path: root/hurd/translator/mtab/discussion.mdwn

[[!meta copyright="Copyright © 2008, 2009, 2013 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]]."]]"""]]

[[!tag open_issue_documentation]] The following text needs to be updated.

# Original Project Description

As a consequence, there is currently no easy way to obtain a listing of all
mounted file systems. This also means that commands like `df` can only work on
explicitly specified mountpoints, instead of displaying the usual listing.

One possible solution to this would be for the translator startup mechanism to
update the `mtab` on any `mount`/`unmount`, like in traditional systems.
However, there are some problems with this approach.  Most notably: what to do
with passive translators, i.e., translators that are not presently running, but
set up to be started automatically whenever the node is accessed?  Probably
these should be counted among the mounted filesystems; but how to handle the
`mtab` updates for a translator that is not started yet?  Generally, being
centralized and event-based, this is a pretty inelegant, non-hurdish solution.

A more promising approach is to have `mtab` exported by a special translator,
which gathers the necessary information on demand.  This could work by
traversing the tree of translators, asking each one for mount points attached
to it.  (Theoretically, it could also be done by just traversing *all* nodes,
checking each one for attached translators.  That would be very inefficient,
though.  Thus a special interface is probably required, that allows asking a
translator to list mount points only.)

There are also some other issues to keep in mind.  Traversing arbitrary
translators set by other users can be quite dangerous -- and it's probably not
very interesting anyways what private filesystems some other user has mounted.
But what about the global `/etc/mtab`?  Should it list only root-owned
filesystems?  Or should it create different listings depending on what user
contacts it?...

That leads to a more generic question: which translators should be actually
listed?  There are different kinds of translators: ranging from traditional
filesystems ([[disks|hurd/libdiskfs]] and other actual
[[stores|hurd/translator/storeio]]), but also purely virtual filesystems like
[[hurd/translator/ftpfs]] or [[hurd/translator/unionfs]], and even things that
have very little to do with a traditional filesystem, like a
[[gzip_translator|hurd/translator/storeio]],
[[mbox_translator|hurd/translator/mboxfs]],
[[xml_translator|hurd/translator/xmlfs]], or various device file translators...
Listing all of these in `/etc/mtab` would be pretty pointless, so some kind of
classification mechanism is necessary.  By default it probably should list only
translators that claim to be real filesystems, though alternative views with
other filtering rules might be desirable.

After taking decisions on the outstanding design questions, the student will
implement both the actual mtab translator, and the
necessary interface(s) for gathering the data.  It requires getting a good
understanding of the translator mechanism and Hurd interfaces in general.


# IRC, freenode, #hurd, 2013-04-17

    <kuldeepdhaka> thinking how to get the listing. traversing would be
      ineffecient,   trying to come up with something better
    <braunr> what listing ?
    <braunr> and traversing what ?
    <kuldeepdhaka> mtab
    <braunr> well i assumed so
    <braunr> be more precise please
    <kuldeepdhaka> when the translator is done initalized      <translation
      info> are written to /etc/mtab      <translation info> will be provided
      by the translator, and when some one want to read the info just read it
      this way if their is some credentials like ftp sites pass username can be
      masked by the translator
    <kuldeepdhaka> if some trans dont want to list them, no need to write to
      file    | while unmounting (sorry i couldnt find the right word) , it
      will pass the mount node address | <translation info> will have special
      structure to remove/add mounts example "a /mount-to /mount-from" = add
      , "r /mount-to" = remove    here "/mount-to" will be unique for every
      mount
    <kuldeepdhaka> this have a draw back , we would have to trust trans for the
      listed data  | also "/mount-to" + "/mount-from" could be used a
      combination for making sure that other trans unable remove others trans
      mount data
    <kuldeepdhaka> sorry but "also "/mount-to" + "/mount-from" could be used a
      combination for making sure that other trans unable remove others trans
      mount data"  this is a bad idea if we had to print the whole thing
    <kuldeepdhaka> braunr, whats ur opinion?
    <pinotree> you don't need a mtab to "unmount" things on hurd
    <braunr> kuldeepdhaka: hum, have you read the project idea ?

    <braunr> A more promising approach is to have mtab exported by a special
      translator, which gathers the necessary information on demand. This could
      work by traversing the tree of translators, asking each one for mount
      points attached to it.
    <kuldeepdhaka> pinotree, not to unmount, i mean is to remove the
      <translation data>
    <braunr> for a first implementation, i'd suggest a recursive traversal of
      root-owned translators
    <kuldeepdhaka> braunr, hum, but it did stated it as inefficient
    <braunr> where ?
    <kuldeepdhaka> para 5 , line 3
    <kuldeepdhaka> and line 6
    <braunr> no
    <braunr> traversing "all" nodes would be inefficient
    <braunr> translators which host the nodes of other translators could
      maintain a simple list of active translators
    <braunr> ext2fs, e.g. (if that's not already the case) could keep the list
      of the translators it started
    <braunr> we can already see that list with pstree for example
    <braunr> but this new list would only retain those relevant for mtab
    <braunr> i.e. root-owned ones
    <pinotree> i would not limit to those though
    <braunr> and then filter on their type (e.g. file system ones)
    <braunr> pinotree: why ?
    <pinotree> this way you could have proper per-user /proc/$pid/mounts info
    <braunr> we could also very easily have a denial of service
    <kuldeepdhaka> but how will the mount point and source point will be
      listed?
    <braunr> they're returned by the translator
    <kuldeepdhaka> k
    <braunr> you ask /, it returns its store and its options, and asks its
      children recursively 
    <braunr> a /home translator would return its store and its options
    <braunr> etc..
    <braunr> each translator would build the complete path before returning it
    <braunr> sort of, it's very basic
    <braunr> but that would be a very hurdish way to do it
    <kuldeepdhaka> shall /etc/mtab should be made seek-able and what should be
      the filesize?  content are generated on demand so, it could arise problem
      (fsize:0 , seek-able:no), ur opinions?
    <braunr> kuldeepdhaka: it should have all the properties of a regular file
    <braunr> the filesize would be determined after it's generated
    <braunr> being empty doesn't imply it's not seekable
    <kuldeepdhaka> content is generated on demand so, could cause problem while
      seeking and filesize, shall i still program as regular file?
    <kuldeepdhaka> in two different read, it could generate different content,
      though same seek pos is used...
    <braunr> what ?
    <braunr> the content is generated on open
    <kuldeepdhaka> ooh, ok


# IRC, freenode, #hurd, 2013-06-04

    <safinaskar> how to see list of all connected translators?
    <braunr> you can't directly
    <braunr> you can use ps to list processes and guess which are translators
    <braunr> (e.g. everything starting with /hurd/)
    <braunr> a recursive call to obtain such a list would be useful
    <braunr> similar to what's needed to implement /proc/mounts


# IRC, freenode, #hurd, 2013-06-25

In context of [[open_issues/mig_portable_rpc_declarations]].

    <teythoon> should I go for an iterator like interface instead?
    <teythoon> btw, what's the expected roundtrip time?
    <braunr> don't think that way
    <braunr> consider the round trip delay as varying
    <teythoon> y, is it that bad?
    <braunr> no
    <braunr> but the less there is the better
    <braunr> we think the same with system calls even if they're faster
    <braunr> the delay itself isn't the real issue
    <braunr> look at how proc provides information
    <braunr> (in procfs for example)


## IRC, freenode, #hurd, 2013-06-26

    <teythoon> so tell me about the more hurdish way of dealing with that issue
    <teythoon> creating a specialized translator for this?

    <braunr> you need to avoid thinking with centralization in mind
    <braunr> the hurd is a distributed system in practice
    <braunr> i think proc is the only centralized component in there
    <teythoon> braunr: would having an mtab translator and having fs
      translators register to thae be acceptable?
    <teythoon> that*
    <braunr> teythoon: why do you want to centralize it ?
    <braunr> translators already register themselves when they get attached to
      a node
    <braunr> we don't want an additional registration
    <braunr> have you read the link we gave you ?
    <teythoon> I did and I got the message, but isn't the concept of
      /proc/mounts or a mtab file already a centralized one?
    <braunr> that doesn't mean the implementation has to be
    <braunr> and no, i don't think it's centralized actually
    <braunr> it's just a file
    <braunr> you can build a file from many sources
    <teythoon> or if we do it your way recursing on fs translators *but*
      restricting this to root owned translators also suffering from
      centralization wrt to the root user? I mean the concept of all mounted
      filesystems does not apply cleanly to the hurd
    <braunr> i don't understand
    <braunr> restricting to the root user doesn't mean it's centralized
    <braunr> trust has nothing to do with being centralized
    <teythoon> I guess I'm not used to thinking this way
    <braunr> teythoon: i guess that's the main reason why so few developers
      work on the hurd
    <teythoon> also the way fs notification is done is also centralized, that
      could also be done recursively
    <braunr> what doyou call fs notification ?
    <teythoon> and the information I need could just be stuffed into the same
      mechanism
    <teythoon> fs translators being notified of system shutdown
    <braunr> right
    <braunr> that gets a bit complicated because the kernel is also a
      centralized component
    <braunr> it knows every memory object and their pagers
    <braunr> it manages all virtual memory
    <braunr> there are two different issues here
    <braunr> syncing memory and shutting down file systems
    <braunr> the latter could be done recursively, yes
    <braunr> i wonder if the former could be delegated to external pagers as
      well
    <braunr> teythoon: but that's not the focus of your work aiui, it would
      take much time
    <teythoon> sure, but missing an mtab file or better yet /proc/mounts could
      be an issue for me, at least a cosmetic one, if not a functional one
    <braunr> i understand
    <teythoon> and hacking up a quick solution for that seemed like a good
      exercise
    <braunr> i suggest you discuss it with your mentors
    <braunr> they might agree to a temporary centralized solution
    <braunr> although i don't think it's much simpler than the recursive one
    <teythoon> braunr: would that be implemented in libdiskfs and friends?
    <braunr> teythoon: i'm not sure, it might be a generic fs operation
    <braunr> libnetfs etc.. are also mount points
    <teythoon> so where would it go if it was generic?
    <braunr> libfshelp perhaps
    <teythoon> translator startup is handled in start-translator-long.c, so in
      case a startup is successful, I'd add it to a list?
    <braunr> i'd say so, yes
    <teythoon> would that cover all cases, passive and active translators?
    <braunr> that's another question
    <braunr> do we consider passive translators as mounted ?
    <teythoon> ah, that was not what i meant
    <braunr> i know
    <braunr> but it's related
    <teythoon> start b/c of accessing a passive one vs. starting an active one
      using settrans
    <braunr> start_translator_xxx only spawn active translators
    <braunr> it's the same
    <teythoon> ok
    <braunr> the definition of a passive translator is that it starts the
      active translator on access
    <teythoon> yeah I can see how that wouldn't be hard to implement
    <braunr> i think we want to include passive translators in the mount table
    <braunr> so registration must happen before starting the active one
    <teythoon> so it's a) keeping a list of active translators and b) add an
      interface to query fs translators for this list and c) an interface to
      query mtab style information?
    <braunr> keeping a list of all translators attached
    <braunr> and yes
    <braunr> well
    <braunr> a is easy
    <braunr> b is the real work
    <braunr> c would be procfs using b
    <teythoon> oh? I thought recursing on the translators and querying info
      would be separate operations?
    <braunr> why so ?
    <braunr> the point is querying recursively :)
    <braunr> and when i say recursively, it's only a logical view
    <teythoon> ok, yes, it can be implemented this way, so we construct the
      list while recursing on the translators
    <braunr> i think it would be better to implement it the way looking up a
      node is done
    <teythoon> in a loop, using a stack?
    <braunr> iteratively
    <braunr> a translator would provide information about itself (if
      supported), and referrences to translators locally registered to it
    <teythoon> could you point me to the node lookup?
    <teythoon> ah, yes
    <braunr> eg., you ask /, it tells you it's on /dev/hd0, read-write, with
      options, and send rights to /home, /proc, etc..
    <braunr> well rights, references
    <braunr> it could be the path itself
    <teythoon> rights as in a port to the translators?
    <braunr> i think the path would be better but i'm not sure
    <braunr> it would also allow you to check the permissions of the node
      before querying
    <teythoon> path would be nicer in the presence of stacked translators
    <braunr> and obviously you'd have the path right away, no need to provide
      it in the reply
    <teythoon> true

