On Sat, 6 Apr 2013 20:08:43 +0200 Michał Górny <mgo...@gentoo.org> wrote:
> Hello, > > As far as I'm aware, we don't really have much of a patch maintenance > policy in Gentoo. There a few loose rules like «don't put awfully big > files into FILESDIR» or the common sense «use unified diff», but no > complete and clear policy. > > Especially considering the late discussion related to the needless > and semi-broken functionality in epatch, I'd like to propose > setting the following rules for patches in tree and in Gentoo-sourced > patchsets: > > 1. Patches have to be either in unified or context diff format. Unified > diff is preferred. Yes. > 2. Patches have to apply to the top directory of the source tree with > 'patch -p1'. If patches are applied to sub-directories, necessary '-p' > argument shall be passed to 'epatch' explicitly. Developers are > encouraged to create patches which are compatible with 'git am'. Bikeshedding below. > 3. Patches have to end with either '.patch' or '.diff' suffix. Yes. > 4. If possible, patches shall be named in a way allowing them to be > applied in lexical order. However, this one isn't necessary if patches > from an older ebuild are applied to a newer one. Even more below. > 5. The patch name shall shortly summarize the changes done by it. Yes. > 6. Patch files shall start with a brief description of what the patch > does. Developers are encouraged to use git-style tags like 'Fixes:' to > point to the relevant bug URIs. Yes. > 7. Patch combining is discouraged. Developers shall prefer multiple > patches following either the upstream commits or a logical commit > sequence (if changes are not committed upstream). Kinda. > > The above-listed policy will apply to the patches kept in the gx86 tree > (in FILESDIRs) and patch archives created by Gentoo developers. They > will not apply to the patch archives created upstream. > > > RATIONALE > --------- > > The main reason for the whole policy is to let Gentoo supply users with > consistent and friendly patches. That is, patches which can be used > directly on a source tree or submitted upstream without any additional > actions from user. > > (1) lists the most common patch formats. The formats shall be generally > both readable and reusable whenever possible. I wanted just 'unified > diff' but ulm suggested that we should also support those upstreams who > use 'context diffs'. > > > (2) is likely to be a bikeshed point here. Long story short, epatch has > this fragile patchlevel guessing, users don't have it. If we keep our > patches consistent to a single patchlevel, we gain: > > * ability for users to apply the patches without having them try all > patchlevels by hand. > > * clean error output if patch stops to apply for some reason. > > * no risk that patch will get applied to the wrong file if patch stops > to apply at expected patchlevel and starts to apply on another. If I can't drop patches from an upstream mailing list with a different repo structure than source, or just a random user who doesn't know better, into my patchdir I'm going to be very grumpy. Our patches should be -p1, I agree. I respin any patches I find that aren't. The first two points are nice, but not worth it to me in particular. The third I don't see the risk. What are the chances of that patch actually applying successfully? It'll fail, and you root it out. And since all our own patches will be -p1 this becomes a tiny corner case. > Also, by creating git-format patches, we gain the ability of rebasing > and updating the patches easily. Even with non-git upstreams, we can > do: > > git init; git add -A; git commit -m 1; git am ... > > (3) should be mostly obvious. The main idea is that if we apply a whole > patchdir, we should be able to easily tell between patches > and auxiliary files like 'README' or Debian's 'series'. > > I have never seen a patch file named other than '*.patch' or '*.diff'. > Therefore, I think that it's better to just require those rather than > trying to provide a sane list of excludes. > > > (4) is mostly about friendliness (again). Since shell does filename > expansion in lexical order, it's just great for user to be able apply > patches like: > > git am /usr/portage/foo/bar/files/198-*.patch > > The other sentence is not to enforce this rule e.g. when the same patch > is applied to different versions of the same package. Although with > a fair of trickery that could be gotten working, I don't think it will > be user-friendly anymore :). That's useful but ${PN}-${PV}-desc.patch is the currently accepted convention. What you're suggesting here is a suggested workflow. I don't think it belongs in a policy doc, at least as it's own section, or at least until it becomes more commonplace. > (5) makes finding a particular patch of interest easier, while (6) > makes sure that the purpose of the patch can be read from patch alone. > In both cases, having described patches is much better than having to > look into ebuilds for explanations. > > > (7) is because merged patches are usually hard to read and completely > not suitable for submitting anywhere. Well, patches should do one thing, agreed. But I'm not adding 4 patches to respect LDFLAGS, use CFLAGS while linking, install into /usr, and fix the version number in a 30 line Makefile. I'm adding foo-0.01-makefile.patch which does one thing (fixes crappy-ass Makefile). I can live with "discouraged" though. -- gcc-porting toolchain, wxwidgets @ gentoo.org
signature.asc
Description: PGP signature