Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On Mon, 2013-11-11 at 10:52 +0800, ChenQi wrote: On 11/10/2013 07:00 AM, Phil Blundell wrote: 1. initscript doesn't obviously rdepend on busybox so it's not obvious that the latter will always be available; Yes. Initscript doesn't rdepend on busybox. But note it also doesn't rdepend on sed or awk or grep. So I think it's reasonable to assume the presence of busybox. I think one could argue that it's also a bug that it doesn't rdepend on the three things you mentioned (and indeed on /bin/sh itself, for non-rpm systems). However, the usage of grep and sed in particular, and perhaps to a slightly lesser extent awk, is so deeply ingrained into so much software that I think it's probably fair to say those utilities will almost always be present. By contrast, it is perfectly feasible to build a system which doesn't use busybox (by using the full GNU implementations of everything that busybox provides) and indeed I think there might even be a task package in oe-core which does exactly that. So it seems entirely possible that /bin/busybox might not be installed unless you have a dependency on it. 2. it should probably be using ${base_bindir} and ${bindir} rather than hardcoding absolute paths. In init scripts, we usually hardcode things, because these scripts are destined to run on target. In recipes we try not to hardcode things because the recipe may need to extend to native or nativesdk. I'm not quite sure I follow the logic here. Even if the script is intended to run on the target it still ought to be respecting the values of ${bindir}, ${sysconfdir} and such like. 3. the whole idea of creating a shadow /usr/bin underneath what's meant to be a mountpoint seems rather dubious to me. Agree. The problem here is that the init scripts under /etc/rcS.d/ need to execute commands like awk, dirname, and readlink which are from /usr. As it's not appropriate to move these commands into /bin, basically there are only two options I can see here. One is to modify these scripts to use only commands from /bin or /sbin; the other is to make use of busybox, as busybox is located under /bin as it provides these commands. I chose to the latter one because I thought that solution would have the less impact. What do you think? Do we need to modify the init scripts? Or any other solution? I thought that last time this topic came up on the mailing list, the eventual conclusion was that Wind River (being more-or-less the only people who seemed to feel strongly that supporting / and /usr on different partitions was important) were going to come up with some sort of overall strategy for solving the whole problem rather than just fixing up individual bits in a piecemeal fashion. I think that strategy would need to include a policy for which utilities initscripts can legitimately expect to be available before /usr is mounted, and if this set is different from what we have today then it would need to include a plan for sorting that out. To answer the particular question at hand it seems that someone would need to do some analysis of things like: - how many scripts actually need those commands - how hard would it be to change them to not need them - what would be the impact of moving them into ${base_bindir} (for example, would this cause a whole load of libraries to get dragged into ${base_libdir} as well?) Offhand I don't know the answer to any of those things. 4. this seems like distro policy and not something that really belongs in oe-core at all. For systems where ${bindir} and ${base_bindir} are on the same filesystem (or even are the same directory) this script will just make bootup slower without achieving anything useful. If /usr and / are on the same file system, this script has no real effect because /usr will always be there. So I think it will not take much time at boot. Just spawning a new copy of the shell and reading the script does take a finite (and measureable) time. If you can determine statically at build time that the script isn't going to do anything useful then I don't think it's appropriate to install it. If ${base_bindir} and ${bindir} are the same, that means that there's no /usr. In this case, there should no /usr/xxx entries in /etc/busybox.links, so this script should also function correctly and it will not take much time at boot. (I just configured ${bindir} and ${base_bindir} to be the same and performed a build, it failed. I'm not sure whether it's valid to make such configurations in OE.) It is a valid configuration, and this is the way that most of the images I build are configured. It probably is true that there are still some number of recipes in oe-core that don't support this properly, but I don't want to see that situation get any worse. p. ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On 11 November 2013 02:52, ChenQi qi.c...@windriver.com wrote: The problem here is that the init scripts under /etc/rcS.d/ need to execute commands like awk, dirname, and readlink which are from /usr. Yes. So why do you really need to support split /usr, and why isn't an initramfs a suitable alternative? People seem to insist that split /usr is trivial but these patches are clearly showing it's not. Ross ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
