glibc's fd_to_filename

IRC, freenode, #hurd, June 2010

<pochu> is there a way (POSIX or Hurdish) to get the corresponding file
  name for a fd or a hurd port?
<marcusb> there is a way
<pochu> marcusb: which one would that be?
<marcusb> I forgot
<marcusb> there is an implementation in libc
<marcusb> realpath has a similar job
<marcusb> but that's not what I mean
<marcusb> pochu: maybe I am misremembering.  But it was something where you
  keep looking up .. and list that directory, looking for the node with the
  ID of the node you had .. for
<marcusb> maybe it works only for directories
<marcusb> yeah
<marcusb> pochu: check the getcwd() implementation of libc
<marcusb> sysdeps/mach/hurd/getcwd.c
<marcusb> _hurd_canonicalize_directory_name_internal 
* pochu looks
<pochu> marcusb: interesting
<pochu> though that is for dirs, and doesn't seem to be extensible to
  files, as you cannot lookup for ".." under a file
<marcusb> right
<pochu> oh you already said that :)
<marcusb> actually, I am not sure that's correct
<marcusb> it's probably correct, but there is no reason why looking .. up
  on a file couldn't return the directory it's contianed in
<pochu> I don't know the interfaces or the Hurd internals very well yet,
  but it would look strange to me if you could do that
<marcusb> the hurd is strange
<pochu> it sounds like if you could `ls getcwd.c/..` to get
  sysdeps/mach/hurd/ :-)
<marcusb> yep
<pochu> ok. interesting
<marcusb> you wouldn't find "ls foo.zip/.." very strange, wouldn't you?
<pochu> I guess not if `ls foo.zip` listed the contents of foo.zip
<marcusb> there you go
<marcusb> or the other way round: would you be surprised if "cat somedir"
  would work?
<pochu> I think so. if it did, what would it do?
<marcusb> originally, cat dir would list the directory content!
<marcusb> in the old unix times
<pochu> I was surprised the first time I typed `vi somedir` by accident
<marcusb> and some early BSDs 
* pochu feels young :-)
<marcusb> he don't worry, I didn't see those times either
<marcusb> technically, files and directories are implemented in the same
  way in the hurd, they both are objects implementing the fs.defs interface
<marcusb> which combines file and directory operations
<marcusb> of course, files and directories implement those functions
  differently
<antrik> marcusb: do you know why this behavior (cat on directories) was
  changed?

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

* pinotree ponders about sending as RFC his patch for /proc/$pid/maps
<tschwinge> Including a scheme for providing the names of mapped files?
  ;-D
<braunr> that would be really great indeed
<tschwinge> I have not yet researched how Linux does this.  Perhaps store
  the filename used for first opening a file as a string somewhere?
<pinotree> tschwinge: eh, indeed that's lacking in my patch
<braunr> i'm not sure we should aim at doing it the same way
<youpi> I was wondering about having interfaces for naming tasks, threads,
  objects
<youpi> that'd be useful for debugging in general
<braunr> yes
<braunr> i don't think we need to take namespaces into account
<braunr> a simple name or path should be quite enough
<tschwinge> Agreed.  "Just something!"
<tschwinge> So, a Java toString() method for ports.
<tschwinge> ;-)
<braunr> yes
<tschwinge> Oh, and could this also work recursively?  The ext2fs instance
  on /home asks its parent fs about its own path -- can it do that?  (And
  then cache that, most likely?)  Would one get rooted filesnames that way?
<braunr> i really don't think we should link it to the VFS
<braunr> it should merely be a name for debugging
<youpi> yep, same for me
<youpi> I'd say it's the linker's task of just setting a sane name
<braunr> first, keeping it isolated prevents increasing complexity
<braunr> next, it doesn't reduce performance
<tschwinge> youpi: Linker?
<tschwinge> braunr: Ack.
<braunr> yes, ld is the one creating the mappings
<youpi> tschwinge: the one that loads libraries
<tschwinge> Ah, for /proc/*/maps, right.  I've been thinking more globally.

task_get_name, task_set_name RPCs

The following needs to be updated now that task set name has been implemented.

id:"518AA5B0.6030409@verizon.net"

IRC, freenode, #hurd, 2013-05-10

<youpi> tschwinge's suggestion to put names on ports instead of tasks would
  be useful too
<braunr> do you get task ports as easily as you get tasks in kdb ?
<youpi> there is task->itk_self & such
<youpi> or itk_space
<youpi> I don't remember which one is used by userspace
<braunr> i mean
<braunr> when you use the debugger, can you easily find its ports ?
<braunr> the task ports i mean
<braunr> or thread ports or whatever
<youpi> once you have a task, it's a matter of getting the itk_self port
<youpi> s/port/field member/
<braunr> so the debugger provides you with the addresses of the structs
<braunr> right ?
<youpi> yes, that's what we have already
<braunr> then ok
<braunr> bddebian: do that :p
<braunr> hehe
<youpi> see show all thread
<braunr> (haven't used kdb in a long time)
<bddebian> So, adding a name to ports like I did with tasks?
<braunr> remove what you did for tasks
<braunr> move it to ports
<braunr> it's very similar
<braunr> but hm
<braunr> i'm not sure where the RPC would be
<braunr> this RPC would exist for *all* ports
<braunr> or only for kernel objects if added to gnumach.defs
<youpi> it's just about moving the char array field to another structure
<youpi> and plugging that
<bddebian> But mach_task_self is a syscal, it looks like itk_self is just a
  pointer to an ipc_port ?
<braunr> so ?
<braunr> you take that pointer and you get the port
<braunr> just like vm_map gets a struct vm_map from a task
<bddebian> So I am just adding ipc_port_name to the ipc_port struct in this
  case?
<braunr> yes
<braunr> actually
<braunr> don't do anything just yet
<braunr> we need to sort a few details out first
<braunr> see bug-hurd

