Fix gitsm networking and mirroring
PATCH 1/1 Fix gitsm networking and mirroring
gitsm would honor the BB_NO_NETWORK and mirroring for the primary repository, but submodules would always attempt to hit the network.
This code reworks the git submodule fetcher by avoiding the "git submodule init".
The code has been tested using my local use-case, as well as the built-in test suite. (You'll notice I had to adjust the test suite slightly in the patch.)
The change is available at:
git://git.openembedded.org/bitbake-contrib mgh/gitsm
Mark Hatle (1):
fetch2/gitsm.py: Rework the git submodule fetcher
lib/bb/fetch2/gitsm.py | 271 +++++++++++++++++++++++++++---------------------- lib/bb/tests/fetch.py | 3 + 2 files changed, 151 insertions(+), 123 deletions(-)
Signature · 1 line
1.8.3.1
PATCH 1/1 fetch2/gitsm.py: Rework the git submodule fetcher
The prior fetcher did not know how to work with MIRRORS, and did not honor BB_NO_NETWORK and similar.
The new fetcher approach recursively calls 'gitsm' download on each submodule detected. This ensures that it will go throug the standard download process.
Each downloaded submodule is then 'attached' to the original download in the 'modules' directory. This mimics the behavior of:
git submodule init
but there is no chance it will contact the network without permission.
It then corrects upstream reference URIs.
The unpack steps simply copies the items from the downloads to the destdir. Once copied the submodules are connected and we then run:
git submodule update
According to the git documentation, git submodule init can and will modify the project configuration and may connect to the network. Doing the work manually prevents this. (This manual process is allowed based on my reading of the documentation.)
See: https://git-scm.com/book/en/v2/Git-Tools-Submodules
The small change to the existing test is due to this new code always assuming the code is from a remote system, and not a 'local' repository. If this assumption proves to be incorrect -- code will need to be added to deal with local repositories without an upstream URI.
--- lib/bb/fetch2/gitsm.py | 271 +++++++++++++++++++++++++++---------------------- lib/bb/tests/fetch.py | 3 + 2 files changed, 151 insertions(+), 123 deletions(-)
lib/bb/fetch2/gitsm.py +148 -123
@@ -34,6 +34,8 @@ import bb from bb.fetch2.git import Git from bb.fetch2 import runfetchcmd from bb.fetch2 import logger +from bb.fetch2 import Fetch +from bb.fetch2 import BBFetchException class GitSM(Git): def supports(self, ud, d):
@@ -42,96 +44,66 @@ class GitSM(Git): """ return ud.type in ['gitsm'] - def uses_submodules(self, ud, d, wd): + def update_submodules(self, ud, d): + submodules = [] + paths = {} + uris = {} + local_paths = {} + for name in ud.names: try: - runfetchcmd("%s show %s:.gitmodules" % (ud.basecmd, ud.revisions[name]), d, quiet=True, workdir=wd) - return True - except bb.fetch.FetchError: - pass - return False + gitmodules = runfetchcmd("%s show %s:.gitmodules" % (ud.basecmd, ud.revisions[name]), d, quiet=True, workdir=ud.clonedir) + except: + # No submodules to update + continue - def _set_relative_paths(self, repopath): - """ - Fix submodule paths to be relative instead of absolute, - so that when we move the repo it doesn't break - (In Git 1.7.10+ this is done automatically) - """ - submodules = [] - with open(os.path.join(repopath, '.gitmodules'), 'r') as f: - for line in f.readlines(): + module = "" + for line in gitmodules.splitlines(): if line.startswith('[submodule'): - submodules.append(line.split('"')[1]) + module = line.split('"')[1] + submodules.append(module) + elif module and line.strip().startswith('path'): + path = line.split('=')[1].strip() + paths[module] = path + elif module and