At Sat, 25 Aug 2012 14:43:32 -0400, Dave Reisner wrote:

On Sat, Aug 25, 2012 at 01:36:41PM -0400, Luke Shumaker wrote:

...

A while ago I started working on a derivative of makepkg to support having 'git://...' type urls in the sources=() array. When preparing to file this patch, I did a `git rebase`, and noticed that Allan McRae began working on a similar feature. Our implementations are in many ways similar. Hopefully mine will be useful.

An interesting approach. As you've noticed, I think we're fairly committed to the implementation that Allan has provided. If you have specific concerns, maybe we can work to fix those.

* Allan's URL parsing has some issues, see point 2 below. This is an easy fix. * Composability. It's not really a "problem", but it seems to me that It's better to create a new general-purpose tool, instead of shoving a bunch of "special case" behavior into makepkg. Allan's implementation could even fairly easily be pulled out of makepkg. And really, does it make sense to have these URL schemes hardcoded into makepkg? Why have DLAGENTS at all, if we're going to hard-code schemes into makepkg? Again, I'm not against Allan's imlementation, but I am for moving it out of makepkg.

...

My implementation makes minimal changes to makepkg itself (only adding blob expansion to DLAGENTS, allowing for things like "git+*::""). Instead I added a `vcsget` tool which generates a tarball from the VCS repo, in a very similar manner to the way Allan's implementation does so within makepkg.

I'm not thrilled with the shell I saw in that patch -- there's extensive use of external commands when simple bash features would have sufficed.

I assume you're speaking of my 1. use of cut in get_field 2. use of sed to parse URLs 3. use of sed to parse URL authorities 4. use of readlink to establish the tarball name 5. use of sed to create a basedir name from the tarball name 6. use of '[' with if statements My defense of each: 1. `cut` in get_field --------------------- This was simply the simplest solution that was obviously correct. Another solution would have been to manipulate IFS, which has all kinds of funny side effects and fringe behaviors. That was a weak argument, and as a coder, I would be thrilled to find out if there is a better solution. 2. `sed` to parse URLs ---------------------- I wanted full URL parsing, for a few cases when the translation is less than than straight-forward. For example, including authentication info in svn URLs. Further, I wanted to make sure that *any* URL parsing done would be robust. Given that the regex used was taken straight from the relevent RFC, I knew I could count on it to work, even in fringe cases. For example a "fragment" *should* be able to contain a '#', but Allan's implementation discards anything before the last '#' in the fragment. This is a problem for (at least) git, as tags and branches may contain the symbol. 3. `sed` to parse URL authorities --------------------------------- I believe that a robust (not having the same problems as fragments for URLs) implementation in bash would be non-trivial. 4. `readlink` to establish the tarball name ------------------------------------------- I used `readlink -m` to turn the tarball name into an absolute path. This was not an absolute must, but it allowed me to avoid worrying about keeping track of the current directory. It would be fairly easy to remove this by useing pushd/popd instead of cd. If you would like this changed, I can do that. 5. `sed` to create a basedir name from the tarball name ------------------------------------------------------- I'll admit, this was me getting a little lazy. A pure bash implementation is: base=${tarball##*/} base=${base%%.tar.*} base=${base%.tar} 6. `[` with if statements ------------------------- Can I call stylistic choice on this one? It is trivial to replace this with bash's '[['.

...

It looks as if Allan's download_*() functions are more verbose than mine about what failed when there is an error. His svn and hg handlers are likely more robust--though my git is pretty solid. I also have a half-written handler for for bzr.

An advantage of my design is that it does allow for integrity checks of VCS packages, rather than inserting 'SKIP' into the md5sums array. This is very important to the derivative distribution Parabola. (However, the 'SKIP' option is still valuable for URLs that track a branch)

I don't see this as an advantage so much as a duplication. The backing VCS takes care of integrity checks. They're only necessary with tarballs because there is no "builtin" checksum to reference.

This is only partially true of the non-distributed VCSs. There was concern from other Parabola developers about being able to use the checksum to unambiguously be able to confirm that a source tree is the correct version, but Allan's implementation seems acceptable on that front. Happy hacking, ~ Luke Shumaker

At Mon, 27 Aug 2012 10:53:32 -0400, Dave Reisner wrote:

On Mon, Aug 27, 2012 at 09:55:57AM -0400, Luke T.Shumaker wrote:

...

At Sat, 25 Aug 2012 14:43:32 -0400, Dave Reisner wrote:

...

On Sat, Aug 25, 2012 at 01:36:41PM -0400, Luke Shumaker wrote:

...

