path: root/open_issues/multithreading.mdwn

[[!meta copyright="Copyright © 2010, 2011, 2012, 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_hurd]]

Hurd servers / VFS libraries are multithreaded.  They can even be said to be
"hyperthreaded".


# Implementation

  * well-known threading libraries

      * [[hurd/libthreads]]

      * [[hurd/libpthread]]

## IRC, freenode, #hurd, 2011-04-20

    <braunr> so basically, a thread should consume only a few kernel resources
    <braunr> in GNU Mach, it doesn't even consume a kernel stack because only
      continuations are used

[[microkernel/mach/gnumach/continuation]].

    <braunr> and in userspace, it consumes 2 MiB of virtual memory, a few table
      entries, and almost no CPU time
    <svante_> What does "hyperthreaded" mean: Do you have a reference?
    <braunr> in this context, it just means there are a lot of threads
    <braunr> even back in the 90s, the expected number of threads could scale
      up to the thousand
    <braunr> today, it isn't much impressive any more
    <braunr> but at the time, most systems didn't have LWPs yet
    <braunr> and a process was very expensive
    <svante_> Looks like I have some catching up to do: What is "continuations"
      and LWP? Maybe I also need a reference to an overview on multi-threading.
    <ArneBab> Lightweight process?
      http://en.wikipedia.org/wiki/Light-weight_process
    <braunr> LWPs are another names for kernel threads usually
    <braunr> most current kernels support kernel preemption though

[[microkernel/mach/gnumach/preemption]].

    <braunr> which means their state is saved based on scheduler decisions
    <braunr> unlike continuations where the thread voluntarily saves its state
    <braunr> if you only have continuations, you can't have kernel preemption,
      but you end up with one kernel stack per processor
    <braunr> while the other model allows kernel preemption and requires one
      kernel stack per thread
    <svante_> I know resources are limited, but it looks like kernel preemption
      would be nice to have. Is that too much for a GSoC student?
    <braunr> it would require a lot of changes in obscure and sensitive parts
      of the kernel
    <braunr> and no, kernel preemption is something we don't actually need
    <braunr> even current debian linux kernels are built without kernel
      preemption
    <braunr> and considering mach has hard limitations on its physical memory
      management, increasing the amount of memory used for kernel stacks would
      imply less available memory for the rest of the system
    <svante_> Are these hard limits in mach difficult to change?
    <braunr> yes
    <braunr> consider mach difficult to change
    <braunr> that's actually one of the goals of my stalled project
    <braunr> which I hope to resume by the end of the year :/
    <svante_> Reading Wikipedia it looks like LWP are "kernel treads" and other
      threads are "user threads" at least in IBM/AIX. LWP in Linux is a thread
      sharing resources and in SunOS they are "user threads". Which is closest
      for Hurd?
    <braunr> i told you
    <braunr> 14:09 < braunr> LWPs are another names for kernel threads usually
    <svante_> Similar to to the IBM definition then? Sorry for not remembering
      what I've  been reading.


# Design

## Application Programs

### [[glibc/signal/signal_thread]]

## Hurd Servers

See [[hurd/libports]]: roughly using one thread per
incoming request.  This is not the best approach: it doesn't really make sense
to scale the number of worker threads with the number of incoming requests, but
instead they should be scaled according to the backends' characteristics.

The [[hurd/Critique]] should have some more on this.

[*Event-based Concurrency
Control*](http://soft.vub.ac.be/~tvcutsem/talks/presentations/T37_nobackground.pdf),
Tom Van Cutsem, 2009.


