Re: 2.6.21-rc1: known regressions (part 2)

[Ingo Molnar <mingo@elte.hu>]

* Ingo Molnar <mingo@elte.hu> wrote:

> I'll now try the following: i'll try to manually apply Len's fix to 
> every tree that git-bisect offers me, in the hope of being able to 
> isolate the /other/ bug.
> 
> [ But really, i'm not expecting any miracles because this is way out of
>   league for git-bisect which really depends on only having a binary 
>   space to search for. ]

this method pointed out the real bug that we are interested in:

| 774c47f1d78e373a6bd2964f4e278d1ce26c21cb is first bad commit
| commit 774c47f1d78e373a6bd2964f4e278d1ce26c21cb
| Author: Avi Kivity <avi@qumranet.com>
| Date:   Mon Feb 12 00:54:47 2007 -0800
|     [PATCH] KVM: cpu hotplug support

undoing the 774c47f1 patch from HEAD gave me a working resume. I'll send 
a fix for this KVM breakage in a separate mail.

here's how the bisection went:

 bad:   [01363220f5d23ef68276db8974e46a502e43d01d] [PARISC] clocksource:
 good:  [fa285a3d7924a0e3782926e51f16865c5129a2f7] Linux 2.6.20
 +good: [bcdb81ae29091f6a66369aabfd8324e4a53d05dc] knfsd: SUNRPC: add a
 bad:   [920841d8d1d61bc12b43f95a579a5374f6d98f81] Merge branch 'for-linus'
 +bad:  [86a71dbd3e81e8870d0f0e56b87875f57e58222b] sysctl: hide the sysctl
 +bad:  [719c91ccadd3ed26570dbb29d54166914832eee9] [POWERPC] Use udbg_early
 +bad:  [ebaf0c6032f525ddb0158fb59848d41899dce8cd] Merge branch 'for-linus'
 +good: [8cd133073f9b5cd335c0b2e4740aceb025d50ca9] kvm: Fix mismatch between
 +bad:  [5b8e8ee6c65a34d8aafaeb8e2eaa97e496c2567c] ps3: add shutdown to
 +bad:  [a524d946bdced73c5fbe60170fb33611491c4211] tgafb: sync-on-green
 +bad:  [a268422de8bf1b4c0cb97987b6c329c9f6a3da4b] fbdev driver for S3
 +good: [8d0be2b3bf4a55606967d7d84e56c52521e94333] KVM: VMX: add vcpu_clear()
 +bad:  [59ae6c6b87711ceb2d1ea5f9e08bb13aee947a29] KVM: Host suspend/resume
 +bad:  [774c47f1d78e373a6bd2964f4e278d1ce26c21cb] KVM: cpu hotplug support

the commits prefixed with '+' were the ones where i had to hand-merge 
the f3ccb06f commit to. Near the end of the bisection it nicely honed in 
on the group of KVM changes that included the bad commit.

but the conclusion is clear: if multiple bugs are present in the search 
area then it gets quite difficult to sort it out via git-bisect - but 
it's not impossible either. The following git-bisect enhancement could 
have made things easier for me:

   git-bisect mark-must-have <tree>

which would mark <tree> as a must-have fix and would attempt to merge it 
to whatever bisection point it ends up with - if that bisection point 
does not have <tree> included already. (Maybe not even the full tree but 
only one particular commit ID.) In this particular case that merge, when 
it had to be done, was always 'clean'.

	Ingo