diff mbox series

[v3] depmod: Handle installing modules under a prefix

Message ID 20230714122111.7528-1-msuchanek@suse.de (mailing list archive)
State New, archived
Headers show
Series [v3] depmod: Handle installing modules under a prefix | expand

Commit Message

Michal Suchanek July 14, 2023, 12:21 p.m. UTC
Some distributions aim at not shipping any files in / outside of usr.

The path under which kernel modules are installed is hardcoded to /lib
which conflicts with this goal.

When kmod provides the config command, use it to determine the correct
module installation prefix.

This is a prefix under which the modules are searched by kmod on the
system, and is separate from the temporary staging location already
supported by INSTALL_MOD_PATH.

With kmod that does not provide the config command empty prefix is used
as before.

Signed-off-by: Michal Suchanek <msuchanek@suse.de>
---
v2: Avoid error on systems with kmod that does not support config
command
v3: More verbose commit message
---
 Makefile          | 4 +++-
 scripts/depmod.sh | 8 ++++----
 2 files changed, 7 insertions(+), 5 deletions(-)

Comments

Jan Engelhardt July 14, 2023, 1:38 p.m. UTC | #1
On Friday 2023-07-14 14:21, Michal Suchanek wrote:

>Some distributions aim at not shipping any files in / outside of usr.
>
>The path under which kernel modules are installed is hardcoded to /lib
>which conflicts with this goal.
>
>+MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)

Ok, so if the problem statement is that hardcoded paths are bad, then why
continue to hardcode the "/lib/modules" fragment? Just make it so that
KERNEL_MODULE_PREFIX can be set to the exact string "/usr/lib/modules" and not
just "/usr".
Michal Suchanek July 14, 2023, 1:57 p.m. UTC | #2
Hello,

On Fri, Jul 14, 2023 at 03:38:18PM +0200, Jan Engelhardt wrote:
> 
> On Friday 2023-07-14 14:21, Michal Suchanek wrote:
> 
> >Some distributions aim at not shipping any files in / outside of usr.
> >
> >The path under which kernel modules are installed is hardcoded to /lib
> >which conflicts with this goal.
> >
> >+MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> 
> Ok, so if the problem statement is that hardcoded paths are bad, then why
> continue to hardcode the "/lib/modules" fragment? Just make it so that
> KERNEL_MODULE_PREFIX can be set to the exact string "/usr/lib/modules" and not
> just "/usr".

That's certainly an option.

The feature is modelled after the installation prefix option that can
move the whole filesystem hierarchy of installed files under /usr/local,
/opt, or any other directory of choice. However, in that case the
subdirectories in the hierarchy can be configured as well while in this
case /lib/modules remains hardcoded.

Making it possible to set the whole path is generally more flexible
although there is no need to set the later part for this particular use
case.

Thanks

Michal
Michal Koutný July 14, 2023, 1:59 p.m. UTC | #3
On Fri, Jul 14, 2023 at 03:38:18PM +0200, Jan Engelhardt <jengelh@inai.de> wrote:
> Ok, so if the problem statement is that hardcoded paths are bad, then why
> continue to hardcode the "/lib/modules" fragment? Just make it so that
> KERNEL_MODULE_PREFIX can be set to the exact string "/usr/lib/modules" and not
> just "/usr".

That sounds cleaner but I'm worried it would be a BC break in setups
that expect the existing layout under INSTALL_MOD_PATH, wouldn't it?

Michal
Nicolas Schier July 14, 2023, 2:05 p.m. UTC | #4
On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> Some distributions aim at not shipping any files in / outside of usr.

For me, preventing negation often makes things easier, e.g.: "... aim at
shipping files only below /usr".

> 
> The path under which kernel modules are installed is hardcoded to /lib
> which conflicts with this goal.
> 
> When kmod provides the config command, use it to determine the correct
> module installation prefix.
> 
> This is a prefix under which the modules are searched by kmod on the
> system, and is separate from the temporary staging location already
> supported by INSTALL_MOD_PATH.
> 
> With kmod that does not provide the config command empty prefix is used
> as before.
> 
> Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> ---
> v2: Avoid error on systems with kmod that does not support config
> command
> v3: More verbose commit message
> ---
>  Makefile          | 4 +++-
>  scripts/depmod.sh | 8 ++++----
>  2 files changed, 7 insertions(+), 5 deletions(-)
> 
> diff --git a/Makefile b/Makefile
> index 47690c28456a..b1fea135bdec 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
>  # makefile but the argument can be passed to make if needed.
>  #
>  
> -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)

All other calls of `jq` that I could find are located at tools/; as this here
is evaluated on each invocation, this should probably be documented in
Documentation/process/changes.rst?

(Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)

> +
> +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
>  export MODLIB
>  
>  PHONY += prepare0
> diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> index 3643b4f896ed..88ac79056153 100755
> --- a/scripts/depmod.sh
> +++ b/scripts/depmod.sh
> @@ -27,16 +27,16 @@ fi
>  # numbers, so we cheat with a symlink here
>  depmod_hack_needed=true
>  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
>  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
>  		depmod_hack_needed=false
>  	fi
>  fi

I'd like to come back to the statement from Masahiro: Is the check above,
against some very old versions of depmod [1], the only reason for this patch?  

If we could remove that, would

    make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install

be sufficient?