### IRC, freenode, #hurd, 2012-07-08

    <youpi> braunr: about limiting number of threads, IIRC the problem is that
      for some threads, completing their work means triggering some action in
      the server itself, and waiting for it (with, unfortunately, some lock
      held), which never terminates when we can't create new threads any more
    <braunr> youpi: the number of threads should be limited, but not globally
      by libports
    <braunr> pagers should throttle their writeback requests
    <youpi> right


### IRC, freenode, #hurd, 2012-07-16

    <braunr> hm interesting
    <braunr> when many threads are creating to handle requests, they
      automatically create a pool of worker threads by staying around for some
      time
    <braunr> this time is given in the libport call
    <braunr> but the thread always remain
    <braunr> they must be used in turn each time a new requet comes in
    <braunr> ah no :(, they're maintained by the periodic sync :(
    <braunr> hm, still not that, so weird
    <antrik> braunr: yes, that's a known problem: unused threads should go away
      after some time, but that doesn't actually happen
    <antrik> don't remember though whether it's broken for some reason, or
      simply not implemented at all...
    <antrik> (this was already a known issue when thread throttling was
      discussed around 2005...)
    <braunr> antrik: ok
    <braunr> hm threads actually do finish ..
    <braunr> libthreads retain them in a pool for faster allocations
    <braunr> hm, it's worse than i thought
    <braunr> i think the hurd does its job well
    <braunr> the cthreads code never reaps threads
    <braunr> when threads are finished, they just wait until assigned a new
      invocation

    <braunr> i don't understand ports_manage_port_operations_multithread :/
    <braunr> i think i get it
    <braunr> why do people write things in such a complicated way ..
    <braunr> such code is error prone and confuses anyone

    <braunr> i wonder how well nested functions interact with threads when
      sharing variables :/
    <braunr> the simple idea of nested functions hurts my head
    <braunr> do you see my point ? :) variables on the stack automatically
      shared between threads, without the need to explicitely pass them by
      address
    <antrik> braunr: I don't understand. why would variables on the stack be
      shared between threads?...
    <braunr> antrik: one function declares two variables, two nested functions,
      and use these in separate threads
    <braunr> are the local variables still "local"
    <braunr> ?
    <antrik> braunr: I would think so? why wouldn't they? threads have separate
      stacks, right?...
    <antrik> I must admit though that I have no idea how accessing local
      variables from the parent function works at all...
    <braunr> me neither

    <braunr> why don't demuxers get a generic void * like every callback does
      :((
    <antrik> ?
    <braunr> antrik: they get pointers to the input and output messages only
    <antrik> why is this a problem?
    <braunr> ports_manage_port_operations_multithread can be called multiple
      times in the same process
    <braunr> each call must have its own context
    <braunr> currently this is done by using nested functions
    <braunr> also, why demuxers return booleans while mach_msg_server_timeout
      happily ignores them :(
    <braunr> callbacks shouldn't return anything anyway
    <braunr> but then you have a totally meaningless "return 1" in the middle
      of the code
    <braunr> i'd advise not using a single nested function
    <antrik> I don't understand the remark about nested function
    <braunr> they're just horrible extensions
    <braunr> the compiler completely hides what happens behind the scenes, and
      nasty bugs could come out of that
    <braunr> i'll try to rewrite ports_manage_port_operations_multithread
      without them and see if it changes anything
    <braunr> but it's not easy
    <braunr> also, it makes debugging harder :p
    <braunr> i suspect gdb hangs are due to that, since threads directly start
      on a nested function
    <braunr> and if i'm right, they are created on the stack
    <braunr> (which is also horrible for security concerns, but that's another
      story)
    <braunr> (at least the trampolines)
    <antrik> I seriously doubt it will change anything... but feel free to
      prove me wrong :-)
    <braunr> well, i can see really weird things, but it may have nothing to do
      with the fact functions are nested
    <braunr> (i still strongly believe those shouldn't be used at all)


