Hi, While peeking over how you do things in crux it, I've found some things that could potentially be adjusted. Short overview: - NAME= is ignored since v154 (excluding renaming network interfaces) - ignore_device no longer exists since v148 (and ignore_remove since v152) - additional 20090523 rules are in large part invalid in context of the above points (and looking at standard udev rules, somewhat superfluous) - trigger's default is 'change' since v152 - for coldplug, 'add' should be specified explicitly. - devtmpfs - if compiled in kernel, should likely be preferred over tmpfs for /dev - is "fixing" libudev.so placement really necessary ? It forces you to keep track of libudev version and it sometimes get out of sync - pkg-config also scans (under default configuration) /usr/share/pkgconfig, so /usr/lib/pkgconfig is potentially not necessary - udevd at one point uses /proc/kmsg, so it could be added to static nodes Also since v155, udevd automatically copies static nodes from /lib/udev/devices and creates symbolic links. I think that Pkgfile could be made a bit more simple, and start_udev a bit more universal (taking into account that some people will use kernels with devtmpfs, boot with a custom initramfs, etc.) Anyway, a few possible tirival changes for your consideration: diff --git a/udev/Pkgfile b/udev/Pkgfile index e0d8b04..e4b7014 100644 --- a/udev/Pkgfile +++ b/udev/Pkgfile @@ -6,7 +6,6 @@ name=udev version=163 release=1 source=(ftp://ftp.kernel.org/pub/linux/utils/kernel/hotplug/$name-$version.tar.bz2 \ - http://crux.nu/files/distfiles/udev-config-20090523.tar.bz2 \ start_udev) build() { @@ -21,29 +20,16 @@ build() { pkgconfigdir=/usr/lib/pkgconfig mkdir -p $PKG/lib/{firmware,udev/devices/{pts,shm}} - mkdir -p $PKG/usr/lib - - # Fix location of devel lib - rm $PKG/lib/libudev.so - ln -sf ../../lib/libudev.so.0.9.1 $PKG/usr/lib/libudev.so # Populate nodes mknod -m 600 $PKG/lib/udev/devices/console c 5 1 + mknod -m 600 $PKG/lib/udev/devices/kmsg c 1 11 mknod -m 666 $PKG/lib/udev/devices/null c 1 3 mknod -m 666 $PKG/lib/udev/devices/zero c 1 5 - ln -s /proc/self/fd $PKG/lib/udev/devices/fd - ln -s /proc/self/fd/0 $PKG/lib/udev/devices/stdin - ln -s /proc/self/fd/1 $PKG/lib/udev/devices/stdout - ln -s /proc/self/fd/2 $PKG/lib/udev/devices/stderr - ln -s /proc/kcore $PKG/lib/udev/devices/core - # Add CRUX items + LFS rules + # Add CRUX items install -m 0755 $SRC/start_udev $PKG/sbin - pushd $SRC/udev-config-20090523 - make DESTDIR=$PKG RULES_DIR=/lib/udev/rules.d install - popd - # Keep udevinfo for legacy support mkdir -p $PKG/usr/bin ln -s /sbin/udevadm $PKG/usr/bin/udevinfo diff --git a/udev/start_udev b/udev/start_udev index 1be7ce9..35cdb2b 100644 --- a/udev/start_udev +++ b/udev/start_udev @@ -1,17 +1,32 @@ #!/bin/sh -# mount /dev as a tmpfs; note: some video drivers require exec access in /dev -/bin/mount -n -t tmpfs udev /dev -o exec,nosuid,mode=0755 +# mount /dev as a devtmpfs or tmpfs, depending on current kernel capability +# mount only, if it's not already mounted (e.g. after handover from initramfs) +# note: some video drivers require exec access in /dev +if /bin/grep -Fvq ' / /dev/ ' /proc/self/mountinfo ; then + if /bin/grep -Fq devtmpfs /proc/filesystems ; then + UDEVFS=devtmpfs + else + UDEVFS=tmpfs + fi + /bin/mount -n -t $UDEVFS udev /dev -o exec,nosuid,mode=0755,size=8M +fi -# unset hotplugger +# make sure hotplugger is not set echo > /proc/sys/kernel/hotplug -# populate /dev with static nodes and directories -/bin/cp -a /lib/udev/devices/* /dev/ +# since v155, udevd automatically copies /lib/udev/devices +# and creates /proc/{kcore,self/fd/{0,1,2}} symlinks # launch udev daemon /sbin/udevd --daemon # coldplug devices and wait for the queue to be processed -/sbin/udevadm trigger +/sbin/udevadm trigger --type=subsystems --action=add +/sbin/udevadm settle +/sbin/udevadm trigger --type=devices --action=add +/sbin/udevadm settle + +# retry any failures +/sbin/udevadm trigger --type=failed --action=add /sbin/udevadm settle
On Tue, Nov 16, 2010 at 09:47:27PM +0100, Michal Soltys wrote:
Hi,
Hello Michal,
While peeking over how you do things in crux it, I've found some things that could potentially be adjusted. Short overview:
Thanks. I'm happy to see that we have a new udev-helper, particularly because udev is a not-so-loved port among the core-ports ;) Needless to say that help is alway welcome for any kind of stuff. See some first comments below, I'll try to do some testing the next days.