A while ago I started working on a derivative of makepkg to support having 'git://...' type urls in the sources=() array. When preparing to file this patch, I did a `git rebase`, and noticed that Allan McRae began working on a similar feature. Our implementations are in many ways similar. Hopefully mine will be useful.

An interesting approach. As you've noticed, I think we're fairly committed to the implementation that Allan has provided. If you have specific concerns, maybe we can work to fix those.

* Allan's URL parsing has some issues, see point 2 below. This is an easy fix.

Great, let's fix those.

...

* Composability. It's not really a "problem", but it seems to me that It's better to create a new general-purpose tool, instead of shoving a bunch of "special case" behavior into makepkg. Allan's implementation could even fairly easily be pulled out of makepkg.

-1 for moving it out of makepkg. Decoupling it means we would need to do a lot of sharing so that the tool is sure to understand makepkg's absurdly complicated environment. Not extensible, and there's a lot of code that would be "duplicated".

...

And really, does it make sense to have these URL schemes hardcoded into makepkg? Why have DLAGENTS at all, if we're going to hard-code schemes into makepkg? Again, I'm not against Allan's imlementation, but I am for moving it out of makepkg.

Let's see what's already established and available for http downloading:

- curl - wget - aria2 - dog - fetch - bash tcp sockets (im sure there's more)

Show me a popular reimplementation of the reference git/svn/bzr/cvs/darcs/hg client, and then we'll talk about a way to make it possible for the user to configure that in place of our "hardcoded default".

I can't speak for the others, but there are multiple implementations of git (though I am only personally familiar with the original).

The nerd in me wants to point out that there is no backing RFC for any of the VCS protocols so it doesn't belong in DLAGENTS, but we've violated that puritanism already by adding a DLAGENT for scp.

Regardless, we *are* mapping the VCS tools to URLs. There are plenty of URL schemes in common use that aren't registered with IANA (per RFC4395/BCP35). Another effect of the Allan's current implementation is that adding another scheme requires modification to makepkg, event if it is just to pass it to DLAGENTS. I'd argue that even if we have the special cases, the correct behavior would be to have download_file (DLAGENTS) as a fallback, instead of only tripping on 'ftp|http|https|rsync|scp'. Further, a use case I see is a user putting some funky stuff in DLAGENTS to either use an offline cache, or do some network stuff to deal with a contrived network setup. Also, the nerd in me wants to point out that the makepkg source always refers to it as a 'protocol' (or 'proto') when RFC3986 says that it should be 'scheme'.

...

...

...

My implementation makes minimal changes to makepkg itself (only adding blob expansion to DLAGENTS, allowing for things like "git+*::""). Instead I added a `vcsget` tool which generates a tarball from the VCS repo, in a very similar manner to the way Allan's implementation does so within makepkg.

I'm not thrilled with the shell I saw in that patch -- there's extensive use of external commands when simple bash features would have sufficed.

I assume you're speaking of my 1. use of cut in get_field 2. use of sed to parse URLs 3. use of sed to parse URL authorities 4. use of readlink to establish the tarball name 5. use of sed to create a basedir name from the tarball name 6. use of '[' with if statements

Probably a good laundry list to start with. I'm happy to give a proper review of vcsget if you'd like, but I'll respond here in general terms.

That would be great!

...

My defense of each:

1. `cut` in get_field ---------------------

This was simply the simplest solution that was obviously correct. Another solution would have been to manipulate IFS, which has all kinds of funny side effects and fringe behaviors.

That was a weak argument, and as a coder, I would be thrilled to find out if there is a better solution.

Cut is useless. IFS manipulation with read is the preferred solution here. I'm curious what you think are the "fringe behaviors" in altering IFS (I suspect you're merely misusing it -- it's a common point of confusion). I've yet to come across any that make me want to use cut. I'd sooner offload a larger portion of work to awk then use cut for several reasons, anyways.

I'll give you that awk might be better--but I was never comfortable with awk. The problem with IFS that stopped me from using it is that it always* will collapse multiple whitespace characters, so it is impossible to have an empty column in the middle if useing a ' ' delimiter. For some reason I was thinking that space was the only safe (printable ASCII) character to use--I now realize that both '<' and '>' would work fine. * perhaps there is a shopt option, but I am unaware of it

...

2. `sed` to parse URLs ----------------------

I wanted full URL parsing, for a few cases when the translation is less than than straight-forward. For example, including authentication info in svn URLs.

Further, I wanted to make sure that *any* URL parsing done would be robust. Given that the regex used was taken straight from the relevent RFC, I knew I could count on it to work, even in fringe cases. For example a "fragment" *should* be able to contain a '#', but Allan's implementation discards anything before the last '#' in the fragment. This is a problem for (at least) git, as tags and branches may contain the symbol.

bash has regex matching which supports ERE with backreferences -- this could easily be adopted to split the URL all at once, simply saving the relevant pieces of BASH_REMATCH after the fact. See the =~ operator in bash(1).

