Create symbolic links atomically in the fetcher

We have occasional failures in our autobuilders where a setscene task fails, causing the original task to be run instead, but bitbake still fails with an error code in the end, causing unnecessary grief. One such case has been identified through the following error log:

The stack trace of python calls that resulted in this exception/failure was:
File: 'exec_python_func() autogenerated', lineno: 2, function: <module>

     0001:
 *** 0002:do_package_write_rpm_setscene(d)
     0003:
File: '${COREBASE}/meta/classes/package_rpm.bbclass', lineno: 757, function: do_package_write_rpm_setscene
     0753:# but we need to stop the rootfs/solver from running while we do...
     0754:do_package_write_rpm[sstate-lockfile-shared] += "${DEPLOY_DIR_RPM}/rpm.lock"
     0755:
     0756:python do_package_write_rpm_setscene () {
 *** 0757:    sstate_setscene(d)
     0758:}
     0759:addtask do_package_write_rpm_setscene
     0760:
     0761:python do_package_write_rpm () {
File: '${COREBASE}/meta/classes/sstate.bbclass', lineno: 648, function: sstate_setscene
     0644:            break
     0645:
     0646:def sstate_setscene(d):
     0647:    shared_state = sstate_state_fromvars(d)

*** 0648: accelerate = sstate_installpkg(shared_state, d)
0649: if not accelerate:
0650: raise bb.build.FuncFailed("No suitable staging package found")
0651:
0652:python sstate_task_prefunc () {
File: '${COREBASE}/meta/classes/sstate.bbclass', lineno: 297, function: sstate_installpkg

     0293:    sstatefetch = d.getVar('SSTATE_PKGNAME', True) + '_' + ss['task'] + ".tgz"
     0294:    sstatepkg = d.getVar('SSTATE_PKG', True) + '_' + ss['task'] + ".tgz"
     0295:
     0296:    if not os.path.exists(sstatepkg):

*** 0297: pstaging_fetch(sstatefetch, sstatepkg, d)
0298:
0299: if not os.path.isfile(sstatepkg):
0300: bb.note("Staging package %s does not exist" % sstatepkg)
0301: return False
File: '${COREBASE}/meta/classes/sstate.bbclass', lineno: 635, function: pstaging_fetch
0631: for srcuri in uris:
0632: localdata.setVar('SRC_URI', srcuri)
0633: try:
0634: fetcher = bb.fetch2.Fetch([srcuri], localdata, cache=False)
*** 0635: fetcher.download()
0636:
0637: # Need to optimise this, if using file:// urls, the fetcher just changes the local path
0638: # For now work around by symlinking
0639: localpath = bb.data.expand(fetcher.localpath(srcuri), localdata)
File: '${COREBASE}/poky/bitbake/lib/bb/fetch2/__init__.py', lineno: 1572, function: download
1568: localpath = ud.localpath
1569: elif m.try_premirror(ud, self.d):
1570: logger.debug(1, "Trying PREMIRRORS")
1571: mirrors = mirror_from_string(self.d.getVar('PREMIRRORS', True))
*** 1572: localpath = try_mirrors(self, self.d, ud, mirrors, False)

     1573:
     1574:                if premirroronly:
     1575:                    self.d.setVar("BB_NO_NETWORK", "1")
     1576:

File: '${COREBASE}/poky/bitbake/lib/bb/fetch2/__init__.py', lineno: 1020, function: try_mirrors
1016:
1017: uris, uds = build_mirroruris(origud, mirrors, ld)
1018:
1019: for index, uri in enumerate(uris):
*** 1020: ret = try_mirror_url(fetch, origud, uds[index], ld, check)
1021: if ret != False:
1022: return ret
1023: return None
1024:
File: '${COREBASE}/poky/bitbake/lib/bb/fetch2/__init__.py', lineno: 978, function: try_mirror_url
0974: if os.path.islink(origud.localpath):
0975: # Broken symbolic link
0976: os.unlink(origud.localpath)
0977:
*** 0978: os.symlink(ud.localpath, origud.localpath)
0979: update_stamp(origud, ld)
0980: return ud.localpath
0981:
0982: except bb.fetch2.NetworkAccess:
Exception: OSError: [Errno 17] File exists

What happens here is that two tasks simultaneously decide to download something, and both come to the conclusion that they need to create a symbolic link. And even if there is a check for whether the link already exists, there is a small window of time where both tasks see the missing link and tries to create it with the result that the second task fails as per above.

The change provided here causes the link creation to be made in an atomic way so that even if two tasks actually do decide that they need to create the same link, neither of them will fail.

