[PATCH v2] dracut-module-setup: Skip initrd-cleanup and initrd-parse-etc in kdump

[Lichen Liu]

When using multipath devices as the target for kdump, if user_friendly_name
is also specified, devices default to names like "mpath*", e.g., mpatha.
In dracut, we obtain a persistent device name via get_persistent_dev. However,
dracut currently believes using /dev/mapper/mpath* could cause issues, thus
alternatively names are used, here it's /dev/disk/by-uuid/<FS_UUID>.

During the kdump boot progress, the /dev/disk/by-uuid/<FS_UUID> will exist as
soon as one of the path devices exists, but it won't be usable by systemd,
since multipathd will claim that device as a path device. Then multipathd will
get stopped before it can create the multipath device.

Without user_friendly_name, /dev/mapper/<WWID> is considered a persistent
device name, avoiding the issue.

The exit of multipathd is due to two dependencies in the current dracut module
90multipath/multipathd.service, "Before=initrd-cleanup.service" and
"Conflicts=initrd-cleanup.service".

As per man 5 systemd.unit, if A.service has "Conflicts=B.service", starting
B.service will stop A.service.

This is useful during normal boot. However, we will never switch-root after
capturing vmcore in kdump.

We need to ensure that multipathd is not killed due to such dependency issue.
Without modifying multipathd.service, we add ConditionPathExists=!/proc/vmcore
to skip initrd-cleanup.service in kdump. This approach is beneficial as
it avoid the potential termination of other services that conflict with
initrd-cleanup.service. Also skip initrd-parse-etc.service as it will try to
start initrd-cleanup.service. Both of these services are used for switch root,
so they can be safely skipped in kdump.

Suggested-by: Benjamin Marzinski <bmarzins(a)redhat.com&gt;
Suggested-by: Dave Young <dyoung(a)redhat.com&gt;
Signed-off-by: Lichen Liu <lichliu(a)redhat.com&gt;
---
 dracut-module-setup.sh | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/dracut-module-setup.sh b/dracut-module-setup.sh
index 1dc88dc..0a901c2 100755
--- a/dracut-module-setup.sh
+++ b/dracut-module-setup.sh
@@ -1089,6 +1089,15 @@ install() {
         's/\(^[[:space:]]*reserved_memory[[:space:]]*=\)[[:space:]]*[[:digit:]]*/\1
1024/' \
         "${initdir}/etc/lvm/lvm.conf" &> /dev/null
 
+    # Skip initrd-cleanup.service and initrd-parse-etc.service becasue we don't
+    # need to switch root. Instead of removing them, we use ConditionPathExists
+    # to check if /proc/vmcore exists to determine if we are in kdump.
+    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
+        "${initdir}/${systemdsystemunitdir}/initrd-cleanup.service" &>
/dev/null
+
+    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
+        "${initdir}/${systemdsystemunitdir}/initrd-parse-etc.service" &>
/dev/null
+
     # Save more memory by dropping switch root capability
     dracut_no_switch_root
 }
-- 
2.43.0

[Philipp Rudo]

Hi Lichen,

On Mon, 22 Jan 2024 15:59:09 +0800
Lichen Liu <lichliu(a)redhat.com&gt; wrote:

>  When using multipath devices as the target for kdump, if
> user_friendly_name
>  is also specified, devices default to names like "mpath*", e.g., mpatha.
>  In dracut, we obtain a persistent device name via get_persistent_dev. However,
>  dracut currently believes using /dev/mapper/mpath* could cause issues, thus
>  alternatively names are used, here it's /dev/disk/by-uuid/<FS_UUID>.
>  
>  During the kdump boot progress, the /dev/disk/by-uuid/<FS_UUID> will exist as
>  soon as one of the path devices exists, but it won't be usable by systemd,
>  since multipathd will claim that device as a path device. Then multipathd will
>  get stopped before it can create the multipath device.
>  
>  Without user_friendly_name, /dev/mapper/<WWID> is considered a persistent
>  device name, avoiding the issue.
>  
>  The exit of multipathd is due to two dependencies in the current dracut module
>  90multipath/multipathd.service, "Before=initrd-cleanup.service" and
>  "Conflicts=initrd-cleanup.service".
>  
>  As per man 5 systemd.unit, if A.service has "Conflicts=B.service", starting
>  B.service will stop A.service.
>  
>  This is useful during normal boot. However, we will never switch-root after
>  capturing vmcore in kdump.
>  
>  We need to ensure that multipathd is not killed due to such dependency issue.
>  Without modifying multipathd.service, we add ConditionPathExists=!/proc/vmcore
>  to skip initrd-cleanup.service in kdump. This approach is beneficial as
>  it avoid the potential termination of other services that conflict with
>  initrd-cleanup.service. Also skip initrd-parse-etc.service as it will try to
>  start initrd-cleanup.service. Both of these services are used for switch root,
>  so they can be safely skipped in kdump.
>  
>  Suggested-by: Benjamin Marzinski <bmarzins(a)redhat.com&gt;
>  Suggested-by: Dave Young <dyoung(a)redhat.com&gt;
>  Signed-off-by: Lichen Liu <lichliu(a)redhat.com&gt; 
Looks good. Thanks for the great description in the commit message!
Reviewed-by: Philipp Rudo <prudo(a)redhat.com&gt;

