Better support for upgrading packages in opkg and update-rc.d.bbclass

init-ifupdown inherits default implementations of prerm, postrm, preinst and postinst from update-rc.d.bbclass. Unfortunately these default implementations don't deal with package upgrades as well as they could.

What ends up happening is that opkg starts downloading and unpacking packages sequentially. As it downloads and unpacks each package, opkg calls prerm, preinst and postrm hook scripts. Once all packages are unpacked, the postinst script is finally called for all packages as part of the "configure" stage.

In the case of init-ifupdown, the default prerm and preinst scripts stop networking. Networking is not brought back up again until the init-ifupdown postinst is called, but that only happens after all the packages have been downloaded, unpacked and installed. This leaves a window where any package that needs to be downloaded after init-ifupdown is encountered will fail.

This patch fixes the problem by enhancing opkg to also provide the "upgrade" argument to prerm and postrm, to (partially) match what dpkg does. See
https://wiki.debian.org/MaintainerScripts for dpkg diagrams that clarify the intended process. opkg lacks the full functionality of dpkg, but for the purpose of this exercise, they are similar enough.

I have submitted a patch to the opkg-devel list to include the "upgrade" argument in future version of opkg.

The second part of the solution is an update to the default implementations of the pre- and post- scripts provided by update-rc.d.bbclass. The scripts are now careful to remove the package init scripts using the old package context and to delay the restart of a service until the configure stage, called from the postinst script.

---
 meta/classes/update-rc.d.bbclass                   | 75 +++++++++++++---------
 .../opkg/upgrade-argument-for-pre_postrm.patch     | 30 +++++++++
 meta/recipes-devtools/opkg/opkg_0.2.2.bb           |  1 +
 3 files changed, 77 insertions(+), 29 deletions(-)
 create mode 100644 meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch

meta/classes/update-rc.d.bbclass +37 -27

index bc1aa7d..c29457f 100644
@@ -14,45 +14,62 @@ INITSCRIPT_PARAMS ?= "defaults"
  INIT_D_DIR = "${sysconfdir}/init.d"
 -updatercd_preinst() {
-if [ -z "$D" -a -f "${INIT_D_DIR}/${INITSCRIPT_NAME}" ]; then
-       ${INIT_D_DIR}/${INITSCRIPT_NAME} stop
+# During an upgrade, the pre/postrm scripts from old package are called
+# and the pre/postinst scripts called are from the new package.
+# See https://wiki.debian.org/MaintainerScripts for dpkg diagrams.
+# opkg uses a subset, which lacks most of the error handling.
+
+# Old package context, step 1
+updatercd_prerm() {
+if [ "x$1" != "xupgrade" ] ; then
+  ${INIT_D_DIR}/${INITSCRIPT_NAME} stop
 fi
-if type update-rc.d >/dev/null 2>/dev/null; then
-       if [ -n "$D" ]; then
-               OPT="-f -r $D"
-       else
-               OPT="-f"
-       fi
-       update-rc.d $OPT ${INITSCRIPT_NAME} remove
+if [ -z "$D" ]; then
+  OPT=""
+else
+  OPT="-r $D"
 fi
-}
-
-updatercd_postinst() {
 if type update-rc.d >/dev/null 2>/dev/null; then
-       if [ -n "$D" ]; then
-               OPT="-r $D"
-       else
-               OPT="-s"
-       fi
-       update-rc.d $OPT ${INITSCRIPT_NAME} ${INITSCRIPT_PARAMS}
+  update-rc.d $OPT ${INITSCRIPT_NAME} remove
 fi
 }
 -updatercd_prerm() {
-if [ -z "$D" ]; then
-       ${INIT_D_DIR}/${INITSCRIPT_NAME} stop
-fi
+# New package context, step 2
+updatercd_preinst() {
+case "$1" in
+  upgrade)
+    ;;
+  *)
+    ;;
+esac
 }
 +# Old package context, step 3
 updatercd_postrm() {
+case "$1" in
+  upgrade)
+    ;;
+  *)
+    ;;
+esac
+}
+
+# N.B. Step 4 runs after all packages have been through steps 1-3 and therefore we
+#      need to delay service restarts during upgrade until here. Otherwise we end
+#      up with situations, like networking going down in the middle of "opkg upgrade",
+#      thus resulting in failures to fetch further packages.
+
+# New package context, step 4
+updatercd_postinst() {
 if type update-rc.d >/dev/null 2>/dev/null; then
-       if [ -n "$D" ]; then
-               OPT="-r $D"
-       else
-               OPT=""
-       fi
-       update-rc.d $OPT ${INITSCRIPT_NAME} remove

meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch new file +11 -0

+diff --git a/libopkg/opkg_remove.c b/libopkg/opkg_remove.c
+index 07401b2..52454f8 100644
+--- a/libopkg/opkg_remove.c
++++ b/libopkg/opkg_remove.c
+@@ -292,7 +292,7 @@ opkg_remove_pkg(pkg_t *pkg, int from_upgrade)
+      pkg->state_want = SW_DEINSTALL;
+      opkg_state_changed++;
+ +-     if (pkg_run_script(pkg, "prerm", "remove") != 0) {
++     if (pkg_run_script(pkg, "prerm", from_upgrade ? "upgrade" : "remove") != 0) {
+          if (!conf->force_remove) {
+              opkg_msg(ERROR, "not removing package \"%s\", "

Quoted 15 lines · Peter Urbanec

meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch new file +15 -0

+diff --git a/libopkg/opkg_remove.c b/libopkg/opkg_remove.c
+index 07401b2..52454f8 100644
+--- a/libopkg/opkg_remove.c
++++ b/libopkg/opkg_remove.c
+@@ -292,7 +292,7 @@ opkg_remove_pkg(pkg_t *pkg, int from_upgrade)
+      pkg->state_want = SW_DEINSTALL;
+      opkg_state_changed++;
+ +-     if (pkg_run_script(pkg, "prerm", "remove") != 0) {
++     if (pkg_run_script(pkg, "prerm", from_upgrade ? "upgrade" : "remove") != 0) {
+          if (!conf->force_remove) {
+              opkg_msg(ERROR, "not removing package \"%s\", "
+                              "prerm script failed\n", pkg->name);
+@@ -310,7 +310,7 @@ opkg_remove_pkg(pkg_t *pkg, int from_upgrade)
+        feel free to fix this. */
+      remove_data_files_and_list(pkg);

meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch line 28

Quoted 6 lines

meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch new file +6 -0

+
+-     err = pkg_run_script(pkg, "postrm", "remove");
++     err = pkg_run_script(pkg, "postrm", from_upgrade ? "upgrade" : "remove");
+
+      remove_maintainer_scripts(pkg);
+      pkg->state_status = SS_NOT_INSTALLED;

I'm happy with this change, both here and upstream.

opkg_remove_pkg is never called with from_upgrade set to a non-zero value. The removal of an old package during an upgrade is handled by directly calling remove_data_files_and_list and remove_maintainer_scripts from opkg_install_pkg. The desired behaviour is given in the comments in prerm_upgrade_old_pkg in opkg_install.c, the behaviour given by the comment in that function just needs implementing.

This is already entered as issue 104 in the opkg issue tracker:
https://code.google.com/p/opkg/issues/detail?id=104

Have you tested this patch before submitting it?

Quoted 8 lines · Paul Barker

opkg_remove_pkg is never called with from_upgrade set to a non-zero value. The removal of an old package during an upgrade is handled by directly calling remove_data_files_and_list and remove_maintainer_scripts from opkg_install_pkg. The desired behaviour is given in the comments in prerm_upgrade_old_pkg in opkg_install.c, the behaviour given by the comment in that function just needs implementing.

Indeed, I got fooled by looking at the call to opkg_install_pkg(pkg, 1)

Quoted 3 lines

This is already entered as issue 104 in the opkg issue tracker:
https://code.google.com/p/opkg/issues/detail?id=104

So it is. Slightly more work than my trivial opkg patch, but then again, my trivial opkg patch is not much use ;-) Please disregard it.

Quoted 1 line

Have you tested this patch before submitting it?

To be honest, I have not tested the opkg "upgrade" patch in isolation. I have tested the opkg patch for multiple package upgrade case together with my changes to update-rc.d.bbclass changes (i.e. the entire patch I posted to the oe-core list) and this has solved the problems that I encountered initially. The "fix" for my use case is due to the changes in the way the prerm, postrm and postinst scripts work, but reviewing those changes, that's not good enough either.

Thanks for the review.

I'll try again later...

Peter

init-ifupdown inherits default implementations of prerm, postrm, preinst and postinst from update-rc.d.bbclass. Unfortunately these default implementations don't deal with package upgrades as well as they could.

