On Fri, Jan 11, 2019 at 12:11:00AM -0600, Steve French wrote:
> In discussing an interesting problem with scp recently with Pavel, we
> found a fairly obvious bug in scp when it is run with a progress
> indicator (which is the default when the source file is remote).
>
> scp triggers "SIGALRM" probably from update_progress_meter() in
> progressmeter.c when executed with "scp localhost:somelargefile /mnt"
> ie with an ssh source path but a local path for the target, as long as
> the flush of the large amount of cached data to disk (which seems to
> be triggered by the ftruncate call in scp.c) takes more than a few
> seconds (which can be common depending on disk or network speed).
Are you saying the SIGALRM interrupts ftruncate() and causes the ftruncate
to fail?
> It is interesting that this can be avoided easily by running in quiet
> mode ("-q") which disables the progress meter in scp, but it seems
> very, very broken that scp of a large file can 'fail' when using a
> progress meter (unless caching were disabled on the file system) due
> to the slight delay in ftruncate due to triggering a flush of a large
> amount of cached data to disk sequentially.
>
> Any thoughts of whether scp is actively maintained and best approach
> to fix progressmeter.c so it doesn't break on Linux?
>
> --
> Thanks,
>
> Steve