Kind regards,
Nicolas


[1]: https://lore.kernel.org/linux-kbuild/1307631448-29848-5-git-send-email-mmarek@suse.cz/

>  rm -rf "$tmp_dir"
>  if $depmod_hack_needed; then
> -	symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
> +	symlink="$INSTALL_MOD_PATH$KERNEL_MODULE_PREFIX/lib/modules/99.98.$KERNELRELEASE"
>  	ln -s "$KERNELRELEASE" "$symlink"
>  	KERNELRELEASE=99.98.$KERNELRELEASE
>  fi
> -- 
> 2.41.0
Michal Suchanek July 14, 2023, 2:05 p.m. UTC | #5
Hello,

On Fri, Jul 14, 2023 at 03:59:17PM +0200, Michal Koutný wrote:
> On Fri, Jul 14, 2023 at 03:38:18PM +0200, Jan Engelhardt <jengelh@inai.de> wrote:
> > Ok, so if the problem statement is that hardcoded paths are bad, then why
> > continue to hardcode the "/lib/modules" fragment? Just make it so that
> > KERNEL_MODULE_PREFIX can be set to the exact string "/usr/lib/modules" and not
> > just "/usr".
> 
> That sounds cleaner but I'm worried it would be a BC break in setups
> that expect the existing layout under INSTALL_MOD_PATH, wouldn't it?

It's a break either way, the expected directory righ now is exactly
/lib/modules. /usr/lib/modules works to some extent for some use cases
only when the compatibility symlink lib -> usr/lib is present.

Thanks

Michal
Michal Suchanek July 14, 2023, 2:30 p.m. UTC | #6
Hello,

On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > Some distributions aim at not shipping any files in / outside of usr.
> 
> For me, preventing negation often makes things easier, e.g.: "... aim at
> shipping files only below /usr".
> 
> > 
> > The path under which kernel modules are installed is hardcoded to /lib
> > which conflicts with this goal.
> > 
> > When kmod provides the config command, use it to determine the correct
> > module installation prefix.
> > 
> > This is a prefix under which the modules are searched by kmod on the
> > system, and is separate from the temporary staging location already
> > supported by INSTALL_MOD_PATH.
> > 
> > With kmod that does not provide the config command empty prefix is used
> > as before.
> > 
> > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > ---
> > v2: Avoid error on systems with kmod that does not support config
> > command
> > v3: More verbose commit message
> > ---
> >  Makefile          | 4 +++-
> >  scripts/depmod.sh | 8 ++++----
> >  2 files changed, 7 insertions(+), 5 deletions(-)
> > 
> > diff --git a/Makefile b/Makefile
> > index 47690c28456a..b1fea135bdec 100644
> > --- a/Makefile
> > +++ b/Makefile
> > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> >  # makefile but the argument can be passed to make if needed.
> >  #
> >  
> > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> 
> All other calls of `jq` that I could find are located at tools/; as this here
> is evaluated on each invocation, this should probably be documented in
> Documentation/process/changes.rst?
> 
> (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)

That's a good point.

> 
> > +
> > +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> >  export MODLIB
> >  
> >  PHONY += prepare0
> > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > index 3643b4f896ed..88ac79056153 100755
> > --- a/scripts/depmod.sh
> > +++ b/scripts/depmod.sh
> > @@ -27,16 +27,16 @@ fi
> >  # numbers, so we cheat with a symlink here
> >  depmod_hack_needed=true
> >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> >  		depmod_hack_needed=false
> >  	fi
> >  fi
> 
> I'd like to come back to the statement from Masahiro: Is the check above,
> against some very old versions of depmod [1], the only reason for this patch?  
> 
> If we could remove that, would
> 
>     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> 
> be sufficient?

No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
the newly added part is not because it's integral part of where the
modules are installed on the system, and not the staging area path.

> 
> Kind regards,
> Nicolas
> 
> 
> [1]: https://lore.kernel.org/linux-kbuild/1307631448-29848-5-git-send-email-mmarek@suse.cz/

Was busybox ever fixed to not require the hack?

Thanks

Michal
Jan Engelhardt July 14, 2023, 2:42 p.m. UTC | #7
On Friday 2023-07-14 16:30, Michal Suchánek wrote:
>> >  
>> > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
>> > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
>> 
>> All other calls of `jq` that I could find are located at tools/; as this here
>> is evaluated on each invocation, this should probably be documented in
>> Documentation/process/changes.rst?
>> 
>> (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
>
>That's a good point.

Speaking of which, "&>" is a bashism and probably should be replaced.
Nicolas Schier July 14, 2023, 2:54 p.m. UTC | #8
On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> Hello,
> 
> On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > Some distributions aim at not shipping any files in / outside of usr.
> > 
> > For me, preventing negation often makes things easier, e.g.: "... aim at
> > shipping files only below /usr".
> > 
> > > 
> > > The path under which kernel modules are installed is hardcoded to /lib
> > > which conflicts with this goal.
> > > 
> > > When kmod provides the config command, use it to determine the correct
> > > module installation prefix.
> > > 
> > > This is a prefix under which the modules are searched by kmod on the
> > > system, and is separate from the temporary staging location already
> > > supported by INSTALL_MOD_PATH.
> > > 
> > > With kmod that does not provide the config command empty prefix is used
> > > as before.
> > > 
> > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > ---
> > > v2: Avoid error on systems with kmod that does not support config
> > > command
> > > v3: More verbose commit message
> > > ---
> > >  Makefile          | 4 +++-
> > >  scripts/depmod.sh | 8 ++++----
> > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > 
> > > diff --git a/Makefile b/Makefile
> > > index 47690c28456a..b1fea135bdec 100644
> > > --- a/Makefile
> > > +++ b/Makefile
> > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > >  # makefile but the argument can be passed to make if needed.
> > >  #
> > >  
> > > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> > 
> > All other calls of `jq` that I could find are located at tools/; as this here
> > is evaluated on each invocation, this should probably be documented in
> > Documentation/process/changes.rst?
> > 
> > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> 
> That's a good point.
> 
> > 
> > > +
> > > +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > >  export MODLIB
> > >  
> > >  PHONY += prepare0
> > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > index 3643b4f896ed..88ac79056153 100755
> > > --- a/scripts/depmod.sh
> > > +++ b/scripts/depmod.sh
> > > @@ -27,16 +27,16 @@ fi
> > >  # numbers, so we cheat with a symlink here
> > >  depmod_hack_needed=true
> > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > >  		depmod_hack_needed=false
> > >  	fi
> > >  fi
> > 
> > I'd like to come back to the statement from Masahiro: Is the check above,
> > against some very old versions of depmod [1], the only reason for this patch?  
> > 
> > If we could remove that, would
> > 
> >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > 
> > be sufficient?
> 
> No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> the newly added part is not because it's integral part of where the
> modules are installed on the system, and not the staging area path.

