bug#27476: guix pull fails on powerful server

[Ricardo Wurmus <rekado@elephly.net>]

Hi Ludo,

> Ricardo Wurmus <rekado@elephly.net> skribis:
>
>> The following derivation will be built:
>>    /gnu/store/z5bhk17nxmdhvj0g4cy038p25mzh1gys-guix-latest.drv
>> copying and compiling to '/gnu/store/s3s7xlqa10mvf8v0ypxz8gzw3lcf1x5z-guix-latest' with Guile 2.2.2...
>> loading...       25.7% of 635 filesrandom seed for tests: 1506720257
>> loading...       99.8% of 635 files
>> compiling...     69.1% of 635 filesice-9/threads.scm:289:22: In procedure loop:
>> ice-9/threads.scm:289:22: Syntax error:
>> guix/scripts/graph.scm:103:10: return: return used outside of 'with-monad' in form (return (package-node-edges a))
>
> The program below crashes with completely surreal backtraces in less
> than a minute on my 4-thread laptop:
>
> --8<---------------cut here---------------start------------->8---
> (use-modules (ice-9 threads)
>              (srfi srfi-1)
>              (guix monads)
>              (guix store))
>
> (define threads
>   (unfold (lambda (x) (> x 100))
>           (lambda (x)
>             (call-with-new-thread
>              (lambda ()
>                (define monad
>                  (symbol-append 'foo-monad
>                                 (string->symbol (number->string x))))
>
>                (while #t
>                  (macroexpand
>                   `(begin
>                      (define-monad ,monad
>                        (bind +)
>                        (return -))
>                      (with-monad ,monad
>                        (return 3))
>                      (mapm ,monad + '(1 2 3))))))))
>           1+
>           0))
>
> (for-each join-thread threads)
> --8<---------------cut here---------------end--------------->8---
>
> Can you check if that also happens on your many-core machine?

It does not crash.  I left it running for more than an hour (without
compiling) and it printed things like this:

--8<---------------cut here---------------start------------->8---
…
GC Warning: Repeated allocation of very large block (appr. size 57528320):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 57528320):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 57528320):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 57528320):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 14385152):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 14385152):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 57528320):
        May lead to memory leak and poor performance
GC Warning: Repeated allocation of very large block (appr. size 28766208):
        May lead to memory leak and poor performance
…
--8<---------------cut here---------------end--------------->8---

That’s on the machine with 1.5T RAM and 192 cores.  Then I ran it again
for 10 minutes after compiling it.  It did not crash.

> The patch below seems to fix the problem: (guix monads) has shared state
> (hash tables) used both at expansion-time and run-time, and it wasn’t
> protected.
>
> My hypothesis is that this was causing random memory corruption.  The
> weird thing, though, is that the errors we were getting were not so
> random.  Also, the load phase of ‘guix pull’ is sequential.
>
> Could you test it and report back?

I’m trying the patch right now with “guix pull”.

--
Ricardo

GPG: BCA6 89B6 3655 3801 C3C6  2150 197A 5888 235F ACAC
https://elephly.net