Re: [PATCH for-maint] apply: gitdiff_verify_name(): accept "/dev/null\r"

Laszlo Ersek <lersek@redhat.com> writes:

This is not entirely correct, though. But it suggests an avenue for a possible enhancement.

Yes, but that is within MTAs and is expected to be invisible at MUA level. Typically mailbox files are stored in platform's native format (e.g. I see LF endings in my mailbox), be it CRLF of LF.

The important thing to note here is that use of text/plain for patches, if you want to have distinction between CRLF and LF in your payload, is not designed to work over e-mails.

On a platform whose native line endings are LF, the stripping may not even happen (i.e. "am/apply" may not see CRLF in the mailbox in the first place).

On a platform whose mbox has CRLF, or if you "unix2dos" a mbox on a platform whose mbox has LF to manufacture a mbox with CRLF line endings, you would have an opposite issue. Information is lost in MTAs while your e-mail is in transit and you cannot tell patch to which paths had CRLF line endings and patch to which paths had LF line endings. If all files you are tracking uniformly use CRLF, you can assume that everything had CRLF but it is still an assumption and that is a special case that is not particularly interesting.

Applying such a CRLF patch with your "/dev/null\r\n" patch applied to "am" will _add_ unwanted CR before LF to files that are meant to use LF line endings.

Now if we accept that this issue is coming from lossy nature of transporting patches via e-mails, we would realize that the right place to solve this is not in "apply"'s parsing of structural part of the "diff" output (e.g. "diff --git ...", "rename from ..." or "--- filename"), but the payload part (i.e. " " followed by context, "-" followed by removed and "+" followed by added). Removal of CR by "am -> mailsplit -> mailinfo -> apply" callchain is not losing any information, as the input does not have useful information to let us answer "are the lines in this path supposed to end with CRLF?" in the first place; "/dev/null\r" patch is barking up a wrong tree.

Our line-endings infrastructure (e.g. core.autocrlf configuration variable, `eol` attribute) is all geared toward supporting cross platform projects in that the BCP is to use LF line endings as the canonical line endings for in-repository data and convert it to CRLF for working tree files when necessary. We do not have direct support (other than declaring contents for paths as "binary" aka "no conversion") to use CRLF in in-repository data (and that is partly deliberate).

But it is conceivable to enhance the attribute system to allow us to mark certain paths (or "all paths", which is a trivial special case) as using CRLF line endings in in-repository and in-working-tree. It could be setting `eol` attribute to `both-crlf` or something.

Then "am -> mailsplit -> mailinfo -> apply" chain could:

 * "mailsplit" and "mailinfo" does not have to do anything special,
   other than stripping CR and make sure "apply" only sees LF
   endings;

 * "apply" is taught a new option "--fix-mta-corruption" which
   causes it to pay attention to the `eol` attribute set to
   `both-crlf`, and when it reads a patch

	diff --git a/one b/one
        --- one
        +++ one
        @@ -4,6 +4,6 @@
         common1
	 common2
        -old1
        -old2
        +new1
        +new2
         common3
         common4

   and notices that path "one" is marked as such, it pretends as if
   the input were

	diff --git a/one b/one
        --- one
        +++ one
        @@ -4,6 +4,6 @@
         common1^M
	 common2^M
        -old1^M
        -old2^M
        +new1^M
        +new2^M
         common3^M
         common4^M

   to compensate for possible breakage during the mail transit.

 * "am" is taught to pass "--fix-mta-corruption" to "apply" perhaps
   by default.

Because code paths that internally run "git am", e.g. "git rebase", work on data that is *not* corrupt by mta (we generate diff and apply ourselves), these places need to tell "am" not to use the "--fix-mta-corruption" when running "apply".