line.strip().startswith('url'): + url = line.split('=')[1].strip() + uris[module] = url for module in submodules: - repo_conf = os.path.join(repopath, module, '.git') - if os.path.exists(repo_conf): - with open(repo_conf, 'r') as f: - lines = f.readlines() - newpath = '' - for i, line in enumerate(lines): - if line.startswith('gitdir:'): - oldpath = line.split(': ')[-1].rstrip() - if oldpath.startswith('/'): - newpath = '../' * (module.count('/') + 1) + '.git/modules/' + module - lines[i] = 'gitdir: %s\n' % newpath - break - if newpath: - with open(repo_conf, 'w') as f: - for line in lines: - f.write(line) - - repo_conf2 = os.path.join(repopath, '.git', 'modules', module, 'config') - if os.path.exists(repo_conf2): - with open(repo_conf2, 'r') as f: - lines = f.readlines() - newpath = '' - for i, line in enumerate(lines): - if line.lstrip().startswith('worktree = '): - oldpath = line.split(' = ')[-1].rstrip() - if oldpath.startswith('/'): - newpath = '../' * (module.count('/') + 3) + module - lines[i] = '\tworktree = %s\n' % newpath - break - if newpath: - with open(repo_conf2, 'w') as f: - for line in lines: - f.write(line) - - def update_submodules(self, ud, d, allow_network): - # We have to convert bare -> full repo, do the submodule bit, then convert back - tmpclonedir = ud.clonedir + ".tmp" - gitdir = tmpclonedir + os.sep + ".git" - bb.utils.remove(tmpclonedir, True) - os.mkdir(tmpclonedir) - os.rename(ud.clonedir, gitdir) - runfetchcmd("sed " + gitdir + "/config -i -e 's/bare.*=.*true/bare = false/'", d) - runfetchcmd(ud.basecmd + " reset --hard", d, workdir=tmpclonedir) - runfetchcmd(ud.basecmd + " checkout -f " + ud.revisions[ud.names[0]], d, workdir=tmpclonedir) - - try: - if allow_network: - fetch_flags = "" - else: - fetch_flags = "--no-fetch" - - # The 'git submodule sync' sandwiched between two successive 'git submodule update' commands is - # intentional. See the notes on the similar construction in download() for an explanation. - runfetchcmd("%(basecmd)s submodule update --init --recursive %(fetch_flags)s || (%(basecmd)s submodule sync --recursive && %(basecmd)s submodule update --init --recursive %(fetch_flags)s)" % {'basecmd': ud.basecmd, 'fetch_flags' : fetch_flags}, d, workdir=tmpclonedir) - except bb.fetch.FetchError: - if allow_network: - raise - else: - # This method was called as a probe to see whether the submodule history - # is complete enough to allow the current working copy to have its - # modules filled in. It's not, so swallow up the exception and report - # the negative result. - return False - finally: - self._set_relative_paths(tmpclonedir) - runfetchcmd("sed " + gitdir + "/config -i -e 's/bare.*=.*false/bare = true/'", d, workdir=tmpclonedir) - os.rename(gitdir, ud.clonedir,) - bb.utils.remove(tmpclonedir, True) + module_hash = runfetchcmd("%s ls-tree -z -d %s %s" % (ud.basecmd, ud.revisions[name], paths[module]), d, quiet=True, workdir=ud.clonedir) + module_hash = module_hash.split()[2] + + # Build new SRC_URI + proto = uris[module].split(':', 1)[0] + url = uris[module].replace('%s:' % proto, 'gitsm:', 1) + url += ';protocol=%s' % proto + url += ";name=%s" % module + url += ";qbareclone=1;nocheckout=1" + + ld = d.createCopy() + # Not necessary to set SRC_URI, since we're passing the URI to + # Fetch. + #ld.setVar('SRC_URI', url) + ld.setVar('SRCREV_%s' % module, module_hash) + + # Workaround for issues with SRCPV/SRCREV_FORMAT errors + # error refer to 'multiple' repositories. Only the repository + # in the original SRC_URI actually matters... + ld.setVar('SRCPV', d.getVar('SRCPV')) + ld.setVar('SRCREV_FORMAT', module) + + newfetch = Fetch([url], ld, cache=False) + newfetch.download() + local_paths[module] = newfetch.localpath(url) + + # Correct the submodule references to the local download version... + runfetchcmd("%(basecmd)s config submodule.%(module)s.url %(url)s" % {'basecmd': ud.basecmd, 'module': module, 'url' : local_paths[module]}, d, workdir=ud.clonedir) + try: + os.mkdir(os.path.join(ud.clonedir, 'modules')) + except OSError: + pass + if not os.path.exists(os.path.join(ud.clonedir, 'modules', paths[module])): + os.symlink(local_paths[module], os.path.join(ud.clonedir, 'modules', paths[module])) return True