I was unaware of that feature. It would save me a lot of effort.

...

3. `sed` to parse URL authorities ---------------------------------

I believe that a robust (not having the same problems as fragments for URLs) implementation in bash would be non-trivial.

See above.

...

4. `readlink` to establish the tarball name -------------------------------------------

I used `readlink -m` to turn the tarball name into an absolute path. This was not an absolute must, but it allowed me to avoid worrying about keeping track of the current directory. It would be fairly easy to remove this by useing pushd/popd instead of cd. If you would like this changed, I can do that.

This is _not_ portable. We've refrained from using readlink at all in our codebase except in the untainted form. The only thing you can portably rely on readlink to do is to wrap the readlink syscall. Better yet, figure out why a potential symlink isn't good enough and fix that. You shouldn't need to know the actual target that a symlink points to, since 90% of a tool will dereference it for you.

My use of `readlink` had nothing to do with symlinks; it is simply to turn a relative path into an absolute path. An 'unintended' use of readlink, but the UNIX authors would say that so is `cat` to display a single file is too. I've seen this in several codebases, and is the most reliable method of doing this in shell, assuming GNU coreutils. As I said, I could have avoided this by being more careful with `cd` and doing a little more directory hopping.

...

5. `sed` to create a basedir name from the tarball name -------------------------------------------------------

I'll admit, this was me getting a little lazy. A pure bash implementation is:

base=${tarball##*/} base=${base%%.tar.*} base=${base%.tar}

And with an extglob, you can more precisely cut that down to:

base=${tarball##*/} base=${base%.tar?(.*)}

The unfortunately named foo.tar.bar.tar.gz wouldn't be "broken" in this way.

I was unaware of extglob offering that. Dang these shopt features I have to turn on! However, given that input, I would have thought that 'foo' is the correct output, so that's what I wrote.

...

6. `[` with if statements -------------------------

Can I call stylistic choice on this one? It is trivial to replace this with bash's '[['.

Apologies for the incoming rant.

No, this isn't merely an issue of "style". [[ is far superior to [ in several ways. It's a bash keyword, making it far faster than the [ builtin, it supports pattern matching with globs and regex, and it has defined behavior when you pass more than 3 arguments. Quoting semantics are far simpler (and safer), too. Consider a contrived example which, actually, you see far too often:

[ -n $foo ]

This is always true, regardless of whether or not 'foo' is defined. since '[' is a builtin, which means it behaves no different than another other standard command. In reality, it's the same as:

[ -n ]

This is _true_, because you've really only passed a single argument to [. In the single argument case, the existance test is performed. Since '-n' always has a length, it's true.

By comparison, the preferred bash construction:

[[ -n $foo ]]

This always does the right thing because the lexer passes over it and sees the variable before its expanded. It understands that even though foo might not be defined, you've passed an argument and exits with an error when the argument is the empty string. The only time you need to quote arguments to tests is when dealing with equations:

[[ $foo = $bar ]]

is not the same as

[[ $foo = "$bar" ]]

Generally, you want the latter since the RHS (right hand side) is subject to glob expansion. Quoting it ensures that any glob characters are not expanded. The LHS is _never_ subject to glob expansion, but will require quoting if it contains whitespace.

OK, 'style' was the wrong word. I am in the habbit of using '[' from having to code for Almquist-family shells, including dash, busybox, and BDS's sh. I realize the performance advantage of '[[' for Bash. For every way I used it (and I know this isn't always the case), the two are identical in result. As for quoting rules--I always disliked the 'simpler' rules for '[[', the seem like they were trying to compensate for bad programmers, (in the first example) I'd have to quote the variable if I were to do a similar task with any other command. ---- Well, I have been taught a few features of Bash, and been made to feel like I noob. Thank you, this is learning! Happy hacking, ~ Luke Shumaker

On 28/08/12 10:17, Luke T.Shumaker wrote:

Another effect of the Allan's current implementation is that adding another scheme requires modification to makepkg, event if it is just to pass it to DLAGENTS. I'd argue that even if we have the special cases, the correct behavior would be to have download_file (DLAGENTS) as a fallback, instead of only tripping on 'ftp|http|https|rsync|scp'.

Fix on its way.

Further, a use case I see is a user putting some funky stuff in DLAGENTS to either use an offline cache, or do some network stuff to deal with a contrived network setup.

This is the only thing I think needs dealt with... maybe... Can you give a concrete example? Thanks, Allan