Going to true paths is expensive and a lot of work, so it does this
only when requested. (It has to do a check at every directory level).
In subversion 1.5 we had some accidental truepath checks which were
directly visible as bottleneck in the profiler.
Bert
Bert Huijben (Cell phone) From: Daniel Shahaf
Sent: donderdag 21 april 2011 17:33
To: Johan Corveleyn
Cc: Philip Martin; Branko Čibej; dev@subversion.apache.org
Subject: Re: Case-only renames on Windows (issue #3702)
Johan Corveleyn wrote on Thu, Apr 14, 2011 at 23:49:16 +0200:
> On Tue, Apr 12, 2011 at 11:14 PM, Johan Corveleyn <jcor...@gmail.com> wrote:
> > On Fri, Mar 25, 2011 at 2:21 PM, Johan Corveleyn <jcor...@gmail.com> wrote:
> >> On Sun, Mar 20, 2011 at 9:32 PM, Johan Corveleyn <jcor...@gmail.com> wrote:
> >>> Some thoughts:
> >>>
> >>> - There is only a problem if the dst_path gets case-normalized to one
> >>> of the source paths. Otherwise, the case-normalization really does
> >>> need to happen.
> >>
> >> Hm, this thought may be incorrect (or at least "unexpected" for
> >> windows users). In fact, if we look at how the native "move" behaves,
> >> it never case-normalizes the target path:
> >>
> >> [[[
> >> C:\Temp>dir /B test
> >> TODO
> >>
> >> C:\Temp>echo anothertest > bla
> >>
> >> C:\Temp>move bla test\toDO
> >> Overwrite C:\Temp\test\toDO? (Yes/No/All): y
> >> 1 file(s) moved.
> >>
> >> C:\Temp>dir /B test
> >> toDO
> >>
> >> C:\Temp>type test\todo
> >> anothertest
> >> ]]]
> >>
> >> So it seems that, if we want "svn mv" to behave more like "move" on
> >> Windows, the target path should never be normalized to the on-disk
> >> casing. Just use it as is ...
> >
> > Finally getting around to this again ... the above is still an open
> > question for specifying the desired behavior of "svn move" on Win32
> > ... If there is a file A on disk, what should "svn move B a" do?
> >
> > In theory it could generate a replace of A by a ("copied from" B).
> > That would be consistent with behavior of Windows' "move" command,
> > after confirming the overwrite.
> >
> > But, lacking the "confirm-for-overwrite", I guess the behavior most
> > consistent with existing svn functionality would be to refuse this
> > move in the same way as "svn move B A" is refused. Currently, this is
> > with the (slightly not-to-the-point) error message: "svn: E155007:
> > Path 'C:\temp\bla\A' is not a directory".
> >
> > Would that be ok?
>
> I think this would be ok: just refuse a move that overwrites another
> local file (need to rm it first).
>
> Now, I just realized that case-only renames (or renames in general)
> only come into play when there are exactly 2 "targets" of the move
> command: the source and the destination (ok, I'm a bit slow, I know
> :-)).
>
> So, for fixing issue #3702 (case-only renames on Windows), if we take
> into account the rest that's been discussed in this thread, we only
> need to do something special:
> - If there are exactly 2 targets
> - If those targets are paths, rather than uris.
> - If normalized dst_path == normalized src_path
>
> In that case I think we should either:
>
> 1) Normalize the original dst_path again, but this time without the
> APR_FILEPATH_TRUENAME flag to apr_filepath_merge, so it isn't
> converted to on-disk-representation.
>
Call me a pedant, but:
Does APR guarantee that it will *not* canonicalize-to-truename if
APR_FILEPATH_TRUENAME is not passed?
That does not follow logically from "APR will canonicalize-to-truename
if APR_FILEPATH_TRUENAME is passed".
> or
>
> 2) Just use the original dst_path as it was passed to the client,
> un-normalized.
>
>
> I have no idea if it should be 1) (other important normalization-stuff
> going on besides converting it to the on-disk-representation, that
> should still be performed in this case) or 2) ("if the user tells us
> to use this particular representation of the path, let's use it").
>
> Option 2) would of course be much easier. For option 1) code needs to
> be written/copy-pasted/refactored to do all the normalization stuff,
> except passing the flag to apr_filepath_merge.