Ah, thanks.  So just for my understanding, could this be a (non-gentle)
alternative version of your patch, w/o modifying top-level Makefile?

diff --git a/scripts/depmod.sh b/scripts/depmod.sh
index 3643b4f896ed..72c819de0669 100755
--- a/scripts/depmod.sh
+++ b/scripts/depmod.sh
@@ -1,4 +1,4 @@
-#!/bin/sh
+#!/bin/bash
 # SPDX-License-Identifier: GPL-2.0
 #
 # A depmod wrapper used by the toplevel Makefile
@@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
        exit 0
 fi
 
+kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
+
 # older versions of depmod require the version string to start with three
 # numbers, so we cheat with a symlink here
 depmod_hack_needed=true
@@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
        fi
 fi
 rm -rf "$tmp_dir"
+
+if [ "${kmod_version}" -gt 32 ]; then
+       kmod_prefix="$(kmod config | jq -r .module_prefix)"
+       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
+       depmod_hack_needed=false
+fi
+
 if $depmod_hack_needed; then
        symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
        ln -s "$KERNELRELEASE" "$symlink"

(untested, and assuming that kmod module prefix is in kmod >= 32)

Or are I am still missing something?

> Was busybox ever fixed to not require the hack?

I haven't checked that, yet.

Kind regards,
Nicolas
Michal Suchanek July 14, 2023, 3:10 p.m. UTC | #9
On Fri, Jul 14, 2023 at 04:54:49PM +0200, Nicolas Schier wrote:
> On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > Hello,
> > 
> > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > Some distributions aim at not shipping any files in / outside of usr.
> > > 
> > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > shipping files only below /usr".
> > > 
> > > > 
> > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > which conflicts with this goal.
> > > > 
> > > > When kmod provides the config command, use it to determine the correct
> > > > module installation prefix.
> > > > 
> > > > This is a prefix under which the modules are searched by kmod on the
> > > > system, and is separate from the temporary staging location already
> > > > supported by INSTALL_MOD_PATH.
> > > > 
> > > > With kmod that does not provide the config command empty prefix is used
> > > > as before.
> > > > 
> > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > ---
> > > > v2: Avoid error on systems with kmod that does not support config
> > > > command
> > > > v3: More verbose commit message
> > > > ---
> > > >  Makefile          | 4 +++-
> > > >  scripts/depmod.sh | 8 ++++----
> > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > > 
> > > > diff --git a/Makefile b/Makefile
> > > > index 47690c28456a..b1fea135bdec 100644
> > > > --- a/Makefile
> > > > +++ b/Makefile
> > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > >  # makefile but the argument can be passed to make if needed.
> > > >  #
> > > >  
> > > > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> > > 
> > > All other calls of `jq` that I could find are located at tools/; as this here
> > > is evaluated on each invocation, this should probably be documented in
> > > Documentation/process/changes.rst?
> > > 
> > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> > 
> > That's a good point.
> > 
> > > 
> > > > +
> > > > +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > >  export MODLIB
> > > >  
> > > >  PHONY += prepare0
> > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > index 3643b4f896ed..88ac79056153 100755
> > > > --- a/scripts/depmod.sh
> > > > +++ b/scripts/depmod.sh
> > > > @@ -27,16 +27,16 @@ fi
> > > >  # numbers, so we cheat with a symlink here
> > > >  depmod_hack_needed=true
> > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > >  		depmod_hack_needed=false
> > > >  	fi
> > > >  fi
> > > 
> > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > against some very old versions of depmod [1], the only reason for this patch?  
> > > 
> > > If we could remove that, would
> > > 
> > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > > 
> > > be sufficient?
> > 
> > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > the newly added part is not because it's integral part of where the
> > modules are installed on the system, and not the staging area path.
> 
> Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> alternative version of your patch, w/o modifying top-level Makefile?
> 
> diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> index 3643b4f896ed..72c819de0669 100755
> --- a/scripts/depmod.sh
> +++ b/scripts/depmod.sh
> @@ -1,4 +1,4 @@
> -#!/bin/sh
> +#!/bin/bash
>  # SPDX-License-Identifier: GPL-2.0
>  #
>  # A depmod wrapper used by the toplevel Makefile
> @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
>         exit 0
>  fi
>  
> +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> +
>  # older versions of depmod require the version string to start with three
>  # numbers, so we cheat with a symlink here
>  depmod_hack_needed=true
> @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
>         fi
>  fi
>  rm -rf "$tmp_dir"
> +
> +if [ "${kmod_version}" -gt 32 ]; then
> +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> +       depmod_hack_needed=false
> +fi
> +
>  if $depmod_hack_needed; then
>         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
>         ln -s "$KERNELRELEASE" "$symlink"
> 
> (untested, and assuming that kmod module prefix is in kmod >= 32)

It can be detected by running the 'kmod config' command first and
ignoring the output when it fails which the above patch already did.
The version check does not sound very reliable.

> Or are I am still missing something?

MODLIB still needs to include the extra prefix so that files are
installed in the correct location. And that's defined in the toplevel
Makefile.

