Hurd servers / VFS libraries are multithreaded. They can even be said to be "hyperthreaded".
Implementation
well-known threading libraries
?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
<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
<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
signal thread
Hurd Servers
See 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 Critique should have some more on this.
Event-based Concurrency Control, 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
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
IRC, freenode, #hurd, 2013-11-30
"Thread storms".
<braunr> if you copy a large file for example, it is loaded in memory, each
page is touched and becomes dirty, and when the file system requests them
to be flushed, the kernel sends one message for each page
<braunr> the file system spawns a thread as soon as a message arrives and
there is no idle thread left
<braunr> if the amount of message is large and arrives very quickly, a lot
of threads are created
<braunr> and they compete for cpu time
<Gerhard> How do you plan to work around that?
<braunr> first i have to merge in some work about pagein clustering
<braunr> then i intend to implement a specific thread pool for paging
messages
<braunr> with a fixed size
<Gerhard> something compareable for a kernel scheduler?
<braunr> no
<braunr> the problem in the hurd is that it spawns threads as soon as it
needs
<braunr> the thread does both the receiving and the processing
<Gerhard> But you want to queue such threads?
<braunr> what i want is to separate those tasks for paging
<braunr> and manage action queues internally
<braunr> in the past, it was attempted to limit the amount ot threads in
servers, but since receiving is bound with processing, and some actions
in libpager depend on messages not yet received, file systems would
sometimes freeze
<Gerhard> that's entirely the task of the hurd? One cannot solve that in
the microkernel itself?
<braunr> it could, but it would involve redesigning the paging interface
<braunr> and the less there is in the microkernel, the better
IRC, freenode, #hurd, 2013-12-03
<braunr> i think our greatest problem currently is our file system and our
paging library
<braunr> if someone can spend some time getting to know the details and
fixing the major problems they have, we would have a much more stable
system
<TimKack> braunr: The paging library because it cannot predict or keep
statistics on pages to evict or not?
<TimKack> braunr: I.e. in short - is it a stability problem or a
performance problem (or both :) )
<braunr> it's a scalability problem
<braunr> the sclability problem makes paging so slow that paging requests
stack up until the system becomes almost completely unresponsive
<TimKack> ah
<TimKack> So one should chase defpager code then
<braunr> no
<braunr> defpager is for anonymous memory
<TimKack> vmm?
<TimKack> Ah ok ofc
<braunr> our swap has problems of its own, but we don't suffer from it as
much as from ext2fs
<TimKack> From what I have picked up from the mailing lists is the ext2fs
just because no one really have put lots of love in it? While paging is
because it is hard?
<TimKack> (and I am not at that level of wizardry!)
<braunr> no
<braunr> just because it was done at a time when memory was a lot smaller,
and developers didn't anticipate the huge growth of data that came during
the 90s and after
<braunr> that's what scalability is about
<braunr> properly dealing with any kind of quantity
<teythoon> braunr: are we talking about libpager ?
<braunr> yes
<braunr> and ext2fs
<teythoon> yeah, i got that one :p
<braunr> :)
<braunr> the linear scans are in ext2fs
<braunr> the main drawback of libpager is that it doesn't restrict the
amount of concurrent paging requests
<braunr> i think we talked about that recently
<teythoon> i don't remember
<braunr> maybe with someone else then
<teythoon> that doesn't sound too hard to add, is it ?
<teythoon> what are the requirements ?
<teythoon> and more importantly, will it make the system faster ?
<braunr> it's not too hard
<braunr> well
<braunr> it's not that easy to do reliably because of the async nature of
the paging requests
<braunr> teythoon: the problem with paging on top of mach is that paging
requests are asynchronous
<teythoon> ok
<braunr> libpager uses the bare thread pool from libports to deal with
that, i.e. a thread is spawned as soon as a message arrives and all
threads are busy
<braunr> if a lot of messages arrive in a burst, a lot of threads are
created
<braunr> libports implies a lot of contention (which should hopefully be
lowered with your payload patch)
<braunr> that contention is part of the scalability problem
<braunr> a simple solution is to use a more controlled thread pool that
merely queues requests until user threads can process them
<braunr> i'll try to make it clearer : we can't simply limit the amout of
threads in libports, because some paging requests require the reception
of future paging requests in order to complete an operation
<teythoon> why would that help with the async nature of paging requests ?
