Re: [PATCH v5 proxmox-backup 10/16] client: treat minus sign as stdin

[Robert Obkircher <r.obkircher@proxmox.com>]

On 2/4/26 15:42, Christian Ebner wrote:
> On 2/4/26 3:00 PM, Robert Obkircher wrote:
>>
>> On 2/2/26 13:14, Christian Ebner wrote:
>>> On 1/30/26 5:45 PM, Robert Obkircher wrote:
>>>> Treat "-" as an alias for "/dev/stdin". If there is an actual file
>>>> with that name it can still be read via "./-".
>>>>
>>>> Signed-off-by: Robert Obkircher <r.obkircher@proxmox.com>
>>>> ---
>>>>    proxmox-backup-client/src/main.rs | 7 +++++++
>>>>    1 file changed, 7 insertions(+)
>>>>
>>>> diff --git a/proxmox-backup-client/src/main.rs
>>>> b/proxmox-backup-client/src/main.rs
>>>> index 7fc711fd..37878b01 100644
>>>> --- a/proxmox-backup-client/src/main.rs
>>>> +++ b/proxmox-backup-client/src/main.rs
>>>> @@ -845,6 +845,13 @@ async fn create_backup(
>>>>            }
>>>>            target_set.insert(target.clone());
>>>>    +        // one can still use ./- to refer to an actual file with
>>>> that name
>>>> +        let filename = if filename == "-" {
>>>> +            String::from("/dev/stdin")
>>>
>>> should we add a warning here?
>>>
>>> This is actually a breaking change, but guarding it behind a flag is
>>> maybe a bit overkill. We could check if the file exists before
>>> continuing and if so fail with an appropriate warning/error?
>>
>> Another option would be to just require /dev/stdin and print a hint
>> about that if the user tries to specify "-". What do you think
>> about that?
>
> Might be the better option for the time being, yes. And must be
> documented correctly of course. The '-' shortcut could still be
> added as breaking change in a future version.
>
> But now I'm wondering, how is the multi archive case handled if
> reading from stdin?
>
> To extend your example in the cover letter: `ssh host cmd |
> proxmox-backup-client backup a.img:/dev/stdin b.img:/dev/stdin`
>
> Without having tested it I guess this will fail on the second
> archive creation since that will be empty? 
Yes, this specific example will fail for that reason.

But in general, running something like `cat - test.txt -` or `cat
/dev/stdin test.txt /dev/stdin` actually works if you run it
interactively in a terminal and hit `ctrl+d` to end the first input.
>
> This should be checked and fail on command invocation instead. 
This is not trivial because of symlinks:

$ file /dev/stdin
/dev/stdin: symbolic link to /proc/self/fd/0

$ file /proc/self/fd/0
/proc/self/fd/0: symbolic link to /dev/pts/5

$ echo | file /proc/self/fd/0
/proc/self/fd/0: symbolic link to pipe:[317534]


I can think of two solutions:

1) Since we already read the metadata of all inputs at the start, we
could check that the inode number is unique for fifo files.

2) Simply check if /dev/stdin and/or /proc/self/fd/0 appear multiple
times as strings. In this case a warning might be better than an error.


I also just realized that my patch didn't actually work interactively,
because pts devices are not fifo pipes. Somehow I've only ever tested
this with my script or non-interactive pipes.

I was able to fix this by allowing character devices, but initially I
got some really weird behavior. For example, a single ctrl+D was not
enough to end the stream and I also got errors about multiple smaller
chunks being sent. The problem was that FixedChunkStream continued to
call (try_)poll_next even after it got a None, which is not allowed.

>
> > The only real advantage of "-" is that it could be implemented in
> a> platform independent way by reading from io::stdin(). That would
>> require a lot of changes though.
>>
>>>
>>>> +        } else {
>>>> +            filename
>>>> +        };
>>>> +
>>>>            use std::os::unix::fs::FileTypeExt;
>>>>              let metadata = std::fs::metadata(&filename)
>>>
>