Hi Phil, First of all, thank you for your careful review and explanation. To conclude, you suggest modifying the init scripts and moving commands around if needed, right? And please see comments inline. On 11/11/2013 07:53 PM, Phil Blundell wrote: On Mon, 2013-11-11 at 10:52 +0800, ChenQi wrote: On 11/10/2013 07:00 AM, Phil Blundell wrote: 1. initscript doesn't obviously rdepend on busybox so it's not obvious that the latter will always be available; Yes. Initscript doesn't rdepend on busybox. But note it also doesn't rdepend on sed or awk or grep. So I think it's reasonable to assume the presence of busybox. I think one could argue that it's also a bug that it doesn't rdepend on the three things you mentioned (and indeed on /bin/sh itself, for non-rpm systems). However, the usage of grep and sed in particular, and perhaps to a slightly lesser extent awk, is so deeply ingrained into so much software that I think it's probably fair to say those utilities will almost always be present. By contrast, it is perfectly feasible to build a system which doesn't use busybox (by using the full GNU implementations of everything that busybox provides) and indeed I think there might even be a task package in oe-core which does exactly that. So it seems entirely possible that /bin/busybox might not be installed unless you have a dependency on it. Agree. I'm gonna drop this patch and try to come up with another solution :) 2. it should probably be using ${base_bindir} and ${bindir} rather than hardcoding absolute paths. In init scripts, we usually hardcode things, because these scripts are destined to run on target. In recipes we try not to hardcode things because the recipe may need to extend to native or nativesdk. I'm not quite sure I follow the logic here. Even if the script is intended to run on the target it still ought to be respecting the values of ${bindir}, ${sysconfdir} and such like. 3. the whole idea of creating a shadow /usr/bin underneath what's meant to be a mountpoint seems rather dubious to me. Agree. The problem here is that the init scripts under /etc/rcS.d/ need to execute commands like awk, dirname, and readlink which are from /usr. As it's not appropriate to move these commands into /bin, basically there are only two options I can see here. One is to modify these scripts to use only commands from /bin or /sbin; the other is to make use of busybox, as busybox is located under /bin as it provides these commands. I chose to the latter one because I thought that solution would have the less impact. What do you think? Do we need to modify the init scripts? Or any other solution? I thought that last time this topic came up on the mailing list, the eventual conclusion was that Wind River (being more-or-less the only people who seemed to feel strongly that supporting / and /usr on different partitions was important) were going to come up with some sort of overall strategy for solving the whole problem rather than just fixing up individual bits in a piecemeal fashion. I think that strategy would need to include a policy for which utilities initscripts can legitimately expect to be available before /usr is mounted, and if this set is different from what we have today then it would need to include a plan for sorting that out. Basically we don't have any problem if /usr is going to be mounted, because the mountall.sh takes place really early. What I want to achieve here is that we can boot into single user mode for recovery or repair if /usr is not available (maybe just because the drive's broken). To answer the particular question at hand it seems that someone would need to do some analysis of things like: - how many scripts actually need those commands - how hard would it be to change them to not need them Only init scripts that have links under /etc/rcS.d/ need to be examined, so that wouldn't be a lot of work. I'm just not sure whether we should make such a restriction of which commands could be used and which could not. After all, commands like `readlink', `dirname' and `awk' are also very commonly used commands. - what would be the impact of moving them into ${base_bindir} (for example, would this cause a whole load of libraries to get dragged into ${base_libdir} as well?) If I can easily modify the init scripts to make things work, I don't want to move things around. But I guess statements using awk are hard to be replaced unless I use a block of codes Offhand I don't know the answer to any of those things. 4. this seems like distro policy and not something that really belongs in oe-core at all. For systems where ${bindir} and ${base_bindir} are on the same filesystem (or even are the same directory) this script will just make bootup slower without achieving anything useful. If /usr and / are on the same file system, this script has no real effect because /usr will always be there. So I think it will not take much time at boot. Just