What ends up happening is that opkg starts downloading and unpacking packages sequentially. As it downloads and unpacks each package, opkg calls prerm, preinst and postrm hook scripts. Once all packages are unpacked, the postinst script is finally called for all packages as part of the "configure" stage.

In the case of init-ifupdown, the default prerm and preinst scripts stop networking.

Networking is not brought back up again until the init-ifupdown postinst is called, but that only happens after all the packages have been downloaded, unpacked and installed. This leaves a window where any package that needs to be downloaded after init-ifupdown is encountered will fail.

This patch fixes the problem by enhancing opkg to also provide the "upgrade" argument to prerm and postrm, to (partially) match what dpkg does. See
https://wiki.debian.org/MaintainerScripts for dpkg diagrams that clarify the intended process. opkg lacks the full functionality of dpkg, but for the purpose of this exercise, they are similar enough.

I have submitted a patch to the opkg-devel list to include the "upgrade" argument in future version of opkg.

The second part of the solution is an update to the default implementations of the pre- and post- scripts provided by update-rc.d.bbclass. The scripts are now careful to remove the package init scripts using the old package context and to delay the restart of a service until the configure stage, called from the postinst script.

---
 meta/classes/update-rc.d.bbclass                   | 75 +++++++++++++---------
 .../opkg/upgrade-argument-for-pre_postrm.patch     | 30 +++++++++
 meta/recipes-devtools/opkg/opkg_0.2.2.bb           |  1 +
 3 files changed, 77 insertions(+), 29 deletions(-)
 create mode 100644 meta/recipes-devtools/opkg/opkg/upgrade-argument-for-pre_postrm.patch

meta/classes/update-rc.d.bbclass +0 -0

index bc1aa7d..c29457f 100644
@@ -14,45 +14,62 @@ INITSCRIPT_PARAMS ?= "defaults"
  INIT_D_DIR = "${sysconfdir}/init.d"
 -updatercd_preinst() {

meta/classes/update-rc.d.bbclass +14 -19

+  ${INIT_D_DIR}/${INITSCRIPT_NAME} stop
 fi
-if type update-rc.d >/dev/null 2>/dev/null; then
-	if [ -n "$D" ]; then
-		OPT="-f -r $D"
-	else
-		OPT="-f"
-	fi
-	update-rc.d $OPT ${INITSCRIPT_NAME} remove
+if [ -z "$D" ]; then
+  OPT=""
+else
+  OPT="-r $D"
 fi
-}
-
-updatercd_postinst() {
 if type update-rc.d >/dev/null 2>/dev/null; then
-	if [ -n "$D" ]; then
-		OPT="-r $D"
-	else
-		OPT="-s"
-	fi
-	update-rc.d $OPT ${INITSCRIPT_NAME} ${INITSCRIPT_PARAMS}
+  update-rc.d $OPT ${INITSCRIPT_NAME} remove
 fi
 }
 -updatercd_prerm() {
-if [ -z "$D" ]; then
-	${INIT_D_DIR}/${INITSCRIPT_NAME} stop
-fi
+# New package context, step 2
+updatercd_preinst() {
+case "$1" in
+  upgrade)
+    ;;
+  *)
+    ;;
+esac
 }

meta/classes/update-rc.d.bbclass +7 -4

+# N.B. Step 4 runs after all packages have been through steps 1-3 and therefore we
+#      need to delay service restarts during upgrade until here. Otherwise we end
+#      up with situations, like networking going down in the middle of "opkg upgrade",
+#      thus resulting in failures to fetch further packages.
+
+# New package context, step 4
+updatercd_postinst() {
 if type update-rc.d >/dev/null 2>/dev/null; then
-	if [ -n "$D" ]; then
-		OPT="-r $D"
-	else
-		OPT=""

Quoted 2 lines · Peter Urbanec

In the case of init-ifupdown, the default prerm and preinst scripts stop networking.

I think prerm and postrm scripts should exit silently on upgrade. At least that's what happens with rpm and with a patch just submitted for deb [1]. I guess ipk should implement the same logic, for which your patch to opkg is a prerequisite.

Quoted 1 line

+if [ "x$1" != "xupgrade" ] ; then

I think using "x..." syntax is obsolete. "$1" != "upgrade" should work with every supported shell and is easier to read.

This part of the patch wouldn't be needed if package_ipk.bbclass implemented the same logic as rpm mentioned above.

