[bug#56218] guix: inferior: Fix the behaviour of open-inferior #:error-port.

  This should be the error port used by the inferior process, but currently it's
either stderr if #:error-port is a file port, or /dev/null otherwise.

I'm looking at this as the Guix Data Service uses this behaviour to record and
display logs from inferior processes.

* guix/inferior.scm (open-bidirectional-pipe): Call dup2 for file descriptor
2, passing either the file number for the current error port, or a file
descriptor for /dev/null.
---
 guix/inferior.scm | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
  

Christopher Baines schreef op za 25-06-2022 om 18:18 [+0100]:
>              (close-port parent)
>              (close-fdes 0)
>              (close-fdes 1)
> +            (close-fdes 2)
>              (dup2 (fileno child) 0)
>              (dup2 (fileno child) 1)
>              ;; Mimic 'open-pipe*'.
> -            (unless (file-port? (current-error-port))
> -              (close-fdes 2)
> -              (dup2 (open-fdes "/dev/null" O_WRONLY) 2))
> +            (dup2 (if (file-port? (current-error-port))
> +                      (fileno (current-error-port))
> +                      (open-fdes "/dev/null" O_WRONLY))
> +                  2)

I don't this would work if (current-error-port) has fd 1. Would
move->fdes be appropriate here?  The following seems less fragile (*)
to me (untested, also I didn't look at the context)

  (move->fdes [child port] 0)
  (move->fdes (dup [child port]) 1)
  (if (file-port? (current-error-port))
      (move->fdes (current-error-port) 2)
      (move->fdes (open-file "/dev/null" O_WRONLY) 2))


(*): move->fdes automatically moves ports out of the way.  Also, if one
of the moves fails, then at least (current-output-port) etc will still
have a correct fd so some error reporting should be possible

Greetings,
Maxime.
  

Maxime Devos schreef op za 25-06-2022 om 19:50 [+0200]:
> I don't this would work if (current-error-port) has fd 1. [...]

TBC, I never worked with file descriptor manipulation and process
forking much, and I didn't look at the surrounding code, so don't take
my word for it.
  

Hi Christopher,

Christopher Baines <mail@cbaines.net> skribis:

> This should be the error port used by the inferior process, but currently it's
> either stderr if #:error-port is a file port, or /dev/null otherwise.

That’s still the case with this patch, no?

The patch does make a difference when (current-error-port) wraps a file
descriptor other than 2 though.

> +++ b/guix/inferior.scm
> @@ -156,12 +156,14 @@ (define (open-bidirectional-pipe command . args)
>              (close-port parent)
>              (close-fdes 0)
>              (close-fdes 1)
> +            (close-fdes 2)
>              (dup2 (fileno child) 0)
>              (dup2 (fileno child) 1)
>              ;; Mimic 'open-pipe*'.
> -            (unless (file-port? (current-error-port))
> -              (close-fdes 2)
> -              (dup2 (open-fdes "/dev/null" O_WRONLY) 2))
> +            (dup2 (if (file-port? (current-error-port))
> +                      (fileno (current-error-port))
> +                      (open-fdes "/dev/null" O_WRONLY))
> +                  2)

If (current-error-port) wraps FD 2 when the function is called, then, by
the time we reach (dup2 … 2), the FD behind (current-error-port) has be
closed; we end up doing (dup2 2 2), but FD 2 is closed, so we get EBADF.

Or am I misunderstanding?

Perhaps we should add one test for each case (error port is a file port
vs. error port is another kind of port) in ‘tests/inferior.scm’.

Thanks,
Ludo’.
  

Maxime Devos <maximedevos@telenet.be> writes:

> [[PGP Signed Part:Undecided]]
> Christopher Baines schreef op za 25-06-2022 om 18:18 [+0100]:
>>              (close-port parent)
>>              (close-fdes 0)
>>              (close-fdes 1)
>> +            (close-fdes 2)
>>              (dup2 (fileno child) 0)
>>              (dup2 (fileno child) 1)
>>              ;; Mimic 'open-pipe*'.
>> -            (unless (file-port? (current-error-port))
>> -              (close-fdes 2)
>> -              (dup2 (open-fdes "/dev/null" O_WRONLY) 2))
>> +            (dup2 (if (file-port? (current-error-port))
>> +                      (fileno (current-error-port))
>> +                      (open-fdes "/dev/null" O_WRONLY))
>> +                  2)
>
> I don't this would work if (current-error-port) has fd 1. Would
> move->fdes be appropriate here?  The following seems less fragile (*)
> to me (untested, also I didn't look at the context)
>
>   (move->fdes [child port] 0)
>   (move->fdes (dup [child port]) 1)
>   (if (file-port? (current-error-port))
>       (move->fdes (current-error-port) 2)
>       (move->fdes (open-file "/dev/null" O_WRONLY) 2))
>
>
> (*): move->fdes automatically moves ports out of the way.  Also, if one
> of the moves fails, then at least (current-output-port) etc will still
> have a correct fd so some error reporting should be possible

Maybe. I haven't actually tried this yet, but the docs seem to suggest
it would work.

Thanks,

Chris
  

Ludovic Courtès <ludo@gnu.org> writes:

> Hi Christopher,
>
> Christopher Baines <mail@cbaines.net> skribis:
>
>> This should be the error port used by the inferior process, but currently it's
>> either stderr if #:error-port is a file port, or /dev/null otherwise.
>
> That’s still the case with this patch, no?
>
> The patch does make a difference when (current-error-port) wraps a file
> descriptor other than 2 though.

Maybe this sentance is a little unclear.

What I'm trying to say is that passing a port as #:error-port doesn't
really work. There's no scenario where the output actually goes to the
port you provide, though it can have some effect.