- NAME= is ignored since v154 (excluding renaming network interfaces) - ignore_device no longer exists since v148 (and ignore_remove since v152) - additional 20090523 rules are in large part invalid in context of the above points (and looking at standard udev rules, somewhat superfluous) - trigger's default is 'change' since v152 - for coldplug, 'add' should be specified explicitly. - devtmpfs - if compiled in kernel, should likely be preferred over tmpfs for /dev - is "fixing" libudev.so placement really necessary ? It forces you to keep track of libudev version and it sometimes get out of sync
if we keep the *.so link in /lib we have to adjust libudev.pc, which is the same/more effort. Having the *.so library links in /usr/lib is reasonable furthermore.
- pkg-config also scans (under default configuration) /usr/share/pkgconfig, so /usr/lib/pkgconfig is potentially not necessary
Sure, but the correct place for libudev.pc is /usr/lib/pkgconfig.
- udevd at one point uses /proc/kmsg, so it could be added to static nodes
Also since v155, udevd automatically copies static nodes from /lib/udev/devices and creates symbolic links.
I think that Pkgfile could be made a bit more simple, and start_udev a bit more universal (taking into account that some people will use kernels with devtmpfs, boot with a custom initramfs, etc.)
Anyway, a few possible tirival changes for your consideration:
diff --git a/udev/Pkgfile b/udev/Pkgfile index e0d8b04..e4b7014 100644 --- a/udev/Pkgfile +++ b/udev/Pkgfile @@ -6,7 +6,6 @@ name=udev version=163 release=1 source=(ftp://ftp.kernel.org/pub/linux/utils/kernel/hotplug/$name-$version.tar.bz2 \ - http://crux.nu/files/distfiles/udev-config-20090523.tar.bz2 \ start_udev)
build() { @@ -21,29 +20,16 @@ build() { pkgconfigdir=/usr/lib/pkgconfig
mkdir -p $PKG/lib/{firmware,udev/devices/{pts,shm}} - mkdir -p $PKG/usr/lib - - # Fix location of devel lib - rm $PKG/lib/libudev.so - ln -sf ../../lib/libudev.so.0.9.1 $PKG/usr/lib/libudev.so
# Populate nodes mknod -m 600 $PKG/lib/udev/devices/console c 5 1 + mknod -m 600 $PKG/lib/udev/devices/kmsg c 1 11 mknod -m 666 $PKG/lib/udev/devices/null c 1 3 mknod -m 666 $PKG/lib/udev/devices/zero c 1 5 - ln -s /proc/self/fd $PKG/lib/udev/devices/fd - ln -s /proc/self/fd/0 $PKG/lib/udev/devices/stdin - ln -s /proc/self/fd/1 $PKG/lib/udev/devices/stdout - ln -s /proc/self/fd/2 $PKG/lib/udev/devices/stderr - ln -s /proc/kcore $PKG/lib/udev/devices/core
- # Add CRUX items + LFS rules + # Add CRUX items install -m 0755 $SRC/start_udev $PKG/sbin
- pushd $SRC/udev-config-20090523 - make DESTDIR=$PKG RULES_DIR=/lib/udev/rules.d install - popd - # Keep udevinfo for legacy support mkdir -p $PKG/usr/bin ln -s /sbin/udevadm $PKG/usr/bin/udevinfo diff --git a/udev/start_udev b/udev/start_udev index 1be7ce9..35cdb2b 100644 --- a/udev/start_udev +++ b/udev/start_udev @@ -1,17 +1,32 @@ #!/bin/sh
-# mount /dev as a tmpfs; note: some video drivers require exec access in /dev -/bin/mount -n -t tmpfs udev /dev -o exec,nosuid,mode=0755 +# mount /dev as a devtmpfs or tmpfs, depending on current kernel capability +# mount only, if it's not already mounted (e.g. after handover from initramfs) +# note: some video drivers require exec access in /dev
This dosn't work, because our grep is in /usr/bin and not in /bin. Might be usefull to have grep in /bin, but if we move grep we have to either move libpcre as well or build grep without pcre. It's something we should consider for the next version of CRUX. But I'd say we can use mountpoint/sed instead, even though the inverted logic of the sed line is a bit messy (untested):
