Re: [9fans] Backgrounding a task
> Think about multiple processes owned by multiple users running on a > cpu server. Which processes should be allowed to join which > namespaces? > > Perhaps allowing only the hostowner to join namespaces for debugging > and administration purposes would be acceptable. Ah, right. What about only allowing a process to join another namespace if the user is the same? > This seems a contrived example. Would you really spend HOURS working > on setting up a namespace by hand? Surely you would instead be > working on a script that builds the namespace for you; make the > computer do the work. Then when you mess up, you can modify the > script, create a new window, and try again. Yes, good point. That's the best way to do it. Also screen file can help add commands to the script post facto. The hours term isn't so much in the development of the environment but more to do with the amount of time you could be working in that shell instance. All the while not remembering all of the things you did to the namespace and environment in that time. Cheers, Chris
Re: [9fans] Backgrounding a task
Here it is: https://github.com/JehanneOS/jehanne/commit/320e6e6f35bfbc2e37dbd079c8d6a9124bd9ac6c The simple test attached confirms that it works as expected: https://github.com/JehanneOS/jehanne/blob/master/qa/kern/nsclone.c Now it's just matter of modifying the plumber to use this facility and add a ns/clone command that take a pid and a command to run so that ns/clone 256 rc would start a new rc in a copy of the name space of the process with pid 256. Giacomo 2017-10-24 21:18 GMT+02:00 Giacomo Tesio: > 2017-10-24 16:21 GMT+02:00 Alex Musolino : >> Creating a child process is something that a process explicitly >> controls and the RFNOTEG flag of rfork(2) allows a process to control >> whether or not it shares its namespace with its children. Allowing >> other, unrelated processes to fiddle with your namespace is quite >> different. >> >> Think about multiple processes owned by multiple users running on a >> cpu server. Which processes should be allowed to join which >> namespaces? >> >> Perhaps allowing only the hostowner to join namespaces for debugging >> and administration purposes would be acceptable. > > I like this idea a lot. I will give it a try in Jehanne. > > However I'm going to use a slightly different design: writing "clone" > to /proc/$pid/ns will cause the current process to replace its own > name space with a *copy* of that of $pid. > If the owner of $pid is different from that of the current process or > if $pid is not running on the same machine as the current process, the > write will fail with an error. > > However any change to the name space after the clone does not impact > the original process. > > As for the plumber, I will add a message that make the plumber clone > the name space of a target process. > > This should address both use-cases without issues for the processes > running in the original name space. > > > Giacomo
Re: [9fans] Backgrounding a task
2017-10-24 16:21 GMT+02:00 Alex Musolino: > Creating a child process is something that a process explicitly > controls and the RFNOTEG flag of rfork(2) allows a process to control > whether or not it shares its namespace with its children. Allowing > other, unrelated processes to fiddle with your namespace is quite > different. > > Think about multiple processes owned by multiple users running on a > cpu server. Which processes should be allowed to join which > namespaces? > > Perhaps allowing only the hostowner to join namespaces for debugging > and administration purposes would be acceptable. I like this idea a lot. I will give it a try in Jehanne. However I'm going to use a slightly different design: writing "clone" to /proc/$pid/ns will cause the current process to replace its own name space with a *copy* of that of $pid. If the owner of $pid is different from that of the current process or if $pid is not running on the same machine as the current process, the write will fail with an error. However any change to the name space after the clone does not impact the original process. As for the plumber, I will add a message that make the plumber clone the name space of a target process. This should address both use-cases without issues for the processes running in the original name space. Giacomo
Re: [9fans] Backgrounding a task
> The namespace join facility looks interesting. Do you have a patch > somewhere for it? I'll see what I can dig up though it wouldn't tbe erribly difficult to reimplement. You basically just need to modify the pgrp pointer of the proc, adjusting ref counts as required. >> Of course, a lot of the isolation that per-process namespaces give >> you is suddenly undone by the introduction of this facility. > > I'm not sure if the lack of isolation is any different than what can > be done with a child process that shares the namespace. Is there a > particular case that you are thinking? Creating a child process is something that a process explicitly controls and the RFNOTEG flag of rfork(2) allows a process to control whether or not it shares its namespace with its children. Allowing other, unrelated processes to fiddle with your namespace is quite different. Think about multiple processes owned by multiple users running on a cpu server. Which processes should be allowed to join which namespaces? Perhaps allowing only the hostowner to join namespaces for debugging and administration purposes would be acceptable. >> At this point I'm not entirely convinced that it's worth the >> trouble. > > I think that it can be depending on how much time you have spent > building up a namespace for a process. Perhaps I have spent hours > working on something slowly customizing the namespace mounting and > binding things. If I end up running a long running command that > blocks and I want to work in parallel with it then I must remember > everything that I have done and repeat in a new window. It seems > like something the computer should do for me or at least help me to do > it. This seems a contrived example. Would you really spend HOURS working on setting up a namespace by hand? Surely you would instead be working on a script that builds the namespace for you; make the computer do the work. Then when you mess up, you can modify the script, create a new window, and try again. One more thing to consider is the #σ device in 9front which seems to address some of the problems that you might otherwise use nsjoin to solve. -- Cheers, Alex Musolino
Re: [9fans] Backgrounding a task
The namespace join facility looks interesting. Do you have a patch somewhere for it? > Of course, a lot of the isolation that per-process namespaces give you > is suddenly undone by the introduction of this facility. I'm not sure if the lack of isolation is any different than what can be done with a child process that shares the namespace. Is there a particular case that you are thinking? > At this > point I'm not entirely convinced that it's worth the trouble. I think that it can be depending on how much time you have spent building up a namespace for a process. Perhaps I have spent hours working on something slowly customizing the namespace mounting and binding things. If I end up running a long running command that blocks and I want to work in parallel with it then I must remember everything that I have done and repeat in a new window. It seems like something the computer should do for me or at least help me to do it.
