[[!meta copyright="Copyright © 2010, 2013, 2014 Free Software Foundation, Inc."]] [[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable id="license" text="Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled [[GNU Free Documentation License|/fdl]]."]]"""]] [[!meta title="AF_BUS, D-Bus, and the Linux kernel"]] Might be interesting to watch how this develops. [[!toc]] # IRC, freenode, #hurd, about 2010-08/2010-09 <neal> check this out: <neal> someone is working on implementing dbus in linux <neal> linux finally gets mach ipc ;-) <marcusb> it's old news though, unless there is an update <marcusb> and I think it was only the client? <neal> youpi : someone is adding dbus ipc to the linux kernel <neal> marcusb: I just heard about it. <youpi> (it's crazy how this drives backward compared to a hurdish approach) <youpi> what is the motivation for moving to the kernel? <neal> context switch overhead <azeem_> they wanna use it to talk to device drivers? :) <kilobug> well, they did that with the in-kernel web server, but they abandonned it later on <neal> azeem: I don't think so. <neal> dbus in the kernel is actually good for the Hurd as dbus IPC is basically neutered Mach IPC <marcusb> I don't think anybody wants to put the dbus server in the kernel <neal> well, there is at least one person <marcusb> maybe this is a different news from the one I read <neal> Alban Crequy (albanc) is working out. He works for collabora, fwiw <http://alban.apinc.org/blog/2010/09/15/d-bus-in-the-kernel-faster/> <marcusb> what I read was about hal etc <marcusb> so that you don't need a user space daemon to glue the kernel to the dbus world <neal> I don't think that is what he is talking about <marcusb> I can't find it anymore though. I mentioned it in this channel at the time though, so it should be in the backlog <marcusb> neal, yeah could very well be a separate thing <marcusb> neal, dbus does have marginal support for fd passing though, and some attempts on the mailing list to make "fds" an official type in the message failed (as far as I could see, I didn't read the whole discussion) <marcusb> so no mach ipc just yet <neal> wrong <neal> FD handling is in 1.4 <neal> type o, if I'm not mistaken <marcusb> then the discussion moved on from initial rejection <neal> no, 'h' <marcusb> I'm out of date by two months <marcusb> ok <guillem> neal: AFAIR Marcel Holtmann talked about dbus in-kernel several years ago, but he never ended up implementing it, or there were rumors he had private "working code" * Related Mailing List Discussion * [\[PATCH 0/5\] RFC: Multicast and filtering features on AF_UNIX](http://article.gmane.org/gmane.linux.kernel/1040481), 2010-09-24 # 2013-02 [AF_BUS, D-Bus, and the Linux kernel](http://www.kroah.com/log/linux/af_bus.html), Greg Kroah-Hartman, 2013-02-08. # kdbus ## IRC, freenode, #hurd, 2014-01-28 <braunr> i would like to see things like dbus and zeromq use an optimized microkernel transport one day <teythoon> we could port kdbus >,< <braunr> why not <braunr> you port cgroups first <teythoon> exactly <braunr> :p [[systemd]]. ## IRC, freenode, #hurd, 2014-02-23 In context of [[linux_as_the_kernel]], *IRC, freenode, #hurd, 2014-02-23*. <desrt> mach seems like this really simple thing when you first explain what a microkernel is <braunr> and because of that, i think it's better to start the right solution directly <braunr> it looks simple, it's clearly not <desrt> but i did a bit of looking into it... it's a bit non-trivial after all :) <braunr> mach ipc is over complicated and error prone <braunr> it leads to unefficient communication compared to other solutions such as what l4 does <desrt> ya -- i hear that this is a big part of the performance hit <braunr> that's why i've started x15 <desrt> i was also doing some reading about how it's based on mapping memory segments between processes <braunr> first, it was a mach clone, but since i've come to know mach better, it's now a "spiritual" mach successor .. :) <desrt> these are two issues that we've been dealing with at another level... in the design of kdbus <braunr> ah kdbus :) <desrt> this is something that started with my masters thesis a long time ago... <braunr> ah you too <desrt> first thing we did is make the serialisation format so that all messages are valid and therefore never need to be checked <desrt> (old dbus format requires checks at every step on the way) <braunr> looks interesting <desrt> then of course we cut the daemon out <desrt> but some other interesting things: security is super-simple... it's based enirely on endpoints <desrt> either you're allowed to send messages between two processes or you're not <desrt> there is no checking for message types, for example <braunr> yes <desrt> and the other thing: memory mapping is usually bad <braunr> that's what i mean when i say mach ipc is over complicated <braunr> it depends <desrt> the kdbus guys did some performance testing and found out that if the message is less than ~512k then the cost of invalidating the TLB in order to change the memory mapping is higher than the cost of just copying the data <braunr> yes, we know that too <braunr> that's why zero copy isn't the normal way of passing small amounts of data over mach either <desrt> nice <desrt> i got the impression in some of my reading (wikipedia, honestly) that memory mapping was being done all the time <braunr> well <braunr> no it's not <braunr> memory mapping is unfortunately a small fraction of the performance overhead <desrt> that's good :) <braunr> that being said <braunr> memory mapping can be very useful <braunr> for example, it's hard for us to comply with posix requirements of being able to read/write at least 2G of data in a single call <braunr> weird bugs occur beyond 512M iirc <braunr> you do want memory mapping for that <desrt> ya... for things of this size.... you don't want to copy that through a socket :) <braunr> monolithic kernels have it naturally, since the kernel is mapped everywhere <braunr> for microkernels, it's a little more complicated <braunr> and the problem gets worse on smp <braunr> again, that's why i preferred starting a new kernel instead of reusing linux