    <teythoon> braunr: if we want to list passive translators (and I agree, we
      should), it isn't sufficient to touch libfshelp, as setting a passive
      translator is not handled there, only the startup
    <braunr> teythoon: doesn't mean you can't add something there that other
      libraries will use
    <braunr> so yes, not sufficient


## IRC, freenode, #hurd, 2013-06-29

    <teythoon> braunr: diskfs_S_fsys_set_options uses diskfs_node_iterate to
      recurse on active translators if do_children is given
    <teythoon> braunr: I wonder how fast that is in practice
    <teythoon> braunr: if it's fast enough, there might not even be a need for
      a new function in fsys.defs
    <teythoon> and no need to keep a list of translators for that reason
    <teythoon> braunr: if it's not fast enough, then diskfs_S_fsys_set_options
      could use the list to speed this up
    <braunr> teythoon: on all nodes ?
    <teythoon> braunr: i believe so, yes, see libdiskfs/fsys-options.c
    <braunr> teythoon: well, if it really is all node, you clearly don't want
      that


## IRC, freenode, #hurd, 2013-07-01

    <teythoon> I've ment to ask, the shiny new fsys_get_translators interface,
      should it return the options for the queried translator or not?
    <braunr> i don't think it should
    <teythoon> ok
    <braunr> let's walk through why it shouldn't
    <teythoon> may I assume that the last argument returned by fsys_get_options
      is the "source"?
    <braunr> how would you know these options ?
    <braunr> the source ?
    <teythoon> I wouldn't actually
    <braunr> yes, you wouldn't
    <braunr> you'd have to ask the translators for that
    <braunr> so the only thing you can do is point to them
    <teythoon> well, the device to include in the mtab file
    <braunr> and the client asks
    <braunr> i don't know fsys_get_options tbh
    <teythoon> well, both tmpfs and ext2fs print an appropriate value for
      "device" as last argument
    <braunr> looks like a bad interface to me
    <braunr> options should be options
    <braunr> there should be a specific call for the device
    <braunr> but if everyone agrees with the options order, you can do it that
      way for now i guess
    <teythoon> one that could be used to recreate the "mount" using either
      mount or settrans
    <braunr> just comment it where appropriate
    <teythoon> I thought that'd be the point?
    <braunr> ?
    <teythoon> % fsysopts tmp
    <teythoon> /hurd/tmpfs --writable --no-inherit-dir-group --no-sync 48K
    <braunr> where is the device ?
    <teythoon> % settrans -ca tmp $(fsysopts tmp)
    <braunr> 15:56 < teythoon> well, both tmpfs and ext2fs print an appropriate
      value for "device" as last argument
    <teythoon> 48K
    <braunr> i don't see it
    <braunr> really ?
    <teythoon> yes
    <braunr> what about ext2fs ?
    <braunr> hm ok i see
    <teythoon> % fsysopts /
    <teythoon> ext2fs --writable --no-inherit-dir-group --sync=10
      --store-type=typed device:hd0s1
    <braunr> i don't think you should consider that as devices
    <braunr> but really translator specific options
    <pinotree> agree
    <teythoon> I don't ;)
    <teythoon> b/c the translator calling convention is hardcoded in the mount
      utility
    <braunr> ?
    <teythoon> I think it's reasonable to assume that this mapping can be
      reversed
    <pinotree> theorically you can write a translator that takes no arguments,
      but just options
    <braunr> the 48K string for tmpfs is completely meaningless
    <braunr> in fstab, it should be none
    <pinotree> "tmpfs"
    <braunr> the linux equivalent is the size option
    <braunr> no, none
    <braunr> it's totally ignored
    <braunr> and it's recommended to set none rather than the type to avoid
      confusion
    <teythoon> u sure?
    <teythoon> % settrans -cga tmp /hurd/tmpfs --mode=666 6M          
    <teythoon> % settrans -cga tmp /hurd/tmpfs --mode=666 6M          
    <teythoon> % fsysopts tmp
    <teythoon> /hurd/tmpfs --writable --no-inherit-dir-group --no-sync 6M
    <braunr> i've not explained myself clearly
    <braunr> it's not ignored by the translator
    <braunr> but in fstab, it should be in the options field
    <braunr> it's not the source
    <braunr> clearly not
    <teythoon> ah
    <braunr> now i'm talking about fstab, but iirc the format is similar in
      mtab/mounts
    <pinotree> close, but not the same
    <braunr> yes, close
    <teythoon> ok, so I'll put a method into libfshelp so that translators can
      explicitly set a device and patch all existing translators to do so?
    <braunr> teythoon: what i meant is that, for virtual vile systems (actually
      file systems with no underlying devices), the device field is normally
      ignored
    <braunr> teythoon: why do you need that for exactly
    <teythoon> right
    <pinotree> do they even have a "device" field?
    <braunr> (i can see why but i'd like more visibility)
    <braunr> pinotree: not yet
    <braunr> pinotree: that's what he wants to add
    <braunr> but i'd like to see if there is another way to get the information
    <braunr> 16:05 < braunr> teythoon: why do you need that for exactly
    <teythoon> well if I'm constructing a mtab entry I need a value for the
      device field
    <braunr> do we actually need it to be valid ?
    <teythoon> not necessarily I guess
    <braunr> discuss it with your mentors then
    <youpi> it has to be valid for e2fsck checks etc.
    <braunr> doesn't e2fsck check fstab actually ?
    <youpi> i.e. actually for the cases where it's trivial
    <youpi> fstab doesn't tell it whether it's mounted
    <youpi> I mean fsck checking whether it's mounted
    <youpi> not fsck -a
    <braunr> oh
    <braunr> couldn't we ask the device instead ?
    <braunr> looks twisted too
    <youpi> that'd mean patching a lot of applications which do similar checks
    <braunr> yes
    <braunr> teythoon: propose an interface for that with your mentors then
    <teythoon> yeah, but couldn't you lay it out a little, I mean would it be
      one procedure or like three?
    <braunr> 16:04 < teythoon> ok, so I'll put a method into libfshelp so that
      translators can explicitly set a device and patch all existing
      translators to do so?
    <teythoon> ok
    <braunr> why three ?
    <teythoon> no, I mean when adding stuff to fsys.defs
    <braunr> i understood that
    <braunr> but why three ? :)
    <teythoon> it'd be more generic
    <braunr> really ?
    <braunr> please show a quick example of what you have in mind
    <teythoon> i honestly don't know, thus I'm asking ;)
    <braunr> well first, this device thing bothers me
    <braunr> when you look at how we set up our ext2fs translators, you can see
      they use device:xxx
    <braunr> and not /dev/xxx
    <braunr> but ok, let's assume it's harmless
    <teythoon> ok, but isn't the first way actually better?
    <braunr> i think it ends up being the same
    <braunr> ideally, that's what we want to use as device path
    <teythoon> but you can recreate a storeio translator using the device:xxx
      info, the node is useless for that
    <braunr> so that we don't need to explicitely set it
    <braunr> ?
    <braunr> what do you mean ?
    <teythoon> well, fsysopts / tells currently tells me device:hd0s1
    <braunr> for /, there isn't much choice
    <braunr> /dev isn't there yet
    <teythoon> ah, got it
    <teythoon> that's why it differs...
    <braunr> differs ?
    <braunr> from what ?
    <braunr> other ext2fs translators are set the same way by the debian
      installer for example
    <teythoon> % fsysopts /media/scratch
    <teythoon> /hurd/ext2fs --writable --no-inherit-dir-group /dev/hd1s1
    <teythoon> here it uses the path to the node
    <braunr> that's weird
    <braunr> was that done by the debian installer ?
    <teythoon> ah no, that was me
    <braunr> :p
    <braunr> $ fsysopts /home
    <braunr> /hurd/ext2fs --writable --no-inherit-dir-group --store-type=device
      hd0s6
    <braunr> so as you can see, it's not that simple to infer the device path
    <teythoon> oho, yet another way ;)
    <teythoon> right then
    <pinotree> isn't device:hd0s1 as shortcut for specifying the store type, as
      done with --store-type=device hd0s1?
    <braunr> but perhaps we don't need to
    <braunr> yes it is
    <pinotree> iirc it's something libstore does, per-store prefixes
    <braunr> ah that sucks
    <braunr> teythoon: you may need to normalize those strings
    <braunr> so that they match what's in fstab
    <braunr> i.e. unix /dev paths
    <braunr> otherwise e2fsck still won't be able to find the translators
      mounting the device
    <braunr> well, if it's mounted actually
    <braunr> it just needs to find the matching line in mtab aiui
    <braunr> so perhaps a libfshelp function for that, yes
    <teythoon> braunr: so you suggest adding a normalizing function to
      libfshelp that creates a /dev/path?
    <braunr> yes
    <braunr> used by the call you intend to add, which returns that device
      string as found in fstab
    <teythoon> found in fstab? so this would only work for translators managed
      by fstab?
    <braunr> no
    <teythoon> ah
    <teythoon> a string like the ones found in fstab?
    <braunr> yes
    <braunr> so that fsck and friends are able to know whether a device is
      mounted or not
    <braunr> i don't see any other purpose for that string in mtab
    <braunr> you'd take regular paths as they are, convert device:xxx to
      /dev/xxx, and return "none" for the rest i suppose
    <teythoon> ok
    <braunr> i'm not even sure it's right
    <braunr> youpi: are you sure it's required ?
    <teythoon> well it's a start and I think it's not too much work
    <braunr> aiui, e2fsck may simply find the mount point in fstab, and ask the
      translator if it's mounted
    <teythoon> we can refine this later on maybe?
    <braunr> or rather, init scripts, using mountpoint, before starting e2fsck
    <braunr> teythoon: sure
    <teythoon> there's this mountpoint issue... I need to run fsysopts /
      --update early in the boot process
    <teythoon> otherwise the device ids returned by stat(2)ing / are wrong and
      mountpoint misbehaves
    <teythoon> i guess b/c it's the rootfs
    <braunr> device ids ?
    <teythoon> % stat / | grep Device
    <teythoon> Device: 3h/3d     Inode: 2           Links: 22
    <braunr> do you mean the major/minor identifiers ?
    <teythoon> I do. if I don't do the --update i get seemingly random values
    <braunr> i guess that's expected
    <braunr> we don't have major/minor values
    <braunr> well, they're emulated
    <teythoon> well, if that's fixable, that'd be really nice ;)
    <braunr> we'll never have major/minor values
    <teythoon> yeah, I understand that
    <braunr> but they could be fixed by MAKEDEV when creating device nodes
    <teythoon> but not having to call fsys_set_options on the rootfs to get the
      emulation up to speed
    <braunr> try doing it from grub
    <braunr> not sure it's possible
    <braunr> but worth checking
    <teythoon> by means of an ext2fs flag?
    <braunr> yes
    <braunr> if there is one
    <braunr> i don't know the --update flag, is it new from your work ?
    <teythoon> braunr: no, it's been there before. -oremount gets mapped to
      that
    <braunr> it's documented by fsysopts, but not by the ext2fs translators
    <teythoon> libdiskfs source says something about flushing buffers iirc
    <braunr> -s
    <braunr> what does it do ?
    <braunr> teythoon: ok
    <teythoon> braunr: so the plan is to automatically generate a device path
      from the translators argz vector but to provide the functionality so
      translators can set a more appropriate value? did I get the last part of
      the discussion right?
    <braunr> not set, return
    <teythoon> yeah return from the procedure but settable using libfshelp?
    <braunr> why settable ?
    <braunr> you'd have a fsys call to obtain the dev string, and the server
      side would call libfshelp on the fly to obtain a normalized value and
      return it
    <teythoon> ah, make a function overrideable that returns an appropriate
      response?
    <braunr> overrideable ?
    <teythoon> like netfs_append_args
    <braunr> you wouldn't change the command line, no
    <teythoon> isn't that done using weak references or something?
    <teythoon> no I know
    <braunr> sorry i'm lost then
    <teythoon> never mind, I'll propose a patch early to get your feedback
    <youpi> braunr: am I sure that _what_ is required?
    <youpi> the device?
    <youpi> e2fsck surely needs it, yes
    <braunr> a valid device path, yes
    <youpi> it can't rely only on fstab
    <braunr> yes
    <youpi> since users may mount things by hand
    <braunr> i've used strace on it and it does perform lookups there
    <braunr> (although i also saw uuid magic that i guess wouldn't work yet on
      the hurd)