### IRC, freenode, #hurd, 2012-08-31

    <braunr> and the hurd is all but scalable
    <gnu_srs> I thought scalability was built-in already, at least for hurd??
    <braunr> built in ?
    <gnu_srs> designed in
    <braunr> i guess you think that because you read "aggressively
      multithreaded" ?
    <braunr> well, a system that is unable to control the amount of threads it
      creates for no valid reason and uses global lock about everywhere isn't
      really scalable
    <braunr> it's not smp nor memory scalable
    <gnu_srs> most modern OSes have multi-cpu support.
    <braunr> that doesn't mean they scale
    <braunr> bsd sucks in this area
    <braunr> it got better in recent years but they're way behind linux
    <braunr> linux has this magic thing called rcu
    <braunr> and i want that in my system, from the beginning
    <braunr> and no, the hurd was never designed to scale
    <braunr> that's obvious
    <braunr> a very common mistake of the early 90s


### IRC, freenode, #hurd, 2012-09-06

    <braunr> mel-: the problem with such a true client/server architecture is
      that the scheduling context of clients is not transferred to servers
    <braunr> mel-: and the hurd creates threads on demand, so if it's too slow
      to process requests, more threads are spawned
    <braunr> to prevent hurd servers from creating too many threads, they are
      given a higher priority
    <braunr> and it causes increased latency for normal user applications
    <braunr> a better way, which is what modern synchronous microkernel based
      systems do
    <braunr> is to transfer the scheduling context of the client to the server
    <braunr> the server thread behaves like the client thread from the
      scheduler perspective
    <gnu_srs> how can creating more threads ease the slowness, is that a design
      decision??
    <mel-> what would be needed to implement this?
    <braunr> mel-: thread migration
    <braunr> gnu_srs: is that what i wrote ?
    <mel-> does mach support it?
    <braunr> mel-: some versions do yes
    <braunr> mel-: not ours
    <gnu_srs> 21:49:03) braunr: mel-: and the hurd creates threads on demand,
      so if it's too slow to process requests, more threads are spawned
    <braunr> of course it's a design decision
    <braunr> it doesn't "ease the slowness"
    <braunr> it makes servers able to use multiple processors to handle
      requests
    <braunr> but it's a wrong design decision as the number of threads is
      completely unchecked
    <gnu_srs> what's the idea of creating more theads then, multiple cpus is
      not supported?
    <braunr> it's a very old decision taken at a time when systems and machines
      were very different
    <braunr> mach used to support multiple processors
    <braunr> it was expected gnumach would do so too
    <braunr> mel-: but getting thread migration would also require us to adjust
      our threading library and our servers
    <braunr> it's not an easy task at all
    <braunr> and it doesn't fix everything
    <braunr> thread migration on mach is an optimization
    <mel-> interesting
    <braunr> async ipc remains available, which means notifications, which are
      async by nature, will create messages floods anyway