But, on the other hand, maybe package_rpm.bbclass is wrong... ;-)

Quoted 13 lines

meta/classes/update-rc.d.bbclass +10 -0

+# New package context, step 2
+updatercd_preinst() {
+case "$1" in
+  upgrade)
+    ;;
+  *)
+    ;;
+esac
  }
  +# Old package context, step 3
  updatercd_postrm() {
+case "$1" in
+  upgrade)

meta/classes/update-rc.d.bbclass INITSCRIPT_PARAMS ?= "defaults" · line 54

Quoted 6 lines

meta/classes/update-rc.d.bbclass +6 -0

+    ;;
+  *)
+    ;;
+esac
+}
+

The two functions above don't do anything, so they should be removed.

I have patches to update-rc.d [2] and update-rc.d.bbclass [3] to improve handling of distributions with both systemd and sysvinit distro features enabled. What they try to avoid is direct calls to init scripts, where systemd may provide own unit files. Would it be feasible to use update-rc.d to issue a restart command?

I have revised the patches and have taken some of your feedback on board. See comments further down...

Quoted 8 lines · Andreas Oberritter

Quoted 2 lines · Peter Urbanec

In the case of init-ifupdown, the default prerm and preinst scripts stop networking.

I think prerm and postrm scripts should exit silently on upgrade. At least that's what happens with rpm and with a patch just submitted for deb [1]. I guess ipk should implement the same logic, for which your patch to opkg is a prerequisite.

I modelled my changes on documented behaviour for dpkg since opkg code often refers to dpkg behaviour. dpkg supports error return codes from these scripts and can do various error handling. opkg is less featured in that respect at this moment, but I think in the interest of future compatibility (and the principle of least astonishment) the scripts should return non-zero to signify that the install/remove/upgrade process can not continue. Having said that, I really don't see many reasons for a prerm or postrm script to try to abort an upgrade.

Quoted 4 lines

This part of the patch wouldn't be needed if package_ipk.bbclass implemented the same logic as rpm mentioned above.

But, on the other hand, maybe package_rpm.bbclass is wrong... ;-)

I think you have me confused here. package_ipk.bbclass tries to grab the prerm/postrm and preinst/postinst script variables from the package and writes the contents into the corresponding files within the *.ipk. These scripts are then run by opkg when the *.ipk files are processed.

Quoted 13 lines

Quoted 13 lines

meta/classes/update-rc.d.bbclass +10 -0

+# New package context, step 2
+updatercd_preinst() {
+case "$1" in
+  upgrade)
+    ;;
+  *)
+    ;;
+esac
  }
  +# Old package context, step 3
  updatercd_postrm() {
+case "$1" in
+  upgrade)

Quoted 8 lines

meta/classes/update-rc.d.bbclass INITSCRIPT_PARAMS ?= "defaults" · line 54

Quoted 6 lines

meta/classes/update-rc.d.bbclass +6 -0

+    ;;
+  *)
+    ;;
+esac
+}
+

The two functions above don't do anything, so they should be removed.

They are used later by update_rcd_package(). They could have been simpler, but I thought that both the presence of comments and the template code would serve as a reminder of the semantics of the upgrade process and the fact that package upgrades are different to package installs or removals.

Quoted 5 lines

I have patches to update-rc.d [2] and update-rc.d.bbclass [3] to improve handling of distributions with both systemd and sysvinit distro features enabled. What they try to avoid is direct calls to init scripts, where systemd may provide own unit files. Would it be feasible to use update-rc.d to issue a restart command?

The patch I submitted uses the "-s" flag to update-rc.d to start the service when it is installed, but as far as I can tell, there is no way to just stop the service using update-rc.d and the restart case will not work for scripts that don't support the "restart" argument (such as urandom).

The restart case is also predicated on a number of conditions that will generally evaluate to false, since the old links would have been removed in prerm.

What happens when generating rpm (and since few days deb) packages is that code gets injected that makes these scripts exit with return code 0 on upgrades. This may be just a workaround to emulate opkg's limitations, but it may also be by design. I guess this should be decided upon first. Whatever we'll agree on, the most important thing is that all three supported package formats should behave identically.

Quoted 1 line

Quoted 1 line

But, on the other hand, maybe package_rpm.bbclass is wrong... ;-)

I meant to say that if package_ipk wanted to behave like package_rpm (and package_deb since a few days), then it would inject '[ "$1" != upgrade ] || exit 0' at the beginning of postrm.