Thanks

Michal
Nicolas Schier July 14, 2023, 7:37 p.m. UTC | #10
On Fri, Jul 14, 2023 at 05:10:42PM +0200 Michal Suchánek wrote:
> On Fri, Jul 14, 2023 at 04:54:49PM +0200, Nicolas Schier wrote:
> > On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > > Hello,
> > > 
> > > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > > Some distributions aim at not shipping any files in / outside of usr.
> > > > 
> > > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > > shipping files only below /usr".
> > > > 
> > > > > 
> > > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > > which conflicts with this goal.
> > > > > 
> > > > > When kmod provides the config command, use it to determine the correct
> > > > > module installation prefix.
> > > > > 
> > > > > This is a prefix under which the modules are searched by kmod on the
> > > > > system, and is separate from the temporary staging location already
> > > > > supported by INSTALL_MOD_PATH.
> > > > > 
> > > > > With kmod that does not provide the config command empty prefix is used
> > > > > as before.
> > > > > 
> > > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > > ---
> > > > > v2: Avoid error on systems with kmod that does not support config
> > > > > command
> > > > > v3: More verbose commit message
> > > > > ---
> > > > >  Makefile          | 4 +++-
> > > > >  scripts/depmod.sh | 8 ++++----
> > > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > > > 
> > > > > diff --git a/Makefile b/Makefile
> > > > > index 47690c28456a..b1fea135bdec 100644
> > > > > --- a/Makefile
> > > > > +++ b/Makefile
> > > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > > >  # makefile but the argument can be passed to make if needed.
> > > > >  #
> > > > >  
> > > > > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)

oh, should this be 'jq -r .prefix' (w/o ".module") to match your other patches?

> > > > 
> > > > All other calls of `jq` that I could find are located at tools/; as this here
> > > > is evaluated on each invocation, this should probably be documented in
> > > > Documentation/process/changes.rst?
> > > > 
> > > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> > > 
> > > That's a good point.
> > > 
> > > > 
> > > > > +
> > > > > +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > > >  export MODLIB
> > > > >  
> > > > >  PHONY += prepare0
> > > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > > index 3643b4f896ed..88ac79056153 100755
> > > > > --- a/scripts/depmod.sh
> > > > > +++ b/scripts/depmod.sh
> > > > > @@ -27,16 +27,16 @@ fi
> > > > >  # numbers, so we cheat with a symlink here
> > > > >  depmod_hack_needed=true
> > > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > > -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > >  		depmod_hack_needed=false
> > > > >  	fi
> > > > >  fi
> > > > 
> > > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > > against some very old versions of depmod [1], the only reason for this patch?  
> > > > 
> > > > If we could remove that, would
> > > > 
> > > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > > > 
> > > > be sufficient?
> > > 
> > > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > > the newly added part is not because it's integral part of where the
> > > modules are installed on the system, and not the staging area path.
> > 
> > Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> > alternative version of your patch, w/o modifying top-level Makefile?
> > 
> > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > index 3643b4f896ed..72c819de0669 100755
> > --- a/scripts/depmod.sh
> > +++ b/scripts/depmod.sh
> > @@ -1,4 +1,4 @@
> > -#!/bin/sh
> > +#!/bin/bash
> >  # SPDX-License-Identifier: GPL-2.0
> >  #
> >  # A depmod wrapper used by the toplevel Makefile
> > @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
> >         exit 0
> >  fi
> >  
> > +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> > +
> >  # older versions of depmod require the version string to start with three
> >  # numbers, so we cheat with a symlink here
> >  depmod_hack_needed=true
> > @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> >         fi
> >  fi
> >  rm -rf "$tmp_dir"
> > +
> > +if [ "${kmod_version}" -gt 32 ]; then
> > +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> > +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> > +       depmod_hack_needed=false
> > +fi
> > +
> >  if $depmod_hack_needed; then
> >         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
> >         ln -s "$KERNELRELEASE" "$symlink"
> > 
> > (untested, and assuming that kmod module prefix is in kmod >= 32)
> 
> It can be detected by running the 'kmod config' command first and
> ignoring the output when it fails which the above patch already did.
> The version check does not sound very reliable.
> 
> > Or are I am still missing something?
> 
> MODLIB still needs to include the extra prefix so that files are
> installed in the correct location. And that's defined in the toplevel
> Makefile.

Well, I think that depends.  Technically, you are right; and if we want
to support system with a non-empty kmod prefix fully transparently, then
patching top-level Makefile will probably be necessary.  

As for me, I am not convinced yet, that the fully transparent way to support
PREFIX/lib/modules/ is the best way forward.  I think it might be better to
first only make script/depmod.sh fit for a kmod prefix and require an adjusted
INSTALL_MOD_PATH for modules_install.

Which concrete distributions did you have in mind while composing the patches?

Kind regards,
Nicolas
Michal Suchanek July 17, 2023, 9:55 a.m. UTC | #11
On Fri, Jul 14, 2023 at 09:37:04PM +0200, Nicolas Schier wrote:
> On Fri, Jul 14, 2023 at 05:10:42PM +0200 Michal Suchánek wrote:
> > On Fri, Jul 14, 2023 at 04:54:49PM +0200, Nicolas Schier wrote:
> > > On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > > > Hello,
> > > > 
> > > > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > > > Some distributions aim at not shipping any files in / outside of usr.
> > > > > 
> > > > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > > > shipping files only below /usr".
> > > > > 
> > > > > > 
> > > > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > > > which conflicts with this goal.
> > > > > > 
> > > > > > When kmod provides the config command, use it to determine the correct
> > > > > > module installation prefix.
> > > > > > 
> > > > > > This is a prefix under which the modules are searched by kmod on the
> > > > > > system, and is separate from the temporary staging location already
> > > > > > supported by INSTALL_MOD_PATH.
> > > > > > 
> > > > > > With kmod that does not provide the config command empty prefix is used
> > > > > > as before.
> > > > > > 
> > > > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > > > ---
> > > > > > v2: Avoid error on systems with kmod that does not support config
> > > > > > command
> > > > > > v3: More verbose commit message
> > > > > > ---
> > > > > >  Makefile          | 4 +++-
> > > > > >  scripts/depmod.sh | 8 ++++----
> > > > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > > > > 
> > > > > > diff --git a/Makefile b/Makefile
> > > > > > index 47690c28456a..b1fea135bdec 100644
> > > > > > --- a/Makefile
> > > > > > +++ b/Makefile
> > > > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > > > >  # makefile but the argument can be passed to make if needed.
> > > > > >  #
> > > > > >  
> > > > > > -MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> 
> oh, should this be 'jq -r .prefix' (w/o ".module") to match your other patches?

