Mach initiating RPCs to userspace tasks

IRC, freenode, #hurd, 2011-06-11

<antrik> I don't think we have a precendence case of Mach initiating RPCs
  to userspace tasks
<braunr> well mach regularly sends RPCs to external pagers
<antrik> hm, right
<antrik> anyways, the ds_ in device.defs is for use *inside* Mach, not for
  the userspace interface
<braunr> what makes you think so ?
<antrik> several things
<antrik> not least the fact that without zhengda's modifications, the
  device handling never calls out to userspace for all I know
<braunr> hm, it does
<braunr> for async I/O
<braunr> when the kernel has finished its I/O, it calls
  ds_device_read_reply/ds_device_write_reply
<antrik> I see
<antrik> I never quite understood the _reply stuff
<braunr> although i wonder how mig is supposed to forge those names
<antrik> braunr: it isn't
<antrik> braunr: there is a separate device_reply.defs
<antrik> braunr: and it sets a *userprefix* of ds_
<antrik> rather than a serverprefix
<braunr> i saw, yes
<braunr> ah right
<antrik> so ds still refers to the in-Mach device server, not anything
  userspace
<braunr> so this is where the patch is supposed to introduce the
  device_intr_notify RPC
<antrik> or at least that's my understanding...
<braunr> no, it doesn't refer to in-mach servers
<braunr> it really forges the right rpcs to be called by mach
<antrik> the definition of "RPC" is rather unclear here
<braunr> why ?
<braunr> mach has its own mach_msg() call for kernel-to-user messaging
<antrik> yes, but this is used only to send the reply message for the RPC
  earlier initiated by userspace AIUI
<antrik> it doesn't look like there is any special RPC for async I/O
<braunr> yes, because this is the only use case they had
<braunr> hence the name "reply"
<braunr> intr_notify isn't a reply, but it uses the same mechanism
<braunr> these are declared as simpleroutine
<antrik> sure. but the fact that it isn't a reply message, but rather
  initiates a new RPC, changes things from MiG point of view I believe
<antrik> right, as there is no reply to the reply :-)
<braunr> :)
<braunr> a simpleroutine is how to turn an rpc into a simple ipc
<antrik> I know
<antrik> so in _reply, we pretend that the reply is actually a new RPC,
  with server and client roles reversed, and no reply
<antrik> (this is actually rather kludgy... apparently MIG has no real
  notion of async replies)
<braunr> i don't understand what you mean
<braunr> simpleroutine is the explicit solution for async replies
<braunr> as stated in
  http://www.cs.cmu.edu/afs/cs/project/mach/public/doc/unpublished/mig.ps
<braunr> it's not a new rpc with roles reversed
<braunr> it's not a reply either
<antrik> it might be an explicit solution for that, but it still seems
  kludgy :-)
<braunr> i don't see why :/
<braunr> would you have expected something like an option to create both
  sync and async versions ?
<antrik> because it requires an extra .defs file
<antrik> yes
<braunr> ok
<braunr> well this seems cumbersome to me :)
<braunr> i prefer the simpleroutine approach
<braunr> but i agree this seems odd since mach has a high level ipc api
<antrik> anyways, my point is that the ds_ in device_reply.defs still
  refers to the Mach side of things
<braunr> npnth: which package fails to build ?
<antrik> though a userspace process that actually handles the replies in an
  async fashion will of course need some kind of device server too, just
  like the DDE stuff...
<antrik> though naming it ds_ is confusing IMHO, because of the name clash
  with the device server in Mach
<braunr> hm again, i fail to see why
<braunr> ds_ just means device_server
<braunr> and as most things in mach, it can be in kernel or not
<braunr> i mean, this is an interface prefix, i don't refer to an actual
  single instance of a "device server" out there
<antrik> oh, right... DDE implements the Mach device protocol, so it *does*
  do the ds_ part... but that makes the interrupt notification stuff even
  more confusing
<braunr> hm
<braunr> because it provides a ds_device_intr_notify() which will never be
  used, just to completely implement the interface ?
<antrik> yeah, that's what I suspect...
<braunr> sounds likely
<antrik> the device interface actually has two parts: one for "generic"
  RPCs on the master device port, and one for device-specific RPCs. DDE
  implements the latter, and uses the former...
<antrik> they live in separate places though I think: the individual device
  RPCs are implemented in libmachdev, while the intr_ stuff is used in
  libddekit probably
<braunr> it would be hairy to build otherwise
<antrik> so we *really* need to know what component npnth gets the error
  with
<antrik> braunr: nah, not really. that's why we always have a separate
  prefix for the server routines in Hurd RPCs
<braunr> right, i really need to read about mig again
<antrik> it's pretty normal for a translator to both implement and use an
  interface

MACH_SEND_INTERRUPT/MACH_RCV_INTERRUPT

IRC, freenode, #hurd, 2013-03-22

<braunr> i'm also testing glibc packages on darnassus with a patch that
  removes the MACH_{SEND,RCV}_INTERRUPT options from mach_msg calls to
  avoid taking the slow path because of them
<braunr> if i got it right, almost every mach_msg call doesn't use any of
  these options, except for select
<braunr> it could slightly improve that, i'm not sure
<youpi> but don't we need that to get EINTR ?
<braunr> the options are purely userspace
<braunr> i'll upload the patch
<braunr>
  http://www.sceen.net/~rbraun/0001-Mask-options-implemented-by-the-userspace-side-of-ma.patch
<youpi> ah, ok, you mean userspace already implements what we need

IRC, freenode, #hurd, 2013-04-23

<braunr> i couldn't measure any difference with the glibc patch that
  removes the mach_msg interrupt related flags
<braunr> which isn't very surprising as it only concerns select as far as i
  can tell