@@ -147,56 +119,109 @@ class GitSM(Git): # Now check that the submodule histories are new enough. The git-submodule command doesn't have # any clean interface for doing this aside from just attempting the checkout (with network # fetched disabled). - return not self.update_submodules(ud, d, allow_network=False) + return not self.update_submodules(ud, d) def download(self, ud, d): Git.download(self, ud, d) if not ud.shallow or ud.localpath != ud.fullshallow: - submodules = self.uses_submodules(ud, d, ud.clonedir) - if submodules: - self.update_submodules(ud, d, allow_network=True) + self.update_submodules(ud, d) + + def copy_submodules(self, submodules, ud, destdir, d): + if ud.bareclone: + repo_conf = destdir + else: + repo_conf = os.path.join(destdir, '.git') + + if submodules and not os.path.exists(os.path.join(repo_conf, 'modules')): + os.mkdir(os.path.join(repo_conf, 'modules')) + + for module in submodules: + srcpath = os.path.join(ud.clonedir, 'modules', module) + modpath = os.path.join(repo_conf, 'modules', module) + + if os.path.exists(srcpath): + if os.path.exists(os.path.join(srcpath, '.git')): + srcpath = os.path.join(srcpath, '.git') + + target = modpath + if os.path.exists(modpath): + target = os.path.dirname(modpath) + + runfetchcmd("cp -fpLR %s %s" % (srcpath, target), d) + elif os.path.exists(modpath): + # Module already exists, likely unpacked from a shallow mirror clone + pass + else: + # This is fatal, as we do NOT want git-submodule to hit the network + raise bb.fetch2.FetchError('Submodule %s does not exist in %s or %s.' % (module, srcpath, modpath)) def clone_shallow_local(self, ud, dest, d): super(GitSM, self).clone_shallow_local(ud, dest, d) - runfetchcmd('cp -fpPRH "%s/modules" "%s/"' % (ud.clonedir, os.path.join(dest, '.git')), d) + # Copy over the submodules' fetched histories too. + repo_conf = os.path.join(dest, '.git') + + submodules = [] + for name in ud.names: + gitmodules = runfetchcmd("%s show %s:.gitmodules" % (ud.basecmd, ud.revision), d, quiet=True, workdir=dest) + + for line in gitmodules.splitlines(): + if line.startswith('[submodule'): + module = line.split('"')[1] + submodules.append(module) + + self.copy_submodules(submodules, ud, dest, d) def unpack(self, ud, destdir, d): Git.unpack(self, ud, destdir, d) - if self.uses_submodules(ud, d, ud.destdir): - runfetchcmd(ud.basecmd + " checkout " + ud.revisions[ud.names[0]], d, workdir=ud.destdir) + # Copy over the submodules' fetched histories too. + if ud.bareclone: + repo_conf = ud.destdir + else: + repo_conf = os.path.join(ud.destdir, '.git') - # Copy over the submodules' fetched histories too. - if ud.bareclone: - repo_conf = ud.destdir - else: - repo_conf = os.path.join(ud.destdir, '.git') - - if os.path.exists(ud.clonedir): - # This is not a copy unpacked from a shallow mirror clone. So - # the manual intervention to populate the .git/modules done - # in clone_shallow_local() won't have been done yet. - runfetchcmd("cp -fpPRH %s %s" % (os.path.join(ud.clonedir, 'modules'), repo_conf), d) - fetch_flags = "--no-fetch" - elif os.path.exists(os.path.join(repo_conf, 'modules')): - # Unpacked from a shallow mirror clone. Manual population of - # .git/modules is already done. - fetch_flags = "--no-fetch" - else: - # This isn't