+if /bin/grep -Fvq ' / /dev/ ' /proc/self/mountinfo ; then + if /bin/grep -Fq devtmpfs /proc/filesystems ; then + UDEVFS=devtmpfs + else + UDEVFS=tmpfs + fi + /bin/mount -n -t $UDEVFS udev /dev -o exec,nosuid,mode=0755,size=8M +fi
if ! /bin/mountpoint -q /dev ; then if ! /bin/sed -n '/devtmpfs/q1' /proc/filesystems ; then UDEVFS=devtmpfs else UDEVFS=tmpfs fi /bin/mount -n -t $UDEVFS udev /dev -o exec,nosuid,mode=0755,size=8M fi
-# unset hotplugger +# make sure hotplugger is not set echo > /proc/sys/kernel/hotplug
-# populate /dev with static nodes and directories -/bin/cp -a /lib/udev/devices/* /dev/ +# since v155, udevd automatically copies /lib/udev/devices +# and creates /proc/{kcore,self/fd/{0,1,2}} symlinks
# launch udev daemon /sbin/udevd --daemon
# coldplug devices and wait for the queue to be processed -/sbin/udevadm trigger +/sbin/udevadm trigger --type=subsystems --action=add +/sbin/udevadm settle +/sbin/udevadm trigger --type=devices --action=add +/sbin/udevadm settle + +# retry any failures +/sbin/udevadm trigger --type=failed --action=add /sbin/udevadm settle
thanks again and best regards Juergen -- Juergen Daubert | mailto:jue@jue.li Korb, Germany | http://jue.li/crux
On 10-11-17 17:06, Juergen Daubert wrote:
On Tue, Nov 16, 2010 at 09:47:27PM +0100, Michal Soltys wrote:
- is "fixing" libudev.so placement really necessary ? It forces you to keep track of libudev version and it sometimes get out of sync
if we keep the *.so link in /lib we have to adjust libudev.pc, which is the same/more effort. Having the *.so library links in /usr/lib is reasonable furthermore.
In such case, to not hunt lib version, maybe something like: ln -sf ../../lib/$(/bin/readlink $PKG/lib/libudev.so) $PKG/usr/lib/libudev.so rm -f $PKG/lib/libudev.so or more roboust: LIBUDEV=$(/bin/readlink $PKG/lib/libudev.so) if [[ -f "$PKG/lib/$LIBUDEV" ]]; then ln -sf ../../lib/$LIBUDEV $PKG/usr/lib/libudev.so rm -f $PKG/lib/libudev.so else echo "/lib/libudev.so is no longer present, please inspect the Pkgfile" exit 1 fi
- pkg-config also scans (under default configuration) /usr/share/pkgconfig, so /usr/lib/pkgconfig is potentially not necessary
Sure, but the correct place for libudev.pc is /usr/lib/pkgconfig.
Gah, I mixed 2 things. Current Pkgfile adjusts /lib/pkgconfig to /usr/lib/pkgconfig, and I was looking at udev.pc (which ends in /usr/share/pkgconfig).
if ! /bin/mountpoint -q /dev ; then if ! /bin/sed -n '/devtmpfs/q1' /proc/filesystems ; then UDEVFS=devtmpfs else UDEVFS=tmpfs fi /bin/mount -n -t $UDEVFS udev /dev -o exec,nosuid,mode=0755,size=8M fi
/bin/mountpoint seems to be fine. Alternatively sed could be used in similar fashion, e.g. sed -n '/ \/ \/dev /q1' /proc/self/mountinfo or more simply: sed -n '/ \/dev /q1' /proc/self/mounts
# coldplug devices and wait for the queue to be processed -/sbin/udevadm trigger +/sbin/udevadm trigger --type=subsystems --action=add +/sbin/udevadm settle +/sbin/udevadm trigger --type=devices --action=add +/sbin/udevadm settle + +# retry any failures +/sbin/udevadm trigger --type=failed --action=add /sbin/udevadm settle
'settle' between 'subsystems' and 'devices' is probably unnecessary (e.g. opensuse doesn't have one). I think I'll ask on udev list about reference way to coldplug these days. Lots of distros still don't bother with --type, or even --action. Looking at libudev/libudev-enumerate.c it seems both 'subsystems' and 'devices' should be done, but I'm not completely sure.
