
+  __GTHREAD_MUTEX_INIT_FUNCTION (&unit_rwlock);
 #endif
 
   if (sizeof (max_offset) == 8)
@@ -731,7 +748,7 @@ close_unit_1 (gfc_unit *u, int locked)
 
   u->closed = 1;
   if (!locked)
-    LOCK (&unit_lock);
+    WRLOCK (&unit_rwlock);
 
   for (i = 0; i < CACHE_SIZE; i++)
     if (unit_cache[i] == u)
@@ -758,7 +775,7 @@ close_unit_1 (gfc_unit *u, int locked)
     destroy_unit_mutex (u);
 
   if (!locked)
-    UNLOCK (&unit_lock);
+    RWUNLOCK (&unit_rwlock);
 
   return rc;
 }
@@ -795,10 +812,10 @@ close_unit (gfc_unit *u)
 void
 close_units (void)
 {
-  LOCK (&unit_lock);
+  WRLOCK (&unit_rwlock);
   while (unit_root != NULL)
     close_unit_1 (unit_root, 1);
-  UNLOCK (&unit_lock);
+  RWUNLOCK (&unit_rwlock);
 
   free (newunits);
 
@@ -905,7 +922,7 @@ finish_last_advance_record (gfc_unit *u)
 int
 newunit_alloc (void)
 {
-  LOCK (&unit_lock);
+  WRLOCK (&unit_rwlock);
   if (!newunits)
     {
       newunits = xcalloc (16, 1);
@@ -919,7 +936,7 @@ newunit_alloc (void)
         {
           newunits[ii] = true;
           newunit_lwi = ii + 1;
-	  UNLOCK (&unit_lock);
+	  RWUNLOCK (&unit_rwlock);
           return -ii + NEWUNIT_START;
         }
     }
@@ -932,12 +949,12 @@ newunit_alloc (void)
   memset (newunits + old_size, 0, old_size);
   newunits[old_size] = true;
   newunit_lwi = old_size + 1;
-    UNLOCK (&unit_lock);
+    RWUNLOCK (&unit_rwlock);
   return -old_size + NEWUNIT_START;
 }
 
 
-/* Free a previously allocated newunit= unit number.  unit_lock must
+/* Free a previously allocated newunit= unit number.  unit_rwlock must
    be held when calling.  */
 
 void
diff --git a/libgfortran/io/unix.c b/libgfortran/io/unix.c
index ba12be08252..d53003919ab 100644
--- a/libgfortran/io/unix.c
+++ b/libgfortran/io/unix.c
@@ -1774,7 +1774,7 @@ find_file (const char *file, gfc_charlen_type file_len)
   id = id_from_path (path);
 #endif
 
-  LOCK (&unit_lock);
+  RDLOCK (&unit_rwlock);
 retry:
   u = find_file0 (unit_root, FIND_FILE0_ARGS);
   if (u != NULL)
@@ -1783,19 +1783,19 @@ retry:
       if (! __gthread_mutex_trylock (&u->lock))
 	{
 	  /* assert (u->closed == 0); */
-	  UNLOCK (&unit_lock);
+	  RWUNLOCK (&unit_rwlock);
 	  goto done;
 	}
 
       inc_waiting_locked (u);
     }
-  UNLOCK (&unit_lock);
+  RWUNLOCK (&unit_rwlock);
   if (u != NULL)
     {
       LOCK (&u->lock);
       if (u->closed)
 	{
-	  LOCK (&unit_lock);
+	  RDLOCK (&unit_rwlock);
 	  UNLOCK (&u->lock);
 	  if (predec_waiting_locked (u) == 0)
 	    free (u);
@@ -1839,13 +1839,13 @@ flush_all_units (void)
   gfc_unit *u;
   int min_unit = 0;
 
-  LOCK (&unit_lock);
+  WRLOCK (&unit_rwlock);
   do
     {
       u = flush_all_units_1 (unit_root, min_unit);
       if (u != NULL)
 	inc_waiting_locked (u);
-      UNLOCK (&unit_lock);
+      RWUNLOCK (&unit_rwlock);
       if (u == NULL)
 	return;
 
@@ -1856,13 +1856,13 @@ flush_all_units (void)
       if (u->closed == 0)
 	{
 	  sflush (u->s);
-	  LOCK (&unit_lock);
+	  WRLOCK (&unit_rwlock);
 	  UNLOCK (&u->lock);
 	  (void) predec_waiting_locked (u);
 	}
       else
 	{
-	  LOCK (&unit_lock);
+	  WRLOCK (&unit_rwlock);
 	  UNLOCK (&u->lock);
 	  if (predec_waiting_locked (u) == 0)
 	    free (u);
-- 
2.39.1