fatal; git-submodule will just fetch it - # during do_unpack(). - fetch_flags = "" - bb.error("submodule history not retrieved during do_fetch()") - - # Careful not to hit the network during unpacking; all history should already - # be fetched. - # - # The repeated attempts to do the submodule initialization sandwiched around a sync to - # install the correct remote URLs into the submodules' .git/config metadata are deliberate. - # Bad remote URLs are leftover in the modules' .git/config files from the unpack of bare - # clone tarballs and an initial 'git submodule update' is necessary to prod them back to - # enough life so that the 'git submodule sync' realizes the existing module .git/config - # files exist to be updated. - runfetchcmd("%(basecmd)s submodule update --init --recursive %(fetch_flags)s || (%(basecmd)s submodule sync --recursive && %(basecmd)s submodule update --init --recursive %(fetch_flags)s)" % {'basecmd': ud.basecmd, 'fetch_flags': fetch_flags}, d, workdir=ud.destdir) + submodules = [] + paths = {} + uris = {} + local_paths = {} + for name in ud.names: + gitmodules = runfetchcmd("%s show HEAD:.gitmodules" % (ud.basecmd), d, quiet=True, workdir=ud.destdir) + + module = "" + for line in gitmodules.splitlines(): + if line.startswith('[submodule'): + module = line.split('"')[1] + submodules.append(module) + elif module and line.strip().startswith('path'): + path = line.split('=')[1].strip() + paths[module] = path + elif module and line.strip().startswith('url'): + url = line.split('=')[1].strip() + uris[module] = url + + self.copy_submodules(submodules, ud, ud.destdir, d) + + for module in submodules: + srcpath = os.path.join(ud.clonedir, 'modules', module) + modpath = os.path.join(repo_conf, 'modules', module) + + # Determine (from the submodule) the correct url to reference + try: + output = runfetchcmd("%(basecmd)s config remote.origin.url" % {'basecmd': ud.basecmd}, d, workdir=modpath) + except bb.fetch2.FetchError as e: + # No remote url defined in this submodule + continue + + local_paths[module] = output + + # Setup the local URL properly (like git submodule init or sync would do...) + runfetchcmd("%(basecmd)s config submodule.%(module)s.url %(url)s" % {'basecmd': ud.basecmd, 'module': module, 'url' : local_paths[module]}, d, workdir=ud.destdir) + + # Ensure the submodule repository is NOT set to bare, since we're checking it out... + runfetchcmd("%s config core.bare false" % (ud.basecmd), d, quiet=True, workdir=modpath) + + if submodules: + # Run submodule update, this sets up the directories -- without touching the config + runfetchcmd("%s submodule update --no-fetch" % (ud.basecmd), d, quiet=True, workdir=ud.destdir)
lib/bb/tests/fetch.py +3 -0
@@ -1344,6 +1344,9 @@ class GitShallowTest(FetcherTest): smdir = os.path.join(self.tempdir, 'gitsubmodule') bb.utils.mkdirhier(smdir) self.git('init', cwd=smdir) + # Make this look like it was cloned from a remote... + self.git('config --add remote.origin.url "%s"' % smdir, cwd=smdir) + self.git('config --add remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*"', cwd=smdir) self.add_empty_file('asub', cwd=smdir) self.git('submodule init', cwd=self.srcdir)
Signature · 1 line
1.8.3.1
Looks good from what I understand of the patch, but I have a question about git/gitsm fetchers:
Is it possible that two recipes end up modifying the same file e.g. via hard links when they use git submodules and share one of the repositories?
I'm seeing very odd build failures on sumo and it looks like recipes are either mixing up files in their sysroots or seeing each others patches on top of the shared git submodule tree. These happen when building the whole system and rebuilding affected recipes alone makes the problem go away.