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On Mon, 2013-11-11 at 20:40 +0800, ChenQi wrote: Hi Phil, First of all, thank you for your careful review and explanation. To conclude, you suggest modifying the init scripts and moving commands around if needed, right? Well, I'm not sure that I'm necessarily suggesting any particular change. What I'm saying is that if we're going to try to solve the split-/usr problem in oe-core (as opposed to Wind River just doing it in their internal distro layers) then: 1. It needs to be done in a way that doesn't have any negative impact for people who don't have a split /usr (or indeed who don't have /usr at all) 2. The changes need to be presented as a coherent package with the reasoning clearly explained. 3. We need to be clear that it's solving the whole problem (whatever that might be) rather than just some part of it. Could you please share your configuration so that when I make V2 I can test the case of ${bindir} and ${base_bindir} being the same. (Currently I really can't build out an image with such configuration ... ) I think it should basically just be a case of setting: prefix = exec_prefix = prefix_native = in your distro configuration. p. ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On 11/11/13, 5:53 AM, Phil Blundell wrote: On Mon, 2013-11-11 at 10:52 +0800, ChenQi wrote: On 11/10/2013 07:00 AM, Phil Blundell wrote: ... I thought that last time this topic came up on the mailing list, the eventual conclusion was that Wind River (being more-or-less the only people who seemed to feel strongly that supporting / and /usr on different partitions was important) were going to come up with some sort of overall strategy for solving the whole problem rather than just fixing up individual bits in a piecemeal fashion. I think that strategy would need to include a policy for which utilities initscripts can legitimately expect to be available before /usr is mounted, and if this set is different from what we have today then it would need to include a plan for sorting that out. There is a bug open for the Yocto Project 1.6 to explicitly do this. Come up with an overall strategy for these kinds of things, as well as a test plan to verify that whatever we end up doing works properly long term. As you said, busybox is not required on a system -- just something that has a reasonable set of software packages. Also a lot of this stuff is simply limited to 'early boot'. At some point we do need to define 'early boot'. (Generally I define it as until /usr would normally be mounted.) And unfortunately you will see patches in piecemeal at this point. Until such a strategy is generated, everything is being done reactively. We see a QA error/warning and someone is assigned to solve it. Yocto Project Compliance (and our own internal) guidelines then indicate since we've patched something it has to go out to the community. To answer the particular question at hand it seems that someone would need to do some analysis of things like: - how many scripts actually need those commands - how hard would it be to change them to not need them - what would be the impact of moving them into ${base_bindir} (for example, would this cause a whole load of libraries to get dragged into ${base_libdir} as well?) That was my thought exactly. Lets look at moving things -or- stop using them. Whichever is more reasonable. Offhand I don't know the answer to any of those things. 4. this seems like distro policy and not something that really belongs in oe-core at all. For systems where ${bindir} and ${base_bindir} are on the same filesystem (or even are the same directory) this script will just make bootup slower without achieving anything useful. If /usr and / are on the same file system, this script has no real effect because /usr will always be there. So I think it will not take much time at boot. Just spawning a new copy of the shell and reading the script does take a finite (and measureable) time. If you can determine statically at build time that the script isn't going to do anything useful then I don't think it's appropriate to install it. If ${base_bindir} and ${bindir} are the same, that means that there's no /usr. In this case, there should no /usr/xxx entries in /etc/busybox.links, so this script should also function correctly and it will not take much time at boot. (I just configured ${bindir} and ${base_bindir} to be the same and performed a build, it failed. I'm not sure whether it's valid to make such configurations in OE.) It is a valid configuration, and this is the way that most of the images I build are configured. It probably is true that there are still some number of recipes in oe-core that don't support this properly, but I don't want to see that situation get any worse. And yes, all scripts need to (at build time) bind to the bindir and base_bindir of that distribution. Hardcoding those values into the scripts before build time is wrong. --Mark p. ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On 11/11/13, 6:12 AM, Burton, Ross wrote: On 11 November 2013 02:52, ChenQi qi.c...