>> +++ b/guix/inferior.scm
>> @@ -156,12 +156,14 @@ (define (open-bidirectional-pipe command . args)
>>              (close-port parent)
>>              (close-fdes 0)
>>              (close-fdes 1)
>> +            (close-fdes 2)
>>              (dup2 (fileno child) 0)
>>              (dup2 (fileno child) 1)
>>              ;; Mimic 'open-pipe*'.
>> -            (unless (file-port? (current-error-port))
>> -              (close-fdes 2)
>> -              (dup2 (open-fdes "/dev/null" O_WRONLY) 2))
>> +            (dup2 (if (file-port? (current-error-port))
>> +                      (fileno (current-error-port))
>> +                      (open-fdes "/dev/null" O_WRONLY))
>> +                  2)
>
> If (current-error-port) wraps FD 2 when the function is called, then, by
> the time we reach (dup2 … 2), the FD behind (current-error-port) has be
> closed; we end up doing (dup2 2 2), but FD 2 is closed, so we get EBADF.
>
> Or am I misunderstanding?

That sounds reasonable, I've only tested this change in the scenario
when the #:error-port isn't stderr, and I mostly adapted this from what
I thought open-pipe* did.

Maxime suggested using move->fdes, so maybe this would be an improved
version:

  ;; Mimic 'open-pipe*'.
  (if (file-port? (current-error-port))
      (unless (eq? (fileno (current-error-port)) 2)
        (move-fdes (current-error-port) 2))
      (move->fdes (open-file "/dev/null" O_WRONLY) 2))

> Perhaps we should add one test for each case (error port is a file port
> vs. error port is another kind of port) in ‘tests/inferior.scm’.

Yep, sounds good.

Thanks,

Chris
  

Hi,

Christopher Baines <mail@cbaines.net> skribis:

> Ludovic Courtès <ludo@gnu.org> writes:
>
>> Hi Christopher,
>>
>> Christopher Baines <mail@cbaines.net> skribis:
>>
>>> This should be the error port used by the inferior process, but currently it's
>>> either stderr if #:error-port is a file port, or /dev/null otherwise.
>>
>> That’s still the case with this patch, no?
>>
>> The patch does make a difference when (current-error-port) wraps a file
>> descriptor other than 2 though.
>
> Maybe this sentance is a little unclear.
>
> What I'm trying to say is that passing a port as #:error-port doesn't
> really work. There's no scenario where the output actually goes to the
> port you provide, though it can have some effect.

OK, I think I got it.

>>> +            (dup2 (if (file-port? (current-error-port))
>>> +                      (fileno (current-error-port))
>>> +                      (open-fdes "/dev/null" O_WRONLY))
>>> +                  2)
>>
>> If (current-error-port) wraps FD 2 when the function is called, then, by
>> the time we reach (dup2 … 2), the FD behind (current-error-port) has be
>> closed; we end up doing (dup2 2 2), but FD 2 is closed, so we get EBADF.
>>
>> Or am I misunderstanding?
>
> That sounds reasonable, I've only tested this change in the scenario
> when the #:error-port isn't stderr, and I mostly adapted this from what
> I thought open-pipe* did.
>
> Maxime suggested using move->fdes, so maybe this would be an improved
> version:
>
>   ;; Mimic 'open-pipe*'.
>   (if (file-port? (current-error-port))
>       (unless (eq? (fileno (current-error-port)) 2)
>         (move-fdes (current-error-port) 2))
>       (move->fdes (open-file "/dev/null" O_WRONLY) 2))

I prefer the original version: I find it clearer (it’s low-level) and
probably more robust (thinking through the port/FD interaction needs is
more demanding :-)).

>> Perhaps we should add one test for each case (error port is a file port
>> vs. error port is another kind of port) in ‘tests/inferior.scm’.
>
> Yep, sounds good.

To sum up: I think it’s a welcome change, and it’s even more welcome
with a couple of tests to make sure it behaves the way we think it does.

Thanks!

Ludo’.
  

Ludovic Courtès <ludo@gnu.org> writes:

>>>> +            (dup2 (if (file-port? (current-error-port))
>>>> +                      (fileno (current-error-port))
>>>> +                      (open-fdes "/dev/null" O_WRONLY))
>>>> +                  2)
>>>
>>> If (current-error-port) wraps FD 2 when the function is called, then, by
>>> the time we reach (dup2 … 2), the FD behind (current-error-port) has be
>>> closed; we end up doing (dup2 2 2), but FD 2 is closed, so we get EBADF.
>>>
>>> Or am I misunderstanding?
>>
>> That sounds reasonable, I've only tested this change in the scenario
>> when the #:error-port isn't stderr, and I mostly adapted this from what
>> I thought open-pipe* did.
>>
>> Maxime suggested using move->fdes, so maybe this would be an improved
>> version:
>>
>>   ;; Mimic 'open-pipe*'.
>>   (if (file-port? (current-error-port))
>>       (unless (eq? (fileno (current-error-port)) 2)
>>         (move-fdes (current-error-port) 2))
>>       (move->fdes (open-file "/dev/null" O_WRONLY) 2))
>
> I prefer the original version: I find it clearer (it’s low-level) and
> probably more robust (thinking through the port/FD interaction needs is
> more demanding :-)).
>
>>> Perhaps we should add one test for each case (error port is a file port
>>> vs. error port is another kind of port) in ‘tests/inferior.scm’.
>>
>> Yep, sounds good.
>
> To sum up: I think it’s a welcome change, and it’s even more welcome
> with a couple of tests to make sure it behaves the way we think it does.

I've gone ahead and pushed a fix plus some tests as
a9fd06121240c78071a398dd1e0ddb47553f3809.

The tests probably aren't great, but I think the do cover the
#:error-port behaviour.