Re: [Qemu-devel] Re: [RESEND][PATCH] gdbstub: Add vCont support

[Jan Kiszka]

Paul Brook wrote:
> On Friday 16 January 2009, Jan Kiszka wrote:
>> Paul Brook wrote:
>>>>  a) Modeling cpus as processes buys us nothing compared to threads given
>>>>     the fact that we cannot provide a stable memory mapping to the gdb
>>>>     frontend anyway. (*)
>>> I disagree. The process model fits perfectly. The whole point is that
>>> each CPU has its own virtual address space. Separate address spaces is
>>> the fundamental difference between a process and a thread.
>>>
>>> If you have a multicore system where several cores share a MMU[1] then
>>> modelling these as threads probably make sense.
>>>
>>> Don't confuse this with OS awareness in GDB (i.e. implementing a
>>> userspace debug environment via a bare metal kernel level debug
>>> interface). That's a completely separate issue.
>> You snipped away my argument under (*):
>>> (*) Process abstraction is, if used at all, guest business. At best we
>>> could try to invent (likely OS-specific) heuristics to identify
>>> identical mappings and call them processes. I don't see a reasonable
>>> benefit compared to the expected effort and unreliability.
> 
> You're doing exactly what I said not do do. You are confusing a GDB process 
> model (i.e. separate address spaces) with actual OS processes.

Well, if gdb can tell them apart... My concern is that gdb's process
model implementation was likely not done with the qemu scenario in mind
(otherwise we would talk about "gdb thread model vs. address space model").

> 
> The point is that each CPU has its own distinct virtual address space. GDB 
> assumes that all threads use the same memory map. To handle these distinct 
> address spaces you need to model CPUs as processes. Your thread hack is 
> dependent on, and limited to, address ranges that happen to mapped the same 
> on all CPUs. This may be sufficient for simple linux kernel debugging, but 
> fails very rapidly when you start trying to do anything clever.

So far it hasn't because the thread model implementation is a bit
smarter than a "hack".

However, you still don't tell us what prevents to _add_ the multiprocess
support to qemu once gdb or some other debugger can handle it according
to our needs. It is an _addon_ feature to the gdb protocol anyway, and
qemu will have to check if the frontend supports it. If not, qemu could
fall back to the thread model to allow at least "basic" debugging and
would not have to tell the frontend "sorry, I'm too smart for you".

> 
>> You cannot simply assign some CPU n to a virtual process n because the
>> mapping you see on that CPU at some point in time may next pop up on
>> some other CPU - and vice versa. You start to make gdb believe it sees
>> consistent processes while you have no clue at all about the scheduling
>> of your guest. So what do you gain?
> 
> Mapping current CPU state to an OS process is the debugger's problem. A 
> sufficiently smart debugger is able to interpret the OS process table, and 
> match that to the address space that's present on a particular CPU. I don't 
> think GDB is capable of doing this right now, but I've seen it done in other 
> debuggers.

So this is still a future requirement on qemu, no?

Let's assume we had multiprocess support implemented in qemu already.
Would I still be able to debug the "corner case" Linux without
additional pain?

> 
>> I can tell you what you loose: If gdb starts to think that there are n
>> separate processes, you will have to set separate breakpoints as well.
>> Bad. And if some breakpoint assigned to process n suddenly hits you on
>> process (CPU) m, only chaos is the result. E.g. kvm would re-inject it
>> as guest-originated. Even worse.
> 
> You need to use multi-process breakpoints.

Does gdb provide them already? Or is there some other free debuggers
that do so?

> 
> An OS aware debugger will take care of migrating userspace breakpoints when 
> OS 
> context switches occur.
> 
>> There was zero practical need for this so far. But maybe you can
>> describe a _concrete_ debugging use case and the related steps in gdb -
>> based on a potential process interface. I'm surely willing to learn
>> about it if there is really such a great practical improvement feasible.
>> What will be the user's benefit?
> 
> In embedded systems it's fairly common to run entirely separate operating 
> systems on different cores. e.g. you may run linux on one core and a RTOS or 
> dedicated codec engine on the other.

Accepted, that can be a valid use case.

Right now the debugger could already handle this scenario /wrt soft
breakpoints by leaving the focus on those CPUs for which it has the code
image.

> 
> The rest of my examples are hypothetical. I don't have actual examples, 
> however it's things that I know people are interested in, and in some cases 
> actively working on. GDB already has a python plugin interface, though 
> currently there aren't any useful hooks for doingn OS awareness stuff.
> 
> I'm not sure offhand how highmem mappings work. Are these per-cpu or global?
> My guess is linux makes them global, but I can certainly imagine them being 
> per-cpu.

highmem is global. I'm not sure if anyone will bother changing this in
the future. Multicore CPU need an increasing amount of memory, thus
generally come with a sufficiently large virtual address spaces.


I admit: If you add OS awareness to the debugger or if you have a fixed
CPU schedule, the multiprocess protocol extension can be useful as it
allows to direct break/watchpoints to certain CPUs.

My suggestion still stands: Add multiprocess support on top of the
thread model, but keep threads as foundation. The former is upcoming
technology while the latter already covers most of todays' kernel
debugging scenarios without problems.

But I do have my problems with providing interfaces _only_ for users
still under development, neglecting what is already possible with
todays' versions. Right now we face a similar issue with x86 due to gdb
not yet being as far as a certain change to gdbstub assumed. I'm working
on this, but it was confirmed by gdb people that there is easy quick
fix. And that currently leaves qemu unusable for certain scenarios
unless you patch it.

So please focus on what is available today - and please let us merge
this patch. vCont is mandatory for the multiprocess feature anyway, just
the thread-id encoding will vary.

Jan

signature.asc
Description: OpenPGP digital signature