>  ---
>   dracut-module-setup.sh | 9 +++++++++
>   1 file changed, 9 insertions(+)
>  
>  diff --git a/dracut-module-setup.sh b/dracut-module-setup.sh
>  index 1dc88dc..0a901c2 100755
>  --- a/dracut-module-setup.sh
>  +++ b/dracut-module-setup.sh
>  @@ -1089,6 +1089,15 @@ install() {
>           's/\(^[[:space:]]*reserved_memory[[:space:]]*=\)[[:space:]]*[[:digit:]]*/\1
> 1024/' \
>           "${initdir}/etc/lvm/lvm.conf" &> /dev/null
>   
>  +    # Skip initrd-cleanup.service and initrd-parse-etc.service becasue we don't
>  +    # need to switch root. Instead of removing them, we use ConditionPathExists
>  +    # to check if /proc/vmcore exists to determine if we are in kdump.
>  +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
>  +        "${initdir}/${systemdsystemunitdir}/initrd-cleanup.service" &>
> /dev/null
>  +
>  +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
>  +        "${initdir}/${systemdsystemunitdir}/initrd-parse-etc.service"
> &> /dev/null
>  +
>       # Save more memory by dropping switch root capability
>       dracut_no_switch_root
>   }

[Coiby Xu]

On Tue, Jan 23, 2024 at 11:07:47AM +0100, Philipp Rudo wrote:
> Hi Lichen,
>
> On Mon, 22 Jan 2024 15:59:09 +0800
> Lichen Liu <lichliu(a)redhat.com&gt; wrote:
>
> > When using multipath devices as the target for kdump, if user_friendly_name
> > is also specified, devices default to names like "mpath*", e.g., mpatha.
> > In dracut, we obtain a persistent device name via get_persistent_dev. However,
> > dracut currently believes using /dev/mapper/mpath* could cause issues, thus
> > alternatively names are used, here it's /dev/disk/by-uuid/<FS_UUID>.
> >
> > During the kdump boot progress, the /dev/disk/by-uuid/<FS_UUID> will exist as
> > soon as one of the path devices exists, but it won't be usable by systemd,
> > since multipathd will claim that device as a path device. Then multipathd will
> > get stopped before it can create the multipath device.
> >
> > Without user_friendly_name, /dev/mapper/<WWID> is considered a persistent
> > device name, avoiding the issue.
> >
> > The exit of multipathd is due to two dependencies in the current dracut module
> > 90multipath/multipathd.service, "Before=initrd-cleanup.service" and
> > "Conflicts=initrd-cleanup.service".
> >
> > As per man 5 systemd.unit, if A.service has "Conflicts=B.service",
> starting
> > B.service will stop A.service.
> >
> > This is useful during normal boot. However, we will never switch-root after
> > capturing vmcore in kdump.
> >
> > We need to ensure that multipathd is not killed due to such dependency issue.
> > Without modifying multipathd.service, we add ConditionPathExists=!/proc/vmcore
> > to skip initrd-cleanup.service in kdump. This approach is beneficial as
> > it avoid the potential termination of other services that conflict with
> > initrd-cleanup.service. Also skip initrd-parse-etc.service as it will try to
> > start initrd-cleanup.service. Both of these services are used for switch root,
> > so they can be safely skipped in kdump.
> >
> > Suggested-by: Benjamin Marzinski <bmarzins(a)redhat.com&gt;
> > Suggested-by: Dave Young <dyoung(a)redhat.com&gt;
> > Signed-off-by: Lichen Liu <lichliu(a)redhat.com&gt;
>
> Looks good. Thanks for the great description in the commit message!
> Reviewed-by: Philipp Rudo <prudo(a)redhat.com&gt;
>
> > ---
> >  dracut-module-setup.sh | 9 +++++++++
> >  1 file changed, 9 insertions(+)
> >
> > diff --git a/dracut-module-setup.sh b/dracut-module-setup.sh
> > index 1dc88dc..0a901c2 100755
> > --- a/dracut-module-setup.sh
> > +++ b/dracut-module-setup.sh
> > @@ -1089,6 +1089,15 @@ install() {
> >         
> 's/\(^[[:space:]]*reserved_memory[[:space:]]*=\)[[:space:]]*[[:digit:]]*/\1 1024/'
> \
> >          "${initdir}/etc/lvm/lvm.conf" &> /dev/null
> >
> > +    # Skip initrd-cleanup.service and initrd-parse-etc.service becasue we don't
> > +    # need to switch root. Instead of removing them, we use ConditionPathExists
> > +    # to check if /proc/vmcore exists to determine if we are in kdump.
> > +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
> > +        "${initdir}/${systemdsystemunitdir}/initrd-cleanup.service"
> &> /dev/null
> > +
> > +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
> > +        "${initdir}/${systemdsystemunitdir}/initrd-parse-etc.service"
> &> /dev/null
> > +
> >      # Save more memory by dropping switch root capability
> >      dracut_no_switch_root
> >  }
>  
Patch merged, thanks to Lichen and Philipp!

