[vfio-users] Interrupt Handling: Regression or config change?

[Laszlo Ersek]

On 01/13/17 18:54, Torbjorn Jansson wrote:
> On 2017-01-12 19:49, Laszlo Ersek wrote:
>> On 01/12/17 19:21, Ede Wolf wrote:
>>> Besides my fighting with git, is it correct, that the working kernel is
>>> labeld "bad", while the non working one "good"? May very well possible
>>> for the workflow of dissect, just want to confirm, please
>>>
>>>> git bisect start
>>>> git bisect bad v4.5.4
>>>> git bisect good v4.6.1
>>
>> No. git-bisect always locates a regression, that is, the first commit
>> that introduces a bug. Where things go from right to wrong. If this
>> matches your situation (old version works, new version breaks), simply
>> use the "good" and "bad" command line arguments verbatim.
>>
>> Now, if you are looking for a commit that fixes a known bug, i.e., old
>> version is broken, new version works, but you don't exactly know what
>> fixed it, then you have to invert the meanings. Whenever the commit
>> under testing works, you have to say "bad", and whenever it breaks, you
>> have to say "good". More modern git versions help automate this (for
>> less boggling of the mind) with "git bisect terms". (See the manual.)
>> With older git versions, git command aliases, shell aliases, or small
>> wrapper scripts can hide the confusion.
>>
>> Given that you are facing a real regression, you should use the good/bad
>> terms verbatim. From your earlier email, "4.5.7" works, "4.6" breaks,
>> hence you should qualify the former with "good", and the latter with
>> "bad".
>>
>> Also, the word is "bisect", not "dissect".
>>
>> https://en.wiktionary.org/wiki/bisect
>> https://en.wiktionary.org/wiki/dissect
>>
>> git-bisect performs a binary search, halving commit ranges. Hence the
>> "bi" in "bisect".
>>
>> Laszlo
> 
> it is also important when bisecting to make 100% sure you tell git the
> right thing after testing each build.
> this can be complicated if some builds don't compile properly and
> prevents you from building and testing that specific version.
> 
> also binary search means you cut the problem into two halves each time
> and when you tell git if it is bad or good, this means you throw away
> that half of the problem and then repeat the process to further isolate
> the issue.
> 
> all of this is automated by git bisect, but it all depends on picking a
> good and bad commit to start with and then making sure you tell git if
> the current version you are testing is really good or bad.
> 
> 
> when i tried git bisect the first time i screwed up because of this.
> then i did it right and also kept track of what commits i was testing
> (the output of each git bisect) since some of the versions did not build
> so i had to restart the bisect at a different point.

Manuel mentions "git bisect skip" in his followup; I'd just like to
point out a generalization: git-bisect doesn't really dictate what
commits you should test at, it merely makes suggestions, and then
records the outcome of the test at *whatever* commit you performed the
test. Then the next suggestion (automatic checkout) will be based on all
the commits tested thus far.

What it boils down to is this: if the commit suggested by git-bisect
doesn't build, or it builds, but execution encounters an early error
that prevents you from testing the issue of interest, not only can you
use "git bisect skip", but even "git checkout" with whatever commit you
deem fit for testing. When you test that commit, and then record the
result with "git bisect good" or "git bisect bad", git-bisect will
adapt, and suggest (auto-checkout) the next commit to test accordingly.

Sometimes this comes very handy (e.g. you know for sure what commit
fixes another, independent bug that masks the bug of your interest).

Laszlo