## IRC, freenode, #hurd, 2013-07-03

    <teythoon> I added a procedure to fsys.defs, added a server stub to my
      tmpfs translator and wrote a simple client, but something hasn't picked
      up the new message yet
    <teythoon> % ./mtab tmp
    <teythoon> ./mtab: get_translator_info: (ipc/mig) bad request message ID
    <teythoon> I guess it's libhurduser.so from glibc, not sure though...
    <braunr> glibc would only have the client calls
    <braunr> what is "% ./mtab tmp" ?
    <teythoon> mtab is my mtab tool/soon to be a translator testing thing, tmp
      is an active tmpfs with the appropriate server stub
    <braunr> so mtab has the client call, right ?
    <teythoon> yes
    <braunr> then tmpfs doesn't
    <teythoon> so what to do about it?
    <teythoon> i set LD_LIBRARY_PATH to my hurd builds lib dir, is that
      preserved by settrans -a?
    <pinotree> not really
    <braunr> not at all
    <braunr> there is a wiki entry about that iirc
    <pinotree> http://darnassus.sceen.net/~hurd-web/hurd/debugging/translator/
    <teythoon> yeah, I read it too once
    <teythoon> ah
    <braunr> on the other hand, using export to set the environment should do
      the work
    <teythoon> yes, that did the trick, thanks :)
    * teythoon got his EOPNOPSUPP... *nomnomnom
    <braunr> ?
    <braunr> same error ?
    <teythoon> well I stubbed it out
    <braunr> oh
    <teythoon> no, that's what I've been expecting ;)
    <pinotree> great
    <braunr> :)
    <braunr> yes that's better than "mig can't find it"
    <teythoon> braunr: in that list of active and passive translators that will
      have to be maintained, do you expect it should carry more information
      other than the relative path to that translator?
    <braunr> like what ?
    <teythoon> dunno, maybe like a port to any active translator there
    <teythoon> should we care if any active translator dies and remove the
      entry if there's no passive translator that could restart it again?
    <braunr> don't add anything until you see it's necessary or really useful
    <braunr> yes
    <braunr> think of something like sshfs
    <braunr> when you kill it, it's not reported by mount any more
    <teythoon> well, for a dynamically allocated list of strings I could use
      the argz stuff, but if we'd ever add anything else, we'd need a linked
      list or something, maybe a hash table
    <teythoon> yes, I thought that'd be useful
    <braunr> use libihash for no
    <braunr> now
    <teythoon> braunr: but what would I use as keys? the relative path should
      be unique (unless translators are stacked... hmmm), but that's the value
      I'd like to store and ihash keys are pointers
    <teythoon> stacked translators are an kinda interesting case for mtab
      anyways...
    <braunr> why not store the string address ?
    <braunr> i suppose that, for stacked translators, the code querying
      information would only return the topmost translator
    <braunr> since this is the one which matters for regular clients (if i'm
      right)
    <teythoon> wouldn't that map strings that are equal but stored at different
      locations to different values?
    <teythoon> that'd defeat the point
    <teythoon> I suppose so, yes
    <braunr> then add a layer that looks for existing strings before adding
    <braunr> the list should normally be small so a linear lookup is fine
    <teythoon> yeah sure, but then there's little advantage of using ihash in
      the first place, isn't it?
    <braunr> over what ?
    <teythoon> over not using it at all
    <braunr> how would you store the list then ?
    <teythoon> it's either ll or ll+ihash
    <braunr> uh no
    <braunr> let me check
    <braunr> there is ihash_iterate
    <braunr> so you don't need a linked list
    <teythoon> so how do I store my list of strings to deduplicate the keys?
    <braunr> you store pointers
    <braunr> and on addition, you iterate over all entries, making sure none
      matches the new one
    <braunr> and if it does, you replace it i guess
    <braunr> depending on how you design the rest
    <teythoon> in an dynamically allocated region of memory?
    <braunr> i don't understand
    <braunr> your strings should be dynmaically allocate, yes
    <teythoon> no the array of char *
    <braunr> your data structure being managed by libihash, you don't care
      about allocation
    <braunr> what array ?
    <teythoon> ah, got it...
    <teythoon> right.
    <braunr> there is only one structure here, an ihash of char *
    <teythoon> yes, I got the picture ;)
    <braunr> goo
    <braunr> d
    <braunr> actually, the lookup wouldn't be linear since usually, hash tables
      have stale entries
    <teythoon> heh... what forest?!?
    <braunr> but that's ok
    <braunr> teythoon: ?
    <teythoon> the one I couldn't make out b/c of all the trees...
    <braunr> ?
    <teythoon> ah, it's not important. there is this saying over here, not sure
      if there's an english equivalent
    <braunr> ok got it
    <braunr> we have the same in french
    <teythoon> I ran into a problem with my prototype
    <teythoon> if an translator is set in e. g. diskfs_S_file_set_translator,
      how do I get the path to that node?
    <teythoon> I believe there cannot be a way to do that, b/c the mapping is
      not bijective
    <braunr> it doesn't have to be
    <teythoon> ok, so how do I get *a* path for this node?
    <braunr> that's another question
    <braunr> do you see how the node is obtained ?
    <braunr> np = cred->po->np;
    <teythoon> yes
    <braunr> the translation occurred earlier
    <braunr> you need to find where
    <braunr> then perhaps, you'll need to carry the path along
    <braunr> or if you're lucky, it will still be there somewhere
    <teythoon> the translation from path to node?
    <braunr> yes
    <teythoon> doesn't that happen in the client? and the client hands a file_t
      to the file_set_translator routine?
    <braunr> the relative lookup can't happen in the client
    <braunr> the server can (and often does) retain information between two
      RPCs
    <teythoon> uh, I can access information from a previous rpc? is that
      considered safe?
    <braunr> think of open() then read()
    <braunr> a simple int doesn't carry enough information
    <braunr> that's why it's a descriptor
    <teythoon> ah, the server retains some state, sure
    <braunr> what it refers to is the state retained between several calls
    <braunr> the object being invoked by clients
    <braunr> teythoon: what is the "passive" parameter passed to
      diskfs_S_file_set_translator ?
    <teythoon> braunr: argz vector of the passive translator
    <braunr> so it is a name
    <braunr> but we also want active translators
    <braunr> and what is active ?
    <teythoon> not the name of the node though
    <teythoon> active is the port (?) to the active translator
    <teythoon> I guess
    <braunr> fsys_t, looks that way yes
    <braunr> i suppose you could add the path to the peropen structure
    <teythoon> ok
    <braunr> see diskfs_make_peropen
    <teythoon> braunr: but translation happens in dir_lookup
    <teythoon> in all places I've seen diskfs_make_peropen used, the path is
      not available
    <teythoon> why did you point me to diskfs_make_peropen?
    <teythoon> s/dir_lookup/diskfs_lookup/
    <teythoon> diskfs_lookup operates on struct node, so the path would have to
      be stored there, right?
    <braunr> teythoon: dir_lookup should call diskfs_make_peropen
    <braunr> at least diskfs_S_dir_lookup does
    <braunr> and the path is present there
    <teythoon> braunr: right

    <teythoon> hrm... I added a path field to struct peropen and initialize it
      properly in diskfs_make_peropen, but some bogus values keep creeping in
      :/
    <braunr> first of all, make it a dynamically allocated string
    <teythoon> it is
    <braunr> not a fixed sized embedded array
    <braunr> good
    <teythoon> yes
    <braunr> if you really need help debugging what's happening, feel free to
      post your current changes somewhere
    <teythoon> there is a struct literal in fsys-getroot.c, but i fixed that as
      well
    <teythoon> % ./mtab tmp 
    <teythoon> none tmp ../tmpfs/tmpfs writable,no-inherit-dir-group,no-sync 0
      0
    <teythoon> none tmp/bar ../tmpfs/tmpfs
      writable,no-inherit-dir-group,no-sync 0 0
    <teythoon> none tmp/foo ../tmpfs/tmpfs
      writable,no-inherit-dir-group,no-sync 0 0
    <teythoon> none tmp/foo/bar ../tmpfs/tmpfs
      writable,no-inherit-dir-group,no-sync 0 0
    <teythoon> :)


