[[!meta copyright="Copyright © 2011 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
# IRC, freenode, #hurd, 2011-09-14
Coming from [[translators_set_up_by_untrusted_users]], 2011-09-14 discussion:
<slpz> antrik: I think a tunable option for preventing non-root users from
creating pagers and attaching translators could also be desirable
<antrik> slpz: why would you want to prevent creating pagers and attaching
<tschwinge> Preventing resource exhaustion, I guess.
<slpz> antrik: security and (as tschwinge says) for prevent a rouge pager
from exhausting the system.
<slpz> antrik: without the ability to use translators for non-root users,
Hurd can provide (almost) the same level of resource protection than
See also: [[translators_set_up_by_untrusted_users]],
<braunr> the hurd is about that though
<slpz> there should be also a limit on the number of outstanding requests
that a task can have, and some other easily traceable values
<braunr> port messages queues have limits
<antrik> slpz: anything can exhaust the system. there are much more basic
limits that are missing... and I don't see how translators or pagers are
special in that regard
<slpz> braunr: that's what I said tunable. If I don't share my computer
with untrusted users, I want full functionality. Otherwise, I can enable
<slpz> braunr: but I think those limits are on reception
<braunr> that's a wrong solution
<slpz> antrik: because pagers are external memory objects, and those are
<braunr> compared to what ?
<braunr> and yes, the limit is on the message queue, on reception
<braunr> why is that a problem ?
<slpz> antrik: forbidding the use of translator was for security, to avoid
the problem of traversing an untrusted FS
<slpz> braunr: compared to anonymous memory
<slpz> braunr: because if the limit is on reception, a task can easily do a
DoS against a server
<braunr> hm actually, the problems we have with swap handling is that
anonymous memory is handled in a very similar way as other objects
<slpz> braunr: I want to limit the number of outstanding (unprocessed
messages in queues) requests
<braunr> slpz: the solution isn't about forbidding the use of translators,
but changing common code (libc i guess) not to use them, they can still
<slpz> braunr: that's because, currently, the external page limit is not
<braunr> i'm also not sure about DoS attacks
<braunr> if i'm right, there is often one port for each managed object,
which usually exist per client
<slpz> braunr: yes, that could an option too (for translators, not for
<braunr> i don't see how pagers wouldn't be translators on the hurd
<slpz> braunr: all pagers are translators, but not all translators are
<braunr> so if it works for translators, it also works for pagers
<slpz> braunr: it would fix the security issue, but not the resource
exhaustion problem, with only affects to pagers
<braunr> i just don't see a point in implementing resource limits before
even fixing other fundamental issues
<braunr> the only way to avoid resource exhaustion is resource limits
<antrik> slpz: just not following untrusted translators is much more useful
than forbidding them alltogether
<braunr> and the main problem of mach is resource accounting
<braunr> so first, fix that, using the critique as a starting point
<slpz> braunr: i'm not saying that this should be implemented right now,
i'm just pointing out this possibility
<braunr> i think we're all mostly aware of it
<slpz> braunr: resource accounting, as it's expressed in the critique,
would be wonderful, but it's just too complex IMHO
<braunr> it requires carefully designed changes to the interface yes
<slpz> to the interface, to the internals, to user space tasks...
<braunr> the internals wouldn't be impacted that much
<braunr> user space tasks would mostly include hurd servers
<braunr> if the changes are centralized in libraries, it should be easy to
provide to the servers
# IRC, freenode, #hurd, 2011-09-22
<slpz> antrik: I've also implemented a simple resource control on dirty
pages and changed pageout_scan to free external pages, and only touch
anonymous memory if it's really needed
<slpz> antrik: those combined make the system work better under heavy load
<slpz> antrik: 1.5 GB of RAM and another 1.5 GB of swap helps a lot, too
<antrik> hm... I'm not sure what these things mean exactly TBH... but I
wonder whether some of these could fix the performance degradation (and
ultimate crash) I described recently...
[[/open_issues/default_pager]], [[system performance degradation
<antrik> care to explain them to a noob like me?
<slpz> probably not. During my tests, I've noticed that, at some points,
the system performance starts to degrade, and this doesn't change until
<slpz> but I wasn't able to create a test case to reproduce the bug...