On Thu, Nov 18, 2010 at 01:32:29AM +0100, Michal Soltys wrote: [...]
'settle' between 'subsystems' and 'devices' is probably unnecessary (e.g. opensuse doesn't have one). I think I'll ask on udev list about reference way to coldplug these days. Lots of distros still don't bother with --type, or even --action. Looking at libudev/libudev-enumerate.c it seems both 'subsystems' and 'devices' should be done, but I'm not completely sure.
Well, I think it's not a big loss of time to call settle, though the README [1] doesn't show it in the recommended setup sequence. If you get new informations, please report back. We've done some tests yesterday and found only a small issue. We have to cp 3 devices manually if tmpfs is used as the filesystem for /dev, without that udevd complains at system start about a missing /dev/null. Not a serious problem, because everything else seems to work fine. That's explained in the README too and not necessary with devtmpfs. I've collected all the changes we've discussed here into a new udev port, which is in my privat repo for testing [2]. Please report any issues. Thanks in advance to all for testing and best regards Juergen [1] http://tinyurl.com/36o5dxa [2] httpup sync http://jue.li/crux/ports/#udev udev -- Juergen Daubert | mailto:jue@jue.li Korb, Germany | http://jue.li/crux
On 10-11-19 12:59, Juergen Daubert wrote:
On Thu, Nov 18, 2010 at 01:32:29AM +0100, Michal Soltys wrote:
[...]
'settle' between 'subsystems' and 'devices' is probably unnecessary (e.g. opensuse doesn't have one). I think I'll ask on udev list about reference way to coldplug these days. Lots of distros still don't bother with --type, or even --action. Looking at libudev/libudev-enumerate.c it seems both 'subsystems' and 'devices' should be done, but I'm not completely sure.
Well, I think it's not a big loss of time to call settle, though the README [1] doesn't show it in the recommended setup sequence. If you get new informations, please report back.
We've done some tests yesterday and found only a small issue. We have to cp 3 devices manually if tmpfs is used as the filesystem for /dev, without that udevd complains at system start about a missing /dev/null. Not a serious problem, because everything else seems to work fine. That's explained in the README too and not necessary with devtmpfs.
After checking udevd.c it looks that the function responsible for creating standard symlinks and static+module exported nodes is called after /dev/null check. I suppose the order could be changed - I'll do small patch and see what udev people think.
On 10-11-19 17:40, Michal Soltys wrote:
On 10-11-19 12:59, Juergen Daubert wrote:
On Thu, Nov 18, 2010 at 01:32:29AM +0100, Michal Soltys wrote:
We've done some tests yesterday and found only a small issue. We have to cp 3 devices manually if tmpfs is used as the filesystem for /dev, without that udevd complains at system start about a missing /dev/null. Not a serious problem, because everything else seems to work fine. That's explained in the README too and not necessary with devtmpfs.
After checking udevd.c it looks that the function responsible for creating standard symlinks and static+module exported nodes is called after /dev/null check. I suppose the order could be changed - I'll do small patch and see what udev people think. _______________________________________________
http://git.kernel.org/?p=linux/hotplug/udev.git;a=commitdiff;h=1a6ab670304dc...
On Tue, Nov 16, 2010 at 09:47:27PM +0100, Michal Soltys wrote:
Hi,
While peeking over how you do things in crux it, I've found some things that could potentially be adjusted.
I've committed a new version of the udev port yesterday. Thanks again for your contribution. regards Juergen -- Juergen Daubert | mailto:jue@jue.li Korb, Germany | http://jue.li/crux
On 03.12.2010 08:28, Juergen Daubert wrote:
On Tue, Nov 16, 2010 at 09:47:27PM +0100, Michal Soltys wrote:
Hi,
While peeking over how you do things in crux it, I've found some things that could potentially be adjusted.
I've committed a new version of the udev port yesterday. Thanks again for your contribution.
regards Juergen
Yea, just saw in git. Btw - once v165 get relased, you can remove extra processing for tmpfs case.
participants (2)
-
Juergen Daubert -
Michal Soltys