On Fri, Jun 15, 2018 at 3:41 AM Herve Jourdain <[email protected]> wrote:
>
> Hi,
>
> Actually, I meant "works" in the sense of "does compile" - as opposed to 
> armv8 where it does not compile, which is why we're having this discussion in 
> the first place.
> So I was merely suggesting to not modify previous oe behavior for the db 
> package for previous architectures, and just add the removal of 'swp' for 
> armv8, where it matters most.
>
> If we want to look at it in details, based on the "ARM Architecture Reference 
> Manual ARMv7-A and ARMv7-R edition":
> 1. SWP is the way to go before ARMv6.
> 2. SWP has been deprecated in ARMv6.
> 3. SWP has been deprecated AND made optional in ARMv7ve.
> 4. "The SWP and SWPB instructions rely on the properties of the system beyond 
> the processor to ensure that no stores from other observers can occur between 
> the load access and the store access, and this might not be implemented for 
> all regions of memory on some system implementations. In all cases, SWP and 
> SWPB do ensure that no stores from the processor that executed the SWP or 
> SWPB instruction can occur between the load access and the store access of 
> the SWP or SWPB."
>
> This latest part means that it may or may not work in SMP environments, it 
> depends on how the system is architecture around the cores - most likely how 
> the bus system is designed I believe. So it may actually be working fine if 
> the system/bus designer has taken that into account.
>
> This said, I believe that from point #3 above, it might make sense to disable 
> SWP for armv7ve as well, since being optional means that it might be compiled 
> correctly, but still fail at runtime, depending on the choices of the SoC 
> manufacturer.
> So my recommendation would be to add:
> MUTEX_armv7ve = ""
> MUTEX_armv8 = ""
>
> To disable 'swp' by default only for those 2 archs, while keeping things like 
> they are for previous architectures.
>

the least intrusive fix it to fix what we need and let the defaults be
as it in this case.

> Cheers,
> Herve
>
> -----Original Message-----
> From: Andre McCurdy [mailto:[email protected]]
> Sent: vendredi 15 juin 2018 09:39
> To: Herve Jourdain <[email protected]>
> Cc: Khem Raj <[email protected]>; Ovidiu Panait 
> <[email protected]>; Patches and discussions about the oe-core 
> layer <[email protected]>
> Subject: Re: [OE-core] [PATCH 1/1] db: disable the ARM assembler mutex code
>
> On Fri, Jun 15, 2018 at 12:10 AM, Herve Jourdain <[email protected]> 
> wrote:
> > Hi,
> >
> > So the issue is whether we want to change the behaviour of previous 
> > architectures, or if we try to fix the issue only for the architectures 
> > that don't work.
> > Until now, the db recipe was enabling the 'swp' optimization, and that 
> > behavior could be disabled on .bbappend if needed.
> > While that works fine until armv7ve, it does not work for armv8, which has 
> > removed support for those instructions.
>
> I don't know if "works fine until armv7ve" is correct. Although the swp 
> instruction exists for armv7, according to the link I posted yesterday, it is 
> not guaranteed to work.
>
>   
> https://community.arm.com/processors/b/blog/posts/locks-swps-and-two-smoking-barriers
>
> Or do you have other evidence to suggest that swp is safe to use for armv7?
>
> > Therefore, there is a need to fix it for armv8 - and armv8 only - whereas 
> > it can be safely used on previous architectures.
> > If we remove the use for all ARM architectures, that might create some 
> > regression/issues.
> > If we just remove the use of 'swp' only for armv8, we ensure it doesn't 
> > break anything that's running on previous ARM architectures.
