2014-12-08 06:37, Neil Horman:
> On Mon, Dec 08, 2014 at 03:37:19AM +0000, Qiu, Michael wrote:
> > On 12/8/2014 11:00 AM, Neil Horman wrote:
> > > On Mon, Dec 08, 2014 at 02:46:51AM +0000, Qiu, Michael wrote:
> > >> On 12/5/2014 11:25 PM, Neil Horman wrote:
> > >>> On Fri, Dec 05, 2014 at 03:02:33PM +0000, Bruce Richardson wrote:
> > >>>> On Fri, Dec 05, 2014 at 09:22:05AM -0500, Neil Horman wrote:
> > >>>>> On Fri, Dec 05, 2014 at 04:31:47PM +0800, Chao Zhu wrote:
> > >>>>>> On 2014/12/4 17:12, Michael Qiu wrote:
> > >>>>>>> lib/librte_eal/linuxapp/eal/eal_memory.c:324:4: error: comparison
> > >>>>>>> is always false due to limited range of data type
> > >>>>>>> [-Werror=type-limits]
> > >>>>>>> || (hugepage_sz == RTE_PGSIZE_16G)) {
> > >>>>>>> ^
> > >>>>>>> cc1: all warnings being treated as errors
> > >>>>>>>
> > >>>>>>> lib/librte_eal/linuxapp/eal/eal.c(461): error #2259: non-pointer
> > >>>>>>> conversion from "long long" to "void *" may lose significant bits
> > >>>>>>> RTE_PTR_ALIGN_CEIL((uintptr_t)addr, RTE_PGSIZE_16M);
> > >>>>>>>
> > >>>>>>> This was introuduced by commit b77b5639:
> > >>>>>>> mem: add huge page sizes for IBM Power
> > >>>>>>>
> > >>>>>>> The root cause is that size_t and uintptr_t are 32-bit in i686
> > >>>>>>> platform, but RTE_PGSIZE_16M and RTE_PGSIZE_16G are always 64-bit.
> > >>>>>>>
> > >>>>>>> Define RTE_PGSIZE_16G only in 64 bit platform to avoid
> > >>>>>>> this issue.
> > >>>>>>>
> > >>>>>>> Signed-off-by: Michael Qiu <michael.qiu at intel.com>
> > >>>>>>> ---
> > >>>>>>> v3 ---> v2
> > >>>>>>> Change RTE_PGSIZE_16G from ULL to UL
> > >>>>>>> to keep all entries consistent
> > >>>>>>>
> > >>>>>>> V2 ---> v1
> > >>>>>>> Change two type entries to one, and
> > >>>>>>> leave RTE_PGSIZE_16G only valid for
> > >>>>>>> 64-bit platform
> > >>>>>>>
> > >>>>> NACK, this is the wrong way to fix this problem. Pagesizes are
> > >>>>> independent of
> > >>>>> architecutre. While a system can't have a hugepage size that exceeds
> > >>>>> its
> > >>>>> virtual address limit, theres no need to do per-architecture special
> > >>>>> casing of
> > >>>>> page sizes here. Instead of littering the code with ifdef RTE_ARCH_64
> > >>>>> everytime you want to check a page size, just convert the size_t to a
> > >>>>> uint64_t
> > >>>>> and you can allow all of the enumerated page types on all
> > >>>>> architecutres, and
> > >>>>> save yourself some ifdeffing in the process.
> > >>>>>
> > >>>>> Neil
> > >>>> While I get your point, I find it distasteful to use a uint64_t for
> > >>>> memory sizes,
> > >>>> when there is the size_t type defined for that particular purpose.
> > >>>> However, I suppose that reducing the number of #ifdefs compared to
> > >>>> using the
> > >>>> "correct" datatypes for objects is a more practical optino - however
> > >>>> distastful
> > >>>> I find it.
> > >>> size_t isn't defined for memory sizes in the sense that we're using
> > >>> them here.
> > >>> size_t is meant to address the need for a type to describe object sizes
> > >>> on a
> > >>> particular system, and it itself is sized accordingly (32 bits on a 32
> > >>> bit arch,
> > >>> 64 on 64), so that you can safely store a size that the system in
> > >>> question might
> > >>> maximally allocate/return. In this situation we are describing memory
> > >>> sizes
> > >>> that might occur no a plurality of arches, and so size_t is
> > >>> inappropriate
> > >>> because it as a type is not sized for anything other than the arch it
> > >>> is being
> > >>> built for. The pragmatic benefits of ennumerating page sizes in a
> > >>> single
> > >>> canonical location far outweigh the desire to use a misappropriated
> > >>> type to
> > >>> describe them.
> > >> Neil,
> > >>
> > >> This patch fix two compile issues, and we need to do *dpdk testing
> > >> affairs*, if it is blocked in build stage, we can do *nothing* for
> > >> testing.
> > >>
> > >> I've get you mind and your concern. But we should take care of changing
> > >> the type of "hugepage_sz", because lots of places using it.
> > >>
> > >> Would you mind if we consider this as hot fix, and we can post a better
> > >> fix later(like in dpdk 2.0)? Otherwise all test cycle are blocked.
> > >>
> > > Honestly, no. Because intels testing schedule shouldn't drive the
> > > inclusion of
> > > upstream fixes. Also, I'm not asking for a major redesign of anything,
> > > I'm
> > > asking for a proper fix for a very straightforward problem. I've
> > > attached the
> > > proper fix below.
> > >
> > > Regards
> > > Neil
> >
> > We test dpdk upstream now as 1,8 rc2 and rc3 released :)
> >
> Yes, I don't take issue with you testing dpdk, on the contrary, I appreciate
> it.
> What I take issue with is that you are asserting that the blockage of your
> testing is reason to ignore a proper fix an issue, rather than some
> substandard
> one.
I agree. Neil's patch seems a lot better.
> > I know that what you mean. but lots of please using "hugepage_sz" do you
> > confirm it will not affect other issue?
> >
> 5. There are 5 placees that use hugepage_sz, as the patch below indicates.
> Thats no alot.
>
> Also, I take issue with the assertion that this patch creates no additional
> problems. I'm sure it creates no additional problems that your patch wouldn't
> also create, arguably less. If we were really being pragmatic here, I would
> point out that this problem was caused by the fact that support for an entire
> new architecture was integrated during the -rc phase of a release, which seems
> extreemely risky to me,
No, it was integrated between -rc1 and -rc2. But -rc1 was not really a release
candidate. It was a first step after mbuf rework. This tag was requested for
validation but it was a bad idea. We won't create such wrong tag anymore.
</digress>
PPC was integrated before the real RC phase.
> and as such, the most appropriate thing to do would be
> to back support for ppc out until after the 1.8 release when it could be
> properly tested. Instead we are slamming in hacked up fixes that throw arch
> specific ifdefs througout the code.
I think we can fix it (without ugly ifdefs) and avoid going back.
Thanks for your help.
--
Thomas
> > On other hand, we use 32 bit address in 32 bit platform for better
> > performance(some of places also use uintptr_t for address check and
> > alignment).
> >
> This has nothing to do with address bus size.
>
> > And it should not acceptable in 32 bit platform to use 64-bit platform
> > specification affairs(like RTE_PGSIZE_16G).
> >
> Ok, so add a single arch specific runtime check during hugepage mapping to
> exit
> on the 16G size use on 32 bit systems. Thats a fair and reasonable thing to
> do,
> though I think the hugepage remap is already ifdeffed for 54 bit arches only.
>
> Neil
