Date: Fri, 29 Aug 2014 17:31:21 +0000 From: "Bjoern A. Zeeb" <bz@FreeBSD.org> To: Konstantin Belousov <kostikbel@gmail.com> Cc: svn-src-head@freebsd.org, svn-src-all@freebsd.org, src-committers@freebsd.org Subject: Re: svn commit: r270806 - head/sys/dev/ixl Message-ID: <DBA55B1E-8381-47D9-A646-B8ADBAAB6073@FreeBSD.org> In-Reply-To: <20140829171452.GO2737@kib.kiev.ua> References: <201408291240.s7TCe1OQ029986@svn.freebsd.org> <20140829171452.GO2737@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On 29 Aug 2014, at 17:14 , Konstantin Belousov <kostikbel@gmail.com> = wrote: > On Fri, Aug 29, 2014 at 12:40:01PM +0000, Bjoern A. Zeeb wrote: >> Author: bz >> Date: Fri Aug 29 12:40:01 2014 >> New Revision: 270806 >> URL: http://svnweb.freebsd.org/changeset/base/270806 >>=20 >> Log: >> Properly handle prefetch only for amd64 and i386 as we do elsewhere. >>=20 >> In general theraven is right that we should factr this out and = provide >> a general and per-arch implementation that everything can use. >>=20 >> MFC after: 3 days >> X-MFC with: r270755 >>=20 >> Modified: >> head/sys/dev/ixl/i40e_osdep.h >>=20 >> Modified: head/sys/dev/ixl/i40e_osdep.h >> = =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D= =3D=3D=3D >> --- head/sys/dev/ixl/i40e_osdep.h Fri Aug 29 11:18:54 2014 = (r270805) >> +++ head/sys/dev/ixl/i40e_osdep.h Fri Aug 29 12:40:01 2014 = (r270806) >> @@ -137,11 +137,15 @@ struct i40e_spinlock { >>=20 >> #define le16_to_cpu=20 >>=20 >> +#if defined(__amd64__) || defined(i386) >> static __inline >> void prefetch(void *x) >> { >> __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); >> } > This only fix the build failure, but the code is still somewhat wrong. >=20 > Availability of the prefetch instruction depends on the presence of = SSE. We have quite a few of these as-are in the tree already; I guess they = all need fixing. > Althought it is probably impossible to find a machine where the = intended > hardware can operate and which does not support SSE _currently_, I am > not sure that it is wise to hard-code SSE instructions in the i386 > kernel. >=20 > Might be, a change of the compilation test from both amd64 and i386 to > just amd64 is due. I doubt that anybody would use 40Gb with 32bit OS, > and even if somebody does, that the performace is at the stake in > this case. >=20 I think going with David=92s suggestion to do the right thing and = centralise them somewhere so everything can use it is the proper way = forward. I think there=92s also a __builtin_prefetch(). >> +#else >> +#define prefetch(x) >> +#endif >>=20 >> struct i40e_osdep >> { =97=20 Bjoern A. Zeeb "Come on. Learn, goddamn it.", WarGames, 1983
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?DBA55B1E-8381-47D9-A646-B8ADBAAB6073>