## IRC, freenode, #hurd, 2013-07-10

    <teythoon> btw, I read getcwd.c and got the idea
    <teythoon> however this situation is different afaict
    <teythoon> getcwd has a port to the current working directory, right?
    <teythoon> so they can do open_dir with .. as relative path
    <teythoon> but all I've got is a port referencing the node the translator
      is being attached to
    <teythoon> s/open_dir/dir_lookup/
    <teythoon> and that is not necessarily a directory, so dir_lookup fails
      with not a directory
    <teythoon> as far as I can see it is not possible to get the directory a
      node is in from a port referencing that node
    <teythoon> dir_lookup has to be handled by all nodes, not just directories
    <teythoon> but file nodes only support "looking up" the empty string
    <teythoon> not empty, but null:
    <teythoon> This call is required to be supported by all files (even
      non-directories) if the filename is null, and should function in that
      case as a re-open of the file.  */
    <braunr> why do you want the directory ?
    <braunr> 10:40 < teythoon> as far as I can see it is not possible to get
      the directory a node is in from a port referencing that node
    <teythoon> to readdir(3) it and figure out the name of the node the
      translator is bound to
    <teythoon> similar to what getcwd does
    <braunr> that's out of the question
    <teythoon> wasn't that was youpi was suggesting?
    <braunr> you may have a lot of nodes in there, such a lookup shouldn't be
      done
    <braunr> i didn't see that detail
    <teythoon> "│   Concerning storing the path, it's a bit sad to have to do
      that, and
    <teythoon> │   it'll become wrong if one moves the mount points.  Another
      way would
    <teythoon> │   be to make the client figure it out by itself from a port to
      the mount
    <teythoon> │   point, much like glibc/sysdeps/mach/hurd/getcwd.c. It'll be
      slower, but
    <teythoon> │   should be safer.  The RPC would thus return an array of
      ports to the
    <teythoon> │   mount points instead of an array of strings.
    <braunr> yes i remember that
    <braunr> but i didn't understand well how getcwd work
    <braunr> s
    <braunr> another scalability issue
    <braunr> not a big one though, we rarely have translators in directories
      with thousands of nodes
    <braunr> so why not
    <braunr> teythoon: do it as youpi suggested
    <braunr> well if you can
    <braunr> eh
    <braunr> if not, i don't know
    <braunr> 10:47 < teythoon> │   it'll become wrong if one moves the mount
      points.  Another way would
    <teythoon> yes, I know... :/
    <teythoon> well, I'm not even sure it is possible to get the directory a
      node is in from the port referencing the node
    <teythoon> as in, I'm not sure if the information is even there
    <teythoon> b/c a filesystem is a tree, directories are nodes and files are
      leafs
    <teythoon> all non-leaf nodes reference their parent to allow traversing
      the tree starting from any directory
    <teythoon> but why would a leaf reference its parent(s - in case of
      hardlinks)?
    <braunr> uh, for the same reason ?
    <teythoon> sure, it would be nice to do that, but I dont think this is
      possible on unixy systems
    <braunr> ?
    <teythoon> you cannot say fchdir(2) to a fd that references a file
    <braunr> do you mean /path/to/file/../ ?
    <teythoon> yes
    <teythoon> only that /path/to/file is given as fd or port
    <braunr> when i pasted
    <braunr> 10:49 < braunr> 10:47 < teythoon> │   it'll become wrong if one
      moves the mount points.  Another way would
    <braunr> i was actually wondering if it was true
    <teythoon> ah
    <braunr> why can't the path be updated at the same time ?
    <braunr> it's a relative path anyway
    <braunr> completely managed by the parent translator
    <teythoon> ah
    <teythoon> right
    <teythoon> it's still kind of hacky, but I cannot see how to do this
      properly
    <braunr> hacky ?
    <teythoon> but yes, updating the path should work I guess
    <teythoon> or sad
    <braunr> what i find hacky is to set translators in two passes
    <braunr> otherwise we'd only keep the translator paths
    <braunr> not all paths
    <teythoon> true
    <braunr> but then, it only concerns open nodes
    <braunr> and again, there shouldn't be too many of them
    <braunr> so actually it's ok
    <teythoon> braunr: I understand the struct nodes are cached in libdiskfs,
      so wouldn't it be easier to attach the path to that struct instead of
      struct peropen so that all peropen objects reference the same node
      object?
    <teythoon> so that the path can be updated if anyone dir_renames it
    <teythoon> *all peropen objects derived from the same file name that is
    <braunr> teythoon: i'm not sure
    <braunr> nodes could be real nodes (i.e. inodes)
    <braunr> there can be several paths for the same inode
    <teythoon> braunr: I'm aware of that, but didn't we agree the other day
      that any path would do?
    <braunr> i don't remember we did
    <braunr> i don't know the details well, but i don't think setting a
      translator on a hard link should set the translator at the inode level
    <braunr> on the other hand, if a new inode is created to replace the
      previous one (or stack over it), then storing the path there should be
      fine
    <teythoon> braunr: I don't think I can update the paths if they're stored
      in the peropen struct
    <teythoon> how would I get a reference to all those peropen objects?
    <braunr> ?
    <braunr> first, what's the context when you talkb about updating paths ?
    <teythoon> well, youpi was concerned about renaming a mount point
    <teythoon> and you implied that this could be managed
    <braunr> can we actually do that btw ?
    <teythoon> what?
    <braunr> renaming a mount point
    <teythoon> yep, just tried
    <braunr> i mean, on a regular unix system like linux
    <braunr> $ mv test blah
    <braunr> mv: cannot move `test' to `blah': Device or resource busy
    <braunr> (using sshfs so YMMV)
    <pinotree> do you have anything (shells, open files, etc) inside it?
    <braunr> no
    <braunr> i'll try with an empty loop-mounted ext4
    <teythoon> I was testing on the Hurd, worked fine there even with a shell
      inside
    <braunr> same thing
    <braunr> i consider it a bug
    <braunr> we may want to check what posix says about it
    <teythoon> o_O
    <braunr> and decide not to support renaming
    <teythoon> why?
    <pinotree> start a discussion in ml, maybe roland can chime in
    <braunr> it complicates things
    <braunr> ah yes
    <teythoon> sure, but I can move or rename a directory, why should it be
      different with a mount point?
    <braunr> because it's two of them
    <braunr> they're stacked
    <braunr> if we do want to support that, we must be very careful about
      atomically updating all the stack
    <teythoon> ok
    <teythoon> braunr: I'm trying to detect dying translators to remove them
      from the list of translators
    <teythoon> what port can I use for that purpose?
    <teythoon> if I use the bootstrap port, can I then use the same method as
      init/init.c uses? just defining a do_mach_notify_dead_name function and
      the multiplexer will call this?
    <braunr> teythoon: possibly
    <teythoon> braunr: we'll see shortly...
    <teythoon> I get KERN_INVALID_CAPABILITY indicating that my bootstrap port
      is invalid
    <teythoon> when calling mach_port_request_notification to get the dead name
      notification I mean
    <braunr> is the translator already started when you do that ?
    <teythoon> yes, at least I think so, I'm hooking into
      diskfs_S_file_set_translator and that gets an active translators port
    <teythoon> also the mach docs suggests that the notification port is
      invalid, not the name port referencing the translator
    <braunr> i guess it shouldn't
    <braunr> oh
    <braunr> please show the code
    <braunr> but beware, if the translator is started, assume it could die
      immediately
    <teythoon> braunr: http://paste.debian.net/15371/ line 87
    <braunr> teythoon: notify can't be bootstrap
    <braunr> what do you have in mind when writing this ?
    <braunr> i'm not sure i follow
    <teythoon> I want to be notified if an active translator goes away to
      remove it from the list of translators
    <braunr> ok but then
    <braunr> create a send-once right
    <braunr> and wait on it
    <braunr> also, why do you want to be notified ?
    <braunr> isn't this already done ?
    <braunr> or can't do it lazily on access attempt ?
    <braunr> +you
    <teythoon> in the client?
    <braunr> in the parent server
    <braunr> what happens currently when a translator dies
    <braunr> is the parent notified ?
    <braunr> or does it give an invalid right ?
    <teythoon> ah, i think so
    <braunr> then you don't need to do it again
    <teythoon> right, I overlooked that


## IRC, freenode, #hurd, 2013-07-12

    <teythoon> recursively traversing all translators from / turns out to be
      more dangerous than I expected
    <teythoon> ... if done by a translator bound somewhere below /...
    <teythoon> my interpretation is that the mtab translator tries to talk to
      itself and deadlocks
    <teythoon> (and as a side effect the whole system kinda just stops...)


## IRC, freenode, #hurd, 2013-07-15

    <youpi> teythoon: did you discuss with braunr about returning port vs path
      in fsys_get_children?
    <teythoon> youpi: we did
    <teythoon> as I wrote I looked at the getcwd source you pointed me at
    <teythoon> and I started to code up something similar
    <teythoon> but as far as I can see there's no way to tell from a port
      referencing a file the directory this file is located in
    <youpi> ah, right, there was a [0] mail
    <youpi> teythoon: because it doesn't have a "..", right
    <teythoon> about Neals concerns, he's right about not covering passive
      translators very well
    <teythoon> but the solution he proposed was similar to what I tried to do
      first
    <youpi> I don't like half-covering passive translators at all, to be honest
      :)
    <youpi> either covering them completely, or not at all, would be fine
    <teythoon> and then braunr convinced me that the "recursive" approach is
      more elegant and hurdish, and I came to agree with him
    <teythoon> youpi: one could scan the filesystem at translator startup and
      populate the list
    <youpi> by "Neal's solution", you mean an mtab registry?
    <teythoon> yes
    <braunr> so, let's see what linux does when renaming parent directories
    <teythoon> mount points you mean?
    <youpi> teythoon: browsing the whole filesystem just to find passive
      translators is costly
    <youpi> teythoon, braunr: and that won't prevent the user from unexpectedly
      starting other translators at will
    <braunr> scary
    <teythoon> youpi: but that requires the privilege to open the device
    <youpi> the fact that a passive translator is set is nothing more than a
      user having the intent of starting a translator
    <braunr> linux retains the original path in the mount table
    <youpi> heh
    <teythoon> youpi: any unprivileged user can trigger a translator startup
    <youpi> sure, but root can do that too
    <youpi> and expect the system to behave nicely
    <teythoon> but if I'm root and want to fsck something, I won't start
      translators accessing the device just before that
    <teythoon> but if there's a passive translator targetting the device,
      someone else might do that
    <youpi> root does not always completely control what he's doing
    <youpi> linux for instance does prevent from mounting a filesystem being
      checked
    <teythoon> but still, including passive translators in the list would at
      least prevent anyone starting an translator by accident, isn't that worth
      doing then?
    <youpi> if there's a way to prevent root too, that's better than having a
      half-support for something which we don't necessarily really want
    <youpi> (i.e. an exclusive lock on the underlying device)
    <teythoon> right, that would also do the trick
    <teythoon> btw, some programs or scripts seem to hardcode /proc/mounts and
      procfs and I cannot bind a translator to /proc/mounts since it is
      read-only and the node does not exist
    <kilobug> IMHO automatically starting translators is a generic feature, and
      passive translator is just a specific instance of it; but we could very
      well have, like an "autofs" that automatically start translators in tar
      archives and iso images, allowing to cd into any tar/iso on the system;
      implementing such things is part of the Hurd flexibility, the "core
      system" shouldn't be too aware on how translators are started
    <youpi> so in the end, storing where the active translator was started
      first seems okayish according to what linux has been exposing for decades
    <youpi> kilobug: indeed
    <teythoon> it could serve a mounts with a passive translator by default, or
      a link to /run/mtab, or an simple file so we could bind a translator to
      that node
    <youpi> I'd tend to think that /proc/mounts should be a passive translator
      and /run/mtab / /etc/mtab a symlink to it
    <youpi> not being to choose the translator is a concern however
    <teythoon> ok, I'll look into that
    <youpi> it could be an empty file, and people be able to set a translator
      on it
    <teythoon> if it had a passive translator, people still could bind their
      own translator to it later on, right?
    <teythoon> afaics the issue currently is mostly, that there is no mounts
      node and it is not possible to create one
    <youpi> right
    <teythoon> cool
    <youpi> so with the actual path, you can even check for caller's permission
      to read the path
    <youpi> i.e. not provide any more information than the user would be able
      to get from browsing by hand
    <teythoon> sure, that concern of Neil's is easy to address
    <youpi> I'm not so much concerned by stale paths being shown in mtab
    <youpi> the worst that can happen is a user not being able to umount the
      path
    <youpi> but he can settrans -g it
    <youpi> (which he can't on linux ;) )
    <teythoon> yes, and the device information is still valid
    <youpi> yes
    <braunr> despite the parent dir being renamed, linux is still able to
      umount the new path
    <teythoon> and so is our current umount
    <braunr> good
    <teythoon> (if one uses the mount point as argument)
    <braunr> what's the current plan concerning /proc/mounts ?
    <teythoon> serving a node with a passive translator record
    <braunr> ?
    <teythoon> so that /hurd/mtab / is started on access
    <braunr> i mean, still planning on using the recursive approach instead of
      a registry ?
    <teythoon> ah
    <teythoon> I do not feel confident enough to decide this, but I agree with
      you, it feels elegant
    <teythoon> and it works :)
    <teythoon> modulo the translator deadlocking if it talks to itself, any
      thoughts on that?
    <youpi> it is a non-threaded translator I guess?
    <teythoon> currently yes
    <youpi> making it threaded should fix the issue
    <teythoon> I tried to make the mtab translator multithreaded but that
      didn't help
    <youpi> that's odd
    <teythoon> maybe I did it wrong
    <braunr> i don't find it surprising
    <braunr> well, not that surprising :p
    <braunr> on what lock does it block ?
    <teythoon> as far as i can see the only difference of hello and hellot-mt
      is that it uses a different dispatcher and has lot's of locking, right?
    <teythoon> braunr: I'm not sure, partly because that wrecked havoc on the
      whole system
    <teythoon> it just freezes
    <teythoon> but it wasn't permanent. once i let it running and it recovered
    <braunr> consider using a subhurd
    <teythoon> ah right, I ment to set up one anyway, but my first attempts
      were not successful, not sure why
    <teythoon> anyway, is there a way to prevent this in the first place?
    <teythoon> if one could compare ports that'd be helpful
    <youpi> Mmm, did you try to simply compare the number?
    <teythoon> with the bootstrap port I presume?
    <youpi> Mmm, no, the send port and the receive port would be different
    <youpi> no, with the receive port
    <teythoon> ah
    <braunr> comparing the numbers should work
    <braunr> youpi: no they should be the same
    <youpi> braunr: ah, then it should work yes
    <braunr> that's why there are user ref counts
    <youpi> ok
    <braunr> only send-once rights have their own names
    <teythoon> btw, I'll push my work to darnassus from now on,
      e.g. http://darnassus.sceen.net/gitweb/?p=teythoon/hurd.git;a=shortlog;h=refs/heads/feature-mtab-translator-v3-wip


