Re: svn commit: r308817 - head/sys/powerpc/include [Still have pmap_t and struct pmap ppowerpc64 problems as of -r308860]

From: Mark Millard <markmi_at_dsl-only.net>
Date: Sun, 20 Nov 2016 00:05:11 -0800
Glad to help.


Going in another direction: What of stable/11 ? (And possibly stable/10 .)

Changing the BROOKE kernel ABI in stable/11 is a no-no. Similarly for AIM.

Is there a reason to prefer which context actually works correctly for
11.0 default builds for accessing pm_stats in pmap?

My guess is that GENERIC64 and GENERIC for powerpc64 and powerpc should
by default work.

Needing to have something special for 11.x BOOKE to build correctly in
seems less of a problem overall: BOOKE likely has a more tailored 11.x
build context anyway and there are no ftp 11.0 snapshots or releases for
KERNCONF=MPC85XX . TARGET_ARCH=powerpcspe (KERNCONF=MPC85XXSPE) does
not exist until head (12.0-CURRENT) from what I can tell.

This might suggest for stable/11 something like:
(notation picked for emphasis)

#if defined(AIM) || !defined(BOOKE)
. . .
struct  pmap {
        struct  mtx     pm_mtx;
        
    #ifdef __powerpc64__
        struct slbtnode *pm_slb_tree_root;
        struct slb      **pm_slb;
        int             pm_slb_len;
    #else
        register_t      pm_sr[16];
    #endif
        cpuset_t        pm_active;

        struct pmap     *pmap_phys;
        struct          pmap_statistics pm_stats;. . .
. . . (the unchanged order above) . . .

#elif defined(BOOKE)

struct pmap {
        struct mtx              pm_mtx;         /* pmap mutex */
        tlbtid_t                pm_tid[MAXCPU]; /* TID to identify this pmap entries in TLB */
        cpuset_t                pm_active;      /* active on cpus */
        struct pmap_statistics  pm_stats;       /* pmap statistics */
. . . (the unchanged order above) . . .
#endif

with some means used for defining BOOKE outside buildkernel
for the BOOKE builds, such as for KERNCONF=MPC85XX .

Then at least GENERIC and GENERIC64 would build correctly
for this subject area by default.

In particular GENERIC64 actually has no MPC85XX (BOOKE)
alternative so far as I know. That makes it too bad that
it is always broken for the issue in stable/11 (and in
stable/10 ).

I choose the above to also cover GENERIC. If only GENERIC64
was to be made to work right by default there are alternatives
that involve __powerpc64__ in at least one more place, such
as:

#if defined(AIM) || defined(__powerpc64__)


===
Mark Millard
markmi at dsl-only.net

On 2016-Nov-19, at 10:12 PM, Justin Hibbits <jhibbits at freebsd.org> wrote:

> My buildworld completed successfully, so it's been fixed in r308873/r308874.
> 
> Thanks for your testing.  I often build just kernel, so wouldn't have seen the fallout until it was far too late.
> 
> - Justin

On Nov 19, 2016, at 10:22 PM, Mark Millard wrote:

> On 2016-Nov-16, at 8:33 PM, Justin Hibbits <jhibbits at freebsd.org> wrote:
> 
>> *sigh* okay, thanks.  I just tested, and vm/vm_page.h, and vm/vm.h can both be removed from memstat_uma.c for it to compile.  I'm kicking off a buildworld myself now, too, and hope to have it ready to commit tomorrow (takes a couple hours to buildworld on my G5).
>> 
>> - Justin
> 
> That will not be the only potential place: umastat.c in tools/tools/umastat/
> also has a include of vm/vm_page.h:
> 
>> # find /usr/src/ -name .svn -prune -o -name sys -prune -o -name man -prune -o -exec grep "vm_page[.]h" {} \; -print | more
>> #include <vm/vm_page.h>
>> /usr/src/lib/libmemstat/memstat_uma.c
>> #define LIBMEMSTAT      /* Cause vm_page.h not to include opt_vmpage.h */
>> #include <vm/vm_page.h>
>> /usr/src/tools/tools/umastat/umastat.c
> 
> 
> ===
> Mark Millard
> markmi at dsl-only.net
> 
> On Nov 19, 2016, at 9:47 PM, Mark Millard wrote:
> 
>> [Top post of bad news.]
>> 
>> With the patch I get a different incomplete type used in libmemstat:
>> 
>> struct md_page
>> 
>> --- all_subdir_lib/libmemstat ---
>> In file included from /usr/src/lib/libmemstat/memstat_uma.c:34:0:
>> /usr/src/sys/vm/vm_page.h:144:17: error: field 'md' has incomplete type
>> struct md_page md;  /* machine dependent stuff */
>>              ^
>> *** [memstat_uma.o] Error code 1
>> 
>> make[5]: stopped in /usr/src/lib/libmemstat
>> 
>> 
>> ===
>> Mark Millard
>> markmi at dsl-only.net
>> 
>> On 2016-Nov-19, at 7:42 PM, Mark Millard <markmi at dsl-only.net> wrote:
>> 
>>> On 2016-Nov-19, at 7:36 PM, Mark Millard <markmi at dsl-only.net> wrote:
>>> 
>>>> On 2016-Nov-19, at 7:32 PM, Justin Hibbits <jhibbits at freebsd.org> wrote:
>>>> 
>>>>> Sorry, I generated the diff from a different tree that wasn't synced to head (had the same change in both trees originally). If that is the only problem, you can ignore it and try the rest. I can generate another diff later too.
>>>>> - Justin
>>>> 
>>>> Yep: I manually did the move of the pm_stats line and am building.
>>> 
>>> If it builds and I install it on a PowerMac G5 and it boots, what do I
>>> do to test if pm_stats and pm_mtx seems to be working well/right for
>>> the out of kernel code? Do you know of a reasonable test?
>>> 
>>> ===
>>> Mark Millard
>>> markmi at  dsl-only.net
>>> 
>> On Nov 19, 2016 21:27, "Mark Millard" <markmi at dsl-only.net> wrote:
>>> [Top post about patch issues.]
>>> 
>>> Looking at the patch it seems to be designed for when #else was in use:
>>> 
>>>> -#else
>>>> +#elif defined(BOOKE)
>>> 
>>> but -r308817 already has the 2nd line (BOOKE). Your patch shows:
>>> 
>>>> Index: sys/powerpc/include/pmap.h
>>>> ===================================================================
>>>> --- sys/powerpc/include/pmap.h  (revision 308718)
>>>> +++ sys/powerpc/include/pmap.h  (working copy)
>>> 
>>> So it looks like you started from before -r308817 .
>>> 
>>> Trying it (I'm at -r308860):
>>> 
>>>> Patching file sys/powerpc/include/pmap.h using Plan A...
>>>> Hunk #1 succeeded at 74.
>>>> Hunk #2 succeeded at 84.
>>>> Hunk #3 succeeded at 132.
>>>> Hunk #4 succeeded at 145.
>>>> Hunk #5 failed at 180.
>>>> Hunk #6 succeeded at 194.
>>>> Hunk #7 succeeded at 210.
>>>> 1 out of 7 hunks failed--saving rejects to sys/powerpc/include/pmap.h.rej
>>> 
>>>> # more sys/powerpc/include/pmap.h.rej
>>>> _at__at_ -179,13 +180,13 _at__at_
>>>> struct slb **slb_alloc_user_cache(void);
>>>> void   slb_free_user_cache(struct slb **);
>>>> 
>>>> -#else
>>>> +#elif defined(BOOKE)
>>>> 
>>>> struct pmap {
>>>> +       struct pmap_statistics  pm_stats;       /* pmap statistics */
>>>>   struct mtx              pm_mtx;         /* pmap mutex */
>>>>   tlbtid_t                pm_tid[MAXCPU]; /* TID to identify this pmap entries in TLB */
>>>>   cpuset_t                pm_active;      /* active on cpus */
>>>> -       struct pmap_statistics  pm_stats;       /* pmap statistics */
>>>> 
>>>>   /* Page table directory, array of pointers to page tables. */
>>>>   pte_t                   *pm_pdir[PDIR_NENTRIES];
>>> 
>>> 
>>> ===
>>> Mark Millard
>>> markmi at dsl-only.net
>>> 
>>> On 2016-Nov-19, at 7:00 PM, Mark Millard <markmi at dsl-only.net> wrote:
>>> 
>>> It may take a little bit but I'll try the patch.
>>> 
>>> It looks like sys/powerpc/include/pmap.h from -r176700 from 2088-Mar-3
>>> is when the BOOKE/E500 split started with the preprocessor use of AIM
>>> and #else . This predates PowerMac G5 support.
>>> 
>>> This is definitely not new for the general structure on the powerpc
>>> side of things. Any place that did not have the AIM vs. not status
>>> available was subject to problems of possibly mismatched definitions.
>>> 
>>> ===
>>> Mark Millard
>>> markmi at dsl-only.net
>>> 
>>> On 2016-Nov-19, at 6:47 PM, Justin Hibbits <jhibbits at freebsd.org> wrote:
>>> 
>>> On Sat, 19 Nov 2016 18:36:39 -0800
>>> Mark Millard <markmi at dsl-only.net> wrote:
>>> 
>>>> [Quick top post I'm afraid.]
>>>> 
>>>> I think that I figured out why there is a problem even earlier
>>>> --that just did not stop the compiles.
>>>> 
>>>> lib/libutil/kinfo_getallproc.c is built here as part of buildworld
>>>> (stage 4.2 "building libraries" instead of buildkernel. It does not
>>>> have the KERNCONF's AIM vs. BOOKE vs. . . . definitions vs. lack of
>>>> them).
>>>> 
>>>> So if it includes machine/pmap.h that binds to
>>>> sys/powerpc/include/pmap.h which has the structure. . .
>>>> 
>>>> . . .
>>>> #if defined(AIM)
>>>> . . . (definitions here)
>>>> #elif defined(BOOKE)
>>>> . . . (definitions here)
>>>> #endif
>>>> . . .
>>>> 
>>>> it gets no definition now.
>>>> 
>>>> With the older:
>>>> 
>>>> . . .
>>>> #if defined(AIM)
>>>> . . . (definitions here)
>>>> #else
>>>> . . . (definitions here)
>>>> #endif
>>>> . . .
>>>> 
>>>> It got a definition, just not necessarily the right one.
>>>> 
>>>> 
>>>> ===
>>>> Mark Millard
>>>> markmi at dsl-only.net
>>> 
>>> Can you try the attached patch?  There was a subtle ABI issue that
>>> r308817 exposed, which is that the pmap structs aren't identical such
>>> that the pm_stats are at different locations, and libkvm ends up
>>> reading with the Book-E pmap, getting different stats than expected for
>>> AIM.  This patch fixes that, bumping version to account for this ABI
>>> change.
>>> 
>>> - Justin<fix_pmap.diff>
>> 
>> 
>> 
> 
> 
Received on Sun Nov 20 2016 - 07:05:20 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:41:08 UTC