On Tue, Apr 12, 2011 at 7:00 PM, Sebastian Nowicki wrote:

On Tue, Apr 12, 2011 at 2:46 AM, Dan McGee wrote:

...

On Sun, Apr 10, 2011 at 5:58 AM, Sebastian Nowicki wrote:

...

In addition to the names of the conflicting packages, the origin and versions will be displayed to the user.

This introduces an API change in the PM_TRANS_CONV_CONFLICT_PKG conversation callback. All three arguments were strings. The first two arguments are now pmpkg_t structures (sync and local, respectively).

Fixes FS#12536 ---

In FS#12536 there was some discussion about changing pmconflict_t to contain pmpkg_t structures rather than just strings. This would be a much larger change, which I didn't think was worthwhile for a cosmetic change.

Even so, this does introduce an API change, and a sneaky one at that. The conversation callback had three string arguments, now it has two pmpkg_t and a string. Due to these being all void pointers, end-users won't get any compiler warnings/errors if they don't update their code!

Any consideration to just building the relevant string in the backend so we don't need 1) an API change and 2) the awfully ugly code in the frontend to print what you are printing?

This would work for pacman, but I don't think it's ideal for other front-ends (especially GUI). As a front-end developer, I'd prefer having the data in the callback rather than pre-processed information.

On second thought, this would be a good compromise. If front-ends require the pmpkg_t structure, it can be added later on. There's little harm in changing the strings, although, in a way, this would still be an API change (if the front-end relies on the string being the name for some reason).

On Sun, Apr 17, 2011 at 5:03 PM, Sebastian Nowicki wrote:

In addition to the names of the conflicting packages, the origin and versions will be displayed to the user.

This introduces a slight API change in the PM_TRANS_CONV_CONFLICT_PKG conversation callback. The format of the first two strings has changed from package names to strings of the format "db/name-version".

Fixes FS#12536 ---

I rewrote the patch to pre-format the package information rather than changing the API completely, as per Dan's suggestion. This still introduces a bit of an API change as the two string can no longer be used as package identifiers, e.g. in alpm_db_get_pkg().

$ sudo ./src/pacman/pacman -S qemu-kvm resolving dependencies... looking for inter-conflicts... :: extra/qemu-kvm-0.14.0-1 and local/qemu-0.14.0-1 are in conflict (qemu). Remove local/qemu-0.14.0-1? [y/N] n

Hmm, just noticed this breaks the pkgname == data3 comparison:

if(strcmp(data1, data3) == 0 || strcmp(data2, data3) == 0) {

So the ".. are in conflict (package_name)" bit will always be shown in the pacman callback. Not sure how to work around this other than to modify the reason to have the same formatting as well. It seems odd that the package name is given in general, but I don't know the reasons why it was done that way.

On Sun, Apr 17, 2011 at 5:50 PM, Rémy Oudompheng wrote:

On Sun 17 April 2011 at 17:03 +0800, Sebastian Nowicki wrote:

...

In addition to the names of the conflicting packages, the origin and versions will be displayed to the user.

This introduces a slight API change in the PM_TRANS_CONV_CONFLICT_PKG conversation callback. The format of the first two strings has changed from package names to strings of the format "db/name-version".

Fixes FS#12536 ---

I rewrote the patch to pre-format the package information rather than changing the API completely, as per Dan's suggestion. This still introduces a bit of an API change as the two string can no longer be used as package identifiers, e.g. in alpm_db_get_pkg().

$ sudo ./src/pacman/pacman -S qemu-kvm resolving dependencies... looking for inter-conflicts... :: extra/qemu-kvm-0.14.0-1 and local/qemu-0.14.0-1 are in conflict (qemu). Remove local/qemu-0.14.0-1? [y/N] n error: unresolvable package conflicts detected error: failed to prepare transaction (conflicting dependencies) :: qemu-kvm and qemu are in conflict

 lib/libalpm/sync.c |   51 ++++++++++++++++++++++++++++++++++++++++++++++++---  1 files changed, 48 insertions(+), 3 deletions(-)

I am totally against this. Really really totally.

I personally think a formatting function for package information has strictly nothing to do in sync.c, and maybe not even in libalpm. Second, it would introduce additional parsing effort from front-ends to split against the passed string if they want to display information in a different way.

Agreed, but the alternatives require changes to the API (char * -> pmpkg_t).

And what output do you expect if a conflict arises when running pacman -U ?? What do you think is the repository from a package loaded from a file.

Just the package name and version. Having a "file/" prefix would be meaningless, and if you're installing from a single file, you're likely to know what you're installing anyway.

I don't even understand why we would want to display any sort of version information. I don't see how it would help me answer yes or no to that conflict question either.

That is in my mind the job of the front-end: if you really want information, it should display the list of selected packages (repository/pkgname-pkgver) that comes between the parsing of the command-line and the beginning of the transaction.

For example:

cb_trans_conv (from pacman) should be patched to display :  :: target qemu-kvm conflicts with installed qemu. Remove qemu ?

which is perfectly clear: qemu is installed and pacman wants you to remove it, "local/qemu" brings no additional information, and the version number is usually meaningless. The callback returns precisely one local package and one target package, it can be reflected in the user output.