## [[open_issues/libnetfs_passive_translators]]


## IRC, freenode, #hurd, 2013-07-16

    <teythoon> which port is the receive port of a translator? I mean, how is
      it called in the source, there is no port in sight named receive anywhere
      I looked.
    <braunr> teythoon: what is the "receive port of a translator" ?
    <teythoon> braunr: we talked yesterday about preventing the mtab deadlock
      by comparing ports
    <teythoon> I asked which one to use for the comparison, youpi said the
      receive port
    <braunr> i'm not sure what he meant
    <braunr> it could be the receive port used for the RPC
    <braunr> but i don't think it's exported past mig stub code
    <teythoon> weird, I just reread it. I asked if i should use the bootstrap
      port, and he said receive port, but it might have been addressed to you?
    <teythoon> you were talking about send and receive ports being singletons
      or not
    <teythoon> umm
    <braunr> no i answered him
    <braunr> he was wondering if the receive port could actually be used for
      comparison
    <braunr> i said it can
    <braunr> but still, i'm not sure what port
    <braunr> if it's urgent, send him a mail
    <teythoon> no, my pipeline is full of stuff I can do instead ;)
    <braunr> :)


## IRC, freenode, #hurd, 2013-07-17

    <teythoon> braunr: btw, comparing ports solved the deadlock in the mtab
      translator rather easily
    <braunr> :)
    <braunr> which port then ?
    <teythoon> currently I'm stuck though, I'm not sure how to address Neals
      concern wrt to access permission checks
    <teythoon> I believe it's called control port
    <braunr> ok
    <teythoon> the one one gets from doing the handshake with the parent
    <braunr> i thought it was the bootstrap port
    <braunr> but i don't know the details so i may be wrong
    <braunr> anyway
    <teythoon> yes
    <braunr> what is the permission problem again ?
    <teythoon> 871u73j4zp.wl%neal@walfield.org
    <braunr> well, you could perform a lookup on the stored path
    <braunr> as if opening the node
    <teythoon> if I look at any server implementation of a procedure from
      fs.defs (say libtrivfs/file-chmod.c [bad example though, that looks wrong
      to me]), there is permission checking being done
    <teythoon> any server implementation of a procedure from fsys.defs lacks
      permission checks, so I guess it's being done somewhere else
    <braunr> i must say i'm a bit lost in this discussion
    <braunr> i don't know :/
    <braunr> can *you* sum up the permission problem please ?
    <braunr> i mean here, now, in just a few words ?
    <teythoon> ok, so I'm extending the fsys api with the get_children
      procedure
    <teythoon> that one should not return any children x/y if the user doing
      the request has no read permissions on x
    <braunr> really ?
    <braunr> why so ?
    <teythoon> the same way ls x would not reveal the existence of y
    <braunr> i could also say unlike cat /proc/mounts
    <braunr> i can see why we would want that
    <braunr> i also can see why we could let this behaviour in place
    <braunr> let's admit we do want it
    <teythoon> true, but I thought this could easily be addressed
    <braunr> what you could do is
    <teythoon> now I'm not sure b/c I cannot even find the permission checking
      code for any fsys_* function
    <braunr> for each element in the list of child translators
    <braunr> perform a lookup on the stored path on behalf of the user
    <braunr> and add to the returned list if permission checks pass
    <braunr> teythoon: note that i said lookup on the path, which is an fs
      interface
    <braunr> i assume there is no permission checking for the fsys interface
      because it's done at the file (fs) level
    <teythoon> i think so too, yes
    <teythoon> sure, if I only knew who made the request in the first place
    <teythoon> the file-* options have a convenient credential handle passed in
      as first parameter
    <teythoon> s/options/procedures/
    <teythoon> surely the fsys-* procedures also have a means of retrieving
      that information, I just don't know how
    <braunr> mig magic
    <braunr> teythoon: see file_t in hurd_types.defs
    <braunr> there is the macro FILE_INTRAN which is defined in subdirectories
      (or not)
    <teythoon> ah, retrieving the control port requires permissions, and the
      fsys-* operations then operate on the control port?
    <braunr> see libdiskfs/fsmutations.h for example
    <braunr> uh yes but that's for < braunr> i assume there is no permission
      checking for the fsys interface because it's done at the file (fs) level
    <braunr> i'm answering < teythoon> sure, if I only knew who made the
      request in the first place
    <braunr> teythoon: do we understand each other or is there still something
      fuzzy ?
    <teythoon> braunr: thanks for the pointers, I'll read up on that a bit
      later
    <braunr> teythoon: ok


## IRC, freenode, #hurd, 2013-07-18

    <teythoon> braunr: back to the permission checking problem for the
      fsys_get_children interface
    <teythoon> I can see how this could be easily implemented in the mtab
      translator, it asks the translator for the list of children and then
      checks if the user has permission to read the parent dir
    <teythoon> but that is pointless, it has to be implemented in the
      fsys_get_children server function
    <braunr> yes
    <braunr> why is it pointless ?
    <teythoon> because one could circumvent the restriction by doing the
      fsys_get_children call w/o the mtab translator
    <braunr> uh no
    <braunr> you got it wrong
    <braunr> what i suggested is that fsys_get_children does it before
      returning a list
    <braunr> the problem is that the mtab translator has a different identity
      from the users accessing it
    <teythoon> yes, but I cannot see how to do this, b/c at this point I do not
      have the user credentials
    <braunr> get them
    <teythoon> how?
    <braunr> 16:14 < braunr> mig magic
    <braunr> 16:15 < braunr> teythoon: see file_t in hurd_types.defs
    <braunr> 16:16 < braunr> there is the macro FILE_INTRAN which is defined in
      subdirectories (or not)
    <braunr> 16:16 < braunr> see libdiskfs/fsmutations.h for example
    <teythoon> i saw that
    <braunr> is there a problem i don't see then ?
    <braunr> i suppose you should define FSYS_INTRAN rather
    <braunr> but the idea is the same
    <teythoon> won't that change all the function signatures of the fsys-*
      family?
    <braunr> that's probably the only reason not to implement this feature
      right now
    <teythoon> then again, that change is probably easy and mechanic in nature,
      might be an excuse to play around with coccinelle
    <braunr> why not
    <braunr> if you have the time
    <teythoon> right, if this can be done, the mtab translator (if run as root)
      could get credentials matching the users credentials to make that
      request, right?
    <braunr> i suppose
    <braunr> i'm not sure it's easy to make servers do requests on behalf of
      users on the hurd
    <braunr> which makes me wonder if the mtab functionality shouldn't be
      implemented in glibc eheheh ....
    <braunr> but probably not
    <teythoon> well, I'll try out the mig magic thing and see how painful it is
      to fix everything ;)
    <braunr> good luck
    <braunr> honestly, i'm starting to think it's deviating too much from your
      initial goal
    <braunr> i'd be fine with a linux-like /proc/mounts
    <braunr> with a TODO concerning permissions
    <teythoon> ok, fine with me :)
    <braunr> confirm it with the other mentors please
    <braunr> we have to agree quickly on this
    <teythoon> y?

    <teythoon> braunr: I actually believe that the permission issue can be
      addressed cleanly and unobstrusively
    <teythoon> braunr: would you still be opposed to the get_children approach
      if that is solved?
    <teythoon> the filesystem is a tree and the translators "creating" that
      tree are a more coarse version of that tree
    <teythoon> having a method to traverse that tree seems natural to me
    <braunr> teythoon: it is natural
    <braunr> i'm just worried it's a bit too complicated, unnecessary, and
      out-of-scope for the problem at hand
    <braunr> (which is /proc/mounts, not to forget it)


## IRC, freenode, #hurd, 2013-07-19

    <teythoon> braunr: I think you could be a bit more optimistic and
      supportive of the decentralized approach
    <teythoon> I know the dark side has cookies and strong language and it's
      mighty tempting
    <teythoon> but both are bad for you :p


## IRC, freenode, #hurd, 2013-07-22

    <youpi> teythoon: AIUI, you should be able to run the mtab translator as
      no-user (i.e. no uid)
    <teythoon> youpi: yes, that works fine