I do not know if this solves the same problem that is solved by commit b8b14d975a254444461ba857fc6fb8c725de8874 on the master-next branch in the bitbake repository. Since I have no way to recreate the failure in a controlled way, I cannot test if the change on the master-next branch also solves it or not. Its description does not exactly match our situation (we do not map file:// URLs to http:// URLs in our SSTATE_MIRRORS), but maybe someone with better knowledge of the code can tell if either or both changes are needed.

//Peter

The following changes since commit 415b72ffcbd26e5f3664370d8b2a9b8105fb6342:

dnf: remove systemd units in nativesdk builds (2017-03-28 10:34:37 +0100)

are available in the git repository at:

git://git.yoctoproject.org/poky-contrib pkj/atomic_symlinks
http://git.yoctoproject.org/cgit.cgi/poky-contrib/log/?h=pkj/atomic_symlinks

Peter Kjellerstedt (1):
fetch2: Create/replace symbolic links atomically

 bitbake/lib/bb/fetch2/__init__.py | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

2.12.0

Under rare circumstances, creating symbolic links could fail because the link already exists. At first glance the code seemed to protect against this, but there was a small window where two separate tasks could decide that a symbolic link needed to be created and then the first task would create it and the second task would fail.

---
 bitbake/lib/bb/fetch2/__init__.py | 20 +++++++++++++-------
 1 file changed, 13 insertions(+), 7 deletions(-)

index 464e66b98a..f891e34eab 100644
@@ -25,7 +25,7 @@ BitBake build tools.
 #
 # Based on functions from the base bb module, Copyright 2003 Holger Schurig
 
-import os, re
+import os, re, tempfile
 import signal
 import logging
 import urllib.request, urllib.parse, urllib.error

@@ -946,6 +946,16 @@ def rename_bad_checksum(ud, suffix):
     bb.warn("Renaming %s to %s" % (ud.localpath, new_localpath))
     bb.utils.movefile(ud.localpath, new_localpath)
 
+def atomic_symlink(src, dst):
+    """Create/replace a symbolic link atomically."""
+
+    tmpname = tempfile.mktemp(prefix=dst)
+    os.symlink(src, tmpname)
+    try:
+        os.rename(tmpname, dst)
+    except OSError:
+        os.remove(tmpname)
+        raise
 
 def try_mirror_url(fetch, origud, ud, ld, check = False):
     # Return of None or a value means we're finished

@@ -983,7 +993,7 @@ def try_mirror_url(fetch, origud, ud, ld, check = False):
                 open(ud.donestamp, 'w').close()
             dest = os.path.join(dldir, os.path.basename(ud.localpath))
             if not os.path.exists(dest):
-                os.symlink(ud.localpath, dest)
+                atomic_symlink(ud.localpath, dest)
             if not verify_donestamp(origud, ld) or origud.method.need_update(origud, ld):
                 origud.method.download(origud, ld)
                 if hasattr(origud.method,"build_mirror_data"):

@@ -991,11 +1001,7 @@ def try_mirror_url(fetch, origud, ud, ld, check = False):
             return origud.localpath
         # Otherwise the result is a local file:// and we symlink to it
         if not os.path.exists(origud.localpath):
-            if os.path.islink(origud.localpath):
-                # Broken symbolic link
-                os.unlink(origud.localpath)
-
-            os.symlink(ud.localpath, origud.localpath)
+            atomic_symlink(ud.localpath, origud.localpath)
         update_stamp(origud, ld)
         return ud.localpath
 

2.12.0

== Series Details ==

Series: Create symbolic links atomically in the fetcher
Revision: 1
URL : https://patchwork.openembedded.org/series/6022/
State : failure

== Summary ==

Thank you for submitting this patch series to OpenEmbedded Core. This is an automated response. Several tests have been executed on the proposed series by patchtest resulting in the following failures:

* Issue Series does not apply on top of target branch [test_series_merge_on_head]
Suggested fix Rebase your series on top of targeted branch
Targeted branch master (currently at d68a86d87a)

If you believe any of these test results are incorrect, please reply to the mailing list ([email protected]) raising your concerns. Otherwise we would appreciate you correcting the issues and submitting a new version of the patchset if applicable. Please ensure you add/increment the version number when sending the new version (i.e. [PATCH] -> [PATCH v2] -> [PATCH v3] -> ...).

---

Test framework: http://git.yoctoproject.org/cgit/cgit.cgi/patchtest
Test suite:     http://git.yoctoproject.org/cgit/cgit.cgi/patchtest-oe

Bitbake has its own list, can you resend to
[email protected]?

Cheers,
Ross

As Ross says, this needs to go to the bitbake list. I'm also curious why not watch for the "AlreadyExists" exception (whatever it really is) and then simply pass over the error assuming the symlink points to the same place? That saves messing with temporary files.

I also worry a little about the locking? Why have you two fetch processes which are changing the same files but not covered by the same lock file? :/

Cheers,

Richard