user interrupt in copy/register step breaks the package database

[cblp]

cblp@krypton:~$ stack --resolver=lts-9.0 install weeder  
clock-0.7.2: using precompiled package
mmorph-1.0.9: using precompiled package
transformers-base-0.4.4: using precompiled package
cmdargs-0.10.17: using precompiled package
extra-1.5.3: using precompiled package
monad-control-1.0.2.1: download
monad-control-1.0.2.1: configure
monad-control-1.0.2.1: build
monad-control-1.0.2.1: copy/register                                                       
Progress: 5/11^Cuser interrupt                                                             

cblp@krypton:~$ stack --resolver=lts-9.0 install weeder                                    
lifted-base-0.2.3.11: configure                                                            
lifted-base-0.2.3.11: build                                                                
Progress: 1/5                                                                              
--  While building package lifted-base-0.2.3.11 using:                                     
      /home/cblp/.stack/setup-exe-cache/x86_64-linux-nopie/Cabal-simple_mPHDZzAJ_1.24.2.0_ghc-8.0.2 --builddir=.stack-work/dist/x86_64-linux-nopie/Cabal-1.24.2.0 build --ghc-options " -ddump-hi -ddump-to-file"                                                                
    Process exited with code: ExitFailure 1                                                
    Logs have been written to: /home/cblp/.stack/global-project/.stack-work/logs/lifted-base-0.2.3.11.log

    Configuring lifted-base-0.2.3.11...
    Building lifted-base-0.2.3.11...
    Preprocessing library lifted-base-0.2.3.11...
    <command line>: cannot satisfy -package-id monad-control-1.0.2.1-CqOqIBYVBjy1GnPfyWnFHC
        (use -v for more information)

cblp@krypton:~$ stack --version
Version 1.4.0, Git revision 5986498a9d1073e8d58c79ec1302bb2a7847a396 (4642 commits) x86_64 hpack-0.17.0

[mgsloan]

Interesting, good to know!

It probably makes sense to wait for the copy to finish before exiting from Async exceptions (in particular ThreadKilled / UserInterrupt). An additional Ctrl-C would force exit, could print a message to warn the user not to do that, and when it happens, print a message letting them know the DB may be corrupted.

It also likely makes sense to have similar protection around ghc-pkg invocations.

[snoyberg]

I like that as a workaround @mgsloan. One question though: should it be the copy step that gets masked, or the register step? AFAICT, an interrupted copy shouldn't hurt anyone. But maybe I'm mistaken.

We don't need such protection around all ghc-pkg invocations, probably just the mutating ones.

[mgsloan]

Ah, yeah, it is probably just the register step. Certainly, just the mutating ghc-pkg calls.

[mgsloan]

Bumping to P1 due to popular demand.

[cblp]

Doesn't ghc-pkg have proper SIGINT handler?

[mgsloan]

@cblp it doesn't appear to be so. So should check this box at some point:

• Report upstream to ghc

At least once we're sure it's ghc-pkg messing up here

[snoyberg]

This is definitely an upstream issue, see #4559. Following up there.