<slpz> antrik: Sure. First, I've changed GNU Mach to:
<slpz> - Classify all pages from data_supply as external, and count them
in vm_page_external_count (previously, this variable was always zero)
<slpz> - Count all pages for which a data_unlock has been requested as
potentially dirty pages
<antrik> there is one important bit I forgot to mention in my recent
report: one "reliable" way to cause growing swap usage is simply
installing a lot of debian packages (e.g. running an apt-get upgrade)
<antrik> some other kinds of I/O also seem to have such an effect, but I
wasn't able to pinpoint specific situations
<slpz> - Establish a limit on how many potentially dirty pages are
allowed. If it's reached, a notification (right now it's just a bogus
m_o_data_unlock, to avoid implementing a new RPC) it's sent to the pager
which has generated the page fault
<slpz> - Establish a hard limit on those dirt pages. If it's reached,
threads asking for a data_unlock are blocked until someone cleans some
pages. This should be improved with a forced pageout, if needed.
<slpz> - And finally, in vm_pageout_scan, run over the inactive queue
searching for clean, external pages, freeing them. If it's not possible
to free enough pages, or if vm_page_external_count is less than 10% of
system's memory, the "normal" pageout is used.
<slpz> I need to clean up things a little, but I want to send a preliminary
patch to bug-hurd ASAP, to have more people testing it.
<slpz> antrik: Do you thing that performance degradation can be related
with the number of threads of your ext2fs translators?
<antrik> slpz: hm... I didn't watch that recently; but in the past, I
observe that the thread count is pretty constant after it reaches
something like 14000 on heavy load...
<antrik> err... wait, 14000 was ports :-)
<antrik> I doubt my system would survive 14000 threads ;-)
<antrik> don't remember thread count... I guess I should start watching
<slpz> antrik: I was thinking that 14000 threads sound like a lot :-)
<slpz> what I know for sure, is that when operating with large files, the
deactivation of all pages of the memory object which is done after every
operation really hurts to performance
<antrik> right now my root FS has 5100 ports and a mere 71 thread... but
then, it's almost freshly booted :-)
<slpz> that's why I've just commented that operation in my code, since it's
not really needed anymore :-)
<slpz> anyway, after submitting all my pending mails to bug-hurd, I'll try
to hunt that bug. Sounds funny.
<antrik> regarding your explanation, I'm still trying to wrap my head
around some of the details. I must admit that I don't remember what
data_unlock does... or maybe I never fully understood it
<antrik> the limit on dirty pages is global?
<slpz> yes, right now it's global
<marcusb> I try to find the old discussion of the thread storm stuff
<marcusb> there was some concern about deadlocks
<slpz> marcusb: yes, because we were talking about putting an static limit
for the server threads of a translators
<slpz> marcusb: and that was wrong (my fault, I was even dumber back then
<marcusb> oh boy digging in old mail is no fun. first I see mistakes in my
english. then I see quite complicated pager stuff I don't ever remember
touching. but there is a patch, and it has my name on it
<marcusb> I think I lost a couple of the early years of my hurd hacking :)
<antrik> hm... I reread the chapter on locking, and it's still above me :-(
<marcusb> not sure what you are talking about, but if there are any
<antrik> marcusb: external pager interface
<marcusb> uuuuh ;)
<antrik> memory_object_lock_request(), memory_object_lock_completed(),
<marcusb> is that from the mach manual?
<antrik> I didn't really understand that part when I first read it a couple
of years ago, and I still don't understand it now :-(
<marcusb> I am sure I didn't understand it either
<marcusb> and maybe I missed my window :)
<marcusb> let's see
<antrik> slpz: what exactly do you mean by "the pager which has generated
the page fault"?
<antrik> marcusb: essentially I'm trying to understand the explanation of
the changes slpz did, but there are several bits totally obscure to me
<slpz> antrik: when a I/O operation is requested to ext2fs, it maps the
object in question to it's own space, and then memcpy's from/to there
<slpz> antrik: so the translator (which is also a pager) is the one who
generates the page fault
<marcusb> antrik: it's important to understand which messages are sent by
the kernel to the manager and which are sent the other way
<marcusb> if the dest port is memory_object_t, that indicates a msg from
kernel to manager. if it is memory_object_control_t, it's a msg from
manager to kernel
<slpz> antrik: m_o_lock_request it's used by the pager to "settle" the
status of a memory object, m_o_lock_completed is the answer from the
kernel when the lock has been completed (only if the client has requested
to be notified), and m_o_data_unlock is a request from the kernel to
change the level of protection for a page (it's called from vm_fault.c)
<marcusb> slpz: but it's not pagers generating page faults, but users of
the memory object on the other side
<antrik> marcusb: well, I think the direction is clear to me... but the
purpose not really :-)
<marcusb> ie a client that mapped a file
<slpz> antrik: in ext2fs, all pages are initially provided to the kernel
(via data_supply) write protected. When a write operation is done over
one of those pages, a page fault it's generated, which sends a
m_o_data_unlock to the pager, which answers (if convenient) which a
page_lock decreasing the protection level
<marcusb> antrik: one use of lock_request is when you want to shut down
cleanly and want to get the dirty pages written back to you from the
<marcusb> antrik: the other thing may be COW strategies
<slpz> marcusb: well, pagers and clients are in the same task for most
translators, like ext2fs
<marcusb> slpz: oh.