No, this aligns perfectly fine, prefix is where kmod is installed.

> 
> > > > > 
> > > > > All other calls of `jq` that I could find are located at tools/; as this here
> > > > > is evaluated on each invocation, this should probably be documented in
> > > > > Documentation/process/changes.rst?
> > > > > 
> > > > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> > > > 
> > > > That's a good point.
> > > > 
> > > > > 
> > > > > > +
> > > > > > +MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > > > >  export MODLIB
> > > > > >  
> > > > > >  PHONY += prepare0
> > > > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > > > index 3643b4f896ed..88ac79056153 100755
> > > > > > --- a/scripts/depmod.sh
> > > > > > +++ b/scripts/depmod.sh
> > > > > > @@ -27,16 +27,16 @@ fi
> > > > > >  # numbers, so we cheat with a symlink here
> > > > > >  depmod_hack_needed=true
> > > > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > > > -	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > > -		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > > +	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > > +		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > >  		depmod_hack_needed=false
> > > > > >  	fi
> > > > > >  fi
> > > > > 
> > > > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > > > against some very old versions of depmod [1], the only reason for this patch?  
> > > > > 
> > > > > If we could remove that, would
> > > > > 
> > > > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > > > > 
> > > > > be sufficient?
> > > > 
> > > > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > > > the newly added part is not because it's integral part of where the
> > > > modules are installed on the system, and not the staging area path.
> > > 
> > > Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> > > alternative version of your patch, w/o modifying top-level Makefile?
> > > 
> > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > index 3643b4f896ed..72c819de0669 100755
> > > --- a/scripts/depmod.sh
> > > +++ b/scripts/depmod.sh
> > > @@ -1,4 +1,4 @@
> > > -#!/bin/sh
> > > +#!/bin/bash
> > >  # SPDX-License-Identifier: GPL-2.0
> > >  #
> > >  # A depmod wrapper used by the toplevel Makefile
> > > @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
> > >         exit 0
> > >  fi
> > >  
> > > +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> > > +
> > >  # older versions of depmod require the version string to start with three
> > >  # numbers, so we cheat with a symlink here
> > >  depmod_hack_needed=true
> > > @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > >         fi
> > >  fi
> > >  rm -rf "$tmp_dir"
> > > +
> > > +if [ "${kmod_version}" -gt 32 ]; then
> > > +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> > > +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> > > +       depmod_hack_needed=false
> > > +fi
> > > +
> > >  if $depmod_hack_needed; then
> > >         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
> > >         ln -s "$KERNELRELEASE" "$symlink"
> > > 
> > > (untested, and assuming that kmod module prefix is in kmod >= 32)
> > 
> > It can be detected by running the 'kmod config' command first and
> > ignoring the output when it fails which the above patch already did.
> > The version check does not sound very reliable.
> > 
> > > Or are I am still missing something?
> > 
> > MODLIB still needs to include the extra prefix so that files are
> > installed in the correct location. And that's defined in the toplevel
> > Makefile.
> 
> Well, I think that depends.  Technically, you are right; and if we want
> to support system with a non-empty kmod prefix fully transparently, then
> patching top-level Makefile will probably be necessary.  
> 
> As for me, I am not convinced yet, that the fully transparent way to support
> PREFIX/lib/modules/ is the best way forward.  I think it might be better to
> first only make script/depmod.sh fit for a kmod prefix and require an adjusted
> INSTALL_MOD_PATH for modules_install.

Nevermind, I will update the patch to change the whole path. That will
make it crystal clear that no amount of fiddling with INSTALL_MOD_PATH
will work.

> Which concrete distributions did you have in mind while composing the patches?

If you STFW for usrmerge you can find that a number of distributions is
in different stages of experimenting with packing all shipped files into
/usr.

In the early experiment stages when the distribution is built with /lib
/bin, and /sbin in place and massaged after the fact to remove these
directories it does not matter much where files were installed
initially.

In later stages when the distribution is built without these extra
directories they are missing in the staging area for building packages,
and creating them is an error because files in these directories are not
to be shipped in packages. At this point the kernel insisting that
modules must be in /lib is sticking out like a sore thumb, it's on every
system.

Sure, there is a number of hacks that can be used to work around the
problem. Still it's a problem the needs to be addressed for usrmerge to
fully work.

openSUSE is maybe 75% there - it's nowhere near complete but the
packages that insist on using these directories outside of /usr are
becoming problematic.

