Re: [uml-user] Info patches please

[Blaisorblade <bla...@ya...>]

On Saturday 14 October 2006 09:10, Flavio wrote:
> 2006/10/14, Blaisorblade <bla...@ya...>:
> > Since UML has been integrated in mainline, it is not any more correct to
> > talk
> > about "UML patch" - that's misleading for people (including Jeff Dike).
>
> I'm sorry, probably you misunderstood me. I didn't want to mislead anyone.
> I simply want to explain what I was doing. I'm writing on my thesis that I
> use a patch
> on vanilla sources, to get an executable guest kernel. I only want motiva=
te
> my actions.
> Why don't write more on what am I doing? If someone ask me "Ok, let's
> explain better about
> your work with patching the guest kernel, what does really do that patch?
> etc...", What can
> I reply?


> I found some information on http://www.honeynet.org/papers/uml/ where I
> read:
>
> Unlike VMware, UML does not require any additional virtualization softwar=
e.
>
> > Instead, you patch the source of the Linux kernel you want to run as yo=
ur
> > Guest OS. This UML patch converts the kernel into a executable binary
> > called 'linux', which allows the Guest kernel to run on your system as a
> > seperate operating system. When you run this UML patched kernel, all you
> > need to do is give it a filesystem to use, and you now have a independe=
nt
> > Linux system running on your computer, two for the price of one! This n=
ew
> > kernel is a userspace application running on the real kernel (Host OS).
> > The UML kernel receives system calls from its applications and
> > sends/requests them to the Host kernel. There are also additional
> > management and networking UML tools you can install on the computer that
> > makes your life easier.
>
> Is it right, when expecially thalks about uml patch??? It says:  "This UML
> patch converts the kernel into a executable binary called 'linux', which
> allows the Guest kernel to run on your system as a seperate operating
> system."

> You said guest kernel works also without patch, but it's slower. Ok, but,
> it seems (reading above sentence) the patch is applied to get an executab=
le
> binary kernel image at the end. Does it means if I don't apply uml patch,=
 I
> cannot get an executable binary kernel even if I do "make linux ARCH=3Dum"
> ???

That was the correct meaning, and the sentence was correct until about 2.6.=
9=20
kernels, but it is now wrong. You can now get a UML kernel from a kernel.or=
g=20
tree with "make ARCH=3Dum".

Ok, let's explain from scratch.
The howto's you found talk about a guest/UML patch. That is now included in=
=20
kernel.org sources; -bs*/-bb* trees have a meaning similar to the -stable=20
tree* (i.e. they deploy some fixes without waiting for a full release cycle=
),=20
even if -bs hasn't very strict rules. Ideally, everything would go=20
in -stable. So the guest -bs patch is simply a set of bugfixes.

* releases like 2.6.18.1 are from the -stable tree.

Also, they talk about a SKAS patch. That is applied to the host kernel, and=
=20
guest kernel transparently detects and uses it.

> parse error - what do you mean? What is the alternative to patches? You

> > mean "why don't you run UML on vanilla hosts"? The answer is "we can, b=
ut
> > it's slower".

> Ok, thanks...  but.. that's all???
No, but it's technical to explain and I didn't have (and don't have) the ti=
me,=20
and I didn't know if that was the exact answer.

> > - Why do they work?
> >
> > Sorry, but the answer to this question is because they were coded by
> > competent
> > programmers, and I can't find your real question (I'm not perfect at
> > mind-reading ;-)).

> Ok, don't warry. I only want to know a simple information: what is the
> foundation of
> patch functioning? (qual'=E8 il principio di funzionamento che sta alla b=
ase
> delle patch?)
Ok, "how do UML code and SKAS patch work? What's the basic principle?"

UML in itself is based on creating a thread/process for each guest process.=
=20
After a fork on the guest a new "process" is created on the host, while aft=
er=20
a pthread_create() (or a clone(CLONE_VM)) a host thread is created, i.e. th=
ey=20
share the memory space or not.

UML then uses ptrace() (the API used by gdb and strace) to intercept system=
=20
calls and page faults done by the process; when it executes pure userspace=
=20
code it works exactly like a host process (no layer is interposed between t=
he=20
code and the CPU). When a system call is done, UML intercepts and blocks it=
,=20
providing the return value after performing requested work. Page faults are=
=20
also intercepted and resolved by intercepting SIGSEGV signals.

Currently after installing mappings on guest page tables, they're made=20
executive by calling mmap()/munmap()/mprotect() on a file representing=20
virtual RAM (the one placed in /dev/shm) - Ingo Molnar and I have written=20
faster APIs to do it and code for UML to use it.

Two problems arise:
a) how does UML execute mmap() and friends for another process?
b) how does UML gather the address for which SIGSEGV is delivered and the=20
exact fault mode?

The SKAS3 patch adds system calls to the host to solve both problems;=20
additionally it allows collapsing all threads representing guest processes=
=20
into a single one, since on an UniProcessor(UP) UML they never execute=20
together, but only one at a time.

When SKAS3 patch is not applied, a different solution has been devised by m=
e=20
and Jeff, called SKAS0: the code to do this is run inside the thread - it=20
is "copied" (remapped actually) there, then when needed we modify EIP via=20
ptrace() and redirect execution there.
=2D-=20
Inform me of my mistakes, so I can keep imitating Homer Simpson's "Doh!".
Paolo Giarrusso, aka Blaisorblade
http://www.user-mode-linux.org/~blaisorblade
Chiacchiera con i tuoi amici in tempo reale! 
 http://it.yahoo.com/mail_it/foot/*http://it.messenger.yahoo.com