### IRC, freenode, #hurd, 2013-02-23

    <braunr> hmm let's try something
    <braunr> iirc, we cannot limit the max number of threads in libports
    <braunr> but did someone try limiting the number of threads used by
      libpager ?
    <braunr> (the only source of system stability problems i currently have are
      the unthrottled writeback requests)
    <youpi> braunr: perhaps we can limit the amount of requests batched by the
      ext2fs sync?
    <braunr> youpi: that's another approach, yes
    <youpi> (I'm not sure to understand what threads libpager create)
    <braunr> youpi: one for each writeback request
    <youpi> ew
    <braunr> but it makes its own call to
      ports_manage_port_operations_multithread
    <braunr> i'll write a new ports_manage_port_operations_multithread_n
      function that takes a mx threads parameter
    <braunr> and see if it helps
    <braunr> i thought replacing spin locks with mutexes would help, but it's
      not enough, the true problem is simply far too much contention
    <braunr> youpi: i still think we should increase the page dirty timeout to
      30 seconds
    <youpi> wouldn't that actually increase the amount of request done in one
      go?
    <braunr> it would
    <braunr> but other systems (including linux) do that
    <youpi> but they group requests
    <braunr> what linux does is scan pages every 5 seconds, and writeback those
      who have been dirty for more than 30 secs
    <braunr> hum yes but that's just a performance issue
    <braunr> i mean, a separate one
    <braunr> a great source of fs performance degradation is due to this
      regular scan happenning at the same time regular I/O calls are made
    <braunr> e.G. aptitude update
    <braunr> so, as a first step, until the sync scan is truley optimized, we
      could increase that interval
    <youpi> I'm afraid of the resulting stability regression
    <youpi> having 6 times as much writebacks to do
    <braunr> i see
    <braunr> my current patch seems to work fine for now
    <braunr> i'll stress it some more
    <braunr> (it limits the number of paging threads to 10 currently)
    <braunr> but iirc, you fixed a deadlock with a debian patch there
    <braunr> i think the case was a pager thread sending a request to the
      kernel, and waiting for the kernel to call another RPC that would unblock
      the pager thread
    <braunr> ah yes it was merged upstream
    <braunr> which means a thread calling memory_object_lock_request with sync
      == 1 must wait for a memory_object_lock_completed
    <braunr> so it can deadlock, whatever the number of threads
    <braunr> i'll try creating two separate pools with a limited number of
      threads then
    <braunr> we probably have the same deadlock issue in
      pager_change_attributes btw
    <braunr> hm no, i can still bring a hurd down easily with a large i/o
      request :(
    <braunr> and now it just recovered after 20 seconds without any visible cpu
      or i/o usage ..
    <braunr> i'm giving up on this libpager issue
    <braunr> it simply requires a redesign


### IRC, freenode, #hurd, 2013-02-28

    <smindinvern> so what causes the stability issues?  or is that not really
      known yet?
    <braunr> the basic idea is that the kernel handles the page cache
    <braunr> and writebacks aren't correctly throttled
    <braunr> so a huge number of threads (several hundreds, sometimes
      thousands) are created
    <braunr> when this pathological state is reached, it's very hard to recover
      because of the various sources of (low) I/O in the system
    <braunr> a simple line sent to syslog increases the load average
    <braunr> the solution requires reworking the libpager library, and probably
      the libdiskfs one too, perhaps others, certainly also the pagers
    <braunr> maybe the kernel too, i'm not sure
    <braunr> i'd say so because it manages a big part of the paging policy


### IRC, freenode, #hurd, 2013-03-02

    <braunr> i think i have a simple-enough solution for the writeback
      instability

[[hurd/libpager]].


### IRC, freenode, #hurd, 2013-03-29

    <braunr> some day i'd like to add a system call that threads can use to
      terminate themselves, passing their stack as a parameter for deallocation
    <braunr> then, we should review the timeouts used with libports management
    <braunr> having servers go away when unneeded is a valuable and visible
      feature of modularity


### IRC, freenode, #hurd, 2013-04-03

    <braunr> youpi: i also run without the libports_stability patch
    <braunr> and i'd like it to be removed unless you still have a good reason
      to keep it around
    <youpi> well, the reason I know is mentioned in the patch
    <youpi> i.e. the box becomes unresponsive when all these threads wake up at
      the same time
    <youpi> maybe we could just introduce some randomness in the sleep time
    <braunr> youpi: i didn't experience the problem
    <youpi> well, I did :)
    <braunr> or if i did, it was very short
    <youpi> for the libports stability, I'd really say take a random value
      between timeout/2 and timeout
    <youpi> and that should just nicely fix the issue
    <braunr> ok


## Alternative approaches:

  * <http://www.concurrencykit.org/>

  * Continuation-passing style

      * [[microkernel/Mach]] internally [[uses
        continuations|microkernel/mach/gnumach/continuation]], too.

  * [[Erlang-style_parallelism]]

      * [[!wikipedia Actor_model]]; also see overlap with
        {{$capability#wikipedia_object-capability_model}}.

  * [libtcr - Threaded Coroutine Library](http://oss.linbit.com/libtcr/)

  * <http://monkey.org/~provos/libevent/>

---

See also: [[multiprocessing]].