If the version of the local package is newer than the one being installed, it would be good to inform the user (as would have been likely with the libdrm-git package). The "local/" prefix can easily be omitted.

If you need to display the version number, then I don't feel like messing with libalpm's internals is the right answer. I would probably prefer have process_targname() patched (in pacman/sync.c), so that after running alpm_find_dbs_satisfier(), it prints something.

I think that changing the API of the callback to use pmpkg_t, (or even pmconflict_t if that's exposed) rather than a string, is the right thing to do. The front-end shouldn't have to find the package again to get its metadata (what if there are multiple packages with the same name?). For instance a GUI front-end may want to show the description of the package or something. It's a pretty drastic change considering the result is cosmetic.

However, I think modifying the callback to return pmpkg_t is not a bad idea, but then the formatting function should go in pacman. I just want to understand the first reason why we are doing this. The bug report is totally inconsistent (the title is inconsistent with the described problem and so on), and it seems we are trying to solve some imaginary problem.

The best thing we can probably do is make libalpm functions return more information about their internal processing. Modifying formatting will not help that. And if people want to know why some package suddenly appears in the target list, I fear there is no solution.

In which case the API needs to change. I don't think cosmetic changes are worth doing so, but that's a decision for the devs.

On Mon, Apr 18, 2011 at 12:49 AM, Rémy Oudompheng wrote:

On Sun 17 April 2011 at 23:40 +0800, Sebastian Nowicki wrote:

...

On Sun, Apr 17, 2011 at 5:50 PM, Rémy Oudompheng wrote:

...

I am totally against this. Really really totally.

I personally think a formatting function for package information has strictly nothing to do in sync.c, and maybe not even in libalpm. Second, it would introduce additional parsing effort from front-ends to split against the passed string if they want to display information in a different way.

Agreed, but the alternatives require changes to the API (char * -> pmpkg_t).

I think we disagree in what we call an API change. All the suggested patches are API changes in the sens that they change the behaviour of a function. Even if its type signature does not change, the behaviour changes in the sense that the nature of what is returned changes.

As you already noticed, going from a simple package name to a complex string changes a lot of things when considering what a user (developer) may want to do with it.

The only advantage is keeping binary compatibility for several applications, e.g. pacman, but that is a coincidence.

We do agree. I mentioned that it's an API change in the patch. I guess I was referring to ABI rather than API as only the semantics changed in this patch.

...

...

And what output do you expect if a conflict arises when running pacman -U ?? What do you think is the repository from a package loaded from a file.

Just the package name and version. Having a "file/" prefix would be meaningless, and if you're installing from a single file, you're likely to know what you're installing anyway.

Is there is a reason you don't know what you are installing in other cases? Possible from the point of view of the user, not very improbable from the point of view of the program using libalpm.

The best example is `pacman -Syu` in which case hundreds of packages can be updated. For me this is the most common instance of replaces taking affect.

...

...

I don't even understand why we would want to display any sort of version information. I don't see how it would help me answer yes or no to that conflict question either.

That is in my mind the job of the front-end: if you really want information, it should display the list of selected packages (repository/pkgname-pkgver) that comes between the parsing of the command-line and the beginning of the transaction.

For example:

cb_trans_conv (from pacman) should be patched to display :  :: target qemu-kvm conflicts with installed qemu. Remove qemu ?

which is perfectly clear: qemu is installed and pacman wants you to remove it, "local/qemu" brings no additional information, and the version number is usually meaningless. The callback returns precisely one local package and one target package, it can be reflected in the user output.

If the version of the local package is newer than the one being installed, it would be good to inform the user (as would have been likely with the libdrm-git package). The "local/" prefix can easily be omitted.

I don't see why the version information is particularly more relevant than other things. Personnally I'm interested in knowing whether my local package comes from an official repository, if so, whether it's in testing... Things that still not solved using such a partial approach.

I never stated it was more relevant. The origin is already shown. What other information are you referring to?

Does knowing the version really helps you in answering yes or no ???

This is what the feature request was for.

...

...

If you need to display the version number, then I don't feel like messing with libalpm's internals is the right answer. I would probably prefer have process_targname() patched (in pacman/sync.c), so that after running alpm_find_dbs_satisfier(), it prints something.

I think that changing the API of the callback to use pmpkg_t, (or even pmconflict_t if that's exposed) rather than a string, is the right thing to do. The front-end shouldn't have to find the package again to get its metadata (what if there are multiple packages with the same name?).

The front-end doesn't have to do anything magic to find what packages are in the transaction. Simple calls to alpm_find_satisfier() and alpm_trans_get_add() suffice to find the culprit package (even if of course having the whole pmpkg_t structure would be better).

In my opinion this is only a workaround to prevent an API change (which is reasonable). It still requires traversing through lists in order to find the package. It may not be "magic", but it is unnecessary effort to find what is already available in the _alpm_sync_prepare() function. I'll write up a patch if this is the preferred method though.