Incorrect dependency resolution if package provider is not in pmaports

Related: https://gitlab.com/postmarketOS/pmaports/issues/404

Building mesa-git fails at the moment:

$ pmbootstrap -v build mesa-git --arch aarch64 --force
[18:36:59] (native) install alpine-base
[18:37:00] (native) install qemu-aarch64
[18:37:00] Register qemu binfmt (aarch64)
[18:37:00] (buildroot_aarch64) install alpine-base
[18:37:02] (buildroot_aarch64) install abuild build-base ccache git
[18:37:09] NOTE: You can edit the 'arch=' line inside the APKBUILD
[18:37:09] NOTE: Alternatively, use --arch to build for another architecture ('pmbootstrap build --arch=armhf libdrm-grate')
[18:37:09] ERROR: Can't build 'libdrm-grate' for architecture aarch64
[18:37:09] See also: <https://postmarketos.org/troubleshooting>
Run 'pmbootstrap log' for details.

pmbootstrap log: log.txt

It works fine with build --strict but I believe abuild handles the dependency resolution in that case.

If I understand the dependency resolution correctly, the following happens: (see pmb/parse/depends.py)

Resolve dependency libdrm:

1. package = package_from_aports(args, pkgname_depend = 'libdrm')
   Look for an pmaport providing libdrm. Returns libdrm-grate, because libdrm is only in Alpine.
2. package = package_from_index(args, pkgname_depend = 'libdrm', pkgnames_install, package_aport = {...'libdrm-grate'...}, suffix)
   Look for binary packages providing libdrm.
   1. provider = package_provider(args, pkgname_depend = 'libdrm', pkgnames_install, suffix)
      Pick libdrm from Alpine as the shortest provider.
   2. if (package_aport and pmb.parse.version.compare(package_aport["version"], provider["version"]) == 1):
      Compare libdrm version 2.4.100-r0 with libdrm-grate version 2.4.100_git20191221-r0
   3. libdrm binary is outdated, return libdrm-grate aport as provider.

So we are comparing package versions that don't really have anything to do with each other.

A naive fix would be:

diff --git a/pmb/parse/depends.py b/pmb/parse/depends.py
index 3010d505..39decc56 100644
--- a/pmb/parse/depends.py
+++ b/pmb/parse/depends.py
@@ -105,6 +105,13 @@ def package_from_index(args, pkgname_depend, pkgnames_install, package_aport,
     if not provider:
         return package_aport
 
+    # Check if we actually have an aport for the package we want to install
+    if package_aport and provider["pkgname"] != package_aport["pkgname"]:
+        logging.verbose(pkgname_depend + ": selected provider is " +
+                        provider["pkgname"] + " but the aport is for " +
+                        package_aport["pkgname"])
+        return provider
+
     # Binary package outdated
     if (package_aport and pmb.parse.version.compare(package_aport["version"],
                                                     provider["version"]) == 1):

to make sure the aport is actually for the binary package we have chosen. But I'm not sure if this works correctly if the provider is a sub-package.

Overall, (in my opinion) the current approach to resolve dependencies, i.e.

1. Look for a single pmaport first (there could be several ones providing the same package!)
2. Pick a binary package provider and compare the version with the (potentially unrelated) aport

doesn't really work correctly, especially if there are multiple providers.

I would suggest something like:

1. Find all package providers
   1. Add binary package providers to the list
   2. Add aport package providers to the list (note: also need to consider arch)
      • If the package is already in the list
        • Replace it? I.e. assume the aport is more up-to-date OR
        • Compare the versions and pick latest
      • Otherwise simply add as provider to the list (= no binary package)
2. Pick a provider with the existing logic in package_provider()
3. Return it. Now we have either a binary package provider or an aport, because both were in the same list.

Does that make sense?