Re: [9fans] Backgrounding a task
> So far, it looks like the closest equivalent is to draw a new window > and inherit the namespace of the original one by reading the namespace > from the proc. The problem with /proc/$pid/ns is entries that can't be "replayed". For example, the following command will not work: mount -b '#|/data' /mnt Nor does it provide any real indication as to what exactly is hooked up to the other end of the 9P pipe. > I wonder if there could be a Rio gesture to draw a new window > inheriting the namespace of the window I pick by clicking. Rio can't access the namespaces of the processes running in its windows. The -m option to window(1) works by rforking a new process with the RFNAMEG flag set and then doing the appropriate mounting and binding to create and use a new rio window. Some months ago I played with the idea of adding support for an "nsjoin" command to the /proc/$pid/ctl file. This allowed you to join the namespace of another process a little something like this: term% ps | grep plumber alex6700:00 0:00 864K Preadplumber alex6710:00 0:00 864K Rendez plumber term% echo nsjoin 670 >/proc/$pid/ctl term% Subsequent changes to the namespace would affect the plumber. If you wanted a copy, just do 'rfork n' after joining the namespace. Of course, a lot of the isolation that per-process namespaces give you is suddenly undone by the introduction of this facility. At this point I'm not entirely convinced that it's worth the trouble. -- Cheers, Alex Musolino
Re: [9fans] Backgrounding a task
“window -m cmd” will run the command in the same namespace, forked, but in new window. “-m” is for “mount”, an alternative way of communication with the window system to /dev/wctl which is default. Yaroslav Kolomiiets 7 жовт. 2017 р. о 15:21 Chris McGeeпише: Thanks for the tip! I'll give that a try. Chris З мобільного > On Oct 7, 2017, at 12:04 AM, Skip Tavakkolian > wrote: > > Spitballing here: in the new window do something like > > cat /proc/123/ns | rc > > Or first massage the ns then generate an output for rc. > >> On Fri, Oct 6, 2017, 4:34 PM Chris McGee wrote: >> Hi All, >> >> When I'm using Unix, there's a workflow that I use for long running commands >> that I'm hoping to find the equivalent in the Plan 9 way of doing things. >> >> I will occasionally run a command, realize that it will take a long time to >> complete. I don't want to kill it. I'll just Ctrl-Z and bg to put it into >> the background using the shell. It's almost as if I had run it with '&' in >> the first place. I can then run other commands in the same working >> directory, environment and shell history. >> >> Is there an equivalent to this workflow in Plan 9? >> >> I realize that the whole job control system dates back to old single session >> terminals, which isn't a problem with Rio where you can draw new windows at >> will. Initially I thought, that you just drag that window to a corner >> somewhere and let it complete. But, if I draw a new window it won't be in >> the same working directory, have the same environment and namespace. Maybe >> there is a way to create a window that inherits these from an existing >> process? >> >> Chris
Re: [9fans] Backgrounding a task
Thanks for the tip! I'll give that a try. Chris > On Oct 7, 2017, at 12:04 AM, Skip Tavakkolian> wrote: > > Spitballing here: in the new window do something like > > cat /proc/123/ns | rc > > Or first massage the ns then generate an output for rc. > >> On Fri, Oct 6, 2017, 4:34 PM Chris McGee wrote: >> Hi All, >> >> When I'm using Unix, there's a workflow that I use for long running commands >> that I'm hoping to find the equivalent in the Plan 9 way of doing things. >> >> I will occasionally run a command, realize that it will take a long time to >> complete. I don't want to kill it. I'll just Ctrl-Z and bg to put it into >> the background using the shell. It's almost as if I had run it with '&' in >> the first place. I can then run other commands in the same working >> directory, environment and shell history. >> >> Is there an equivalent to this workflow in Plan 9? >> >> I realize that the whole job control system dates back to old single session >> terminals, which isn't a problem with Rio where you can draw new windows at >> will. Initially I thought, that you just drag that window to a corner >> somewhere and let it complete. But, if I draw a new window it won't be in >> the same working directory, have the same environment and namespace. Maybe >> there is a way to create a window that inherits these from an existing >> process? >> >> Chris
[9fans] Backgrounding a task
Hi All, When I'm using Unix, there's a workflow that I use for long running commands that I'm hoping to find the equivalent in the Plan 9 way of doing things. I will occasionally run a command, realize that it will take a long time to complete. I don't want to kill it. I'll just Ctrl-Z and bg to put it into the background using the shell. It's almost as if I had run it with '&' in the first place. I can then run other commands in the same working directory, environment and shell history. Is there an equivalent to this workflow in Plan 9? I realize that the whole job control system dates back to old single session terminals, which isn't a problem with Rio where you can draw new windows at will. Initially I thought, that you just drag that window to a corner somewhere and let it complete. But, if I draw a new window it won't be in the same working directory, have the same environment and namespace. Maybe there is a way to create a window that inherits these from an existing process? Chris