Thanks

Michal
Masahiro Yamada July 17, 2023, 7:13 p.m. UTC | #12
On Fri, Jul 14, 2023 at 11:55 PM Nicolas Schier <nicolas@fjasle.eu> wrote:
>
> On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > Hello,
> >
> > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > Some distributions aim at not shipping any files in / outside of usr.
> > >
> > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > shipping files only below /usr".
> > >
> > > >
> > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > which conflicts with this goal.
> > > >
> > > > When kmod provides the config command, use it to determine the correct
> > > > module installation prefix.
> > > >
> > > > This is a prefix under which the modules are searched by kmod on the
> > > > system, and is separate from the temporary staging location already
> > > > supported by INSTALL_MOD_PATH.
> > > >
> > > > With kmod that does not provide the config command empty prefix is used
> > > > as before.
> > > >
> > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > ---
> > > > v2: Avoid error on systems with kmod that does not support config
> > > > command
> > > > v3: More verbose commit message
> > > > ---
> > > >  Makefile          | 4 +++-
> > > >  scripts/depmod.sh | 8 ++++----
> > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > >
> > > > diff --git a/Makefile b/Makefile
> > > > index 47690c28456a..b1fea135bdec 100644
> > > > --- a/Makefile
> > > > +++ b/Makefile
> > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > >  # makefile but the argument can be passed to make if needed.
> > > >  #
> > > >
> > > > -MODLIB   = $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> > >
> > > All other calls of `jq` that I could find are located at tools/; as this here
> > > is evaluated on each invocation, this should probably be documented in
> > > Documentation/process/changes.rst?
> > >
> > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> >
> > That's a good point.
> >
> > >
> > > > +
> > > > +MODLIB   = $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > >  export MODLIB
> > > >
> > > >  PHONY += prepare0
> > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > index 3643b4f896ed..88ac79056153 100755
> > > > --- a/scripts/depmod.sh
> > > > +++ b/scripts/depmod.sh
> > > > @@ -27,16 +27,16 @@ fi
> > > >  # numbers, so we cheat with a symlink here
> > > >  depmod_hack_needed=true
> > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > - if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > -         -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > + if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > +         -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > >           depmod_hack_needed=false
> > > >   fi
> > > >  fi
> > >
> > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > against some very old versions of depmod [1], the only reason for this patch?
> > >
> > > If we could remove that, would
> > >
> > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > >
> > > be sufficient?
> >
> > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > the newly added part is not because it's integral part of where the
> > modules are installed on the system, and not the staging area path.
>
> Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> alternative version of your patch, w/o modifying top-level Makefile?
>
> diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> index 3643b4f896ed..72c819de0669 100755
> --- a/scripts/depmod.sh
> +++ b/scripts/depmod.sh
> @@ -1,4 +1,4 @@
> -#!/bin/sh
> +#!/bin/bash
>  # SPDX-License-Identifier: GPL-2.0
>  #
>  # A depmod wrapper used by the toplevel Makefile
> @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
>         exit 0
>  fi
>
> +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> +
>  # older versions of depmod require the version string to start with three
>  # numbers, so we cheat with a symlink here
>  depmod_hack_needed=true
> @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
>         fi
>  fi
>  rm -rf "$tmp_dir"
> +
> +if [ "${kmod_version}" -gt 32 ]; then
> +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> +       depmod_hack_needed=false
> +fi
> +
>  if $depmod_hack_needed; then
>         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
>         ln -s "$KERNELRELEASE" "$symlink"
>
> (untested, and assuming that kmod module prefix is in kmod >= 32)
>
> Or are I am still missing something?
>
> > Was busybox ever fixed to not require the hack?
>
> I haven't checked that, yet.



I believe we can revert

8fc62e59425389a6d48429b9d146223122743435
bfe5424a8b31624e7a476f959d552999f931e7c7

There is no good reason to keep such old hacky code.


The old module-init-tools project was replaced by kmod.
I do not know whether busybox fixed the issue or not.
Anyway, Linux 3.0 was released 12 years ago.
There was plenty of time to fix the issue if we wanted.
If busybox is still not able to handle the X.Y version form,
it is just that nobody is caring about the busybox's depmod.




--
Best Regards
Masahiro Yamada
Masahiro Yamada July 17, 2023, 7:14 p.m. UTC | #13
On Sat, Jul 15, 2023 at 12:10 AM Michal Suchánek <msuchanek@suse.de> wrote:
>
> On Fri, Jul 14, 2023 at 04:54:49PM +0200, Nicolas Schier wrote:
> > On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > > Hello,
> > >
> > > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > > Some distributions aim at not shipping any files in / outside of usr.
> > > >
> > > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > > shipping files only below /usr".
> > > >
> > > > >
> > > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > > which conflicts with this goal.
> > > > >
> > > > > When kmod provides the config command, use it to determine the correct
> > > > > module installation prefix.
> > > > >
> > > > > This is a prefix under which the modules are searched by kmod on the
> > > > > system, and is separate from the temporary staging location already
> > > > > supported by INSTALL_MOD_PATH.
> > > > >
> > > > > With kmod that does not provide the config command empty prefix is used
> > > > > as before.
> > > > >
> > > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > > ---
> > > > > v2: Avoid error on systems with kmod that does not support config
> > > > > command
> > > > > v3: More verbose commit message
> > > > > ---
> > > > >  Makefile          | 4 +++-
> > > > >  scripts/depmod.sh | 8 ++++----
> > > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > > >
> > > > > diff --git a/Makefile b/Makefile
> > > > > index 47690c28456a..b1fea135bdec 100644
> > > > > --- a/Makefile
> > > > > +++ b/Makefile
> > > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > > >  # makefile but the argument can be passed to make if needed.
> > > > >  #
> > > > >
> > > > > -MODLIB = $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> > > >
> > > > All other calls of `jq` that I could find are located at tools/; as this here
> > > > is evaluated on each invocation, this should probably be documented in
> > > > Documentation/process/changes.rst?
> > > >
> > > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> > >
> > > That's a good point.
> > >
> > > >
> > > > > +
> > > > > +MODLIB = $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > > >  export MODLIB
> > > > >
> > > > >  PHONY += prepare0
> > > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > > index 3643b4f896ed..88ac79056153 100755
> > > > > --- a/scripts/depmod.sh
> > > > > +++ b/scripts/depmod.sh
> > > > > @@ -27,16 +27,16 @@ fi
> > > > >  # numbers, so we cheat with a symlink here
> > > > >  depmod_hack_needed=true
> > > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > > -       if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > -               -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > +       if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > +               -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > >                 depmod_hack_needed=false
> > > > >         fi
> > > > >  fi
> > > >
> > > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > > against some very old versions of depmod [1], the only reason for this patch?
> > > >
> > > > If we could remove that, would
> > > >
> > > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > > >
> > > > be sufficient?
> > >
> > > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > > the newly added part is not because it's integral part of where the
> > > modules are installed on the system, and not the staging area path.
> >
> > Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> > alternative version of your patch, w/o modifying top-level Makefile?
> >
> > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > index 3643b4f896ed..72c819de0669 100755
> > --- a/scripts/depmod.sh
> > +++ b/scripts/depmod.sh
> > @@ -1,4 +1,4 @@
> > -#!/bin/sh
> > +#!/bin/bash
> >  # SPDX-License-Identifier: GPL-2.0
> >  #
> >  # A depmod wrapper used by the toplevel Makefile
> > @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
> >         exit 0
> >  fi
> >
> > +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> > +
> >  # older versions of depmod require the version string to start with three
> >  # numbers, so we cheat with a symlink here
> >  depmod_hack_needed=true
> > @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> >         fi
> >  fi
> >  rm -rf "$tmp_dir"
> > +
> > +if [ "${kmod_version}" -gt 32 ]; then
> > +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> > +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> > +       depmod_hack_needed=false
> > +fi
> > +
> >  if $depmod_hack_needed; then
> >         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
> >         ln -s "$KERNELRELEASE" "$symlink"
> >
> > (untested, and assuming that kmod module prefix is in kmod >= 32)
>
> It can be detected by running the 'kmod config' command first and
> ignoring the output when it fails which the above patch already did.
> The version check does not sound very reliable.
>
> > Or are I am still missing something?
>
> MODLIB still needs to include the extra prefix so that files are
> installed in the correct location. And that's defined in the toplevel
> Makefile.
>
> Thanks
>
> Michal