With svn fetcher we see the same, but that's one of the problems we knew about from past. svn fetcher does not lock the download cached tree correctly and if multiple recipes use the same repo, one of them can see the repo in an inconsistent state and fail with various errors. Our workaround is to switch to http/s fetcher instead.
-Mikko
Quoted 5 lines · On 9/21/18 2:02 AM, [email protected] wrote:
Looks good from what I understand of the patch, but I have a question about git/gitsm fetchers:
Is it possible that two recipes end up modifying the same file e.g. via hard links when they use git submodules and share one of the repositories?
The symlinks (not hard links in this case) are between the components from the submodule and whatever the fetcher returned.
The only modified element(s) are the references to the submodules within the bare repositories.
Assuming all users use the same bare repositories, it's my understanding/expectation that the code in the 'downloads' will be the same. The only difference will then be in working dir.
Quoted 5 lines
I'm seeing very odd build failures on sumo and it looks like recipes are either mixing up files in their sysroots or seeing each others patches on top of the shared git submodule tree. These happen when building the whole system and rebuilding affected recipes alone makes the problem go away.
Before or after my patch? The gitsm changes here should ONLY affect the repository and any submodules affected. The code affects the fetch and unpack behaviors of the system. So no patches would be applied to the tree.. (well they are not supposed to be!)
The previous (current) gitsm fetcher makes it's submodules present using either copies or git submodule updates. In the case of a copy, it might be possible for one workdir to have a symlink back to the download or other shared location.
(I never encountered this, so this is purely speculation.)
If you can get me a gitsm reproducer for the problems you were seeing, I can run them through this code and see if I can make it work.
(As an aside, I think git submodules are one of the worst things ever. I really hate them...)
Quoted 5 lines
With svn fetcher we see the same, but that's one of the problems we knew about from past. svn fetcher does not lock the download cached tree correctly and if multiple recipes use the same repo, one of them can see the repo in an inconsistent state and fail with various errors. Our workaround is to switch to http/s fetcher instead.
With my patch set, I'm not doing an explicit locking. The download process itself, should be locked by the existing fetch2 system, and git fetcher. The module setup/symlinking process might be slightly different.
One place there COULD be an issue is if two different recipes refer to the same gitsm 'source', but different srcrev, and the different srcrev point to different submodules. I don't know if this is really a concern in practice though.
--Mark
Quoted 2 lines
-Mikko
Quoted 23 lines · Mark Hatle
Quoted 5 lines · On 9/21/18 2:02 AM, [email protected] wrote:
Looks good from what I understand of the patch, but I have a question about git/gitsm fetchers:
Is it possible that two recipes end up modifying the same file e.g. via hard links when they use git submodules and share one of the repositories?
The symlinks (not hard links in this case) are between the components from the submodule and whatever the fetcher returned.
The only modified element(s) are the references to the submodules within the bare repositories.
Assuming all users use the same bare repositories, it's my understanding/expectation that the code in the 'downloads' will be the same. The only difference will then be in working dir.
Quoted 5 lines
I'm seeing very odd build failures on sumo and it looks like recipes are either mixing up files in their sysroots or seeing each others patches on top of the shared git submodule tree. These happen when building the whole system and rebuilding affected recipes alone makes the problem go away.
Quoted 4 lines
Before or after my patch? The gitsm changes here should ONLY affect the repository and any submodules affected. The code affects the fetch and unpack behaviors of the system. So no patches would be applied to the tree.. (well they are not supposed to be!)
Before, on sumo and I was wondering if your patch could fix it.
But with help from #yocto I've now debugged this into openssl do_configure and am testing a patch. With luck my problem had nothing to do with git submodules.
Quoted 17 lines
The previous (current) gitsm fetcher makes it's submodules present using either copies or git submodule updates. In the case of a copy, it might be possible for one workdir to have a symlink back to the download or other shared location.
(I never encountered this, so this is purely speculation.)If you can get me a gitsm reproducer for the problems you were seeing, I can run them through this code and see if I can make it work.