-- 
Best regards,
Coiby

[Benjamin Marzinski]

On Mon, Jan 22, 2024 at 03:59:09PM +0800, Lichen Liu wrote:
>  When using multipath devices as the target for kdump, if
> user_friendly_name
>  is also specified, devices default to names like "mpath*", e.g., mpatha.
>  In dracut, we obtain a persistent device name via get_persistent_dev. However,
>  dracut currently believes using /dev/mapper/mpath* could cause issues, thus
>  alternatively names are used, here it's /dev/disk/by-uuid/<FS_UUID>.
>  
>  During the kdump boot progress, the /dev/disk/by-uuid/<FS_UUID> will exist as
>  soon as one of the path devices exists, but it won't be usable by systemd,
>  since multipathd will claim that device as a path device. Then multipathd will
>  get stopped before it can create the multipath device.
>  
>  Without user_friendly_name, /dev/mapper/<WWID> is considered a persistent
>  device name, avoiding the issue.
>  
>  The exit of multipathd is due to two dependencies in the current dracut module
>  90multipath/multipathd.service, "Before=initrd-cleanup.service" and
>  "Conflicts=initrd-cleanup.service".
>  
>  As per man 5 systemd.unit, if A.service has "Conflicts=B.service", starting
>  B.service will stop A.service.
>  
>  This is useful during normal boot. However, we will never switch-root after
>  capturing vmcore in kdump.
>  
>  We need to ensure that multipathd is not killed due to such dependency issue.
>  Without modifying multipathd.service, we add ConditionPathExists=!/proc/vmcore
>  to skip initrd-cleanup.service in kdump. This approach is beneficial as
>  it avoid the potential termination of other services that conflict with
>  initrd-cleanup.service. Also skip initrd-parse-etc.service as it will try to
>  start initrd-cleanup.service. Both of these services are used for switch root,
>  so they can be safely skipped in kdump.
>  
>  Suggested-by: Benjamin Marzinski <bmarzins(a)redhat.com&gt;
>  Suggested-by: Dave Young <dyoung(a)redhat.com&gt;
>  Signed-off-by: Lichen Liu <lichliu(a)redhat.com&gt; 
This patch fixes my tests as well.

Reviewed-by: Benjamin Marzinski <bmarzins(a)redhat.com&gt;

>  ---
>   dracut-module-setup.sh | 9 +++++++++
>   1 file changed, 9 insertions(+)
>  
>  diff --git a/dracut-module-setup.sh b/dracut-module-setup.sh
>  index 1dc88dc..0a901c2 100755
>  --- a/dracut-module-setup.sh
>  +++ b/dracut-module-setup.sh
>  @@ -1089,6 +1089,15 @@ install() {
>           's/\(^[[:space:]]*reserved_memory[[:space:]]*=\)[[:space:]]*[[:digit:]]*/\1
> 1024/' \
>           "${initdir}/etc/lvm/lvm.conf" &> /dev/null
>   
>  +    # Skip initrd-cleanup.service and initrd-parse-etc.service becasue we don't
>  +    # need to switch root. Instead of removing them, we use ConditionPathExists
>  +    # to check if /proc/vmcore exists to determine if we are in kdump.
>  +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
>  +        "${initdir}/${systemdsystemunitdir}/initrd-cleanup.service" &>
> /dev/null
>  +
>  +    sed -i '/\[Unit\]/a ConditionPathExists=!\/proc\/vmcore' \
>  +        "${initdir}/${systemdsystemunitdir}/initrd-parse-etc.service"
> &> /dev/null
>  +
>       # Save more memory by dropping switch root capability
>       dracut_no_switch_root
>   }
>  -- 
>  2.43.0