<braunr> it wouldn't
<teythoon> right
<braunr> thaht's a solution to the scalability problem, not to reliability
<teythoon> well, that kind of queue could also be useful for the other hurd
servers, no ?
<braunr> i don't think so
<teythoon> why not ?
<braunr> teythoon: why would it ?
<braunr> the only other major async messages in the hurd are the no sender
and dead name notification
<braunr> notifications*
<teythoon> we could cap the number of threads
<braunr> two problems with that solution
<teythoon> does not solve the dos issue, but makes it less interruptive,
no?
<braunr> 1/ it would dynamically scale
<braunr> and 2/ it would prevent the reception of messages that allow
operations to complete
<teythoon> why would it block the reception ?
<teythoon> it won't be processed, but accepting it should be possilbe
<braunr> because all worker threads would be blocked, waiting for a future
message to arrive to complete, and no thread would be available to
receive that message
<braunr> accepting, yes
<braunr> that's why i was suggesting a separate pool just for that
<braunr> 15:35 < braunr> a simple solution is to use a more controlled
thread pool that merely queues requests until user threads can process
them
<braunr> "user threads" is a poor choice
<braunr> i used that to mirror what happens in current kernels, where
threads are blocked until the system tells them they can continue
<teythoon> hm
<braunr> but user threads don't handle their own page faults on mach
<teythoon> so how would the threads be blocked exactly, mach_msg ?
phread_locks ?
<braunr> probably a pthread_hurd_cond_wait_np yes
<braunr> that's not really the problem
<teythoon> why not ? that's the point where we could yield the thread and
steal some work from our queue
<braunr> this solution (a specific thread pool of a limited number of
threads to receive messages) has the advantage that it solves one part of
the scalability issue
<braunr> if you do that, you loose the current state, and you have to use
something like continuations instead
<teythoon> indeed ;)
<braunr> this is about the same as making threads uninterruptible when
waiting for IO in unix
<braunr> it makes things simpler
<braunr> less error prone
<braunr> but then, the problem has just been moved
<braunr> instead of a large number of threads, we might have a large number
of queued requests
<braunr> actually, it's not completely asynchronous
<braunr> the pageout code in mach uses some heuristics to slow down
<braunr> it's ugly, and is the reason why the system can get extremely slow
when swap is used
<braunr> solving that probably requires a new paging interface with the
kernel
<teythoon> ok, we will postpone this
<teythoon> I'll have to look at libpager for the protected payload series
anyways
<braunr> 15:38 < braunr> 1/ it would dynamically scale
<braunr> + not
<teythoon> why not ?
<braunr> 15:37 < teythoon> we could cap the number of threads
<braunr> to what value ?
<teythoon> we could adjust the number of threads and the queue size based
on some magic unicorn function
<braunr> :)
<braunr> this one deserves a smiley too
<teythoon> ^^
IRC, freenode, #hurd, 2014-03-02
<youpi> braunr: what is the status of the thread storm issue? do you have
pending code changes for this?
<youpi> I was wondering whether to make ext2fs use adaptative locks,
i.e. spin a bit and then block
<youpi> I don't remember whether anybody already did something like that
<braunr> youpi: no i don't
<braunr> youpi: i attempted switch from spin locks to mutexes once but it
doesn't solve the problem
<braunr> switching*
<gg0> found another storm maker:
<gg0> $ dpkg-reconfigure gnome-accessibility-themes
<gg0> aka update-icon-caches /usr/share/icons/HighContrast
<youpi> braunr: ok
Alternative approaches:
Continuation-passing style
- Mach internally uses continuations, too.
-
- Actor model; also see overlap with Wikipedia, object-capability model.
See also: multiprocessing.