IRC, freenode, #hurd, 2013-12-05

<teythoon> braunr: no more room for vm_map_find_entry in 80220a40
<teythoon> 80220a40 <- is that a task ?
<braunr> or a vm_map, not sure
<braunr> probably a vm_map
<teythoon> hm
<teythoon> let's fix this kind of reporting
<braunr> :)
<teythoon> let one process register for kernel log messages
<teythoon> make a rich interface, say klog_thread and friends
<teythoon> a userspace process gets the port name, looks it up in proc,
  logs nicely to syslog
<teythoon> if noone registered for this notifications, fall back to the old
  reporting
<braunr> i tend to think using internal names is probably better
<teythoon> how would i use them to see wich process caused the issue ?
<braunr> you give the name of the task
<braunr> (which means tasks have names, yes)
<teythoon> ok
<braunr> the reason is that reporting is often used for debugging
<braunr> and debugging usually means there is a bug
<braunr> if the bug prevents from reporting, it's not very useful
<braunr> and we're talking about the kernel here, the low level stuff
<teythoon> incidentally, i got myself a stuck process
<teythoon> ah, got it killed
<teythoon> braunr: so you propose to add a task rpc to set a name ?
<braunr> i don't want to push such things
<braunr> which is why this hasn't been done until now
<braunr> but that's what i'd do in x15, yes
<teythoon> y not ?
<braunr> and instead of a process registered to gather kernel messages, i'd
  use a dmesg-like interface, where the kernel manages its message buffer
  itself
<braunr> i didn't feel the need to
<braunr> the tools i've had until now were sufficient
<braunr> don't forget you still need to fix mtab :p
<braunr> or is it done ?
<teythoon> i sometimes see tasks deallocating invalid ports
<teythoon> no
<teythoon> there is an un-acked patche series on the list
<braunr> ok
<teythoon> so, i want to identify which process caused it
<teythoon> is that possible right now ?
<braunr> not easily, no
<teythoon> so that's a valid use case
<braunr> it is
<teythoon> good
<teythoon> :)
<teythoon> so proc would register a string describing each task and mach
  would use this for printing nicer messages ?
<braunr> for example, yes
<braunr> one problem with that approach is that it doesn't fit well with
  subhurds
<teythoon> *bingbingbing
<braunr> but i personally wouldn't care much, they're kernel messages
<braunr> in the future, we could make mach more a hypervisor, and register
  names for each domains
<teythoon> yet unanswered proposal about hierachical proc servers on the
  list...
<teythoon> that'd also fix subhurds, so that the parents processes won't
  appear in the subhurd
<teythoon> making it sandboxier
<teythoon> and killall5 couldn't slaughter the host system if the subhurd
  shuts down with sysvinit

IRC, freenode, #hurd, 2014-01-20

<teythoon> i wonder if it would not be best to add a description to mach
  tasks
<braunr> i think it would
<teythoon> to aid fixing these kind of issues
<braunr> in x15, i actually add descriptions (names) to all kernel objects
<teythoon> that's probably a good idea, yes
<braunr> well, not all, but many

IRC, OFTC, #debian-hurd, 2014-02-05

<teythoon> youpi: about that patch implementing task_set_name, may i merge
  the amended version ?
<youpi> yes

IRC, freenode, #hurd, 2011-07-13

A related issue:

<braunr> rbraun@nordrassil:~$ vminfo $$ | wc -l
<braunr> 1039
<braunr> any idea why a shell would consume more than 1039 map entries ?
<braunr> (well, not more actually)
<braunr> even the kernel and ext2fs have around 100
<braunr> (the kernel has actually only 23, which is very good and expected)
<tschwinge> braunr: I agree that having some sort of debugging information
  for memory objects et al. would be quite hand.  To see where they're
  coming from, etc.
<braunr> tschwinge: this would require naming objects at the mach level
<braunr> e.g. when creating an object
<braunr> giving it the path of a file for example
<tschwinge> braunr: I have recently seen something (due to youpi fixing a
  bug) that bash is doing its own memory management.  Perhaps all these are
  such regions?
<tschwinge> braunr: For example, yes.
<braunr> what ?
<braunr> ?!
<tschwinge> braunr:
  http://lists.gnu.org/archive/html/bug-bash/2011-04/msg00097.html
<braunr> i see

Also see email thread starting at id:"20110714082216.GA8335@sceen.net".

Justus: Once these patches are merged, there will be a way to map from ports to file names, at least for libdiskfs and libnetfs, one would only have to make this information available somehow.