> >
> > Cheers,
> > Herve
> >
> > -----Original Message-----
> > From: Andre McCurdy [mailto:[email protected]]
> > Sent: vendredi 15 juin 2018 00:03
> > To: Khem Raj <[email protected]>
> > Cc: Herve Jourdain <[email protected]>; Ovidiu Panait
> > <[email protected]>; Patches and discussions about the
> > oe-core layer <[email protected]>
> > Subject: Re: [OE-core] [PATCH 1/1] db: disable the ARM assembler mutex
> > code
> >
> > On Thu, Jun 14, 2018 at 2:48 PM, Khem Raj <[email protected]> wrote:
> >> On Thu, Jun 14, 2018 at 1:01 PM Andre McCurdy <[email protected]> wrote:
> >>> On Thu, Jun 14, 2018 at 12:24 PM, Khem Raj <[email protected]> wrote:
> >>> > On Thu, Jun 14, 2018 at 12:12 PM Andre McCurdy
> >>> > <[email protected]>
> >>> > wrote:
> >>> >>
> >>> >> On Thu, Jun 14, 2018 at 9:40 AM, Khem Raj <[email protected]> wrote:
> >>> >> > On 6/14/18 5:10 AM, Herve Jourdain wrote:
> >>> >> >> Hi,
> >>> >> >>
> >>> >> >> I believe I solved that same problem by just adding, in the
> >>> >> >> case of
> >>> >> >> armv8
> >>> >> >> (which I believe may be the new architecture you're referring to):
> >>> >> >> MUTEX_armv8 = ""
> >>> >> >> This way, it allows previous versions to work just like they
> >>> >> >> did before, without having to disable ARM assembler mutex code
> >>> >> >> for architectures that support it correctly - up to armv7ve I
> >>> >> >> believe.
> >>> >> >> Of course, we might need to also have a good definition for
> >>> >> >> armv8, which is the object of another thread.
> >>> >> >
> >>> >> > right thats a better approach.
> >>> >>
> >>> >> SWP is not guaranteed to work on SMP systems... and even if it
> >>> >> does, performance is likely to be worse than the pthreads version
> >>> >> (which can take advantage of the newer instructions).
> >>> >>
> >>> >>
> >>> >> https://community.arm.com/processors/b/blog/posts/locks-swps-and-
> >>> >> t
> >>> >> wo-smoking-barriers
> >>> >>
> >>> >> In general, use of hand optimised assembler mutex implementations
> >>> >> in user space isn't something to be encouraged - use pthreads (or
> >>> >> maybe a gcc intrinsic) instead.
> >>> >>
> >>> >
> >>> > question is about disabling it on old arm machines, do we have
> >>> > data where we know that other sync methods without swp works
> >>> > better on
> >>> > armv5 and lower ?
> >>>
> >>> On armv5 and below a hand optimised implementation using SWP is
> >>> likely to be faster than pthreads. No one is suggesting otherwise.
> >>>
> >>> On SMP (highly likely nowadays for armv7 and above), SWP simply
> >>> might not work (aside from the fact that if it does work, it's
> >>> likely to be slower than pthreads). It's not really a question of
> >>> performance there, so the proposal to only disable SWP for armv8
> >>> doesn't seem like a safe solution.
> >>
> >> Suggestion is not to just do it for armv8 but To keep it there where
> >> its beneficial
> >
> > You can always argue that micro optimisations are beneficial. The question 
> > is whether they make a big enough difference in some real world use case to 
> > be worth the maintenance effort.
> >
> >>> Using pthreads unconditionally is safe and easy. Unless you can
> >>> prove that hand optimised SWP is really a big win for armv5 (is
> >>> anyone really running a performance critical database on an armv5
> >>> system?) why not keep the recipe simple and use pthreads everywhere?
> >>>
> >>> >> I think the original patch is good.
> >>> >>
> >>> >> >> Cheers,
> >>> >> >> Herve
> >>> >> >>
> >>> >> >> -----Original Message-----
> >>> >> >> From: [email protected]
> >>> >> >> [mailto:[email protected]] On
> >>> >> >> Behalf Of Ovidiu Panait
> >>> >> >> Sent: jeudi 14 juin 2018 13:55
> >>> >> >> To: [email protected]
> >>> >> >> Subject: [OE-core] [PATCH 1/1] db: disable the ARM assembler
> >>> >> >> mutex code
> >>> >> >>
> >>> >> >> The swpb in macro MUTEX_SET will cause "undefined instruction"
> >>> >> >> error on the new arm arches which don't support this assembly
> >>> >> >> instruction any more. If use ldrex/strex to replace swpb, the
> >>> >> >> old arm arches don't support them. So to avoid this issue,
> >>> >> >> just disable the ARM assembler mutex code, and use the default
> >>> >> >> pthreads mutex.