    <youpi> teythoon: so there is actually no need to define FSYS_INTRAN, doing
      it by hand as you did is fine, right?
    <youpi> (/me backlogs mails...)
    <teythoon> youpi: yes, the main challenge was to figure out what mig does
      and how the cpp is involved
    <youpi> heh :)
    <teythoon> my patch does exactly the same, but only for this one server
      function
    <teythoon> youpi: I'm confused by your mail, why are read permissions on
      all path components necessary?
    <braunr> teythoon: only execution normally
    <youpi> teythoon: to avoid letting a user discover a translator running on
      a hidden directory
    <teythoon> braunr: exactly, and that is tested
    <youpi> e.g. ~/home/foo is o+x, but o-r
    <youpi> and I have a translator running on ~/home/foo/aZeRtYuyU
    <youpi> I don't want that to show up on /proc/mounts
    <braunr> youpi: i don't understand either: why isn't execution permission
      enough ?
    <teythoon> youpi: but that requires testing for read on the *last*
      component of the *dirname* of your translator, and that is tested
    <youpi> let me take another example :)
    <youpi> e.g. ~/home/foo/aZeRtYuyU is o+x, but o-r
    <youpi> and I have a translator running on ~/home/foo/aZeRtYuyU/foo
    <youpi> ergl sorry, I meant this actually:
    <teythoon> yes, that won't show up then in the mtab for users that are not
      you and not root
    <youpi> e.g. ~/home/foo is o+x, but o-r
    <youpi> and I have a translator running on ~/home/foo/aZeRtYuyU/foo
    <teythoon> ah
    <teythoon> hmm, good point
    <braunr> ?
    * braunr still confused
    <teythoon> well, qwfpgjlu is the secret
    <teythoon> and that is revealed by the fsys_get_children procedure
    <braunr> then i didn't understand the description of the call right
    <braunr> > +  /* check_access performs the same permission check as is
      normally                                                                                                                    
    <braunr> > +     done, i.e. it checks that all but the last path components
      are                                                                                                                    
    <braunr> > +     executable by the requesting user and that the last
      component is                                                                                                                  
    <braunr> > +     readable.       */                                                                                                                                                                
    <teythoon> braunr: youpi argues that this is not enough in this case
    <braunr> from that, it looks ok to me
    <youpi> the function and the documentation agree, yes
    <youpi> but that's not what we want
    <braunr> and that's where i fail to understand
    <youpi> again, see my example
    <braunr> i am
    <braunr> 10:43 < youpi> e.g. ~/home/foo is o+x, but o-r
    <braunr> ok
    <youpi> so the user is not supposed to find out the secret
    <braunr> then your example isn't enough to describe what's wron
    <braunr> g
    <youpi> checking read permission only on ~/home/foo/aZeRtYuyU will not
      garantee that
    <braunr> ah
    <braunr> i thought foo was the last component
    <youpi> no, that's why I changed my example
    <braunr> hum
    <braunr> 10:43 < youpi> e.g. ~/home/foo is o+x, but o-r
    <braunr> 10:43 < youpi> and I have a translator running on
      ~/home/foo/aZeRtYuyU/foo
    <braunr> i meant, the last foo
    <teythoon> still, this is easily fixed
    <youpi> sure
    <youpi> just has to be :)
    <teythoon> youpi, braunr: so do you think that this approach will work?
    <youpi> I believe so
    <braunr> i still don't see the problem, so don't ask me :)
    <braunr> i've been sick all week end and hardly slept, which might explain
    <braunr> in the example, "all but the last path components" is
      "~/home/foo/aZeRtYuyU"
    <braunr> right ?
    <youpi> braunr: well, I haven't looked at the details
    <youpi> but be it the last, or but-last  doesn't change the issue
    <youpi> if my ~/hidden is o-r,o+x
    <youpi> and I have a translator on ~/hidden/a/b/c/d/e
    <youpi> checking only +x on hidden is not ok
    <braunr> but won't the call also check a b c d ?
    <youpi> yes, but that's not what matters
    <youpi> what matters is that hidden is o-r
    <braunr> hm
    <youpi> so the mtab translator is not supposed to reveal that there is an
      "a" in there
    <braunr> ok i'm starting to understand
    <braunr> so r must be checked on all components too
    <youpi> yes
    <braunr> right
    <youpi> to simulate the user doing ls, cd, ls, cd, etc.
    <braunr> well, not cd
    <braunr> ah
    <youpi> for being able to do ls, you have to be able to do cd
    <braunr> as an ordered list of commands
    <braunr> ok
    <teythoon> agreed. can you think of any more issues?
    <braunr> so both x and r must be checked
    <youpi> so in the end this RPC is really a shortcut for a find + fsysopts
      script
    <youpi> teythoon: I don't see any
    <braunr> teythoon: i couldn't take a clear look at the patch but
    <braunr> do you perform a lookup on all nodes ?
    <teythoon> yes, all nodes on the path from the root to the one specified by
      the mount point entry in the active translator list
    <braunr> let me rephrase
    <braunr> do you at some point do a lookup, similar to a find, on all nodes
      of a translator ?
    <teythoon> no
    <braunr> good
    <teythoon> yes
    <braunr> iirc, neal raised that concern once
    <teythoon> and I'll also fix settrans --recursive not to iterate over *all*
      nodes either
    <braunr> great
    <braunr> :)
    <teythoon> fsys_set_options with do_children=1 currently does that (I've
      only looked at the diskfs version)


## IRC, freenode, #hurd, 2013-07-27

    <teythoon> youpi: ah, I just found msg_get_init_port, that should make the
      translator detection feasible


## IRC, freenode, #hurd, 2013-07-31

    <teythoon> braunr: can I discover the sender of an rpc message?
    <braunr> teythoon: no
    <braunr> teythoon: what do you mean by "sender" ?
    <teythoon> braunr: well, I'm trying to do permission checks in the
      S_proc_mark_essential server function
    <braunr> ok so, the sending user
    <braunr> that should be doable
    <teythoon> I've got a struct proc *p courtesy of a mig intran mutation and
      a port lookup
    <teythoon> but that is not necessarily the sender, right?
    <braunr> proc is really the server i know the least :/
    <braunr> there is permission checking for signals
    <braunr> it does work
    <braunr> you should look there
    <teythoon> yes, there are permission checks there
    <teythoon> but the only argument the rpc has is a mach_port_t refering to
      an object in the proc server
    <braunr> yes
    <teythoon> anyone can obtain such a handle for any process, no?
    <braunr> can you tell where it is exactly please ?
    <braunr> i don't think so, no
    <teythoon> what?
    <braunr> 14:42 < teythoon> but the only argument the rpc has is a
      mach_port_t refering to an object in the proc server
    <teythoon> ah
    <braunr> the code you're referring to
    <braunr> a common way to give privileges to public objects is to provide
      different types of rights
    <braunr> a public (usually read-only) right
    <braunr> and a privileged one, like host_priv which you may have seen
    <braunr> acting on (modifying) a remote object normally requires the latter
    <teythoon> http://paste.debian.net/20795/
    <braunr> i thought you were referring to existing code
    <teythoon> well, there is existing code doing permission checks the same
      way I'm doing it there
    <braunr> where is it please ?
    <braunr> mgt.c ?
    <teythoon> proc/mgt.c (S_proc_setowner) for example
    <teythoon> yes
    <braunr> that's different
    <teythoon> but anyone can obtain such a reference by doing proc_pid2proc
    <braunr> the sender is explicitely giving the new uid
    <braunr> yes but not anyone is already an owner of the target process
    <braunr> (although it may look like anyone has the right to clear the owner
      oO)
    <teythoon> see, that's what made me worry, it is not checked who's the
      sender of the message
    <teythoon> unless i'm missing something here
    <teythoon> ah
    <teythoon> I am
    <teythoon> pid2proc returns EPERM if one is not the owner of the process in
      question
    <teythoon> all is well
    <braunr> ok
    <braunr> it still requires the caller process though
    <teythoon> what?
    <braunr> see check_owner
    <braunr> the only occurrence i find in the hurd is in libps/procstat.c
    <braunr> MGET(PSTAT_PROCESS, PSTAT_PID, proc_pid2proc (server, ps->pid,
      &ps->process));
    <braunr> server being the proc server AIUI
    <teythoon> yes, most likely
    <braunr> but pid2proc describes this first argument to be the caller
      process
    <teythoon> ah but it is
    <braunr> ?
    <teythoon> mig magic :p
    <teythoon> MIGSFLAGS="-DPROCESS_INTRAN=pstruct_t reqport_find (process_t)"
      \
    <teythoon> MIGSFLAGS="-DPROCESS_INTRAN=pstruct_t reqport_find (process_t)"
      \
    <braunr> ah nice
    <braunr> hum no
    <braunr> this just looks up the proc object from a port name, which is
      obvious
    <braunr> what i mean is
    <braunr> 14:53 < braunr> MGET(PSTAT_PROCESS, PSTAT_PID, proc_pid2proc
      (server, ps->pid, &ps->process));
    <braunr> this is done in libps
    <braunr> which can be used by any process
    <braunr> server is the proc server for this process (it defines the process
      namespace)
    <teythoon> yes, but isn't the port to the proc server different for each
      process?
    <braunr> no, the port is the same (the name changes only)
    <braunr> ports are global non-first class objects
    <teythoon> and the proc server can thus tell with the lookup which process
      it is talking to?
    <braunr> that's the thing
    <braunr> from pid2proc :
    <braunr> S_proc_pid2proc (struct proc *callerp
    <braunr> [...]
    <braunr> if (! check_owner (callerp, p))
    <braunr> check_owner (struct proc *proc1, struct proc *proc2)
    <braunr> "Returns true if PROC1 has `owner' privileges over PROC2 (and can
      thus get its task port &c)."
    <braunr> callerp looks like it should be the caller process
    <braunr> but in libps, it seems to be the proc server
    <braunr> this looks strange to me
    <teythoon> yep, to me too, hence my confusion
    <braunr> could be a bug that allows anyone to perform pid2proc
    <teythoon> braunr: well, proc_pid2proc (getproc (), 1, ...) fails with
      EPERM as expected for me
    <braunr> ofc it does with getproc()
    <braunr> but what forces a process to pass itself as the first argument ?
    <teythoon> braunr: nothing, but what else would it pass there?
    <braunr> 14:53 < braunr> MGET(PSTAT_PROCESS, PSTAT_PID, proc_pid2proc
      (server, ps->pid, &ps->process));
    <braunr> everyone knows the proc server
    <braunr> ok now, that's weird
    <braunr> teythoon: does getproc() return the proc server ?
    <teythoon> I think so, yes
    <teythoon> damn those distributed systems, all of their sources are so
      distributed too
    <braunr> i suspect there is another layer of dark glue in the way
    <teythoon> I cannot even find getproc :/
    <braunr> hurdports.c:GETSET (process_t, proc, PROC)
    <braunr> that's the dark glue :p
    <teythoon> ah, so it must be true that the ports to the proc server are
      indeed process specific, right?
    <braunr> ?
    <teythoon> well, it is not one port to the proc server that everyone knows
    <braunr> it is
    <braunr> what makes you think it's not ?
    <teythoon> proc_pid2proc (getproc (), 1, ...) fails with EPERM for anyone
      not being root, but succeeds for root
    <braunr> hm right
    <teythoon> if getproc () were to return the same port, the proc server
      couldn't distinguish these
    <braunr> indeed
    <braunr> in which case getproc() actually returns the caller's process
      object at its proc server
    <teythoon> yes, that is better worded
    <braunr> teythoon: i'm not sure it's true actually :/
    <teythoon> braunr: well, exploit or it didn't happen
    <braunr> teythoon: getproc() apparently returns a bootstrap port
    <braunr> we must find the code that sets this port
    <braunr> i have a hard time doing that :/
    <pinotree> isn't part of the stuff which is passed to a new process by
      exec?
    <teythoon> braunr: I know that feeling
    <braunr> pinotree: probably
    <braunr> still hard to find ..
    <pinotree> search in glibc
    <teythoon> braunr: exec/exec.c:1654 asks the proc server for the proc
      object to use for the new process
    <teythoon> so how much of hurd do I have to rebuild once i changed struct
      procinfo in hurd_types.h?
    <teythoon> oh noez, glibc uses it too :/