@windriver.com wrote: The problem here is that the init scripts under /etc/rcS.d/ need to execute commands like awk, dirname, and readlink which are from /usr. Yes. So why do you really need to support split /usr, and why isn't an initramfs a suitable alternative? People seem to insist that split /usr is trivial but these patches are clearly showing it's not. Just to be clear, I've never said it was trivial -- but what I have said is that it -should- be trivial if people pay attention to this stuff early. The only thing that has to be done for the split is 'early boot'. (See my email response to Phil..) As for why, we still have people developing systems with small per-device memory that need to do the early boot from a small system, then mount the /usr partition from a shared location. (/usr is often times still mounted RO.. and is shared between multiple CPUs within a single device or chasis.) The usage of this is getting less and less, but we do still get issues from customers when things don't work right, so we know people are still doing it. --Mark Ross ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On 11/10/2013 07:00 AM, Phil Blundell wrote: On Sat, 2013-11-09 at 13:28 +0800, qi.c...@windriver.com wrote: +for dir in /usr/bin /usr/sbin; do +if [ ! -e $dir ]; then + if [ $VERBOSE != no ]; then + echo WARN: $dir missing, setting up links to busybox + fi + mkdir -p $dir + for suffix in .nosuid .suid ; do + if [ ! -e /etc/busybox.links${suffix} ]; then + continue + fi + usr_commands=`grep $dir /etc/busybox.links${suffix}` + for command in $usr_commands; do + ln -sf /bin/busybox${suffix} $command + done Hi Phil, This script is an effort to try to make system boot up even if /usr is not there. It relies on busybox because in most cases, busybox is there in an OE based system. And please see comments below. This seems slightly bogus for a number of reasons: 1. initscript doesn't obviously rdepend on busybox so it's not obvious that the latter will always be available; Yes. Initscript doesn't rdepend on busybox. But note it also doesn't rdepend on sed or awk or grep. So I think it's reasonable to assume the presence of busybox. 2. it should probably be using ${base_bindir} and ${bindir} rather than hardcoding absolute paths. In init scripts, we usually hardcode things, because these scripts are destined to run on target. In recipes we try not to hardcode things because the recipe may need to extend to native or nativesdk. 3. the whole idea of creating a shadow /usr/bin underneath what's meant to be a mountpoint seems rather dubious to me. Agree. The problem here is that the init scripts under /etc/rcS.d/ need to execute commands like awk, dirname, and readlink which are from /usr. As it's not appropriate to move these commands into /bin, basically there are only two options I can see here. One is to modify these scripts to use only commands from /bin or /sbin; the other is to make use of busybox, as busybox is located under /bin as it provides these commands. I chose to the latter one because I thought that solution would have the less impact. What do you think? Do we need to modify the init scripts? Or any other solution? 4. this seems like distro policy and not something that really belongs in oe-core at all. For systems where ${bindir} and ${base_bindir} are on the same filesystem (or even are the same directory) this script will just make bootup slower without achieving anything useful. p. If /usr and / are on the same file system, this script has no real effect because /usr will always be there. So I think it will not take much time at boot. If ${base_bindir} and ${bindir} are the same, that means that there's no /usr. In this case, there should no /usr/xxx entries in /etc/busybox.links, so this script should also function correctly and it will not take much time at boot. (I just configured ${bindir} and ${base_bindir} to be the same and performed a build, it failed. I'm not sure whether it's valid to make such configurations in OE.) Best Regards, Chen Qi ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core
Re: [OE-core] [PATCH 2/8] initscripts: add setup-commands.sh
On Sat, 2013-11-09 at 13:28 +0800, qi.c...@windriver.com wrote: +for dir in /usr/bin /usr/sbin; do +if [ ! -e $dir ]; then + if [ $VERBOSE != no ]; then + echo WARN: $dir missing, setting up links to busybox + fi + mkdir -p $dir + for suffix in .nosuid .suid ; do + if [ ! -e /etc/busybox.links${suffix} ]; then + continue + fi + usr_commands=`grep $dir /etc/busybox.links${suffix}` + for command in $usr_commands; do + ln -sf /bin/busybox${suffix} $command + done This seems slightly bogus for a number of reasons: 1. initscript doesn't obviously rdepend on busybox so it's not obvious that the latter will always be available; 2. it should probably be using ${base_bindir} and ${bindir} rather than hardcoding absolute paths. 3. the whole idea of creating a shadow /usr/bin underneath what's meant to be a mountpoint seems rather dubious to me. 4. this seems like distro policy and not something that really belongs in oe-core at all. For systems where ${bindir} and ${base_bindir} are on the same filesystem (or even are the same directory) this script will just make bootup slower without achieving anything useful. p. ___ Openembedded-core mailing list Openembedded-core@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-core