> >>> >> >>
> >>> >> >> Signed-off-by: Li Zhou <[email protected]>
> >>> >> >> Signed-off-by: Catalin Enache <[email protected]>
> >>> >> >> Signed-off-by: Ovidiu Panait <[email protected]>
> >>> >> >> ---
> >>> >> >>  meta/recipes-support/db/db_5.3.28.bb | 13 +------------
> >>> >> >>  1 file changed, 1 insertion(+), 12 deletions(-)
> >>> >> >>
> >>> >> >> diff --git a/meta/recipes-support/db/db_5.3.28.bb
> >>> >> >> b/meta/recipes-support/db/db_5.3.28.bb
> >>> >> >> index 093ee44909..15b4155a29 100644
> >>> >> >> --- a/meta/recipes-support/db/db_5.3.28.bb
> >>> >> >> +++ b/meta/recipes-support/db/db_5.3.28.bb
> >>> >> >> @@ -59,18 +59,7 @@ FILES_SOLIBSDEV = "${libdir}/libdb.so
> >>> >> >> ${libdir}/libdb_cxx.so"
> >>> >> >>  # All the --disable-* options replace --enable-smallbuild,
> >>> >> >> which breaks a bunch of stuff (eg. postfix)  DB5_CONFIG ?=
> >>> >> >> "--enable-o_direct --disable-cryptography --disable-queue
> >>> >> >> --disable-replication --disable-verify --disable-compat185
> >>> >> >> --disable-sql"
> >>> >> >>
> >>> >> >> -EXTRA_OECONF = "${DB5_CONFIG} --enable-shared --enable-cxx
> >>> >> >> --with-sysroot"
> >>> >> >> -
> >>> >> >> -# Override the MUTEX setting here, the POSIX library is -#
> >>> >> >> the default - "POSIX/pthreads/library".
> >>> >> >> -# Don't ignore the nice SWP instruction on the ARM:
> >>> >> >> -# These enable the ARM assembler mutex code, this won't -#
> >>> >> >> work with thumb compilation...
> >>> >> >> -ARM_MUTEX = "--with-mutex=ARM/gcc-assembly"
> >>> >> >> -MUTEX = ""
> >>> >> >> -MUTEX_arm = "${ARM_MUTEX}"
> >>> >> >> -MUTEX_armeb = "${ARM_MUTEX}"
> >>> >> >> -EXTRA_OECONF += "${MUTEX} STRIP=true"
> >>> >> >> +EXTRA_OECONF = "${DB5_CONFIG} --enable-shared --enable-cxx
> >>> >> >> --with-sysroot
> >>> >> >> STRIP=true"
> >>> >> >>  EXTRA_OEMAKE += "LIBTOOL='./${HOST_SYS}-libtool'"
> >>> >> >>
> >>> >> >>  EXTRA_AUTORECONF += "--exclude=autoheader  -I
> >>> >> >> ${S}/dist/aclocal -I${S}/dist/aclocal_java"
> >>> >> >> --
> >>> >> >> 2.17.1
> >>> >> >>
> >>> >> >> --
> >>> >> >> _______________________________________________
> >>> >> >> Openembedded-core mailing list
> >>> >> >> [email protected]
> >>> >> >> http://lists.openembedded.org/mailman/listinfo/openembedded-co
> >>> >> >> r
> >>> >> >> e
> >>> >> >>
> >>> >> >
> >>> >> >
> >>> >> >
> >>> >> > --
> >>> >> > _______________________________________________
> >>> >> > Openembedded-core mailing list
> >>> >> > [email protected]
> >>> >> > http://lists.openembedded.org/mailman/listinfo/openembedded-cor
> >>> >> > e
> >>> >> >
> >
>
-- 
_______________________________________________
Openembedded-core mailing list
[email protected]
http://lists.openembedded.org/mailman/listinfo/openembedded-core

Reply via email to