## IRC, freenode, #hurd, 2013-08-01

    <teythoon> I need some pointers on building the libc, specifically how to
      point libcs build system to my modified hurd headers
    <teythoon> nlightnfotis: hi
    <teythoon> nlightnfotis: you rebuild the libc right? do you have any hurd
      specific pointers for doing so?
    <nlightnfotis> teythoon, I have not yet rebuild the libc (I was planning
      to, but I followed other courses of action) Thomas had pointed me to some
      resources on the Hurd website. I can look them up for you
    <nlightnfotis> teythoon, here are the instructions
      http://darnassus.sceen.net/~hurd-web/open_issues/glibc/debian/
    <nlightnfotis> and the eglibc snapshot is here
      http://snapshot.debian.org/package/eglibc/
    <teythoon> nlightnfotis: yeah, I found those. the thing is I changed a
      struct in the hurd_types.h header, so now I want to rebuild the libc with
      that header
    <teythoon> and I cannot figure out how to point libcs build system to my
      hurd headers
    <teythoon> :/
    <nlightnfotis> can you patch eglibc and build that one instead?
    <pochu> teythoon: put your header in the appropriate /usr/include/ dir
    <teythoon> pochu: is there no other way?
    <pinotree> iirc nope
    <pochu> teythoon: you may be able to pass some flag to configure, but I
      don't know if that will work in this specific case
    <teythoon> ouch >,< that explains why I haven't found one
    <pochu> check ./configure --help, it's usually FOO_CFLAGS (so something
      like HURD_CFLAGS maybe)
    <pochu> but then you may need _LIBS as well depending on how you changed
      the header... so in the end it's just easier to put the header in
      /usr/include/
    <braunr> teythoon: did you find the info for your libc build ?
    <teythoon> braunr: well, i firmlinked my hurd_types.h into /usr/include/...
    <braunr> ew
    <braunr> i recommend building debian packages
    <teythoon> but the build was not successful, looks unrelated to my changes
      though
    <teythoon> I tried that last week and the process took more than eight
      hours and did not finish
    <braunr> use darnassus
    <braunr> it takes about 6 hours on it
    <teythoon> I shall try again and skip the unused variants
    <braunr> i also suggest you use ./debian/rules build
    <braunr> and then interrupt the build process one you see it's building
      object files
    <braunr> go to the hurd-libc-i386 build dir, and use make lib others
    <braunr> make lib builds libc, others is for companion libraries lik
      libpthread
    <braunr> actually building libc takes less than an hour
    <braunr> so once you validate your build this way, you know building the
      whole debian package will succedd
    <braunr> succeed*
    <teythoon> so how do I get the build system to pick up my hurd_types.h?
    <braunr> sorry if this is obvious to you, you might be more familiar with
      debian than i am :)
    <braunr> patch the hurd package
    <braunr> append your own version string like +teythoon.hurd.1
    <braunr> install it
    <braunr> then build libc
    <braunr> i'll reboot darnassus so you have a fresh and fast build env
    <braunr> almost a month of uptime without any major issue :)
    <teythoon> err, but I cannot install my hurd package on darnassus, can I? I
      don't think that'd be wise even if it were possible
    <braunr> teythoon: rebooted, enjoy
    <braunr> why not ?
    <braunr> i often do it for my own developments
    <braunr> teythoon: screen is normally available
    <braunr> teythoon: be aware that fakeroot-tcp is known to hang when pfinet
      is out of ports (that's a bug)
    <braunr> it takes more time to reach that bug since a patch that got in
      less than a year ago, but it still happens
    <braunr> the hurd packages are quick to build, and they should only provide
      the new header, right ?
    <braunr> you can include the functionality too in the packages if you're
      confident enough
    <teythoon> but my latest work on the killing of essential processes issues
      involves patching hurd_types.h and that in a way that breaks the ABI,
      hence the need to rebuild the libc (afaiui)
    <braunr> teythoon: yes, this isn't uncommon
    <teythoon> braunr: this is much more intrusive than anything I've done so
      far, so I'm not so confident in my changes for now
    <braunr> teythoon: show me the patch please
    <teythoon> braunr: it's not split up yet, so kind of messy:
      http://paste.debian.net/21403/
    <braunr> teythoon: did you make sure to add RPCs at the end of defs files ?
    <teythoon> yes, I got burned by this one on my very first attempt, you
      pointed out that mistake
    <braunr> :)
    <braunr> ok
    <braunr> you're changing struct procinfo
    <braunr> this really breaks the abi
    <teythoon> yes
    <braunr> i.e. you can't do that
    <teythoon> I cannot put it at the end b/c of that variable length array
    <braunr> you probably should add another interface
    <teythoon> that'd be easier, sure, but this will slow down procfs even
      more, no?
    <braunr> that's secondary
    <braunr> it won't be easier, breaking the abi may break updates
    <braunr> in which case it's impossible
    <braunr> another way would be to ues a new procinfo struct
    <braunr> like struct procinfo2
    <braunr> but then you need a transition step so that all users switch to
      that new version
    <braunr> which is the best way to deal with these issues imo, but this time
      not the easiest :)
    <teythoon> ok, so I'll introduce another rpc and make sure that one is
      extensible
    <braunr> hum no
    <braunr> this usually involves using a version anyway
    <teythoon> no? but it is likely that we need to save more addresses of this
      kind in the future
    <braunr> in which case it will be hanlded as an independant problem with a
      true solution such as the one i mentioned
    <teythoon> it could return an array of vm_address_ts with a length
      indicating how many items were returned
    <braunr> it's ugly
    <braunr> the code is already confusing enough
    <braunr> keep names around for clarity
    <teythoon> ok, point taken
    <braunr> really, don't mind additional RPCs when first adding new features
    <braunr> once the interface is stable, a new and improved version becomes a
      new development of its own
    <braunr> you're invited to work on that after gsoc :)
    <braunr> but during gsoc, it just seems like an unnecessary burden
    <teythoon> ok cool, I really like that way of extending Hurd, it's really
      easy
    <teythoon> and feels so natural
    <braunr> i share your concern about performances, and had a similar problem
      when adding page cache information to gnumach
    <braunr> in the end, i'll have to rework that again
    <braunr> because i tried to extend it beyond what i needed
    <teythoon> true, I see how that could happen easily
    <braunr> the real problem is mig
    <braunr> mig limits subsystems to 100 calls
    <braunr> it's clearly not enough
    <braunr> in x15, i intend to use 16 bits for subsystems and 16 bits for
      RPCs, which should be plenty
    <teythoon> that limit seems rather artificial, it's not a power of two
    <braunr> yes it is
    <teythoon> so let's fix it
    <braunr> mach had many artificial static limits
    <braunr> eh :D
    <braunr> not easy
    <braunr> replies are encoded by taking the request ID and adding 100
    <teythoon> uh
    <braunr> "uh" indeed
    <teythoon> so we need an intermediate version of mig that accepts both
      id+100 and dunno id+2^x as replies for id
    <teythoon> or -id - 1
    <braunr> that would completely break the abi
    <teythoon> braunr: how so? the change would be in the *_server functions
      and be compatible with the old id scheme
    <braunr> how do you make sure id+2^x doesn't conflict with another id ?
    <teythoon> oh, the id is added to the subsystem id?
    <teythoon> to obtain a global message id?
    <braunr> yes
    <teythoon> ah, I see
    <teythoon> ah, but the hurd subsystems are 1000 ids apart
    <teythoon> so id+100 or id +500 would work
    <braunr> we need to make sure it's true
    <braunr> always true
    <teythoon> so how many bits do we have for the message id in mach?
    <teythoon> (mig?)
    <braunr> mach shouldn't care, it's entirely a mig thing
    <braunr> well yes and no
    <braunr> mach defines the message header, which includes the message id
    <braunr> see mach/message.h
    <braunr>     mach_msg_id_t       msgh_id;
    <braunr> typedef integer_t mach_msg_id_t;
    <teythoon> well, if that is like a 32 bit integer, then allow -id-1 as
      reply and forbid ids > 2^x / 2
    <braunr> yes
    <braunr> seems reasonable
    <teythoon> that'd give us an smooth upgrade path, no?
    <braunr> i think so


## IRC, freenode, #hurd, 2013-08-28

    <youpi> teythoon: Mmm, your patch series does not make e.g. ext2fs provide
      a diskfs_get_source, does it?


## IRC, freenode, #hurd, 2013-08-29

    <teythoon> youpi: that is correct
    <youpi> teythoon: Mmm, I must be missing something then: as such the patch
      series introduces an RPC, but only EOPNOTSUPP is ever returned in all
      cases for now?
    <youpi> ah
    <youpi>  /* Guess based on the last argument. */
    <youpi> since ext2fs & such report their options with store last, it seems
      ok indeed
    <youpi> it still seems a bit lame not to return that information in
      get_source
    <teythoon> yes
    <teythoon> well, if it had been just for me, I would not have created that
      rpc, but only guessing was frowned uppon iirc
    <teythoon> then again, maybe this should be used and then the mtab
      translator could skip any translators that do not provide this
      information to filter out non-"filesystem" translators
    <youpi> guessing is usually trap-prone, yes
    <youpi> if it is to be used by mtab, then maybe it should be documented as
      being used by mtab
    <youpi> otherwise symlink would set a source, for instance
    <youpi> while we don't really want it here
    <teythoon> why would the symlink translator answer to such requests? it is
      not a filesystem-like translator
    <youpi> no, but the name & documentation of the RPC doesn't tell it's only
      for filesystem-like translators
    <youpi> well, the documentation does say "filesystem"
    <youpi> but it does not clearly specify that one shouldn't implement
      get_source if one is not a filesystme
    <youpi> "If the concept of a source is applicable" works for a symlink
    <youpi> that could be the same for eth-filter, etc.
    <teythoon> right
    <youpi> Mmm, that said it's fsys.defs
    <youpi> not io.defs
    <youpi> teythoon: it is the fact that we get EOPNOTSUPP (i.e. fsys
      interface supported, just not that call), and not MIG_BAD_ID (i.e. fsys
      interface not supported), that filters out symlink & such, right?
    <teythoon> that's what I was thinking, but that's based on my
      interpretation of EOPNOPSUPP of course ;)
    <youpi> teythoon: I believe that for whatever is a bit questionable, even
      if you put yourself on the side that people will probably agree on, the
      discussion will still take place so we make sure it's the right side :)
    <youpi> (re: start/end_code)
    <teythoon> I'm not sure I follow
    <teythoon> youpi: /proc/pid/stat seems to be used a lot:
      http://codesearch.debian.net/search?q=%22%2Fproc%2F.*%2Fstat%22
    <teythoon> that does not mean that start/endcode is used, but still it
      seems like a good thing to mimic Linux closely
    <youpi> stat is used a lot for cpu usage for instance, yes
    <youpi> start/endcode, I really wonder who is using it
    <youpi> using it for kernel thread detection looks weird to me :)
    <youpi> (questionable): I mean that even if you take the time to put
      yourself on the side that people will probably agree on, the discussion
      will happen
    <youpi> it has to happen so people know they agree on it
    <youpi> I've seen that a lot in various projects (not only CS-related)
    <teythoon> ok, I think I got it
    <teythoon> it's to document the reasons for (not) doing something?
    <youpi> something like this, yes
    <youpi> even if you look right, people will try to poke holes
    <youpi> just to make sure :)
    <teythoon> btw, I think it's rather unusual that our storeio experiments
      would produce such different results
    <teythoon> you're right about the block device, no idea why I got a
      character file there
    <teythoon> I used settrans -ca /tmp/hello.unzipped /hurd/storeio -T
      gunzip:file /tmp/hello
    <teythoon> also I tried stacking the translator on /tmp/hello directly,
      from what I've gathered that should be possible, but I failed
    <teythoon> ftr I use the exec server with all my patches, so the unzipping
      code has been removed from it
    <youpi> ah, I probably still have it
    <youpi> it shouldn't matter here, though
    <teythoon> I agree
    <youpi> how would you stack it?
    <youpi> I've never had a look at that
    <youpi> I'm not sure attaching the translator to the node is done before or
      after the translator has a change to open its target
    <teythoon> right
    <teythoon> but it could be done, if storeio used the reference to the
      underlying node, no?
    <youpi> yes
    <youpi> btw, you had said at some point that you had issues with running
      remap. Was the issue what you fixed with your patches?
    * youpi realizes that he should have shown the remap.c source code during
        his presentation
    <teythoon> well, I tried to remap /servers/exec (iirc) and that failed
    <teythoon> then again, I recently played with remap and all seemed fine
    <teythoon> but I'm sure it has nothing to do with my patches
    <youpi> ok
    <teythoon> those I came up with investigating fakeroot-hurd
    <teythoon> and I saw that this also aplies to remap.sh
    <teythoon> *while
    <youpi> yep, they're basically the same
    <teythoon> btw, I somehow feel settrans is being abused for chroot and
      friends, there is no translator setting involved
    <youpi> chroot, the command? or the settrans option?
    <youpi> I don't understand what you are pointing at
    <teythoon> the settrans option being used by fakeroot, remap and (most
      likely) our chroot
    <youpi> our chroot is just a file_reparent call
    <youpi> fakeroot and remap do start a translator
    <teythoon> yes, but it is not being bound to a node, which is (how I
      understand it) what settrans does
    <teythoon> the point being that if settrans is being invoked with --chroot,
      it does something completely different (see the big if (chroot) {...}
      blocks)
    <teythoon> to a point that it might be better of in a separate command
    <youpi> Mmm, indeed, a lot of the options don't make sense for chroot


## IRC, freenode, #hurd, 2013-09-06

    <braunr> teythoon: do you personally prefer /proc being able to implement
      /proc/self on its own, or using the magic server to tell clients to
      resolve those specific cases themselves ?
    <pinotree> imho solving the "who's the sender of an rpc" could solve both
      the SCM_CREDS implementation and the self case in procfs

[[open_issues/SENDMSG_SCM_CREDS]],
[[hurd/translator/procfs/jkoenig/discussion]], *`/proc/self`*.