As Documentation/kbuild/kbuild.rst mentions,
you can override MODLIB.

Kbuild already provides the maximum flexibility.


$ make modules_install \
  INSTALL_MOD_PATH=/path/to/whatever/staging/dir \
  MODLIB=/path/to/whatever/real/install/destination
Michal Suchanek July 18, 2023, 8:52 a.m. UTC | #14
On Tue, Jul 18, 2023 at 04:14:11AM +0900, Masahiro Yamada wrote:
> On Sat, Jul 15, 2023 at 12:10 AM Michal Suchánek <msuchanek@suse.de> wrote:
> >
> > On Fri, Jul 14, 2023 at 04:54:49PM +0200, Nicolas Schier wrote:
> > > On Fri, Jul 14, 2023 at 04:30:02PM +0200, Michal Such�nek wrote:
> > > > Hello,
> > > >
> > > > On Fri, Jul 14, 2023 at 04:05:10PM +0200, Nicolas Schier wrote:
> > > > > On Fri, Jul 14, 2023 at 02:21:08PM +0200 Michal Suchanek wrote:
> > > > > > Some distributions aim at not shipping any files in / outside of usr.
> > > > >
> > > > > For me, preventing negation often makes things easier, e.g.: "... aim at
> > > > > shipping files only below /usr".
> > > > >
> > > > > >
> > > > > > The path under which kernel modules are installed is hardcoded to /lib
> > > > > > which conflicts with this goal.
> > > > > >
> > > > > > When kmod provides the config command, use it to determine the correct
> > > > > > module installation prefix.
> > > > > >
> > > > > > This is a prefix under which the modules are searched by kmod on the
> > > > > > system, and is separate from the temporary staging location already
> > > > > > supported by INSTALL_MOD_PATH.
> > > > > >
> > > > > > With kmod that does not provide the config command empty prefix is used
> > > > > > as before.
> > > > > >
> > > > > > Signed-off-by: Michal Suchanek <msuchanek@suse.de>
> > > > > > ---
> > > > > > v2: Avoid error on systems with kmod that does not support config
> > > > > > command
> > > > > > v3: More verbose commit message
> > > > > > ---
> > > > > >  Makefile          | 4 +++-
> > > > > >  scripts/depmod.sh | 8 ++++----
> > > > > >  2 files changed, 7 insertions(+), 5 deletions(-)
> > > > > >
> > > > > > diff --git a/Makefile b/Makefile
> > > > > > index 47690c28456a..b1fea135bdec 100644
> > > > > > --- a/Makefile
> > > > > > +++ b/Makefile
> > > > > > @@ -1165,7 +1165,9 @@ export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
> > > > > >  # makefile but the argument can be passed to make if needed.
> > > > > >  #
> > > > > >
> > > > > > -MODLIB = $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
> > > > > > +export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
> > > > >
> > > > > All other calls of `jq` that I could find are located at tools/; as this here
> > > > > is evaluated on each invocation, this should probably be documented in
> > > > > Documentation/process/changes.rst?
> > > > >
> > > > > (Absence of `jq` will cause error messages, even with CONFIG_MODULES=n.)
> > > >
> > > > That's a good point.
> > > >
> > > > >
> > > > > > +
> > > > > > +MODLIB = $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
> > > > > >  export MODLIB
> > > > > >
> > > > > >  PHONY += prepare0
> > > > > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > > > > index 3643b4f896ed..88ac79056153 100755
> > > > > > --- a/scripts/depmod.sh
> > > > > > +++ b/scripts/depmod.sh
> > > > > > @@ -27,16 +27,16 @@ fi
> > > > > >  # numbers, so we cheat with a symlink here
> > > > > >  depmod_hack_needed=true
> > > > > >  tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
> > > > > > -mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
> > > > > > +mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
> > > > > >  if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > > > > > -       if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > > -               -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > > +       if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
> > > > > > +               -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
> > > > > >                 depmod_hack_needed=false
> > > > > >         fi
> > > > > >  fi
> > > > >
> > > > > I'd like to come back to the statement from Masahiro: Is the check above,
> > > > > against some very old versions of depmod [1], the only reason for this patch?
> > > > >
> > > > > If we could remove that, would
> > > > >
> > > > >     make INSTALL_MOD_PATH="$(kmod config | jq -r .module_prefix)" modules_install
> > > > >
> > > > > be sufficient?
> > > >
> > > > No, the INSTALL_MOD_PATH is passed as the -b argument to depmod while
> > > > the newly added part is not because it's integral part of where the
> > > > modules are installed on the system, and not the staging area path.
> > >
> > > Ah, thanks.  So just for my understanding, could this be a (non-gentle)
> > > alternative version of your patch, w/o modifying top-level Makefile?
> > >
> > > diff --git a/scripts/depmod.sh b/scripts/depmod.sh
> > > index 3643b4f896ed..72c819de0669 100755
> > > --- a/scripts/depmod.sh
> > > +++ b/scripts/depmod.sh
> > > @@ -1,4 +1,4 @@
> > > -#!/bin/sh
> > > +#!/bin/bash
> > >  # SPDX-License-Identifier: GPL-2.0
> > >  #
> > >  # A depmod wrapper used by the toplevel Makefile
> > > @@ -23,6 +23,8 @@ if [ -z $(command -v $DEPMOD) ]; then
> > >         exit 0
> > >  fi
> > >
> > > +kmod_version=$(( $(kmod --version | sed -rne 's/^kmod version ([0-9]+).*$/\1/p') ))
> > > +
> > >  # older versions of depmod require the version string to start with three
> > >  # numbers, so we cheat with a symlink here
> > >  depmod_hack_needed=true
> > > @@ -35,6 +37,13 @@ if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
> > >         fi
> > >  fi
> > >  rm -rf "$tmp_dir"
> > > +
> > > +if [ "${kmod_version}" -gt 32 ]; then
> > > +       kmod_prefix="$(kmod config | jq -r .module_prefix)"
> > > +       INSTALL_MOD_PATH="${INSTALL_MOD_PATH#${kmod_prefix}"
> > > +       depmod_hack_needed=false
> > > +fi
> > > +
> > >  if $depmod_hack_needed; then
> > >         symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
> > >         ln -s "$KERNELRELEASE" "$symlink"
> > >
> > > (untested, and assuming that kmod module prefix is in kmod >= 32)
> >
> > It can be detected by running the 'kmod config' command first and
> > ignoring the output when it fails which the above patch already did.
> > The version check does not sound very reliable.
> >
> > > Or are I am still missing something?
> >
> > MODLIB still needs to include the extra prefix so that files are
> > installed in the correct location. And that's defined in the toplevel
> > Makefile.
> >
> > Thanks
> >
> > Michal
> 
> 
> 
> As Documentation/kbuild/kbuild.rst mentions,
> you can override MODLIB.
> 
> Kbuild already provides the maximum flexibility.
> 
> 
> $ make modules_install \
>   INSTALL_MOD_PATH=/path/to/whatever/staging/dir \
>   MODLIB=/path/to/whatever/real/install/destination