(As an aside, I think git submodules are one of the worst things ever. I really hate them...)
Quoted 5 lines
With svn fetcher we see the same, but that's one of the problems we knew about from past. svn fetcher does not lock the download cached tree correctly and if multiple recipes use the same repo, one of them can see the repo in an inconsistent state and fail with various errors. Our workaround is to switch to http/s fetcher instead.
Quoted 7 lines
With my patch set, I'm not doing an explicit locking. The download process itself, should be locked by the existing fetch2 system, and git fetcher. The module setup/symlinking process might be slightly different.
One place there COULD be an issue is if two different recipes refer to the same gitsm 'source', but different srcrev, and the different srcrev point to different submodules. I don't know if this is really a concern in practice though.
Given my luck, we will hit this sooner or later. I think with svn fetcher we have already seen cases like this. In the best case build fails and in the worst case wrong things end up in the build...
-Mikko
Quoted 56 lines · On 9/21/18 9:15 AM, [email protected] wrote:
Quoted 23 lines · Mark Hatle
Quoted 5 lines · On 9/21/18 2:02 AM, [email protected] wrote:
Looks good from what I understand of the patch, but I have a question about git/gitsm fetchers:
Is it possible that two recipes end up modifying the same file e.g. via hard links when they use git submodules and share one of the repositories?
The symlinks (not hard links in this case) are between the components from the submodule and whatever the fetcher returned.
The only modified element(s) are the references to the submodules within the bare repositories.
Assuming all users use the same bare repositories, it's my understanding/expectation that the code in the 'downloads' will be the same. The only difference will then be in working dir.
Quoted 5 lines
I'm seeing very odd build failures on sumo and it looks like recipes are either mixing up files in their sysroots or seeing each others patches on top of the shared git submodule tree. These happen when building the whole system and rebuilding affected recipes alone makes the problem go away.
Quoted 4 lines
Before or after my patch? The gitsm changes here should ONLY affect the repository and any submodules affected. The code affects the fetch and unpack behaviors of the system. So no patches would be applied to the tree.. (well they are not supposed to be!)
Before, on sumo and I was wondering if your patch could fix it.
But with help from #yocto I've now debugged this into openssl
do_configureand am testing a patch. With luck my problem had nothing to do with git submodules.Quoted 21 lines
The previous (current) gitsm fetcher makes it's submodules present using either copies or git submodule updates. In the case of a copy, it might be possible for one workdir to have a symlink back to the download or other shared location.
(I never encountered this, so this is purely speculation.)If you can get me a gitsm reproducer for the problems you were seeing, I can run them through this code and see if I can make it work.
(As an aside, I think git submodules are one of the worst things ever. I really hate them...)
Quoted 5 lines
With svn fetcher we see the same, but that's one of the problems we knew about from past. svn fetcher does not lock the download cached tree correctly and if multiple recipes use the same repo, one of them can see the repo in an inconsistent state and fail with various errors. Our workaround is to switch to http/s fetcher instead.
With my patch set, I'm not doing an explicit locking. The download process itself, should be locked by the existing fetch2 system, and git fetcher. The module setup/symlinking process might be slightly different.
Quoted 7 lines
Quoted 3 lines
One place there COULD be an issue is if two different recipes refer to the same gitsm 'source', but different srcrev, and the different srcrev point to different submodules. I don't know if this is really a concern in practice though.
Given my luck, we will hit this sooner or later. I think with svn fetcher we have already seen cases like this. In the best case build fails and in the worst case wrong things end up in the build...
I looked through my code. There is an iteration step that happens in the unpack. If the module being asked for (in a given revision) is not present it will raise an exception. This should prevent the 'wrong things' at least.
--Mark
Quoted 2 lines
-Mikko
Quoted 3 lines · Mark Hatle
I looked through my code. There is an iteration step that happens in the unpack. If the module being asked for (in a given revision) is not present it will raise an exception. This should prevent the 'wrong things' at least.
Thanks! Looking forward to this patch in sumo branch too...
-Mikko