Quoted 5 lines

Quoted 3 lines

Quoted 2 lines

+updatercd_preinst() {
   updatercd_postrm() {

The two functions above don't do anything, so they should be removed.

They are used later by update_rcd_package().

If they are empty, they should be removed from update_rcd_package(), too. There are many more places in OE-Core where postinst scripts are used, and the common policy is to implement only those snippets which are used. Otherwise it's just a waste of space on the target systems.

Quoted 5 lines

The patch I submitted uses the "-s" flag to update-rc.d to start the service when it is installed, but as far as I can tell, there is no way to just stop the service using update-rc.d and the restart case will not work for scripts that don't support the "restart" argument (such as urandom).

Isn't restart required to be implemented by LSB? Maybe we could just fix urandom instead. Does urandom even use update-rc.d? I don't know which recipe actually installs a urandom init script, to be honest.

Quoted 3 lines

The restart case is also predicated on a number of conditions that will generally evaluate to false, since the old links would have been removed in prerm.

I don't understand which conditions you're referring to.

Sorry for being slow to respond, I've been snowed under and dealing with hardware failures :-(

I'll be brief...

Quoted 6 lines · Andreas Oberritter

What happens when generating rpm (and since few days deb) packages is that code gets injected that makes these scripts exit with return code 0 on upgrades. This may be just a workaround to emulate opkg's limitations, but it may also be by design. I guess this should be decided upon first. Whatever we'll agree on, the most important thing is that all three supported package formats should behave identically.

I don't think this is a good idea. Consider the case where a package is being upgraded and the old package was configured to run in runlevels 3 and 4. The new package is configured to run in runlevels 2 and 3. You need to run the old prerm script so that you correctly remove the script from all old runlevels, then you need to run the new postinst script to add the script to the correct runlevels. If you prevent prerm from running on upgrade, you will never remove the script from runlevel 4.

Quoted 1 line

Quoted 1 line

Quoted 1 line

But, on the other hand, maybe package_rpm.bbclass is wrong... ;-)

I have not studied RPM packaging. Does it use the same scripts and more importantly does RPM expect the same install logic? Perhaps it's not possible to apply the same logic to all packaging and installation systems.

Quoted 3 lines

I meant to say that if package_ipk wanted to behave like package_rpm (and package_deb since a few days), then it would inject '[ "$1" != upgrade ] || exit 0' at the beginning of postrm.

See my two points above. Perhaps one size does not fit all in this case.

Quoted 10 lines

Quoted 5 lines

Quoted 3 lines

Quoted 2 lines

+updatercd_preinst() {
   updatercd_postrm() {

The two functions above don't do anything, so they should be removed.

They are used later by update_rcd_package().

If they are empty, they should be removed from update_rcd_package(), too. There are many more places in OE-Core where postinst scripts are used, and the common policy is to implement only those snippets which are used. Otherwise it's just a waste of space on the target systems.

I guess the logic could be changed so that if the user does not provide their own implementation, then nothing is generated. I haven't studied the code in depth to establish whether this may have a cascading effect down the line. I opted for the option that is less likely to cause unintended breakage.

Quoted 9 lines

Quoted 5 lines

The patch I submitted uses the "-s" flag to update-rc.d to start the service when it is installed, but as far as I can tell, there is no way to just stop the service using update-rc.d and the restart case will not work for scripts that don't support the "restart" argument (such as urandom).

Isn't restart required to be implemented by LSB? Maybe we could just fix urandom instead. Does urandom even use update-rc.d? I don't know which recipe actually installs a urandom init script, to be honest.

I'm not sure what the LSB requirements are, but I did take a quick look through a few (random) existing scripts to determine the state of affairs and concluded that one can not rely on a "restart" implementation. The comments in init-ifupdown/init actually say that "restart" is deprecated and should not be used.

Quoted 5 lines

Quoted 3 lines

The restart case is also predicated on a number of conditions that will generally evaluate to false, since the old links would have been removed in prerm.

I don't understand which conditions you're referring to.

It's a bit confusing to follow, but have a look at the part of the code that checks the existence of the links in the context of the first point I made in this email. If you want to support the upgrade logic properly (at least as far as doing things like changing the runlevels), you should remove the old links in prerm, which will later cause the code to incorrectly assume that this was not an upgrade.

Cheers,
Peter