That's great that we have another workaround for the case when the
kernel build system and kmod do not agree on the location of the kernel
modules.

However, they should by default agree, and if this feature is merged
into kmod the kernel should also adapt to make sure the modules are
installed where kmod expects them.

Thanks

Michal
diff mbox series

Patch

diff --git a/Makefile b/Makefile
index 47690c28456a..b1fea135bdec 100644
--- a/Makefile
+++ b/Makefile
@@ -1165,7 +1165,9 @@  export INSTALL_DTBS_PATH ?= $(INSTALL_PATH)/dtbs/$(KERNELRELEASE)
 # makefile but the argument can be passed to make if needed.
 #
 
-MODLIB	= $(INSTALL_MOD_PATH)/lib/modules/$(KERNELRELEASE)
+export KERNEL_MODULE_PREFIX := $(shell kmod config &> /dev/null && kmod config | jq -r .module_prefix)
+
+MODLIB	= $(INSTALL_MOD_PATH)$(KERNEL_MODULE_PREFIX)/lib/modules/$(KERNELRELEASE)
 export MODLIB
 
 PHONY += prepare0
diff --git a/scripts/depmod.sh b/scripts/depmod.sh
index 3643b4f896ed..88ac79056153 100755
--- a/scripts/depmod.sh
+++ b/scripts/depmod.sh
@@ -27,16 +27,16 @@  fi
 # numbers, so we cheat with a symlink here
 depmod_hack_needed=true
 tmp_dir=$(mktemp -d ${TMPDIR:-/tmp}/depmod.XXXXXX)
-mkdir -p "$tmp_dir/lib/modules/$KERNELRELEASE"
+mkdir -p "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE"
 if "$DEPMOD" -b "$tmp_dir" $KERNELRELEASE 2>/dev/null; then
-	if test -e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep" -o \
-		-e "$tmp_dir/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
+	if test -e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep" -o \
+		-e "$tmp_dir$KERNEL_MODULE_PREFIX/lib/modules/$KERNELRELEASE/modules.dep.bin"; then
 		depmod_hack_needed=false
 	fi
 fi
 rm -rf "$tmp_dir"
 if $depmod_hack_needed; then
-	symlink="$INSTALL_MOD_PATH/lib/modules/99.98.$KERNELRELEASE"
+	symlink="$INSTALL_MOD_PATH$KERNEL_MODULE_PREFIX/lib/modules/99.98.$KERNELRELEASE"
 	ln -s "$KERNELRELEASE" "$symlink"
 	KERNELRELEASE=99.98.$KERNELRELEASE
 fi