<slpz> marcusb: but yes, a read operation in a mmap'ed file would trigger
the fault in a client user task
<marcusb> slpz: I think I forgot everything about pagers :)
<slpz> marcusb: pager-memcpy.c is the key :-)
<marcusb> slpz: what becomes of the fault then? the kernel sees it's a
mapped memory object. will it then talk to the manager or to a pager?
<antrik> slpz: the translator causes the faults itself when it handles
io_read()/io_write() requests I suppose, as opposed to clients accessing
mmap()ed objects which then generate the faults?...
<antrik> ah, that's actually what you already said above :-)
<slpz> marcusb: I'm not sure what do you mean by "manager"...
<marcusb> manager == memory object
<slpz> marcusb: for all external objects, it will ask to their current
<marcusb> slpz: I think I am missing a couple of details, so nevermind.
It's starting to come back to me, but I am a bit afraid of that ;)
<marcusb> what I love about the Hurd is how damn readable the code is
<marcusb> considering it's an object system, it's so much nicer to read
than gtk stuff
<slpz> when you get the big picture, it's actually somewhat fun to see how
data moves around just to fulfill a simple read()
<marcusb> you should make a diagram!
<marcusb> bonus point for animated video ;)
<slpz> marcusb: heh, take a look at the hurd specific parts of glibc... I
cry in pain every time a do that...
<marcusb> slpz: oh yeah, rdwr-internal.
<marcusb> oh man
<marcusb> slpz: funny thing, I just looked at them the other day because of
the security issue
<slpz> marcusb: I think there was one, maybe a slice from someone's
<marcusb> I think I was always confused about the pager/memobj/kernel
<slpz> marcusb: I'm barely able to read Roland's glibc code. I think it's
out of my reach.
<antrik> marcusb: I think part of the problem is confusing terminology
<marcusb> it's good that you are instrumenting the mach kernel to see
what's actually going on in there. it was a black book for me, but neal
too a peek and got a much better understanding of the performance issues
than I ever did
<antrik> when talking about "pager", we usually mean the process doing the
paging; but in mach terminology this actually seems to be the "manager",
while a "pager" is an individual object in the manager process... or
something like that ;-)
<marcusb> antrik: I just never took a look at the big picture. I look at
<marcusb> I knew the tail, ears, and legs of the elephant.
<marcusb> it's a lot of code for a beginner
<antrik> I never understood the distinction between "pager" and "memory
<antrik> maybe "pager" refers to the object in the external pager, while
"memory object" is the part managed in Mach itself?...
<marcusb> memory object is a real object, to which you can send messages.
it's implemented in the server
<antrik> hm... maybe it's the other way around then ;-)
<marcusb> there is also the default pager
<marcusb> I think the pager is just another name for the process that
serves the memory object (default pager == memory object for anonymous
memory == swap)
<marcusb> there is also libpager
<marcusb> and that's a more complicated beast
<antrik> actually, the correct term seems to be "default memory manager"...
<marcusb> from mach's pov
<marcusb> we always called it default pager in the Hurd
<antrik> marcusb: problem is that "pager" is sometimes used in the Mach
documentation to refer to memory object ports IIRC
<marcusb> isn't it defpager executable?
<marcusb> could be
<marcusb> it's the same thing, really
<antrik> indeed, the program implementing the default memory manager is
called "default pager"... so the terminology is really inconsistent
<marcusb> the hurd's pager library is a high level abstraction for mach's
external memory object interface.
<marcusb> i wouldn't worry about it too much
<antrik> I never looked at libpager
<marcusb> you should!
<marcusb> it's an important beast
<antrik> never seemed relevant to anything I did so far...
<antrik> though maybe it would help understanding
<marcusb> it's related to what you are looking now :)