    <braunr> pinotree: yes
    <braunr> but that would require servers impersonating users to some extent
    <braunr> and this seems against the hurd philosophy
    <pinotree> and there was also the fact that you could create a
      fake/different port when sending an rpc
    <braunr> to fake what ?
    <pinotree> the sender identiy
    <pinotree> *identity
    <braunr> what ?
    <braunr> you mean intermediate servers can do that
    <teythoon> braunr: I don't know if I understand all the implications of
      your question, but the magic server is the only hurd server that actually
      implements fsys_forward (afaics), so why not use that?
    <braunr> teythoon: my question was rather about the principle
    <braunr> do people find it acceptable to entrust a server with their
      authority or not
    <braunr> on the hurd, it's clearly wrong
    <braunr> but then it means you need special cases everywhere, usually
      handled by glibc
    <braunr> and that's something i find wrong too
    <braunr> it restricts extensibility
    <braunr> the user can always change its libc at runtime, but in practice,
      it's harder to perform than simply doing it in the server
    <teythoon> braunr: then I think I didn't get the question at all
    <braunr> teythoon: it's kind of the same issue that you had with the mtab
      translator
    <braunr> about showing or not some entries the user normally doesn't have
      access to
    <braunr> this problem occurs when there is more than one server on the
      execution path and the servers beyond the first one need credentials to
      reply something meaningful
    <braunr> the /proc/self case is a perfect one
    <braunr> (conceptually, it's client -> procfs -> symlink)
    <braunr> 1/ procfs tells the client it needs to handle this specially,
      which is what the hurd does with magic
    <braunr> 2/ procfs assumes the identity of the client and the symlink
      translator can act as expected because of that
    <braunr> teythoon: what way do you find better ?
    <teythoon> braunr: by "procfs assumes the identity" you mean procfs
      impersonating the user?
    <braunr> yes
    <teythoon> braunr: tbh I still do not see how this can be implemented at
      all b/c the /proc/self symlink is not about identity (which can be
      derived from the peropen struct initially created by fsys_getroot) but
      the pid of the callee (which afaics is nowhere to be found)
    <teythoon> s/callee/caller/
    <teythoon> the one doing the rpc
    <braunr> impersonating the user isn't only about identity
    <braunr> actually, it's impersonating the client
    <teythoon> yes, client is the term >,<
    <braunr> so basically, asking proc about the properties of the process
      being impersonated
    <teythoon> proc o_O
    <braunr> it's not hard, it's just a big turn in the way the system would
      function
    <braunr> teythoon: ?
    <teythoon> you lost me somewhere
    <braunr> the client is the process
    <braunr> not the user
    <teythoon> in order to implement /proc/self properly, one has to get the
      process id of the process doing the /proc/self lookup, right?
    <braunr> yes
    <braunr> actually, we would even slice it more and have the client be a
      thread
    <teythoon> so how do you get to that piece of information at all?
    <braunr> the server inherits a special port designating the client, which
      allows it to query proc about its properties, and assume it's identity in
      servers such as auth
    <braunr> its*
    <teythoon> ah, but that kind of functionality isn't there at the moment, is
      it?
    <braunr> it's not, by design
    <teythoon> right, hence my confusion
    <braunr> instead, servers use the magic translator to send a "retry with
      special handling" message to clients
    <teythoon> right, so the procfs could bounce that back to the libc handler
      that of course knows its pid
    <braunr> yes
    <teythoon> right, so now at last I got the whole question :)
    <braunr> :)
    <teythoon> ugh, I just found the FS_RETRY_MAGICAL handler in the libc :-/
    <braunr> ?
    <braunr> why "ugh" ?
    <teythoon> well, I'm inclined to think this is the bad kind of magic ;)
    <braunr> do i need to look at the code to understand ?
    <teythoon> ok, so I think option 1/ is easily implemented, option 2/ has
      consequences that I cannot fully comprehend
    <braunr> same for me
    <teythoon> no, but you yourself said that you do not like that kind of
      logic being implemented in the libc
    <braunr> well
    <braunr> easily
    <braunr> i'm not so sure
    <braunr> it's easy to code, but i assume checking for magic replies has its
      cost
    <teythoon> why not? the code is doing a big switch over the retryname
      supplied by the server
    <teythoon> we could stuff getpid() logic in there
    <braunr> 14:50 < braunr> it's easy to code, but i assume checking for magic
      replies has its cost
    <teythoon> what kind of cost? computational cost?
    <braunr> yes
    <braunr> the big switch you mentioned
    <braunr> run every time a client gets a reply
    <braunr> (unless i'm mistaken)
    <teythoon> a only for RETRY_MAGICAL replies
    <braunr> but you need to test for it
    <teythoon> 	  switch (retryname[0])
    <teythoon> 	    {
    <teythoon> 	    case '/':
    <teythoon> ...
    <teythoon> that should compile to a jump table, so the cost of adding
      another case should be minimal, no?
    <braunr> yes
    <braunr> but
    <braunr> it's even less than that
    <braunr> the real cost is checking for RETRY_MAGICAL
    <braunr> 14:55 < teythoon> a only for RETRY_MAGICAL replies
    <braunr> so it's basically a if
    <braunr> one if, right ?
    <teythoon> no, it's switch'ing over doretry
    <teythoon> you should pull up the code and see for yourself. it's in
      hurd/lookup-retry.c
    <braunr> ok
    <braunr> well no, that's not what i'm looking for
    <teythoon> it's not o_O
    <braunr> i'm looking for what triggers the call to lookup_retry
    <braunr> teythoon: hm ok, it's for lookups only, that's decent
    <braunr> teythoon: 1/ has the least security implications
    <teythoon> yes
    <braunr> it could slightly be improved with e.g. a well defined interface
      so a user could preload a library to extend it
    <teythoon> extend the whole magic lookup thing?
    <braunr> yes
    <teythoon> but that is no immediate concern, you are trying to fix
      /proc/self, right?
    <braunr> no, i'm thinking about the big picture for x15/propel, keeping the
      current design or doing something else
    <teythoon> oh, okay
    <braunr> solving /proc/self looks actually very easy
    <teythoon> well, I'd say this depends a lot on your trust model then
    <teythoon> do you consider servers trusted?
    <teythoon> (btw, will there be mutual authentication of clients/servers in
      propel?)
    <braunr> there were very interesting discussions about that during the
      l4hurd project
    <braunr> iirc, shapiro insisted that using a server without trusting it
      (and there were specific terminology about trusting/relying/etc..) is
      nonsense
    <braunr> teythoon: i haven't thought too much about that yet, for now it's
      supposed to be similar to what the hurd does
    <teythoon> hm, then again trust is not an on/off thing imho
    <braunr> ?
    <teythoon> trusting someone to impersonate yourself is a very high level of
      trust
    <teythoon> s/is/requires/
    <teythoon> the mobile code paper suggests that mutual authentication might
      be a good thing, and I tend to agree
    <braunr> i'll have to read that again
    <braunr> teythoon: for now (well, when i have time to work on it again
      .. :))
    <braunr> i'm focusing on the low level stuff, in a way that won't disturb
      such high level features
    <braunr> teythoon: have you found something related to a thread-specific
      port in the proc server ?
    <braunr> hurd/process.defs:297: /* You are not expected to understand
      this. */
    <braunr> \o/
    <teythoon> braunr: no, why would I (the thread related question)
    <teythoon> braunr: yes, that comment also cought my eye :/
    <braunr> teythoon: because you read a lot of the proc code lately
    <braunr> so maybe your view of it is better detailed than mine


## IRC, freenode, #hurd, 2013-09-13

    * youpi crosses fingers
    <youpi> yay, still boots
    <youpi> teythoon: I'm getting a few spurious entries in /proc/mounts
    <youpi> none /servers/socket/26 /hurd/pfinet interface=/dev/eth0, etc.
    <youpi> /dev/ttyp0 /dev/ttyp0 /hurd/term name,/dev/ptyp0,type,pty-master 0
      0
    <youpi> /dev/sd1 /dev/cons ext2fs
      writable,no-atime,no-inherit-dir-group,store-type=typed 0 0
    <youpi> fortunately mount drops most of them
    <youpi> but not /dev/cons
    <youpi> spurious entries in df are getting more and more common on linux
      too anyway...
    <youpi> ah, after a console restart, I don't have it any more
    <youpi> I'm getting df: `/dev/cons': Operation not supported instead


## IRC, freenode, #hurd, 2013-09-16

    <youpi> teythoon: e2fsck does not seem to be seeing that a given filesystem
      is mounted
    <youpi> /dev/sd0s1 on /boot type ext2 (rw,no-inherit-dir-group)
    <youpi> and still # e2fsck -C 0 /dev/sd0s1
    <youpi> e2fsck 1.42.8 (20-Jun-2013)
    <youpi> /dev/sd0s1 was not cleanly unmounted, check forced.
    <youpi> (yes, both /etc/mtab and /run/mtab point to /proc/mounts)
    <tschwinge> Yes, that is a "known" problem.
    <youpi> tschwinge: no, it's supposed to be fixed by the mtab translator :)
    <pinotree> youpi: glibc's paths.h points to /var/run/mtab (for us)
    <tschwinge> youpi: Oh.  But this is by means of mtab presence, and not by
      proper locking?  (Which is at least something, of course!)
    <youpi>  /var/run points to /run
    <youpi> tschwinge: yes
    <youpi> anyway, got to run


## IRC, freenode, #hurd, 2013-09-20

    <braunr> teythoon: how come i see three mtab translators running ?
    <braunr> 6 now oO
    <braunr> looks like df -h spawns a few every time
    <teythoon> yes, weird...
    <braunr> accessing /proc/mounts does actually
    <braunr> teythoon: more bug fixing for you :)


## IRC, freenode, #hurd, 2013-09-23

    <teythoon> so it might be a problem with either libnetfs (which afaics has
      never supported passive translator records before) or procfs, but tbh I
      haven't investigated this yet

[[open_issues/libnetfs_passive_translators]].